Understanding how cities are perceived from on-site visitors' perspectives can provide valuable insights for urban planning and development applications. However, existing studies estimated people's perceptions by having them view photographed landscape images; the scores derived by these methods were thus merely quantified impressions of specific viewpoints that do not necessarily represent perceptions people would have were they at the site. To address this issue, we developed a framework, named Omni-CityMood, for quantifying people's on-site perceptions of urban atmospheres. Based on the idea that the viewpoint influences the perception of an urban landscape, the proposed framework identifies critical viewpoints of a location by using both visual-based features of landscape images and geographical characteristics of the site. In particular, Omni-CityMood enables the mood of a location to be evaluated from viewpoints over a range of 360 degrees by leveraging the techniques of neural recommendation systems. We evaluated Omni-CityMood on a dataset we built that includes perceived atmosphere experiences in various cities. Experiments and extensive analyses demonstrate the promising capability of modeling landscape viewpoints to quantify urban on-site atmospheres.

Individual mobility prediction holds significant importance in urban computing, supporting various applications such as place recommendations. Current studies primarily focus on frequent mobility patterns including commuting trips to residential and workplaces. However, such studies do not accurately forecast irregular trips, which incorporate journeys that end at locations other than residences and workplaces. Despite their usefulness in recommendations and advertising, the stochastic, infrequent, and spontaneous nature of irregular trips makes them challenging to predict. To address the difficulty, this study proposes a web search-driven bipartite graph neural network, namely WS-BiGNN, for the individual irregular mobility prediction (IIMP) problem. Specifically, we construct bipartite graphs to represent mobility and web search records, formulating the IIMP problem as a link prediction task. First, WS-BiGNN employs user-user edges and POI-POI edges (POI: point-of-interest) to bolster information propagation within sparse bipartite graphs. Second, the temporal weighting module is created to discern the influence of past mobility and web searches on future mobility. Lastly, WS-BiGNN incorporates the search-mobility memory module, which classifies four interpretable web search-mobility patterns and harnesses them to improve prediction accuracy. We perform experiments utilizing real-world data in Tokyo from October 2019 to March 2020. The results showcase the superior performance of WS-BiGNN compared to baseline models, as supported by higher scores in Recall and NDCG. The exceptional performance and additional analysis reveal that infrequent behavior may be effectively predicted by learning search-mobility patterns at the individual level.

When visiting universities, you might notice the distinctive atmospheres of each university, such as a calm and serious environment or a lively enthusiasm for sports. Capturing these atmospheres could help in promoting universities and fostering development in the communities around the universities. To explore the atmospheres of universities, we analyze the thoughts and interests of university community members, such as students and faculty members. Specifically, we use a large-scale dataset derived from search queries and GPS logs to quantify visitors’ interests. Additionally, to extract the meaningful atmospheres of universities, we apply topic modeling to the dataset.

In risk management and investment strategy formulation within financial markets, successfully predicting the Volatility Index (VIX) is crucial. While prediction methods leveraging social media texts have been considered promising, they predominantly rely on Twitter data, leaving the potential of online forums underexplored. However, online forums are rich sources of investor discussions and can provide valuable information for VIX prediction. In this study, we propose a deep learning architecture called ForumPFN that effectively captures the complex discussion structures and contextual dependencies unique to online forums, thereby enhancing financial market predictions. At its core, the Discussion Aggregator Module comprises two main components: a Topic Align Algorithm that classifies and reorganizes posts by topic, and a Multi-Scale 1D-Convolutional Path that integrates features at different scales. This design allows for precise modeling of the discussion flows and dynamics specific to Online Forum, maximizing the utilization of information obtained from online forums. We conduct experiments on directional prediction of the Nikkei 225 Volatility Index (Nikkei 225VI)—a representative VIX of Japanese stocks—using data from Yahoo Finance Message Boards, Japan’s largest online forum. The experimental results confirm that ForumPFN outperforms traditional baseline methods. Furthermore, through ablation studies, we demonstrate the effectiveness of each module in detail and explain the module’s operation via visualization of the attention matrix.

As the capabilities of smart sensing and mobile technologies continue to evolve and expand, storing diverse sensor data on smartphones and cloud servers becomes increasingly challenging. Effective data compression is crucial to alleviate these storage pressures. Compressed sensing (CS) offers a promising approach, but traditional CS methods often struggle with the unique characteristics of sensor data-like variability, dynamic changes, and different sampling rates-leading to slow processing and poor reconstruction quality. To address these issues, we developed Mob-ISTA-1DNet, an innovative CS framework that integrates deep learning with the iterative shrinkage-thresholding algorithm (ISTA) to adaptively compress and reconstruct smartphone sensor data. This framework is designed to manage the complexities of smartphone sensor data, ensuring high-quality reconstruction across diverse conditions. We developed a mobile application to collect data from 30 volunteers over one month, including accelerometer, gyroscope, barometer, and other sensor measurements. Comparative analysis reveals that Mob-ISTA-1DNet not only enhances reconstruction accuracy but also significantly reduces processing time, consistently outperforming other methods in various scenarios.

