Improving collaborative recommendations using vector quantization and clustering

The Internet provides its users with a wide variety of resources. However, the huge volume of data makes finding relevant data a difficult task for the users. By using item filtering algorithms, the recommendation systems can effectively offer several items similar to what users search for. However, most of the existing algorithms do not consider the impact of item features in calculating the similarity between users. To overcome this problem, a new method based on vector quantization and clustering is presented here. The proposed method uses features of the items as weighting factors to calculate a vector of user rates instead of a single value for rating a particular item. Then, the vector is normalized and the user rates are employed to establish a user-item matrix in which each element of a row vector indicates the interestingness of the given feature for the selected user. Using this matrix, the users are grouped into a number of clusters and their preferred items are identified by calculating the mean value of their rating vectors. The result of the performance evaluation experiments shows that the suggested method provides a remarkable improvement in handling “cold-start” problem, lower time complexity and reasonable accuracy of recommendations.