580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can component-wise and higher-order dissimilarity measures enhance similarity assessments in heterogeneous and complex data spaces?
This research area focuses on devising and analyzing dissimilarity/similarity measures tailored for complex real-world objects represented by heterogeneous, multi-component data. Traditional metric or Euclidean assumptions often fail in such unconventional spaces where data comprise mixed types (numerical, categorical, time series, graphs). Component-wise dissimilarities allow each heterogeneous component to be compared using domain-appropriate submeasures, combined often through weighted convex combinations. Theoretical and experimental studies explore how these weighted measures affect metric properties and Euclidean embeddability. Further, the concept of meta-distances introduces higher-order similarities that consider the relative similarities of objects with respect to the entire dataset, thereby capturing richer relational patterns beyond pairwise comparisons. These measures prove essential for improving pattern recognition and local classification performance in complex domains.
Key finding: The paper formalizes component-wise dissimilarity measures to accommodate heterogeneous real-world data described by mixed features and demonstrates that such dissimilarities often produce non-Euclidean matrices, limiting... Read more
Key finding: Introducing meta-distances constructed from primary classical distances by incorporating an adjunct dissimilarity factor encoding higher-order similarity relationships among all objects in a dataset, this study demonstrates... Read more
Key finding: The 'brsim' R package operationalizes the Brainerd-Robinson similarity coefficient designed for compositional data, facilitating significance testing through permutation methods and hierarchical clustering analyses. Its... Read more
2. How can semantic and fuzzy similarity measures be parametrically adapted and combined to improve conceptual reasoning and decision-making?
This theme covers theoretical and applied advancements in parametrically flexible similarity measures designed for semantic resources and fuzzy set representations. Semantic similarity methods leverage information content and ontology-based taxonomies with weights informed by either resource frequency or ontology structure, allowing improved assessment of concept relatedness capturing both statistical and domain-specific knowledge. Similarly, in fuzzy logic, combining distance and similarity measures into unified parametric forms addresses challenges in fuzzy set comparison, avoiding ambiguous interpretations when sets are disjoint or partially overlapping. Parametric adjustments and combinations enable tailoring similarity measures to better reflect nuanced semantic or fuzzy relationships, thereby enhancing applications such as semantic retrieval, multi-attribute decision making, and reasoning under uncertainty.
Key finding: The paper presents SemSim p, a parametric semantic similarity method that improves upon its predecessor by adjusting ontology concept weights and normalization factors. Experiments using the ACM Computing Classification... Read more
Key finding: This work formulates a novel combined measure of similarity and distance between fuzzy sets using an ordered weighted averaging (OWA) operator. It overcomes limitations when similarity or distance measures are used... Read more
Key finding: Addressing limitations in existing picture fuzzy set similarity measures, the paper introduces a parametric similarity measure with three adjustable parameters (m1, m2, m3) enabling flexible decision-making styles. Analytical... Read more
3. What novel similarity measures improve performance and interpretability in collaborative filtering and image similarity tasks?
This research cluster investigates new or hybrid similarity metrics tailored to enhance the effectiveness of collaborative filtering (CF) recommender systems and image similarity assessment. For CF, combining classical numerical similarity measures (e.g., cosine, Pearson correlation) with Jaccard similarity—which emphasizes presence/absence of ratings rather than rating magnitude—has been shown to produce superior neighbor identification and recommendation accuracy. In image similarity, beyond traditional pixel-wise metrics (PSNR, SSIM), novel approaches leverage fuzzy set solutions derived via max–min and min–max compositions or convolutional neural networks (CNNs) to capture nuanced perceptual similarities and increase robustness to noise. These advances address key challenges including sparsity, noise sensitivity, and semantic expressiveness, advancing both theory and practical applications in recommendation systems and image quality assessment.
Key finding: Experiments on MovieLens and FilmTrust datasets demonstrate that hybrid similarity measures combining Jaccard similarity—which captures rating presence—and numerical measures like cosine and Pearson outperform any single... Read more
Key finding: Extending prior findings, this paper empirically validates that fusing Jaccard similarity with classical numerical similarity metrics yields significant improvements in collaborative filtering performance. Rigorous testing on... Read more
Key finding: The study proposes a novel fuzzy-based image similarity measure that computes similarity using the greatest and smallest fuzzy sets derived as symmetrical solutions of fuzzy relation equations for image blocks. Evaluation on... Read more
Key finding: Introducing a deep learning-based approach, this paper develops a CNN model to assess similarity between UML class diagram images, thereby enabling automatic, objective evaluation of student diagrams in education. The model... Read more