Forecasting rail congestion is crucial for efficient mobility in transport systems. We present rail congestion forecasting using reports from passengers collected through a transit application. Although reports from passengers have received attention from researchers, ensuring a sufficient volume of reports is challenging due to passenger's reluctance. The limited number of reports results in the sparsity of the congestion label, which can be an issue in building a stable prediction model. To address this issue, we propose a semi-supervised method for congestion forecasting for trains, or SURCONFORT. Our key idea is twofold: firstly, we adopt semi-supervised learning to leverage sparsely labeled data and many unlabeled data. Secondly, in order to complement the unlabeled data from nearby stations, we design a railway network-oriented graph and apply the graph to semi-supervised graph regularization. Empirical experiments with actual reporting data show that SURCONFORT improved the forecasting performance by 14.9% over state-of-the-art methods under the label sparsity.

Forecasting the number of visitors at a public event, termed event crowd forecasting (ECF), has recently garnered attention due to its social significance. Although existing ECF methods have pioneered successful feature design by considering event contents with contexts (e.g., weather, type of day, time), their scalability across different event types is limited due to the necessity of costly feature engineering. To address this issue, we propose a novel ECF framework, named EventOutlook. Based on our observation of various events, online event announcements indicate the factors that induce crowded events. Thus, we incorporate event announcements into ECF methods. To handle such unstructured data, which have no unified format among events, we leverage large language models (LLM) to extract crowding factors and embed them into an LLM-driven crowding-indicator feature (LCIF). Empirical experiments with real-world event data show that EventOutlook significantly improved ECF performance compared to state-of-the-art methods.

We propose and demonstrate the DisasterNeedFinder framework in order to provide appropriate information support for the Noto Peninsula Earthquake. In the event of a large-scale disaster, it is essential to accurately capture the ever-changing information needs.As a data-driven approach, we aim to pick up precise information needs at the site by integrally analyzing the location information of disaster victims and search information. The idea of assuming that the magnitude of information needs is not the volume of searches, but the degree of abnormalities in searches, enables an appropriate understanding of the information needs of the disaster victims in low population area. DNF has been continuously clarifying the information needs of disaster areas since the disaster strike, and has been recognized as a new approach to support disaster areas by being featured in the major Japanese media on several occasions. (For more information on this short paper, please refer to the Tsubouchi et al.[8].)

This study proposes a novel hybrid feedback approach called ”Boosting Feedback” to address the challenges of collecting correct data in machine learning research. Implicit feedback, derived from implicit behavioral logs, provides sufficient data quantity but may lack data quality due to various factors. Explicit feedback, obtained directly from users through surveys, offers high-quality data but is resource-intensive. The Boosting Feedback approach leverages implicit logs to augment the quantity of correct data from a single explicit feedback, doubling the available data by estimating opposite states from implicit logs. The method’s effectiveness is validated in actual recommendation experiment in the wild. Boosting Feedback offers a promising solution to improve data collection in machine learning research.

Predicting the density and flow of the crowd or traffic at a citywide level becomes possible by using the big data and cutting-edge AI technologies. It has been a very significant research topic with high social impact, which can be widely applied to emergency management, traffic regulation, and urban planning. In particular, by meshing a large urban area to a number of fine-grained mesh-grids, citywide crowd and traffic information in a continuous time period can be represented with 4D tensor (Timestep, Height, Width, Channel). Based on this idea, a series of methods have been proposed to address grid-based prediction for citywide crowd and traffic. In this study, we revisit the density and in-out flow prediction problem and publish a new aggregated human mobility dataset generated from a real-world smartphone application. Comparing with the existing ones, our dataset holds several advantages including large mesh-grid number, fine-grained mesh size, and high user sample. Towards this large-scale crowd dataset, we propose a novel deep learning model called DeepCrowd by designing pyramid architectures and high-dimensional attention mechanism based on Convolutional LSTM. Lastly, thorough and comprehensive performance evaluations are conducted to demonstrate the superiority of the proposed DeepCrowd comparing to multiple state-of-the-art methods.

