The prediction objectives of online advertisement ranking models are evolving from probabilistic metrics like conversion rate (CVR) to numerical business metrics like post-click gross merchandise volume (GMV). Unlike the well-studied delayed feedback problem in CVR prediction, delayed feedback modeling for GMV prediction remains unexplored and poses greater challenges, as GMV is a continuous target, and a single click can lead to multiple purchases that cumulatively form the label.

In industrial recommender systems, conversion rate (CVR) is often used for traffic allocation, but fails to fully reflect recommendation effectiveness as it does not account for refund rate (RFR). Thus, net conversion rate (NetCVR), the probability that a clicked item is purchased and not refunded, is proposed to better show true user satisfaction and business value. Unlike CVR, NetCVR prediction involves a more complex multi-stage cascaded delay feedback phenomenon. The two cascaded delays Click->Conversion and Conversion->Refund in NetCVR have opposite effects. Therefore, traditional CVR methods cannot be directly applied. At present, the lack of relevant open-source datasets and online continuous training schemes poses a challenge. To address these, we first introduce CAscadal Sequences of Conversion And Delayed rEfund (CASCADE), the first large-scale open dataset derived from Taobao app for online continuous NetCVR prediction. We further analyze CASCADE and derive three key insights: (1) NetCVR exhibits clear temporal patterns necessitating online continuous modeling; (2) Cascaded modeling CVR and RFR for NetCVR outperforms directly modeling NetCVR; and (3) delay time, which correlated with both CVR and RFR, is an important feature for NetCVR prediction. Based on these insights, we propose neT convErsion caScaded modeLing and debiAsing method (TESLA). This continuous method features a CVR-RFR cascaded architecture, stage-wise debiasing, and a delay-time-aware ranking loss for efficient NetCVR prediction. Experiments show that TESLA outperforms state-of-the-art methods on CASCADE, achieving an absolute improvement of 12.41% in RI-AUC and 14.94% in RI-PRAUC on NetCVR over the strongest baseline. We hope this work provides a new direction for online delayed feedback modeling in NetCVR prediction. Our code and dataset are available at https://github.com/alimama-tech/NetCVR.

Recent advances in large language models (LLMs) have enabled more semantic-aware recommendations through natural language generation. Existing LLM for recommendation (LLM4Rec) methods mostly operate in a System 1-like manner, relying on superficial features to match similar items based on click history, rather than reasoning through deeper behavioral logic. This often leads to superficial and erroneous recommendations. Motivated by this, we propose ThinkRec, a thinking-based framework that shifts LLM4Rec from System 1 to System 2 (rational system). Technically, ThinkRec introduces a thinking activation mechanism that augments item metadata with keyword summarization and injects synthetic reasoning traces, guiding the model to form interpretable reasoning chains that consist of analyzing interaction histories, identifying user preferences, and making decisions based on target items. On top of this, we propose an instance-wise expert fusion mechanism to reduce the reasoning difficulty. By dynamically assigning weights to expert models based on users' latent features, ThinkRec adapts its reasoning path to individual users, thereby enhancing precision and personalization. Extensive experiments on real-world datasets demonstrate that ThinkRec significantly improves the accuracy and interpretability of recommendations. Our implementations are available at https://github.com/Yu-Qi-hang/ThinkRec.

An ideal index advisor needs to effectively manage changes in workload and data, but current approaches fall short in both effectiveness and efficiency because of intrinsic limitations in their frameworks. Heuristic-based methods struggle with efficiency due to their rigid algorithms and lack of adaptive learning capabilities. Reinforcement learning techniques often fail to consistently reach an optimal policy. Classification methods require vast amounts of labeled workloads that include optimal indexes. Additionally, none of the learning-based strategies are equipped to handle shifts in data. To overcome these limitations, this paper presents a new index advisor for dynamic workloads and data, GenIA, which learns to generate a sequence of the recommended index configuration based on historical experience. The generative framework of GenIA avoids erroneous trials to explore bad actions and reliance on high-quality positive and negative examples. Specifically, its novelty exhibits in three aspects. (1) GenIA is empowered with novel attention mechanisms to capture implicit relationships between indexable columns. (2) GenIA combines comprehensive features extracted from workloads, data manipulation statements, and underlying data to effectively capture workload shifts and subtle data shifts. (3) GenIA adopts a novel perturbation-based training strategy to enhance the diversity of training samples and to improve the model parameters' robustness. Extensive experiments on various benchmarks under varying levels of workload and data shifts demonstrate that GenIA outperforms SOTA heuristic-based IA Extend on average by about 7.5%, while utilizing less than 1% of the inference time, and surpasses SOTA learning-based IA SWIRL by 25% − 30% in scenarios with significant workload and data shifts.

