Abstract
Most semi-supervised learning (SSL) models entail complex structures and iterative training processes as well as face difficulties in interpreting their predictions to users. To address these issues, this paper proposes a new interpretable SSL model using the supervised and unsupervised Adaptive Resonance Theory (ART) family of networks, which is denoted as SSL-ART. Firstly, SSL-ART adopts an unsupervised fuzzy ART network to create a number of prototype nodes using unlabeled samples. Then, it leverages a supervised fuzzy ARTMAP structure to map the established prototype nodes to the target classes using labeled samples. Specifically, a one-to-many (OtM) mapping scheme is devised to associate a prototype node with more than one class label. The main advantages of SSL-ART include the capability of: (i) performing online learning, (ii) reducing the number of redundant prototype nodes through the OtM mapping scheme and minimizing the effects of noisy samples, and (iii) providing an explanation facility for users to interpret the predicted outcomes. In addition, a weighted voting strategy is introduced to form an ensemble SSL-ART model, which is denoted as WESSL-ART. Every ensemble member, i.e., SSL-ART, assigns a different weight to each class based on its performance pertaining to the corresponding class. The aim is to mitigate the effects of training data sequences on all SSL-ART members and improve the overall performance of WESSL-ART. The experimental results on eighteen benchmark data sets, three artificially generated data sets, and a real-world case study indicate the benefits of the proposed SSL-ART and WESSL-ART models for tackling pattern classification problems.
Introduction
Semi-supervised learning (SSL) is a critical paradigm in machine learning, particularly when labeled data is expensive or difficult to obtain while unlabeled data is abundant. Applications range from medical diagnosis to industrial fault detection. However, existing SSL models often suffer from high computational complexity due to iterative training, a lack of transparency (the "black box" problem), and susceptibility to catastrophic forgetting during incremental learning.
This paper introduces SSL-ART and its ensemble variant WESSL-ART, designed to overcome these hurdles. By leveraging the Adaptive Resonance Theory (ART) framework, the proposed models offer efficient one-pass learning, stability, and human-interpretable explanations.
Methodology
SSL-ART Architecture
The core of the SSL-ART model is a two-stage process that combines unsupervised and supervised learning within an ARTMAP-like structure.
- Unsupervised Clustering (Stage 1): A Fuzzy ART network ($ART_a$) processes unlabeled samples to establish "prototype nodes." These nodes represent clusters in the input space.
- Supervised Mapping (Stage 2): A Fuzzy ARTMAP structure links these established prototypes to target class labels using a smaller set of labeled data ($ART_b$).
One-to-Many (OtM) Mapping
A key innovation in SSL-ART is the One-to-Many (OtM) mapping scheme. In standard ARTMAP, a prototype node is strictly mapped to a single class. This often leads to "category proliferation," where many redundant nodes are created to handle noisy or overlapping data. The OtM scheme allows a prototype to initially associate with multiple class labels. The final class assignment is determined by the most frequent label encountered during training. This approach significantly reduces model complexity and enhances robustness to noise.
WESSL-ART Ensemble
ART models are inherently sensitive to the order of training data. To mitigate this, the authors propose WESSL-ART, an ensemble of SSL-ART models. Each member is trained on a different data sequence. A weighted voting strategy then combines their outputs, where weights are assigned based on each member's class-specific performance on a validation set.
Rule Extraction and Explanation
Unlike many neural networks, SSL-ART provides an explanation facility. The internal prototype nodes can be translated into fuzzy "If-Then" linguistic rules. For example, a rule might state: "If Feature A is High and Feature B is Low, then the class is X." This transparency is vital for high-stakes applications like healthcare.
Experimental Evaluation
The models were rigorously tested across diverse scenarios:
- 18 UCI Benchmarks: Including Iris, Wine, MNIST, and Pima Indians Diabetes.
- Anomaly Detection: Evaluated on the Tennessee Eastman process (TEP), Water Distribution (WADI), and UNSW-NB15 datasets.
- Real-World Case Study: Human motion detection using smartphone accelerometer data.
Key Results
- Performance: WESSL-ART consistently outperformed established SSL methods like SemiBoost and RegBoost. In anomaly detection tasks (TEP and WADI), it achieved F1-scores of 0.9378 and 0.9572, respectively, surpassing state-of-the-art models like GANF.
- Robustness: The OtM mapping scheme proved more resilient to label noise than standard one-to-one mapping in the human motion study.
- Efficiency: The model's one-pass learning capability makes it suitable for real-time streaming data.
- Interpretability: Rule extraction produced a compact set of rules—often three times fewer than the number of unlabeled samples—facilitating easy auditing of the model's logic.
Conclusion
SSL-ART and WESSL-ART represent a significant step toward efficient and transparent semi-supervised learning. By combining the stability of ART with a novel OtM mapping scheme and ensemble voting, the framework achieves high accuracy while remaining interpretable. Future work will focus on refining the OtM scheme to better handle complex decision boundaries and exploring integration with reinforcement learning for uncertain environments.
Summary updated