Abstract
We consider techniques based on artificial neural networks for combining collaborative (social) and content information in a recommender system in order to enhance recommendation performance. We find that the recommendation quality achieved by a feedforward multilayer perceptron network operating on combined collaborative and content-based information (preprocessed using the singular value decomposition) is statistically significantly better than that of a network that is provided with the collaborative data alone, assuming that dimensionality reduction is performed on the collaborative and content-based data components separately. We propose a mixture of attribute experts neural network architecture that exploits the natural division between content and social information in order to reduce the number of network connections, resulting in more efficient training and recommendation than a standard fully connected network. We characterize the set of functions that can be expressed by mixture of attribute experts networks. The top 3 precision achieved by a recommender system based on our mixture of attribute experts architecture is superior to that of a purely collaborative system at a strong statistical significance level (P< 0.01). A random restarting technique reduces the average running time without affecting recommendation precision.
CR Categories and Subject Descriptors. H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval – Information Filtering; H.2.8 [Database Management]: Database Applications – Data Mining; I.2.6 [Artificial Intelligence]: Learning – Connectionism and neural nets; I.5.1 [Pattern Recognition]: Models – Neural nets; I.5.5 [Pattern Recognition]: Implementation – Special architectures.