Sequential recommendation systems have become essential for personalized services in e-commerce and content platforms. While recent research has extended these systems with multi-modal features, existing approaches face three major challenges. First, they inadequately model fine-grained temporal interval distributions, failing to discriminate between high-frequency short intervals and low-frequency long intervals. Second, uniform fusion in the time domain leads to semantic misalignment across modalities because it ignores their inherent differences in the frequency domain. Third, rigid fusion strategies without self-supervised constraints lead to limited representation quality and semantic drift from pretrained embeddings. To address these issues, we propose Adaptive Temporal Expert Routing with Hierarchical Wavelet Enhancement (ATHWE) framework. ATHWE employs exponential saturation time mapping to generate temporally adaptive embeddings. These embeddings guide a sparse mixture of experts to model multi-scale user behavior dynamics. A hierarchical wavelet decomposition with band-specific gating selectively fuses complementary frequency components across modalities. Furthermore, contrastive learning and cluster-preserving objectives preserve semantic information during multi-modal fusion. Extensive experiments on multiple datasets validate the effectiveness of our framework. Our code is available at https://github.com/lulusiyuyu/ATHWE .

Aligning the multi-modal content and ID embeddings is crucial in multi-modal recommendation systems. Existing solutions typically adopt a bidirectional alignment paradigm. Our prior work, FETTLE , challenges this paradigm by proposing a one-way directional alignment at the item level, thus reducing the negative impact of low-quality modalities. However, FETTLE leaves two open questions: (1) when is one-way directional alignment optimal, and (2) how to incorporate collaborative signals to enhance alignment? We present CROSS (feedba C k-o R iented multi-m O dal alignment in recommendation S y S tem), a plug-and-play framework that extends FETTLE by introducing three major advancements. First, we introduce Dynamic Item-Level Alignment , which dynamically calibrates the “strength” of each modality via a variance-based compensation mechanism, mitigating the risk of overshadowing weaker modalities in the early stages of training. Second, we develop Multi-grained Collaborative Alignment , which introduces a medium-granularity alignment strategy based on neighboring items that share similar user feedback profiles. This neighbor-level alignment effectively balances noisy user interactions and excessive smoothing across items. Third, we conduct extensive experiments on more real-world datasets and show that CROSS significantly boosts the performance of both collaborative filtering (CF) models and multi-modal recommendation (MRS) approaches, achieving 21.52%–70.78% average improvement on CF backbones and 8.70%–20.73% on MRS backbones. Compared with FETTLE , CROSS achieves additional improvements of 3.82%–5.24%.

Recently, large-language models (LLMs) have demonstrated strong potential to solve database problems. However, LLMs still face two challenges in solving the index selection problem: (1) representing the workload in an LLM-friendly form and (2) finding the optimal index set. To solve these challenges, we propose LLM4IA, an LLM-based index selection method that can recommend indexes for any analytical workload directly on any database instance. LLM4IA proposes a concise description of natural language by extracting and sorting predicates and completely avoiding numerical input. LLM4IA adopts an iterative index selection process by repeatedly improving previous index candidates and summarizing effective candidates. Experiments on TPC-H and TPC-DS show that LLM4IA surpasses the near-optimal index advisor Extend by 5%-10%. Our demonstration highlights how LLM4IA recommends high-quality indexes for a new database instance without expensive retraining or fine-tuning.

The development of Large Language Models (LLMs) has revolutionized QA across various industries, including the database domain. However, there lacks a thorough evaluation regarding the capabilities of different LLMs in database QA. To this end, we introduce DQABench, the first comprehensive database QA benchmark for LLMs. DQABench features an innovative LLM-based method to automate the generation, cleaning, and rewriting of evaluation dataset, resulting in over 200,000 QA pairs in English and Chinese. These QA pairs cover a wide range of database-specific knowledge extracted from manuals, online communities, and DB instances, allowing for assessment of LLMs' Retrieval-Augmented Generation (RAG) and Tool Invocation Generation (TIG) capabilities in the database QA task. Furthermore, we propose a highly modular and scalable testbed DQATestbed, with basic and advanced components such as Fine-tuning, Question Classification Routing (QCR), RAG, TIG, and Prompt Template Engineering (PTE). Finally, we provide an evaluation pipeline that computes various metrics throughout a standardized evaluation process to ensure the accuracy and fairness. Our evaluation reveals the strengths and limitations of nine open-source and commercial LLMs, and the impact of various service components (e.g., fine-tuning, QCR, RAG, TIG). The proposed benchmark dataset is available at https://github.com/XMUDM/DQABench.

Many index advisors have recently been proposed to build indexes automatically to improve query performance. However, they mainly consider performance improvement in static scenarios. Their robustness, i.e., stable performance in dynamic scenarios (e.g., with minor workload changes), has not been well investigated. This paper addresses the challenges of assessing the index advisor's robustness from the following aspects. First, we introduce perturbation-based workloads for robustness assessment and identify three typical perturbation constraints that occur in real scenarios. Second, with the perturbation constraints, we formulate the generation of perturbed queries as a sequence-to-sequence problem and propose TRAP (Tailored Robustness assessment via Adversarial Perturbation) to pinpoint the performance loopholes of index advisors. Third, to generalize to various index advisors, we place TRAP in an opaque-box setting (i.e., with little knowledge of the index advisors' internal design), and we propose a two-phase training paradigm to efficiently train TRAP without elaborately annotated data. Fourth, we conduct comprehensive robustness assessments on standard benchmarks and real workloads for ten existing index advisors. Our findings reveal that these index advisors are vulnerable to the workloads generated by TRAP. Finally, based on the assessment results, we shed light on insights to enhance the robustness of different index advisors. For example, learning-based index advisors can benefit from adopting a fine-grained state representation and a candidate pruning strategy.

