Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency during autoregressive generation.Speculative Decoding (SD) mitigates this by applying a draft-and-verify paradigm, yet existing methods are constrained by rigid exact-match rules, severely limiting the acceleration potential.To bridge this gap, we propose LVSPEC, the first training-free loosely SD framework tailored for Video-LLMs.Grounded in the insight that generation is governed by sparse visual-relevant anchors (mandating strictness) amidst abundant visualirrelevant fillers (permitting loose verification), LVSPEC employs a lightweight visual-relevant token identification scheme to accurately pinpoint the former.To further maximize acceptance, we augment this with a position-shift tolerant mechanism that effectively salvages positionally mismatched but semantically equivalent tokens.Experiments demonstrate that LVSPEC achieves high fidelity and speed: it preserves >99.8% of target performance while accelerating Qwen2.5-VL-32B by 2.70× and LLaVA-OneVision-72B by 2.94×.Notably, it boosts the mean accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.Code is at https://github.com/zju-jiyicheng/LVSpec.

Accelerating the inference of large language models (LLMs) has been a critical challenge in generative AI.Speculative decoding (SD) substantially improves LLM inference efficiency.However, its utility is limited by a fundamental constraint: the draft and target models must share the same vocabulary, thus limiting the herd of available draft models and often necessitating the training of a new model from scratch.Inspired by Dynamic Time Warping (DTW), a classic algorithm for aligning time series, we propose the algorithm TokenTiming for universal speculative decoding.It operates by re-encoding the draft token sequence to get a new target token sequence, and then uses DTW to build a mapping to transfer the probability distributions for speculative sampling.Benefiting from this, our method accommodates mismatched vocabularies and works with any off-the-shelf models without re-training and modification.We conduct comprehensive experiments on various tasks, demonstrating 1.57× speedup.This work enables a universal approach for draft model selection, making SD a more versatile and practical tool for LLM acceleration.The code is available at the link.

Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches.Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs.A common solution is to compress caches under a fixed allocated budget at different granularities: tokenlevel uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads.Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies.We propose HYBRIDKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval.Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7Bshow that HYBRIDKV reduces KV cache memory by up to 7.9× and achieves 1.52× faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.

Recently, at Xiaohongshu, the rapid expansion of e-commerce and advertising demands real-time business analytics with high accuracy and low latency. To meet this demand, systems typically rely on converting natural language (NL) queries into Domain-Specific Languages (DSLs) to ensure semantic consistency, validation, and portability. However, existing multi-stage LLM pipelines for this NL-to-DSL task suffer from prohibitive latency, high cost, and error propagation, rendering them unsuitable for enterprise-scale deployment. In this paper, we propose RedParrot, a novel NL-to-DSL framework that accelerates inference via a semantic cache. Observing the high repetition and stable structural patterns in user queries, RedParrot bypasses the costly pipeline by matching new requests against cached "query skeletons" (normalized structural patterns) and adapting their corresponding DSLs. Our core technical contributions include (1) an offline skeleton construction strategy, (2) an online, entity-agnostic embedding model trained via contrastive learning for robust matching, and (3) a heterogeneous Retrieval-Augmented Generation (RAG) method that integrates diverse knowledge sources to handle unseen entities. Experiments on six real enterprise datasets from Xiaohongshu show RedParrot achieves an average 3.6x speedup and an 8.26% accuracy improvement. Furthermore, on new public benchmarks adapted from Spider and BIRD, it boosts accuracy by 34.8%, substantially outperforming standard in-context learning baselines.

Deep neural network (DNN) models are crucial for Internet-of-Things (IoT) applications. In a multi-access edge computing (MEC) system, IoT devices dispatch their requests to a nearby edge server to accelerate DNN inference. Recently, processing-in-memory (PIM) accelerators have emerged for DNN inference due to their power and area efficiency. However, current scheduling methods primarily focus on single-DNN workloads and do not fully utilize PIM’s potential for managing multi-DNN workloads in MEC scenarios. Additionally, PIM accelerators face challenges such as inflexible resource allocation and costly write operations, constraining the direct application of GPU scheduling techniques. In this paper, we define the multi-DNN scheduling problem on the PIM accelerator. To reduce the problem’s complexity, we propose a heuristic optimization algorithm, called Leader-Follower. Using this algorithm, we present PimShare, a scheduler for the PIM accelerator to process multi-DNN workloads. PimShare schedules multi-DNN inference requests to enable both temporal and spatial multiplexing, which can enhance hardware utilization. Compared with baseline scheduling methods, including those adapted from GPU scheduling, PimShare can achieve up to a two-order-of-magnitude improvement in throughput and scale the high throughput to workloads containing 18 DNN models. In addition, PimShare reduces write operations and processing latency.

Large language models (LLMs) have significantly advanced the performance of various natural language processing tasks, including text-to-SQL. Current LLM-based text-to-SQL schemes mainly focus on improving the understanding of natural language questions (NLQs) or refining the quality of generated SQLs. While these strategies are effective, they often address specific, nuanced aspects. In contrast, humans approach text-to-SQL with a holistic view, applying transitional logical reasoning across multiple steps to arrive at the final answer. We believe LLMs can leverage human cognitive processes to achieve greater accuracy in text-to-SQL. In this paper, we present COGSQL, a framework featuring a suite of tailored models and strategies aimed at replicating human cognitive processes for enhanced LLM-based text-to-SQL. COGSQL consists of three key modules: (1) SQL preparation: we employ a coarse-to-fine schema linking and syntax keyword prediction, akin to how human recall and align key concepts for better understanding. (2) SQL generation: we introduce a concept-enhanced chain-of-thought prompting, enhancing NLQ interpretation and SQL composition of LLMs, similar to humans drafting SQL query. (3) SQL correction: we develop NLQ consistency and result consistency techniques to correct various errors, mirroring how humans evaluate and refine reasoning. We conduct extensive experiments using diverse benchmarks and LLMs. The results and analysis verify the effectiveness and generalizability of COGSQL.

Video large language models (Vid-LLMs) have shown strong capabilities in understanding video content.However, their reliance on dense video token representations introduces substantial memory and computational overhead in both prefilling and decoding.To mitigate the information loss of recent video token reduction methods and accelerate the decoding stage of Vid-LLMs losslessly, we introduce SPECVLM, a training-free speculative decoding (SD) framework tailored for Vid-LLMs that incorporates staged video token pruning.Building on our novel finding that the draft model's speculation exhibits low sensitivity to video token pruning, SPECVLM prunes up to 90% of video tokens to enable efficient speculation without sacrificing accuracy.To achieve this, we perform a two-stage pruning process: Stage I selects highly informative tokens guided by attention signals from the verifier (target model), while Stage II prunes the remaining redundant ones in a spatially uniform manner.Extensive experiments on four video understanding benchmarks demonstrate the effectiveness and robustness of SPECVLM, which achieves up to 2.68× decoding speedup for LLaVA-OneVision-72B and 2.11× speedup for Qwen2.5-VL-32B.Code is available at https: //github.com/zju-jiyicheng/SpecVLM.

Data analysts nowadays are keen to have analytical capabilities involving deep learning (DL). Prediction queries, which combine relational operations with DL models to analyze multi-modal data, provide a powerful facility for smart in-database analysis. However, loose integration systems, which support such queries via User-Defined Functions (UDFs) and external runtimes, often impose high economic costs, particularly in cloud-based environments; while tight integration systems, which implement model inference through a sequence of explic-itly written SQL queries, incur heavy user burdens and huge optimization space. In this paper, we introduce PEPS, an end-to-end analytical database for automatic and efficient prediction query processing. PEPS automates the process of prediction query synthesis and ensures usability through declarative schemes. Additionally, it improves query performance with offline optimization using the DB-oriented Computation Graph Optimization (DBCGO) algorithm and online optimization via heuristic query rewriting. Empirical evaluations show that PEPS offers better usability than baseline methods, provides lower economic costs, and achieves performance speedup compared to advanced Python UDFs.

Large Language Models (LLMs) have significantly advanced natural language processing with exceptional task generalization capabilities. Low-Rank Adaption (LoRA) offers a cost-effective fine-tuning solution, freezing the original model parameters and training only lightweight, low-rank adapter matrices. However, the memory footprint of LoRA is largely dominated by the original model parameters. To mitigate this, we propose LoRAM, a memory-efficient LoRA training scheme founded on the intuition that many neurons in over-parameterized LLMs have low training utility but are essential for inference. LoRAM presents a unique twist: it trains on a pruned (small) model to obtain pruned low-rank matrices, which are then recovered and utilized with the original (large) model for inference. Additionally, minimal-cost continual pre-training, performed by the model publishers in advance, aligns the knowledge discrepancy between pruned and original models. Our extensive experiments demonstrate the efficacy of LoRAM across various pruning strategies and downstream tasks. For a model with 70 billion parameters, LoRAM enables training on a GPU with only 20G HBM, replacing an A100-80G GPU for LoRA training and 15 GPUs for full fine-tuning. Specifically, QLoRAM implemented by structured pruning combined with 4-bit quantization, for LLaMA-3.1-70B (LLaMA-2-70B), reduces the parameter storage cost that dominates the memory usage in low-rank matrix training by 15.81× (16.95×), while achieving dominant performance gains over both the original LLaMA-3.1-70B (LLaMA-2-70B) and LoRA-trained LLaMA-3.1-8B (LLaMA-2-13B). Code is available at https://github.com/junzhang-zj/LoRAM.

With the widespread adoption of Mixture-of-Experts (MoE) models, there is a growing demand for efficient inference on memory-constrained devices. While offloading expert parameters to CPU memory and loading activated experts on demand has emerged as a potential solution, the large size of activated experts overburdens the limited PCIe bandwidth, hindering the effectiveness in latency-sensitive scenarios. To mitigate this, we propose FloE, an on-the-fly MoE inference system on memory-constrained GPUs. FloE is built on the insight that there exists substantial untapped redundancy within sparsely activated experts. It employs various compression techniques on the expert's internal parameter matrices to reduce the data movement load, combined with low-cost sparse prediction, achieving perceptible inference acceleration in wall-clock time on resource-constrained devices. Empirically, FloE achieves a 9.3$\times$ compression of parameters per expert in Mixtral-8$\times$7B; enables deployment on a GPU with only 11GB VRAM, reducing the memory footprint by up to 8.5$\times$; and delivers a 48.7$\times$ inference speedup compared to DeepSpeed-MII on a single GeForce RTX 3090—all with only a 4.4\% $\sim$ 7.6\% average performance degradation.

