Prompt tuning for pre-trained language models (PLMs) has been an effective approach for few-shot text classification. To make a prediction, a typical prompt tuning method employs a template wrapping the input text into a cloze question, and a verbalizer mapping the output embedding to labels. However, current methods typically depend on handcrafted templates and verbalizers, which require much domain-specific prior knowledge by human efforts. In this work, we investigate how to build a good human-free prompt tuning using soft prompt templates and soft verbalizers, which can be learned directly from data. To address the challenge of data scarcity, we integrate a set of trainable bases for sentence representation to transfer the contextual information into a low-dimensional space. By jointly pre-training the soft prompts and the bases using contrastive learning, the projection space can catch critical semantics at the sentence level, which could be transferred to various downstream tasks. To better bridge the gap between downstream tasks and the pre-training procedure, we formulate the few-shot classification tasks as another contrastive learning problem. We name this Jointly Pretrained Template and Verbalizer (JPTV). Extensive experiments show that this human-free prompt tuning can achieve comparable or even better performance than manual prompt tuning.

Auto-scaling is crucial for achieving elasticity in cloud databases as well as other cloud systems. Predictive auto-scaling, which leverages forecasting techniques to adjust resources based on predicted workload, has been widely adopted. However, the inherent inaccuracy of forecasting presents a significant challenge, potentially causing resource under-provisioning. To address this challenge, we propose robust predictive auto-scaling that considers the uncertainty in forecasts. Unlike previous predictive approaches that rely on single-valued forecasts, we leverage probabilistic forecasting techniques to generate quan-tile forecasts, providing a more comprehensive understanding of the potential future workloads. By formulating the auto-scaling problem as a robust optimization problem, we enable the implementation of auto-scaling strategies with customizable levels of robustness, which can be determined by considering various quantile levels of forecasts. Moreover, we enhance the adaptability of our strategy by incorporating different quantile levels through-out the entire decision horizon, allowing for dynamic adjustments in the conservatism of our auto-scaling decisions. This enables us to strike a balance between resource efficiency and system robustness. Through extensive experiments, we demonstrate the effectiveness of our approach in achieving robust auto-scaling in cloud databases, while maintaining reasonable resource efficiency.

Recent years have witnessed the success of introducing deep learning models to time series forecasting. From a data generation perspective, we illustrate that existing models are susceptible to distribution shifts driven by temporal contexts, whether observed or unobserved. Such context-driven distribution shift (CDS) introduces biases in predictions within specific contexts and poses challenges for conventional training paradigms. In this paper, we introduce a universal calibration methodology for the detection and adaptation of CDS with a trained model. To this end, we propose a novel CDS detector, termed the "residual-based CDS detector" or "Reconditionor", which quantifies the model's vulnerability to CDS by evaluating the mutual information between prediction residuals and their corresponding contexts. A high Reconditionor score indicates a severe susceptibility, thereby necessitating model adaptation. In this circumstance, we put forth a straightforward yet potent adapter framework for model calibration, termed the "sample-level contextualized adapter" or "SOLID". This framework involves the curation of a contextually similar dataset to the provided test sample and the subsequent fine-tuning of the model's prediction layer with a limited number of steps. Our theoretical analysis demonstrates that this adaptation strategy can achieve an optimal bias-variance trade-off. Notably, our proposed Reconditionor and SOLID are model-agnostic and readily adaptable to a wide range of models. Extensive experiments show that SOLID consistently enhances the performance of current forecasting models on real-world datasets, especially on cases with substantial CDS detected by the proposed Reconditionor, thus validating the effectiveness of the calibration approach.

Graph contrastive learning often faces challenges when data augmentations compromise the graph's critical attributes, introducing the risk of generating noise-positive pairs. Although recent methods have attempted to address these issues, they either fall short of ensuring effective data augmentation or suffer from excessive computational demands. The advent of full-attention graph Transformers, with their enhanced capacity for graph representation learning, has sparked significant interest. Despite their potential, employing full-attention graph Transformers for contrastive learning can introduce issues such as noisy redundancies. In this work, we propose the Graph Attention Contrastive Learning (GACL) model, which innovatively combines a full-attention transformer with a message-passing graph neural network as its encoder. To mitigate the noise associated with full-attention mechanisms, we apply a denoising modification. Our GACL model effectively tackles the challenges associated with full-attention mechanisms and introduces a novel approach for data augmentation. Moreover, we propose the concept of effective mutual information to theoretically underpin our methodology. Utilizing this framework, we explore the impact of the denoising matrix within GACL's contrastive learning process and delve into comprehensive discussions on its implications. Empirical assessments underscore GACL's exceptional performance, establishing it as a state-of-the-art solution in graph contrastive learning.

