Enhancing Performance of Recommender Systems

Abstract

Freedom of choice is our right. Coincidentally, the right to choose is the freedom. It may also be said that it is very difficult to make right choices when many options are available. In terms of Internetworking applications, this century can be dedicated to Social Networking. Information sharing across the globe, is rampant. Most of the options are also already available online. It is crucial to get information suiting ones requirements from the available data. There is a need for practical applications that help users to deal with information overload and provide personalized recommendations, content and services to them. These recommendations are highlighted by vendors in terms of advertisements, short interviews during tele-marketing and through feedbacks. Every transaction also records the choice of an individual. Servers record these transactions and use them in commercial ratings. Still, there exist confusion among the users to choose the best tting option within their de ned behavioural constraints. Recommender system is one of the applications to predict rating or preference for the items that have not been seen by a user. This system typically produces a list of recommendations. Recommending books, CDs, and other products at amazon. com, movies by MovieLens, and news at VERSIFI Technologies (formerly adaptiveInfo. com) are examples of such applications to name a few. However, despite these developments, the current generation of recommender systems still requires further improvements to make recommendation methods more accurate and applicable to an even broader range of real-life needs including recommending vacations, research articles, URLs, music artists, social tags and certain types of financial services to investors. These improvements also cater to mapping the behaviour of an individual with the corresponding choice of an item. Hence, advanced recommendation modelling methods, incorporation of various contextual information into the recommendation process, and measures to determine performance of recommender systems are considered. Raw data used for processing is available under various categories. These categories are positive and unlabelled data, combination of labelled and unlabelled data and time-series data to name a few. Different applications generate different categories of data. Data generated through social bookmarking systems (Bibsonomy or Delicious) is an example of positive and unlabelled data. Labelled and unlabelled data is aggregated in feedback application for any item, where an item may be liked or disliked by individuals (E.g. movie feedback system). Continuously changing data (time-series) is available for example from stock market. Depending on type of data, processing in recommender system changes. Hence, this thesis focuses on optimizing performance of recommender system, catering to the requirements of these different categories of data. Recommending on the basis of positive and unlabelled data is crucial, especially if the information is available through Social Networking sites. Social bookmarking sites such as Bibsonomy or Delicious allow user to bookmark URLs and submit the research articles. User bookmarking a resource (URL) or submitting a resource (research article) on this system, implicitly indicates his likings to this resource. These resources are considered as positive examples of the user preference. Other resources (URLs/research articles), however, do not imply negative preference of the user about them. This leads to the situation where we have positive examples but no negative examples for user preference. A recommender which is learnt from positive and unlabeled examples is devised to recommend these social resources to user. The work also focuses on recommending tags for the resources being submitted first time to the social bookmarking site. The task is modeled as multi-label text classification. Naive-Bayes classifier is used as the base learner of the multi-label text classifier and multinomial distribution is fetted to the data for experimentation. Recommendation system is then developed considering a variety of tags associated with an item. A movie recommendation application is one such application. A movie recommendation system is hence proposed, that combines ratings, tags, genres and star cast of movies. In a recommender system, the items which are liked or disliked by the user are the labelled examples about the users' preference. However, there are many other items which are not rated by the user. These items form the set of unlabelled examples. One of the major problem with recommender system is that it does not have adequate number of labelled examples for the new user. E cacy of co-training algorithm in learning a recommender by exploiting unlabelled examples in the presence of small number of labelled examples is critical. Recommendation system is also one of the outcomes of prediction in several applications. The best application of such a recommendation system is in stock market where data keeps changing with time. The problem of predicting direction of movement of stock price and stock market index is handled by rst converting the data into trend deterministic data and then learning through various machine learning techniques using this data. To predict future values of stock market indices, hybrid techniques combining support vector regression with other machine learning techniques is developed. The techniques predict value of stock price indices for 1 to 10, 15 and 30 days in advance. In a nutshell, these recommendations are not limited to a list of unseen items, but also include some predictions or forecasts to help users in making appropriate decisions.