Model Input-Output Configuration Search with Embedded Feature Selection for Sensor Time-series and Image Classification

TL;DR

The paper tackles learning when input-output mappings and model architecture are unknown by integrating input-output configuration search with embedded feature selection. It introduces MICS-EFS, a two-level framework built around a modified encoder-decoder IO autoencoder and a CNN classifier, using Sequential Forward Search to explore IO configurations and embedded FS to prune features. On diverse 1D sensor and 2D image datasets, plus an industrial machining case, MICS-EFS achieves average accuracy gains around 1.5% and reduces the feature set to 2-5% of original, improving both accuracy and computational efficiency. The work demonstrates practical viability for real-world tasks and contributes a publicly available implementation.

Abstract

Machine learning is a powerful tool for extracting valuable information and making various predictions from diverse datasets. Traditional machine learning algorithms rely on well-defined input and output variables; however, there are scenarios where the separation between the input and output variables and the underlying, associated input and output layers of the model are unknown. Feature Selection (FS) and Neural Architecture Search (NAS) have emerged as promising solutions in such scenarios. This paper proposes MICS-EFS, a Model Input-Output Configuration Search with Embedded Feature Selection. The methodology explores internal dependencies in the complete input parameter space for classification tasks involving both 1D sensor time-series and 2D image data. MICS-EFS employs a modified encoder-decoder model and the Sequential Forward Search (SFS) algorithm, combining input-output configuration search with embedded feature selection. Experimental results demonstrate the superior performance of MICS-EFS compared to other FS algorithms. Across all tested datasets, MICS-EFS delivered an average accuracy improvement of 1.5% over baseline models, with the accuracy gains ranging from 0.5% to 5.9%. Moreover, the algorithm reduced feature dimensionality to just 2-5% of the original data, significantly enhancing computational efficiency. These results highlight the potential of MICS-EFS to improve model accuracy and efficiency in various machine learning tasks. Furthermore, the proposed method has been validated in a real-world industrial application focused on machining processes, underscoring its effectiveness and practicality in addressing complex input-output challenges.

Figures (20)

• Figure 1: Given parameters: $x$ and $x^2$. a) The relationship $x \xrightarrow{} x^2$ (with x as input and $x^2$ as output) is attempted to be learned. This configuration is well-chosen, allowing the function to be effectively learned. b) The relationship $x^2 \xrightarrow{} x$, (with $x^2$ as input and $x$ as output), is attempted to be learned. This configuration is poorly chosen, resulting in inaccurate learning of the function.
• Figure 2: The concept of the Input-Output Configuration Search. The input-output properties of elements within the parameter space are initially unknown. Through this search process, the parameters are separated into two distinct sets, identifying which parameters behave more as inputs or outputs.
• Figure 3: Modified adaptation of the classic autoencoder model with input $x_{in}$ and output $x_{out}$. Blue color indicates available and utilized, white color indicates missing (ignored) variables.
• Figure 4: Modified adaptations of the classic convolutional autoencoder model with input $x_{in}$ and output $x_{out}$ for image (upper) and sensor time-series data (lower).
• Figure 5: The process of the Sequential Forward Search algorithm. SFS iteratively builds an input feature subset by adding one feature at each iteration, selecting the one that has optimal input-output dependency, until a stopping criterion is met.