Citywide crowd prediction can be of great importance for emergency management, traffic regulation, and urban planning. By meshing a large urban area to a number of fine-grained mesh-grids as illustrated in Fig. 1 , citywide crowd in a continuous time period can be represented with a four-dimensional tensor ${\mathbb{R}^{Timestep{\text{ }}p \times {\text{ }}Height{\text{ }} \times {\text{ }}Width{\text{ }} \times {\text{ }}Channel{\text{ }}}}$ in an analogous manner to video data, where each Timestep can be seen as one video frame, Height , Width is two-dimensional index for mesh-grids, and each Channel stores an aggregated scalar value for each mesh-grid. Specifically, given historical observations of crowd density and in-out flow x d = d 1 ,…,d t , xf = f 1 ,…, f t at timestamp t , we aim to build prediction models for the next-step density and in-out flow y d = d t +1, yf = f t +1, where y d means how many people will be in each mesh-grid at the next timestamp, and yf means how many people will flow into or out from each mesh-grid in next time interval. Al-though many deep models [1] – [6] have been proposed to address such tasks, their actual effects are still not well validated on large-scale and high-quality datasets. The datasets used in most of the works so far are originally generated based on taxi or bicycle trip data, which don’t cover and reflect the citywide crowd density and flow. Thus, we first publish new crowd flow data called BousaiTYO and BousaiOSA [7] . These new datasets are created using the GPS log data collected from a popular smartphone app of Yahoo! Japan Corporation, which can well reflect the real-world crowd flow in Tokyo and Osaka. As shown by Table 1 , our dataset has: (1) larger spatial area; (2) finer mesh size; (3) higher user sample.

In recent geospatial research, the importance of modeling and generating human mobility trajectories is rising. Whereas there are already plenty of feasible approaches applicable to geospatial sequence modeling itself, there seems to be room to improve with regard to evaluation, specifically about measuring the similarity between generated and reference trajectories. In this work, we propose a novel similarity measure, GEO-BLEU, which can be especially useful in the context of geospatial sequence modeling and generation. As the name suggests, this work is based on BLEU, one of the most popular measures used in machine translation research, while introducing spatial proximity to the idea of n-gram. We compare this measure with an established method, dynamic time warping, applying both measures to simple artificial sequences and examining differences in their characteristics.

Predicting the density and flow of the crowd at a citywide level is significant for city management. By meshing a large urban area to a number of fine-grained mesh-grids, citywide crowd and traffic information in a continuous time period can be represented with 4D tensor (Timestep, Height, Width, Channel). Based on this, we revisit the density and in-out flow prediction problem and publish a new aggregated human mobility dataset generated from a real-world smartphone application. Compared with the existing ones, our dataset has larger mesh-grid number, finer-grained mesh size, and higher user sample. Towards such kind of large-scale crowd dataset, we propose a novel deep learning model called DeepCrowd by designing pyramid architectures and high-dimensional attention mechanism based on Convolutional LSTM. Both the datasets and codes are made available at https://github.com/deepkashiwa20/DeepCrowd.

Systems that utilize and manage predictive models have become increasingly significant in industry. In the services offered by Yahoo! JAPAN, once a predictive model is utilized for recommendations, it is thrown away. Such models could however be reused for expanding the coverage of other recommendations. Here, our goal is to construct recommendation systems that expand the coverage of recommendations by effectively utilizing models which would otherwise be discarded. Another goal is to deploy such a recommendation system on real services and make practical use of it. In this paper, we describe a recommendation system that achieves these two goals by overcoming the challenges facing its deployment on real services. Specifically, we developed an optimization method that alleviates the psychological barrier against using the recommendation system and clarified the performance of our method in making real recommendations. An offline test and a large-scale online test on making real recommendations showed that our method substantially expands the coverage of recommendations. As a highlight of the results, our method made recommendations to 76.9 times more users at the same level of recommendation performance as the currently used recommendation system by the service. Overall, the results show that our method has a huge impact on services and can be applied to real recommendations.

Recurring outbreaks of COVID-19 have posed enduring effects on global society, which calls for a predictor of pandemic waves using various data with early availability. Existing prediction models that forecast the first outbreak wave using mobility data may not be applicable to the multiwave prediction, because the evidence in the USA and Japan has shown that mobility patterns across different waves exhibit varying relationships with fluctuations in infection cases. Therefore, to predict the multiwave pandemic, we propose a Social Awareness-Based Graph Neural Network (SAB-GNN) that considers the decay of symptom-related web search frequency to capture the changes in public awareness across multiple waves. Our model combines GNN and LSTM to model the complex relationships among urban districts, inter-district mobility patterns, web search history, and future COVID-19 infections. We train our model to predict future pandemic outbreaks in the Tokyo area using its mobility and web search data from April 2020 to May 2021 across four pandemic waves collected by Yahoo Japan Corporation under strict privacy protection rules. Results demonstrate our model outperforms state-of-the-art baselines such as ST-GNN, MPNN, and GraphLSTM. Though our model is not computationally expensive (only 3 layers and 10 hidden neurons), the proposed model enables public agencies to anticipate and prepare for future pandemic outbreaks.

