Temporal Data Clustering Techniques

Introduction

• Conventional action recognition approaches are designed to recognize videos that contain a single action.
• A preprocessing segmentation process is necessary for such approaches.
• Temporal data clustering is complex due to data dimensionality and temporal relationships.

Temporal Data Clustering

• Three categories of temporal clustering methods exist: model-based, subspace clustering, and kernel methods.
• Subspace clustering methods, such as Sparse Subspace Clustering (SSC), Least-Square Regression (LSR), and Low-Rank Representation (LRR) aim to learn a distinctive and low-dimensional representation of data.
• Semi-Markov K-means clustering is designed to find repeated patterns in temporal data.
• Hierarchical aligned cluster analysis utilizes a dynamic time alignment kernel to cluster time series data.
• Maximum-margin clustering method simultaneously recognizes the length and position of corresponding segments.
• Temporal Subspace Clustering (TSC) jointly learns a dictionary and representations with a regulation to decode temporal information.

Transfer Learning in Temporal Data Clustering

• Transfer learning techniques are used to transfer knowledge from one task to another, even if the tasks are different but related.
• The proposed approach is a transductive transfer learning scenario, where the source and target tasks are the same, but the domains are different.
• The goal of the proposed approach is to explore the use of source knowledge (related data) to improve the segmentation performance in the target domain.
• The approach learns a domain-shared subspace using a reconstruction-based strategy to guide knowledge transfer.

Proposed Approach

• The proposed approach utilizes a reconstruction-based strategy to guide knowledge transfer.
• A domain-invariant projection is learned to mitigate data distribution differences between source and target domains.
• A graph regularizer is built to capture the temporal information of source and target for better clustering.
• The approach only constrains the representation samples belonging to the same group with temporal constraint in the source part.
• No requirement for the class of source data to be overlapped with target data.
• The approach uses a conventional clustering algorithm, Normalized Cuts (Shi and Malik 2000), after generating an undirected graph G.

Results

• The approach was evaluated using three datasets, where each target dataset was segmented based on the other two as source data.
• The performance of the proposed approach was compared with SSC, LSR, LRR, OSC, and TSC.
• The proposed approach outperformed other methods in terms of clustering accuracy, particularly for datasets with similar visual appearance, such as actions performed under consistent illumination and subject appearance.

Limitations

• The improvement of the proposed approach is not significant compared to other methods when the data is distinctive due to the lack of temporal information preserved in the model.

Conclusion

• The proposed approach is a novel transfer learning based subspace clustering method for temporal data.
• The approach utilizes information from related data to improve clustering performance.
• The approach addresses the challenge of segmenting temporal data into meaningful groups when labeled data is scarce or unavailable.