Parallel Higher-order Svd for Tag-recommendations

Social Tagging has become a method of choice for enriching data (like pictures) with meta-data which in turn can be used for searching (like retrieving art pictures) or tag recommendations relying on Singular Value Decompositions (SVD) to reduce dimensionality. We observed that social tagging-based search or tag-recommendation is more successful, if several dimensions describing the tagging community (like age, interests, cultural background) are considered. In other words, social tagging-based search calls for higher-order, or tensor-based, SVD instead of standard, matrix-based, SVD. This article reports on a parallelization of higher-order SVD which achieves a significantly better time complexity by comparison to single processor approaches. This time complexity is possible thanks to a careful use of Hestenes' SVD. This article also reports on a first experimental evaluation of the approach for a tag recommendation system.