Papers - INUZUKA Nobuhiro
-
Store Clustering Based on Product Genre Representation in POS Data Reviewed
Yuta Marui,Atsuko Mutoh,Kosuke Shima,Koichi Moriyama,Tohgoroh Matsui,Nobuhiro Inuzuka
Program for 2024 IEEE 13th Global Conference on Consumer Electronics (GCCE) 2024.10
Authorship：Last author Language：English Publishing type：Research paper (international conference proceedings)
-
Martial Arts Demonstration Evaluation System Using Machine Learning to Reflect the Actual Evaluation Methods of Instructors Reviewed
Takeo Ueda, Kosuke Shima, Atsuko Mutoh, Koichi Moriyama, Tohgoroh Matsui, Nobuhiro Inuzuka
Procedia Computer Science (Prod. KES 2024) 2024.09
Authorship：Last author Language：English Publishing type：Research paper (international conference proceedings)
-
Enhancing Retrieval Processes for Language Generation with Augumented Queries to Provide Factual Ingormation on Schizophrenia Reviewed International coauthorship
Julien Pierre Edmond Ghali, Kosuke Shima, Atsuko Mutoh, Koichi Moriyama, Nobuhiro Inuzuka
Procedia Computer Science (Prod. KES 2024) 2024.09
Authorship：Last author Language：English Publishing type：Research paper (international conference proceedings)
-
Learning to Avoid Multiple Moving Obstacles Using Artificial Dangerous Fields
Shohei Yamanaka, Atsuko Mutoh, Kousuke Shima, Koichi Moriyama, Tohgoroh Matsui, Nobuhiro Inuzuka
2024-ICS-215 ( 5 ) 1 - 6 2024.09
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
-
Acquisition of Cooperative Behavior Using Curriculum Learning in Geometry Friends
2024-GI-51 ( 21 ) 1 - 7 2024.03
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
-
SACEQ: 連続空間における成功確率と収益に基づく安全な強化学習
大橋 宥斗,松井 藤五郎,武藤 敦子,森山 甲一,島 孔介,犬塚 信博
研究報告数理モデル化と問題解決（MPS） 2024-MPS-147 ( 1 ) 1 - 8 2024.02
Language：Japanese Publishing type：Research paper (conference, symposium, etc.) Publisher：情報処理学会
本論文では，連続行動空間を扱うことができる Soft Actor-Critic (SAC) を成功確率と収益に基づく強化学習 (EQ) に拡張する手法を提案する．近年，宇宙や被災地をはじめとした危険な環境においてロボットが突然危険な状況に陥ることを自律的に回避するために，強化学習により危険回避行動を学習する安全な強化学習という枠組みが提案されている．安全な強化学習の手法の一つに，高次元連続観測空間を扱う成功確率と収益に基づく強化学習 (HDEQ) がある．HDEQ は，本来離散空間のみで行えた成功確率と収益に基づく強化学習 (EQ) を拡張した手法であるが，連続行動空間を扱うことができなかった．そこで本論文では，連続観測空間のみならず連続行動空間においても EQ を扱えるよう，連続行動空間を扱うことができる SAC を拡張する．本論文では，この手法を Soft Actor-Critic for EQ (SACEQ) と呼ぶ．危険の回避が必要な環境における実験で，SACEQの効果を確認した．
-
Community Activity Model Reflected Community Consciousness
Takuya Sadasue,Kohsuke Shima,Atsuko Mutoh,Koichi,Moriyama,Tohgoroh Matsui,Nobuhiro Inuzuka
2024-MPS-147 ( 4 ) 1 - 6 2024.02
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
In recent years, the weakening of local community bonds has become a societal issue. Moreover, community societies are diverse groups composed of individuals with varying ages and values, and their composition varies depending on the region, such as rural or urban areas. Existing community activity models design the decision-making of agents based on the framework of planned action theory, making it difficult to express the diversity of local communities and their specific characteristics. In this study, we introduce a community consciousness that represents individuals' attitudes towards community into the existing models. Additionally, using the proposed model, we observe and analyze the impact of community consciousness on the formation of community activities. Through this analysis, we contemplate operational strategies to increase community activity participation by engaging individuals with low interest in community activities.
