Since its launch in November 2015, the LIFULL HOME'S dataset has opened up new possibilities for data utilization in the real estate field. The dataset has been used for research activities not only in Japan but also internationally, including in Asia, the U.S., and Europe, and more than approximately 150 academic presentations have been made. In this presentation, we will analyze research usage trends over the approximately 9 years since the dataset was first provided, and show the evolution of artificial intelligence research in the real estate field and its impact. It will also demonstrate the diversity of research fields using the dataset and discuss the potential and challenges of real estate data in the future.

One important task for a rental apartment management company is the assessment and setting of the rent of vacant apartments. Not only do the features of the apartment need to be taken into account, but also the duration of the vacancy. Setting a high rent may result in an unacceptably long vacancy while setting a low rent will negatively affect the cash flow. It therefore becomes necessary to jointly consider rent and vacancy duration. We use both vacancy listings from public websites and proprietary data from one of Japan's largest management companies of one-room rental apartments to estimate the relationship between vacancy duration and rent. We use a combination of a LightGBM model and a Bayesian hierarchical survival data model to predict the vacancy duration using the characteristics and advertised rent of the apartment. The results show that the vacancy duration can be predicted with good precision. The evaluation metrics are as good as can be expected from this type of data, and the predictions are well-calibrated. We also see that the vacancy duration varies depending on, for example, region and season.

In the medical industry, location selection is crucial for business success, yet many decisions are still based on experience and intuition. For clinics, with limited access to sales data and available samples, implementing data-driven decision-making for small sample sizes remains a significant challenge. This study proposes a sales prediction method combining geographic information with characteristics extracted from satellite imagery using GPT-4o, such as urbanization levels and building ratios. We adopted Support Vector Regression (SVR) to achieve accurate predictions while preventing overfitting with small samples. Furthermore, we developed a data-driven approach that optimizes prediction model variables through correlation analysis for confounding factor exclusion and feature selection algorithms. Experimental results confirmed that regional characteristics enable accurate sales predictions with limited datasets. We verified that characteristics derived from satellite imagery improved prediction accuracy compared to baseline methods using only geographic information. This methodology shows promise for location selection across various industries beyond the medical field.

The age of real estate properties is used not only for property value estimation but also for assessing disaster risk. However, obtaining accurate building age information remains challenging. To address this issue, it is necessary to develop techniques for predicting building age using exterior images of real estate properties. In this study, we aim to evaluate multiple deep learning architectures for such prediction. Using approximately 9.5 million images from the LIFULL HOME'S dataset, we conducted experiments to analyze the prediction accuracy of five architectures, including, Vision Transformer (ViT), VGG16, and ResNet101\_V2. In addition, we compared prediction accuracy across different regions. Experimental results indicated that ViT achieved the best performance with a mean absolute error of 2.856, confirming its high accuracy and generalizability regardless of regional differences.

When extracting the features of a real estate property from its floor plan, if the floor plan includes a directional symbol, the orientation angle indicated by the symbol can be estimated, allowing for the prediction of orientation-related features. This paper proposes a method for estimating the orientation angle by detecting and analyzing the directional symbol in the floor plan. The method combines results from multiple image processing techniques. The paper also reports the verification results, achieving an average error of 3.68º. Additionally, it describes a method for predicting the property's main daylight-facing surfaces based on the estimated orientation angle and the window orientations extracted from the floor plan.

Understanding the history of land transactions and inheritance is a crucial initial step and an indispensable process in sustainable urban development, municipal urban policies, and infrastructure development. However, the full registry of real estate or land ledgers used before the current registration system, which documents these transaction and inheritance histories, remains non-digitized document media, requiring significant human and economic costs for interpretation. Against this backdrop, this study aims to develop an image recognition model for layout detection of complex table structures found in the full real estate registry and land ledgers. The image recognition model was constructed to enable comparative evaluation between CNN-based models, such as Faster-RCNN, which includes line processing, and YOLO-based models. I have shown that while all of the models have high detection accuracy, the YOLO-based model is superior in terms of learning speed and prediction accuracy.