With the growing abundance of repositories containing tabular data, discovering relevant tables for in-depth analysis remains a challenging task.Existing table discovery methods primarily retrieve desired tables based on a query table or several vague keywords, leaving users to manually filter large result sets.To address this limitation, we propose a new task: NL-conditional table discovery (nlcTD), where users combine a query table with natural language (NL) requirements to refine search results.To advance research in this area, we present nlcTables, a comprehensive benchmark dataset comprising 627 diverse queries spanning NL-only, union, join, and fuzzy conditions, 22,080 candidate tables, and 21,200 relevance annotations.Our evaluation of six state-ofthe-art table discovery methods on nlcTables reveals substantial performance gaps, highlighting the need for advanced techniques to tackle this challenging nlcTD scenario.The dataset, construction framework, and baseline implementations are publicly available at https://github.com/SuDIS-ZJU/nlcTables to foster future research.

Matrix-vector multiplication (MVM) operations, essential for modern hardware architectures, suffer from heavy I/O overheads and costly serial multiply-add operations. The emerging Compute-in-Memory (CIM) architecture alleviates these issues by enabling in situ MVM operations with O(1) time complexity, eliminating the need to move matrices. However, current storage schemes are still inefficient on CIM due to limited optimization objectives and inflexible support for various access patterns and matrix structures. To address this, we propose HyperMR, a hypergraph-enhanced matrix storage scheme for CIM architectures. First, we identify two performance optimization objectives that are tailored to CIM and prove their NP-hardness. We then introduce a hypergraph modeling approach with a novel access-aware hypergraph generation algorithm to handle diverse matrix structures and access patterns. Moreover, we present a two-phase hypergraph partitioning method to efficiently tackle the NP-hard optimization objectives. Experimental results show that HyperMR outperforms multiple state-of-the-art storage schemes, offering valid optimization for all evaluated matrices, compared to the best-performing baseline which optimizes only 75%. HyperMR also achieves the best average optimization performance for matrix storage layouts, significantly improving efficiency in varied workload scenarios, with a 29.65% improvement on synthetic queries and up to 34.9% on scientific image filtering.

The rise of LLM has enabled natural language-based table assistants, but existing systems assume users already have a well-formed table, neglecting the challenge of table discovery in large-scale table pools. To address this, we introduce TableCopilot, an LLM-powered assistant for interactive, precise, and personalized table discovery and analysis. We define a novel scenario, nlcTD, where users provide both a natural language condition and a query table, enabling intuitive and flexible table discovery for users of all expertise levels. To handle this, we propose Crofuma, a cross-fusion-based approach that learns and aggregates single-modal and cross-modal matching scores. Experimental results show Crofuma outperforms SOTA single-input methods by at least 12% on NDCG@5. We also release an instructional video, codebase, datasets, and other resources on GitHub to encourage community contributions. TableCopilot sets a new standard for interactive table assistants, making advanced table discovery accessible and integrated.

Over the recent years, Shapley value (SV), a solution concept from cooperative game theory, has found numerous applications in data analytics (DA). This paper presents the first comprehensive study of SV used throughout the DA workflow, clarifying the key variables in defining DA-applicable SV and the essential functionalities that SV can provide for data scientists. We condense four primary challenges of using SV in DA, namely computation efficiency, approximation error, privacy preservation, and interpretability, disentangle the resolution techniques from existing arts in this field, then analyze and discuss the techniques w.r.t. each challenge and the potential conflicts between challenges. We also implement SVBench , a modular and extensible open-source framework for developing SV applications in different DA tasks, and conduct extensive evaluations to validate our analyses and discussions. Based on the qualitative and quantitative results, we identify the limitations of current efforts for applying SV to DA and highlight the directions of future research and engineering.

Active learning (AL) reduces human annotation costs for machine learning systems by strategically selecting the most informative unlabeled data for annotation, but performing it individually may still be insufficient due to restricted data diversity and annotation budget. Federated Active Learning (FAL) addresses this by facilitating collaborative data selection and model training, while preserving the confidentiality of raw data samples. Yet, existing FAL methods fail to account for the heterogeneity of data distribution across clients and the associated fluctuations in global and local model parameters, adversely affecting model accuracy. To overcome these challenges, we propose${\sf CHASe}$(Client Heterogeneity-Aware Data Selection), specifically designed for FAL.${\sf CHASe}$focuses on identifying those unlabeled samples with high epistemic variations (EVs), which notably oscillate around the decision boundaries during training. To achieve both effectiveness and efficiency,${\sf CHASe}$encompasses techniques for 1) tracking EVs by analyzing inference inconsistencies across training epochs, 2) calibrating decision boundaries of inaccurate models with a new alignment loss, and 3) enhancing data selection efficiency via a data freeze and awaken mechanism with subset sampling. Experiments show that${\sf CHASe}$surpasses various established baselines in terms of effectiveness and efficiency, validated across diverse datasets, model complexities, and heterogeneous federation settings.

Multi-label few-shot aspect category detection (FS-ACD) is a challenging sentiment analysis task, which aims to learn a multi-label learning paradigm with limited training data. The difficulty of this task is how to use limited data to generalize effective discriminative representations for different categories. Nowadays, all advanced FS-ACD works utilize the prototypical network to learn label prototypes to represent different aspects. However, such point-based estimation methods are inherently noise-susceptible and bias-vulnerable. To this end, this paper proposes a novel Variational Hybrid-Attention Framework (VHAF) for the FS-ACD task. Specifically, to alleviate the data noise, we adopt a hybrid-attention mechanism to generate more discriminative aspect-specific embeddings. Then, based on these embeddings, we introduce the variational distribution inference to obtain the aspect-specific distribution as a more robust aspect representation, which can eliminate the scarce data bias for better inference. Moreover, we further leverage an adaptive threshold estimation to help VHAF better identify multiple relevant aspects. Extensive experiments on three datasets demonstrate the effectiveness of our VHAF over other state-of-the-art methods. Code is available at https://github.com/chengzju/VHAF.

Multi-modal multi-label emotion recognition (MMER) aims to identify relevant emotions from multiple modalities. The challenge of MMER is how to effectively capture discriminative features for multiple labels from heterogeneous data. Recent studies are mainly devoted to exploring various fusion strategies to integrate multi-modal information into a unified representation for all labels. However, such a learning scheme not only overlooks the specificity of each modality but also fails to capture individual discriminative features for different labels. Moreover, dependencies of labels and modalities cannot be effectively modeled. To address these issues, this paper presents ContrAstive feature Reconstruction and AggregaTion (CARAT) for the MMER task. Specifically, we devise a reconstruction-based fusion mechanism to better model fine-grained modality-to-label dependencies by contrastively learning modal-separated and label-specific features. To further exploit the modality complementarity, we introduce a shuffle-based aggregation strategy to enrich co-occurrence collaboration among labels. Experiments on two benchmark datasets CMU-MOSEI and M3ED demonstrate the effectiveness of CARAT over state-of-the-art methods. Code is available at https://github.com/chengzju/CARAT.

We present a novel inference scheme, selfspeculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model.This approach is characterized by a two-stage process: drafting and verification.The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting.Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass.This process ensures the final output remains identical to that produced by the unaltered LLM.Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration.Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99×. 1 * Huan Li and Lidan Shou are the corresponding authors. 1 Code is available at https://github.com/dilab-zju/ self-speculative-decoding.

Non- Volatile Memory (NVM) offers low-latency, non-volatility, and byte-addressability, positioning it as a highly promising device for database performance enhancement. Cur-rent research primarily focuses on utilizing LSM-Tree in con-junction with NVM to reduce write amplification and alleviate write stall issues. However, the comprehensive potential of NVM in simultaneously augmenting both read and write performances remains underexplored. And the previous NVM-enhanced LSM-Tree also ignores the sensitivity of NVM to small-grained random reads and writes, which we believe is the key to further improving read and write performance. To address these issues, we propose BushStore, an innovative LSM-Tree variant specifically optimized for NVM. BushStore is designed with a three-level architecture, where the higher levels of BushStore contain a group of immutable, non-clustered B+Trees, replacing traditional SSTables. By storing the non-leaf nodes in the DRAM and the leaf nodes in NVM, and separating the data pages from the indexes, these B+Trees are able to exhibit high performance for diverse read and write operations. Our approach encompasses four key techniques to significantly boost system efficiency: First, we develop novel data structures that localize read/write operations to confined NVM areas, enhancing access speed. Second, we optimize the key-value data handling during flushing and compaction phases, leveraging the superior scanning and sequantial writing capabilities of B+Trees to ex-pedite write and compaction processes. Third, we dynamically adjust the B+Tree sizes, enabling a balanced and optimized flushing and compaction process, thereby improving overall write performance. Fourth, we implement a lazy-delete Cuckoo filtering and lazy-persistent allocation strategy to accelerate query and compaction processes. Evaluations show that BushStore exhibits high performance and scalability under synthetic and real work-loads, and achieves an average performance improvement of 3.3x in random write throughput and 4.3x in random read throughput compared to the state-of-the-art MioDB system.

Privacy concerns have led to the rise of federated recommender systems (FRS), which can create personalized models across distributed clients. However, FRS is vulnerable to poisoning attacks, where malicious users manipulate gradients to promote their target items intentionally. Existing attacks against FRS have limitations, as they depend on specific models and prior knowledge, restricting their real-world applicability. In our exploration of practical FRS vulnerabilities, we devise a model-agnostic and prior-knowledge-free attack, named PIECK (Popular Item Embedding based Attack). The core module of PIECK is popular item mining, which leverages embedding changes during FRS training to effectively identify the popular items. Built upon the core module, PIECK branches into two diverse solutions: The PIECKIPE solution employs an item popularity enhancement module, which aligns the embeddings of targeted items with the mined popular items to increase item exposure. The PIECKUEA further enhances the robustness of the attack by using a user embedding approximation module, which approximates private user embeddings using mined popular items. Upon identifying PIECK, we evaluate existing federated defense methods and find them ineffective against PIECK, as poisonous gradients inevitably overwhelm the cold target items. We then propose a novel defense method by introducing two regularization terms during user training, which constrain item popularity enhancement and user embedding approximation while preserving FRS performance. We evaluate PIECK and its defense across two base models, three real datasets, four top-tier attacks, and six general defense methods, affirming the efficacy of both PIECK and its defense.

Multi-label text classification (MLTC) allows a given text to be associated with multiple labels, which well suits many real-world data mining scenarios. However, the annotation effort of MLTC is inevitably expensive and time-consuming. Although multi-label active learning provides a cost-effective solution, it still faces two major challenges: (i) constructing decent feature space to distinguish the confusing semantics of different labels; (ii) defining proper sampling criteria to measure a sample's joint effect over the entire label space. To bridge these gaps, we propose a Contrastive Multi-label Active Learning framework (CoMAL) that gives an effective data acquisition strategy. Specifically, a contrastive decoupling mechanism is introduced to fully release the semantic information of multiple labels into the latent space. Then, we devise a hybrid criterion that balances two data value measures: (i) similarity-enhanced label cardinality inconsistency reflects the uncertainty of data predictions. (ii) positive feature diversity evaluates the positive-propensity semantic diversity to handle the label sparsity. Extensive experiments demonstrate that our CoMAL outperforms the current state-of-the-art multi-label active learning approaches. Code for CoMAL is available at https://github.com/chengzju/CoMAL.

