The sixth edition of the Intelligent Cross-Data Analysis and Retrieval (ICDAR) workshop continues to serve as a forum for researchers and practitioners addressing the integration, analysis, and retrieval of heterogeneous data sources. While individual modalities such as wearable sensors, lifelogging cameras, and social media have been well studied, analyzing cross-data that incorporates multiple perspectives remains a crucial yet challenging task for advancing human-centered applications. In 2025, the workshop received 19 submissions, of which 7 were accepted following a careful peer-review process, resulting in an acceptance rate of 37%. The accepted papers covered a wide range of topics, including zero-shot composed image retrieval, vision-language scene understanding, adaptive modality fusion, lightweight fine-tuning with truncated SVD, and real-world federated split learning on mobile devices. By fostering interdisciplinary collaboration across domains such as well-being, disaster mitigation, mobility, food computing, and smart cities, the workshop continues to highlight emerging challenges and solutions for building intelligent, sustainable, and human-centric systems driven by cross-modal and multimodal data analytics.

Recent advances in video captioning are driven by large-scale pretrained models, which follow the standard “pre-training followed by fine-tuning” paradigm, where the full model is fine-tuned for downstream tasks. Although effective, this approach becomes computationally prohibitive as the model size increases. The Parameter-Efficient Fine-Tuning (PEFT) approach offers a promising alternative, but primarily focuses on the language components of Multimodal Large Language Models (MLLMs). Despite recent progress, PEFT remains underexplored in multimodal tasks and lacks sufficient understanding of visual information during fine-tuning the model. To bridge this gap, we propose Query-Adapter (Q-Adapter), a lightweight visual adapter module designed to enhance MLLMs by enabling efficient fine-tuning for the video captioning task. Q-Adapter introduces learnable query tokens and a gating layer into Vision Encoder, enabling effective extraction of sparse, caption-relevant features without relying on external textual supervision. We evaluate Q-Adapter on two well-known video captioning datasets, MSR-VTT and MSVD, where it achieves state-of-the-art performance among the methods that take the PEFT approach across BLEU@4, METEOR, ROUGE-L, and CIDEr metrics. Q-Adapter also achieves competitive performance compared to methods that take the full fine-tuning approach while requiring only 1.4% of the parameters. We further analyze the impact of key hyperparameters and design choices on fine-tuning effectiveness, providing insights into optimization strategies for adapter-based learning. These results highlight the strong potential of Q-Adapter in balancing caption quality and parameter efficiency, demonstrating its scalability for video–language modeling.

Our workshop aims to provide a platform for both academic and industrial professionals engaged in the analysis and retrieval of cross-data from diverse perspectives, with a particular emphasis on wearable and ambient sensors, lifelog cameras, social networks, and surrounding sensors.Despite numerous studies exploring individual viewpoints, there remains a significant gap in the analysis and retrieval of cross-data to maximize benefits for humanity.Additionally, challenges such as data security and distributed learning for cross-modal model training and inference arise when dealing with large and distributed datasets.We invite researchers to contribute to this initiative, with the overarching goal of fostering the development of a smart and sustainable society through the efficient utilization of intelligent cross-data analysis and retrieval techniques.

Multi-label multi-view action recognition aims to recognize multiple concurrent or sequential actions from untrimmed videos captured by multiple cameras. Existing work has focused on multi-view action recognition in a narrow area with strong labels available, where the onset and offset of each action are labeled at the frame-level. This study focuses on real-world scenarios where cameras are distributed to capture a wide-range area with only weak labels available at the video-level. We propose the method named Multi-view Action Selection Learning (MultiASL), which leverages action selection learning to enhance view fusion by selecting the most useful information from different viewpoints. The proposed method includes a Multi-view Spatial-Temporal Transformer video encoder to extract spatial and temporal features from multi-viewpoint videos. Action Selection Learning is employed at the frame-level, using pseudo ground-truth obtained from weak labels at the video-level, to identify the most relevant frames for action recognition. Experiments in a real-world office environment using the MM-Office dataset demonstrate the superior performance of the proposed method compared to existing methods. The source code is available at https://github.com/thanhhff/MultiASL/.