How are humans deceived by fake news? To elucidate the neural mechanisms underlying susceptibility to fake news, this study investigated brain regions associated with individual differences in judgments of real versus fake news. For this purpose, we conducted a psychological experiment where participants evaluated the authenticity of various news articles, including those synthesized by AI. Additionally, we measured participants' brain responses during movie viewing to assess the intrinsic response properties of individual brains. We found that individual differences in judgment patterns were correlated with those in brain-response patterns within several brain regions. Among these, the precuneus showed a correlation only with judgments of fake news, not real news. These findings suggest that the precuneus plays a critical role in forming an individual's susceptibility of fake news.

When we believe something, the belief is usually based on available evidence. If the evidence supporting the belief is proven false, or if contradictory evidence is presented, it would be reasonable to reject the belief. However, this does not always happen in practice. People may hold onto their existing beliefs by attempting to justify evidence that is objectively false, or by ignoring contradictory evidence. Such apparently irrational responses to evidence are particularly common within filter bubble and echo chambers. This presentation aims to explore how people come to believe disseminated misinformation and disinformation, and how these beliefs are maintained and reinforced. It focuses specifically on reactions to evidence in belief formation within filter bubble and echo chambers, seeking to uncover the mechanisms behind these processes from a philosophical perspective. Additionally, it examines how a philosophical understanding of these phenomena can contribute to effective strategies for addressing misinformation and disinformation that may be spread by AI.

In cancer treatment, peer support among patients with the same illness plays a crucial role in acquiring cancer-related information and fostering empowerment. However, it is also important to recognize the risk of misinformation spreading through peer support networks. Unlike face-to-face interactions, the acceptance of testimonies on social media lacks clear conditions for ensuring information reliability, requiring special consideration. Moreover, in self-help-oriented peer support on social media, the credibility of testimonies from fellow patients tends to be amplified, which may increase the risk of accepting misinformation, especially within the context of empowerment. Given that existing misinformation regulations on social media platforms are often insufficient, it is necessary to explore measures to prevent patients from being adversely affected by misinformation through peer support.

This paper examines the impact of an information environment rife with disinformation and misinformation on individual autonomy, drawing on the philosophical framework of "relational autonomy". Several AI ethics guidelines uphold respect for autonomy, or the ability to make judgments and take actions based on one's own will and values, a core ethical principle. However, the spread of dis-/mis-information threatens individual autonomy. Theories of relational autonomy hold that autonomy involves three dimensions: (i) self-governance; (ii) self-determination; and (iii) self-authorization. Current measures against dis-/mis-information primarily focus on educational initiatives aimed at strengthening self-governance. However, this new information environment can also threaten individual autonomy in the dimensions of self-determination and self-authorization, necessitating more multifaceted countermeasures. This presentation specifically argues for the need for measures to address what I call "literacy burden," "information skepticism," and "indifference to truth."

Crime forecasting is crucial for maintaining societal safety. This study interprets the criminological theory of Repeat and Near Repeat Victimization (RNRV) using a deep learning model, the Hierarchical Graph Attention Network (HGAT), to forecast occurrence of traffic, violent, sexual, income-generating and child-target crimes nationwide (6,220,112 sections) at a one-block level (0.25 km × 0.25 km) one week in advance. The Predictive Accuracy Index and Root Mean Square Error scores were used as forecast performance index. Result shows that the HGAT model achieved a significant improvement in forecast accuracy compared to state-of-the-art forecasting techniques. The HGAT model continued to forecast 1,724 municipalities with high accuracy for 14 consecutive weeks for the above five types of crimes. Because of its improved forecast accuracy and generalizability, the HGAT model will become the new state-of-the-art model for crime forecasting.

In the field of online restaurant reviews, examining how regional differences and cultural backgrounds influence consumer behavior has emerged as a critical research topic. However, there has been little statistical discussion regarding the psychological and cultural mechanisms that lead to variations in star ratings across different regions and discrepancies between star ratings and review content. This study compares large-scale restaurant reviews from Tokyo and New York to reveal how cultural and psychological biases shape star ratings. We find that Tokyo reviewers often exhibit moderate evaluations, while New York reviewers tend to provide more polarized feedback. By analyzing aspects such as taste, price, ambiance, and service, we demonstrate that star ratings can deviate from underlying sentiments, indicating the influence of social norms. These findings highlight the importance of considering cultural contexts in the design and interpretation of global review platforms.

