Kazuyuki Shoji
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kazuyuki Shoji has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Haru Terashima
Graduate School of Engineering, Nagoya University
8 shared publications - Kenta Urano
Graduate School of Engineering, Nagoya University
8 shared publications - Naoki Tamura
Graduate School of Engineering, Nagoya University
8 shared publications - Nobuo Kawaguchi
Institutes of Innovation for Future Society, Nagoya University
8 shared publications - Takuro Yonezawa
Graduate School of Engineering, Nagoya University
8 shared publications - Shin Katayama
Graduate School of Engineering, Nagoya University
7 shared publications
Latest Publications
- Open Access
- 29 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Spatial Algorithms and Systems
- Vol. 12 (3) , 1-40
Additive Compositionality in Urban Area Embeddings Based on Human Mobility Patterns
Understanding human mobility patterns in urban areas is crucial for a wide range of applications, including urban planning, commercial area analysis, tourism strategies, and disease control measures. In response to these diverse needs, several area embedding techniques have been actively researched, aiming at modeling mobility patterns for individual urban areas as vector representations within latent spaces. On the other hand, word embeddings are known to exhibit a property called additive compositionality, where arithmetic operations on embeddings correspond to meaningful semantic manipulations (e.g., “king” - “man” + “woman” ≈ “queen”). However, it remains unclear whether this property also holds in area embeddings, and if so, how it can be practically applied. Therefore, this study introduces and formulates the concept of additive compositionality from linguistics into area embeddings, defining it as a frequency-weighted average where the weights correspond to the frequency of occurrence for each area in the mobility dataset. Furthermore, we propose practical methods leveraging this property for flexible post-hoc transformations of area embeddings into different spatial shapes without retraining, and quantitative modeling of temporal mobility shifts as arithmetic operations between embeddings from different periods. Our results demonstrate that additive compositionality in area embeddings can effectively support operations on mobility patterns, spatial embedding transformations, and analysis of temporal mobility dynamics.
- 29 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Spatial Algorithms and Systems
- Vol. 12 (3) , 1-40
Open AccessAdditive Compositionality in Urban Area Embeddings Based on Human Mobility Patterns
Understanding human mobility patterns in urban areas is crucial for a wide range of applications, including urban planning, commercial area analysis, tourism strategies, and disease control measures. In response to these diverse needs, several area embedding techniques have been actively researched, aiming at modeling mobility patterns for individual urban areas as vector representations within latent spaces. On the other hand, word embeddings are known to exhibit a property called additive compositionality, where arithmetic operations on embeddings correspond to meaningful semantic manipulations (e.g., “king” - “man” + “woman” ≈ “queen”). However, it remains unclear whether this property also holds in area embeddings, and if so, how it can be practically applied. Therefore, this study introduces and formulates the concept of additive compositionality from linguistics into area embeddings, defining it as a frequency-weighted average where the weights correspond to the frequency of occurrence for each area in the mobility dataset. Furthermore, we propose practical methods leveraging this property for flexible post-hoc transformations of area embeddings into different spatial shapes without retraining, and quantitative modeling of temporal mobility shifts as arithmetic operations between embeddings from different periods. Our results demonstrate that additive compositionality in area embeddings can effectively support operations on mobility patterns, spatial embedding transformations, and analysis of temporal mobility dynamics.
- Open Access
- 6 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Spatial Algorithms and Systems
- Vol. 12 (2) , 1-24
CLIFT: Cross-City Mobility-Derived Lifestyle Pattern Transfer for Improved Multi-City Next Location Prediction
We propose CLIFT ( C ross-City Lif estyle Pattern T ransfer for Human Mobility Prediction), a novel framework that enhances human mobility prediction by integrating general lifestyle patterns shared across cities with city-specific mobility patterns. Accurate human mobility prediction in urban environments is critical for transportation planning, marketing strategies, and disaster response. However, most existing deep learning approaches use only single-city data and exhibit significant performance degradation in small cities with limited training data. These limitations motivate methods that jointly leverage cross-city behavioral patterns and city-specific mobility characteristics. CLIFT addresses this challenge through dual complementary encoders: one captures general lifestyle patterns shared across cities, and the other captures city-specific mobility patterns; their outputs are integrated with a Transformer-based mobility predictor (LP-BERT). This architecture enables the model to jointly capture cross-city transferable behavioral patterns and city-specific mobility characteristics. We evaluated the effectiveness of CLIFT through experiments on the multi-city human mobility dataset LYMob-4Cities, comparing its performance with both single-city and multi-city deep learning-based methods. On average, CLIFT improved GEOBLEU and Top-1 accuracy by 11.1% and 10.6% over the single-city baseline, and by 5.0% and 7.9% over the multi-city baseline, respectively. Furthermore, CLIFT outperformed the top-ranked teams in the international competition, Human Mobility Prediction Challenge 2024 , demonstrating superior predictive performance under the same dataset and task setting.
