A user rating based collaborative filtering approach to predict movie preferences
Md. Asif Shahjalal, Zubaer Ahmad, Mohammad Shamsul Arefin, Mohammad Rubaiyat Tanvir Hossain · 2017 3rd International Conference on Electrical Information and Communication Technology (EICT) · 2017
Recommender systems are arguably one the most successful and widely popular commercial application of machine learning. This paper discusses and presents a collaborative filtering based movie recommender system. One of the most challenging issues in machine learning is choosing perfect features for learning algorithms. However, choosing features manually can be quite difficult and tiresome work in many complex situations. The traditional content-based recommender systems require well-defined feature matrix for the movies. However, such features are not readily available for a large dataset of movies. This is where collaborative filtering comes to rescue. It can select and learn the features all by itself. Here prediction is based on user behavior. The real advantage is that the features learned by the algorithm do not need to be human defined. A user rating based low-rank matrix factorization collaborative filtering approach has been selected for this research. MovieLens 100k dataset [1] from GroupLens research has been used in this research. To minimize the squared error of the cost function gradient descent algorithm has been used in this approach. Here all the users collaboratively contribute to the development of different features that will help new users in the future. This approach saves the time consuming, expensive and difficult nature of content based recommender as well as saving the user from a lot of hassles like filling up a long survey form. This paper is based on the study of popular Coursera course “Machine Learning” [2] by professor Andrew Ng from Stanford University. The cost function minimization to minimize the squared error is done using gradient descent algorithm which ensured the visual confirmation of the cost minimization and to provide a visual representation of the cost minimization. Also, the outliers in the prediction matrix are eliminated using a threshold value. A prototype has been proposed and implemented with the help of Matlab.