With breakthroughs in pretrained language models, a large number of finetuned models specialized in distinct domains have surfaced online. Yet, when faced with a fresh dataset covering multiple (sub)domains, their performance might degrade. Reusing these available finetuned models to train a new model is a more feasible solution than the finetuning method that demands extensive manual labeling. Knowledge Amalgamation (KA) is such a model reusing technique, which derives a new model (termed student model) by amalgamating those trained models (termed teacher models) tailored for distinct domains, bypassing the need for manual labeling. However, when the domains of text samples are unknown, selecting a number of appropriate teacher models (simply called a combination) for reuse becomes complicated. To learn an accurate student model, the classical KA method resorts to manual selections, a process both tedious and inefficient. Our study pioneers the automation of this combination selection process for KA in the fundamental text classification task, an area previously unexplored.

An effective video summary should encapsulate the entire narrative and highlight its most critical content. However, supervised learning heavily relies on labor-intensive and time-consuming manual annotations. To tackle the issue, we propose a Convolutional Attentive Adversarial Network that fundamentally aims to create a deep summarization model in an unsupervised manner. We employ a Generative Adversarial Network (GAN) framework to solve the task, where a generator assigns importance scores to all video frames, while a discriminator distinguishes between score-weighted frame features and the original counterparts. We introduce a novel representative loss function, complemented by adversarial, sparsity, and reconstruction losses, to guide the prediction of frame importance scores. In order to substantiate our proposed method's efficacy, we have conducted extensive experiments across two public benchmark datasets, SumMe and TVSum. The outcomes demonstrate that our approach surpasses other state-of-the-art methods.

Maximal biclique enumeration (MBE) is crucial in bipartite graph analysis. Recent studies rely on extensive set intersections on static bipartite graphs to solve the MBE problem. However, the computational subgraphs dynamically change during enumeration, leading to redundant memory accesses and degraded set intersection performance. To overcome this limitation, we propose an AdaMBE algorithm. First, we redesign its core operations using local neighborhood information derived from computational subgraphs to minimize redundant memory accesses. Second, we dynamically create computational subgraphs using bitmaps leveraging its fast bitwise operations to accelerate set intersections. Finally, we integrate them in AdaMBE. Our experimental results show that AdaMBE is $1.6 \times-49.7 \times$ faster than its closest CPU-based competitor and successfully enumerates all 19 billion maximal bicliques on the TVTropes dataset, a large task beyond the capabilities of existing algorithms. Notably, on certain datasets, our parallel version, ParAdaMBE, on CPUs even outperforms GMBE on GPUs by up to $5.07 \times$.

Abstract Federated learning (FL) is a promising approach for learning a model from data distributed on massive clients without exposing data privacy. It works effectively in the ideal federation where clients share homogeneous data distribution and learning behavior. However, FL may fail to function appropriately when the federation is not ideal, amid an unhealthy state called Negative Federated Learning (NFL), in which most clients gain no benefit from participating in FL. Many studies have tried to address NFL. However, their solutions either (1) predetermine to prevent NFL in the entire learning life-cycle or (2) tackle NFL in the aftermath of numerous learning rounds. Thus, they either (1) indiscriminately incur extra costs even if FL can perform well without such costs or (2) waste numerous learning rounds. Additionally, none of the previous work takes into account the clients who may be unwilling/unable to follow the proposed NFL solutions when using those solutions to upgrade an FL system in use. This paper introduces FL-GUARD, a holistic framework that can be employed on any FL system for tackling NFL in a run-time paradigm. That is, to dynamically detect NFL at the early stage (tens of rounds) of learning and then to activate recovery measures when necessary. Specifically, we devise a cost-effective NFL detection mechanism, which relies on an estimation of performance gain on clients. Only when NFL is detected, we activate the NFL recovery process, in which each client learns in parallel an adapted model when training the global model. Extensive experiment results confirm the effectiveness of FL-GUARD in detecting NFL and recovering from NFL to a healthy learning state. We also show that FL-GUARD is compatible with previous NFL solutions and robust against clients unwilling/unable to take any recovery measures.

Large-scale multi-label text classification (LMTC) aims at tagging each text with multiple relevant labels from a large label space, which typically demonstrates high sparsity, diversity, and skewness. To learn text representations in LMTC, a straightforward strategy is to learn a single vector to represent the whole text, yet limiting good generalization to diverse labels; another popular one is to learn specific representation per label via attention weighting, but excessively emphasizing tail labels restricts the overall performance. To cope with these limitations, we propose a novel LMTC framework, dubbed LADAR, which learns label-adaptive text representations to ensure high performance on large-scale labels. Specifically, we construct a representation pool for each text by collecting multi-layer features of the deep model as well as multi-granularity features of the text. Furthermore, all labels are adaptively matched to their most relevant representations to predict the final scores. Experiments over five benchmark datasets demonstrate the LADAR achieves highly superior results to state-of-the-art LMTC approaches. In particular, LADAR achieves significantly better performance on tail labels, e.g., 5.09% relative improvement on PSP@5 on the Amazon-670K dataset than the best baseline.

Continual learning is known for suffering from catastrophic forgetting, a phenomenon where previously learned concepts are forgotten upon learning new tasks. A natural remedy is to use trained models for old tasks as ‘teachers’ to regularize the update of the current model to prevent such forgetting. However, this requires storing all past models, which is very space-consuming for large models, e.g. BERT, thus impractical in real-world applications. To tackle this issue, we propose to construct snapshots of seen tasks whose key knowledge is captured in lightweight adapters. During continual learning, we transfer knowledge from past snapshots to the current model through knowledge distillation, allowing the current model to review previously learned knowledge while learning new tasks. We also design representation recalibration to better handle the class-incremental setting. Experiments over various task sequences show that our approach effectively mitigates catastrophic forgetting and outperforms all baselines.

Sequential recommendation has long been challenged by data sparsity issues. Most recently, cross-domain sequential recommendation (CDSR) techniques have been proposed to leverage sequential interaction data from other domains. However, accessing raw data from source domains is often restricted due to privacy concerns. To tackle this issue, we introduce TPUF, a novel CDSR model that transfers pre-trained latent user features from the source domain (UFS) instead of the original interaction data. By doing so, TPUF improves recommendation effectiveness while maintaining practicality. TPUF has three functional characteristics: (1) It is a feature mapping-and-aggregation framework that does not impose specific constraints on the nature of pre-trained UFS. (2) It incorporates a temporal feature mapping unit to effectively extract domain-shared information from UFS with temporal information recovered. (3) It additionally employs an adversarial feature alignment unit to align features across domains to combat feature transfer bias. Experimental results on real-world datasets demonstrate that TPUF outperforms other state-of-the-art cross-domain recommendation models and is compatible with multiple UFS types.

This study proposes FEDBFPT (Federated BERT Further Pre-Training), a Federated Learning (FL) framework for further pre-training the BERT language model in specialized domains while addressing privacy concerns. FEDBFPT enables multiple clients to collaboratively train the shallower layers of BERT, which are crucial in the pre-training stage, without the need to share private data. To achieve this, FEDBFPT involves building a local model for each client, progressively training the shallower layers of local models while sampling deeper layers, and aggregating trained parameters on a server to create the final global model. This approach utilizes multiple smaller local models to further pre-train a global model targeted at specific tasks via fine-tuning, resulting in a reduction in resource usage while maintaining model accuracy. Theoretical analysis is conducted to support the efficiency of FEDBFPT, and experiments are conducted on corpora across domains such as medicine, biology, and computer science. Results indicate that FEDBFPT achieves performance levels comparable to traditional FL methods while reducing computation and communication costs by 46.70% and 7.04%, respectively, even approaching the performance of centralized training models. The Source code is released at https://github.com/Hanzhouu/FedBFPT.

Data analysts nowadays are keen to have analytical capabilities involving deep learning (DL). Collaborative queries, which employ relational operations to process structured data and DL models to process unstructured data, provide a powerful facility for DL-based in-database analysis. The classical approach to support collaborative queries in relational databases is to integrate DL models with user-defined functions (UDFs) in a general-purpose language (e.g., C++) to process unstructured data. This approach suffers from suboptimal performance as the opaque UDFs preclude the generation of an optimal query plan. A recent work, DL2SQL, addresses the problem of collaborative query optimization by first converting DL computations into SQL subqueries and then using a classical relational query optimizer to optimize the entire collaborative query. However, the DL2SQL approach compromises usability by requiring data analysts to manually manage DL-related data and tune query performance. To this end, this paper introduces AQUA, an analytical database designed for efficient collaborative query processing. Built on DL2SQL, AQUA automates translations from collaborative queries into SQL queries. To enhance usability, AQUA introduces two techniques: 1) a declarative scheme for DL-related data management, and 2) DL-specific optimizations for collaborative query processing, eliminating the burden of manual data management and performance tuning from the data analysts. We demonstrate the key contributions of AQUA via a web APP that allows the audience to perform collaborative queries on the CIFAR-10 dataset.

We propose HOlistic MEmory management (HOME), a new framework for performing tensor placements in large DNN training when GPU memory space is not enough. HOME combines tensor swapping with tensor recomputation to reduce GPU memory footprint. Different from existing work that only considers partial DNN model information, HOME takes the holistic DNN model information into account in tensor placement decisions. More specifically, HOME uses a custom-designed particle swarm optimization algorithm to achieve the globally optimized placement for each tensor of the DNN model with a greatly reduced searching space. This holistic awareness of the whole model information enables HOME to obtain high performance under the given GPU memory constraint. We implement HOME in PyTorch and conduct our experiments using six popular DNN models. Experimental results show that HOME can outperform vDNN and Capuchin by up to 5.7× and 1.3× in throughput. Furthermore, HOME can improve the maximum batch size by up to 2.8× than the original PyTorch and up to 1.3× than Capuchin.