Graph label propagation (LP) is a core component in many downstream applications such as fraud detection, recommendation and image segmentation. In this paper, we proposeGLP, a GPU-based framework to enable efficientLPprocessing on large-scale graphs. By investigating the data processing pipeline in a large e-commerce platform, we have identified two key challenges on integrating GPU-acceleratedLPprocessing to the pipeline: (1) programmability for evolving application logics; (2) demand for real-time performance. Motivated by these challenges, we offer a set of expressive APIs that data engineers can customize and deploy efficientLPalgorithms on GPUs with ease. To achieve better performance, we propose novel GPU-centric optimizations by leveraging the community as well as power-law properties of large graphs. Further, we significantly reduce the expensive data transfer cost between CPUs and GPUs by enablingLPprocessing on compressed graphs. Extensive experiments have confirmed the effectiveness of our proposed approaches over the state-of-the-art GPU methods. Furthermore, our proposed solution supports a real billion-scale graph workload for fraud detection and achieves 13.2× speedup to the current in-house solution running on a high-end multicore machine with compressed graphs.

Recent studies have highlighted the limitations of message-passing based graph neural networks (GNNs), e.g., limited model expressiveness, over-smoothing, over-squashing, etc. To alleviate these issues, Graph Transformers (GTs) have been proposed which work in the paradigm that allows message passing to a larger coverage even across the whole graph. Hinging on the global range attention mechanism, GTs have shown a superpower for representation learning on homogeneous graphs. However, the investigation of GTs on heterogeneous information networks (HINs) is still under-exploited. In particular, on account of the existence of heterogeneity, HINs show distinct data characteristics and thus require different treatment. To bridge this gap, in this paper we investigate the representation learning on HINs with Graph Transformer, and propose a novel model named HINormer, which capitalizes on a larger-range aggregation mechanism for node representation learning. In particular, assisted by two major modules, i.e., a local structure encoder and a heterogeneous relation encoder, HINormer can capture both the structural and heterogeneous information of nodes on HINs for comprehensive node representations. We conduct extensive experiments on four HIN benchmark datasets, which demonstrate that our proposed model can outperform the state-of-the-art.

Cloud-native databases have become the de-facto choice for mission-critical applications on the cloud due to the need for high availability, resource elasticity, and cost efficiency. Meanwhile, driven by the increasing connectivity between data generation and analysis, users prefer a single database to efficiently process both OLTP and OLAP workloads, which enhances data freshness and reduces the complexity of data synchronization and the overall business cost. In this paper, we summarize five crucial design goals for a cloud-native HTAP database based on our experience and customers' feedback, i.e., transparency, competitive OLAP performance, minimal perturbation on OLTP workloads, high data freshness, and excellent resource elasticity. As our solution to realize these goals, we present PolarDB-IMCI, a cloud-native HTAP database system designed and deployed at Alibaba Cloud. Our evaluation results show that PolarDB-IMCI is able to handle HTAP efficiently on both experimental and production workloads; notably, it speeds up analytical queries up to ×149 on TPC-H (100GB). PolarDB-IMCI introduces low visibility delay and little performance perturbation on OLTP workloads (<5%), and resource elasticity can be achieved by scaling out in tens of seconds.

Natural language to SQL (NL2SQL) techniques provide a convenient interface to access databases, especially for non-expert users, to conduct various data analytics. Existing methods often employ either a rule-base approach or a deep learning based solution. The former is hard to generalize across different domains. Though the latter generalizes well, it often results in queries with syntactic or semantic errors, thus may be even not executable. In this work, we bridge the gap between the two and design a new framework to significantly improve both accuracy and runtime. In particular, we develop a novel CatSQL sketch, which constructs a template with slots that initially serve as placeholders, and tightly integrates with a deep learning model to fill in these slots with meaningful contents based on the database schema. Compared with the widely used sequence-to-sequence-based approaches, our sketch-based method does not need to generate keywords which are boilerplates in the template, and can achieve better accuracy and run much faster. Compared with the existing sketch-based approaches, our CatSQL sketch is more general and versatile, and can leverage the values already filled in on certain slots to derive the rest ones for improved performance. In addition, we propose the Semantics Correction technique, which is the first that leverages database domain knowledge in a deep learning based NL2SQL solution. Semantics Correction is a post-processing routine, which checks the initially generated SQL queries by applying rules to identify and correct semantic errors. This technique significantly improves the NL2SQL accuracy. We conduct extensive evaluations on both single-domain and cross-domain benchmarks and demonstrate that our approach significantly outperforms the previous ones in terms of both accuracy and throughput. In particular, on the state-of-the-art NL2SQL benchmark Spider, our CatSQL prototype outperforms the best of the previous solutions by 4 points on accuracy, while still achieving a throughput up to 63 times higher.