Early crowd dynamics forecasting, such as one week in advance, plays an important role in risk-aware decision-making in urban regions such as congestion mitigation or crowd control for public safety. Although previous approaches have addressed crowd dynamics prediction, they have failed to deal with the scarcity of anomalous events, which results in a large model bias and could not quantify the number of visitors in anomalous crowd gathering. To provide an elaborate early forecast, we focus on the successive properties of importance weighting (IW) to penalize the anomalous data in terms of model bias; however, leveraging the concept of IW is challenging because dividing dataset into normal and abnormal sets is difficult. Motivated by these challenges, we propose CityOutlook, a novel forecasting model based on unbiased regression with importance-based reweighting. To make IW applicable to our approach, we design an anomaly-aware data annotation scheme by utilizing the heterogeneous property of mobility data to determine the data anomaly. We evaluate CityOutlook using the datasets of large-scale mobility and transit search logs. The experimental results show that CityOutlook outperforms the state-of-the-art models on crowd anomaly forecast, providing the same level accuracy in forecasting normal dynamics.

With the pandemic of COVID-19, indoor crowd density monitoring has become one of the most critical responsibilities of public space managers. Beacon placement optimization has been tackled as fundamental research work as the performance of crowd density monitoring highly depends on how BLE beacons are allocated. In this research, we propose a novel beacon placement optimization approach to incrementally place the beacon on the updated detection status adaptively in favor of Bayesian optimization, which can help to provide the optimal beacon placement. Our proposed method can optimize the beacon placement effectively to improve the signal coverage quality in the given environment and minimize human workload.

Recommendation systems help to predict user demand and improve the quality of services offered. While the performance of a recommendation system depends on the quality and quantity of feedback from users, the two major approaches to feedback sacrifice quality for quantity or vice versa; implicit feedback is more abundant but less reliable, while explicit feedback is more credible but harder to collect. Although a hybrid approach has the potential to combine the strengths of both kinds of feedback, the existing approaches using explicit feedback are not suitable for such a combination. In this study, we design a novel feedback suitable for the hybrid approach and use it improve the performance of a recommendation system. The system enables us to collect more varied and less biased feedback from users. It improves performance without requiring major changes to the inference model. It also provides a unique and rich source of information of the model itself. We demonstrate an application of Colorful Feedback showing how it can improve an existing recommendation model.

These days, many apps acquire location data as a way of estimating the user's behavior. As such, there are privacy concerns in using location data. In particular, users who are concerned about privacy may reduce the frequency of location acquisition or turn off the function, even though it degrades the quality of service. On the other hand, the only options available to users are yes-no or either-or ones such as "Always permit background acquisition" or "Permit only while using the app". For example, users who give permission to "Permit only while using the app" are themselves unable to understand how far their own veil of privacy will be lifted. That is, there are no metrics that can help users to understand the value of their own location data. How should the value of location data be determined? This study attempts to answer that question. The difficulty is that the value of a single point of location data depends on the context, such as how much other location data the app holds or when the location data was obtained. We propose a "Location YardStick" (LYS) that calculates the value of location information fairly in context. We confirmed that the LYS score is close to the user's expectations by comparing its results with those of a large online survey of 1300 people, and we conducted case studies in which we calculated LYS on location data acquired in various actual contexts.

Many different services acquire feedback from users, and many predictive models of users interests have been made for recommending items by using data acquired from different services. Here, although we would be able to get deep insights about a user's interest by connecting predictive models, interactions among services and potential synergies of their predictive models have been neglected. Moreover, in the era of big data, it should be relatively easy to develop a recommendation system that works across services by connecting predictive models and using existing technologies. However, the effect and conditions under which such the recommendation system works effectively have not been analyzed in detail.We have developed ModelRecycling FrameWork (MRFW) as an implementation of our idea of connecting predictive models to make recommendations across services. We experimentally evaluated the performance of MRFW by using an online questionnaire survey. The results show that MRFW significantly outperforms predictive models that make recommendations without considering the interaction between services or synergies of predictive models. We also clarified the conditions under which MRFW works effectively by comparing its performance with that of the individual models on which it is based in order to make use of the results of analysis for designing the method that connects predictive models.

The analysis of crowd flow in urban regions (urban dynamics) from GPS traces has been actively explored over the last decade. However, the existing prediction models assume that the population density in the analysis area is almost uniform, making it difficult to analyze fine-grained urban dynamics on a nationwide scale, where urban and rural areas coexist. In this paper, we propose a predictive model, called mixed-order Poisson regression (MOIRE), to capture changes in active populations nationwide by combining lower-order patterns and higher-order interaction effects. The proposed method utilizes multiple pieces of contextual information that greatly affect crowd flows (e.g., time-of-day, day-of-the-week, weather situation, holiday calendar information). We evaluated MOIRE on two massive GPS datasets gathered in urban regions at different scales. The results show that it has better predictive performance than the state-of-the- art method. Moreover, we implemented an anomaly detection system in urban dynamics for the whole nation of Japan in accordance with MOIRE specifications. This application enabled us to confirm MOIRE's performance intuitively.

