In recent years, large-scale discussions on theWeb have attracted much attention. In this situation, research on large-scale online consensus building support systems has been actively conducted. Facilitators are essential to discuss smoothly in the discussion. However, it is difficult for humans to play the role of facilitator in a large-scale discussion because participants post hundreds or thousands of opinions and the discussion may be held all day long. Therefore, developing automated facilitation agent that can automatically facilitate discussions is required. In order to develop automated facilitation agent, there are previous studies on extracting discussion as a graph structure. The purpose of this study is to perform node classification, which is one of the subtask of discussion structure extraction, with high accuracy. In the proposed method, we use Bidirectional Encoder Representations from Transformers (BERT), which was the state of the art in the field of natural language processing at the time of its release and is still widely studied and used today, to obtain distributed representations from sentences. We also use Gated Attention Network (GaAN), which has an attention mechanism and can propagate and calculate the information of neighboring nodes including the importance, in order to classify the obtained distributed representations. We adopt the Issue-Based Information System (IBIS) structure, which is designed to promote creative and constructive discussions, for the discussion structure. The experimental results show that our method can classify opinions with higher accuracy than previous studies and other general classification methods.

To improve the accuracy to prevent from sharing incorrect opinion, this paper proposes a method which can share correct opinions based on majority decision for multi-opinion, named Gradient Descent Weight Tuning (GDWT). In the experiment, this paper compares GDWT with AAT and Self-information Weight Tuning (SWT) which weights the opinion from the agent has clear information from the environment to share correct opinion based on majority decision from environment-to-agent information as the previous methods. From the result of the experiment to investigate the proposed methods on some complex networks with multi-opinion, this paper reveals that (1) though the previous method performs worse in order from less kinds of opinions, the proposed methods performs well (SWT: 0.8, GDWT: 0.9 accuracy); (2) GDWT performs the best without incorrect opinions; and (3) It is clear that GDWT is sensitive for the received incorrect opinions and the own parameters, especially target opinion formation rate, as comparing with SWT. These issues are discussed for future works.

Post-hoc explanation methods for machine learning models have been widely used to support decision-making. Counterfactual Explanation (CE), also known as Actionable Recourse, is one of the post-hoc explanation methods that provides a perturbation vector that alters the prediction result obtained from a classifier. Users can directly interpret the perturbation as an “action” to obtain their desired decision results. However, actions extracted by existing methods often become unrealistic for users because they do not adequately consider the characteristics corresponding to the data distribution, such as feature-correlations and outlier risk. To suggest an executable action for users, we propose a new framework of CE, which we refer to as Distribution-Aware Counterfactual Explanation (DACE), that extracts a realistic action by evaluating its reality on the empirical data distribution. Here, the key idea is to define a new cost function based on the Mahalanobis distance and the local outlier factor. Then, we propose a mixed-integer linear optimization approach to extracting an optimal action by minimizing the defined cost function. Experiments conducted on real datasets demonstrate the effectiveness of the proposed method compared with existing CE methods.

One of the AI systems introduced into our society in recent years is a system that negotiates and collaborates with users by presenting and recommending information through dialogues with multiple users. In general, users have different cultures and values. Therefore, the system must be able to estimate those values and then adapt its behavior to the user. In this study, we focus on the "werewolf game" as a benchmark for this type of technology. The werewolf game is an imperfect information game in which the game proceeds through communication. A werewolf game is a team game and has both cooperative and adversarial characteristics. In werewolf games, it is important not to cause conflicts among allies due to differences in culture and way of thinking. In this study, we first create multiple agent groups with different cultures. Then, we show that there is no specific strong strategy, and that the optimal strategy is different for each group. Then, we build an agent that can estimate the culture of the strategy that players other than itself are following, and can act in such a way that it adapts itself to that culture, and conduct an evaluation based on the winning rate. The results show that the proposed agent is able to adapt to the group and increase its winning rate under certain limitations.