-
Game Recommend System Considering Playing Time With Friends
Kazuki Iwadare,Kohsuke Shima,Atsuko Mutoh,Koichi Moriyama,Tohgoroh Matsui,Nobuhiro Inuzuka
2024-MPS-147 ( 6 ) 1 - 6 2024.02
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
Information overload has become a problem because of the rapid development of the Internet, and recommendation systems have attracted attention and research as a solution. The use of game play time data has become mainstream in the gaming field, often focusing on the game play time of a single player and data about the game. However, many recent games require interaction with other players, and it is highly likely that players are influenced by players other than themselves. In this study, we proposed a new game recommendation system based on the relationship between in-game friends and game selection, focusing on co-play time, which is the time spent playing games together with in-game friends. Furthermore, we confirmed the effectiveness of the proposed method through experiments.
-
Opinion formation model considering fact-checking on SNS
2024-MPS-147 ( 8 ) 1 - 6 2024.02
Authorship：Last author
In recent years, information exchange through Social Networking Systems (SNS) has become popular, but it has been pointed out that it creates a polarized information environment known as an echo chamber. Prior research has proposed an opinion formation model that mimics SNS, and has shown that the basic structure of SNS promotes the occurrence of echo chambers and polarization of opinions. Although this model assumes political opinions among people with different partisanship, it has been confirmed that in the real world, the echo chamber phenomenon also occurs due to misinformation spread on SNSs. In addition, although fact-checking activities to correct misinformation have been conducted, it is considered possible that corrected information is not shared in an environment where echo chambers are formed. In this study, we propose a model of opinion formation on SNS that takes into account fact-checking and examine the impact of fact-checking in an environment where echo chambers are formed.
-
Evaluation System for Martial Arts Demonstration Using Acceleration Sensor Considering Uncertainty of Actual Evaluation Reviewed
Shohei Yamanaka, Atsuko Mutoh, Kousuke Shima, Koichi Moriyama, Tohgoroh Matsui, Nobuhiro Inuzuka
17 ( 1 ) 23 - 28 2024.02
Authorship：Last author Language：Japanese Publishing type：Research paper (scientific journal)
The recent proliferation of smartphones has made it possible for the common user to analyze movements without the use of special devices such as Kinect. A system that uses decision tree learning to evaluate the performance of a martial arts demonstration based on features obtained from the acceleration information of a smartphone attached to the performer's waist has been proposed, but the prediction accuracy of the system is not high. The reasons for this may be that the feature selection method is not appropriate and is not in line with the actual evaluation method of martial arts demonstrations. In the actual evaluation of martial arts demonstrations, which consist of multiple movements, the evaluator intuitively derives a single overall evaluation by summarizing the evaluations for each movements, so it is inappropriate to apply machine learning directly to each movement and the overall evaluation. In the proposed model, we realized a model that is closer to the human evaluation method by using machine learning on noisy labels to calculate the overall evaluation based on the majority vote of each movement evaluation, and confirmed that it can reproduce the evaluator's evaluation with high accuracy in experiments.
-
Applying Degree of Contribution to Curiosity-driven Search for Multi-agent Pursuit of Learning Preys
2024-ICS-212 ( 2 ) 1 - 7 2024.02
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
-
Enhancing Retrieval Processes for Language Generation with Augmented Queries International coauthorship International journal
Julien Pierre Edmond Ghali, Kosuke Shima, Koichi Moriyama, Atsuko Mutoh, Nobuhiro Inuzuka
2402 ( 16874 ) 1 - 28 2024.02
Authorship：Last author Language：English Publishing type：Research paper (other academic) Publisher：arXiv
In the rapidly changing world of smart technology, searching for documents has become more challenging due to the rise of advanced language models. These models sometimes face difficulties, like providing inaccurate information, commonly known as "hallucination." This research focuses on addressing this issue through Retrieval-Augmented Generation (RAG), a technique that guides models to give accurate responses based on real facts. To overcome scalability issues, the study explores connecting user queries with sophisticated language models such as BERT and Orca2, using an innovative query optimization process. The study unfolds in three scenarios: first, without RAG, second, without additional assistance, and finally, with extra help. Choosing the compact yet efficient Orca2 7B model demonstrates a smart use of computing resources. The empirical results indicate a significant improvement in the initial language model's performance under RAG, particularly when assisted with prompts augmenters. Consistency in document retrieval across different encodings highlights the effectiveness of using language model-generated queries. The introduction of UMAP for BERT further simplifies document retrieval while maintaining strong results.
