Manifold Learning for Multivariate Variable-Length Sequences With an Application to Similarity Search.

Area of Science:

• Data Science
• Machine Learning
• Pattern Recognition

Background:

• Multivariate variable-length sequences from mobile devices and sensor networks are increasingly common.
• Analyzing this data is challenging due to nonmetric similarity measures.
• Existing methods struggle with the complexity and variability of sequence data.

Purpose of the Study:

• To propose a general manifold learning framework for arbitrary-length multivariate data sequences.
• To develop a method for learning similarity and distance measures in both original and manifold spaces.
• To enable robust similarity search for complex sequence data.

Main Methods:

• A semisupervised manifold learning framework is proposed.
• Instance-level constraints (similarity/dissimilarity information) are utilized.
• Data sequences are transformed into feature vectors in a learned manifold.
• A consensus voting scheme is used for similarity search.

Main Results:

• The framework successfully transforms nonmetric sequence data into a structured manifold.
• Similar sequences are mapped to nearby points, dissimilar ones to distant points.
• Experimental results validate the framework's feasibility on synthetic and real-world data (tropical cyclones).

Conclusions:

• The proposed manifold learning framework effectively handles variable-length multivariate sequences.
• It enables robust similarity search by preserving data structure in the learned manifold.
• This approach offers a powerful tool for analyzing complex sequential data in various applications.