IIAI Letters on Informatics and Interdisciplinary Research

Optimization of Recommendation System by Improving Serendipity and Grouping Users Based on Their Number of Data Points

Haruto Domoto — 2025-02-22

Information recommendation systems aim to deliver optimal content to users, but conventional methods often only suggest similar items, leading to user boredom and reduced recommendation effectiveness. This study addresses this limitation by focusing on “serendipity”, enhancing the unexpectedness of recommendations. We compared conventional methods, existing techniques, and six newly proposed approaches. Users were grouped into three categories based on the number of data points they evaluated to analyze the impact on recommendation performance. For users with fewer data points, the best approach was to recommend items significantly different from the average user preferences. For users with more data points, recommending items that other users disliked but held high value for the target user was most effective. These strategies improved diversity and unexpectedness without sacrificing usefulness, thereby successfully enhancing serendipity. This method shows promise in increasing user satisfaction by providing a more engaging recommendation experience.

Estimating Human Difficulty for SameGame Puzzles

Motoki Miura — 2025-02-22

SameGame is a computer puzzle game in which players must select adjacent blocks of the same color vertically and horizontally to remove them from the screen. Generally, initial blocks are arranged randomly, but it is known that the difficulty varies depending on the initial blocks. It is also known that it can be impossible to completely remove all blocks on small boards. As a way to quantitatively estimate the difficulty of SameGame puzzles, we devised an index that quantifies the correspondence between blocks when they are removed and the discrete state of those blocks. We compared our proposed index with a random success rate, which examines the ratio of the number of answer steps divided by the number of all possible steps. By investigating the correlation with the clear rate based on human play data, we analyzed the effectiveness of several indices including the proposed method.

How Does the Persona Given to Large Language Models Affect the Idea Evaluations?

Hiroaki FURUKAWA — 2025-02-22

This paper investigates the effect of personas in Large Language Models (LLMs) on idea evaluation. The language comprehension ability of LLMs has recently reached a level comparable to that of humans. Consequently, LLMs are being explored for their potential application in idea evaluation. However, LLMs face several challenges in their outputs, including hallucinations and biases. To address these issues, prompt engineering is utilized to guide LLMs toward producing desired results. This study focuses on Persona as a factor in prompt engineering for LLMs. Personas enable the reproduction and control of specific personalities within LLMs. The objective of this study is to validate the relationship between personas and idea evaluation using GPT-4. The results suggest that variations in personas influence the evaluation of ideas. Furthermore, a relationship was observed between evaluation scores and the evaluation criteria deemed important by the LLM.

Quantitative Evaluation of Perceived Acceptability of Accentual Patterns in Four-Mora Japanese Words Based on Lexical Attributes

Ikuyo Masuda-Katsuse — 2025-03-03

We generated four-mora Japanese word utterances with various pitch accent features and mapped the perceived pitch accent acceptability of native Japanese speakers onto a distribution of pitch accent features. The results confirmed that even for words classified as having the same accent type, the acceptability distribution differed depending on the word types. These results demonstrate the diversity of pitch accent perception and provide quantitative support from the perspective of auditory perception for knowledge about accent rules reported in linguistics. We also discussed the potential application of these results to CALL systems for Japanese language learners.

Digital Knowledge Twin: Bridging the Gap Between Physical and Cyber Knowledge Spaces by Generative AI

NAOSHI UCHIHIRA — 2025-02-22

The extraction, sharing, and utilization of explicit, latent, and tacit field knowledge (called “Gen-Ba knowledge”) possessed by skilled workers in fields such as healthcare, caregiving, agriculture, maintenance and inspection, and manufacturing, have long been recognized as important yet challenging tasks. Organizations and companies capable of achieving this through human systems hold a competitive advantage. With the advent of generative AI, explicit knowledge that can be codified (documented) has become dramatically more accessible. However, the challenge remains in how to efficiently and effectively handle the latent and tacit Gen-Ba knowledge possessed by humans, which AI cannot directly process. This study proposes a “Digital Knowledge Twin”, which facilitates externalization, combination, and internalization of Gen-Ba knowledge for bridging the gap between physical and cyber spaces. In the externalization phase, Gen-Ba knowledge is accumulated in the physical space as fragments of knowledge, integrating human voice messages, photos, and physical sensor data. Then, these Gen-Ba knowledge fragments are linked with structure of operations in the cyber space in a combination phase. Finally, Gen-Ba knowledge is collaboratively internalized in the physical space through workshops, where internalization means sensemaking and symbol grounding among humans. This paper presents the necessary technologies and system architecture required for the implementation of the “Digital Knowledge Twin.” Gen-Ba knowledge management using generative AI is a challenging issue, and the theoretical contribution of this study is that it provides one promising research direction of future knowledge management.

Investigation of Effective Heart Rate Variability Indices for Emotion Estimation During Long-term Images Gazing

Tetsuya Nishibe — 2025-02-22

In recent years, demand for emotion estimation using heart rate has increased in fields such as life-log and healthcare. HRV indices can estimate the activity level of the autonomic nervous system, but their sensitivity varies depending on the task and stimuli, sometimes leading to inconsistent results. Therefore, selecting the optimal HRV indices for each task and stimuli is important, and this study focused on emotional induction during long-term visual stimuli. We constructed the experiment in which subjects gazed at images of consistent emotional evaluation values, and identified the effective HRV indices for emotion estimation. The results showed that SDRR and L were effective for estimating excitement, Mean and rMSSD for estimating disgust, and LF/HF for classifying boredom and relaxation. These results are expected to serve as a basis for emotion estimation in situations where people are exposed to visual stimuli for long periods, such as when viewing videos or scenery.