Other Link： https://arxiv.org/abs/2402.16874
-
Automating Lexicon Generation: A Comprehensive Review of Alternative Approaches Reviewed International coauthorship International journal
Julien Pierre Edmond Ghali, Nobuhiro Inuzuka, Kosuke Shima, Koichi Moriyama, Atsuko Mutoh
Procedia Computer Science 225 1142 - 1150 2023.12
Language：English Publishing type：Research paper (scientific journal) Publisher：ScienceDirect
Lexicon-based approaches to Document Classification are widely used, but the manual construction of lexicons can be time-consuming and resource-intensive. In this paper, we propose methods for automating the generation of lexicons later used for Document Classification. We explored diverse methods for generating lexicons, including semantic matches, frequency-based approaches, machine learning algorithms, and large language model techniques. We, later, used these lexicons to classify documents based on their content. By comparing our different lexicons results on a same task, based on criteria such as scalability and the F1 score, we determine optimized use-case for those methods. We show that our automated approaches are effective and efficient, producing accurate classifications with minimal human intervention. Some approaches have the potential to streamline the document classification process, reducing the time and resources required for manual lexicon generation, it also gives optimized use-case for the different methods. Thereafter, we discussed the obtained results.
DOI： https://doi.org/10.1016/j.procs.2023.10.102
Other Link： https://www.sciencedirect.com/science/article/pii/S1877050923012607
-
Store Cluster Analysis by Extracting Order Trends Using POS Data Reviewed
2023-BIO-73 ( 12 ) 1 - 6 2023.10
Authorship：Last author Language：Japanese Publishing type：Research paper (scientific journal)
POS data analysis is a technique for analyzing consumer purchasing behavior. In order to improve product sales, there is a technique for clustering stores based on the ratio of product sales by genre to total retail store sales, and clarifying the relationship between location and the tendency to order at stores within a cluster. However, in restaurants, the trend is not in the number of product sales, but in the content of individual customer orders. In this study, we propose a method to classify stores by extracting ordering trends from the orders of each receipt using non-negative matrix factorization. Since this method classifies stores based on frequently occurring order trends, it can be used to identify the relationship between store needs and location/environment. In an experiment, we applied the proposed method to POS data of several restaurants and confirmed its effectiveness.
-
Evacuation Planning Model Considering Diurnal Variation in the Number of People Staying
16 ( 2 ) 103 - 109 2023.10
Authorship：Last author Language：Japanese Publishing type：Research paper (scientific journal)
In recent years, we have seen many large-scale disasters such as earthquakes, and overflows of shelter capacity often occurred due to the day-night population gap at the time of the Great East Japan Earthquake. Therefore, it has become an important issue to allocate shelters in consideration of the day-night population gap (diurnal variation). In the shelter assignment problem, the use of zero-suppressed dichotomous decision graphs (ZDD) makes it possible to enumerate all patterns of shelter assignment. Studies using ZDD evaluated the distance to the shelter and the capacity factor, but did not consider the dwell time on the travel route, and used the fixed number of evacuees. In this study, we add an evaluation index using a network flow simulation method in order to take into account the number of people who stay in shelters along travel routes. In addition, we propose a method for obtaining shelter allocation that takes into account the diurnal variation in the number of people staying in shelters using stay history data for all time segments. In our experiments, we used three evaluation values: distance, occupancy rate, and evacuation completion time, and confirmed the effectiveness of the method through experiments using real environments.
-
Effect of Curiosity-driven Search in Multi-agent Deep Reinforcement Learning Reviewed
16 ( 2 ) 80 - 90 2023.10
Authorship：Last author Language：Japanese
In a multi-agent environment where multiple agents exist, the agents may achieve better results by cooperating one another than by acting self-interestedly. Curiosity-driven exploration, which is used in a single-agent environment with sparse rewards, offers rewards to unknown states, and the agent learns a policy to seek such unknown states accordingly. Using it in a multi-agent environment may allow the agents to acquire cooperative behavior by promoting exploration and changing the reward structure. The experiments we conducted in a predator-prey problem show that the curiosity-driven exploration allows the predators to catch more preys.