With the advancement of information recommendation systems, users' news consumption has increasingly become biased toward specific perspectives, exacerbating social issues such as polarization and division. This has raised concerns about the deterioration of “informational health”, where users originally maintained a balanced intake of diverse information. To address this, it is essential to understand user information consumption tendencies from the perspective of diversity. In this study, we analyze a large-scale dataset comprising news articles and user browsing logs. We employ BERTopic to convert news articles into topic distributions and evaluate users' news consumption diversity by applying Jensen-Shannon (JS) divergence to the traditionally used GS-score. This approach enables a refined assessment of users' browsing tendencies. Our results demonstrate that the proposed method outperforms the conventional GS-score in evaluating diversity. Furthermore, through topic-level analysis, we provide a more granular and detailed understanding of the relationship between news diversity and user browsing tendencies.

With the rise of digital media, news consumption has become increasingly diversified. While online platforms offer personalized content, concerns about filter bubbles persist. In response, traditional newspapers have adapted by providing both fixed-edition (morning/evening) and user-selected content. Using data from Nikkei Inc., we identified distinct reading patterns. Fixed-edition readers tend to browse quickly and engage with a wide range of topics, whereas user-selected readers spend more time per article but focus on a narrower set of content. We also analyzed beginner-oriented articles designed to offer diverse perspectives. Our findings indicate that fixed-edition readers rarely engage with these articles, while user-selected readers actively seek them out. This limited reach suggests inefficiencies in the distribution of such content. Optimizing recommendations for fixed-edition readers could enhance news delivery, ensuring that specialized content reaches a broader audience. Such improvements may contribute to a more balanced information environment in the digital age.

The role of social media in elections has become increasingly significant. In this study, we analyze data from the Tokyo gubernatorial election of July 2024 using the social media platform X. We specifically compare and examine the characteristics of the seven main candidates by integrating text analysis and network analysis. Our findings revealed substantial differences in social media strategies, post content, and network positioning among the candidates. These findings are valuable for understanding the current state of social media strategies in elections.

Constructing meta-datasets that capture the relationships between action labels across different datasets typically requires extensive human annotations. To address this challenge, we propose an approach that automates this process by leveraging common sense knowledge graphs. Specifically, our method aims to extract equivalence relationships between action labels by linking labels from multiple action recognition datasets to the corresponding nodes in common sense knowledge graphs. In this study, we focus on six widely used datasets. To evaluate the effectiveness of our approach, we use precision, recall, and F-measure with the manually constructed MetaVD serving as the ground truth. Additionally, we introduce a simplified evaluation metric, the Average Transferred Precision Gain to assess the performance. Our preliminary results demonstrate the potential of the proposed method for automating the equivalence relationship extraction. Future work will extend this approach to identify additional relationships, such as similarity and hierarchical (is-a) relationships, to enable a more comprehensive automated meta-dataset construction.

In recent years, the CtoC market has rapidly expanded, leading to increased peer-to-peer transactions. However, fraudulent activities on these platforms have become a critical issue, particularly the manipulation of ratings through the misuse of the "favorite" function. This study analyzes data from the handmade CtoC platform "minne" to identify the behavioral characteristics of rate-manipulating users and their impact on the market. Users who meet specific behavioral criteria are defined as rate-manipulating users, and network analysis is employed to extract community structures. The results reveal that rate-manipulating users are concentrated within specific communities, with their influence being largely confined to these groups. Furthermore, a time-series analysis using Sankey diagrams demonstrates that rate-manipulating users tend to remain within the same community over time. This study aims to clarify the nature and characteristics of fraudulent behavior and provide guidelines for enhancing the integrity of CtoC platforms.

In real soccer, players predict the movements of the ball and opponents to construct effective offensive and defensive plays. Similarly, in the RoboCup Soccer Simulation 2D League, which is the focus of this study, such predictions are necessary to take optimal actions. This research aims to predict ball movements and opponent players’ movements, particularly those involved in goal-scoring situations, using LSTM models in the RoboCup Soccer. Numerical experiments demonstrated that predictions for one cycle ahead could be made with high accuracy.

