Abstract: Inspired by the information processing mechanism of the human brain, the artificial neural network (ANN) is a classic data mining method and a powerful soft computing technique. The ANN provides a valuable tool for information processing and pattern recognition, thanks to its advantages in distributed storage, parallel processing, fast problem-solving and adaptive learning. The constructive neural network (CNN) is a popular emerging neural network model suitable for processing largescale data. In this paper, a novel neural network classification model was established based on the covering algorithm (CA) and the immune clustering algorithm (ICA). The CA is highly comprehensible, and enjoys fast computing speed, and high recognition rate. However, the learning effect of this algorithm is rather poor, because the training set is randomly selected from the original data, and the number of nodes (covering number) and area being covered are greatly affected by the learning sequence. To solve the problem, the ICA was introduced to preprocess the data samples, and calculate the cluster centers based on the antibody-antigen affinity. The CA and the ICA work together to determine the covering center and radius automatically, and convert them into the weights and thresholds of the hidden layer of neural network. The number of hidden layer neurons equals the number of covering. In addition, the McCulloch-Pitts (M-P) neurons were adopted for the output layer. Based on the input feature of the hidden layer, the output feature completes the mapping from input to output, creating the final classes of the original data. The introduction of the ICA fully solves the defect of the CA. Finally, our neural network classification model was verified through experiments on real-world datasets.
Introduction
Traditional Artificial Neural Networks (ANNs) often rely on backpropagation, which requires extensive weight tuning and a pre-defined architecture. Constructive Neural Networks (CNNs) offer an alternative by "building" the network based on the data's geometry—specifically using a Covering Algorithm (CA) to partition the data space into spherical "covers."
However, the standard CA has a major flaw: it is highly sensitive to the order in which data is presented. Because it selects initial covering centers randomly, the resulting network can be inconsistent and inefficient. This paper aims to solve this "learning sequence" problem by using a more intelligent, data-driven way to find the starting points for these geometric covers.
Methodology
The proposed model integrates two distinct algorithms to create a robust three-layer architecture:
- Immune Clustering Algorithm (ICA): Inspired by biological immune systems, this algorithm treats data points as "antigens" and cluster centers as "antibodies." It calculates "affinity" (Euclidean distance) to find natural clusters in the data without needing to know the number of clusters in advance.
- Covering Algorithm (CA): Once the ICA identifies the cluster centers, the CA determines the optimal radius for each spherical field. This ensures that the entire sample space is covered by these high-dimensional spheres.
The Architecture
- Input Layer: Receives the raw feature vectors.
- Classification Layer (Hidden): Each neuron represents one "cover" (a sphere in $n$-dimensional space). The number of neurons is automatically set to the number of covers found by the ICA.
- Output Layer: Uses McCulloch-Pitts (M-P) neurons, which act as simple logic gates to map the activated hidden neurons to the final class labels.
Unlike standard deep learning, this model does not require iterative weight adjustments via gradient descent. Instead, the "weights" are derived directly from the geometric centers and radii of the covers.
Key Insights: Geometry over Gradients
This research shifts the focus from "learning weights" to "covering domains." By viewing classification as a problem of high-dimensional geometry, the model becomes more interpretable—each hidden neuron has a clear physical meaning as a specific region in the feature space.
By using ICA as a front-end processor, the authors removed the element of luck from the Covering Algorithm. The ICA finds the "skeleton" of the data, ensuring the neural network starts with the most representative points. Furthermore, the data's inherent clustering structure dictates the network's size, automating the topology design.
Experimental Results
The proposed model was evaluated using three classic benchmarks from the UCI Machine Learning Repository: Iris, Wine, and Zoo. The results (averaged over 50 tests) showed a clear advantage over standard k-means clustering and the traditional Covering Algorithm.
| Dataset | Our Method (Accuracy %) | k-means (Accuracy %) | Traditional CA (Accuracy %) |
|---|---|---|---|
| Iris | 95.80 | 88.67 | 5.58 |
| Wine | 92.26 | 86.52 | 91.49 |
| Zoo | 88.14 | 84.31 | 89.87 |
The most striking result is on the Iris dataset, where the traditional CA performed poorly (5.58%), likely due to overlapping classes confusing the random initialization. The ICA-based approach remained robust, maintaining high accuracy across all tests.
Conclusion
While the model shows strong performance, future work will focus on scalability for "big data" contexts and deeper integration of the CA with specialized neural networks to enhance recognition rates in complex environments. Although it solves the "sequence" problem, the ICA's suppression thresholds may still require tuning for specific data types.
Summary updated