Abstract
Text document clusters are usually represented with respect to cluster centers called centroids. A clustering algorithm tries to find a good combination of centroids that satisfies the objective of clustering, i.e., assigning similar documents together and dissimilar documents in different clusters. The success of such an algorithm mainly depends on the ability to generate different potential solutions, i,e., combinations of centroids. However, most of the existing approaches such as k-means clustering and genetic algorithm (GA) depends heavily on the randomly generated initial solution(s). At times, such methods arc found to reach premature convergences due to getting stuck at local extrema, from which no more potential solutions can be explored. This paper proposes a modification to the crossover operations, which is the heart of GA to generate potential solutions, to overcome such limitations. Experimental evaluation also shows the performance improvements of available GA approaches when using the modified crossover operation.