- 6 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Spatial Algorithms and Systems
- Vol. 12 (2) , 1-24
Open AccessCLIFT: Cross-City Mobility-Derived Lifestyle Pattern Transfer for Improved Multi-City Next Location Prediction
We propose CLIFT ( C ross-City Lif estyle Pattern T ransfer for Human Mobility Prediction), a novel framework that enhances human mobility prediction by integrating general lifestyle patterns shared across cities with city-specific mobility patterns. Accurate human mobility prediction in urban environments is critical for transportation planning, marketing strategies, and disaster response. However, most existing deep learning approaches use only single-city data and exhibit significant performance degradation in small cities with limited training data. These limitations motivate methods that jointly leverage cross-city behavioral patterns and city-specific mobility characteristics. CLIFT addresses this challenge through dual complementary encoders: one captures general lifestyle patterns shared across cities, and the other captures city-specific mobility patterns; their outputs are integrated with a Transformer-based mobility predictor (LP-BERT). This architecture enables the model to jointly capture cross-city transferable behavioral patterns and city-specific mobility characteristics. We evaluated the effectiveness of CLIFT through experiments on the multi-city human mobility dataset LYMob-4Cities, comparing its performance with both single-city and multi-city deep learning-based methods. On average, CLIFT improved GEOBLEU and Top-1 accuracy by 11.1% and 10.6% over the single-city baseline, and by 5.0% and 7.9% over the multi-city baseline, respectively. Furthermore, CLIFT outperformed the top-ranked teams in the international competition, Human Mobility Prediction Challenge 2024 , demonstrating superior predictive performance under the same dataset and task setting.
- Open Access
- 3 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 1218-1221
Multi-City Next Location Prediction through Mobility-Derived Multi-Pattern Transfer Learning
- 3 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 1218-1221
Open AccessMulti-City Next Location Prediction through Mobility-Derived Multi-Pattern Transfer Learning
- Open Access
- 23 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 131803-131822
Life Pattern-Based Human Attribute Estimation Using Only GPS Data
Analysis of large-scale GPS data enables the summarization and structuring of human mobility characteristics, providing insights into residents’ behavior patterns. However, existing methods analyze sequences of individual-derived stay labels (e.g., home or workplace), which do not reflect actual area functions and may fail to distinguish individuals based on differences in workplaces. This limitation arises because annotating stay locations and purposes in raw GPS data is challenging due to the complexity of urban structures and constraints on data quality. In this study, we propose ArLPH, a framework that takes only raw GPS data as input, models mobility from the perspective of "when and what types of places he/she visited," and enables the interpretation of individuals’ life patterns. ArLPH consists of two modules: "Area Modeling," which represents urban areas based on usage patterns, and "Human Modeling," which represents individuals based on behavior patterns. By clustering the representations generated by these modules, it becomes possible to estimate occupational and behavioral attributes through the interpretation of life patterns. Through an evaluation of a large-scale, real-world GPS dataset collected from tens of thousands of smartphone users, we demonstrate that ArLPH can estimate occupational attributes, such as office workers or homemakers, as well as behavioral attributes, including those who frequently work on weekends.