Unbiased learning to rank (ULTR) aims to train an unbiased ranking model from biased user click logs. Most of the current ULTR methods are based on the examination hypothesis (EH), which assumes that the click probability can be factorized into two scalar functions, one related to ranking features and the other related to bias factors. Unfortunately, the interactions among features, bias factors and clicks are complicated in practice, and usually cannot be factorized in this independent way. Fitting click data with EH could lead to model misspecification and bring the approximation error.

The prevalence of graph structures attracts a surge of investigation on graph data, enabling several downstream tasks such as multi-graph classification. However, in the multi-graph setting, graphs usually follow a long-tailed distribution in terms of their sizes, i.e., the number of nodes. In particular, a large fraction of tail graphs usually have small sizes. Though recent graph neural networks (GNNs) can learn powerful graph-level representations, they treat the graphs uniformly and marginalize the tail graphs which suffer from the lack of distinguishable structures, resulting in inferior performance on tail graphs. To alleviate this concern, in this paper we propose a novel graph neural network named SOLT-GNN, to close the representational gap between the head and tail graphs from the perspective of knowledge transfer. In particular, SOLT-GNN capitalizes on the co-occurrence substructures exploitation to extract the transferable patterns from head graphs. Furthermore, a novel relevance prediction function is proposed to memorize the pattern relevance derived from head graphs, in order to predict the complements for tail graphs to attain more comprehensive structures for enrichment. We conduct extensive experiments on five benchmark datasets, and demonstrate that our proposed model can outperform the state-of-the-art baselines.

A significant fraction of data in cloud storage is rarely accessed, referred to as cold data. Accurately identifying and efficiently managing cold data on cost-effective storages is one of the major challenges for cloud providers, which balances between reducing the cost and improving the system performance. To this end, we propose SA-LSM to use (S)urvival (A)nalysis for Log-Structure Merge Tree (LSM-tree) key-value (KV) stores. Conventionally, the data layout of LSM-tree is determined jointly by the write and the compaction operations. However, this process by default does not fully utilize the access information of data records, leading to a suboptimal data layout that negatively impacts the system performance. SA-LSM utilizes the survival analysis, a statistical learning algorithm commonly used in biostatistics, to optimize the data layout. When put into perspective of LSM-tree with proper adoptions, SA-LSM can accurately predict cold data using the historical semantic information and access traces. As a concrete realization, we implement our proposal in X-Engine, a commercial-strength open-source LSM-tree storage engine. To make the deployment more flexible, we also design a non-intrusive architecture that offloads CPU-intensive work, e.g., model training and inference, to an external service. Extensive experiments on real-world workloads show that it can decrease the tail latency by up to 78.9% compared to the state-of-the-art techniques. The generality of this approach and the significant performance improvement show great potentials in a variety of related applications.

The hash table is a fundamental structure that has been implemented on graphics processing units (GPUs) to accelerate a wide range of analytics workloads. Most existing works have focused on static scenarios and occupy large GPU memory to maximize the insertion efficiency. In many cases, data stored in hash tables get updated dynamically, and existing approaches use unnecessarily large memory resources. One naïve solution is to rebuild a hash table (known as rehashing) whenever it is either filled or mostly empty. However, this approach renders significant overheads for rehashing. In this paper, we propose a novel dynamic cuckoo hash table technique on GPUs, known as DyCuckoo. We devise a resizing strategy for dynamic scenarios without rehashing the entire table that ensures a guaranteed filled factor. The strategy trades search performance with resizing efficiency, and this tradeoff can be configured by users. To further improve efficiency, we propose a 2-in-d cuckoo hashing scheme that ensures a maximum of two lookups for find and delete operations, while retaining similar performance for insertions as a general cuckoo hash. Extensive experiments have validated the proposed design's effectiveness over several state-of-the-art hash table implementations on GPUs. DyCuckoo achieves superior efficiency while enables fine-grained memory control, which is not available in existing GPU hash table approaches.

