Seiki Ubukata

A semi-supervised fuzzy co-clustering framework and application to twitter data analysis

Semi-supervised clustering is an efficient scheme for utilizing data with partial class information, where unsupervised data distributions are estimated under some supports of partial supervised class information. In this paper, a novel framework for performing fuzzy co-clustering of cooccurrence information with partial supervision is proposed, which is induced by multinomial mixture concept. Co-clustering is useful for extracting object-item pair-wise clusters from cooccurrence information and has been utilized in various applications such as document-keyword analysis and customer-products purchase history data analysis. Several experimental results including a twitter data analysis demonstrate the ability of improving the classification quality of the fuzzified co-cluster structural knowledge. Then, the proposed semi-supervised framework is expected to be a powerful tool in Big Data analysis with huge volumes of data but partial supervisions only.

Imprecise Rules for Data Privacy

When rules are induced, some rules can be supported only by a very small number of objects. Such rules often correspond to special cases so that supporting objects may be easily estimated. If the rules with small support include some sensitive data, this estimation of objects is not very good in the sense of data privacy. Considering this fact, we investigate utilization of imprecise rules for privacy protection in rule induction. Imprecise rules are rules classifying objects only into a set of possible classes. Utilizing imprecise rules, we propose an algorithm to induce k-anonymous rules, rules with k or more supporting objects. We demonstrate that the accuracy of the classifier with rules induced by the proposed algorithm is not worse than that of the classifier with rules induced by the conventional method. Moreover, the advantage of the proposed method with imprecise rules is examined by comparing other conceivable method with precise rules.

A Noise Fuzzy Co-Clustering Scheme in MMMs-Induced Clustering

Noise fuzzy clustering is a practical model for handling noisy data sets in FCM clustering, where an additional noise cluster is introduced for dumping noise objects into it. Because the noise cluster is assumed to have an equal (fixed) distance from all objects, noise objects having larger distances from all clusters can be assigned to the noise cluster. In this paper, a novel scheme for implementing noise clustering in fuzzy coclustering is proposed, where noise cluster is defined in a slightly different manner from the FCM-type one because co-clustering is a prototype-less clustering. A noise co-cluster is defined with homogeneous item memberships for drawing noise objects, whose cooccurrence features are dissimilar from all general clusters. The characteristics of the proposed scheme are demonstrated in numerical experiments.

The Rough Set k-Means Clustering

In the field of clustering, soft computing approaches which deal with vague cluster memberships are effective. Clustering based on rough set theory is considered to be a promising approach as a way to represent vague cluster memberships as well as fuzzy clustering. In this paper, we propose the Rough Set k-Means (RSKM) clustering which is based on rough sets. The RSKM clustering is a type of the k-means clustering based on rough approximations and it enables certain clustering detecting the boundary and positive regions of temporal clusters. We carried out some numerical experiments and confirmed the performance and characteristics of the proposed method.

A fuzzy co-clustering interpretation of probabilistic latent semantic analysis

Several fuzzy clustering models were proposed by extending intrinsic fuzzy partition mechanisms of probabilistic mixture models and have been shown to have ability of improving the partition quality and interpretability of probabilistic partitions. In this paper, a novel fuzzy clustering interpretation of probabilistic latent semantic analysis (pLSA) is discussed and a fuzzy co-clustering model is proposed by introducing adjustable fuzzification penalty to the pseudo-log-likelihood function of pLSA. Several numerical experiments demonstrate the advantage of tuning the intrinsic fuzziness of pLSA likelihood function.

MMMs-Induced Fuzzy Co-clustering with Exclusive Partition Penalty on Selected Items

Fuzzy co-clustering is a powerful tool for summarizing co-occurrence information while some intrinsic knowledge on meaningful items may be concealed by the dominant items shared by multiple clusters. In this paper, the conventional fully exclusive item partition model is modified such that exclusive penalties are forced only on some selected items. Its advantages are demonstrated through two numerical experiments. In a document clustering task, the proposed model is utilized for emphasizing cluster-wise meaningful keywords, which are useful for effectively summarizing document clusters. In an unsupervised classification task, the classification quality is improved by efficiently selecting promising items based on the item-wise single penalization test.

An Agent Control Method Based on Variable Neighborhoods

In this paper, we propose a model that an agent selects actions based on variable neighborhoods. We formulate relationships among variable neighborhoods, the agents observations, and the agents behaviors in a framework of rough set theory and topological spaces. The main task is to explore a method by which we can select sizes of neighborhoods under given contexts. We also show simulation results of the proposed method.

Cluster Validation in Multinomial Mixtures-Induced Fuzzy Co-Clustering

Cluster validation is an important process in FCM-type clustering, where the optimal cluster partition should be selected from candidate solutions derived with various initialization and cluster numbers. In the standard FCM clustering, some validity measures such as partition quality-based or geometric features-based ones have been proposed. In this paper, fuzzy cluster validation is discussed in the multinomial mixtures-induced fuzzy co-clustering context. Partition quality-based indices are directly applied to the prototype-less co-clustering model with a brief modification. The concept of a geometric-based index is achieved supported by the object-item aggregation measure. The characteristics of the proposed indices are demonstrated in numerical experiments.

Fuzzy DA Clustering-Based Improvement of Probabilistic Latent Semantic Analysis

Probabilistic latent semantic analysis (pLSA) can be interpreted as a soft co-clustering model with an intrinsic fuzzification penalty and the partition quality was shown to be improved by tuning the degree of intrinsic partition fuzziness while the model is not supported by probabilistic constraints. In this paper, the mechanism of intrinsic fuzziness tuning is utilized for improving the partition quality of pLSA under the strict probabilistic constraints. The proposed deterministic annealing approach first initializes a co-cluster partition with a slightly fuzzier penalty weight and then gradually reduces the intrinsic fuzziness until it reaches the strict probabilistic constraints. Supported by the robust feature of fuzzier models against random initialization, the derived pLSA partition is demonstrated to be more stable in several numerical experiments.

A Semi-Supervised Framework for MMMs-Induced Fuzzy Co-Clustering with Virtual Samples

Although the goal of clustering is to reveal structural information from unlabeled datasets, in cases with partial structural supervisions, semi-supervised clustering is expected to improve partition quality. However, in many real applications, it may cause additional costs to provide an enough amount of supervised objects with class labels. A virtual sample approach is a practical technique for improving classification quality in semi-supervised learning, in which additional virtual samples are generated from supervised objects. In this research, the virtual sample approach is adopted in semi-supervised fuzzy co-clustering, where the goal is to reveal object-item pairwise cluster structures from cooccurrence information among them. Several experimental results demonstrate the characteristics of the proposed approach.