Semi-supervised Heterogeneous Evolutionary Co-clustering

One of the challenges of the machine learning problem is the absence of sufficient number of labeled instances or training instances. At the same time generating labeled data is expensive and time consuming. The semi-supervised approach has shown promising results to solve the problem of insufficient or fewer labeled instance datasets. The key challenge is incorporating the semi-supervised knowledge into the heterogeneous data which is evolving in nature. Most of the prior work that uses semi-supervised knowledge has been performed on heterogeneous static data. The semi-supervised knowledge is incorporated into data which aid the clustering algorithm to obtain better clusters. The semi-supervised knowledge is provided as constrained based or distance based. I am proposing a framework to incorporate prior knowledge to perform co-clustering on the evolving heterogeneous data. This framework can be used to solve a wide range of problems dealing with text analysis, web analysis and image grouping. In the semi-supervised approach we incorporate the domain knowledge by placing the constraints which aid the clustering process in performing effective clustering of the data. In the proposed framework, I am using the constraint based semi-supervised non-negative matrix factorization approach to obtain the co-clustering on the heterogeneous evolving data. The constraint based semi-supervised approach uses the user provided must-link or cannot-link constraints on the central data type before performing co-clustering. To process the original datasets efficiently in terms of time and space I am using the low rank approximation technique to obtain the sparse representation of the input data matrix using the Dynamic Colibri approach.