RECOMMENDATIONS SYSTEM BASED ON MATRIXFACTORIZATION FOR DIGITAL EDUCATION PLATFORM

• Main Author: Nabil Fadhlurrahman, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79770
• Institution: Institut Teknologi Bandung

The utilization of matrix factorization methods in an effort to create a recommendation system is explored in this final project. The challenge of creating a recommendation system arose with the advent of the World Wide Web or WWW, which later became accessible to the public. Since then, all kinds of information have become highly accessible on the internet. This is believed to have positive impacts such as easy access to information and the acceleration of the escalation of knowledge. However, there are also negative impacts from this. The ease and openness of access are believed to be one of the causes of Information Overload and Decision Fatigue. In this final project, we will attempt to address the mentioned problems by creating a recommendation system. Specifically, this recommendation system will be designed for users of the Dicoding Indonesia service in an effort to enhance their experience. The improvement in experience is based on the approach of Personalized Learning Experience in the form of class recommendations from a recommendation system that we create. The algorithm we use in building the recommendation system is based on matrix factorization methods. Of course, we will create algorithms ranging from simple or traditional ones to those utilizing artificial neural networks. Not forgetting, after creating a recommendation system, we undergo an evaluation process. Various evaluation processes are used, both in terms of model building and user aspects. From various experiments we conducted by testing the recommendation system with various metrics such as regression and classification, the SVD model has the most balanced results in both regression and classification metrics, namely 0.3136 for RMSE, 0.8758 for Precision@K, and 0.9037 for Recall@K. However, if we only prioritize regression metrics, the model that utilizes low-high feature interactions and artificial neural networks has the best result, which is RMSE 0.2493.