Behavior-Sequence-Transformer-Pytorch

This is a pytorch implementation for the BST model from Alibaba https://arxiv.org/pdf/1905.06874.pdf

This model is a novel recommender architecture based on seq2seq models. We translate user behaviour into sequences and predict a rating for each target item (movie).

Dataset

For this implementation we used Movielens 1M Dataset that contains timestamps per each rating, making it perfect to test in the sequence recommendation model.

Running

You can run it in colab here. If you prefer to run locally the model architecture is contained on pytorch-best.ipynb while data processing is on the prepare_data.ipynb notebook and should be run first.

Results

Training on all-1 user ratings and leaving the latest rating for test we obtain the following results

Here is a screenshot of training logs we we see overfitting from epoch 12-15.

References

• Original paper 1
• Keras implementation 2
• Tensorflow implementation 3