Feature Selection in Learning Common Sense Associations Using Matrix Factorization

We propose a computational model to learn the common sense association between a pair of concept classes based on a bipartite network and matrix factorization methods. We view the concept-pair association as a bipartite network so that the autoassociation mappings can become similarity constraints. We impose the additional similarity and regularity constraints on the optimization objectives so that a mapping matrix can be found in the matrix factorization to best fit the observation data. We extract 139 locations and 436 activities and 667 location-activity pairs from ConceptNet (http://conceptnet5.media.mit.edu/). We evaluate the performance in terms of F-factor, precision, and recall using a common sense association problem between locations and activities against four feature selection strategies in the matrix factorization optimization. The comparison between the performances with and without human judgment reveals that matrix factorization method tends to show good generalization even under very little observation evidence. Among the four feature selection methods, the maximal entropy method performs better in terms of F-score and recall when the feature number is more than 30 % while SVD method performs better in terms of F-score and recall when the feature number is less than 30 %. Random selection can have a higher precision given "enough'' features, but it tends to be the worst performer in the recall and F-score.