Multi-modal recommender systems (MRSs) are pivotal in diverse online web platforms and have garnered considerable attention in recent years. However, previous studies overlook the challenges of (1)noisy multi-modal content, (2) noisy user feedback, and (3) aligning multi-modal content and user feedback. To tackle these challenges, we propose Denoising and Aligning Multi-modal Recommender System (DA-MRS). To mitigate noise in multi-modal content, DA-MRS first constructs item-item graphs determined by consistent content similarity across modalities. To denoise user feedback, DA-MRS associates the probability of observed feedback with multi-modal content and devises a denoised BPR loss. Furthermore, DA-MRS implements Alignment guided by User preference to enhance task-specific item representation and Alignment guided by graded Item relations to provide finer-grained alignment. Extensive experiments verify that DA-MRS is a plug-and-play framework and achieves significant and consistent improvements across various datasets, backbone models, and noisy scenarios.

Multi-modal Recommendation Systems (MRSs) utilize diverse modalities, such as image and text, to enrich item representations and enhance recommendation accuracy. Current MRSs overlook the large misalignment between multi-modal content features and ID embeddings. While bidirectional alignment between visual and textual modalities has been extensively studied in large multi-modal models, this study suggests that multi-modal alignment in MRSs should be in a one-way direction. A plug-and-play framework is presented, called FEedback-orienTed mulTi-modal aLignmEnt (FETTLE). FETTLE contains three novel solutions: (1) it automatically determines item-level alignment direction between each pair of modalities based on estimated user feedback; (2) it coordinates the alignment directions among multiple modalities; (3) it implements cluster-level alignment from both user and item perspectives for more stable alignments. Extensive experiments on three real datasets demonstrate that FETTLE significantly improves various backbone models. Conventional collaborative filtering models are improved by 24.79%-62.79%, and recent MRSs are improved by 5.91% - 20.11%.

Despite the promising performance of recent learning-based Index Advisors (IAs), they exhibited the robustness issue when poisoning attacks polluted training data. This paper presents the first attempt to study the robustness of updatable learning-based IAs against poisoning attack, i.e., whether the IAs can maintain robust performance if their training/updating is disturbed by injecting an extraneous toxic workload. The goal is to provide an opaque-box stress test that is generally effective in evaluating the robustness of different learning-based IAs without using the users' private data. There are three challenges, i.e., how to probe "index preference" from opaque-box IAs, how to design effective injecting strategies even if the IAs can be fine-tuned, and how to generate queries to meet the specific constraints for IA probing and injecting. The presented stress-test framework PIPA consists of a probing stage, an injecting stage, and a query generator. To address the first challenge, the probing stage estimates the IA's indexing preference by observing its responses to the probing workload. To address the second challenge, the injecting stage injects workloads that spoof the IA to demote the top-ranked indexes in the estimated indexing preference and promote mid-ranked indexes. The stress test is effective because the IA is trapped in a local optimum even after fine-tuning. To address the third challenge, PIPA utilizes IABART (Index Aware BART) to generate queries that can be optimized by building indexes on a given set of indexes. Extensive experiments on different benchmarks against various learning-based IAs demonstrate the effectiveness of PIPA and that existing learning-based IAs are non-robust when faced with even a subtle amount of injected extraneous toxic workloads.

Current index advisors often struggle to balance efficiency and effectiveness when dealing with workload shifts. This arises from ignorance of the continual similarity and distant variety in workloads. This paper proposes a novel learning-based index advisor called BALANCE, which boosts indexing performance by leveraging knowledge obtained from dynamic and heterogeneous workloads. Our approach consists of three components. First, we build separate Lightweight Index Advisors (LIAs) on sequential chunks of similar workloads, where each LIA is trained with a small batch of workloads drawn from the chunk, and it provides direct index recommendations for all workloads in the same chunk. Second, we perform a policy transfer mechanism by adapting the LIA's index selection strategy from historical knowledge, substantially reducing the training overhead. Third, we employ a self-supervised contrastive learning method to provide an off-the-shelf workload representation, enabling the LIA to generate more accurate index recommendations. Extensive experiments across various benchmarks demonstrate that BALANCE improves the state-of-the-art learning-based index advisor, SWIRL, by 10.03% while reducing training overhead by 35.70% on average.