- 23 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 131803-131822
Open AccessLife Pattern-Based Human Attribute Estimation Using Only GPS Data
Analysis of large-scale GPS data enables the summarization and structuring of human mobility characteristics, providing insights into residents’ behavior patterns. However, existing methods analyze sequences of individual-derived stay labels (e.g., home or workplace), which do not reflect actual area functions and may fail to distinguish individuals based on differences in workplaces. This limitation arises because annotating stay locations and purposes in raw GPS data is challenging due to the complexity of urban structures and constraints on data quality. In this study, we propose ArLPH, a framework that takes only raw GPS data as input, models mobility from the perspective of "when and what types of places he/she visited," and enables the interpretation of individuals’ life patterns. ArLPH consists of two modules: "Area Modeling," which represents urban areas based on usage patterns, and "Human Modeling," which represents individuals based on behavior patterns. By clustering the representations generated by these modules, it becomes possible to estimate occupational and behavioral attributes through the interpretation of life patterns. Through an evaluation of a large-scale, real-world GPS dataset collected from tens of thousands of smartphone users, we demonstrate that ArLPH can estimate occupational attributes, such as office workers or homemakers, as well as behavioral attributes, including those who frequently work on weekends.
- Open Access
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 5-9
Time-series Stay Frequency for Multi-City Next Location Prediction using Multiple BERTs
Human Mobility Prediction Challenge 2024 was organized to compare human future movement prediction methods using a unified dataset. The challenge focuses on human movement prediction in multiple cities with varying numbers of users. Many existing movement prediction methods train deep learning models using large-scale movement histories from cities and make predictions. While cities with large users have sufficient training data, smaller cities with fewer users may face challenges due to insufficient data, raising concerns about lower prediction accuracy. Additionally, it is difficult to treat stay locations between different cities' movement histories due to differences in spatial area arrangement, making it challenging to share data across cities. To address this issue, we propose human movement prediction method that utilizes time-series stay frequency patterns, which can be commonly applied across different cities. This method demonstrated superiority in predicting movements in cities with fewer users compared to models trained and predicted using only the movement histories of the target city. Furthermore, the method achieved top 10 prediction accuracy in HuMob Challenge 2024.
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 5-9
Open AccessTime-series Stay Frequency for Multi-City Next Location Prediction using Multiple BERTs
Human Mobility Prediction Challenge 2024 was organized to compare human future movement prediction methods using a unified dataset. The challenge focuses on human movement prediction in multiple cities with varying numbers of users. Many existing movement prediction methods train deep learning models using large-scale movement histories from cities and make predictions. While cities with large users have sufficient training data, smaller cities with fewer users may face challenges due to insufficient data, raising concerns about lower prediction accuracy. Additionally, it is difficult to treat stay locations between different cities' movement histories due to differences in spatial area arrangement, making it challenging to share data across cities. To address this issue, we propose human movement prediction method that utilizes time-series stay frequency patterns, which can be commonly applied across different cities. This method demonstrated superiority in predicting movements in cities with fewer users compared to models trained and predicted using only the movement histories of the target city. Furthermore, the method achieved top 10 prediction accuracy in HuMob Challenge 2024.
- Open Access
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 577-580
Additive Compositionality in Urban Area Embeddings Based on Human Mobility Patterns
Understanding the characteristics of various urban areas is crucial for applications such as urban planning, tourism policies, market analysis, and infection control. Techniques for embedding areas as vectors in a latent space based on human mobility patterns are actively researched. Many of these area embedding methods define areas as points, grids, or polygons on a geospatial plane and then embed them. However, existing methods do not allow for mutual transformation between these forms and sizes after the initial embedding. Additionally, if the characteristics of an area change due to events such as the opening of new buildings, re-embedding is necessary. Meanwhile, the Word2Vec technique, a representative word embedding method, has a property called additive compositionality. This property allows for the arithmetic operation of word meanings through the arithmetic operations of word embeddings. In this paper, we propose a method to apply this property to existing area embedding techniques, leveraging it for practical tasks such as area shape transformation and searching for areas with trends change.
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 577-580
Open AccessAdditive Compositionality in Urban Area Embeddings Based on Human Mobility Patterns
Understanding the characteristics of various urban areas is crucial for applications such as urban planning, tourism policies, market analysis, and infection control. Techniques for embedding areas as vectors in a latent space based on human mobility patterns are actively researched. Many of these area embedding methods define areas as points, grids, or polygons on a geospatial plane and then embed them. However, existing methods do not allow for mutual transformation between these forms and sizes after the initial embedding. Additionally, if the characteristics of an area change due to events such as the opening of new buildings, re-embedding is necessary. Meanwhile, the Word2Vec technique, a representative word embedding method, has a property called additive compositionality. This property allows for the arithmetic operation of word meanings through the arithmetic operations of word embeddings. In this paper, we propose a method to apply this property to existing area embedding techniques, leveraging it for practical tasks such as area shape transformation and searching for areas with trends change.
