Sequential Recommenders

Modeling users' dynamic interests has drawn increasingly attention. In this project, we try to implement two kinds of sequential recommenders, i.e., Markov Chains-based methods and neural network-based methods. Currently, we include following models:

Neural models:

The task of sequential recommendation is to predict next item(s) that a user is likely to choose based on his/her ordered behavior sequence. Therefore, neural networks that are sequence-aware can be used to encode the user's history.

* RNN [1]. This is a modified version of GRU4Rec. We train the RNN(LSTM) model via standard BBTT, instead of the per-step training in original paper[1].
* TCN [2]. Use temporal convolutional neural network (a.k.a. dilated causal convolutional neural network) to model sequence data.
* Transformer [3]. Use Transformer (a.k.a. self-attention neural network) to model sequence data.

Markov Chains:

Methods which rely on Markov Chains (MCs) also assume users' next item depend on their previous item(s). We implement two main MCs-based methods.

* FPMC [4]. FPMC models users's dynamic interests by factorizaing item-to-item transitions.
* Fossil [5]. Fossil follows the similar idea of FPMC in modeling users' dynamic interests.

DISCLAIMER: Since we intend to unify these methods into the same framework, we cannot guarantee that all models are implemented the same as authors' official implementations.

References

Hidasi et al. Session-based recommendations with recurrent neural networks. ICLR'16.
Yuan et al. A Simple Convolutional Generative Network for Next Item Recommendation. WSDM'19.
Kang and McAuley. Self-Attentive Sequential Recommendation. ICDM'18.
Rendle et al. Factorizing Personalized Markov Chains for Next-Basket Recommendation. WWW'10.
He et al. Fusing similarity models with markov chains for sparse sequential recommendation. ICDM'16.