Index advisors aim to improve workload performance by judiciously selecting an appropriate set of indexes. Various heuristic-based and learning-based methods have been proposed. However, there lacks a comprehensive assessment of existing index advisors, i.e., their advantages, limitations, and application scenarios. In this work, we conduct an in-depth study of existing index advisors in five key aspects. First, we initiate an end-to-end analysis, i.e., a completed analysis throughout the entire workflow of index advisors. We decompose index advisors into three essential building blocks, establish a taxonomy to classify methods used in each block, and analyze the strengths and weaknesses of these methods. Second, we develop a unified open-source testbed, implementing seventeen index advisors across eleven open-source or real-world datasets. We enable customizable configurations to meet diverse testing requirements. Third, we conduct an extensive assessment of index advisors across database systems in various scenarios. We evaluate their adaptability and robustness, identifying practical application scenarios. Fourth, we undertake a fine-grained ablation study by investigating variants of each building block. We identify effective variants and pinpoint significant factors impacting index advisors' performance via explainable machine-learning techniques. Lastly, we consolidate our findings that could shed light on research directions to advance the future development of index advisors.

Popularity bias significantly impacts news recommendation systems, as popular news articles receive more exposure and are often delivered to irrelevant users, resulting in unsatisfactory performance. Existing methods have not adequately addressed the issue of popularity bias in news recommendations, largely due to the neglect of the time factor and the impact of news content on popularity. In this paper, we propose a novel approach called Time and Content-aware Causal Model, namely TCCM. It models the effects of three factors on user interaction behavior, i.e., the time factor, the news popularity, and the matching between news content and user interest. TCCM also estimates news popularity more accurately by incorporating the news content, i.e., the popularity of entity and words. Causal intervention techniques are applied to obtain debiased recommendations. Extensive experiments on well-known benchmark datasets demonstrate that the proposed approach outperforms a range of state-of-the-art techniques.

Mobile Taobao Application delivers search services on multiple scenarios that take textual, visual, or product queries. This paper aims to propose a unified graph neural network for these search scenarios to leverage data from multiple scenarios and jointly optimize search performances with less training and maintenance costs. Towards this end, this paper proposes BOMGraph, BOosting Multi-scenario E-commerce Search with a unified Graph neural network. BOMGraph is embodied with several components to address challenges in multi-scenario search. It captures heterogeneous information flow across scenarios by inter-scenario and intra-scenario metapaths. It learns robust item representations by disentangling specific characteristics for different scenarios and encoding common knowledge across scenarios. It alleviates label scarcity and long-tail problems in scenarios with low traffic by contrastive learning with cross-scenario augmentation. BOMGraph has been deployed in production by Alibaba's E-commerce search advertising platform. Both offline evaluations and online A/B tests demonstrate the effectiveness of BOMGraph.

Index advisors have become an essential tool to optimize index selection and accelerate query processing. Various index advisors have been developed in recent years, and comprehensively assessing their performance from multiple aspects is necessary. In this demonstration, we introduce VITA, a user-friendly and informative tool for interactively Visualizing, Testing, and Analyzing index advisors. For a user-given workload, VITA can visualize the main steps of the index selection procedure in ten existing index advisors to facilitate the management of index advisors. Moreover, VITA can assess the index advisor's robustness w.r.t. workload drift by generating testing workloads, i.e., potentially future workloads that may damage the index advisor's performance. Finally, VITA provides a comparative analysis across index advisors on four aspects, including the index advisor's utility (i.e., the ratio of the reduced workload cost), robustness (i.e., the performance under dynamic workload), overhead (i.e., the time to acquire the final configuration), and scalability (i.e., the volume of the enumerated index candidates). Therefore, VITA can thoroughly compare existing index advisors to help users determine the most suitable index advisor that meets their requirements. VITA is now being integrated into the openGauss platform as a plug-in.

Recently, many E-commerce search models are based on Graph Neural Networks (GNNs). Despite their promising performances, they are (1) lacking proper semantic representation of product contents; (2) less efficient for industry-scale graphs; and (3) less accurate on long-tail queries and cold-start products. To address these problems simultaneously, this paper proposes CC-GNN, a novel Content Collaborative Graph Neural Network. Firstly, CC-GNN enables content phrases to participate explicitly in graph propagation to capture the proper meaning of phrases and semantic drifts. Secondly, CC-GNN presents several efforts towards a more scalable graph learning framework, including efficient graph construction, MetaPath-guided Message Passing, and Difficulty-aware Representation Perturbation for graph contrastive learning. Furthermore, CC-GNN adopts Counterfactual Data Supplement at both supervised and contrastive learning to resolve the long-tail/cold-start problems. Extensive experiments on a real E-commerce dataset of 100-million-scale nodes show that CC-GNN produces significant improvements over existing methods (i.e., more than 10% improvements in terms of several key evaluation metrics for overall, long-tail queries and cold-start products) while reducing computational complexity. The proposed components of CC-GNN can be applied to other models for search and recommendation tasks. Experiments on a public dataset show that applying the proposed components can improve the performance of different recommendation models.

