Recommendation Framework via Matrix Factorization and Translation

Abstract

According to the recent technical report from Amazon.com, estimated 30 percent of their page views were from recommendations, and likewise as for Netflix, more than 80 percent of movies watched on Netflix came through recommendations, and the value of Netflix recommendations is estimated at more than US$1 billion per year. Therefore, research on recommender systems is of great value to players in the industry, and they are striving hard to build successful recommender systems.

Among various recommendation techniques, collaborative filtering (CF) has been shown to be the most successful; it assumes that users who have had similar interests in the past will tend to share similar interests in the future. While CF is a promising direction of recommender system, it fails to achieve a high performance without sufficient amount of user-item interaction records or user feedback; this phenomenon is also known as the data sparsity problem. In this dissertation, we focus on data-centric approaches to overcoming the data sparsity problem prevalent in recommender systems. We postulate that to effectively alleviate the data sparsity problem, understanding characteristics of data related to users and items should be preceded, and that the recommendation models should be built upon the insights obtained from them. Our data-centric approaches can be categorized into two different streams; auxiliary data-driven approach and algorithmic approach.

The auxiliary data-driven approach refers to methods that leverage side information related to users and items to make up for the lack of user feedback on items; i.e., to solve the data sparsity problem. As the first approach along this line, we propose a social network-based recommendation method that incorporates users’ social network into recommendation by modeling two different roles of users as trusters and trustees while considering the structure of social network. Then, we propose an image-based recommendation method that leverages information hidden in the so-called “also-viewed” products, to also account for the non-visual item aspects overlooked by previous visually-aware recommender systems. Lastly, we propose an implicit feedback-based recommendation method that focuses on dealing with the missing user-item interactions of purchase records by leveraging users’ past click records. The algorithmic approach aims at fully exploiting the hidden properties of data under the algorithmic perspective, whose ultimate goal is to solve the data sparsity problem. As the first approach along this line, we first propose a data sampling scheme called decision boundary focused under-sampling (DBFUS), which under-samples majority class regarding the distance to the decision boundary to retain the classification accuracy, to cope with class imbalance problem among positive and negative instances. Then, we propose a metric learning-based method to discover the intensity and the heterogeneity of user-item relationships embodied in implicit user–item interactions.

We validate the benefit of the above data-centric approaches through extensive experiments. To summarize, this dissertation sheds light on the power of data-centric approaches for solving the data sparsity problem.

본 논문은 추천 시스템의 고전적인 문제인 데이터 부족 현상을 해결하기 위한 데이터 중심적 접근 방법을 제안한다. 특히, 행렬분해 기법, 메트릭 학습 기법을 기반으로 한 방법들을 소개한다. 데이터 중심적 접근 방법은 크게 두가지로 나뉜다. 첫번째는, 사용자 및 품목에 관련된 보조 데이터를 기반으로 한 접근 방법이며, 두번째는, 알고리즘적인 접근 방법이다. 본 학위 논문에서는 다양한 실험을 통해 본 논문에서 제안한 방법들의 우수성을 입증하며, 이는 추천 시스템의 데이터 부족 현상을 해결하는데 새로운 방향을 제시 한다.