Recipe is a set of instructions that describes how to make food. It can help people from the preparation of ingredients, food cooking process, etc. to prepare the food, and increasingly in demand on the Web. To help users find the vast amount of recipes on the Web, we address the task of recipe recommendation. Due to multiple data types and relationships in a recipe, we can treat it as a heterogeneous network to describe its information more accurately. To effectively utilize the heterogeneous network, metapath was proposed to describe the higher-level semantic information between two entities by defining a compound path from peer entities. Therefore, we propose a metapath-enhanced recipe recommendation framework, RecipeMeta, that combines GNN (Graph Neural Network)-based representation learning and specific metapath-based information in a recipe to predict User-Recipe pairs for recommendation. Through extensive experiments, we demonstrate that the proposed model, RecipeMeta, outperforms state-of-the-art methods for recipe recommendation.

We propose a method that assists legislative drafters in locating inappropriate legal terms in Japanese statutory sentences and suggests corrections. We focus on sets of mistakable legal terms whose usages are defined in legislation drafting rules. Our method predicts suitable legal terms using a classifier based on a BERT (Bidirectional Encoder Representations from Transformers) model. We apply three techniques in training the BERT classifier, specifically, preliminary domain adaptation, repetitive soft undersampling, and classifier unification. These techniques cope with two levels of infrequency: legal term-level infrequency that causes class imbalance and legal term set-level infrequency that causes underfitting. Concretely, preliminary domain adaptation improves overall performance by providing prior knowledge of statutory sentences, repetitive soft undersampling improves performance on infrequent legal terms without sacrificing performance on frequent legal terms, and classifier unification improves performance on infrequent legal term sets by sharing common knowledge among legal term sets. Our experiments show that our classifier outperforms conventional classifiers using Random Forest or a language model, and that all three training techniques contribute to performance improvement.

Data integration is still laboursome task when integrating data are not consistently managed. Such inconsistency can happen easily in real-world situations, such as properties of objects are managed by a central organization and trajectories (or logs) of the objects are recorded by other peripheral organizations. This paper deals with a case of missing ordering information. Integrating property data and log data without ordering information causes duplicated results. In order to solve this problem, this paper proposes a join algorithm, called implicit order join, which discovers implicit ordering information from both property data and log data with help of partial true integrated results from human assistance. With the discovered ordering information, the implicit order join enables to integrate the property data and log data. In order to discover the implicit ordering information, ordering correlation between attribute sequences of property data and log data should be found from comprehensive examination of possible attribute sequence pairs. The potential number of sequence pairs is as high as factorial order of the number of attributes. Therefore, this paper develops a heuristic approach to prune unnecessary examinations based on ordering dependency between attribute sequences. Experimental evaluation in this paper indicates that implicit order join can reduce 77% labouring tasks for integration and the pruning method reduces the number of attribute sequences in orders of magnitude.

Analyzing and feeding back the results on real-world services are important missions in the Big Data era to realize smart city. However, analyzing real-world data is still challenging because of dirtiness of data and large variety of analytic requirements. To cope with the challenges, this paper proposes and develops an analytical toolbox for smart city applications. The analytical toolbox consists of three phases: preparation, analysis, and visualization. The preparation phase deals with the dirtiness of the data by including fundamental data cleansing techniques and data integration techniques. The analysis phase is responsible for ETL (extract, transform and load) process and analytical query processing from the next phase. The visualization phase deals with analytical requirements from users and visualization of analytical results. This paper showcases a real-world use case of the proposed analytical toolbox. The use case is now open in public with help of Fujisawa city, Japan, and this fact indicates that the proposed analytical toolbox is feasible for real-world data analysis and feeding back to citizens.

Searching over Linked Data requires large efforts to users, which include knowing locations of suitable SPARQL endpoints and writing appropriate SPARQL queries in terms of language standards as well as the underlying structure of Linked Data. This situation degrades usability of Linked Data, thus is highly problematic. To resolve this problem, this paper proposes CROISSANT which is a centralized view management system for SPARQL endpoints on the Web. CROISSANT stores pre-defined view definitions, and provides a searchable interface for the views to users. To realize CROISSANT, query processing performance is a big issue, because CROISSANT has to communicate with remote SPARQL endpoints and it takes time to receive results. To cope with this issue, this paper proposes four optimization techniques, namely, view materialization, selection push-down, projection push-down, and view query merge. Experimental evaluation demonstrates these optimizations improve query processing performance.