For reinforcement learning to acquire an appropriate policy, the designer must predate a properly designed reward function. However, in complex problem settings, the burden of designing such a reward function increases significantly. An improperly designed reward function can lead the agent to learn policies that deviate from the designer’s intent, becoming a bottleneck for applying reinforcement learning in real-world scenarios.In this study, we propose an approach to address this challenge by labeling the trajectories that the reinforcement learning agent transitions through during the learning process as successes or failures. We train a discriminator in parallel with the reinforcement learning agent to distinguish between these labeled trajectories and use its output as an additional reward. The discriminator outputs the probability of a given state being labeled as successful based on the states encountered by the agent during its interaction with the environment. By feeding this output back to the agent as an additional reward, we aim to reduce the burden of reward design while enabling more efficient learning.

Reservoir computing is a machine learning framework that exploits nonlinear dynamics, exhibiting significant computational capabilities. One of the defining characteristics of reservoir computing is that only linear output, given by a linear combination of reservoir variables, is trained. Inspired by recent mathematical studies of generalized synchronization, we propose a novel reservoir computing framework with a generalized readout, including a nonlinear combination of reservoir variables. Learning prediction tasks can be formulated as an approximation problem of a target map that provides true prediction values. Analysis of the map suggests an interpretation that the linear readout corresponds to a linearization of the map, and further that the generalized readout corresponds to a higher-order approximation of the map. Numerical study shows that introducing a generalized readout, corresponding to the quadratic and cubic approximation of the map, leads to a significant improvement in accuracy and an unexpected enhancement in robustness in the short- and long-term prediction of Lorenz and Rössler chaos. Towards applications of physical reservoir computing, we particularly focus on how the generalized readout effectively exploits low-dimensional reservoir dynamics.

Time series anomaly detection remains a crucial topic in data mining. Recent studies indicate that methods based on identifying anomaly subsequences, known as time series discords, are still effective for univariate time series anomaly detection. Despite the frequent occurrence of missing data in real-world time series, there are no established methods for discovering discords in data with missing values. In this study, we propose a novel approach that efficiently identifies discord candidates by simultaneously performing a new metric that can be calculated for subsequences with missing values and a nearest-neighbor-based time series imputation. Our experimental results demonstrate that this method achieves superior anomaly detection accuracy compared to other distance-based anomaly detection approaches in the presence of missing values.

Anomaly detection in time series data is essential across various industries, such as communication networks. Recently, Autoencoder (AE) has been widely used for anomaly detection. However, the training of AE heavily depends on the geometric features of input data, which makes it challenging to train AE for time series data with complicated periodic patterns. In this paper, we propose an Autoencoder that uses the frequencies of time series data, which we call AE-FED(AE with Frequency-Encoded Decoder). We prove that AE-FED can reconstruct the input time series data with periods. Experimental results on synthetic time series datasets demonstrate that AE-FED outperforms existing approaches in detecting anomalies in periodic time series data, achieving the highest AUC scores.

Coronavirus molecular pathways are activated upon coronaviral infection. An artificial intelligence (AI) approach based on machine learning was utilized to develop models with images of the coronavirus pathogenesis pathway to predict the activation states of the coronavirus molecular pathways. Among more than 100,000 analyses and datasets in the Ingenuity Pathway Analysis (IPA) database, 106 analyses and 106 datasets were involved in severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) as of 2021. A total of 22 analyses in SARS-CoV-2 infected lung adenocarcinoma (LUAD) were identified to be related to the terms “human” and “SARS coronavirus 2” in the database. The coronavirus pathogenesis pathway was activated in SARS-CoV-2-infected LUAD cells. The prediction model was developed in Python using images of coronavirus pathogenesis pathways in different conditions. The prediction model of activation states of coronavirus molecular pathways may be useful for treatment identification.

In real-world reinforcement learning applications, communication delays or sensor failures often cause frame dropping, in which the agent cannot receive dropped states and rewards. Because frame dropping degrades the agent's performance, the decision transformer under random frame dropping (DeFog) was developed by incorporating additional mechanisms into the decision transformer to tackle frame dropping. Although DeFog can mitigate performance degradation in frame-dropping environments, since DeFog is an offline learning method, it is difficult to select appropriate actions for the states not included in the dataset. In this study, we propose OnDeFog, which integrates the mechanisms in DeFog with the online decision transformer (ODT), an online reinforcement learning method that learns policy through direct interaction with the environment. The experimental results show that OnDeFog outperforms ODT in environments with a high dropping frame rate and is superior to DeFog with datasets containing a large amount of low-reward data.

