An Adaptive Spreading Activation Scheme for Performing More Effective Collaborative Recommendation

While Spread Activation has shown its effectiveness in solving the problem of cold start and sparsity in collaborative recommendation, it will suffer a decay of performance (over activation) as the dataset grows denser. In this paper, we first introduce the concepts of Rating Similarity Matrix (RSM) and Rating Similarity Aggregation (RSA), based on which we then extend the existing spreading activation scheme to deal with both the binary (transaction) and the numeric ratings. After that, an iterative algorithm is proposed to learn RSM parameters from the observed ratings, which makes it automatically adaptive to the user similarity shown through their ratings on different items. Thus the similarity calculations tend to be more reasonable and effective. Finally, we test our method on the EachMovie dataset, the most typical benchmark for collaborative recommendation and show that our method succeeds in relieving the effect of over activation and outperforms the existing algorithms on both the sparse and dense dataset.