New perspectives on gray sheep behavior in recommender systems

Abstract

With the rapid rise in popularity of e-commerce applications and the dramatic increase in the size of data present in such applications along with the related social media data generated by customers, information filtering technique like Recommender Systems are widely used to help the customer in finding the item they might find useful. This prediction also helps the companies in cross-selling, upselling and to increase the loyalty of their customers. Collaborative Filtering algorithms, the most widely used technique in Recommender Systems, are primarily based on the phenomenon of homophily which means that humans tend to associate and bond with other humans who are similar to them. This also gets reflected in their preferences for products and the feedback they give to them. The similarity between users in Collaborative Filtering approach is primarily measured through the inverse of the distance between their rating vectors. However, due to lack of sufficient amount of relevant ratings data despite having vast numbers of users and products in the system, there exists a sparsity problem which impacts the quality of recommendations. There also exist few users whose opinions do not consistently agree or disagree with any group of people. Hence such users don’t benefit much from the system. Moreover, their presence also reduces the overall efficiency of the system impacting the quality of recommendations made to other users. This is referred to as Gray Sheep problem, and their identification is an important area of research in Recommender systems. For increasing the efficiency of recommendations and cross-selling, Cross Domain Recommender Systems are used , which use the knowledge from domain to make recommendations in different domain. With the increasing capability to capture more and more contextual data about the users, Context Aware Recommender Systems are also gaining popularity, particularly to make real-time recommendations. In this thesis work, five research gaps have been identified which are present in the existing research literature related to Recommender Systems. They are primarily related to lack of any work done to study the Gray Sheep behavior in different domains and contexts. Since Gray Sheep behavior is a reflection of a user’s eccentric taste, we have also explored research related to the usage of a user’s personality data in an attempt to resolve Gray Sheep problem and also sparsity problem through Transfer Learning approach. Transfer Learning is a machine learning paradigm that makes use of knowledge learned in one task in a different but related source domain, to solve the task in other target domain. This is particularly useful when there is a scarcity of fewer high-quality training data. For Transfer Learning, auxiliary data sources used in this work comprise of users’ social media data and their ratings in other domains. Users’ characteristics are learned using the latest advancement in natural language processing which makes this approach more practical and scalable. Based on our observation about Gray Sheep users, we also propose the concept of Gray Sheep Items, in scenarios where item-based collaborative filtering is used. We have also empirically verified their existence using various publicly available dataset.