In this paper, we propose SkipBERT to accelerate BERT inference by skipping the computation of shallow layers.To achieve this, our approach encodes small text chunks into independent representations, which are then materialized to approximate the shallow representation of BERT.Since the use of such approximation is inexpensive compared with transformer calculations, we leverage it to replace the shallow layers of BERT to skip their runtime overhead.With off-the-shelf early exit mechanisms, we also skip redundant computation from the highest few layers to further improve inference efficiency.Results on GLUE show that our approach can reduce latency by 65% without sacrificing performance.By using only two-layer transformer calculations, we can still maintain 95% accuracy of BERT. 1

As people spend large parts of their lives in indoor venues like shopping malls, airports, and office buildings, there are increasing demands of indoor keyword-aware routing, i.e., finding an indoor path that covers interesting keywords. In this work, we demonstrate an Indoor Keyword-Aware Routing System (IKAROS) which efficiently answers the indoor top-$k$keyword-aware routing query (IKRQ). Given two indoor points$s$and$t$, an IKRQ returns$k\ s$-to-$t$routes that do not exceed a given distance constraint but have optimal ranking scores integrating keyword relevance and spatial distance. To enable cross-platform IKRQ services for end-users, IKAROS adopts the Browser/Server system architecture. The browser provides the interface for users to specify queries and view the results, while the server manages the indoor spatial and textual information and processes queries. Specifically, IKAROS implements two search algorithms with different routing expansions through a tailored indoor space model. Our demonstration covers the major system components such as configuring a query's parameters, choosing a query search algorithm, and visualizing the returned route.

Federated learning (FL) is a promising approach for learning a shared global model on decentralized data owned by multiple clients without exposing their privacy. In real-world scenarios, data accumulated at the client-side varies in distribution over time. As a consequence, the global model tends to forget the knowledge obtained from previous tasks while learning new tasks, showing signs of "catastrophic forgetting". Previous studies in centralized learning use techniques such as data replay and parameter regularization to mitigate catastrophic forgetting. Unfortunately, these techniques cannot adequately solve the non-trivial problem in FL. We propose Continual Federated Learning with Distillation (CFeD) to address catastrophic forgetting under FL. CFeD performs knowledge distillation on both the clients and the server, with each party independently having an unlabeled surrogate dataset, to mitigate forgetting. Moreover, CFeD assigns different learning objectives, namely learning the new task and reviewing old tasks, to different clients, aiming to improve the learning ability of the model. The results show that our method performs well in mitigating catastrophic forgetting and achieves a good trade-off between the two objectives.

In recent years, there has been increasing interest in adopting published neural retrieval models learned from corpora for text retrieval. Although these models achieve excellent retrieval performance, in terms of popular accuracy metrics, on datasets they have been trained, their performance on new text data might degrade. To obtain the desired retrieval performance on both the data used in training and the latest data collected after training, the simple approach of learning a new model from both datasets is not always feasible since the annotated dataset used in training is often not published along with the learned model. Knowledge amalgamation (KA) is an emerging technique to deal with this problem of inaccessibility of data used in previous training. KA learns a new model (called a student model) from new data by reusing (called amalgamating) a number of trained models (called teacher models) instead of accessing the teachers' original training data. However, in order to efficiently learn an accurate student model, the classical KA approach requires manual selection of an appropriate subset of teacher models for amalgamation. This manual procedure for selecting teacher models prevents the classical KA from being scaled to retrieval tasks for which a large number of candidate teacher models are ready to be reused.

Complex data mining algorithms are processed in multiple iterations, where output of one iteration is used as input for the subsequent iterations. Existing parallel programming frameworks, e.g., MapReduce, Pregel and Spark, adopt the breadth first search (BFS) strategy to process those iterative jobs. They invoke the user-defined functions for every key-value pair or vertex to produce all possible intermediate results for the next iteration. Such BFS strategy incurs high I/O overheads, because normally, the size of intermediate search results of BFS is exponential to the size of original data, making it impossible to maintain those intermediate results in memory. In this paper, we present a new type of parallel programming model, the stack-centric model, where all computations are defined for a stack maintained in the distributed shared memory. The stack can be adaptively split into multiple stacks and disseminated to different compute nodes for parallel processing. The most distinguished feature of the stack-centric model is its support for the depth first search (DFS) algorithm which incurs much less memory overhead than its BFS counterpart. The maximal memory usage of DFS algorithm is determined by the height of its search tree, and hence, it is possible to conduct the computation of DFS algorithm mostly in memory. Our stack-centric model is not a pure DFS framework. It supports the hybrid BFS and DFS algorithms by tuning the trade-off between memory usage and parallelism. To show the advantages of stack-centric model, we implement two algorithms, frequent pattern mining algorithm and DNA sequence matching algorithm, on both stack-centric model and Spark. The memory usage of stack-centric model is 10 times less than the Spark, resulting in a significant performance improvement.

To support indoor spatial queries and indoor location-based services (LBS), multiple techniques including model/indexes and search algorithms have been proposed. In this work, we conduct an extensive experimental study on existing proposals for indoor spatial queries. We survey five model/indexes, compare their algorithmic characteristics, and analyze their space and time complexities. We also design an in-depth benchmark with real and synthetic datasets, evaluation tasks and performance metrics. Enabled by the benchmark, we obtain and report the performance results of all model/indexes under investigation. By analyzing the results, we summarize the pros and cons of all techniques and suggest the best choice for typical scenarios.

Membership Inference Attack (MIA) in deep learning is a common form of privacy attack which aims to infer whether a data sample is in a target classifier's training dataset or not. Previous studies of MIA typically tackle either a black-box or a white-box adversary model, assuming an attacker not knowing (or knowing) the structure and parameters of the target classifier while having access to the confidence vector of the query output. With the popularity of privacy protection methods such as differential privacy, it is increasingly easier for an attacker to obtain the defense method adopted by the target classifier, which poses extra challenge to privacy protection. In this paper, we name such attacker a crystal-box adversary. We present definitions for utility and privacy of target classifier, and formulate the design goal of the defense method as an optimization problem. We also conduct theoretical analysis on the respective forms of the optimization for three adversary models, namely black-box, white-box, and crystal-box, and prove that the optimization problem is NP-hard. Thereby we solve a surrogate problem and propose three defense methods, which, if used together, can make trade-off between utility and privacy. A notable advantage of our approach is that it can be used to resist attacks from three adversary models, namely black-box, white-box, and crystal-box, simultaneously. Evaluation results show effectiveness of our proposed approach for defending privacy against MIA and better performance compared to previous defense methods.

Indoor venues accommodate many people who collectively form crowds. Such crowds in turn influence people's routing choices, e.g., people may prefer to avoid crowded rooms when walking from A to B. This paper studies two types of crowd-aware indoor path planning queries. The Indoor Crowd-Aware Fastest Path Query (FPQ) finds a path with the shortest travel time in the presence of crowds, whereas the Indoor Least Crowded Path Query (LCPQ) finds a path encountering the least objects en route. To process the queries, we design a unified framework with three major components. First, an indoor crowd model organizes indoor topology and captures object flows between rooms. Second, a time-evolving population estimator derives room populations for a future timestamp to support crowd-aware routing cost computations in query processing. Third, two exact and two approximate query processing algorithms process each type of query. All algorithms are based on graph traversal over the indoor crowd model and use the same search framework with different strategies of updating the populations during the search process. All proposals are evaluated experimentally on synthetic and real data. The experimental results demonstrate the efficiency and scalability of our framework and query processing algorithms.

Modern mobile applications often produce decentralized data, i.e., a huge amount of privacy-sensitive data distributed over a large number of mobile devices. Techniques for learning models from decentralized data must properly handle two natures of such data, namely privacy and massive engagement. Federated learning (FL) is a promising approach for such a learning task since the technique learns models from data without exposing privacy. However, traditional FL methods assume that the participating mobile devices are honest volunteers. This assumption makes traditional FL methods unsuitable for applications where two kinds of participants are engaged: 1) self-interested participants who, without economical stimulus, are reluctant to contribute their computing resources unconditionally, and 2) malicious participants who send corrupt updates to disrupt the learning process. This paper proposes Refiner, a reliable federated learning system for tackling the challenges introduced by massive engagements of self-interested and malicious participants. Refiner is built upon Ethereum, a public blockchain platform. To engage self-interested participants, we introduce an incentive mechanism which rewards each participant in terms of the amount of its training data and the performance of its local updates. To handle malicious participants, we propose an audit scheme which employs a committee of randomly chosen validators for punishing them with no reward and preclude corrupt updates from the global model. The proposed incentive and audit scheme is implemented with cryptocurrency and smart contract, two primitives offered by Ethereum. This paper demonstrates the main features of Refiner by training a digit classification model on the MNIST dataset.

This paper presents Pyramid, a novel layered model for Nested Named Entity Recognition (nested NER).In our approach, token or text region embeddings are recursively inputted into L flat NER layers, from bottom to top, stacked in a pyramid shape.Each time an embedding passes through a layer of the pyramid, its length is reduced by one.Its hidden state at layer l represents an l-gram in the input text, which is labeled only if its corresponding text region represents a complete entity mention.We also design an inverse pyramid to allow bidirectional interaction between layers.The proposed method achieves state-of-the-art F1 scores in nested NER on ACE-2004, ACE-2005, GENIA, and NNE, which are 80.27, 79.42, 77.78, and 93.70 with conventional embeddings, and 87.74, 86.34, 79.31, and 94.68 with pre-trained contextualized embeddings.In addition, our model can be used for the more general task of Overlapping Named Entity Recognition.A preliminary experiment confirms the effectiveness of our method in overlapping NER.

Indoor mobility semantics analytics can greatly benefit many pertinent applications. Existing semantic annotation methods mainly focus on outdoor space and require extra knowledge such as POI category or human activity regularity. However, these conditions are difficult to meet in indoor venues with relatively small extents but complex topology. This work studies the annotation of indoor mobility semantics that describe an object's mobility event (what ) at a semantic indoor region (where ) during a time period (when ). A coupled conditional Markov network (C2MN) is proposed with a set of feature functions carefully designed by incorporating indoor topology and mobility behaviors. C2MN is able to capture probabilistic dependencies among positioning records, semantic regions, and mobility events jointly. Nevertheless, the correlation of regions and events hinders the parameters learning. Therefore, we devise an alternate learning algorithm to enable the parameter learning over correlated variables. The extensive experiments demonstrate that our C2MN-based semantic annotation is efficient and effective on both real and synthetic indoor mobility data.

People have many activities indoors and there is an increasing demand of keyword-aware route planning for indoor venues. In this paper, we study the indoor top-k keyword-aware routing query (IKRQ). Given two indoor points s and t, an IKRQ returns k s-to-t routes that do not exceed a given distance constraint but have optimal ranking scores integrating keyword relevance and spatial distance. It is challenging to efficiently compute the ranking scores and find the best yet diverse routes in a large indoor space with complex topology. We propose prime routes to diversify top-k routes, devise mapping structures to organize indoor keywords and compute route keyword relevances, and derive pruning rules to reduce search space in routing. With these techniques, we design two search algorithms with different routing expansions. Experiments on synthetic and real data demonstrate the efficiency of our proposals.

