An improved LDA dimension reduction algorithm for multivariate time series classification.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multivariate time series (MTS) classification is a growing research area.
• High dimensionality of MTS data often necessitates dimension reduction for effective classification.
• Existing dimension reduction techniques struggle with unequal-length MTS datasets, causing information loss or redundancy.

Purpose of the Study:

• To develop a novel method for transforming unequal-length MTS datasets into equal-length datasets, minimizing information loss.
• To propose a supervised dimension reduction technique that addresses feature point variations across time moments in MTS data.
• To enhance the effectiveness of dimension reduction for MTS classification.

Main Methods:

• A novel extraction method to convert unequal-length MTS datasets into equal-length datasets.
• A supervised dimension reduction method based on Linear Discriminant Analysis (LDA).
• The LDA method optimizes projection planes at each time point to minimize within-class scatter and maximize between-class scatter.

Main Results:

• The proposed method successfully transforms unequal-length MTS data into an equal-length format.
• Experiments on 16 public datasets demonstrate improved classification performance after dimension reduction.
• The supervised LDA-based dimension reduction effectively captures relevant features for classification.

Conclusions:

• The novel extraction and supervised dimension reduction methods significantly enhance MTS classification performance.
• The approach effectively handles unequal-length MTS data, overcoming limitations of traditional methods.
• This research contributes a more effective dimension reduction strategy for high-dimensional MTS classification problems.