Our laboratory studies how complex real-world data can be made reliable, analyzed systematically, and transformed into interpretable knowledge.
Rather than focusing on a single type of data, we investigate fundamental problems in data engineering and apply the resulting methods to healthcare, cultural, multimedia, and other real-world datasets.
Our current research is organized around closely connected themes.
1. Data Quality & Imputation
Problem
Real-world datasets are rarely complete.
Missing observations, irregular measurement intervals, inconsistent records, and heterogeneous data collection processes can substantially affect subsequent statistical analysis and machine-learning results.
Simply filling missing values is therefore not sufficient. We also need to understand whether an estimated value is reasonable and how imputation influences downstream analysis.
Approach
We investigate statistical and machine-learning approaches to missing-data imputation, including methods that consider:
- temporal relationships;
- correlations among variables;
- characteristics of individual records;
- differences among data sources or institutions;
- uncertainty in estimated values.
A particular interest is explainable imputation: estimating missing values while also providing information about which observations and factors contributed to those estimates.
Representative Research
Our research on electronic medical record data has compared multiple strategies for missing-value estimation and evaluated their accuracy using quantitative measures such as RMSE and MAE.
We are extending this work toward interpretable and time-aware imputation methods for longitudinal real-world data.
Current / Open Topics
- Explainable data imputation
- Time-aware imputation for longitudinal data
- Missingness mechanisms
- Uncertainty-aware estimation
- Effects of imputation on downstream data mining
- Cross-institutional robustness of imputation methods
2. Pattern Discovery
Problem
Large-scale real-world data often contain repeated structures and behaviors, but these patterns are difficult to identify because individual records differ in length, timing, order, and context.
Frequent patterns alone may also hide important variations.
Approach
We develop methods based on:
- sequential pattern mining;
- interactive sequential pattern mining;
- similarity analysis;
- Dynamic Time Warping and related sequence comparison techniques;
- clustering;
- statistical analysis;
- visualization.
Our goal is to discover not only representative patterns, but also meaningful variations among individuals, institutions, and time periods.
Representative Research
Using electronic medical records collected from multiple healthcare institutions, we have studied medical-order sequences and treatment patterns.
Our work investigates how frequently occurring treatment processes can be extracted and how their similarities and differences can be quantified across institutions.
Current / Open Topics
- Sequential pattern discovery
- Multi-institutional pattern analysis
- Sequence similarity
- Clustering of process data
- Temporal changes in patterns
- Visualization of complex sequences
- Pattern discovery from multimodal data
3. Variation Explanation
Problem
Discovering that two groups follow different patterns is useful, but often the more important question is:
Why does the difference occur?
For example, a process may branch into different paths depending on individual characteristics, observed conditions, organizational differences, or contextual information.
Approach
We combine pattern discovery with statistical and machine-learning models to identify factors associated with different branches and variations.
We emphasize methods whose results can be interpreted and connected back to original observations.
Examples include:
- logistic regression;
- odds ratios;
- interpretable predictive models;
- feature attribution;
- statistical comparison across groups.
Representative Research
In healthcare data, we investigate how treatment sequences branch and whether these branches are associated with patient conditions, laboratory information, or characteristics of healthcare institutions.
This allows us to move from simply identifying different treatment patterns toward explaining factors potentially associated with those differences.
Current / Open Topics
- Explanation of branching processes
- Patient-level and institution-level factors
- Statistical modeling of variations
- Explainable AI
- Relationship between pattern discovery and causal analysis
- Explanation of multimodal behavior patterns
4. Trustworthy Data Management
Problem
Data are increasingly stored and exchanged across multiple systems and organizations.
In such environments, it is important to ensure that data have not been improperly modified and that their history can be verified.
Traditional centralized mechanisms may also be insufficient when data are shared among mutually independent parties.
Approach
We investigate data-management mechanisms for maintaining:
- data integrity;
- provenance;
- traceability;
- tamper detection;
- verification in distributed environments.
Our work includes blockchain-based approaches for detecting data modification and managing trustworthy shared data.
Representative Research
We have studied blockchain-based mechanisms for maintaining data integrity in distributed applications, including approaches for efficient tamper detection and recovery of modified data.
We also investigate data structures and verification mechanisms that reduce the cost of detecting inconsistencies in distributed data.
Current / Open Topics
- Data integrity verification
- Blockchain-based data management
- Tamper detection and recovery
- Efficient integrity structures
- Data provenance
- Distributed data management
- Traceable data processing
Emerging Research Direction
Evidence-Aware AI
We are exploring how data engineering can support AI systems that not only generate answers, but also identify, verify, and trace the evidence behind them.
Our current interests include reasoning-oriented retrieval, integration of structured data and external knowledge, evidence verification, and data provenance.
Topics: Reasoning-oriented RAG · Evidence retrieval · Evidence verification · Provenance · Traceability · Trustworthy AI
Application Areas
The same data-engineering problems appear across many different domains. We therefore separate our methodological research themes from the types of data to which they are applied.
Healthcare Data
Healthcare provides an important environment for studying complex, heterogeneous, and sequential real-world data.
Examples of research questions include:
- What treatment patterns commonly appear across healthcare institutions?
- How much do treatment processes differ among institutions?
- What factors are associated with different treatment branches?
- How should missing clinical measurements be estimated?
- Can discovered patterns be linked to supporting evidence?
Healthcare data are an important application area of the laboratory, but our research questions are not limited to it.
Cultural & Multimodal Data
We are also developing research topics involving cultural and multimodal data, particularly for students interested in Humanities Data Engineering.
Possible data include:
Text · Images · Music · Audio · Film · Video · Dance · Performing Arts · Human Activities
Examples of research questions include:
- How can similarities among music, images, or videos be represented and searched?
- Can repeated structures or characteristic patterns be discovered from multimedia data?
- How do different performances of the same work vary?
- How can text, image, audio, and video information be analyzed together?
- How should large collections of cultural content be organized and searched?
- Can temporal patterns in dance or performing arts be discovered automatically?
- How can generative AI support the exploration of multimedia collections while preserving links to original data?
The objective is not simply to apply an existing AI model to cultural content. We are interested in identifying data-engineering problems arising from these datasets and developing methods to address them.