Analyzing vehicle recorder data of electric vehicles (EVs) reveals how the EVs are used. This paper proposes an OLAP framework to support analyzing trajectories in vehicle recorder data and applies the framework to vehicle recorder data of EVs. The framework consists of ETL (extract, transform, and load) process for trajectory data and visualization for analyzing the data. The ETL process includes hierarchy definitions for spatial and temporal dimensions, as well as aggregation functions for trajectory data. In the subsequent visualization phase, the framework displays results of OLAP operations on map interface. To ensure the applicability of the framework for real applications, we apply the framework to vehicle recorder data of micro-sized EVs (or μEVs), which are smaller EVs with one or two passengers including one driver and can drive at most 100km distance without charging on the way. The application realizes that the framework successfully enables analyses on the trajectory data for real analytic requirements.

Linked Data (or LD) has promoted publishing information, and links published information (e.g., vocabularies and facts) for utilization. There are increasing number of LD datasets containing numerical data such as statistics. Analyses using such data require dedicated programs to extract, transform, and load (or ETL) for preparation. Thus, a large effort of developers is required. Also, the LD datasets tend to be large and the dumps (or snapshots) for the datasets easily become not up-to-date due to update frequency of the datasets. Hence, downloading dumps of LD datasets to ETL for OLAP can miss latest records. This paper proposes a framework called SPOOL, which attempts to reduce the effort and to ETL latest numerical records data from LD datasets for OLAP through SPARQL endpoints without downloading whole datasets. SPOOL provides series of SPARQL queries extracting objects and attributes from LD datasets, and converts them into star/snowflake schemas, and materialize relevant triples as fact and dimension tables for OLAP. The applicability of SPOOL is evaluated using exiting LD datasets on the Web, and SPOOL successfully processes the LD datasets to ETL for OLAP.

Applying faceted search for XML data enables users to search XML data in an interactive manner. However, applying faceted search is challenging, because faceted search requires target subtrees (objects) and facets to be defined before-hand. To this problem, existing works assume that such objects and/or facets are defined manually, but it is infeasible to manually specify objects and facets in particular when the XML data are huge and/or its structure is quite complicated. To address this problem, this paper proposes an automatic extraction scheme of objects and facets from XML data. We propose two approaches, namely frequency-based approach and semantic-based approach, and also hybrid approach of them. The basic ideas of these approaches are that the frequently occurring XML elements seem to be objects and facets, and such XML elements may have semantically meaningful name. Although the proposed approaches are rather simple, the experiments using real world XML data show that the proposed approaches can automatically extract objects and facets from the XML data.

Faceted search for XML data is one of the promising exploration methods with high usability to find desired subtrees from a given XML data. This paper proposes improved approach of faceted search over XML data by utilizing facets containing unique and longer textual values, like titles of papers in bibliographic database. Our approach is to extract suitable terms which categorize the current results into several groups. Also we propose a task designing method for evaluating exploratory search by defining specificity of tasks called specification level, and we introduce how to generate tasks with given specification level as well. With this task design, we evaluate our proposed approach and the results show our proposed approach improves search performance comparing with the previous approaches, especially when tasks have low specification levels.

In this paper, we propose a framework of faceted navigation over XML data. General faceted navigation schemes are used to browse objects (or records) containing multiple properties. However, because XML is semi-structured in nature, it is not straightforward to apply faceted navigation to XML data. Specifically, we need to cope with three major technical issues: 1) objects in XML data are not predetermined, 2) objects may have flexible and/or recursive structure, and 3) properties of an object need to be automatically detected and extracted. To these problems, in this paper, we formulate faceted navigation over XML data by giving definitions of class, property, object, and facet in XML data. We then formulate typical user interactions in faceted navigation as operations over aforementioned concepts (class, object, and facet). We also propose a framework based on these definitions and operations, and construct a prototype system based on the framework. Finally, we show experimental evaluations using the prototype system to show the effectiveness of our proposed scheme.