A weight-incorporated similarity-based clustering ensemble method based on swarm intelligence Yue Ming NJIT#: 31351707
About my project
Clustering algorithms, one important and normal used part of unsupervised machine learning methods, aim to discover the natural groupings of patterns, points, or objects. They divide a given dataset into multiple classes according to data similarity. In this work, I choose three clustering algorithms, k-MWO, AP clustering and DBSCAN. k-MWO clustering algorithm combines mussels wandering optimization (MWO), a heuristic method based on swarm intelligence, with classic k-means clustering algorithm. Affinity propagation (AP) takes as input measures of similarity between pairs of data points. DBSCAN is a density based algorithm for dealing with arbitrary shapes.
Simulation & Result
Dataset 1 2-dimension. 2 classes. Each class has 1000 elements. Conform to Gaussian distribution.
Initial data
After using K-MWO
Dataset 2 2-dimension. 2 classes. Arbitrary cluster ship. 1000 elements total.
Initial data
After using K-MWO
Analyze Dataset 1 Dataset 2 Algorithm time 12.536s 7.025s Cluster accuracy rate 90.6% 45.2%
Performance Comparison Dataset 1 (Arbitrary cluster ship with 1000 points) Dataset 2 (Gaussian Distribution with 2000 points) DBSCAN K-MWO AP Running Time 2.906s 7.025s 37.638s Cluster Number 2 24 Accuracy Rate 99.6% 45.2% 9.76% DBSCAN K-MWO AP Running Time 5.864s 12.536s 67.274s Cluster Number 1 2 Accuracy Rate 42.7% 90.6% 89.3%
Conclusion This paper presents a new clustering algorithm based on a new swarm intelligence model, called mussels wandering optimization (MWO). It makes full use of the global optimization ability of MWO and the local search ability of K-means. We test K-MWO, DBSCAN, AP algorithm in 2 same dataset.