Listwise collaborative filtering, which directly predicts a ranking list of items for the given user, achieves superior accuracy performance since it is aligned with the ultimate goals of recommender systems.However, in corpus with the enormous number of items, the calculation cost for the learnt model to predict all user-item preferences is tremendous, which makes full corpus retrieval extremely difficult.In this paper, we propose a binarized collaborative filtering method, called Discrete Listwise Collaborative Filtering (DLCF), to represent users and items as binary codes for fast recommendation.As such, the proposed method could accelerate the retrieval procedure, since the user-item similarity could be efficiently computed via Hamming distance.We further adopt the discrete coordinate descent method to jointly optimize our proposed model.Extensive experiments performed on three real-world datasets demonstrate that 1) DLCF significantly outperforms the state-of-the-art binarized recommendation methods, and 2) DLCF shows very competitive ranking accuracy compared to its real-valued version while significantly improving the retrieval efficiency.

Counterfactual Learning to Rank (CLTR) becomes an attractive research topic due to its capability of training ranker with click logs. However, CLTR inherently suffers from a large amount of bias caused by confounders, variables that affect both the observation (examination) behavior and click behavior. Recent efforts to correct bias mostly focus on position bias, which assumes that each observation in a ranking list is isolated and only depends on the position. Though effective, users often engage with documents in an interactive manner. Ignoring the interactions between observations/clicks would incur a large interactional observation bias no matter how much data is collected. In this work, we leverage the embedding method to develop an Interactional Observation-Based Model (IOBM) to estimate the observation probability. We argue that while there exist complex observed and unobserved confounders for observation/click interactions, it is sufficient to use the embedding as a proxy confounder to uncover the relevant information for the prediction of the observation propensity. Moreover, the embedding could offer an alternative to the fully specified generative model for observation and decouples the complex interaction structure of observations/clicks. In our IOBM, we first learn the individual observation embedding to capture position and click information. Then, we learn the interactional observation embedding to uncover their local interaction structure. To filter out irrelevant information and reduce contextual bias, we utilize query context information and propose the intra-observation attention and the inter-observation attention, respectively. We conduct extensive experiments on two LTR benchmark datasets, demonstrating that the proposed IOBM consistently achieves better performance over the baseline models in various click situations and verifying its effectiveness of eliminating interactional observation bias.

Fraud detection is a pressing challenge for most financial and commercial platforms. In this paper, we study the processing pipeline of fraud detection in a large e-commerce platform of TaoBao. Graph label propagation (LP) is a core component in this pipeline to detect suspicious clusters from the user-interaction graph.Furthermore, the run-time of the LP component occupies 75% overhead of TaoBao's automated detection pipeline. To enable real-time fraud detection, we propose a GPU-based framework, called GLP, to support large-scale LP workloads in enterprises.We have identified two key challenges when integrating GPU acceleration into TaoBao's data processing pipeline: (1)programmability for evolving fraud detection logics; (2)demand for real-time performance. Motivated by these challenges, we offer a set of expressive APIs that data engineers can customize and deploy efficient LP algorithms on GPUs with ease. We propose novel GPU-centric optimizations by leveraging the community as well as power-law properties of large graphs. Extensive experiments have confirmed the effectiveness of our proposed optimizations. With a single GPU, GLP supports a real billion-scale graph workload from the fraud detection pipeline of TaoBao and achieves 8.2x speedup to the current in-house distributed solution running on high-end multicore machines.

Sequential user behavior modeling plays a crucial role in online user-oriented services, such as product purchasing, news feed consumption, and online advertising. The performance of sequential modeling heavily depends on the scale and quality of historical behaviors. However, the number of user behaviors inherently follows a long-tailed distribution, which has been seldom explored. In this work, we argue that focusing on tail users could bring more benefits and address the long tails issue by learning transferrable parameters from both optimization and feature perspectives. Specifically, we propose a gradient alignment optimizer and adopt an adversarial training scheme to facilitate knowledge transfer from the head to the tail. Such methods can also deal with the cold-start problem of new users. Moreover, it could be directly adaptive to various well-established sequential models. Extensive experiments on four real-world datasets verify the superiority of our framework compared with the state-of-the-art baselines.

