Mining complex data in highly streaming environments

Razavi Hesabi, Z 2019, Mining complex data in highly streaming environments, Doctor of Philosophy (PhD), Science, RMIT University.

Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Razavi_Hesabi.pdf Thesis application/pdf 11.12MB
Title Mining complex data in highly streaming environments
Author(s) Razavi Hesabi, Z
Year 2019
Abstract Data is growing at a rapid rate because of advanced hardware and software technologies and platforms such as e-health systems, sensor networks, and social media. One of the challenging problems is storing, processing and transferring this big data in an efficient and effective way. One solution to tackle these challenges is to construct synopsis by means of data summarization techniques. Motivated by the fact that without summarization, processing, analyzing and communicating this vast amount of data is inefficient, this thesis introduces new summarization frameworks with the main goals of reducing communication costs and accelerating data mining processes in different application scenarios. Specifically, we study the following big data summarizaion techniques:(i) dimensionality reduction;(ii)clustering,and(iii)histogram, considering their importance and wide use in various areas and domains.

In our work, we propose three different frameworks using these summarization techniques to cover three different aspects of big data:"Volume","Velocity"and"Variety" in centralized and decentralized platforms. We use dimensionality reduction techniques for summarizing large 2D-arrays, clustering and histograms for processing multiple data streams.

With respect to the importance and rapid growth of emerging e-health applications such as tele-radiology and tele-medicine that require fast, low cost, and often lossless access to massive amounts of medical images and data over band limited channels,our first framework attempts to summarize streams of large volume medical images (e.g. X-rays) for the purpose of compression. Significant amounts of correlation and redundancy exist across different medical images. These can be extracted and used as a data summary to achieve better compression, and consequently less storage and less communication overheads on the network. We propose a novel memory-assisted compression framework as a learning-based universal coding, which can be used to complement any existing algorithm to further eliminate redundancies/similarities across images. This approach is motivated by the fact that, often in medical applications, massive amounts of correlated images from the same family are available as training data for learning the dependencies and deriving appropriate reference or synopses models. The models can then be used for compression of any new image from the same family. In particular, dimensionality reduction techniques such as Principal Component Analysis (PCA) and Non-negative Matrix Factorization (NMF) are applied on a set of images from training data to form the required reference models. The proposed memory-assisted compression allows each image to be processed independently of other images, and hence allows individual image access and transmission.

In the second part of our work,we investigate the problem of summarizing distributed multidimensional data streams using clustering. We devise a distributed clustering framework, DistClusTree, that extends the centralized ClusTree approach. The main difficulty in distributed clustering is balancing communication costs and clustering quality. We tackle this in DistClusTree through combining spatial index summaries and online tracking for efficient local and global incremental clustering. We demonstrate through extensive experiments the efficacy of the framework in terms of communication costs and approximate clustering quality.

In the last part, we use a multidimensional index structure to merge distributed summaries in the form of a centralized histogram as another widely used summarization technique with the application in approximate range query answering. In this thesis, we propose the index-based Distributed Mergeable Summaries (iDMS) framework based on kd-trees that addresses these challenges with data generative models of Gaussian mixture models (GMMs) and a Generative Adversarial Network (GAN). iDMS maintains a global approximate kd-tree at a central site via GMMs or GANs upon new arrivals of streaming data at local sites. Experimental results validate the effectiveness and efficiency of iDMS against baseline distributed settings in terms of approximation error and communication costs.
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Science
Subjects Database Management
Keyword(s) data summarization
big data
machine learning
distributed computing
data management
dimensionality reduction
index data structures
Version Filter Type
Access Statistics: 43 Abstract Views, 102 File Downloads  -  Detailed Statistics
Created: Fri, 15 Nov 2019, 10:33:31 EST by Keely Chapman
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us