This repository provides a summary of our research on Recommender Systems. It includes our code base on different recommendation topics, a comprehensive reading list and a set of bechmark data sets.
Currently, we are interested in sequential recommendation, feature-based recommendation and social recommendation.
Since users' interests are naturally dynamic, modeling users' sequential behaviors can learn contextual representations of users' current interests and therefore provide more accurate recommendations. In this project, we include some state-of-the-art sequential recommenders that empoly advanced sequence modeling techniques, such as Markov Chains (MCs), Recurrent Neural Networks (RNNs), Temporal Convolutional Neural Networks (TCN) and Self-attentive Neural Networks (Transformer).
A general method for recommendation is to predict the click probabilities given users' profiles and items' features, which is known as CTR prediction. For CTR prediction, a core task is to learn (high-order) feature interactions because feature combinations are usually powerful indicators for prediction. However, enumerating all the possible high-order features will exponentially increase the dimension of data, leading to a more serious problem of model overfitting. In this work, we propose to learn low-dimentional representations of combinatorial features with self-attention mechanism, by which feature interactions are automatically implemented. Quantitative results show that our model have good prediction performance as well as satisfactory efficiency.
Online social communities are an essential part of today's online experience. What we do or what we choose may be explicitly or implicitly influenced by our friends. In this project, we study the social influences in session-based recommendations, which simultaneously model users' dynamic interests and context-dependent social influences. First, we model users' dynamic interests with recurrent neural networks. In order to model context-dependent social influences, we propose to employ attention-based graph convolutional neural networks to differentiate friends' dynamic infuences in different behavior sessions.
We maintain a reading list of RecSys papers to keep track of up-to-date research.
We provide a summary of existing benchmark data sets for evaluating recommendation methods.
We contribute a new large-scale dataset, which is collected from a popular movie/music/book review website Douban (www.douban.com). The data set could be useful for researches on sequential recommendation, social recommendation and multi-domain recommendation. See details here.
