The task of text-to-SQL parsing, which aims at converting natural language questions into executable SQL queries, has garnered increasing attention in recent years. One of the major challenges in text-to-SQL parsing is domain generalization, i.e., how to generalize well to unseen databases. Recently, the pre-trained text-to-text transformer model, namely T5, though not specialized for text-to-SQL parsing, has achieved state-of-the-art performance on standard benchmarks targeting domain generalization. In this work, we explore ways to further augment the pre-trained T5 model with specialized components for text-to-SQL parsing. Such components are expected to introduce structural inductive bias into text-to-SQL parsers thus improving the model’s capacity on (potentially multi-hop) reasoning, which is critical for generating structure-rich SQLs. To this end, we propose a new architecture GRAPHIX-T5, a mixed model with the standard pre-trained transformer model augmented by specially-designed graph-aware layers. Extensive experiments and analysis demonstrate the effectiveness of GRAPHIX-T5 across four text-to-SQL benchmarks: SPIDER, SYN, REALISTIC and DK. GRAPHIX-T5 surpasses all other T5-based parsers with a significant margin, achieving new state-of-the-art performance. Notably, GRAPHIX-T5-large reaches performance superior to the original T5-large by 5.7% on exact match (EM) accuracy and 6.6% on execution accuracy (EX). This even outperforms the T5-3B by 1.2% on EM and 1.5% on EX
24 March 2024 in Proceedings of the AAAI Conference on Artificial Intelligence
Personalized recommender systems have found widespread applications for effective information filtering. Conventional models engage in knowledge mining within the static setting to reconstruct singular historical data. Nonetheless, the dynamics of real-world environments are in a constant state of flux, rendering acquired model knowledge inadequate for accommodating emergent trends and thus leading to notable recommendation performance decline. Given the typically prohibitive cost of exhaustive model retraining, it has emerged to study incremental learning for recommender systems with ever-growing data. In this paper, we propose an effective model-agnostic framework, namely INFluential Exemplar Replay (INFER). INFER facilitates recommender models in retaining the earlier assimilated knowledge, e.g., users' enduring preferences, while concurrently accommodating evolving trends manifested in users' new interaction behaviors. We commence with a vanilla implementation that centers on identifying the most representative data samples for effective consolidation of early knowledge. Subsequently, we propose an advanced solution, namely INFERONCE, to optimize the computational overhead associated with the vanilla implementation. Extensive experiments on four prototypical backbone models, two classic recommendation tasks, and four widely used benchmarks consistently demonstrate the effectiveness of our method as well as its compatibility for extending to several incremental recommender models.
13 June 2023 in Proceedings of the ACM on Management of Data
In this paper, for the first time, we introduce the concepts of window-CCs and window-SCCs on undirected and directed temporal graphs, respectively. We then study the queries of window-CC and window-SCC by developing several efficient index-based query solutions. The space costs of the best indices are linear to the sizes of the temporal graphs. The extensive experimental evaluation on 12 real-world datasets demonstrates the high efficiency and effectiveness of the proposed solutions. In the future, we will develop distributed index construction algorithms, which would be useful for very large temporal graphs containing billions of edges. In the future, we will implement our algorithms by using a distributed computing platform (e.g., Pregel), which would be very useful when the temporal graph is too large to be kept by a single machine.
13 June 2023 in Proceedings of the ACM on Management of Data
The densest subgraph problem (DSP) is of great significance due to its wide applications in different domains. Meanwhile, diverse requirements in various applications lead to different density variants for DSP. Unfortunately, existing DSP algorithms cannot be easily extended to handle those variants efficiently and accurately. To fill this gap, we first unify different density metrics into a generalized density definition. We further propose a new model, c-core, to locate the general densest subgraph and show its advantage in accelerating the searching process. Extensive experiments show that our c-core-based optimization can provide up to three orders of magnitude speedup over baselines. Moreover, we study an important variant of DSP under a size constraint, namely the densest-at-least-k-subgraph (DalkS) problem. We propose an algorithm based on graph decomposition, and it is likely to give a solution that is at least 0.8 of the optimal density in our experiments, while the state-of-the-art method can only ensure a solution with density at least 0.5 of the optimal density. Our experiments show that our DalkS algorithm can achieve at least 0.99 of the optimal density for over one-third of all possible size constraints.
29 May 2024 in Proceedings of the ACM on Management of Data
Densest subgraph discovery (DSD) is a fundamental topic in graph mining. It has been extensively studied in the literature and has found many real applications in a wide range of fields, such as biology, finance, and social networks. As a typical problem of DSD, the k-clique densest subgraph (CDS) problem aims to detect a subgraph from a graph, such that the ratio of the number of k-cliques over the number of its vertices is maximized. This problem has received plenty of attention in the literature, and is widely used in identifying larger ''near-cliques''. Existing CDS solutions, either k-core or convex programming based solutions, often need to enumerate almost all the k-cliques, which is very inefficient because real-world graphs usually have a vast number of k-cliques. To improve the efficiency, in this paper, we propose a novel framework based on the Frank-Wolfe algorithm, which only needs k-clique counting, rather than k-clique enumeration, where the former one is often much faster than the latter one. Based on the framework, we develop an efficient approximation algorithm, by employing the state-of-the-art k-clique counting algorithm and proposing some optimization techniques. We have performed extensive experimental evaluation on 14 real-world large graphs and the results demonstrate the high efficiency of our algorithms. Particularly, our algorithm is up to seven orders of magnitude faster than the state-of-the-art algorithm with the same accuracy guarantee.
19 August 2024 in IEEE Transactions on Knowledge and Data Engineering
Searching on bipartite graphs serves as a fundamental task for various real-world applications, such as recommendation systems, database retrieval, and document querying. Conventional approaches rely on similarity matching in continuous euclidean space of vectorized node embeddings. To handle intensive similarity computation efficiently, hashing techniques for graph-structured data have emerged as a prominent research direction. However, despite the retrieval efficiency in Hamming space, previous studies have encountered catastrophic performance decay . To address this challenge, we investigate the problem of hashing with Graph Convolutional Network for effective Top-N search. Our findings indicate the learning effectiveness of incorporating hashing techniques within the exploration of bipartite graph reception fields, as opposed to simply treating hashing as post-processing to output embeddings. To further enhance the model performance, we advance upon these findings and propose B ipartite G raph C ontrastive H ashing ( BGCH+ ). BGCH+ introduces a novel dual augmentation approach to both intermediate information and hash code outputs in the latent feature spaces, thereby producing more expressive and robust hash codes within a dual self-supervised learning paradigm. Comprehensive empirical analyses on six real-world benchmarks validate the effectiveness of our dual feature contrastive learning in boosting the performance of BGCH+ compared to existing approaches.