This study focuses on relational data obtained through object relations. Traditional analysis of relational data often ignores attribute information. Therefore, Mikawa et al. proposed a method to estimate the latent structure of continuous relational data using a generative model and parameter estimation. However, real-world relational data can be discrete, and therefore, we propose a new model for binary relational data using a generative model based on the Bernoulli distribution and the Monte Carlo Expectation-Maximization (EM) algorithm for parameter estimation. We also clarify the effectiveness of the proposed model through simulation experiments using artificial data and real data.

抄録全体を表示

Recent work on temporal relation identification has focused on three types of relations between events: temporal relations between an event and a time expression, between a pair of events and between an event and the document creation time. These types of relations have mostly been identified in isolation by event pairwise comparison. However, this approach neglects logical constraints between temporal relations of different types that we believe to be helpful. We therefore propose a Markov Logic model that jointly identifies relations of all three relation types simultaneously. By evaluating our model on the TempEval data we show that this approach leads to about 2% higher accuracy for all three types of relations ---and to the best results for the task when compared to those of other machine learning based systems.

抄録全体を表示

Clustering word co-occurrences has been studied to discover clusters as latent concepts. Previous work has applied the semantic aggregate model (SAM), and reports that discovered clusters seem semantically significant. The SAM assumes a co-occurrence arises from one latent concept. This assumption seems moderately natural. However, to analyze latent concepts more deeply, the assumption may be too restrictive. We propose to make clusters for each part of speech from co-occurrence data. For example, we make adjective clusters and noun clusters from adjective--noun co-occurrences while the SAM builds clusters of ``co-occurrences.'' The proposed approach allows us to analyze adjectives and nouns independently.

抄録全体を表示

テレビのユーザーインターフェースが視聴者の好みに基づいて番組を推薦出来れば良い。それを実現することを目指して、テレビ内蔵のカメラが視聴者を撮影し、映像の解析により視聴者を認識して、視聴中の番組やシーンにどれ程の興味があるかを推定する興味度推定システムを試作した。システムの仕組みと動作を紹介し、ユーザー実験の結果について述べる。

抄録全体を表示

Clustering word co-occurrences has been studied to discover clusters as latentconcepts. Previous work has applied the semantic aggregate model (SAM), and reports that discovered clusters seem semantically significant. The SAM assumes a co-occurrence arises from one latent concept. This assumption seems moderately natural. However, to analyze latent concepts more deeply, the assumption may be too restrictive. We propose to make clusters for each part of speech from co-occurrence data. For example, we make adjective clusters and noun clusters from adjective—noun co-occurrences while the SAM builds clusters of “co-occurrences.” The proposed approach allows us to analyze adjectives and nouns independently.
To take this approach, we propose a frequency-based infinite relational model (FIRM) for word co-occurrences. The FIRM is a stochastic block model that takes into account the frequency of observations although traditional stochastic blockmodels ignore it. The FIRM also utilizes the Dirichlet process so that the number of clusters is inferred. We derive a variational inference algorithm for the model to apply to a large dataset. Experimental results show that the FIRM is more helpful to analyze adjectives and nouns independently, and the FIRM clusters capture the SAM clusters better than a stochastic blockmodel.

抄録全体を表示

In this paper, we propose a new Estimation of Distribution Algorithms which can cope with Reinforcement Learning Problems. Basic procedure of the EDAs is that 1) select better individuals, 2) estimate probabilistic models, and 3) sample new individuals. In the proposed method, instead of use of individuals, input-output records in episodes are directory treated. Moreover the estimated probabilistic model is regarded as policy so that new input-output records are generated by the interaction between the policy and environments. Computer simulations confirm the effectiveness of the proposed method.

抄録全体を表示

Link prediction is the task of inferring the existence or absence of certain relationships among data objects such as identity, interaction, and collaboration. Link prediction is found in various applications in the fields of information integration, recommender systems, bioinformatics, and social network analysis. The increasing interest in dynamically changing networks has led to growing interest in a more general link prediction problem called temporal link prediction in the data mining and machine learning communities. However, only links among nodes at the same time point are considered in temporal link prediction. We propose a new link prediction problem called cross-temporal link prediction in which the links among nodes at different time points are inferred. A typical example of cross-temporal link prediction is cross-temporal entity resolution to determine the identity of real entities represented by data objects observed in different time periods. In dynamic environments, the features of data change over time, making it difficult to identify cross-temporal links by directly comparing observed data. Other examples of cross-temporal links are asynchronous communications in social networks such as Facebook and Twitter, where a message is posted in reply to a previous message. We adopt a dimension reduction approach to cross-temporal link prediction; that is, data objects in different time frames are mapped into a common low-dimensional latent feature space, and the links are identified on the basis of the distance between the data objects. The proposed method uses different low-dimensional feature projections in different time frames, enabling it to adapt to changes in the latent features over time. Using multi-task learning, it jointly learns a set of feature projection matrices from the training data, given the assumption of temporal smoothness of the projections. The optimal solutions are obtained by solving a single generalized eigenvalue problem. Experiments using a real-world set of bibliographic data for cross-temporal entity resolution and a real-world set of emails for unobserved asynchronous communication inference showed that introducing time-dependent feature projections improved the accuracy of link prediction.

