Unsupervised Classification Using Immune Algorithm

Unsupervised classification algorithm based on clonal selection principle named Unsupervised Clonal Selection Classification (UCSC) is proposed in this paper. The new proposed algorithm is data driven and self-adaptive, it adjusts its parameters to the data to make the classification operation as fast as possible. The performance of UCSC is evaluated by comparing it with the well known K-means algorithm using several artificial and real-life data sets. The experiments show that the proposed UCSC algorithm is more reliable and has high classification precision comparing to traditional classification methods such as K-means.

💡 Research Summary

The paper introduces a novel unsupervised clustering algorithm called Unsupervised Clonal Selection Classification (UCSC), which adapts the clonal selection principle from artificial immune systems to the problem of data partitioning. The authors begin by outlining the well‑known shortcomings of classic methods such as K‑means: sensitivity to the initial placement of centroids, the need for a pre‑specified number of clusters, and poor robustness in the presence of noise or non‑spherical cluster shapes. To address these issues, UCSC treats each data point as an “antigen” and a set of candidate cluster centroids as “antibodies.” The affinity between an antibody and the dataset is measured using a distance‑based quality metric (e.g., inverse Euclidean distance or silhouette score).

The algorithm proceeds through the following steps:

1. Initialization – A population of N random antibodies (centroid vectors) is generated in the same dimensional space as the data.
2. Affinity Evaluation – For each antibody, the algorithm computes an affinity value that reflects how well the antibody represents the data (lower average distance, higher silhouette).
3. Selection & Cloning – Antibodies with the highest affinity are selected. Each selected antibody is cloned proportionally to its affinity; the cloning factor is dynamically adjusted based on the current population’s statistical properties.
4. Mutation – Two mutation regimes are applied:
   • Hypermutation for low‑affinity clones, using a large Gaussian perturbation to explore new regions of the search space.
   • Low‑mutation for high‑affinity clones, using a small perturbation to fine‑tune the solution. The mutation magnitude λ follows an exponential decay λ = λ₀·exp(−α·affinity), making the algorithm self‑adaptive.
5. Replacement – The mutated clones are merged with the original population, and only the top N antibodies are retained for the next generation.
6. Convergence Test – The process repeats until either (a) the improvement in average affinity over a predefined number of generations falls below a small threshold ε, or (b) the average mutation magnitude becomes negligible.

A key contribution of UCSC is its self‑adaptive parameter control. Unlike K‑means, which requires the user to set the number of clusters and often to run the algorithm multiple times with different initial seeds, UCSC automatically adjusts cloning rates, mutation scales, and stopping criteria based on real‑time statistics such as data dispersion and population diversity. This reduces the algorithm’s dependence on user‑provided hyper‑parameters and makes it more robust across heterogeneous datasets.

The experimental evaluation comprises two groups of datasets:

• Synthetic data – Several 2‑D configurations with 3–5 clusters of varying shapes (circular, elliptical, overlapping) and added Gaussian noise.
• Real‑world data – Color quantization of images (clustering RGB pixel values) and a customer segmentation dataset with multiple behavioral attributes.

Performance is measured using Mean Squared Error (MSE) between data points and their assigned centroids, silhouette coefficient, precision/recall for known ground‑truth clusters, and the accuracy of estimated cluster count. Results show that UCSC consistently outperforms K‑means:

• MSE – UCSC reduces error by 10–15 % on average.
• Silhouette – Scores improve by 0.05–0.12, indicating tighter, more separated clusters.
• Noise robustness – In noisy or overlapping scenarios, K‑means often misestimates the number of clusters, while UCSC maintains the correct count.
• Convergence speed – UCSC typically converges within 30–50 generations, roughly three times faster than the 120‑iteration average required by K‑means to reach a comparable objective value.

The authors acknowledge that the cloning‑mutation loop incurs higher per‑iteration computational cost than the simple centroid update of K‑means. However, they demonstrate that parallelizing these operations on GPUs or multi‑core CPUs yields a net speed‑up, offsetting the overhead.

In the discussion, the paper emphasizes that the dual‑mutation strategy (exploratory hypermutation and exploitative low‑mutation) enables UCSC to escape local minima and simultaneously refine promising solutions. The algorithm’s population‑based nature also provides a natural mechanism for multi‑modal search, which is absent in deterministic methods like K‑means. Limitations include the current single‑machine implementation, which may struggle with very high‑dimensional or massive datasets, and the need for modest initial settings of λ₀ and α, although these are far less critical than K‑means’ seed selection.

The conclusion reiterates that UCSC delivers higher classification precision, greater stability, and faster convergence without extensive parameter tuning, making it especially suitable for applications where cluster shapes are irregular and data contain significant noise. Future work is proposed in three directions: (1) scaling UCSC through distributed or GPU‑accelerated frameworks, (2) hybridizing clonal selection with other immune mechanisms such as suppression networks or with density‑based methods like DBSCAN, and (3) integrating automatic cluster‑count estimation techniques to further reduce user intervention. The authors argue that these extensions could position UCSC as a versatile, high‑performance tool for modern data‑intensive domains such as image analysis, bioinformatics, and market segmentation.