Knowing popular indoor locations can benefit many applications like exhibition planning and location-based advertising, among others. In this work, we use uncertain historical indoor mobility data to find the top-k popular indoor semantic locations with the highest flow values. In the data we use, an object positioning report contains a set of samples, each consisting of an indoor location and a corresponding probability. The problem is challenging due to the difficulty in obtaining reliable flow values and the heavy computational workload on probabilistic samples for large numbers of objects. To address the first challenge, we propose an indoor flow definition that takes into account both data uncertainty and indoor topology. To efficiently compute flows for individual indoor semantic locations, we design data structures for facilitating accessing the relevant data, a data reduction method that reduces the intermediate data to process, and an overall flow computing algorithm. Furthermore, we design search algorithms for finding the top-k popular indoor semantic locations. All proposals are evaluated extensively on real and synthetic data. The evaluation results show that our data reduction method significantly reduces the data volume in computing, our search algorithms are efficient and scalable, and the top-k popular semantic locations returned are in good accord with ground truth.

As people spend significant parts of daily lives indoors, it is useful and important to measure indoor densities and find the dense regions in many indoor scenarios like space management and security control. In this paper, we propose a data-driven approach that finds top-k indoor dense regions by using indoor positioning data. Such data is obtained by indoor positioning systems working at a relatively low frequency, and the reported locations in the data are discrete, from a preselected location set that does not continuously cover the entire indoor space. When a search is triggered, the object positioning information is already out-of-date and thus object locations are uncertain. To this end, we first integrate object location uncertainty into the definitions for counting objects in an indoor region and computing its density. Subsequently, we conduct a thorough analysis of the location uncertainty in the context of complex indoor topology, deriving upper and lower bounds of indoor region densities and introducing distance decaying effect into computing concrete indoor densities. Enabled by the uncertainty analysis outcomes, we design efficient search algorithms for solving the problem. Finally, we conduct extensive experimental studies on our proposals using synthetic and real data. The experimental results verify that the proposed search approach is efficient, scalable, and effective. The top-k indoor dense regions returned by our search are considerably consistent with ground truth, despite that the search uses neither historical data nor extra knowledge about objects.

A crowdsourcing task in real-world applications often consists of thousands of atomic tasks. A common practice to distribute a large-scale crowdsourcing task is to pack atomic tasks into task bins and send to crowd workers in batches. It is challenging to decompose a large-scale crowdsourcing task into task bins to ensure reliability at a minimal total cost. In this paper, we propose the Smart Large-scAle task DEcomposer (SLADE) problem, which aims to decompose a large-scale crowdsourcing task to achieve the desired reliability at a minimal cost. We prove its NP-hardness and study two variants of the problem. For the homogeneous SLADE problem, we propose a greedy algorithm and an approximation framework using an optimal priority queue (OPQ) structure with provable approximation ratio. For the heterogeneous SLADE problem, we extend this framework and prove its approximation guarantee. Extensive experiments validate the effectiveness and efficiency of the solutions.

Attributed entity is an entity defined by its structural attributes. Extracting attributed entities from textual documents is an important problem for a variety of big-data applications. We propose a system called EXACT for extracting attributed entities from textual documents by performing explorative annotation tasks, which create attributes and bind them to tag values. To support efficient annotation, we propose a novel tag recommendation technique based on a few-shot learning scheme which can suggest tags for new annotation tasks given very few human-annotated samples. We also propose a document recommendation scheme to provide run-time context for the user. Using a novel attribute index, the system can generate the task-relevant attributed entities on-the-fly. We demonstrate how these techniques can be integrated behind a novel user interface to enable productive and efficient extraction of attributed entities at limited cost in human annotation.

The emergence of persistent memory (PM) spurs on redesigns of database system components to gain full exploitation of the persistence and speed of the hardware. One crucial component studied by researchers is persistent indices. However, such studies to date are unsatisfactory in terms of the number of expensive PM writes required for crash-consistency. In this paper, we propose a new persistent index called DPTree (Differential Persistent Tree) to address this. DPTree's core idea is to batch multiple writes in DRAM persistently and later merge them into a PM component to amortize persistence overhead. DPTree includes several techniques and algorithms to achieve crash-consistency, reduce PM writes significantly, and maintain excellent read performance. To embrace multi-core processors, we present the design of concurrent DPTree. Our experiments on Intel's Optane DIMMs show that DPTree reduces PM writes by a factor of 1.7x-3x compare to state-of-the-art counterparts. Besides, DPTree has a competitive or better read performance and scales well in multi-core environment.

Knowing popular indoor locations can benefit many applications like exhibition planning and location-based advertising, among others. In this work, we use uncertain historical indoor mobility data to find the top-k popular indoor semantic locations with the highest flow values. In the data we use, an object positioning report contains a set of samples, each consisting of an indoor location and a corresponding probability. The problem is challenging due to the difficulty in obtaining reliable flow values and the heavy computational workload on probabilistic samples for large numbers of objects. To address the first challenge, we propose an indoor flow definition that takes into account both data uncertainty and indoor topology. To efficiently compute flows for individual indoor semantic locations, we design data structures for facilitating accessing the relevant data, a data reduction method that reduces the intermediate data to process, and an overall flow computing algorithm. Furthermore, we design search algorithms for finding the top-k popular indoor semantic locations. All proposals are evaluated extensively on real and synthetic data. The evaluation results show that our data reduction method significantly reduces the data volume in computing, our search algorithms are efficient and scalable, and the top-k popular semantic locations returned are in good accord with ground truth.

Big data analytics has been revolutionizing the fashion industry in recent years. This is evidenced by the fact that popular fashion brands and designers have relied on big data analytics to trace fashion trends and predict market patterns. In this paper, we propose learning a common latent feature representation from heterogeneous fashion data. Specifically, we design a multi-modal and multi-domain embedding learning framework for fashion analysis and data retrieval. Unlike most of the existing multi-view embedding methods, which only consider the heterogeneous similarity constraint, our proposed framework jointly considers both the homogeneous and heterogeneous similarity constraints to capture cross-view similarity and preserve the similarity of the same view. The proposed framework is comprised of two projection steps. In the first projection, a quintuplet-based ranking loss is proposed for multi-domain fashion data to preserve the homogeneous similarity. In the second projection, a cross-view similarity ranking loss is designed for multi-modal fashion data to capture heterogeneous similarity. By utilizing the learned common latent feature representation, the distance between any vector pairs from same or different modalities can reflect its semantic similarity. Quantitative evaluation on a new large-scale dataset and a fashion analysis case study demonstrate the effectiveness of our proposed method.

The rapid accumulation of indoor positioning data is increasingly booming the interest in indoor mobility analyses. As a fundamental analysis, it is highly relevant to translate raw indoor positioning data into mobility semantics that describe what, where and when in a more concise and semantics-oriented way. Such a translation is challenging as multiple data sources are involved, raw indoor positioning data is of low quality, and translation results are hard to assess. We demonstrate a system TRIPS that streamlines the entire translation process by three functional components. The Configurator provides a standard but concise means to configure multiple input sources, including the indoor positioning data, indoor space information, and relevant contexts. The Translator cleans the indoor positioning data and exports reliable mobility semantics without manual interventions. The Viewer offers a suite of flexible operations to trace the input, output and intermediate data involved in the translation. Data analysts can interact with TRIPS to obtain the desired mobility semantics in a visual and convenient way.

As people spend significant parts of daily lives indoors, it is useful and important to measure indoor densities and find the dense regions in many indoor scenarios like space management and security control. In this paper, we propose a data-driven approach that finds top-k indoor dense regions by using indoor positioning data. Such data is obtained by indoor positioning systems working at a relatively low frequency, and the reported locations in the data are discrete, from a preselected location set that does not continuously cover the entire indoor space. When a search is triggered, the object positioning information is already out-of-date and thus object locations are uncertain. To this end, we first integrate object location uncertainty into the definitions for counting objects in an indoor region and computing its density. Subsequently, we conduct a thorough analysis of the location uncertainty in the context of complex indoor topology, deriving upper and lower bounds of indoor region densities and introducing distance decaying effect into computing concrete indoor densities. Enabled by the uncertainty analysis outcomes, we design efficient search algorithms for solving the problem. Finally, we conduct extensive experimental studies on our proposals using synthetic and real data. The experimental results verify that the proposed search approach is efficient, scalable, and effective. The top-k indoor dense regions returned by our search are considerably consistent with ground truth, despite that the search uses neither historical data nor extra knowledge about objects.

Crowdsourcing has been shown to be effective in a wide range of applications, and is seeing increasing use. A large-scale crowdsourcing task often consists of thousands or millions of atomic tasks, each of which is usually a simple task such as binary choice or simple voting. To distribute a large-scale crowdsourcing task to limited crowd workers, a common practice is to pack a set of atomic tasks into a task bin and send to a crowd worker in a batch. It is challenging to decompose a large-scale crowdsourcing task and execute batches of atomic tasks, which ensures reliable answers at a minimal total cost. Large batches lead to unreliable answers of atomic tasks, while small batches incur unnecessary cost. In this paper, we investigate a general crowdsourcing task decomposition problem, called the Smart Large-scAle task DEcomposer (SLADE) problem, which aims to decompose a large-scale crowdsourcing task to achieve the desired reliability at a minimal cost. We prove the NP-hardness of the SLADE problem and propose solutions in both homogeneous and heterogeneous scenarios. For the homogeneous SLADE problem, where all the atomic tasks share the same reliability requirement, we propose a greedy heuristic algorithm and an efficient and effective approximation framework using an optimal priority queue (OPQ) structure with provable approximation ratio. For the heterogeneous SLADE problem, where the atomic tasks can have different reliability requirements, we extend the OPQ-based framework leveraging a partition strategy, and also prove its approximation guarantee. Finally, we verify the effectiveness and efficiency of the proposed solutions through extensive experiments on representative crowdsourcing platforms.

Driven by the increasing popular image-dominated social networks, such as Instagram, Pinterest and Chictopica, sharing of daily-life street photos now plays an influential role in fashion adoption between fashion trend-setters and followers. In this work, we propose a deep learning based fine-grained embedding learning approach for street fashion analysis by leveraging user-generated street fashion data. Specifically, we present QuadNet, an effective CNN based image embedding network driven by both multi-task classification loss and neighbor-constrained similarity loss. The latter loss function is computed with a novel quadruplet loss function, which considers both hard and soft positive neighbors as well as a negative neighbor for each anchor image. The embedded feature learned from co-optimization is effective for both fine-grained classification task and image retrieval task. Quantitative evaluation on a newly collected large-scale multi-task street photo dataset shows that our QuadNet outperforms the state-of-the-art triplet network by a significant margin. In order to further evaluate the effectiveness of the learned embedding, we analyze and trace the fashion trends of New York City from 2011 to 2016. In our analysis, we are able to identify some short-term and long-term fashion styles.