Many researchers and companies have engaged in estimating users' interests so that an online shopping system can tell what he/she wants now. This paper tackles the next challenge in online shopping, i.e., predicting the times that users go shopping online. To predict the timing of online shopping, we focus on "wandering behavior" in web search activities and propose a "search wandering score" (SWS). Online shopping behavior can be categorized into three states: "wandering shop-ping", "focused shopping", and others. Wandering shopping is a state in which users make purchases in high SWS situations; focused shopping is a state in which users buy things in low SWS situations. Unlike previous studies, our work is based on an analysis of large-scale data containing shopping and search logs produced by approximately 200,000 users of a real web portal site for over a year. The results of an extensive evaluation show that our methodology can predict user's future shopping behavior types with 86% accuracy. This research is the first step towards understanding the relationship between users' mental states and their online shopping behavior.

Forecasting anomalies in urban areas is of great importance for the safety of people. In this paper, we propose Supervised-CityProphet (SCP), an anomaly score matching-based method towards accurate prediction of anomalous crowds. We re-formulate CityProphet as a regression model via data source association with mobility logs and transit search logs to leverage user's schedules and the actual number of visitors. We evaluate Supervised-CityProphet using the datasets of real mobility and transit search logs. Experimental results show that Supervised-CityProphet can predict anomalous crowds 1 week in advance more accurately than baselines.

Nowadays, GPS devices have increased explosively and produced huge amounts of trajectory data related to people's outgoing. Through those big location data, many researches aim to analyze human mobility for urban development, such as human movement prediction/modeling, POI (Point-Of-Interest) recommendation. However, trajectory data only contains timestamp and location information. The intention of human movement is not explicit so that it is hard to understand why people go to somewhere. The intention prior to the activity could be of great significance for analyzing and predicting human mobility, which has not been taken into consideration by the existing researches until the present. Thus, in this study, we propose a brand-new concept called human intentional mobility, aiming to employ intention information to predict people's outgoing. We carefully utilize user's search query to sense his intention as well as the intensity. For instance, if a user searches a certain POI for many times in a short period, it will represent a relatively high intention to go there. Then, to fully utilize this intention representation for predicting whether user will visit searched POI or not, we specially design Dual Sequential Interaction Network (DualSIN) as a novel and unique deep-learning model, which can effectively capture the sophisticated interactions among two kinds of sequential information (i.e., search sequence and mobility sequence) and typical categorical information (i.e., user attributes). Last, we evaluate our model on real-world dataset collected from Yahoo! Japan portal application, and demonstrate that it can achieve superior satisfactory performances to the-state-of-the-art models on multiple POI search queries.

Agent-based simulations, combined with large scale mobility data, have been an effective method for understanding urban scale human dynamics. However, collecting such large scale human mobility datasets are especially difficult during rare events (e.g., natural disasters), reducing the performance of agent-based simulations. To tackle this problem, we develop an agent-based model that can simulate urban dynamics during rare events by learning from other cities using inverse reinforcement learning. More specifically, in our framework, agents imitate real human-beings' travel behavior from areas where rare events have occurred in the past (source area) and produce synthetic people movement in different cities where such rare events have never occurred (target area). Our framework contains three main stages: 1) recovering the reward function, where the people's travel patterns and preferences are learned from the source areas; 2) transferring the model of the source area to the target areas; 3) simulating the people movement based on learned model in the target area. We apply our approach in various cities for both normal and rare situations using real-world GPS data collected from more than 1 million people in Japan, and show higher simulation performance than previous models.

Place embeddings generated from human mobility trajectories have become a popular method to understand the functionality of places, and could be applied as essential resources to various downstream tasks including land use classification and human mobility prediction. Place embeddings with high spatial resolution are desirable for many applications, however, downscaling the spatial resolution could degrade the quality of embeddings due to data sparsity, especially in less populated areas. Our proposed method addresses this issue by leveraging the hierarchical nature of spatial information, according to the local density of observed data points. We evaluated the effectiveness of our fine grained place embeddings via next place prediction tasks using real world trajectory data from 3 cities in Japan, and compared it with non-hierarchical baseline methods. Our technique of incorporating spatial hierarchical structure can complement and reinforce various other geospatial models using place embedding generation methods.

