Reimplement BST model on movie recommendation

Inspiration

We are interested in movie recommendation model and want to find a NN model in this field to reimplement it for our final project. We choose the BST model because it utilize the transformer architecture. Since transformer is SOA model in many fields, we want to see if it is also good at movie recommendation.

Paper Link:

Behavior Sequence Transformer for E-commerce Recommendation in Alibaba link Multimodal Movie Recommendation System Using Deep Learning link

Model Architecture

BST takes as input the user’s behavior sequence, including the target item, and “Other Features”. The first layer is embedding layers which embeds the input features. The second layer is transformer layers which is used to learn deeper representation of the input features. (only ENCODER is used because we do not need to generate sequence output) Then by concatenating the embeddings of Other Features and the output of the transformer layer, the three-layer MLPs are used to learn the interactions of the hidden features, and sigmoid function is used to generate the final output.

Second Paper is about a method to enhance the movie recommendation model by combining other features efficiently (combing the poster)

Dataset

Movielens: link IMDB: link The movie dataset: link

Propose Base

1. Reimplement the BST model
2. Apply it on movie recommendation task ## Target
3. Enhance Model with Rich Feature
4. Optimize Model Parameters ## stretch goal
5. Extend to Other Domains
6. Making the model able to generate movies user will like

Evaluation metrics:

RMSE and MSE

Final Deliverables:

Uploaded in additional info and also linked below

Built With

• python
• tensorflow