Knowledge distillation is an effective way to transfer knowledge from a strong teacher to an efficient student model. Ideally, we expect the better the teacher is, the better the student performs. However, this expectation does not always come true. It is common that a strong teacher model results in a bad student via distillation due to the nonnegligible gap between teacher and student. To bridge the gap, we propose PROD, a PROgressive Distillation method, for dense retrieval. PROD consists of a teacher progressive distillation and a data progressive distillation to gradually improve the student. To alleviate catastrophic forgetting, we introduce a regularization term in each distillation process. We conduct extensive experiments on seven datasets including five widely-used publicly available benchmarks: MS MARCO Passage, TREC Passage 19, TREC Document 19, MS MARCO Document, and Natural Questions, as well as two industry datasets: Bing-Rel and Bing-Ads. PROD achieves the state-of-the-art in the distillation methods for dense retrieval. Our 6-layer student model even surpasses most of the existing 12-layer models on all five public benchmarks. The code and models are released in https://github.com/microsoft/SimXNS.

Extracting expressive visual features is crucial for accurate Click-Through-Rate (CTR) prediction in visual search advertising systems. Current commercial systems use off-the-shelf visual encoders to facilitate fast online service. However, the extracted visual features are coarse-grained and/or biased. In this paper, we present a visual encoding framework for CTR prediction to overcome these problems. The framework is based on contrastive learning which pulls positive pairs closer and pushes negative pairs apart in the visual feature space. To obtain fine-grained visual features, we present contrastive learning supervised by click-through data to fine-tune the visual encoder. To reduce sample selection bias, firstly we train the visual encoder offline by leveraging both unbiased self-supervision and click supervision signals. Secondly, we incorporate a debiasing network in the online CTR predictor to adjust the visual features by contrasting high impression items with selected, low impression items. We deploy the framework in a mobile E-commerce app. Offline experiments on billion-scale datasets and online experiments demonstrate that the proposed framework can make accurate and unbiased predictions.

Review-based recommender systems (RRS) have received an increasing interest since reviews greatly enhance recommendation quality and interpretability. However, existing RRS suffer from high computational complexity, biased recommendation and poor generalization. The three problems make them inadequate to handle real recommendation scenarios. Previous studies address each issue separately, while none of them consider solving three problems together under a unified framework. This paper presents LUME (a Lightweight Unbiased Multi-teacher Ensemble) for RRS. LUME is a novel framework that addresses the three problems simultaneously. LUME uses multi-teacher ensemble and debiased knowledge distillation to aggregate knowledge from multiple pretrained RRS, and generates a small, unbiased student recommender which generalizes better. Extensive experiments on various real-world benchmarks demonstrate that LUME successfully tackles the three problems and has superior performance than state-of-the-art RRS and knowledge distillation based RS.

Configuration tuning is vital to optimize the performance of big data analysis platforms like Spark. Existing methods (e.g. auto-tuning relational databases) are not effective for tuning Spark, because the unique characteristics of Spark pose new challenges to configuration tuning. (C1) The Spark applications own various code structures and semantics, and the code features significantly affect Spark performance and configuration selection; (C2) Spark applications are extremely time-consuming on big data. It is infeasible for approaches such as Bayesian Optimization and Reinforcement Learning to collect sufficient training instances or repeatedly execute the applications; (C3) Spark supports various analytical applications and the tuning system needs to adapt to different applications. To address these challenges, we propose a LIghtweighT knob rEcommender system (LITE) for auto-tuning Spark configurations on various analytical applications and large-scale datasets. We first propose a code learning framework that can utilize code features to learn complex correlations between application performance and knob values (addressing C1). We then propose a lightweight auto-tuning method that migrates the knowledge learned from small-scale datasets to large-scale datasets (addressing C2). Next, to generalize to different Spark applications, we propose an adaptive model update approach to fine-tune the model via adversarial learning with newly collected feedback (addressing C3). Extensive experiments showed that LITE achieves much better performance compared with state-of-the-art auto-tuning methods.

The importance of modeling contextual information within a search session has been widely acknowledged. However, learning representations of multi-query multi-modal (MM) search, in which Mobile Taobao users repeatedly submit textual and visual queries, remains unexplored in literature. Previous work which learns task-specific representations of textual query sessions fails to capture diverse query types and correlations in MM search sessions. This paper presents to represent MM search sessions by heterogeneous graph neural network (HGN). A multi-view contrastive learning framework is proposed to pretrain the HGN, with two views to model different intra-query, inter-query, and inter-modality information diffusion in MM search. Extensive experiments demonstrate that, the pretrained session representation can benefit state-of-the-art baselines on various downstream tasks, such as personalized click prediction, query suggestion, and intent classification.