Point Of Interest (POI) categorization is to group POIs into several categories and make them easy-to-use in geospatial applications. Previous studies mainly used geospatial features, such as check-in sequences and satellite images, to group POIs into pre-defined rough categories. However, each POI has its own "atmosphere" beyond its geospatial features, which represents what kinds of people tend to visit it and how they spend their time there. This subtle atmosphere is important for users to decide whether to visit the POI, so considering it may be critical when providing commercial services, such as a property search service. In this paper, we propose a new POI categorization method that can capture the POI atmosphere by using user behavior on a web search engine. Our key observation is that the next queries of a search query about a POI tend to contain the user's purpose for visiting it. We harness this observation to train a neural encoder that maps POIs to continuous vectors (called embeddings) via next-query prediction with a deep structured semantic model (DSSM). Experimental results indicate that our method performs well for POI atmosphere categorization of parks as a case study. We believe that our method complements the existing POI categorization methods.

Unsupervised translation has become a popular task in natural language processing (NLP) due to difficulties in collecting large scale parallel datasets. In the urban computing field, place embeddings generated using human mobility patterns via recurrent neural networks are used to understand the functionality of urban areas. Translating place embeddings across cities allow us to transfer knowledge across cities, which may be used for various downstream tasks such as planning new store locations. Despite such advances, current methods fail to translate place embeddings across domains with different scales (e.g. Tokyo to Niigata), due to the straightforward adoption of neural machine translation (NMT) methods from NLP, where vocabulary sizes are similar across languages. We refer to this issue as the domain imbalance problem in unsupervised translation tasks. We address this problem by proposing an unsupervised translation method that translates embeddings by exploiting common hierarchical structures that exist across imbalanced domains. The effectiveness of our method is tested using place embeddings generated from mobile phone data in 6 Japanese cities of heterogeneous sizes. Validation using landuse data clarify that using hierarchical anchors improves the translation accuracy across imbalanced domains. Our method is agnostic to input data type, thus could be applied to unsupervised translation tasks in various fields in addition to linguistics and urban computing.

Deep neural networks are powerful for text sentiment analysis; however, in the real world, they cannot be used in situations where explanations are required owing to their black-box property. In response, we propose a novel neural network model called sentiment shift neural network (SSNN) that can explain the process of its sentiment analysis prediction in a way that humans find natural and agreeable. The SSNN has the following three interpretable layers: the word-level original sentiment layer, sentiment shift layer, and word-level contextual sentiment layer. Using these layers, the SSNN can explain the process of its document-level sentiment analysis results in a human-like way. Realizing the interpretability of these layers is a crucial problem. To realize this interpretability, we propose a novel learning strategy called joint sentiment propagation (JSP) learning. Using real textual datasets, we experimentally demonstrate that the proposed JSP learning is effective for improving the interpretability of layers in SSNN and that both the predictability and explanation ability of the SSNN are high.

Abstract Although deep neural networks are excellent for text sentiment analysis, their applications in real-world practice are occasionally limited owing to their black-box property. In this study, we propose a novel neural network model called contextual sentiment neural network (CSNN) model that can explain the process of its sentiment analysis prediction in a way that humans find natural and agreeable and can catch up the summary of the contents. The CSNN has the following interpretable layers: the word-level original sentiment layer, word-level sentiment shift layer, word-level global importance layer, word-level contextual sentiment layer, and concept-level contextual sentiment layer. Because of these layers, this network can explain the process of its document-level sentiment analysis results in a human-like way using these layers. Realizing the interpretability of each layer in the CSNN is a crucial problem in the development of this CSNN because the general back-propagation method cannot realize such interpretability. To realize this interpretability, we propose a novel learning strategy called initialization propagation (IP) learning. Using real textual datasets, we experimentally demonstrate that the proposed IP learning is effective for improving the interpretability of each layer in CSNN. We then experimentally demonstrate that the CSNN has both the high predictability and high explanation ability.

Extracting identifiable information from human trajectories is a fundamental task in many location-based services (LBS). However, various mobility patterns underlain in human trajectories are difficult to model by existing models. Moreover, we could hardly define a clear user set for user identification because the set of users are dynamic and changing everyday. Bearing these in mind, we apply a deep multiple instance learning method to handle the multimodal mobility patterns in a weak-supervised learning way, and address the dynamic user set problems via a pairwise loss with negative sampling. We utilize a multi-head attention mechanism to automatically extract multiple aspects and match the corresponding information between query trajectories and historical trajectories. Our method shows a good identification accuracy on three human GPS trajectory data sets comparing with baseline methods.

Large mobility datasets collected from various sources have allowed us to observe, analyze, predict and solve a wide range of important urban challenges. In particular, studies have generated place representations (or embeddings) from mobility patterns in a similar manner to word embeddings to better understand the functionality of different places within a city. However, studies have been limited to generating such representations of cities in an individual manner and has lacked an inter-city perspective, which has made it difficult to transfer the insights gained from the place representations across different cities. In this study, we attempt to bridge this research gap by treating cities and languages analogously. We apply methods developed for unsupervised machine language translation tasks to translate place representations across different cities. Real world mobility data collected from mobile phone users in 2 cities in Japan are used to test our place representation translation methods. Translated place representations are validated using landuse data, and results show that our methods were able to accurately translate place representations from one city to another.

