Deep Learning for Recommendation System
Project Name: Deep Learning for Recommendation System
Project Sponsor: AI Singapore and Linksure Pte Ltd
Grant Amount: S$280,000
Duration: 18 months
Description: In this project, NUS collaborates with Linksure Pte Ltd to research solutions to improve the quality of the news recommendation system.
Commonly, the two main techniques applied for recommendations are Collaborative Filtering (CF) and Content-Based (CB) Recommendation. The CF technique relies on usage patterns: the combinations of items that users have consumed or rated provide information about their preferences, and how they relate to each other; Another technique is to predict user preferences from item content and metadata. The consensus is that the CF will generally outperform the CB Recommendation. However, CF could not recommend new items that have not been consumed before and unable to recommend specific things that are of interest to niche users.
The content-based recommendation, on the other hand, utilize more information to make the prediction. The user profile, user reviews, and item information can all be part of the recommender system’s input. The linear model, either logistic regression or the FTRL proposed by Google, are commonly applied in industrial companies. However, the linear model gives weight for each input feature without considering their interactions. The factorization machine(FM) takes this into account while keeping the computational cost low. However, the methods mentioned above are shallow networks that automatically cannot learn high abstract features from the input.
In this project, we develop a novel deep neural network method as well as new CF algorithms based on the industry dataset to improve the overall news recommendation systems.