New data has been generated in an unexpected high speed. To get insight of those data, data analysts will perform a thorough study using state-of-the-art big data analytical tools. Before the analysis starts, a preprocessing is conducted, where data analyst tends to issue a few ad-hoc queries on a new dataset to explore and gain a better understanding. However, it is costly to perform such ad-hoc queries on large scale data using traditional data management systems, e.g., DBMS, because data loading and indexing are very expensive. In this demo, we propose a novel visual data explorer system, FlashView, which omits the loading process by directly querying raw data. FlashView applies approximate query processing technique to achieve real-time query results. It builds both in-memory index and disk index to facilitate the data scanning. It also supports tracking and updating multiple queries concurrently. Note that FlashView is not designed as a replacement of full-fledged DBMS. Instead, it tries to help the analysts quickly understand the characteristics of data, so he/she can selectively load data into the DBMS to do more sophisticated analysis.

We demonstrate iGlasses, a novel recommendation system that accepts a frontal face photo as the input and returns the best-fit eyeglasses as the output. As conventional recommendation techniques such as collaborative filtering become inapplicable in the problem, we propose a new recommendation method which exploits the implicit matching rules between human faces and eyeglasses. We first define fine-grained attributes for human faces and frames of glasses respectively. Then, we develop a recommendation framework based on a probabilistic graphical model, which effectively captures the correlation among these fine-grained attributes. Ranking of the frames (glasses) is done by their similarity to the query facial attributes. Finally, we produce a synthesized image for the input face to demonstrate the visual effect when wearing the recommended glasses.

We demonstrate a generic, user-configurable toolkit for generating different types of indoor mobility data for real-world buildings. Our prototype generates the desired data in a three-layer pipeline. The Infrastructure Layer accepts industry-standard digital building information (DBI) files to generate the host indoor environment, allowing users to configure the generation of a variety of positioning devices, such as Wi-Fi, Bluetooth, RFID, etc. The Moving Object Layer offers the functionality of defining objects or trajectories, with configurable indoor moving patterns, distribution models, and sampling frequencies. The Positioning Layer generates synthetic signal strength measurements known as raw RSSI 1 measurements according to the positioning device data and trajectory data generated at relevant layers. It also generates different types of indoor positioning data through the customization of all typical indoor positioning methods on the raw RSSI data.

This paper presents a project called KnowIng camera prototype SyStem (KISS) for real-time places-of-interest (POI) recognition and annotation for smartphone photos, with the availability of online geotagged images for POIs as our knowledge base. We propose a “Spatial+Visual” (S+V) framework which consists of a probabilistic field-of-view (pFOV) model in the spatial phase and sparse coding similarity metric in the visual phase to recognize phone-captured POIs. Moreover, we put forward an offline Collaborative Salient Area (COSTAR) mining algorithm to detect common visual features (called Costars) among the noisy photos geotagged on each POI, thus to clean the geotagged image database. The mining result can be utilized to annotate the region-of-interest on the query image during the online query processing. Besides, this mining procedure also improves the efficiency and accuracy of the S+V framework. Furthermore, we extend the pFOV model into a Bayesian FOV($\beta$FOV) model which improves the spatial recognition accuracy by more than 30 percent and also further alleviates visual computation. From a bayesian point of view, the likelihood of a certain POI being captured by phones is a prior probability in pFOV model which is represented as a posterior probability in$\beta$FOV model.Our experiments in the real-world and Oxford 5K datasets show promising recognition results. In order to provide a fine-grained annotation ground truth, we labeled a new dataset based on Oxford 5K and make it public available on the web. Our COSTAR mining techniqueoutperforms state-of-the-art approach on both dataset.

In this work, we present a novel project called DeepCamera(DC) for recognizing places-of-interest(POI) with smartphones. Our framework is based on deep convolutional neural networks(ConvNets) which are currently state-of-the-art solutions to vision recognition tasks such as our mission. We propose a novel ConvNet by introducing a new layer called "spatial layer" which captures spatial knowledge from a geographic view. As a result, both spatial and visual knowledge contribute to generating a hybrid probability distribution over all possible POI candidates. Furthermore, we compress multiple trained deep ConvNets into one single shallow net called "shNet" which achieves competitive performance with ensemble methods. Our preliminary experiments conducted on real-world dataset have shown promising POI recognition results.

In this paper, we present FGSS (Fashion Glasses Search System), an innovative cross-scenario eyeglasses retrieval system which automatically recognizes eyeglasses in real-world photos, e.g. the photo of a fashion girl with a stylish pair of eyeglasses, and retrieves a ranking list of visually similar product instances from the database. We propose a novel segmentation-free framework for FGSS to bridge two search gaps, semantic gap and feature gap, where a new type of keypoint-based scheme called EGYPT is tailored for eyeglasses to facilitate the search. In the EGYPT, we use the hybrid descriptors which combine the shape, color and texture features as a feature representation for eyeglasses. The experimental study on the real-world photo dataset and eyeglasses product dataset demonstrates the effectiveness of EGYPT model.

Short-text messages such as tweets are being created and shared at an unprecedented rate. Tweets, in their raw form, while being informative, can also be overwhelming. For both end-users and data analysts, it is a nightmare to plow through millions of tweets which contain enormous amount of noise and redundancy. In this paper, we propose a novel continuous summarization framework called Sumblr to alleviate the problem. In contrast to the traditional document summarization methods which focus on static and small-scale data set, Sumblr is designed to deal with dynamic, fast arriving, and large-scale tweet streams. Our proposed framework consists of three major components. First, we propose an online tweet stream clustering algorithm to cluster tweets and maintain distilled statistics in a data structure called tweet cluster vector (TCV). Second, we develop a TCV-Rank summarization technique for generating online summaries and historical summaries of arbitrary time durations. Third, we design an effective topic evolution detection method, which monitors summary-based/volume-based variations to produce timelines automatically from tweet streams. Our experiments on large-scale real tweets demonstrate the efficiency and effectiveness of our framework.

The solid-state hybrid drive (SSHD) incorporates a small NAND flash memory into a hard drive, resulting in an integrated device with combined Hard Disk Drive (HDD ) and Solid State Disk (SSD) storage. By identifying the data highly associated with the performance and buffering them in the SSD part, SSHD can deliver a better performance than the standard hard drive. However, that requires a significant redesign for existing data processing systems. In this paper, we examine the problem of efficiently processing relational data using MapReduce on a cluster using SSHDs as the underlying storage devices. We present the design of Hybrid MapReduce (HM ), a column-oriented MapReduce system, which adopts different storage layout, query optimizer, data index, and compression algorithm from previous MapReduce systems. In HM, the Distributed File System (DFS ) is deployed on SSHDs, and data layout (how data chunks are disseminated to HDDs and SSDs) plays a key role for the performance. Hence, an approximate algorithm is used to tune the data layout adaptively to maximize the query performance. We evaluate HM using TPC-H benchmark and the results show that with our new design, the hybrid system can provide a similar performance as the SSD-only system.

Geo-textual data are generated in abundance. Recent studies focused on the processing of spatial keyword queries which retrieve objects that match certain keywords within a spatial region. To ensure effective retrieval, various extensions were done including the allowance of errors in keyword matching and auto completion using prefix matching. In this paper, we propose INSPIRE, a general framework, which adopts a unifying strategy for processing different variants of spatial keyword queries. We adopt the auto completion paradigm that generates an initial query as a prefix matching query. If there are few matching results, other variants are performed as a form of relaxation that reuses the processing done in the earlier phase. The types of relaxation allowed include spatial region expansion and exact/approximate prefix/substring matching. Moreover, since the auto completion paradigm allows appending characters after the initial query, we look at how query processing done for the initial query and relaxation can be reused in such instances. Compared to existing works which process variants of spatial keyword query as new queries over different indexes, our approach offers a more compelling way to efficient and effective spatial keyword search. Extensive experiments substantiate our claims.

Singing is a popular social activity and a pleasant way of expressing one’s feelings. One important reason for unsuccessful singing performance is because the singer fails to choose a suitable song. In this paper, we propose a novel competence-based song recommendation framework for the purpose of singing. It is distinguished from most existing music recommendation systems which rely on the computation of listeners’ interests or similarity. We model a singer’s vocal competence as a singer profile, which takes voice pitch, intensity, and quality into consideration. Then we propose techniques to acquire singer profiles. We also present a song profile model which is used to construct a human annotated song database. Then we propose a learning-to-rank scheme for recommending songs by a singer profile. Finally, we introduce a reduced singer profile which can greatly simplify the vocal competence modelling process. The experimental study on real singers demonstrates the effectiveness of our approach and its advantages over two baseline methods.

In this paper, we consider complex pattern matching over event data generated from error-prone sources such as low-cost wireless motes, RFID. Such data are often imprecise in both their values and their timestamps. While there are existing works addressing the problem of spatial uncertainty (i.e. the uncertainty of the data values), relatively little attention has been paid to the problem of temporal uncertainty (i.e. the uncertainty of the event timestamps). As a step to fill this gap, we formulate the problem of matching complex sequence patterns over time-series data with temporal uncertainty and propose a new indexing structure to organize the information of the uncertain sequences and a set of efficient pattern query processing algorithms. We conduct an extensive experimental study on both synthetic and real datasets. The results indicate that the query processing algorithms based on our index structure can dramatically improve the query performance.

Nowadays, an increasing number of singing enthusiasts upload their cover songs and share their performances in online social singing communities. They can also listen and rate other users' song renderings. An important feature of the social singing communities is to recommend appropriate singing-songs which users are able to perform excellently.

This paper presents a project called Knowing Camera for real-time recognizing and annotating places-of-interest(POI) in smartphone photos, with the availability of online geotagged images of such places. We propose a`"Spatial+Visual" (S+V) framework which consists of a probabilistic field-of-view model in the spatial phase and sparse coding similarity metric in the visual phase to recognize phone-captured POIs. Moreover, we put forward an offline Collaborative Salient Area (COSTAR) mining algorithm to detect common visual features (called Costars) among the noisy photos geotagged on each POI, thus to clean the geotagged image database. The mining result can be utilized to annotate the region-of-interest on the query image during the online query processing. Besides, this mining procedure further improves the efficiency and accuracy of the S+V framework. Our experiments in the real-world and Oxford 5K datasets show promising recognition and annotation performances of the proposed approach, and that the proposed COSTAR mining technique outperforms state-of-the-art approach.

