Clustering: A neural network approach

In this paper, a comprehensive overview of competitive learning based clustering methods is given. Importance is attached to a number of competitive learning based clustering neural networks such as the self-organizing map (SOM), the learning vector quantization (LVQ), the neural gas, and the ART model.

About this project

In this project, two of the most famous NN-based clustering methods, SOM and LVQ, will be implemented. Their performance and efficacy, will then be investigated through analysing different criteria such as Calinski-Harabasz index, Davies-Bouldin index, and silhouette values as well as silhouette plot. The programming language for this project is chosen to be MATLAB and the dataset used is artificially created by sampling Gaussian distribution. Moreover, a linear dimensionality reduction based on PCA (Principal Component Analysis) is performed as a pre-processing step.

NOTE: Self-organizing maps learn to cluster data based on similarity, topology, with a preference (but no guarantee) of assigning the same number of instances to each class; competitive layers learn to classify input vectors into a given number of classes, according to similarity between vectors, (AGAIN!) with a preference for equal numbers of vectors per class. (excerpted from MATLAB help)
Hence, if you are sure that your dataset is imbalanced and your expected clusters are not supposed to be of similar population, then I recommend that you try to change the loss function of the neural network (which I feel would be arduous) or implement other clustering methods such as hierarchical clustering and k-means clustering.

Files' description:

• sofm.m >> in this MATLAB script, a SOFM neural network is trained to cluster the data into 3 classes.

• lvq.m >> in this MATLAB script, a LVQ-based neural network is trained to cluster the data into 3 classes.

• Performance_Evaluation.m >> in this MATLAB script, two clustering algorithms, SOFM and LVQ, are compared together through well-known quantitative criteria such as Calinski-Harabasz index and Davies-Bouldin index, as well as silhouette values and silhouette plot.