抄録全体を表示

In this paper, a power-regularization-based fuzzy clustering method is proposed for spherical data. Power regularization has not been previously applied to fuzzy clustering for spherical data. The proposed method is transformed to the conventional fuzzy clustering method, entropy-regularized fuzzy clustering for spherical data (eFCS), for a specified fuzzification parameter value. Numerical experiments on two artificial datasets reveal the properties of the proposed method. Furthermore, numerical experiments on four real datasets indicate that this method is more accurate than the conventional fuzzy clustering methods: standard fuzzy clustering for spherical data (sFCS) and eFCS.

抄録全体を表示

The interpretability of fuzzy co-cluster partitions were shown to be improved by introducing exclusive penalties on both object and item memberships although the conventional fuzzy co-clustering adopted exclusive natures only on object memberships. In real applications, however, fully exclusive constraints may bring inappropriate influences to some items, and partially exclusive penalties should be forced reflecting the characteristics of each item. For example, in customer-product analysis, the degree of popularity of each product may be a measure of compatibility in multiple customer groups, and exclusive penalties should be forced only to some specific products. In this paper, the conventional exclusive constraint model is further modified by forcing exclusive penalties only to some selected items, and the effects of partially exclusive partition are demonstrated from the view points of not only partition quality but also collaborative filtering applicability. In a document-keyword analysis experiment, word class is shown to be useful for exclusively selecting keywords so that the interpretability of document cluster is improved. In a collaborative filtering experiment, the recommendation capability is demonstrated to be improved by considering intrinsic differences of popularity of each product.

抄録全体を表示

Estimation of Distribution Algorithms (EDAs) are a promising evolutionary computation method. Due to the use of probabilistic models, EDAs can outperform conventional evolutionary computation. In this paper, EDAs are extended to solve reinforcement learning problems which are a framework for autonomous agents. In the reinforcement learning problems, we have to find out better policy of agents such that it yields a large amount of reward for the agents in the future. In general, such policy can be represented by conditional probabilities of agents' actions, given the perceptual inputs. In order to estimate such a conditional probability distribution, Conditional Random Fields (CRFs) by Lafferty (2001) are introduced into EDAs. The reason why CRFs are adopted is that CRFs are able to learn conditional probabilistic distributions from a large amount of input-output data, i.e., episodes in the case of reinforcement learning problems. Computer simulations on Probabilistic Transition Problems and Perceptual Aliasing Maze Problems show the effectiveness of EDA-RL.

抄録全体を表示

Spoken language understanding (SLU) aims to map a user's speech into a semantic frame. Since most of the previous works use the semantic structures for SLU, we verify that the structure is useful even for noisy input. We apply a structured prediction method to SLU problem and compare it to an unstructured one. In addition, we present a combined method to embed long-distance dependency between entities in a cascaded manner. On air travel data, we show that our approach improves performance over baseline models.

抄録全体を表示

In the field of cluster analysis, fuzzy theory including the concept of fuzzy sets has been actively utilized to realize flexible and robust clustering methods. Fuzzy C-means (FCM), which is the most representative fuzzy clustering method, has been extended to achieve more robust clustering. For example, noise FCM (NFCM) performs noise rejection by introducing a noise cluster that absorbs noise objects and possibilistic C-means (PCM) performs the independent extraction of possibilistic clusters by introducing cluster-wise noise clusters. Similarly, in the field of co-clustering, fuzzy co-clustering induced by multinomial mixture models (FCCMM) was proposed and extended to noise FCCMM (NFCCMM) in an analogous fashion to the NFCM. Ubukata et al. have proposed noise clustering-based possibilistic co-clustering induced by multinomial mixture models (NPCCMM) in an analogous fashion to the PCM. In this study, we develop an NPCCMM scheme considering variable cluster volumes and the fuzziness degree of item memberships to investigate the specific aspects of fuzzy nature rather than probabilistic nature in co-clustering tasks. We investigated the characteristics of the proposed NPCCMM by applying it to an artificial data set and conducted document clustering experiments using real-life data sets. As a result, we found that the proposed method can derive more flexible possibilistic partitions than the probabilistic model by adjusting the fuzziness degrees of object and item memberships. The document clustering experiments also indicated the effectiveness of tuning the fuzziness degree of object and item memberships, and the optimization of cluster volumes to improve classification performance.

