Similarity based regularization for online matrix-factorization problem: An application to course recommender systems

• Published in: TENCON 2017 - 2017 IEEE Region 10 Conference pp. 1874 - 1879
• Main Authors: Shah, Dhruv, Shah, Pratik, Banerjee, Asim
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2017
• Subjects: Computational modeling; Course Recommender System Recommender System; Kernel; Low-rank matrix factorization; Matrix converters; Matrix decomposition; Online Matrix Completion; Optimization; Prediction algorithms; Recommender systems
• ISSN: 2159-3450
• DOI: 10.1109/TENCON.2017.8228164

The design of a recommender system is largely influenced by its domain of application. A recommender system for niche application requires more accuracy as it targets a specific audience or a specific genre of products to recommend. Certain examples of niche domains include course recommendation for university courses, text recommendation for translators etc. In this paper, we address the problem of designing a recommender system for one such niche domain, a course recommender system. It generates recommendation of university courses for students based on the courses previously preferred by the student. Since such recommendations play a role similar to decision support systems for students, it is evident that it has to be relevant in predicting preferences. Also, every new choice made by the user unfolds additional information about a user which was previously unknown to the system. Literature suggests that course recommender systems have been developed mostly without the use of machine learning techniques. Treating student preference information as a general recommendation problem, it can be represented in a matrix format and generating a low-rank matrix representation have provided encouraging results. Inclusion of additional information in order to make more accurate predictions, leads to higher computational complexity and instability in learning parameters. To overcome such hurdles in designing a Course Recommender System, we propose a similarity based regularization for low-rank matrix factorization algorithm which learns the prediction matrix very fast and is stable.