Instant messaging client (IMC) is now an essential tool for mobile users. In the representative IMC We Chat, cybercriminals deceive frauds, causing financial loss to normal users. Through statistical analysis, we find that certain fraud interactions commonly occur among WeChat users who are not k-hop neighbors. Therefore, efficiently answering whether the distance between two vertices is not longer than k at a certain time point (i.e., k-hop reachability queries) over the dynamic social graph of WeChat becomes a crucial task for fraud feature extraction in the detection system: it can help human experts quickly identify suspicious user interactions and the query results can be further used as the input feature to the downstream machine learning based detection methods. In this paper, we illustrate Bidirectional k-hop Reachability Query Processing over a Dynamic Graph (BREAD) that is used in WeChat for extracting the k-hop reachability feature for fraud detection. BREAD adopts the idea of estimating Personalized PageRank value. It first conducts the backward search from the destination vertex to construct an intermediate vertex set. Then, it performs a certain amount of random walks from the start vertex to see whether they can hit the intermediate vertex set, and the results are returned to answer k-hop reachability queries. We further propose$\text{BREAD}++$that leverages the massive parallel processing power of GPU to achieve a considerable performance gain. Experiments on several large-scale dynamic graph benchmarks and the social graph of WeChat have demonstrated that$\text{BREAD}/\text{BREAD}++$is superior than existing index-free competitors: our methods provide not only fast but also accurate responses and they are of practical value to k-hop reachability feature extraction in the fraud detection system of WeChat. Our implementation is available at https://github.com/XMUDM/BREAD.

Most existing recommendation models learn vectorized representations for items, i.e., item embeddings to make predictions. Item embeddings inherit popularity bias from the data, which leads to biased recommendations. We use this observation to design two simple and effective strategies, which can be flexibly plugged into different backbone recommendation models, to learn popularity neutral item representations. One strategy isolates popularity bias in one embedding direction and neutralizes the popularity direction post-training. The other strategy encourages all embedding directions to be disentangled and popularity neutral. We demonstrate that the proposed strategies outperform state-of-the-art debiasing methods on various real-world datasets, and improve recommendation quality of shallow and deep backbone models.

It is crucial to model missing ratings in recommender systems since user preferences learnt from only observed ratings are biased. One possible explanation for missing ratings is motivated by the spiral of silence theory. When the majority opinion is formed, a spiral process is triggered where users are more and more likely to show their ratings if they perceive that they are supported by the opinion climate. In this paper we first verify the existence of the spiral process in recommender systems by using a variety of different real-life datasets. We then study the characteristics of two key factors in the spiral process: opinion climate and the hardcore users who will give ratings even when they are minority opinion holders. Based on our empirical findings, we develop four variants to model missing ratings. They mimic different components of the spiral of silence based on the spiral process with global opinion climate, local opinion climate, hardcore users, relationships between hardcore users and items, respectively. We experimentally show that, the presented variants all outperform state-of-the-art recommendation models with missing rating components.

Social Recommender Systems (SRS) have attracted considerable attention since its accompanying service, social networks, helps increase user satisfaction and provides auxiliary information to improve recommendations. However, most existing SRS focus on social influence and ignore another essential social phenomenon, i.e., social homophily. Social homophily, which is the premise of social influence, indicates that people tend to build social relations with similar people and form influence propagation paths. In this article, we propose a generic framework Social PathExplorer (SPEX) to enhance neural SRS. SPEX treats the neural recommendation model as a black box and improves the quality of recommendations by modeling the social recommendation task, the formation of social homophily, and their mutual effect in the manner of multi-task learning. We design a Graph Neural Network based component for influence propagation path prediction to help SPEX capture the rich information conveyed by the formation of social homophily. We further propose an uncertainty based task balancing method to set appropriate task weights for the recommendation task and the path prediction task during the joint optimization. Extensive experiments have validated that SPEX can be easily plugged into various state-of-the-art neural recommendation models and help improve their performance. The source code of our work is available at: https://github.com/XMUDM/SPEX.

Biases and de-biasing in recommender systems (RS) have become a research hotspot recently. This paper reveals an unexplored type of bias, i.e., sentiment bias. Through an empirical study, we find that many RS models provide more accurate recommendations on user/item groups having more positive feedback (i.e., positive users/items) than on user/item groups having more negative feedback (i.e., negative users/items). We show that sentiment bias is different from existing biases such as popularity bias: positive users/items do not have more user feedback (i.e., either more ratings or longer reviews). The existence of sentiment bias leads to low-quality recommendations to critical users and unfair recommendations for niche items. We discuss the factors that cause sentiment bias. Then, to fix the sources of sentiment bias, we propose a general de-biasing framework with three strategies manifesting in different regularizers that can be easily plugged into RS models without changing model architectures. Experiments on various RS models and benchmark datasets have verified the effectiveness of our de-biasing framework. To our best knowledge, sentiment bias and its de-biasing have not been studied before. We hope that this work can help strengthen the study of biases and de-biasing in RS.