Factorization Machine (FM) is a general supervised learning framework for many AI applications due to its powerful capability of feature engineering. Despite being extensively studied, existing FM methods have several limitations in common. First of all, most existing FM methods often adopt the squared loss in the modeling process, which can be very sensitive when the data for learning contains noises and outliers. Second, some recent FM variants often explore the low-rank structure of the feature interactions matrix by relaxing the low-rank minimization problem as a trace norm minimization, which cannot always achieve a tight approximation to the original one. To address the aforementioned issues, this paper proposes a new scheme of Robust Factorization Machine (RFM) by exploring a doubly capped norms minimization approach, which employs both a capped squared trace norm in achieving a tighter approximation of the rank minimization and a capped ℓ1-norm loss to enhance the robustness of the empirical loss minimization from noisy data. We develop an efficient algorithm with a rigorous convergence proof of RFM. Experiments on public real-world datasets show that our method outperforms the state-of-the-art FM methods significantly.

Efficiency is crucial to the online recommender systems, especially for the ones which needs to deal with tens of millions of users and items. Because representing users and items as binary vectors for Collaborative Filtering (CF) can achieve fast user-item affinity computation in the Hamming space, in recent years, we have witnessed an emerging research effort in exploiting binary hashing techniques for CF methods. However, CF with binary codes naturally suffers from low accuracy due to limited representation capability in each bit, which impedes it from modeling complex structure of the data.

This paper aims to investigate efficient and scalable machine learning algorithms for resolving Non-negative Matrix Factorization (NMF), which is important for many real-world applications, particularly for collaborative filtering and recommender systems. Unlike traditional batch learning methods, a recently proposed online learning technique named "NN-PA" tackles NMF by applying the popular Passive-Aggressive (PA) online learning, and found promising results. Despite its simplicity and high efficiency, NN-PA falls short in at least two critical limitations: (i) it only exploits the first-order information and thus may converge slowly especially at the beginning of online learning tasks; (ii) it is sensitive to some key parameters which are often difficult to be tuned manually, particularly in a practical online learning system. In this work, we present a novel family of online Adaptive Passive-Aggressive (APA) learning algorithms for NMF, named "NN-APA", which overcomes two critical limitations of NN-PA by (i) exploiting second-order information to enhance PA in making more informative updates at each iteration; and (ii) achieving the parameter auto-selection by exploring the idea of online learning with expert advice in deciding the optimal combination of the key parameters in NMF. We theoretically analyze the regret bounds of the proposed method and show its advantage over the state-of-the-art NN-PA method, and further validate the efficacy and scalability of the proposed technique through an extensive set of experiments on a variety of large-scale real recommender systems datasets.

As location-based social network (LBSN) services become increasingly popular, trip recommendation that recommends a sequence of points of interest (POIs) to visit for a user emerges as one of many important applications of LBSNs. Personalized trip recommendation tailors to users' specific tastes by learning from past check-in behaviors of users and their peers. Finding the optimal trip that maximizes user's experiences for a given time budget constraint is an NP hard problem and previous solutions do not consider two practical and important constraints. One constraint is POI availability where a POI may be only available during a certain time window. Another constraint is uncertain traveling time where the traveling time between two POIs is uncertain. This work presents efficient solutions to personalized trip recommendation by incorporating these constraints to prune the search space. We evaluated the efficiency and effectiveness of our solutions on real life LBSN data sets.

User preferences change over time and capturing such changes is essential for developing accurate recommender systems. Despite its importance, only a few works in collaborative filtering have addressed this issue. In this paper, we consider evolving preferences and we model user dynamics by introducing and learning a transition matrix for each user's latent vectors between consecutive time windows. Intuitively, the transition matrix for a user summarizes the time-invariant pattern of the evolution for the user. We first extend the conventional probabilistic matrix factorization and then improve upon this solution through its fully Bayesian model. These solutions take advantage of the model complexity and scalability of conventional Bayesian matrix factorization, yet adapt dynamically to user's evolving preferences. We evaluate the effectiveness of these solutions through empirical studies on six large-scale real life data sets.

Information propagation in a microblog network aims to identify a set of seed users for propagating a target message to as many interested users as possible. This problem differs from the traditional influence maximization in two major ways: it has a content-rich target message for propagation and it treats each link in the network as communication on certain topics and emphasizes the topic relevance of such communication in propagating the target message. In realistic situations, however, the topics associated with a link are not explicitly expressed but are hidden in the microblogs previously exchanged through the link. In this paper, we present a topic-aware solution to information propagation in a microblog network. We first model the latent topic structure of the network using observed microblog messages published in the network. We then present two methods for estimating the propagation probability based on the topic relevance between a link and the target message. Once the propagation probability is estimated, we adopt the standard greedy algorithm for influence maximization to find seed users. This approach is topic-aware in that the target message finds its way of propagation according to its topic relevance to the latent topic structure in the network. Experiments conducted on real Twitter datasets suggest that the proposed methods are able to select right seed users.