Large Language Models are Text Renders In order to lend interaction wings to recommender systems that are trained on offline historical user behaviors, we propose COnversational agents and REcommender systems (CORE) to bridge conversational agents and recommender systems via an unified uncertainty minimization framework. It can allow any recommendation platform to online query user preference with a plug-and-play conversational agent. In this use case, large language models act as renders to enable our conversational agent communicate as a human. [NeurIPS 2023] We propose Multi-round Auto Guess-and-Update System (MAGUS) organizing a multi-round guess-and-update system that can be applied to any recommender system to allow the recommendation of both queries and items. In this regard, MAGUS can be regarded as a simplified version of free-text conversational agents. [PREPRINT 2023a]

Large Language Models are Data Filters We notice that large language models are capable to memorize knowledge and hold chain of throughs. We propose Golden plUg-iN DAta Manager (GUNDAM), a novel framework that measures sufficiency and necessity of plugging-in demonstrations conditioned on a specific large language models. In this use case, large language models perform as filters to extract high-quality plug-in data, which can be further used for both unsupervised learning setting (a.k.a., few-shot inference) and supervisied learning setting (a.k.a., fine-tuning). [PREPRINT 2023b]

Mining Interactive Patterns on Sequences for Recommendations We define "interactive patterns" as "AND" (e.g., if a user is Chinese AND the date is Chinese New Year, then she is likely to buy dumplings). We propose Heterogeneous INteract and aggreGatE (GraphHINGE). In GraphHINGE, we uese the predefined metapaths to sample the sequences of users and items. We design a convolutional block with fast Fourier transform to get interactions before aggregating the representations to final predcition scores for item recommendations. [KDD 2020] [TOIS 2021] [🏆 AWS Machine Learning Research Project Award] With the prevalence of live broadcast business nowadays, live broadcast recommendation, is widely used in many mobile e-commerce Apps. We propose TWo-side Interactive NetworkS (TWINS) to fully mine the interactive patterns of both static and dynamic information on user and anchor sides, which is the first work for live broadcast recommendation to our knowledge. TWINS is verified by online A/B tests and has been deployed on Diantao anchor recommendation platform. [WWW 2022a] [🔌 Real-world Deployment]

Mining Relevant Patterns on Sequences for Recommendations We define "relevant patterns" by the browsed items relevant to the items to predict. We propose Search-based Time-Aware Recommendation (STARec) which captures the evolving demands of users over time through a unified search-based time-aware model. STARec can be simply deployed into any sequential network to scaling up it to handle long sequence data. STARec is verified by online A/B tests and has been successfully deployed on China Merchants Bank item recommendation platform. [WWW 2022b] [🔌 Real-world Deployment]

Mining Unbiased Patterns on Sequences for Recommendations Implicit feedback (e.g., user clicks) is an important source of data for modern search engines but heavily biased. We propose Deep Recurrent Survival Ranking (DRSR) that incorporates survival analysis techniques with recurrent neural network to derive unbiased estimations of user feedbacks. [SIGIR 2020a] As DRSR models the relevance of click-through rate, we further it to Hierarchical rEcurrent Ranking On the Entire Space (HEROES) to jointly optimize click-through rate and conversation predictions through behavior path "observation→ click→conversion". HEROES could incorporate the contextual information to estimate these behaviors in a multi-scale fashion. [CIKM 2022] Despite elaborate architecture design, we propose a simple yet sufficient unbiased learning-to-rank paradigm named unbiased Ranking via mutual Information minimization (InfoRank) that can simultaneously mitigate the effects of both position and popularity biases by summarizing the impacts of those biases into a single observation factor, therefore providing a unified treatment of the bias problem. [PREPRINT 2023c]

Towards Next Generation of Top-N Recommendation and Ranking Models We notice that Probabilistic Ranking Principle would neglect the contextual dependence among the candidate items, and make the whole pipeline non-differentiable due to the sort operation. We propose Set-To-Arrangement Ranking (STARank), a new framework directly generates the permutations of the candidate items without the need for individually scoring and sort operations; and is end-to-end differentiable. [CIKM 2023]

Mining Relevant Patterns on Graphs for Edge Predictions We refine "relevant patterns" by symbolic semantics on graphs, representing with meta-paths. In order to bridge between symbolic semantics and subgraph topology, we propose ANalogy subGraph Embedding Learning (GraphANGEL) to make a prediction of each node pair based on the subgraphs containing the pair, as well as other (analogy) subgraphs with the same graph patterns. [ICLR 2022a]

Simple Enhancements on Graph Neural Networks for Node and Edge Predictions Graph neural networks and label propagation represent two interrelated modeling strategies designed to exploit graph structure. We propose a Label Trick to accommodate the parallel use of features and labels, which uses a randomly-selected portion of the training labels as model inputs, concatenated with the original node features for making predictions on the remaining labels. [DLG-KDD 2021] [🏆 Best Paper Award] It is foundational to many of the top-ranking submissions on the Open Graph Benchmark leaderboard. We further theoretically show that under certain simplifying assumptions, the stochastic label trick can be reduced to an interpretable, deterministic training objective. [ICLR 2022b] We reveal that the idea of Label Trick would be extended to the edge prediction task. To this end, we introduce an edge splitter to specify the use of training edges where each edge is solely used as either topology or supervision. Based on splitted edges, we develop Edge-aware Message PassIng based REcommendation (EMPIRE) that generates the messages to the source nodes from their neighbor nodes (through topology edges) being aware of the target nodes (through supervision edges). [PREPRINT 2023d]

New Graph Enhanced Sequential Decision-Making Systems Goal-oriented reinforcement learning is a promising approach for scaling up reinforcement learning techniques on sparse reward environments requiring long horizon planning. We propose Graph-Enhanced reinforcement learning (GrapE), a new framework for effective exploration and efficient training based on the state-transition graph. [PREPRINT 2022]

Learning and Planning Algorithms on Ride-Hailing Platform In a general view, there are two major decision-making tasks for such ride-hailing platforms, namely order dispatching and fleet management. Most of the previous work deals with either order dispatching or fleet management without regarding the high correlation of these two tasks. We propose CoRide to jointly model order dispatching and fleet management with hierarchical reinforcement learning. This research is supported by DiDi Athena Joint Talent Program. [CIKM 2019a] [🏆 Gaia Young Scholar Outstanding Project Award]

Large Language Model Empowered Data-driven Systems
We have organized GUNDAM Labet and CORE Labet. GUNDAM is a data manager that utilizes language models to effectively handle textual data; and CORE is a plug-and-play conversational agent designed to seamlessly integrate with any recommender system. We have pre-release versions (a.k.a., version 0.0) of GUNDAM and CORE. We are currently working on exuberating their capabilities.

Simple Computer Game and Website Game Our team (as the Team Leader) build an operational game [Code] in C++ without any engine (over 10,000 lines) on Windows. Be a Hero! Inspired by Get MIT, we develop 2048 game [Game]. Use your arrow keys to move the tiles. When two tiles with the same college touch, they merge into one! NOTE: The order of colleges is not intended to be a ranking, but a rough north-to-south geographical path to SJTU.