Online content clustering using variant K-Means Algorithms

Abstract

We live at a time when so much information is created. Unfortunately, much of the information is redundant. There is a huge amount of online information in the form of news articles that discuss similar stories. The number of articles is projected to grow. The growth makes it difficult for a person to process all that information in order to update themselves on a subject matter. There is an overwhelming amount of similar information on the internet. There is need for a solution that can organize this similar information into specific themes. The solution is a branch of Artificial intelligence (AI) called machine learning (ML) using clustering algorithms. This refers to clustering groups of information that is similar into containers. When the information is clustered people can be presented with information on their subject of interest, grouped together. The information in a group can be further processed into a summary. This research focuses on unsupervised learning. Literature has it that K-Means is one of the most widely used unsupervised clustering algorithm. K-Means is easy to learn, easy to implement and is also efficient. However, there is a horde of variations of K-Means. The research seeks to find a variant of K-Means that can be used with an acceptable performance, to cluster duplicate or similar news articles into correct semantic groups. The research is an experiment. News articles were collected from the internet using gocrawler. gocrawler is a program that takes Universal Resource Locators (URLs) as an argument and collects a story from a website pointed to by the URL. The URLs are read from a repository. The stories come riddled with adverts and images from the web page. This is referred to as a dirty text. The dirty text is sanitized. Sanitization is basically cleaning the collected news articles. This includes removing adverts and images from the web page. The clean text is stored in a repository, it is the input for the algorithm. The other input is the K value. All K-Means based variants take K value that defines the number of clusters to be produced. The stories are manually classified and labelled. The labelling is done to check the accuracy of machine clustering. Each story is labelled with a class to which it belongs. The data collection process itself was not unsupervised but the algorithms used to cluster are totally unsupervised. A total of 45 stories were collected and 9 manual clusters were identified. Under each manual cluster there are sub clusters of stories talking about one specific event. The performance of all the variants is compared to see the one with the best clustering results. Performance was checked by comparing the manual classification and the clustering results from the algorithm. Each K-Means variant is run on the same set of settings and same data set, that is 45 stories. The settings used are, • Dimensionality of the feature vectors, • Window size, • Maximum distance between the current and predicted word in a sentence, • Minimum word frequency, • Specified range of words to ignore, • Number of threads to train the model. • The training algorithm either distributed memory (PV-DM) or distributed bag of words (PV-DBOW), • The initial learning rate. The learning rate decreases to minimum alpha as training progresses, • Number of iterations per cycle, • Final learning rate, • Number of clusters to form, • The number of times the algorithm will be run, • The method used for initialization. The results obtained show that K-Means can perform better than K-Modes. The results are tabulated and presented in graphs in chapter six. Clustering can be improved by incorporating Named Entity (NER) recognition into the K-Means algorithms. Results can also be improved by implementing multi-stage clustering technique. Where initial clustering is done then you take the cluster group and further cluster it to achieve finer clustering results.