Parallelization Strategies for Distributed Non Negative Matrix Factorization

Dimensionality reduction and clustering have been the subject of intense research efforts over the past few years [2]. They offer an approach of knowledge extraction from huge amounts of data. Although some of these techniques are effective at achieving lower data dimensions, very few focused on scaling the techniques to tackle data sets that might not fit into memory. Non negative matrix factorization is (NMF) one of the effective techniques that can be used to achieve dimensionality reduction, missing data prediction and clustering. NMF has been parallelized through shared memory and distributed memory. Our contribution lies in reaching a higher level of parallelism through proposing a new block division technique on a hadoop framework. Furthermore, we use the block-based technique to design an enhanced cascaded NMF [6]. We compare the division techniques that we propose, block-based and cascaded over block-based division to the column-based technique which exists in the literature [12]. The block-based technique performs 18% percent faster than the column based. It achieves higher convergence value than the cascaded technique by 23%