A Distributed Recommender System Based on Graded Multi-label Classification

Abstract

International audienceRecommender systems are designed to find items in which each user has most likely the highest interest. Items can be of any type such as commercial products, e-learning resources, movies, songs, and jokes. Successful web and mobile applications can collect easily thousands of users, thousands of items, and millions of item ratings in only few months. A solution to store and to process these continuously growing data is to build distributed recommender systems. The challenging task is to find the appropriate distribution strategy allowing an efficient retrieval of needed information. Considering the similarity between the task of predicting a rating, and the task of predicting a membership grade in graded multi-label classification (GMLC), we propose an adapted distribution strategy to efficiently build a decentralized recommender system based on GMLC

An Effective Random Forest Approach for Mining Graded Multi-Label Data

“…Recently, in [12], the graded multi-label setting was demonstrated to be a more fitting paradigm for Music Emotion Recognition than the standard single-label and multilabel approaches. Furthermore, GMLC was successfully deployed as a framework for recommendation systems [13][14][15].…”

“…Albeit its simplicity, BR completely ignores the underlying dependencies between the labels which lead to a loss in information and a decline in accuracy. Different transformationbased approaches were proposed with the aim of taking into account these interdependencies [13][14][15][18][19][20]. These methods relied on intuitive transformation schemas, thus decomposing the original problem into a number of multi-label classification tasks that are then solved using various approaches based on RPC (ranking by pairwise comparison), CLR (calibrated label ranking), IBLR (instance based logistic regression), etc.…”