This paper proposes an automatic construction method of a concept network based on neural network. The proposed method constructs a network by connecting weights between a word (word) and a concept (synset) and between a concept and a concept that are described in Japanese WordNet. Two construction methods are proposed: one uses a collocation dictionary, the other uses corpus. The ow of both methods is as follows: (1) sentence analysis is performed (according to necessity) and a set of words that are considered to have semantic relevance is obtained. (2)candidates of concept network based on the set of words are generated. (3) the concept network is modified. (4) cost of the concept network is calculated and it is optimized. Calculation of the relation of words and concepts is possible by generating words and a concept vector based on Japanese Wikipedia text corpus in advance. We implemented a concept network to show the feasibility of the proposed method. According to experiments, it has shown that the subjective accuracy was about 48.7%.

In this paper, we propose a moral judgment system. Moral judgment means right and wrong judgment. The proposed system consists of the learning phase and the moral judgment phase. In the learning phase, the positive words and the negative words (evaluation expressions) are extracted based on the co-occurrence frequency of the learning data and the words in the polarity dictionary. Then, the words having high score are extracted. They are used to calculate the score. In the moral judgment phase, first, the determination of whether the input sentence relates to morality or not is performed. Second, if the input is determined to relate to morality, the scoring based on the co-occurrence frequency of the input and the important evaluation expressions is carried out. At this point, when the number of words in the input sentence is large, moral judgment is difficult. In such a case, such sentences are simplified based on the TF-IDF method. In the experiments, we compared the proposed system with judgments by humans. As a result, the effectiveness of the proposed system is confirmed.

Recently, anomaly detection attracts much attention. In this paper, to verify the validity of various exception criteria based on pattern mining techniques for social media data, we conduct preliminary experiments of mining outliers from three real world tweet datasets. The results show significant difference between pattern-based and rule-based measures.

While user population of SNS is increasing recently, SNS risks such as flaming or leak of personal information have become social issue. In this research, we investigated the process of SNS flaming or information leaking, and we implemented some prototype systems in order to management such risks by oneself. This paper reports our approach and outcomes of the investigation of SNS risks. Additionally, we introduce the SNS monitoring and risk management systems.

We address a problem of identifying high centrality nodes in a large social network based on approximated centrality values derived from a small portion of nodes sampled uniformly at random from the whole set. To this end, we apply our resampling-based framework to estimate the approximation error, and detect gaps between nodes with a given confidence level. Here, a gap means a clear difference between two nodes in terms of a centrality measure, and gap detection means, given two nodes, determining which node has a greater centrality value than the other with a given confidence level. On two real world social networks, we empirically show that the proposed method can successfully detect more gaps only from several tens of percent of the node, compared to the one adopting a standard error estimation framework, and that the resulting gaps enable us to correctly identify a set of nodes having a high centrality value.

In this paper, we attempt to detect change points of a dynamic network structure. We focus on the nodes functions in a network and define the nodes function as the convergence curve of the PageRank score. For each node, we calculate the correlation coeffcients between the convergence curves in adjacent two snapshots of a time-varying network. Then, we propose the average of correlation coeffcients of all nodes as a measure of the change point of a network strucuture and refer to this measure as average similarity. Especially, when the average similarity shows the lower value, we assume that the network structure changes significantly. In our experiments using synthetic and real networks with artificial changes, we evaluate the eectiveness of our proposed measure.

In this paper, we focus on rambling activity, for example, visiting tourist spots or barhopping on shopping districts, and propose a user behavior model for predicting rambling activity. Moreover we evaluate availability of the model. Specifically, we adopt Levy flights as basic probability model, and extend the model by introducing 2 parameters depending on distance between spots and popularity of spots. In our experiments using data from 'The 6th Shizuoka Omachi Bar,' which is a barhopping event, we demonstrate suitability of the user behavior model for predicting rambling activity with comparing results of simulations varied distance and popularity parameters.

For clinical decision support systems designed to help physicians to make diagnostic decisions, "disease similarity" data is highly valuable in that they realize continuous presentation of diagnostic candidates. Toward such a recommendation system, calculation of disease similarity between diseases is a key component, and thus, this paper explores the method to measure the similarity on a simplified disease knowledge base. Our disease knowledge base comprises disease master data, symptom master data, and disease-symptom relations that include clinical information of 1550 disorders. The calculation of the disease similarity is performed on this knowledge base, with i) disease classification, ii) probabilistic calculation, and iii) machine learning, and the results are evaluated with a gold standard list audited by a physician. A comparative study revealed that the machine learning approach outperforms the others. The result suggests that even a superficial calculation on a simplified knowledge base can satisfy the clinical needs in this problem domain.