An Efficient Cross-Domain Recommendation Technique in Cold-Start Situations

Most of the recent studies on recommender systems are focused on single domain recommendation systems. In the single domain recommendation systems, the items that are used for training and test data set are belongs to within the same domain. Cross-site domains or item recommendations in multi-domain environment are available in Amazon i.e. it incorporate two or more domains. Few research studies are done on the cross-site recommendation systems. Cross-site recommendations provide the relationship between the two sets of items from various domains. They can provide the extra information about the users of a target domain and recommendations will be done based on that. In this paper, we will study cross-site recommendation model on the cold start situation, where the purchase history is not available for the new user. Cold-start is the well-known issue in the area of recommendation systems. It seriously affect the recommendations in the collaborative filtering approaches. In this paper, we propose a new solution to recommend products from e-commerce websites to users at social networking sites. a noteworthy issue is how to leverage knowledge from social networking websites when there is no purchase history for a customer especially in cold start situations.in particular we proposed the solution for cold start recommendation by linking the users across social networking sites and e-commerce websites i.e. customers who have social network identities and have purchased on e-commerce websites as a bridge to map user’s social networking features in to another feature representation which can be easier for product recommendation. Here we propose to learn by using recurrent neural networks both user’s and product’s feature representations called user embedding and product embedding from the data collected from e-commerce website and then apply a modified gradient boosting trees method to transform user’s social networking features in to user embedding. Once found, then develop a feature-based matrix factorization approach which can leverage the learnt user embedding for the cold-start product recommendation. Experimental results shows that our approach effectively works and gives the best recommended results in cold start situations