抄録全体を表示

Abduction is desirable for many natural language processing (NLP) tasks. While recent advances in large-scale knowledge acquisition warrant applying abduction with large knowledge bases to real-life NLP problems, as of yet, no existing approach to abduction has achieved the efficiency necessary to be a practical solution for large-scale reasoning on real-life problems. In this paper, we propose an efficient solution for large-scale abduction. The contributions of our study are as follows: (i) we propose an efficient method of cost-based abduction in first-order predicate logic that avoids computationally expensive grounding procedures; (ii) we formulate the best-explanation search problem as an integer linear programming optimization problem, making our approach extensible; (iii) we show how cutting plane inference, which is an iterative optimization strategy developed in operations research, can be applied to make abduction in first-order logic tractable; and (iv) the abductive inference engine presented in this paper is made publicly available.

抄録全体を表示

We introduce a new approach to the problem of link prediction for network structured domains, such as the Web, social networks, and biological networks. Our approach is based on the topological features of network structures, not on the node features. We present a novel parameterized probabilistic model of network evolution and derive an efficient incremental learning algorithm for such models, which is then used to predict links among the nodes. We show some promising experimental results using biological network data sets.

抄録全体を表示

The research field of human like agents that are often represented by an animation character is becoming increasingly active in recent years. As the motion of such agents influences the users' impression, it is easy to expect that the ability of the human like agent to make appropriate gestures could improve the understandability of the utterance contents. The load of the content creator, however, increases if he/she needs to determine when and what gestures the agent should make. This paper attempts to estimate the appropriate gestures for a given utterance text using conditional random fields (CRF), which can be used to reduce the effort spent by contents creators. We create the dataset consisting of the utterance text and the corresponding gesture labels from the educational movie contents and construct a gesture-labeling model using CRF in a supervised learning manner. The estimation performance of appearing the gestures is evaluated and compared with the simple existing model. Especially, we focus on the metaphoric gesture, often representing an abstract concept. This is because the metaphoric gesture is expected to facilitate the users' understanding of the abstract concepts. We empirically confirmed that the proposed model can distinctly estimate the metaphoric and other gestures.

抄録全体を表示

本研究では、帰納論理プログラミングを用いて、ネットワークデータからルールを獲得することを目標とする。具体的には、関係が複数ある場合の分類規則の抽出や、動的に変化するネットワークからのルールの獲得を行う。これら、既存のネットワーク分析ではあまり扱うことのできなかったデータやルールの獲得について比較を通じて検証する。

抄録全体を表示

This paper presents a method for categorizing named entities in Wikipedia. In Wikipedia, an anchor text is glossed in a linked HTML text. We formalize named entity categorization as a task of categorizing anchor texts with linked HTML texts which glosses a named entity. Using this representation, we introduce a graph structure in which anchor texts are regarded as nodes. In order to incorporate HTML structure on the graph, three types of cliques are defined based on the HTML tree structure. We propose a method with Conditional Random Fields (CRFs) to categorize the nodes on the graph. Since the defined graph may include cycles, the exact inference of CRFs is computationally expensive. We introduce an approximate inference method using Tree-based Reparameterization (TRP) to reduce computational cost. In experiments, our proposed model obtained significant improvements compare to baseline models that use Support Vector Machines.

抄録全体を表示

Recently, the research area of mining in structured data has been actively studied. However, since most techniques for structured data mining so far specialize in mining from single structured data, it is difficult for these techniques to handle more realistic data which is related to various types of attribute and which consists of plural kinds of structured data. Since such kind of data is expected to be going to rapidly increase, we need to establish a flexible and highly accurate technique that can inclusively treat such kind of data. In this paper, as one of the techniques to deal with such kind of data, we propose data mining algorithms of mining classification rules in multidimensional structured data. First, an algorithm with two pruning capabilities of mining correlated patterns is introduced. Then, top-k multidimensional correlated patterns are discovered by using this algorithm repeatedly in the fashion like a beam search. We also show the algorithms for constructing classifiers based on the discovered patterns. Experiments with real world data were conducted to assess the effectiveness of the proposed algorithms. The results show that the proposed algorithms can construct comprehensible and accurate classifiers within a reasonable running time.

抄録全体を表示

As semi-structured data is used widely in several fields, the importance of structured data mining is increasing recently. Although mining frequent patterns in structured data is one of the most fundamental tasks, frequent pattern miners often discover huge number of patterns. To overcome this problem, two major approaches, condensed representation mining and constraint-based mining, have been proposed. In this paper, as a technique for integrating these two approaches, we propose three algorithms, RC_LOCOT, posC_LOCOT, and negC_LOCOT, for discovering closed ordered subtrees under anti-monotone constraints about the structure of patterns to be discovered. The proposed algorithms discover closed constrained subtrees efficiently not by post-processing but by pruning and skipping the search space based on the occurrence matching and the patterns on the border.

抄録全体を表示