Although deep neural networks are excellent for text sentiment analysis, their applications in real-world practice are occasionally limited owing to their black-box property. In response, we propose a novel neural network model called contextual sentiment neural network (CSNN) model that can explain the process of its sentiment analysis prediction in a way that humans find natural and agreeable. The CSNN has the following interpretable layers: the word-level original sentiment layer, word-level sentiment shift layer, word-level local contextual sentiment layer, word-level global importance layer, and word-level global contextual sentiment layer. Because of these layers, this network can explain the process of its document-level sentiment analysis results in a human-like way using these layers. Realizing the interpretability of each layer in the CSNN is a crucial problem in the development of this CSNN because the general back-propagation method cannot realize such interpretability. To realize this interpretability, we propose a novel learning strategy called initialization propagation (IP) learning. Using real textual datasets, we experimentally demonstrate that the proposed IP learning is effective for improving the interpretability of each layer in CSNN. We then experimentally demonstrate that both the predictability and explanation ability of the CSNN are high.

The limited attentional resource of users is a bottleneck to delivery of push notifications in today's mobile and ubiquitous computing environments. Adaptive mobile notification scheduling, which detects opportune timings based on mobile sensing and machine learning, has been proposed as a way of alleviating this problem. However, it is still not clear if such adaptive notifications are effective in a large-scale product deployment with real-world situations and configurations, such as users' context changes, personalized content in notifications, and sudden external factors that users commonly experience (such as breaking news). In this paper, we construct a new interruptibility estimation and adaptive notification scheduling with redesigned technical components. From the deploy study of the system to the real product stack of Yahoo! JAPAN Android application and evaluation with 382,518 users for 28 days, we confirmed several significant results, including the maximum 60.7% increase in the users' click rate, 10 times more gain compared to the previous system, significantly better gain in the personalized notification content, and unexpectedly better performance in a situation with exceptional breaking news notifications. With these results, the proposed system has officially been deployed and enabled to all the users of Yahoo! JAPAN product environment where more than 10 million Android app users are enjoying its benefit.

Predicting the evacuation decisions of individuals before the disaster strikes is crucial for planning first response strategies. In addition to the studies on post-disaster analysis of evacuation behavior, there are various works that attempt to predict the evacuation decisions beforehand. Most of these predictive methods, however, require real time location data for calibration, which are becoming much harder to obtain due to the rising privacy concerns. Meanwhile, web search queries of anonymous users have been collected by web companies. Although such data raise less privacy concerns, they have been under-utilized for various applications. In this study, we investigate whether web search data observed prior to the disaster can be used to predict the evacuation decisions. More specifically, we utilize a session-based query encoder that learns the representations of each user's web search behavior prior to evacuation. Our proposed approach is empirically tested using web search data collected from users affected by a major flood in Japan. Results are validated using location data collected from mobile phones of the same set of users as ground truth. We show that evacuation decisions can be accurately predicted (84%) using only the users' pre-disaster web search data as input. This study proposes an alternative method for evacuation prediction that does not require highly sensitive location data, which can assist local governments to prepare effective first response strategies.

We developed a simulation program for trading Nikkei stock index options and verifies the validity of the volatility index (VIX) prediction model proposed by Suwa et al. (2017). We simulated two cases from 18 Nov. 2014 to 29 Jun. 2016. One case involved a benchmark of trading every day during that period and the other was in accordance with the buy/sell/hold instructions of Suwa et al.'s VIX prediction model. When using the call option butterfly spread according to their model's instructions, profit increased from -3,926 to 536 yen. When using the put option butterfly spread according to their model's instructions, profit increased from -4,818 to -799 yen. Therefore, Suwa et al.'s VIX prediction model is effective.

Recently, the frequency and intensity of weather-related disasters are increasing and are becoming more ubiquitous, often devastating vulnerable rural areas. To prepare for speedy and effective first response, we need a flood detection method that works much faster and is able to cover a wider area compared to conventional methods that use CCTV cameras and low cost sensors, which are costly to distribute ubiquitously in all areas with possible flood threats. With the spread of mobile phones, we are able to obtain real time anonymized location information of individuals in a ubiquitous, low cost, and a continuous manner from users that have agreed to provide their location data for disaster relief purposes. Here we propose a novel method that infers flooded areas in real time by detecting anomalous behaviors of individuals using mobile phone location data. We are motivated in applying our method to rural areas that are costly to cover using cameras and sensors. To overcome the sparseness of mobile phone location signals in such rural areas, our method combines mobile phone location data with terrain information including the digital elevation model and river trajectory data. We evaluated our method using real world data from 2 severe floods in the rural parts of Japan and verified that our method is more accurate and has numerous advantages compared to conventional methods. This work presents the potential use of mobile phone data as a complementary, if not an alternative method for flood detection especially in rural areas.

