OPTIMIZED COLLABORATIVE RECOMMENDATION ALGORITHM FOR AUTO DETECTED GROUPS

Abstract

Recommender Systems (RSs) are software tools and techniques used for providing suggestions to user based on either user’s individual taste or are based on resemblance of a user to another user. Considering large number of items available to choose from it becomes difficult for user to make a choice, recommender system helps a user to make an informed decision by calculating the likelihood of user liking an item. Recommender systems are categorised into 6 different branches named: Content based, Collaborative filtering based recommender systems, Demographic, Knowledge-based, Community based recommender systems, Hybrid recommender systems. Collaborative filtering is the most successful and widely used technique in recommendation systems. This approach recommends to user the items that other similar users have liked in the past. Similarity is calculated between users based on similarity of rating history of users. In this project the main focus has been on group recommendations based on collaborative filtering method incorporated with various group modelling strategies [1]. These group modelling strategies combines various user model into a single model and represent the available knowledge about user preference belonging to a group. The users are clustered into groups according to their ratings using a variant of k-means clustering algorithm EZ Hybrid [4]. After clustering the users and group formation the group ratings are predicted for all the items and finally the prediction accuracy is tested for the different strategies. Additive utilitarian, Approval voting with threshold 1 and 2, least misery and most pleasure strategy from existing literature were implemented. A new group modelling strategy “Median strategy” is proposed and its performance is compared with those present in the literature. 1M dataset from Movielens is used in the experiment. RMSE, MAE, Precision and Recall are the parameters used to measure the performance of prediction accuracy for auto detected groups. From results we come to a conclusion that the new proposed strategy gives better Precision, Recall, and MAE compared to already present in literature. MAE is improved by 6.67%, precision by 7.77% and recall by 9.69%. Also the RMSE values are better than all other strategies except additive utilitarian. Hence we conclude that using median strategy helps in making more accurate predictions.