Social Media continues to evolve as an important place of people's communication. Cookingrecipe sharing websites have recently emerged in Social Media, and help to enrich food culture in people's daily lives. In this paper, when a cooking-recipe is posted to a cooking-recipe sharing website, we consider a problem of predicting which users present positive messages about it in the future. In order to solve this problem, we propose a method of exploiting the data of cooking-recipes that those users have posted or highly evaluated in the past. Using real-world data, we experimentally demonstrate the effectiveness of the proposed method.

Acquiring networks of trust relations among users in social media sites such as item-review sites is important for analyzing users' behaviour and efficiently finding reliable information on the Web. For an item-review site, we address the problem of predicting trust-links among users. Recently, non-negative matrix factorization (NMF) methods have been shown to be useful for trust-link predition in such an site, where both the link and activity information is employed. Here, a user activity in an item-review site means posting a review and giving a rating for an item. Aiming to improve NMF methods for trust-link prediction, in this paper, we propose such an NMF method that incorporates information of people's evaluations for users' activities as well as information of trust-links and users' activities. Also, we apply it to an analysis of users' behaviour. Using real data of an item-review site, we experimentally demonstrate the effectiveness of the proposed method.

We adopt pKTN, the Petri-net-based Knowledge Transitive Network, which is an AI approach representing first-order predicate logic (FOL) in the form of a network, and present a successful algorithm for deductive inference. During deduction, the pKTN is expanded to a network with AND/OR-tree structure, in which we need to solve an 'ill-posed' problem, finding a symbolically consistent subtree out of a given redundant tree. For this problem, a token-based algorithm named ELISE (ELiminating Inconsistency by SElection) was proposed, but existing ELISE was only applicable to non-redundant (well-posed) problems. In this paper, we seriously revise the algorithm of ELISE, make it applicable to redundant (ill-posed) problems of the pKTN, and examine the validity of the algorithm with some preliminary experiments. Since ELISE is a distributed algorithm that makes only transitions relevant to a query fire, we expect that the pKTN can now be a powerful method to construct a large logical database equipped with an efficient reasoning tool.

There are 3 main methods of inference for Bayesian network from data. We compared their properties using actual example. Score base method by MDL(Minimum Description Length) can infer more appropriate structure than others. But this method occurs explosion of combination. So we proposed sparse network by significant link due to high MDL value and proved less problems for representation of structure by comparison with full link network.

Because of the recent increase in health consciousness and a rapid growth of user generated recipe data in social networking services, recipe mining, i.e. data mining in recipe databases, has been paid a wide attention as one of attractive research fields. In this paper, as a first step towards comprehensive and interpretable knowledge discovery from recipe databases, we investigate the applicability of the techniques in association rule mining to the recipe domain. We attempt to find interesting patterns which represent important structures among ingredients and characteristic words based on advanced association rules.

We tried to apply a Deep Learning to diagnose the lung cancer from a gas chromatography mass spectrometry data of human urine. The mother data consists of 28 healthy people and 39 lung cancer patient urine data sets. Each data set has 394 pieces of peak value as a feature. We applied unsupervised and supervised learning to four-layer neural network (NN). We got 97.0% accuracy of the diagnosis. We also used the trained NN for search the target substance.

In our research, a new method for sleep pattern characterization based on cluster analysis is proposed. After recording audio data during sleep, we extracted audio clips of events from the audio data and applied various types of Self-Organizing Map (SOM) algorithms on these data, including general SOM, Kullback-Leibler Kernel SOM, or Sequence-based Kernel SOM to be compared, and obtained clusters of sleep related events for a whole night. The sleep related events include snore, bruxism, limb movement, etc. Visualization of events distribution and transition aid in characterization of individual sleep pattern and comprehension of sleep quality evaluation.