Categorical sequential data often contain redundant information and become difficult to interpret visually when their variations are complex. In this study, we assume that the observed data follow different multinomial distributions across multiple regimes and attempt to detect regime breakpoints using dynamic programming based on maximum likelihood estimation. While a similar method, the Pruned Exact Linear Time (commonly known as the PELT method), exists, it is primarily designed for one-dimensional data; when handling multidimensional data such as categorical data, the objective function must be expanded according to the number of dimensions. The proposed method is expected to contribute to the understanding of sequential variations and enhance interpretability in categorical data analysis, serving as a useful analytical foundation, especially for datasets with complex variation characteristics. In the evaluation experiments, we investigate appropriate evaluation criteria (AIC, BIC, MDL, and the elbow method) to mitigate model overfitting, clarifying the tendencies of model selection for each criterion. In particular, we focus on the L method, an automated technique for the elbow method, and examine its behavior.

The Traveling Salesman Problem (TSP) is a combinatorial optimization problem that seeks the shortest route visiting each given city exactly once. As an NP-hard problem, finding the optimal solution becomes infeasible for large instances within a reasonable time. Heuristic approaches are widely used to search for an approximate solution, but computational time remains a challenge, especially in real-time applications. Recently, deep reinforcement learning (DRL) has gained attention for solving large-scale TSP efficiently. H-TSP, a DRL-based method, improves both search time and accuracy by dividing large TSP instances into smaller subproblems. However, H-TSP assumes a uniform node distribution, limiting its applicability to real-world problems with clustered structures, such as logistics and store networks. This study proposes an improved DRL-based approach incorporating an enhanced k-NN graph and a partial initial path strategy to address clustered TSP instances. Comparative experiments with H-TSP demonstrate the effectiveness of the proposed method.

In counterfactual machine learning, off-policy evaluation (OPE) aims to estimate the true performance of decision-making policies using logged data. Traditional estimators evaluate policy performance based solely on rewards directly induced by the policy. However, in real-world scenarios like e-commerce recommendation systems, users often take actions (e.g., purchases) outside the recommended list, leading to unaccounted rewards. To address this, estimators must evaluate performance beyond the recommended items.Existing methods struggle as action spaces grow, with accuracy deteriorating in large-scale environments. For example, e-commerce platforms may have action spaces ranging from thousands to millions of items, requiring robust methods to maintain accuracy. This study proposes a novel estimator extending the OffCEM framework to mitigate accuracy degradation, achieving high performance in large action spaces. Theoretical analysis and experiments show that the proposed method outperforms previous estimators, delivering enhanced accuracy in large-scale settings.

Supervised learning is applied in various domains. However, human instructional errors are inevitable especially when large amount of expert instruction is required. Standard supervised learning with noisy labels suffers from overfitting, which degrades generalization performance of the model. In this study, we analyze the learning process of deep learning models under label noise and investigate how to correct (relabel) erroneous labels for learning with noisy labels. We find that it is important to perform relabeling at a relatively high learning rate for appropriate number of iterations. By incorporating these properties, we demonstrated that classification accuracy can be maintained under label noise.

Deep learning models achieve strong generalization performance but often degrade under distribution shifts between training and test data. Test-time adaptation (TTA) has emerged as a promising approach to mitigate this issue by adapting models during inference using unlabeled test data. A widely used TTA method, Test Entropy Minimization (Tent), improves robustness by minimizing the entropy of output predictions. However, its adaptation is limited to updating only the affine parameters of batch normalization, restricting its ability to handle complex distribution shifts. To address this limitation, we propose integrating Funnel Activation (FReLU), an activation function with an adjustable receptive field, into Tent to enhance its adaptability. Experimental results demonstrate that our method outperforms conventional approaches, achieving improved performance under distribution shifts.