The purpose of this research is to verify the factors that promote creation of new outcomes through co-creation activities. The following three hypotheses were made; 1. The higher degree of achievements of aims set by the academic workshop, the more user activity on social media and the number of outcomes through co-creation projects created, 2. The more positive feedback on user activities on social media, the more user activity increases, 3. As the number of user activities on social media increases, the number of outcomes generated through co-creation projects increases.

The target of this research was Special Interest Group on Crowd Co-Creation Intelligence at The Japanese Society for Artificial Intelligence. The feature of this group is to support creation of co-creation project and to promote activities by combining the academic workshop and social media. The following four data were analyzed; (a) Result of questionnaire conducted after academic workshops, (b) Log data of social media, (c) The number of new outcomes generated by co-creation activities. Each data of (a) to (c) were aggregated for each co-creation project. Spearman's rank correlation coefficient was calculated. A causal relationship model that affected the number of outcomes was constructed by path analysis using items which correlation was significant.

As results, the following four findings were obtained; 1. Promoting the sharing of participants' activities and interests in the academic workshop activate the number of discussion on social media and increase the number of outcomes generated through co-creation projects, 2. Increasing the number of cheer for co-creation projects contributes to increasing the number of comments on social media and outcome of co-creation projects, 3. Supporting increase the number of comments on social media increases the number of outcomes generated by co-creation projects, 4. The causal relationship model that affects the number of outcomes had a certain validity.

”Phantom Genji Scrolls ”are one of the most debatable artifacts whose painter school is still unknown in the history of art. In this study, the authors have reached a conclusion that the Phantom Genji Scrolls would have been painted by the painters who studied Kyoto-Kano and Tosa painter schools but not Edo-Kano painter school.

As a learning data set, more than 1500 face images are extracted from well-known the Tale of Genji pictures painted from Heian (12 Century) to Edo (17 Century) periods whose painter schools are all established in the history of art. The face images are written in typical old painting style so called“line-eye and hook nose (hikime-kagihana)” which often represents the characteristics of the painter schools.

The authors not only identified the painter school of the Phantom Genji Scrolls but also discovered the inconsistency in Iwasa painter school by means of artifact-based painter school learning model. The t-SNE scatter plots clearly indicated that the 266 face images extracted from the Phantom Genji Scrolls were surrounded by the Kyoto- Kano painter school learning data sets. It should be also noted that the 266 face images were far from the learning data set of Ujinobu Kano, one of the typical Edo-Kano painter. Interestingly, the Phantom Genji Scrolls and Ujinobu Kano were intercepted by Mitsuyoshi Tosa, one of the typical Tosa painter in scatter plot. This suggests that the painters of the Phantom Genji Scrolls may have been affected by both Kyoto-Kano and Tosa painter schools, instead of Edo-Kano painter school.

As for the Iwasa painter school, the authors came across strange behavior that the artifacts of Katsutomo Iwasa were identified as the Tosa painter school, even if the validation data of Katsutomo Iwasa were all included in the learning data set of Iwasa painter school, mostly constituted by Katsutomo Iwasa (244 face images). After careful observation and discussion, the authors have concluded that Matabei Iwasa and Kastutomo Iwasa may be too different to be categorized into the same Iwasa painter school with regards to face characteristics.

In this paper, we introduce a user ’s pre-declaration for her demand in order to improve a on-demand ridesharing operation plan by decreasing the number of vehicles. While a reservation requires a user to declare a precise riding time, we allow users to report a certain time period in advance. We evaluate the effectiveness of pre-declaration by computational simulations. In more detail, we evaluate the number of vehicles based on users’ pre-declarations to satisfy the threshold condition for average pick-up time by running computational simulations. However, A detailed simulation requires huge computation time. To solve this problem, we propose a method to determine the number of vehicles to run a simulation by applying Bayesian optimization. In the proposed method, the acquisition function is used to search for the number of vehicles whose average pickup time is around a threshold value. The experimental results show that the proposed method works well in reducing the number of vehicles.