TopicVAE

TopicVAE

This is the Tensorflow implementation for our ACM MM 2022 paper:

Zhiqiang Guo, Guohui Li, Jianjun Li, Huaicong Chen. TopicVAE: Topic-aware Disentanglement Representation Learning for Enhanced Recommendation. In MM 2022. Paper

Introduction

In this work, we propose a novel Topic-aware Disentangled Variational AutoEncoder (TopicVAE) to disentangle and infer representations of users and items from content information for enhanced recommendation. As shown in the framework, TopicVAE mainly contains four modules: attention-based topic extraction (ATE), topic-level variational autoencoder (TVAE), topic-guided contrastive loss (TCL) and heuristic regularization (HREG).

Environment

pip install -r requirements.txt

Data processing

We provide one processed dataset: Industrial. To process other amazon datasets, you first need to download the json file with reviews and the word embedding vectors generated by FastText model pre-trained on Wikipedia, and put them in the data folder.

Then you can run process_fasttext.py to get the final processed datasets.

python process_fasttext.py --data=data/Industrial.json

NOTE: other word vertor models can also be used to process the dataset, such as Word2Vec, GloVe.

Run

After processing the datasets, you can test TopicVAE/TopicDAE on Industrial dataset by:

python TopicVAE.py --data=data/fasttext/Industrial --mode=tst
python TopicDAE.py --data=data/fasttext/Industrial --mode=tst

or retrain TopicVAE/TopicDAE by:

python TopicVAE.py --data=data/fasttext/Industrial --mode=trn
python TopicDAE.py --data=data/fasttext/Industrial --mode=trn

You also can set different parameters to train this model.