Interval Type-2 Fuzzy Logic for Semisupervised Multimodal Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Retrieving nearest neighbors in large-scale multimodal datasets is computationally challenging.
• Existing multimodal hashing methods often require extensive labeled data (supervised) or lack performance (unsupervised).

Purpose of the Study:

• To develop a semisupervised multimodal hashing method that leverages partially available labels for efficient data retrieval.
• To address the limitations of unsupervised and fully supervised approaches in multimodal data analysis.

Main Methods:

• Proposed a semisupervised multimodal hashing approach utilizing interval type-2 fuzzy sets to estimate labels for unlabeled data.
• Employed hard partitioning for defuzzification and a supervised hashing method to generate binary codes.
• Evaluated the method on a multilabeled MIRFlickr dataset for hash lookup tasks.

Main Results:

• The semisupervised method achieved medium performance with 50% labels, comparable to supervised methods using 90% labels.
• Even with only 10% labels, the proposed method competed with the least effective supervised methods.
• Demonstrated the feasibility of the label estimation technique for multilabeled datasets in hash lookup.

Conclusions:

• Semisupervised multimodal hashing offers a viable solution for large-scale data retrieval when labeled data is scarce.
• The proposed interval type-2 fuzzy set-based label estimation enhances the efficiency and applicability of multimodal hashing.
• This method provides a practical approach for improving search performance in complex, multimodal datasets.