top of page
Frame 5cover.png
Group 4icon.png

2021.3 - 2021.5

Transformer VS LSTM 

in Predicting User Ratings

Social Computing /  Deep Learning / Algorithm

Team members

Lixing Liu / Han Bao / Xingyu Li / Jiarui Xu

Role

Programmer / Researcher

Tools

project12.png

Paper

The project aims to find out whether the recently popular attention-based model of Transformer could perform better than the previous baseline proposed in works. We will compare the performance of the transformer model using a similar input setup to the work of Behavior Sequence Transformer to a baseline Long Short-Term Memory model using the same input setup. We measure the performance of the two models on the dataset MovieLens which contains 3,900 movies, 6,040 users, and 1,000,209 ratings, and each user has rated at least 20 movies, with integer scores ranging from 1 to 5. MovieLens is widely used to test and develop recommendation algorithms, especially collaborative filtering algorithms. The two models are trained and validated on predicting user ratings based on sequences of feature inputs. The Long Short-Term memory model achieves an average root mean square error of approximately 1.3547 on validation data, and the transformer model achieves an average root mean square error of approximately 1.2122 on validation data.

bottom of page