580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can rank aggregation and component-wise dissimilarities improve the efficiency and robustness of similarity search in complex, high-dimensional, or heterogeneous data spaces?
This theme explores methodologies that generalize the classical nearest neighbor search to more computationally efficient or conceptually robust similarity search strategies, particularly in complex data spaces characterized by multiple heterogeneous features. It emphasizes how rank aggregation techniques and component-wise similarity measures can resolve challenges related to the computational complexity of finding nearest neighbors and managing diverse data structures, thus enhancing pattern recognition and classification applications.
Key finding: This paper introduces a novel similarity search method grounded on rank aggregation where each coordinate dimension is treated as a 'voter' ranking candidates by their proximity to a query point. The method generalizes... Read more
Key finding: This study analyzes component-wise dissimilarity measures as weighted convex combinations applied to heterogeneous data features within pattern recognition. It demonstrates that while such measures are often non-Euclidean and... Read more
Key finding: This paper applies similarity relations between objects within an information system to construct pretopological spaces, facilitating attribute reduction in high-dimensional data. It defines a similarity matrix derived from... Read more
2. What formal properties should similarity functions satisfy to accurately represent psychological, linguistic, and semantic similarities, and how can parameterized models optimize performance across diverse data types?
This research theme addresses the formalization and empirical evaluation of similarity functions that capture semantic, psychological, and linguistic similarity judgments in a principled manner. It examines parameterized models developed to enforce desirable theoretical properties (e.g., symmetry, self-similarity, monotonicity), and empirical methods to estimate optimal parameters. The aim is to reconcile representational adequacy with computational tractability, improving similarity assessments in natural language processing and cognitive modeling contexts.
Key finding: This paper contrasts two structural psychological models of similarity—structural alignment (SA) and representational distortion (RD)—and shows via experiments that RD, operationalized through complexity of transformations,... Read more
Key finding: This comprehensive review links recent trends in deep learning using similarity functions to estimate mutual information (e.g., MINE, InfoNCE) with the author’s semantic information G theory developed decades ago. It... Read more
3. How can structural similarity measures be adapted and improved for advanced application domains such as medical image analysis and biomedical knowledge graph embeddings?
This theme investigates the adaptation of similarity measures, particularly structural similarity, to specialized domains including medical image quality assessment and biomedical knowledge discovery. It focuses on developing novel fuzzy set-based measures, embedding methods informed by semantic similarity, and evaluating the sensitivity and robustness of these measures against distortions or noise. These advancements empower precise quantitative evaluation and predictive modeling in fields where domain-specific data characteristics demand tailored similarity functions.
Key finding: This work proposes SimTransE, a translation-embedding-based machine learning method that integrates semantic similarity between biomedical entities (drugs and targets) into knowledge graph embeddings for link prediction. By... Read more
Key finding: The paper presents a new fuzzy set-based image similarity index founded on the greatest and smallest fuzzy set solutions of fuzzy relation equations extracted from image blocks. Experimental evaluations with Gaussian noise... Read more