The purpose of this study was to clarify the characteristics of the weights in each layer of a convolutional neural network, focusing on the relationship between the weights and the depth of the layers. We performed weight replacement and linear combination experiments using CNNs with the same structure trained on different datasets. We also analyzed the effect of retraining the Batch Normalization layer. Experimental results showed that the weights of the shallow layer are highly dependent on the dataset, and their characteristics appear as differences in the output distribution, which can be handled by appropriate normalization. On the other hand, the weights of the deeper layers showed a linear similarity between the layers and a relatively small dependence on the dataset.

Large Language Models (LLMs) are being utilized for various tasks in software development, including code generation and code completion. While pre-trained and fine-tuned coding-specialized LLMs like Code Llama achieve advanced code generation capabilities, they have limitations in adapting to specific development methodologies and updates in modern programming languages. In this research, to address this challenge, we conducted experiments and evaluations of a RAG system using hypothetical document embeddings. Through this verification, we compared the proposed method with conventional RAG systems and demonstrated the utility and limitations of this new approach in complementing the constraints of coding-specialized models. This approach presents a promising methodology that can contribute to the development of more flexible and accurate code generation systems.

Deep learning, and in particular Transformer, has been successful in the fields of computer vision and natural language processing, where unstructured data is predominant. In recent years, there has been a transformation of structured tabular data into unstructured strings, which has been applied against Transformer, which is used in large language models. In this case, tabular data consists of a sequence of label names and their value pairs, which are a mixture of text and numerical values. However, the computational expense of these methods on large scales hinders their practical application. This study proposes a novel Decoder-Based Tabular Transformer, utilising sentence embedding and piecewise linear embedding of numerical values, to address this challenge. The efficacy of this approach is validated through its application to tabular data comprising both sentences and numerical values. The proposed model demonstrated a correct response rate of 0.856 on the US annual income prediction benchmark set of the UC Irvine Repository, which is comparable to the performance of the existing method (0.876). Future work should compare the proposed model with previous studies that utilised methods to convert tabular data to strings.

Tabular data is widely used in various fields such as healthcare and finance, containing multiple data types, including numerical, categorical, and textual values. Proper tokenization and embedding methods are essential, especially when both labels and values coexist. This study focuses on Masked Prediction techniques for applying Transformers to variable-length tabular data, comparing two masking strategies: masking labels and values together versus masking them independently. We conducted evaluation experiments using the Adult Dataset from the UC Irvine Repository, performing transfer learning and fine-tuning after pretraining. The results showed that masking labels and values together achieved a higher AUROC score in transfer learning, while independent masking led to lower accuracy. However, in fine-tuning, both methods performed similarly with no significant difference. These findings suggest that independent masking is not advantageous for transfer learning. Future work should explore other datasets and different masking probabilities for a more comprehensive evaluation.

This paper proposes a novel high-dimensional Bayesian optimization framework that combines a linear bandit approach for direction selection with Gaussian process-based one-dimensional searches. Standard Bayesian optimization often faces severe computational challenges in dimensions exceeding ten, due to the “curse of dimensionality” and the difficulty of directly optimizing high-dimensional acquisition functions. To mitigate these issues, our method—Directional Bandit BO—automatically learns effective search directions via a linear bandit (LinUCB) formulation, which treats each potential direction vector as an arm. By defining the bandit’s reward based on the discrepancy between the Gaussian process prediction and the true function value, the algorithm actively explores under-modeled directions. Once a direction is chosen, a one-dimensional Bayesian optimization step refines the search along that axis. We demonstrate the effectiveness of this approach on a 50-dimensional Styblinski-Tang function with only five influential dimensions. Experimental results show faster convergence compared to other high-dimensional methods, validating our proposed strategy.

In recent years, corporate management has placed significant emphasis on ESG (Environmental, Social, and Governance) indicators, in addition to financial metrics. Efforts related to ESG need to be implemented by companies over the long term. However, the impact of ESG initiatives on a company's finances and management is often indirect and complex, making data analysis essential for estimating their effects. From the statistical perspecitve, while indicators are typically obtained on an annual basis from companies, the wide range of indicators makes it challenging to provide accurate analyses. Therefore, this research aims to continuously improve corporate management while enhancing the accuracy of ESG-based analyses by developing a monitoring tool for ESG and financial indicators. This tool offers a means to improve the analytical model by comparing the analysis and predictive simulations between ESG and financial indicators with actual measured values.