Cross-domain recommendation through implicit extraction of psychological factors

Abstract

Internet has become one the main sources of information for a majority of people. The problem of information overload affects the capacity of processing and discerning useful knowledge from the received information. Recommender systems were created with the objective of facilitating and supporting the process of decision making in online environments. These systems help in a technical way by bringing up the right knowledge in an effective and efficient manner. However, psychological aspects must be considered as well in order to design human-computer interfaces that enhance user’s trust in the system. Traditional recommender systems are built up exploiting the existing similarities in behavior be-tween users, or by clustering and suggesting items that share certain common attributes. In most applications, hybrid solutions are a common implementation, where a mix of techniques is employed. To center the focus on the user, these systems are further improved taking contextual, situational, demographic data as well as personality facts. This information can be extracted in an explicit way, via questionnaires or asking customers to establish their preferences and personal data. On the other hand, obtaining implicit, previously unknown and useful information through data mining and machine learning techniques is a method widely used nowadays. Modern e-commerce sites do not focus only in a certain market, but encourages customers to buy products from different domains. Customers also express their opinions on diverse social media and use diverse providers in order to satisfy their needs. Cross recommender systems try to address both issues taking advantage of the relationship between domains and their current users. At the same time, help to tackle the so called “cold start” problem, when a new user or a new item appears but there is not enough information to apply traditional techniques. In this study, a hybrid system solution is offered for users of a previous domain getting recommendations in a sparse new domain. The idea is to leverage preferences by aggregating implicit information from unstructured data using text mining techniques over reviews. Polarity sentiment over items was discerned using NLTK (Natural Language Tool Kit), a tool that uses semantic analysis to extract subjective sentiments about customer’s opinions. Psychological traits were extracted from the text by using LIWC (Language In-quiry and Word Count), and transformed into domains described by the Five Factor Model. The values from the Five Factor Model were loaded into feature vectors in order to be used by the system. For the final solution data from books as a source domain and movies as a target domain were used. Firstly, in order to extract psychological factors only those customers with a certain quantity of overall total words could be used. This restriction was alleviated by the fact that certain users reviewed items in both domains. Those users were selected as the neighbor, their psychological factors extracted and preferences created. Secondly, for neighbor user’s reviews, item genres were crawled, and item ratings were predicted using user similarity within the same genre. The solution was tested using 10 cross fold validation and compared with a baseline K-NN collaborative filtering algorithm.