Transactions of the Japanese Society for Artificial Intelligence Vol. 30(2015) No. 2
Regular, Special "Selected Short Papers from JSAI2014"

Cognitive psychology and behavioral economics have shown that humans have cognitive biases that deviate from normative systems such as classical logic and probability theory. Considering that humans have the ability to understand the world from sparse and/or imprecise data, it is natural to assume that the biases in human have some ecological merits in adaptation. We focus on two cognitive biases, symmetry and mutual exclusivity, that are considered peculiar to human. In this study, with the framework of empirical Bayes, we clarify the implication of a model of human causal cognition, the loosely symmetric (LS) model [Shinohara 07]) that implements the cognitive biases. We show that LS has great descriptive validity in inductive inference of causal relationship (causal induction) with a meta-analysis and an experiment in causal induction. The result of another experiment strongly suggests that humans use the inductively inferred causal relationship to decision-making. Then we show that LS effectively works in sequential decision-making under uncertainty (N-armed bandit problems). Operating LS as a simple value function under the greedy method in the framework of reinforcement learning, we analyze its behavior in terms of cognitive biases or heuristics under uncertainty. The three cognitive properties resulting from the loose symmetry, comparative valuation, satisficing, and prospect theory-like risk attitudes, are shown to be the key of the performance of LS. We parameterize the reference for satisficing and show that the quite intuitive parameter enables optimization.

We propose a system called Haptic Pad for Impressive Text Communication for creating text messages with haptic stimuli using the SPIDAR-tablet haptic interface. This system helps users indicate emotion and action of characters in text messages by attaching physical feedback. We evaluated the effectiveness of the system experimentally in three scenarios: storytelling, business e-mail and text messaging. We also conducted experiment for comparing effectiveness of haptic pattern, bold text and emoticon. We found that haptic stimuli are attached to facilitate three functions: emphasis in text messages, expression of emotion, and scene description. We also found that it is effective to apply HAPPicom not only to text messaging but to storytelling.

The spread of Video on Demand (VOD) services is driving the creation of video viewing environments in which users can watch whatever they like, whenever they like. The recent appearance of social networking services (SNSs), moreover, is bringing big changes to the world of media by enabling anyone to become a disseminator of information. We are studying a platform that combines VOD and SNS to create ``horizontal links'' between program viewers, and to facilitate encounters with new programs. To investigate user behaviors on this platform, we built an SNS site called ``teleda,'' that enables program viewing by VOD, and conducted a large-scale, three-month verification trial with about 1000 participants. In this paper, we report on the relational analysis of teleda users' viewing and communication behaviors. In order to clarify how the users' communication structures relate to the viewing and posting behaviors on this system, we described the communication structures in terms of network structures, and ran a correlation analysis of network indicators and user behavior indicators such as viewing and posting. As a result, we revealed that relationships between the users' communication structures, viewing behaviors, and posting actions are characteristics of each program's genre.

Linked Open Data (LOD) has a graph structure in which nodes are represented by URIs, and thus LOD sets are connected and searched through different domains. In fact, however, 5% of the values are literal (string without URI) even in DBpedia, which is a de facto hub of LOD. Therefore, this paper proposes a method of identifying and aggregating literal nodes in order to give a URI to literals that have the same meaning and to promote data linkage. Our method regards part of the LOD graph structure as a block image, and then extracts image features based on Scale-Invariant Feature Transform (SIFT), and performs ensemble learning, which is well known in the field of computer vision. In an experiment, we created about 30,000 literal pairs from a Japanese music category of DBpedia Japanese and Freebase, and confirmed thatthe proposed method correctly determines literal identity with F-measure of 76--85%.

Various notifications on a display for e-mail arrival, micro-blog updates, and application updates are becoming increasingly important. Thus many studies have been done to make notifications not to prevent a user to achieve a task. In this paper, we propose a novel notification method, the peripheral notification that uses the human cognitive property, visual field narrowing, that a human does not recognize subtle changes in a peripheral visual field when he/she concentrates on a task and that he/she automatically recognizes the changes when not concentrating on the task. By only setting a notification icon in the area of visual field narrowing, a user automatically and easily accepts the notification only when his/her concentration breaks. We conducted two experiments to investigate a visual field narrowing region in a controlled environment and evaluate the effectiveness of peripheral notifications by comparing with a traditional method in a more realistic environment. As a result, we confirmed the visual field narrowing in a display, and our peripheral notification had advantages.

We propose a prediction method for higher-order relational data from multiple sources. The high-dimensional property of higher-order relations causes problems associated with sparse observations. To cope with this problem, we propose a method to integrate higher-order relational data from multiple sources. Our target task is the simultaneous decomposition of higher-order, multi-relational data, which corresponds to the simultaneous decomposition of multiple tensors. However, we transform each tensor into an incidence matrix for the corresponding hypergraph and apply a nonlinear dimensionality reduction technique that results in a generalized eigenvalue problem guaranteeing global optimal solutions. We also extend our method to incorporate objects' attribute information to improve prediction for unseen/unobserved objects. To the best of our knowledge, this is the first reported method that can make predictions for (1) higher-order relations (2) with multi-relational data (3) with object attribute information and which (4) guarantees global optimal solutions. Using real-world datasets from social web services, we demonstrate that our proposed method is more robust against data sparsity than state-of-the-art methods for higher-order, single/multi-relational data including nonnegative multiple tensor factorization.

This paper proposes a mixture model that considers dependence to multiple topics. In time series documents such as news, blog articles, and SNS user posts, topics evolve with depending on one another, and they can die out, be born, merge, or split at any time. The conventional models cannot model the evolution of all of the above aspects because they assume that each topic depends on only one previous topic. In this paper, we propose a new mixture model which assumes that a topic depends on previous multiple topics. This paper shows that the proposed model can capture the topic evolution of death, birth, merger, and split and can model time series documents more adequately than the conventional models.

In AI communities, many applications utilize PageRank. To obtain high PageRank score nodes, the original approach iteratively computes the PageRank score of each node until convergence from the whole graph. If the graph is large, this approach is infeasible due to its high computational cost. The goal of this study is to find top-k PageRank score nodes efficiently for a given graph without sacrificing accuracy. Our solution, F-Rank, is based on two ideas: (1) It iteratively estimates lower/upper bounds of PageRank scores, and (2) It constructs subgraphs in each iteration by pruning unnecessary nodes and edges to identify top-k nodes. Experiments show that F-Rank finds top-k nodes much faster than the original approach.

We propose a new Japanese electronic text format with phrase-based line breaking for tablet computer to improve reading speed. The new text format prohibits splitting of a phrase and breaks a line between phrases. We measured reading speeds and eye movements using both the new text format and a conventional text format. Reading speeds for the new text formats are faster compared to the conventional text formats at all line lengths tested. The enhancement of reading speed in the new text format seems to be caused by the optimization of eye movements at the beginning of a long-length line, and the increase of short-length lines that can be recognized by a single fixation without horizontal saccade.