Online gaming is growing faster than ever before, with increasing challenges of providing better user experience. Recommender systems (RS) for online games face unique challenges since they must fulfill players’ distinct desires, at different user levels, based on their action sequences of various action types. Although many sequential RS already exist, they are mainly single-sequence, single-task, and single-user-level. In this paper, we introduce a new sequential recommendation model for multiple sequences, multiple tasks, and multiple user levels (abbreviated as M3Rec) in Tencent Games platform, which can fully utilize complex data in online games. We leverage Graph Neural Network and multi-task learning to design M3Rec in order to model the complex information in the heterogeneous sequential recommendation scenario of Tencent Games. We verify the effectiveness of M3Rec on three online games of Tencent Games platform, in both offline and online evaluations. The results show that M3Rec successfully addresses the challenges of recommendation in online games, and it generates superior recommendations compared with state-of-the-art sequential recommendation approaches.

Online text streams such as Twitter are the major information source for users when they are looking for ongoing events. Realtime event summarization aims to generate and update coherent and concise summaries to describe the state of a given event. Due to the enormous volume of continuously coming texts, realtime event summarization has become the de facto tool to facilitate information acquisition. However, there exists a challenging yet unexplored issue in current text summarization techniques: how to preserve the integrity, i.e., the accuracy and consistency of summaries during the update process. The issue is critical since online text stream is dynamic and conflicting information could spread during the event period. For example, conflicting numbers of death and injuries might be reported after an earthquake. Such misleading information should not appear in the earthquake summary at any timestamp. In this article, we present a novel realtime event summarization framework called IAEA (i.e., Integrity-Aware Extractive-Abstractive realtime event summarization). Our key idea is to integrate an inconsistency detection module into a unified extractive–abstractive framework. In each update, important new tweets are first extracted in an extractive module, and the extraction is refined by explicitly detecting inconsistency between new tweets and previous summaries. The extractive module is able to capture the sentence-level attention which is later used by an abstractive module to obtain the word-level attention. Finally, the word-level attention is leveraged to rephrase words. We conduct comprehensive experiments on real-world datasets. To reduce efforts required for building sufficient training data, we also provide automatic labeling steps of which the effectiveness has been empirically verified. Through experiments, we demonstrate that IAEA can generate better summaries with consistent information than state-of-the-art approaches.

Recommendation Systems (RS) have become an essential part of many online services. Due to its pivotal role in guiding customers towards purchasing, there is a natural motivation for unscrupulous parties to spoof RS for profits. In this paper, we study the shilling attack: a subsistent and profitable attack where an adversarial party injects a number of user profiles to promote or demote a target item. Conventional shilling attack models are based on simple heuristics that can be easily detected, or directly adopt adversarial attack methods without a special design for RS. Moreover, the study on the attack impact on deep learning based RS is missing in the literature, making the effects of shilling attack against real RS doubtful. We present a novel Augmented Shilling Attack framework (AUSH) and implement it with the idea of Generative Adversarial Network. AUSH is capable of tailoring attacks against RS according to budget and complex attack goals, such as targeting a specific user group. We experimentally show that the attack impact of AUSH is noticeable on a wide range of RS including both classic and modern deep learning based RS, while it is virtually undetectable by the state-of-the-art attack detection model.

Modeling sentiment evolution for social incidents in Microblogs is of vital importance for both enterprises and government officials. Existing works on sentiment tracking are not satisfying, due to the lack of entity-level sentiment extraction and accurate sentiment shift detection. Identifying entity-level sentiment is challenging as Microbloggers often use multiple opinion expressions in a sentence which targets different entities. Moreover, the evolution of the background sentiment, which is essential to shift detection, is ignored in the previous study. To address these issues, we leverage the proximity information to obtain more precise entity-level sentiment extraction. Based on it, we propose to simultaneously model the evolution of background opinion and the sentiment shift using a state space model on the time series of sentiment polarities. Experiments on real data sets demonstrate that our proposed approaches outperform state-of-the-art methods on the task of modeling sentiment evolution for social incidents.

It has been established that, ratings are missing not at random in recommender systems. However, little research has been done to reveal how the ratings are missing. In this paper we present one possible explanation of the missing not at random phenomenon. We verify that, using a variety of different real-life datasets, there is a spiral process for a silent minority in recommender systems where (1) people whose opinions fall into the minority are less likely to give ratings than majority opinion holders; (2) as the majority opinion becomes more dominant, the rating possibility of a majority opinion holder is intensifying but the rating possibility of a minority opinion holder is shrinking; (3) only hardcore users remain to rate for minority opinions when the spiral achieves its steady state. Our empirical findings are beneficial for future recommendation models. To demonstrate the impact of our empirical findings, we present a probabilistic model that mimics the generation process of spiral of silence. We experimentally show that, the presented model offers more accurate recommendations, compared with state-of-the-art recommendation models.

Secure top-k inner product retrieval allows the users to outsource encrypted data vectors to a cloud server and at some later time find the k vectors producing largest inner products giving an encrypted query vector. Existing solutions suffer poor performance raised by the client's filtering out top-k results. To enable the server-side filtering, we introduce an asymmetric inner product encryption AIPE that allows the server to compute inner products from encrypted data and query vectors. To solve AIPE's vulnerability under known plaintext attack, we present a packing approach IP Packing that allows the server to obtain the entire set of inner products between the query and all data vectors but prevents the server from associating any data vector with its inner product. Based on IP Packing, we present our solution SKIP to secure top-k inner product retrieval that further speeds up retrieval process using sequential scan. Experiments on real recommendation datasets demonstrate that our protocols outperform alternatives by several orders of magnitude.

