Improving Android game recommendations

Explaining the science

Recommender systems

When considering recommender systems, the problem can be framed in terms of two types of entities: users - individuals to whom the recommendations are made - and items - entities that are recommended to the users such as movies, news articles or in our case, mobile games. The relationship between users and items can be represented by a so-called user-item matrix, where every element represents the preference of a given user for a specific item, here this relationship is the preference of a game player given an Android game. These preferences can be provided by users explicitly in the form of a rating (e.g., a number between 0 and 5).

Alternatively, preferences can be induced from implicit feedback provided by the users, for instance, you can use the number of times a user has listened to a particular song to deduce whether the user likes the song or not. However, in either cases, there isn't preference information available for every user and for every item. Take users of the online shopping site Amazon, it is quite unlikely for every user user to have a feedback for every single item on the website. The main goal of a recommender tool is therefore to fill in the missing preference information for users. Items with higher preference score will then be recommended to the users.

Getting the right information

As people spend more time online and on their mobile phones, the quantity of data they generate is reaching staggering proportions. Although the number of users and items do increase, the number of different items that users interact with does not necessarily increase, which leads to almost entirely incomplete user-item matrices. Traditional approaches attempt to infer user preference by using solely the known interaction between the users and items, however these methods perform poorly when these known user-item interactions, i.e. preferences, are extremely limited.

Recent developments therefore are investigating whether incorporating auxiliary information can lead to better methods to estimate preference and hence provide better recommendations. For example, if creating a tool for recommending songs, it might be useful to consider additional information coming from the songs such as audio and lyrics or to add extra information about the users. This is a very intuitive approach, and allows the recommender system to learn directly which users are similar to one another, and likewise for items.

Deep learning

The striking performance of deep learning methods have placed them at the very heart of modern recommender tools. A number of studies have been proposed using deep learning methods together with more traditional approaches since these hybrid recommender systems provide a deeper understanding of different aspects of users and items as well as the relationship between them. In this project, a hybrid model is exploited based on the approach proposed in Kim et al. This method has a proven advantage especially when there is text-based auxiliary information which in this case is the descriptions of the mobile games.