YZStack is our developing solution which implements many well-established big data techniques as selectable modules and allows users to customize their systems as a process of module selection. In particular, it includes an openstack based IaaS (Infrastructure as a Service) layer, a distributed file system based DaaS (Data as a Service) layer, a PaaS (Platform as a Service) layer equipped with parallel processing techniques and a SaaS (Software as a Service) layer with popular data analytic algorithms. Layers of YZStack are loosely connected, so that customization of one layer does not affect the other layers and their interactions. In this paper, we use a smart financial system developed for the Zhejiang Provincial Department of Finance to demonstrate how to leverage YZStack to speed up the implementation of big data system. We also introduce two popular applications of the financial system, economic prediction and detection of improper payment.

Driven by the advance of positioning technology and the popularity of location-sharing services, semantic-enriched trajectory data have become unprecedentedly available. The sequential patterns hidden in such data, when properly defined and extracted, can greatly benefit tasks like targeted advertising and urban planning. Unfortunately, classic sequential pattern mining algorithms developed for transactional data cannot effectively mine patterns in semantic trajectories, mainly because the places in the continuous space cannot be regarded as independent "items". Instead, similar places need to be grouped to collaboratively form frequent sequential patterns. That said, it remains a challenging task to mine what we call fine-grained sequential patterns , which must satisfy spatial compactness, semantic consistency and temporal continuity simultaneously. We propose Splitter to effectively mine such fine-grained sequential patterns in two steps. In the first step, it retrieves a set of spatially coarse patterns, each attached with a set of trajectory snippets that precisely record the pattern's occurrences in the database. In the second step, Splitter breaks each coarse pattern into fine-grained ones in a top-down manner, by progressively detecting dense and compact clusters in a higher-dimensional space spanned by the snippets. Splitter uses an effective algorithm called weighted snippet shift to detect such clusters, and leverages a divide-and-conquer strategy to speed up the top-down pattern splitting process. Our experiments on both real and synthetic data sets demonstrate the effectiveness and efficiency of Splitter.

Geo-textual data are ubiquitous these days. Recent study on spatial keyword search focused on the processing of queries which retrieve objects that match certain keywords within a spatial region. To ensure effective data retrieval, various extensions were done including the tolerance of errors in keyword matching and the search-as-you-type feature using prefix matching. We present MESA, a map application to support different variants of spatial keyword query. In this demonstration, we adopt the autocompletion paradigm that generates the initial query as a prefix matching query. If there are few matching results, other variants are performed as a form of relaxation that reuses the processing done in earlier phases. The types of relaxation allowed include spatial region expansion and exact/approximate prefix/substring matching. MESA adopts the client-server architecture. It provides fuzzy type-ahead search over geo-textual data. The core of MESA is to adopt a unifying search strategy, which incrementally applies the relaxation in an appropriate order to maximize the efficiency of query processing. In addition, MESA equips a user-friendly interface to interact with users and visualize results. MESA also provides customized search to meet the needs of different users.

Personalized web search (PWS) has demonstrated its effectiveness in improving the quality of various search services on the Internet. However, evidences show that users' reluctance to disclose their private information during search has become a major barrier for the wide proliferation of PWS. We study privacy protection in PWS applications that model user preferences as hierarchical user profiles. We propose a PWS framework called UPS that can adaptively generalize profiles by queries while respecting user-specified privacy requirements. Our runtime generalization aims at striking a balance between two predictive metrics that evaluate the utility of personalization and the privacy risk of exposing the generalized profile. We present two greedy algorithms, namely GreedyDP and GreedyIL, for runtime generalization. We also provide an online prediction mechanism for deciding whether personalizing a query is beneficial. Extensive experiments demonstrate the effectiveness of our framework. The experimental results also reveal that GreedyIL significantly outperforms GreedyDP in terms of efficiency.

The amount of data managed in today's Cloud systems has reached an unprecedented scale. In order to speed up query processing, an effective mechanism is to build indexes on attributes that are used in query predicates. However, conventional indexing schemes fail to provide a scalable service: as the size of these indexes are proportional to the data size, it is not space efficient to build many indexes. As such, it becomes more crucial to develop effective index to provide scalable database services in the Cloud. In this paper, we propose a compact bitmap indexing scheme for a large-scale data store. The bitmap indexing scheme combines state-of-the-art bitmap compression techniques, such as WAH encoding and bit-sliced encoding. To further reduce the index cost, a novel and query efficient partial indexing technique is adopted, which dynamically refreshes the index to handle updates and process queries. The intuition of our indexing approach is to maximize the number of indexed attributes, so that a wider range of queries, including range and join queries, can be efficiently supported. Our indexing scheme is light-weight and its creation can be seamlessly grafted onto the MapReduce processing engine without incurring significant running cost. Moreover, the compactness allows us to maintain the bitmap indexes in memory so that performance overhead of index access is minimal. We implement our indexing scheme on top of the underlying Distributed File System (DFS) and evaluate its performance on an in-house cluster. We compare our index-based query processing with HadoopDB to show its superior performance. Our experimental results confirm the effectiveness, efficiency and scalability of the indexing scheme.

This paper presents a framework called Knowing Camera for real-time recognizing places-of-interest in smartphone photos, with the availability of online geotagged images of such places. We propose a probabilistic field-of-view model which captures the uncertainty in camera sensor data. This model can be used to retrieve a set of candidate images. The visual similarity computation of the candidate images relies on the sparse coding technique. We also propose an ANN filtering technique to speedup the sparse coding. The final ranking combines an uncertain geometric relevance with the visual similarity. Our preliminary experiments conducted in an urban area of a large city show promising results. The most distinguishing feature of our framework is its ability to perform well in contaminated, real-world online image database. Besides, our framework is highly scalable as it does not incur any complex data structure.

Singing is a popular social activity and a good way of expressing one's feelings. One important reason for unsuccessful singing performance is because the singer fails to choose a suitable song. In this paper, we propose a novel singing competence-based song recommendation framework. It is distinguished from most existing music recommendation systems which rely on the computation of listeners' interests or similarity. We model a singer's vocal competence as singer profile, which takes voice pitch, intensity, and quality into consideration. Then we propose techniques to acquire singer profiles. We also present a song profile model which is used to construct a human annotated song database. Finally, we propose a learning-to-rank scheme for recommending songs by singer profile. The experimental study on real singers demonstrates the effectiveness of our approach and its advantages over two baseline methods. To the best of our knowledge, our work is the first to study competence-based song recommendation.

With the explosive growth of microblogging services, short-text messages (also known as tweets) are being created and shared at an unprecedented rate. Tweets in its raw form can be incredibly informative, but also overwhelming. For both end-users and data analysts it is a nightmare to plow through millions of tweets which contain enormous noises and redundancies. In this paper, we study continuous tweet summarization as a solution to address this problem. While traditional document summarization methods focus on static and small-scale data, we aim to deal with dynamic, quickly arriving, and large-scale tweet streams. We propose a novel prototype called Sumblr (SUMmarization By stream cLusteRing) for tweet streams. We first propose an online tweet stream clustering algorithm to cluster tweets and maintain distilled statistics called Tweet Cluster Vectors. Then we develop a TCV-Rank summarization technique for generating online summaries and historical summaries of arbitrary time durations. Finally, we describe a topic evolvement detection method, which consumes online and historical summaries to produce timelines automatically from tweet streams. Our experiments on large-scale real tweets demonstrate the efficiency and effectiveness of our approach.

Similarity search on spatiotemporal trajectories has a wide range of applications. Most of existing research focuses on certain trajectories. However, trajectories often are uncertain due to various factors, for example, hardware limitations and privacy concerns. In this paper, we introduce p-distance, a novel and adaptive measure that is able to quantify the dissimilarity between two uncertain trajectories. Based on this measure of dissimilarity, we define top-k similarity query (KSQ) on uncertain trajectories. A KSQ returns the k trajectories that are most similar to a given trajectory in terms of p-distance. To process such queries efficiently, we design UTgrid for indexing uncertain trajectories and develop query processing algorithms that make use of UTgrid for effective pruning. We conduct an extensive experimental study on both synthetic and real data sets. The results indicate that UTgrid is an effective indexing method for similarity search on uncertain trajectories. Our query processing using UTgrid dramatically improves the query performance and scales well in terms of query time and I/O.

Time series is an important form of data available in numerous applications and often contains vast amount of personal privacy. The need to protect privacy in time-series data while effectively supporting complex queries on them poses nontrivial challenges to the database community. We study the anonymization of time series while trying to support complex queries, such as range and pattern matching queries, on the published data. The conventional k-anonymity model cannot effectively address this problem as it may suffer severe pattern loss. We propose a novel anonymization model called (k, P)-anonymity for pattern-rich time series. This model publishes both the attribute values and the patterns of time series in separate data forms. We demonstrate that our model can prevent linkage attacks on the published data while effectively support a wide variety of queries on the anonymized data. We propose two algorithms to enforce (k, P)-anonymity on time-series data. Our anonymity model supports customized data publishing, which allows a certain part of the values but a different part of the pattern of the anonymized time series to be published simultaneously. We present estimation techniques to support query processing on such customized data. The proposed methods are evaluated in a comprehensive experimental study. Our results verify the effectiveness and efficiency of our approach.

The emerging location-based social network (LBSN) services not only allow people to maintain cyber links with their friends, but also enable them to share the events happening on them at different locations. The geo-social correlations among event participants make it possible to quantify mutual user influence for various events. Such a quantification of influence could benefit a wide spectrum of real-life applications such as targeted advertising and viral marketing.

No abstract available.

In this demo, we present myDJ, a karaoke recommendation system which recommends the songs people are capable to sing. Different from the existing song recommendation systems which recommend songs people like to listen, myDJ can recommend proper songs according to a subject's physical phonation area. It consists of a singer profiler to analyze the subject's phonation characters. In addition, the song profile for each song in database is extracted. To learn a ranking function, the learning to rank algorithm Listnet is applied under a list of predefined features extracted from each singer-song profile pair. In the results, proper songs which are suitable but challenging for the subject are recommended.

The wide proliferation of powerful smart phones equipped with multiple sensors, 3D graphical engine, and 3G connection has nurtured the creation of a new spectrum of visual mobile applications. These applications require novel data retrieval techniques which we call What-You-Retrieve-Is-What-You-See (WYRIWYS). However, state-of-the-art spatial retrieval methods are mostly distance-based and thus inapplicable for supporting WYRIWYS. Motivated by this problem, we propose a novel query called spatio-visual keyword (SVK) query, to support retrieving spatial Web objects that are both visually conspicuous and semantically relevant to the user. To capture the visual features of spatial Web objects with extents, we introduce a novel visibility metric which computes object visibility in a cumulative manner. We propose an incremental method called Complete Occlusion-map based Retrieval (COR) to answer SVK queries. This method exploits effective heuristics to prune the search space and construct a data structure called Occlusion-Map. Then the method adopts the best-first strategy to return relevant objects incrementally. Extensive experiments on real and synthetic data sets suggest that our method is effective and efficient when processing SVK queries.

