Search

We propose a novel approach for constructing statistical preference models for context-aware recommender systems. To do so, one of the most important but difficult problems is acquiring sufficient training data in various contexts/situations. Particularly, some situations require a heavy workload to set them up or to collect subjects under those situations. To avoid this, often a large amount of data in a supposed situation is collected, i.e., a situation where the subject pretends/imagines that he/she is in a specific situation. Although there may be difference between the preference in the real situation and the supposed situation, this has not been considered in existing researches. Here, to study the difference, we collected a certain amount of corresponding data. We asked subjects the same question about preference both in the real and the supposed situation. Then we proposed a new model construction method using a difference model constructed from the correspondence data and showed the effectiveness through the experiments. [ABSTRACT FROM AUTHOR]

In order to achieve highly accurate content-based music information retrieval (MIR), it is necessary to compensate the various bit rates of encoded songs which are stored in the music collection, since the bit rate differences are expected to apply a negative effect to content-based MIR results. In this paper, we examine how the bit rate differences affect MIR results, propose methods to normalize MFCC features extracted from encoded files with various bit rates, and show their effects to stabilize MIR results. [ABSTRACT FROM AUTHOR]

In this paper, we propose an algorithm for reducing the number of unknown words on blog documents by replacing peculiar expressions with formal expressions. Japanese blog documents contain many peculiar expressions regarded as unknown sequences by morphological analyzers. Reducing these unknown sequences improves the accuracy of morphological analysis for blog documents. Manual registration of peculiar expressions to the morphological dictionaries is a conventional solution, which is costly and requires specialized knowledge. In our algorithm, substitution candidates of peculiar expressions are automatically retrieved from formally written documents such as newspapers and stored as substitution rules. For the correct replacement, a substitution rule is selected based on three criteria; its appearance frequency in retrieval process, the edit distance between substituted sequences and the original text, and the estimated accuracy improvements of word segmentation after the substitution. Experimental results show our algorithm reduces the number of unknown words by 30.3%, maintaining the same segmentation accuracy as the conventional methods, which is twice the reduction rate of the conventional methods. [ABSTRACT FROM AUTHOR]

In support vector learning, computational complexity of testing phase scales linearly with number of support vectors (SVs) included in the solution – support vector machine (SVM). Among different approaches, reduced set methods speed-up the testing phase by replacing original SVM with a simplified one that consists of smaller number of SVs, called reduced vectors (RV). In this paper we introduce an extension of the bottom-up method for binary-class SVMs to multi-class SVMs. The extension includes: calculations for optimally combining two multi-weighted SVs, selection heuristic for choosing a good pair of SVs for replacing them with a newly created vector, and algorithm for reducing the number of SVs included in a SVM classifier. We show that our method possesses key advantages over others in terms of applicability, efficiency and stability. In constructing RVs, it requires finding a single maximum point of a one-variable function. Experimental results on public datasets show that simplified SVMs can run faster original SVMs up to 100 times with almost no change in predictive accuracy. [ABSTRACT FROM AUTHOR]