Pal, Gautam 
ORCID: 0000-0002-2594-9699
(2021)
Multimodal Approach for Big Data Analytics and Applications.
PhD thesis, University of Liverpool.
|
Text
201261055_Sep2021.pdf - Unspecified Download (50MB) | Preview |
Abstract
The thesis presents multimodal conceptual frameworks and their applications in improving the robustness and the performance of big data analytics through cross-modal interaction or integration. A joint interpretation of several knowledge renderings such as stream, batch, linguistics, visuals and metadata creates a unified view that can provide a more accurate and holistic approach to data analytics compared to a single standalone knowledge base. Novel approaches in the thesis involve integrating multimodal framework with state-of-the-art computational models for big data, cloud computing, natural language processing, image processing, video processing, and contextual metadata. The integration of these disparate fields has the potential to improve computational tools and techniques dramatically. Thus, the contributions place multimodality at the forefront of big data analytics; the research aims at mapping and under- standing multimodal correspondence between different modalities. The primary contribution of the thesis is the Multimodal Analytics Framework (MAF), a collaborative ensemble framework for stream and batch processing along with cues from multiple input modalities like language, visuals and metadata to combine benefits from both low-latency and high-throughput. The framework is a five-step process: Data ingestion. As a first step towards Big Data analytics, a high velocity, fault-tolerant streaming data acquisition pipeline is proposed through a distributed big data setup, followed by mining and searching patterns in it while data is still in transit. The data ingestion methods are demonstrated using Hadoop ecosystem tools like Kafka and Flume as sample implementations. Decision making on the ingested data to use the best-fit tools and methods. In Big Data Analytics, the primary challenges often remain in processing heterogeneous data pools with a one-method-fits all approach. The research introduces a decision-making system to select the best-fit solutions for the incoming data stream. This is the second step towards building a data processing pipeline presented in the thesis. The decision-making system introduces a Fuzzy Graph-based method to provide real-time and offline decision-making. Lifelong incremental machine learning. In the third step, the thesis describes a Lifelong Learning model at the processing layer of the analytical pipeline, following the data acquisition and decision making at step two for downstream processing. Lifelong learning iteratively increments the training model using a proposed Multi-agent Lambda Architecture (MALA), a collaborative ensemble architecture between the stream and batch data. As part of the proposed MAF, MALA is one of the primary contributions of the research.The work introduces a general-purpose and comprehensive approach in hybrid learning of batch and stream processing to achieve lifelong learning objectives. Improving machine learning results through ensemble learning. As an extension of the Lifelong Learning model, the thesis proposes a boosting based Ensemble method as the fourth step of the framework, improving lifelong learning results by reducing the learning error in each iteration of a streaming window. The strategy is to incrementally boost the learning accuracy on each iterating mini-batch, enabling the model to accumulate knowledge faster. The base learners adapt more quickly in smaller intervals of a sliding window, improving the machine learning accuracy rate by countering the concept drift. Cross-modal integration between text, image, video and metadata for more comprehensive data coverage than a text-only dataset. The final contribution of this thesis is a new multimodal method where three different modalities: text, visuals (image and video) and metadata, are intertwined along with real-time and batch data for more comprehensive input data coverage than text-only data. The model is validated through a detailed case study on the contemporary and relevant topic of the COVID-19 pandemic. While the remainder of the thesis deals with text-only input, the COVID-19 dataset analyzes both textual and visual information in integration. Post completion of this research work, as an extension to the current framework, multimodal machine learning is investigated as a future research direction.
| Item Type: | Thesis (PhD) |
|---|---|
| Uncontrolled Keywords: | Multimodal Analysis, Lambda Architecture, Real-Time Machine Leaning, Real-Time Data Ingestion, Lifelong Machine Learning, Multi-agent Decision Making, Implicit Feedback Based Recommender System, Clickstream Analytics, E-commerce Analytics, Social Media Analytics, Streaming K-means Clustering, NoSQL Databases, Dimension Reduction, Apache Storm, Apache Kafka, Apache Flume |
| Divisions: | Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 23 Sep 2021 10:50 |
| Last Modified: | 18 Jan 2023 21:28 |
| DOI: | 10.17638/03137807 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3137807 |
Dimensions
Dimensions