This paper is the first work to replicate and simulate urban dynamics by learning individuals' decision-making processes and creating human-like agents from GPS data. We develop a novel agent model by learning from historical data via reinforcement learning techniques. We test our methodology in different scenarios at the citywide level using real world smartphone GPS data. Simulation results show that our agents can successfully learn and generate human-like travel activities. Furthermore, the performance of synthetic urban dynamics significantly outperforms existing methods.

Predicting behaviors of a population from location-oriented log data from smartphones, i.e., urban population dynamics, has become more common in mobile and pervasive computing. A bilinear representation approach has been proposed to improve the prediction accuracy of urban population dynamics by adding contexts such as geographical information and day of the week. However, this approach has a strong limitation in that additional contexts can not be directly utilized in this representation with a unified manner. To resolve this issue, we propose a new predictive model for urban population dynamics based on multilinear Poisson regression so as to handle multiple contexts in a systematic manner. The model is parameterized using a tensor and can be optimized by using an efficient convex optimization with a sequence of matrix parameter optimizations. An empirical evaluation with large-scale smartphone location data showed that our model outperforms conventional approaches.

We propose a method of predicting an increase in the Nikkei VI index by analyzing social media based on the premise that investor sentiment is posted on social media. Since the VI index expresses the fear of investors, it is a closely related index to the risk of depression. Therefore, the VI index is an important indicator as an instrument for investment judgment. To predict the increase in the VI index more accurately, we divide messages by topic models specific to social media of stock trading and predict such the increase by machine learning using those topics. As a result of leave-one-day-out cross-validation, precision of our method was 0.45. We also found that the daily fluctuation in the VI index and the number of messages are as effective as feature quantities as the topic-posting frequency.

This research aims to develop a method for maximizing the accuracy of next place prediction (NPP) in places that are unfamiliar to each mobile phone users. NPP is a problem of predicting the next place of the user given his/her current place and current time. In places that are unfamiliar to the person, it is difficult to predict the next place based on the person's historical location data because there are just a few or no data in such places for each user. Furthermore, it is also difficult to rely on the regularity of human mobility because tourists' mobility is easily affected by many external factors, such as weather. Our research aims to solve the difficulties in NPP in unfamiliar places by focusing on contextual factors such as weather, transportation means, place of residence, and time.

Large scale disasters cause severe social disorder and trigger mass evacuation activities. Managing the evacuation shelters efficiently is crucial for disaster management. Kumamoto prefecture, Japan, was hit by an enormous (Magnitude 7.3) earthquake on 16th of April, 2016. As a result, more than 10,000 buildings were severely damaged and over 100,000 people had to evacuate from their homes. After the earthquake, it took the decision makers several days to grasp the locations where people were evacuating, which delayed of distribution of supply and rescue. This situation was made even more complex since some people evacuated to places that were not designated as evacuation shelters. Conventional methods for grasping evacuation hotspots require on-foot field surveys that take time and are difficult to execute right after the hazard in the confusion.

Mid-level disasters that frequently occur, such as typhoons and earthquakes, heavily affect human activities in urban areas by causing severe congestion and economic loss. Predicting the irregular movement of individuals following such disasters is crucial for managing urban systems. Past survey results show that mid-level disasters do not force many individuals to evacuate away from their homes, but do cause irregular movement by significantly delaying the movement timings, resulting in severe congestion in urban transportation. We propose a novel method that predicts such irregularity of individuals' movements in several mid-level disasters using various types of features including the victims' usual movement patterns, disaster information, and geospatial information of victims' locations. Using real GPS data of 1 million people in Tokyo, we show that our method can predict mobility delay with high accuracy,

This paper describes a pedestrian population trend estimation method using location data of smartphone users. This technique is intended to be an alternative to traffic censuses using tally counters. Traffic censuses using tally counters are still commonly used to survey the number of pedestrians despite their cost and limitations in area and time.

An on-demand bus is like a shared taxi that operates only when riders want to travel between the origin and destination locations. It offers many advantages over fixed-route buses, but the riders are bothered by the need to tediously enter such data as origins, destinations, and deadlines. A location recommendation system that predicts such data would help riders during the reservation process and help target potential riders when buses are idle. In this paper, a general and scalable framework for such location recommendation algorithms is presented. It is based on users' location histories and spatio-temporal correlations among the locations by combining prediction methods of the collaborative filtering algorithms, which are widely used in e-commerce, with a popular method in data mining called link propagation. Experiments on real-world data demonstrate that the accuracy of recommendations with the spatio-temporal information is better than those without.