New technique to alleviate the cold start problem in recommender systems using information from social media and random decision forests

Abstract

The aim of recommender systems is to provide users with items that could be of their interest. However one of the biggest drawbacks from recommender systems is the so called cold start problem, which occurs when new users or products are added to the system and therefore there is no previous information about them. There are many proposals in the literature that aim to deal with this issue. In some cases the user is required to provide some explicit information about them, which demands some effort on their part. Because of that and due to the great boom of social networks, we will focus on extracting implicit information from user’s social stream. In this paper we will present an approach on which social media data will be used to create a behavioural profile to classify the users and based on this classification will create predictions making use of machine learning techniques such as classification trees and random forests. Thus the user will not have to provide actively any kind of data explicitly but their social media source, alleviating in this way the cold start problem since the system would use this data in order to create user profiles, which will be the input for the engine of the recommender systems. We have carried out numerous experiments, as well as a comparison with some other state-of-the-art new user cold-start algorithms, obtaining very satisfactory results.