Study on the Application of Weak Go AI in Go Education: the Balance between Go Progress and Emotional Value

CHI LIU — 2025-02-22

With the continuous advancement of artificial intelligence (AI) technology, the application of AI in Go education and training has become a hot research topic. This article explores the important role of weak Go AI in Go education, especially its application in beginner teaching. We analyzed the current Go education methods and research status, and based on this, proposed a technological innovation: by adjusting the difficulty of AI, it can promote the improvement of beginners’ Go skills while maintaining their enthusiasm for Go. Further experiments have shown that when the AI’s Go power is appropriately higher than the tester’s, it can effectively improve Go power while also considering emotional value. The research in this article proves that this innovation has profound significance for Go education.

An Automatic Bottom-up Idea Grouping Method based on LLMs

Takayuki Ito — 2025-02-22

This paper presents a novel automatic Bottom-up idea grouping method leveraging Large Language Models (LLMs) to facilitate the organization of ideas and concepts. Traditional methods of idea grouping, such as the KJ method, require significant manual effort to group and synthesize information effectively. Our approach utilizes the contextual understanding and grouping capabilities of LLMs to automate this process, aiming to reduce cognitive load and improve efficiency. By performing unsupervised grouping based on the insight of each idea, the model automatically generates cohesive idea groups from diverse sets of inputs in the bottom-up way. We evaluate the effectiveness of our method on qualitative datasets, comparing it with top-down categorizing method. Results indicate that the proposed bottom-up method not only aligns well with human clustering but also demonstrates a high level of interpretability and accuracy in grouping similar ideas. This study highlights the potential of LLMs in transforming qualitative analysis by offering a scalable and intuitive solution for idea grouping.

Exploring the Modality of Network-Type Thinking in Young Children with Picture Books

Ryosuke Kozaki — 2025-03-03

This study investigated the concrete modality of network-type thinking and knowledge entity formation of young children aged from four to six, employing the picture books. More precisely, we analyzed the techniques with which they associate words and concepts from read stories. As a result, it was demonstrated that the children substantially hire verbs and syntagmatic association when developing a knowledge network. In addition, it was also shown that six-year-olds recalled significantly more keywords than younger children when it came to syntagmatic (or “serial”) thinking method. Furthermore, this paper indicates some of the future possibilities of the findings that are to be applied to the more effective selection of picture books in nursery school settings for their language development.

Extracting Factors through Additional Impression Evaluation Experiment Assessing Both High-rated and Low-rated Reviews Posted at EC Sites

Yuya Yokoyama — 2025-02-22

The global diffusion Internet has established electronic commerce (EC) sites where anyone can purchase online. In order to avoid a mismatch between user and products, users can write a review on the product they purchased, helping users refer to the review of the commodity and make decisions. Nevertheless, with more users and items flooded on EC sites, the issues of mismatches are becoming conspicuous. In order to solve these issues, the authors conducted impression evaluation experiment to extract the impression of low-rated reviews. However, the previous analysis yielded only three factors due to insufficient experimental materials. Therefore, this paper reports further experiment with appending high-rated reviews. As a result of the analysis, eight factors are obtained under the fifty impression words. It could be concluded that the approach of extracting impression from the statements can be applicable to the review statements of EC sites under the unbiased experimental materials.

Constraint-Based Protocol for Efficient Automated Negotiation in Multi-issue Negotiation

Shun Okuhara — 2025-02-22

This study investigates the impact of the number of fixed constraints on automated negotiation outcomes in a multi-issue negotiation scenario. Specifically, we analyze how the variation in the fixed constraints set by each negotiating agent influences social welfare, agreement frequency, and negotiation time. The negotiation protocol considers fixed constraints and evaluates their effects on the efficiency and feasibility of reaching agreements. Simulation results reveal that as the number of fixed constraints increases, social welfare and agreement frequency tend to decrease while negotiation time remains relatively unchanged. These findings highlight the importance of understanding the role of fixed constraints in optimizing negotiation outcomes and provide insights for designing more effective automated negotiation protocols.

How Computer Vision Complements Health Care/Nursing Environment

Atsuo Yoshitaka — 2025-02-22

Aging of the population revealed that people of younger generation are facing to take care of more of old generation. It may not be satisfactory to arrange enough number of nursing staffs even in late-evening in hospitals or nursing and personal care facility, or it is not possible to watch every old people continuously in daytime. Detecting anomaly of human motion such as falling down, while no personal care assistant or nurse is accompanied, is expected to complement the insufficiency or lack of care by such staff. However, simple detection of anomaly of human motion such as falling down or sudden stoppage of the staff should be ignored as ‘false positive’ but adaptive detection of such motion by the persons who need assistance is preferable, in order to improve the operability of a system. This paper presents human anomaly detection based on human posture recognition by means of two-stage classification for the solution of abovementioned issues. Proposed method first classifies a person appeared in video into two classes, one corresponds to caregiver and another corresponds to aged people to be taken care of, and then detects anomaly motion only for the latters.