Collaborative recommendation algorithm based on MI clustering

In a personalized recommendation system, the context feature of item is an important factor affecting recommendation accuracy, but traditional collaborative recommendation algorithms cannot take context feature of item into account effectively. To solve this problem we constructed a user-item ratings information representation model with multiple instance learning(MIL) based on traditional user-item ratings information, considering the context features of an item. Exploiting a characteristic trait of MIL, its strong tolerance to fault, a collaborative recommendation algorithm based on MI clustering was designed which computes users' nearest neighbors by MI clustering and predicates user ratings according to the neighbors. Experimental results confirmed that collaborative recommendation algorithm based on MI clustering improved accuracy of predictions and alleviated the problem of sparse data effectively.