Efficient recommender system for cold start items

Abstract

Recommender system is a particular type of intelligent system which predicts the rating or preference of an item that a user would like to give. It is based on information filtering process which simply filter out the undesirable material. It provides only most relevant products or information in which a particular user is interested through personalized recommendation. Currently, it is playing an important role in numerous e-commerce services, social networking applications and recommending products to people such as books, movies, news, garments, research articles, and web pages. The recommender system suffer from cold start problem of different degrees where no rating records or few ones exist for newly coming items or users into the system. Many recommender systems have been proposed in literature but their efficiency reduces in terms of accuracy when cold start items or users comes into the system. Thus, there is a need of recommender technique which can generate more efficient and accurate recommendations under cold start problem. In this thesis, an accurate hybrid recommender technique is proposed. The proposed technique is based on the concept of word embedding to produce the distributed representation of items description along with natural language processing technique to get higher representation for an item. The content embedding are incorporated in memory based collaborative filtering technique. The proposed system is capable to generate more accurate and efficient recommendations under cold start item problem. The content embedding based feedforward neural network technique is also proposed in this thesis. The technique is built on the model based approach which utilizes the neural network. The neural network is used as a profile learner which identifies the hidden trends and patterns to build a model based on auxiliary information along with user ratings and it predict the rating for cold start items more accurately. The experimental results and quantitative comparison with other state-of the-art techniques is provided which demonstrate the importance of proposed techniques.