- Open Access
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 621-624
Unveiling Human Attributes through Life Pattern Clustering using GPS Data Only
Clustering people by their life patterns is valuable in government and business fields. Existing studies often rely on semantic data such as Point of Interest or stay purpose. However, they have the problem that obtaining large datasets is difficult due to the need for annotation work. Some studies try to use only location data. However, they do not reveal the semantics of the area where visitors stay because they only label visited areas by significance according to duration and frequency of stay. In this paper, we propose a framework, LPSeL, for clustering people's Life Patterns at a Semantic Level using only raw GPS location data. LPSeL is based on the idea that analyzing human mobility first requires understanding urban space. Therefore, it begins with area modeling, which models areas in a city based on people's activities. Then, treating human mobility as a sequence of area representations makes it possible to model individuals by semantic-level characteristics of their life patterns. We showed that LPSeL is capable of estimating people's attributes from their life patterns using a real-world dataset consisting of GPS data collected from tens of thousands of smartphone users.
- 29 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 621-624
Open AccessUnveiling Human Attributes through Life Pattern Clustering using GPS Data Only
Clustering people by their life patterns is valuable in government and business fields. Existing studies often rely on semantic data such as Point of Interest or stay purpose. However, they have the problem that obtaining large datasets is difficult due to the need for annotation work. Some studies try to use only location data. However, they do not reveal the semantics of the area where visitors stay because they only label visited areas by significance according to duration and frequency of stay. In this paper, we propose a framework, LPSeL, for clustering people's Life Patterns at a Semantic Level using only raw GPS location data. LPSeL is based on the idea that analyzing human mobility first requires understanding urban space. Therefore, it begins with area modeling, which models areas in a city based on people's activities. Then, treating human mobility as a sequence of area representations makes it possible to model individuals by semantic-level characteristics of their life patterns. We showed that LPSeL is capable of estimating people's attributes from their life patterns using a real-world dataset consisting of GPS data collected from tens of thousands of smartphone users.
- Open Access
- 13 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Human Mobility Prediction Challenge: Next Location Prediction using Spatiotemporal BERT
Understanding, modeling, and predicting human mobility patterns in urban areas has become a crucial task from the perspectives of traffic modeling, disaster risk management, urban planning, and more. HuMob Challenge 2023 aims to predict future movement trajectories based on the past movement trajectories of 100,000 users[1]. Our team, "uclab2023", considered that model design significantly impacts training and prediction times in the task of human mobility trajectory prediction. To address this, we proposed a model based on BERT, commonly used in natural language processing, which allows parallel predictions, thus reducing both training and prediction times. In this challenge, Task 1 involves predicting the 15-day daily mobility trajectories of target users using the movement trajectories of 100,000 users. Task 2 focuses on predicting the 15-day emergency mobility trajectories of target users with data from 25,000 users. Our team achieved accuracy scores of GEOBLEU: 0.3440 and DTW: 29.9633 for Task 1 and GEOBLEU: 0.2239 and DTW: 44.7742 for Task 2[2][3].
- 13 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessHuman Mobility Prediction Challenge: Next Location Prediction using Spatiotemporal BERT
Understanding, modeling, and predicting human mobility patterns in urban areas has become a crucial task from the perspectives of traffic modeling, disaster risk management, urban planning, and more. HuMob Challenge 2023 aims to predict future movement trajectories based on the past movement trajectories of 100,000 users[1]. Our team, "uclab2023", considered that model design significantly impacts training and prediction times in the task of human mobility trajectory prediction. To address this, we proposed a model based on BERT, commonly used in natural language processing, which allows parallel predictions, thus reducing both training and prediction times. In this challenge, Task 1 involves predicting the 15-day daily mobility trajectories of target users using the movement trajectories of 100,000 users. Task 2 focuses on predicting the 15-day emergency mobility trajectories of target users with data from 25,000 users. Our team achieved accuracy scores of GEOBLEU: 0.3440 and DTW: 29.9633 for Task 1 and GEOBLEU: 0.2239 and DTW: 44.7742 for Task 2[2][3].