| 笹崎 海利 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 荒牧 英治 | 諏訪 博彦 | 松井 智一 |
|
title: Exploration of in-home micro-activity class representation through speech for activity recognition
abstract: The development of sensor and machine learning technologies is enhancing the accuracy of activity recognition; however, many studies remain focused on the “coarse” granularity of macro activity recognition. For instance, providing context-aware in-home services that are highly convenient, like cooking activity support services, requires recognizing the resident’s current activities at a more detailed, micro-activity level. Traditional micro-activity recognition methods necessitate pre-defining activity classes, which poses challenges in terms of scalability. This study aims to develop a scalable micro-activity representation method with practical utility in mind, as well as a method for activity recognition using this representation. By utilizing distributed representations, it becomes possible to estimate micro-activities through calculations based on the similarity with neighboring data. Specifically, I propose an activity recognition method using distributed representations that define activities comprehensively with finite-dimensional vectors. To leverage the expressive power of natural language, I constructed a language-activity encoder that generates activity distributed representations from natural language. This allows for the creation of a dataset pairing video data with distributed representations. Ultimately, I use a model with the pretrained Expanding Architectures for Efficient Video Recognition (X3D) as a feature extractor to infer activity distributed representations from video data and perform activity recognition. I report on the creation of a language-activity encoder that acquires activity distributed representations from descriptive activity texts, evaluates its accuracy using the public STAIR activities dataset, and presents initial experimental results from the activity recognition model that infers activity distributed representations from video data. language of the presentation: Japanese | ||||||
| 洲澤 春樹 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 荒牧 英治 | 諏訪 博彦 | 松田 裕貴 |
|
title:
Leveraging LLM for Managing Productive Daily Routines and Smartphone Usage
abstract: Smartphone applications, like social media or mobile games, can be highly addictive, often hindering the development of productive habits such as exercising or studying. Apps designed to limit smartphone usage aim to address this issue by blocking access to addictive apps. However, these tools typically operate on a fixed schedule, which can result in restrictions even when users have made progress in completing their daily routines. This paper introduces MutterHabits, an iOS application that monitors the real-time completion of daily routines and regulates access to addictive apps through an LLM-based chatbot. The chatbot allows users to negotiate app usage based on their progress. A two-week study (n=47) was conducted to assess the potential of LLMs as personal coaches for habit building. The results indicate that MutterHabits' negotiation feature lowers psychological resistance to starting exercise and study habits, facilitating the formation of these habits. language of the presentation: Japanese | ||||||
| 平良 繁幸 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 荒牧 英治 | 諏訪 博彦 | 松田 裕貴 |
|
title:Estimation of Response Attitudes Based on Interaction Logs in Conversational Interfaces for Improving Response Quality in Crowdsourcing
abstract:Crowdsourcing is widely used as a method to collect large-scale data at low cost. However, a challenge arises when some respondents prioritize task completion over providing accurate answers, which can undermine the reliability of the survey results. To address this issue, this study focuses on the ease of intervention through the use of conversational interfaces. We propose a method that facilitates task execution in a conversational format while simultaneously detecting inappropriate answering attitudes in real time by analyzing users' device interaction logs. The proposed method aims to improve respondents' attitudes through intervention. To evaluate the feasibility of detecting inappropriate respondents using machine learning based on interaction logs, we conducted an image-based questionnaire task involving 40 participants. We estimated respondents' attitudes during the task based on their device interactions. As a result, the findings suggest that it is possible to estimate respondents' attitudes using machine learning models trained on interaction data. language of the presentation: Japanese | ||||||
| 上田 健太郎 | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 荒牧 英治 | 諏訪 博彦 | |
|
title: Exploring the Potential of Machine Learning and Text Data in Financial Markets
abstract: Many systems utilizing machine learning techniques have been proposed to support decision-making in investment. In particular, leveraging unstructured data for financial market prediction and understanding has been providing new value for investment decisions. This study focuses on text data and proposes methods to enhance such systems. Specifically, we propose: "(1) Social media text embedding method that considers sentiment and topic information for financial market prediction," "(2) Feature extraction architecture that takes into account the characteristics of post formats in online forum (online message board)," and "(3) Development of a large language model (LLM) capable of handling tables and numerical representations within financial texts." In this presentation, we will report on the results of (1) and (2). language of the presentation: Japanese | ||||||