Logo image
基於非負矩陣分解作常識聯想式學習和特徵選取
Thesis

基於非負矩陣分解作常識聯想式學習和特徵選取

陳姿君
Masters, 國立清華大學, 資訊系統與應用研究所
2015

Abstract

特徵選取 非負矩陣分解 常識聯想學習 feature selection non-negative matrix factorization common sense association learning
Nowadays, more and more news and reports about a robot or a computer will appear in our life. It seems that the information technology is really close to us. However, even now we have the good development of information technology, It is still difficult for computers to learn the common sense possessed by mainkind via things encountered in the daily life as a human. Besides, it is non-trivial to formulate all common sense in terms of definite rules. There are many models of common sense reasoning, and the association reasoning is a fundamental ability of common sense. We focus on the learning on the common sense association reasoning. We develop a computational model to learn the common sense association between a pair of concept classes based on a bipartite network and matrix factorization methods. In this model, we view the concept-pair association as a bipartite network so that the auto-association 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 667 location-activity pairs and 625 goal-action pairs from ConceptNet [12] as our training data and test data. We evaluate the performance in terms of F-factor, precision and recall using a common sense association problem between locations and activities against six feature selection strategies in the matrix factorization optimization. We reach performance of precision up to 77.2%, recall up to 31.2% and F-score up to 42.3% in the location-activity association domain. In the goal-activity association domain, we reach performance of precision up to 93.5%, recall up to 25.3%, and F-score up to 39.5%. For understanding the feature selection effects on the association learning, we assumed six feature selection methods including random, most-link (rank), entropy, singular value decomposition (SVD), H-frequency, and H-weight. We found that all performances of recall and F-score in the five feature selection methods are better than the performance of recall and F-score in the random method.

Metrics

1 Record Views

Details

Logo image