-
An Analysis of Stylistic Feature Changes in Language Use with Conversation Volume on Twitter Reviewed
16 ( 1 ) 1 - 7 2023.03
Authorship：Last author Language：Japanese Publishing type：Research paper (scientific journal)
In the field of sociolinguistics, it has been shown that people manipulate their social distance from others and increase the efficiency of communication by changing their language in conversation based on accommodation theory. In social psychology, the mere exposure effect, in which repeated contact improves the impression the other person, has been identified, and some studies have shown that the greater the number of contacts in a short period, the better the impression. Conversation is one way people make contact, and we can assume that people manipulate their relationships with others by improving their impressions of others and changing their language through increased conversation. On the other hand, the platform Twitter enables the observation of people's behaviour on a large scale. `Reply' on Twitter can be regarded as conversations with others. In this study, we propose a model to quantitatively extract changes in stylistic features of language with the volume of conversation per unit time on Twitter. In the proposed model, we convert the stylistic features of the conversation data into a textual feature matrix. Then, we analyze the conversation data by extracting important textual features that increase or decrease with the conversation volume per unit time using non-negative matrix factorization. Experimental results showed a number of linguistic findings, such as the fact that increased volume of conversations leads to fewer polite expressions and more straightforward sentences.
Other Link： https://ipsj.ixsq.nii.ac.jp/ej/?action=repository_uri&item_id=225568&file_id=1&file_no=1
-
隠れインデックスファンド検出のための One-class SVM 出力校正法
籔内 陽斗, 松井 藤五郎, 武藤 敦子, 島 孔介, 森山 甲一, 犬塚 信博
研究報告バイオ情報学（BIO） 2023 ( FIN-030 ) 32 - 39 2023.03
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
ファンド(投資信託)の中には目論見書に記された運用方針と実際の運用方針が異なる場合がある。先行研究では実際の運用方針に基づくファンドのグループ分けを目的に、月次リターンの時系列に対してk-meansとUMAPを組み合わせた手法を用いてファンドをクラスタリングする方法が提案された。しかし、k-meansによるクラスタリングではラベル情報を一切使用しないという問題がある。本論文では、インデックスファンドのラベルは正しいと仮定し、ラベル情報を使用可能な手法であるOne-class SVMを導入する。ところが、通常のOne-class SVMでは、ラベルが付与されたインデックスファンドだけをインデックスと判別するモデルを作成するため、インデックスに類似した隠れインデックスファンドをインデックスと判別することができない。この問題を解決するために、本論文ではOne-class SVMモデルの出力値に着目し、One-class SVMの出力を校正する新しい方法を提案する。提案手法を同一の指標をベンチマークとするファンド群に対して適用することで隠れインデックスファンドを見つけることができた。
-
マルチエージェント深層強化学習における好奇心探索の影響
岩科 亨, 森山 甲一, 松井 藤五郎, 武藤 敦子, 犬塚 信博
研究報告数理モデル化と問題解決（MPS） 2023-MPS-142 ( 14 ) 1 - 6 2023.03
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
マルチエージェント環境には複数のエージェントが存在するが，それぞれが利己的な行動を獲得してしまうよりも，協力することでより良い結果が得られる可能性がある．報酬が疎なシングルエージェント環境で利用される好奇心探索は，未知の状態に対し報酬を与えることで最大化すべき報酬が変化し，それに伴い方策も未知の状態を求めるよう変化する．それをマルチエージェント環境で用いることで，探索拡大に加え，報酬の構造の変化により，獲得できていなかった協力的な行動に繋がる可能性がある．好奇心探索を用いて追跡問題で実験したところ，より多くの獲物を捕まえることができた．
-
滞在人数の日内変動を考慮した避難計画モデル
山本 正也, 武藤 敦子, 島 孔介, 森山 甲一, 松井 藤五郎, 犬塚 信博
研究報告バイオ情報学（BIO） 2023-BIO-73 ( 25 ) 1 - 5 2023.03
Authorship：Last author Language：Japanese Publishing type：Research paper (conference, symposium, etc.)
近年，地震などの大規模な災害の発生が多く見られ，東日本大震災の際には，昼夜人口のギャップにより容量溢れが多く発生した．そのため，昼夜人口のギャップ（日内変動）を考慮して避難所を割り当てることは重要な課題になっている．避難所割り当て問題において，ゼロサプレス型二分決定グラフ（ZDD）を用いることで避難所割り当てのパターンを全列挙することが可能になっている．ZDD を用いた研究では，避難所までの距離と収容率で評価をしているが，移動経路での滞留は考慮しておらず，避難対象者は常に固定した人数で行っている．そこで本研究では，移動経路での滞留を考慮するためにネットワークフローによるシミュレーション手法での評価を追加し，滞在履歴データを用いて全時間区分で評価を行うことで滞在人数の日内変動を考慮した避難所割り当てを求める手法を提案する．距離・収容率・避難完了時間の 3 つを評価値とし，実環境を用いた実験により効果を確認した．
PREV - 