We describe our demonstration of CLEar ( Cl airaudient Ear ), a real-time online platform for detecting, monitoring, summarizing, contextualizing and visualizing bursty and viral events, those triggering a sudden surge of public interest and going viral on micro-blogging platforms. This task is challenging for existing methods as they either use complicated topic models to analyze topics in a off-line manner or define temporal structure of fixed granularity on the data stream for online topic learning, leaving them hardly scalable for real-time stream like that of Twitter. In this demonstration of CLEar, we present a three-stage system: First, we show a real-time bursty event detection module based on a data-sketch topic model which makes use of acceleration of certain stream quantities as the indicators of topic burstiness to trigger efficient topic inference. Second, we demonstrate popularity prediction for the detected bursty topics and event summarization based on clustering related topics detected in successive time periods. Third, we illustrate CLEar's module for contextualizing and visualizing the event evolution both along time-line and across other news media to offer an easier understanding of the events.

Microblogging service has emerged to be a dominant web medium for billions of individuals sharing and spreading instant news and information, therefore monitoring the event evolution on microblog sphere is crucial for providing both better user experience and deeper understanding on real-time events. In this paper we explore the problem of generating storylines from microblogs for user input queries. This problem is challenging due to the sparse, dynamic and social nature of microblogs. Given a query of an ongoing event, we propose to sketch the real-time storyline of the event by a two-level solution. We first propose a language model with dynamic pseudo relevance feedback to obtain relevant tweets, and then generate storylines via graph optimization. Comprehensive experiments on Twitter data sets demonstrate the effectiveness of the proposed methods in each level and the overall framework.

A variety of news recommender systems based on different strategies have been proposed to provide news personalization services for online news readers. However, little research work has been reported on utilizing the implicit "social" factors (i.e., the potential influential experts in news reading community) among news readers to facilitate news personalization. In this paper, we investigate the feasibility of integrating content-based methods, collaborative filtering and information diffusion models by employing probabilistic matrix factorization techniques. We propose PRemiSE, a novel Personalized news Recommendation framework via implicit Social Experts, in which the opinions of potential influencers on virtual social networks extracted from implicit feedbacks are treated as auxiliary resources for recommendation. Empirical results demonstrate the efficacy and effectiveness of our method, particularly, on handling the so-called cold-start problem.

The huge amount of knowledge in web communities has motivated the research interests in threaded discussions. The dynamic nature of threaded discussions poses lots of challenging problems for computer scientists. Although techniques such as semantic models and structural models have been shown to be useful in a number of areas, they are inefficient in understanding threaded discussions due to three reasons: (I) as most of users read existing messages before posting, posts in a discussion thread are temporally dependent on the previous ones; It causes the semantics and structure to be coupled with each other in threaded discussions; (II) in online discussion threads, there are a lot of junk posts which are useless and may disturb content analysis; and (III) it is very hard to judge the quality of a post. In this paper, we propose a sparse coding-based model named SMSS to Simultaneously Model Semantics and Structure of threaded discussions. The model projects each post into a topic space, and approximates each post by a linear combination of previous posts in the same discussion thread. Meanwhile, the model also imposes two sparse constraints to force a sparse post reconstruction in the topic space and a sparse post approximation from previous posts. The sparse properties effectively take into account the characteristics of threaded discussions. Towards the above three problems, we demonstrate the competency of our model in three applications: reconstructing reply structure of threaded discussions, identifying junk posts, and finding experts in a given board/sub-board in web communities. Experimental results show encouraging performance of the proposed SMSS model in all these applications.

The abundant knowledge in web communities has motivated the research interests in discussion threads. The dynamic nature of discussion threads poses interesting and challenging problems for computer scientists. Although techniques such as semantic models or structural models have been shown to be useful in a number of areas, they are inefficient in understanding discussion threads due to the temporal dependence among posts in a discussion thread. Such dependence causes that semantics and structure coupled with each other in discussion threads. In this paper, we propose a sparse coding-based model named SMSS to Simultaneously Model Semantic and Structure of discussion threads.

It is an important issue to get support from experts in our daily life. Expert finding is challenging. In previous commercial and academic systems, the users may not get what they expect. In this contribution, we address the problem of finding experts in communities. A novel model REC is presented to solve the expert finding problem in dynamic environment. The model ranks experts by textural and social information. Starting with the most familiar communities, the expert seeker may find appropriate experts, by considering both their local rankings in each community and the difficulty to get their help. Experiments are done on real data sets, including DBLP data set and W3C corpora. Compared with other existing methods, REC achieves promising results. It demonstrates the model's competencies in various search applications.