Short Sequence Classification Through Discriminable Linear Dynamical System.

Area of Science:

• Machine Learning
• Data Science
• Systems Biology

Background:

• Linear dynamical systems (LDS) are valuable for uncovering hidden data structures and aiding in data representation and analysis.
• Traditional LDS training algorithms often neglect class labels, leading to models that are indistinguishable across classes and perform poorly in classification tasks.

Purpose of the Study:

• To address the limitation of indistinguishable LDS across classes in traditional training methods.
• To propose a novel approach that enhances the diversity of LDS models for different classes, thereby improving classification accuracy.

Main Methods:

• Revisiting the limitations of standard LDS training algorithms.
• Introducing a diversity measure based on determinantal point processes (DPP) applied to LDS.
• Utilizing DPP to mitigate the greedy behavior of the electromagnetic algorithm during training.
• Developing a training objective that balances model fit with inter-class distinguishability.

Main Results:

• Experiments on synthetic data demonstrated the effectiveness of the proposed method in generating discriminative LDS under supervisory information.
• Classification tasks on short time-span datasets showed that the developed LDS models generalize well to unseen data.
• The approach successfully balances goodness of fit with class distinguishability.

Conclusions:

• The proposed method effectively enhances the discriminative power of linear dynamical systems for classification tasks.
• By incorporating class-aware diversity, the trained LDS models exhibit improved generalization to new data.
• This approach offers a promising direction for improving model-based classification using sequential data.