The main task of the recommendation system is to mine information based on the user’s visit history and recommend items that may be of interest to the user and improve user satisfaction. Recommendation systems have many research purposes, such as recommendation accuracy, recommendation diversity, recommendation explanation etc. Recommendation accuracy is the most basic goal, aiming to predict and recommend the items that users are most likely to click on in future. Recommendation diversity requires the system to consider multiple interests of users. For example, users may like action movies and science fiction movies, so the recommendation list needs to include both kinds of movies, not just one type. Recommendation explanation points out that when recommending items, the user may want to know why these items are recommended to me and not items. The recommendation system has a lot of goals waiting for us to explore.
Lightweight Integration of Geographical Information into Graph Convolution Network for Point-of-Interest Recommendation
In the research Lightweight Integration of Geographical Information into Graph Convolution Network for Point-of-Interest Recommendation, we extend the definition of neighbors in graph neural networks. We propose the use of geographic neighbors to increase the accuracy of the recommendation and maintain the ease of training of the network. In our method, the target user not only aggregate information from the checked items, but also from so called geographical neighbors. After training the unique embeddings of users and items, the recommendation task can be completed by using an inner product of users’ and items’ embeddings.
Explaining Recommendation with Aspect-based Textual Explanation
In the research of Explaining Recommendation with Aspect-based Textual Explanation, we believe that different users would be interested in different aspects of items. We extract user’s preference from his/her textual reviews and generate reasonable explanations for the recommended results that fits user’s preference. In this research, we adopt an Attention-based Aspect Extraction method to filter out non-aspect related words and re-conduct the sentence embedding as a linear combination of all aspect embeddings. We generate explanations in this way to make users more satisfied.
Hero Recommendation for MOBA Games —-Amateur and Professional Level
MOBA games are becoming more and more popular. International championships attract millions of audiences. Hero recommendation system in drafting phase has the potential to improve the experience and gain strategic advantage for both amateur and professional players. In the research of Hero recommendation for MOBA games, we propose to use generative adversarial network to make hero recommendations to increase the probability of winning. This is a new area of machine learning. The use of machine learning will bring new ideas to the decision making in MOBA games.
Item recommendation explainability
By using LIME, which is a interpretation algorithm, for a recommendation system using a machine learning model, we aim to generate explanations for presenting any recommendation model to the user.
Point of Interest (POI) recommendation
Point of Interest (POI) recommendation is recommending some new POIs to users where they have not checked before in order to help them find some new locations and have a better understanding of the city. To get higher recommendation accuracy, we modeled users’ geographical information and combined it with collaborative filtering(CF).