Skyline query is important in the circumstances that require the support of decision making. The existing work on skyline queries is based mainly on the assumption that the datasets are static. Querying skylines over moving objects, however, is also important and requires more attention. In this paper, we propose a framework, namely PRISMO, for processing predictive skyline queries over moving objects that not only contain spatio-temporal information, but also include non-spatial dimensions, such as other dynamic and static attributes. We present two schemes, RBBS (branch-and-bound skyline with rescanning and repacking) and TPBBS (time-parameterized branchand-bound skyline), each with two alternative methods, to handle predictive skyline computation. The basic TPBBS is further extended to TPBBSE (TPBBS with expansion) to enhance the performance of memory space consumption and CPU time. Our schemes are flexible and thus can process point, range, and subspace predictive skyline queries. Extensive experiments show that our proposed schemes can handle predictive skyline queries effectively, and that TPBBS significantly outperforms RBBS.

In many applications, such as online dating or job hunting websites, users often need to search for potential matches based on the requirements or preferences imposed by both sides.We refer to this type of queries as matching queries. In spite of their wide applicabilities, there has been little attention devoted to improve their performance. As matching queries often appear in various forms even within a single application, we, in this paper, propose a general processing framework, which can efficiently process various forms of matching queries. Moreover, we elaborate the detailed processing algorithms for two particular forms of matching queries to illustrate the applicability of this framework. We conduct an extensive experimental study with both synthetic and real datasets. The results indicate that, for various matching queries, our techniques can dramatically improve the query performance, especially when the dimensionality is high.

In recent years, personalized web search (PWS) has demonstrated effectiveness in improving the quality of search service on the Internet. Unfortunately, the need for collecting private information in PWS has become a major barrier for its wide proliferation. We study privacy protection in PWS engines which capture personalities in user profiles. We propose a PWS framework called UPS that can generalize profiles in for each query according to user-specified privacy requirements. Two predictive metrics are proposed to evaluate the privacy breach risk and the query utility for hierarchical user profile. We develop two simple but effective generalization algorithms for user profiles allowing for query-level customization using our proposed metrics. We also provide an online prediction mechanism based on query utility for deciding whether to personalize a query in UPS. Extensive experiments demonstrate the efficiency and effectiveness of our framework.

The cyber-physical systems (CPS) are envisioned as a class of real-time systems integrating the computing, communication and storage facilities with monitoring and control of the physical world. One interesting CPS application in the mobile Internet is to provide Web search "on the spot" regarding the physical world that a user sees, or literally WYRIWYS (What-You-Retrieve-Is-What-You-See). The objective of our work is to develop server/browser software for supporting WYRIWYS search in our prototype cyber-physical search engine. A WYRIWYS search retrieves visible Web objects and ranks them by their cyber-physical relevances (term, visual, spatial, temporal etc.). This work is distinguished from previous LWS as it provides quality Web search geared with the physical world. Therefore it suggests a very promising solution to cyber-physical Web search.

In this poster, we address the problem of location disambiguation for geotagged Web photo resources. We propose an approach for analyzing and partitioning large geotagged photo collections using geographic and semantic information. By organizing the dataset in a structural scheme, we resolve the location ambiguity and clutter problem yield by massive volume of geotagged photos.

In multi-criterion decision making applications, a skyline query narrows the search range, as it returns only the points that are not dominated by others. Unfortunately, in high-dimensional/large-cardinal datasets there exist too many skyline points to offer interesting insights. In this paper, we propose a novel structure, called the dominance tree (Do-Tree), to effectively index and retrieve the skyline. Do-Tree is a straightforward and flexible tree structure, in which skyline points are resident on leaf nodes, while the internal nodes contain the entries that dominate their children. As Do-Tree is built on a dominance relationship, it is suitable for the retrieval of specified skyline via dominance-based predicates customized by users. We discuss the topology of Do-Tree and propose the construction methods. We also present the scan scheme of Do-Tree and some useful queries based on it. Extensive experiments confirm that Do-Tree is an efficient and scalable index structure for the skyline.

The challenges with privacy protection of time series are mainly due to the complex nature of the data and the queries performed on them. We study the anonymization of time series while trying to support complex queries, such as range and pattern similarity queries, on the published data. The conventional k-anonymity cannot effectively address this problem as it may suffer severe pattern loss. We propose a novel anonymization model called (k,P)-anonymity for pattern-rich time series. This model publishes both the attribute values and the patterns of time series in separate data forms. We demonstrate that our model can prevent linkage attacks on the published data while effectively support a wide variety of queries on the anonymized data. We also design an efficient algorithm for enforcing (k,P)-anonymity on time series data.

Many data stream monitoring applications involve rank queries and hence a number of efficient evaluation algorithms are proposed recently. Most of these techniques assume that rank queries are executed directly over the whole data space. However, we observe that many applications often require to perform clustering over the data streams before rank queries are run on each cluster. To address the problem, we propose a novel algorithm for integral clustering and ranking processing and we refer to such integrated queries as cluster-based rank queries. The algorithm includes two phases, namely the online phase which maintains the required data structures and statistics, and the query phase which uses these data structures to process queries. Extensive experiments indicate that the proposed algorithm is efficient in both space consumption and query processing.

Thanks to standard specifications for Web services, more and more data is represented and exchanged in XML. Efficiently finding all occurrences of a twig pattern is a core operation of XML query processing. Recently, a large number of holistic twig join algorithms have been proposed, however ordered XML twig join has been ignored in most of them. In this paper we design a novel algorithm called OTJFast to process ordered XML twig joins efficiently which only need to access the elements of the leaf query nodes. Experimental results show the benefits of our algorithm in term of the size of intermediate results, the number of element reads and the execution time.

Efficiently finding all occurrences of a twig pattern is a core operation of XML query processing. In this paper, we first propose a generic twig join algorithm TJFGeneric which only scans the leaf query nodes using a cursor interface to return elements. And then we present TJFGeneric+algorithm incorporating two optimization rules into TJFGeneric to explore potential benefit of disk read and CPU cost. In particular, available indices (such as B+-tree) are applied to skip the elements that do not participate in final results. Experimental results on various datasets indicate that our algorithms outperform previous approaches significantly.

In today's information society, it is important to retrieve fresh information. Many of the vertical search results are valid to users in only a short period of time. But due to resource constraints, it is not possible to keep the entire local storage synchronized with the Web. We implemented a TSVS (time sensitive vertical search engine) prototype named Velocisaurus focused on time-critical airfare discount information search to investigate the time critical requirements of vertical search and proposed a QTC (query triggered crawling) strategy to coordinate the crawling systems by real-time user queries and solve this problem. Experiment shows that QTC driven crawlers significantly improves the freshness of the search results and utilizes the resources more efficiently compared to regular search engine crawlers..

In this paper, we propose CYBER, a CommunitY Based sEaRch engine, for information retrieval utilizing community feedback information in a DHT network. In CYBER, each user is associated with a set of user profiles that capture his/her interests. Likewise, a document is associated with a set of profiles - one for each indexed term. A document profile is updated by users who query on the term and consider the document as a relevant answer. Thus, the profile acts as a consolidation of users feedback from the same community, and reflects their interests. In this way, as one user finds a document to be relevant, another user in the same community issuing a similar query will benefit from the feedback provided by the earlier user. Hence, the search quality in terms of both precision and recall is improved. We conduct a comprehensive experimental study and the results show the effectiveness of our scheme.

We propose a novel iterative searching and refining prototype for tagged images. This prototype, named PivotBrowser, captures semantically similar tag sets in a structure called pivot. By constructing a pivot for a textual query, PivotBrowser first selects candidate images possibly relevant to the query. The tags contained in these candidate images are then selected in terms of their tag relevances to the pivot. The shortlisted tags are clustered and one of the tag clusters is used to select the results from the candidate images. Ranking of the images in each partition is based on their relevance to the tag cluster. With the guidance of the tag clusters presented, a user is able to perform searching and iterative query refinement.

In this paper, we propose an image semantic model based on the knowledge and criteria in the field of linguistics and taxonomy. Our work bridges the "semantic gap" by seamlessly exploiting the synergy of both visual feature processing and semantic relevance computation in a new way, and provides improved query efficiency and effectiveness for large general image databases. Our main contributions are as follows: we design novel data structures, namely, a lexical hierarchy, an image-semantic hierarchy, and a number of atomic semantic domains, to capture the semantics and the features of the database, and to provide the indexing scheme. We present a novel image query algorithm based on the proposed structures. In addition, we propose a novel term expansion mechanism to improve the lexical processing. Our extensive experiments indicate that our proposed techniques are effective in achieving high runtime performance with improved retrieval accuracy. The experiments also show that the proposed method has good scalability.

This paper presents a framework for design of online courses and collaborative learning based on Web technologies. It provides the following features: construction of structured learning materials in a distributed environment; personalization of course contents in the run-time that adapts to the progress of individual students; it can supply relevant Web resources to the students by analyzing the learning history dynamically; students in a group could exchange their learning experiences and discover relevant knowledge via the communication channels provided by the framework. The proposed techniques could be deployed to a variety of online computer-assisted modules to improve the effectiveness of Web-based e-learning

In this paper, we present a novel structure called the hierarchical degree-of-visibility tree (HDoV-tree) for visibility query processing in visualization systems. The HDoV-tree builds on and extends the R-tree such that 1) the search space is pruned based on the degree of visibility of objects and 2) internal nodes store level-of-details (LoDs) that represent a collection of objects in a coarser form. We propose two tree traversal algorithms that balance performance and visual fidelity, explore three storage structures for the HDoV-tree, and develop novel caching techniques for disk-based HDoV-tree. We implemented the HDoV-tree in a prototype walkthrough system called VISUAL. Our experimental study shows that VISUAL can lead to high frame rates without compromising visual fidelity.

In a visualization system, one of the key issues is to optimize performance and visual fidelity. This is especially critical for large virtual environments where the models do not fit into the memory. Here, we present a novel structure called HDoV-tree that can be tuned to provide excellent visual fidelity and performance based on the degree of visibility of objects. HDoV-tree also exploits internal level-of-details (LoDs) that represent a collection of objects in a coarser form. We also propose three storage structures for the HDoV-tree. We implemented HDoV-tree in a prototype walkthrough system called VISUAL. We have evaluated the HDoV-tree on visibility queries, and also compared the performance of VISUAL against REVIEW, a walkthrough system based on R-tree. Our results show that the HDoV-tree is an efficient structure. Moreover, VISUAL can lead to high frame rates without compromising visual fidelity.

No abstract available.