Evolutionary algorithm based feature extraction for enhanced recommendations

A major challenge to Collaborative Filtering systems is high dimensional and sparse data which they have to deal with. Feature selection techniques partly address this problem by reducing the feature space and retaining only a representative subset of features. However these techniques do not address the sparsity problem which affects both quality and quantity of recommendations. A more promising direction would be to construct/extract new features which are low dimensional, dense and have more discriminative power. Content based construction of features has been explored in the past. This work proposes a evolutionary algorithm based feature extraction techniques which discover hidden features with high discriminative capacity. Such an approach offers the advantage of discovering features even in the absence of additional information such as item contents etc. The proposed approach is contrasted with content based feature extraction techniques through experiments and the ability of the new approach in discovering interesting and useful features is established.