{"id":{"repo_id":"umkc","oai_identifier":"oai:mospace.umsystem.edu:10355/95107"},"canonical_url":"https://search.dev.ndltd.org/etd/umkc/oai:mospace.umsystem.edu:10355/95107","repository":{"repo_id":"umkc","name":"University of Missouri - Kansas City","base_url":"https://mospace.umsystem.edu/oai/request"},"display":{"title":"Multimedia Big Data Analytics and Fusion for Data Science","abstract":"Big data is becoming increasingly prevalent in people's everyday lives due to the enormous quantity of data generated from social and economic activities worldwide. As a result, extensive research has been undertaken to support the big data revolution. However, as data grows in volume, traditional data analytic methods face various challenges—especially when raw data comes in multiple forms and formats. This dissertation proposes a multimodal big data analytics and fusion framework that addresses several challenges in data science for handling and learning from multimodal big data. The proposed framework addresses issues during a standard data science project workflow, including data fusion, spatio-temporal deep feature extraction, and model training optimization strategy. First, a hierarchical graph fusion network is presented to capture the inter-modality correlations among modalities. The network hierarchy models the modality-wise combinations with gradually increased complexity to explore all n-modality interactions. Next, an adaptive spatio-temporal graph network is proposed to capture the hidden patterns from spatio-temporal data. It exploits local and global node correlations by improving the pre-defined graph Laplacian and automatically generates the graph adjacency matrix based on a data-driven method. In addition, a dynamic multi-task learning method is introduced to optimize the model training progress by dynamically adjusting the loss weights assigned to each task. It systematically monitors the sample-level prediction errors, task-level weight parameter changing rate, and iteration-level total loss to adjust the weight balance among tasks. The proposed framework has been evaluated on various datasets, including disaster event videos, social media, traffic flow, and other public datasets.","abstract_html":"Big data is becoming increasingly prevalent in people's everyday lives due to the enormous quantity of data generated from social and economic activities worldwide. As a result, extensive research has been undertaken to support the big data revolution. However, as data grows in volume, traditional data analytic methods face various challenges—especially when raw data comes in multiple forms and formats. This dissertation proposes a multimodal big data analytics and fusion framework that addresses several challenges in data science for handling and learning from multimodal big data. The proposed framework addresses issues during a standard data science project workflow, including data fusion, spatio-temporal deep feature extraction, and model training optimization strategy. First, a hierarchical graph fusion network is presented to capture the inter-modality correlations among modalities. The network hierarchy models the modality-wise combinations with gradually increased complexity to explore all n-modality interactions. Next, an adaptive spatio-temporal graph network is proposed to capture the hidden patterns from spatio-temporal data. It exploits local and global node correlations by improving the pre-defined graph Laplacian and automatically generates the graph adjacency matrix based on a data-driven method. In addition, a dynamic multi-task learning method is introduced to optimize the model training progress by dynamically adjusting the loss weights assigned to each task. It systematically monitors the sample-level prediction errors, task-level weight parameter changing rate, and iteration-level total loss to adjust the weight balance among tasks. The proposed framework has been evaluated on various datasets, including disaster event videos, social media, traffic flow, and other public datasets.","abstract_has_math":false,"creators":["Wang, Tianyi"],"institution":"University of Missouri--Kansas City","degree_name":null,"degree_level":"Doctoral","degree_discipline":"Computer Science (UMKC)","degree_department":null,"school":null,"contributors":[],"advisors":["Chen, Shu-Ching, 1963-"],"committee_chairs":[],"committee_members":[],"year":2023,"date_issued":"2023","date_published":"2023","updated_at":"2026-07-24T05:16:57Z","subjects":[],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/10355/95107","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Chen, Shu-Ching, 1963-"]},{"key":"dc:creator","label":"Author","values":["Wang, Tianyi"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2023-05-19T15:12:57Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2023-05-19T15:12:57Z"]},{"key":"dc:date.issued","label":"Date","values":["2023"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computer Science (UMKC)","Electrical and Computer Engineering (UMKC)"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Doctoral"]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Missouri--Kansas City","Ph.D. (Doctor of Philosophy)"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/10355/95107"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Title from PDF of title page, viewed May 24, 2023","Dissertation advisor: Shu-Ching Chen","Vita","Includes bibliographical references (pages 178-212)","Dissertation (Ph.D.)--Department of Computer Science and Electrical Engineering. University of Missouri--Kansas City, 2023"]},{"key":"dc:description.abstract","label":"Abstract","values":["Big data is becoming increasingly prevalent in people's everyday lives due to the enormous quantity of data generated from social and economic activities worldwide. As a result, extensive research has been undertaken to support the big data revolution. However, as data grows in volume, traditional data analytic methods face various challenges—especially when raw data comes in multiple forms and formats. This dissertation proposes a multimodal big data analytics and fusion framework that addresses several challenges in data science for handling and learning from multimodal big data. The proposed framework addresses issues during a standard data science project workflow, including data fusion, spatio-temporal deep feature extraction, and model training optimization strategy. First, a hierarchical graph fusion network is presented to capture the inter-modality correlations among modalities. The network hierarchy models the modality-wise combinations with gradually increased complexity to explore all n-modality interactions. Next, an adaptive spatio-temporal graph network is proposed to capture the hidden patterns from spatio-temporal data. It exploits local and global node correlations by improving the pre-defined graph Laplacian and automatically generates the graph adjacency matrix based on a data-driven method. In addition, a dynamic multi-task learning method is introduced to optimize the model training progress by dynamically adjusting the loss weights assigned to each task. It systematically monitors the sample-level prediction errors, task-level weight parameter changing rate, and iteration-level total loss to adjust the weight balance among tasks. The proposed framework has been evaluated on various datasets, including disaster event videos, social media, traffic flow, and other public datasets."]},{"key":"dc:title","label":"Title","values":["Multimedia Big Data Analytics and Fusion for Data Science"]}]}],"canonical_facts":{"dc:contributor.advisor":["Chen, Shu-Ching, 1963-"],"dc:creator":["Wang, Tianyi"],"dc:date.accessioned":["2023-05-19T15:12:57Z"],"dc:date.available":["2023-05-19T15:12:57Z"],"dc:date.issued":["2023"],"dc:description":["Title from PDF of title page, viewed May 24, 2023","Dissertation advisor: Shu-Ching Chen","Vita","Includes bibliographical references (pages 178-212)","Dissertation (Ph.D.)--Department of Computer Science and Electrical Engineering. University of Missouri--Kansas City, 2023"],"dc:description.abstract":["Big data is becoming increasingly prevalent in people's everyday lives due to the enormous quantity of data generated from social and economic activities worldwide. As a result, extensive research has been undertaken to support the big data revolution. However, as data grows in volume, traditional data analytic methods face various challenges—especially when raw data comes in multiple forms and formats. This dissertation proposes a multimodal big data analytics and fusion framework that addresses several challenges in data science for handling and learning from multimodal big data. The proposed framework addresses issues during a standard data science project workflow, including data fusion, spatio-temporal deep feature extraction, and model training optimization strategy. First, a hierarchical graph fusion network is presented to capture the inter-modality correlations among modalities. The network hierarchy models the modality-wise combinations with gradually increased complexity to explore all n-modality interactions. Next, an adaptive spatio-temporal graph network is proposed to capture the hidden patterns from spatio-temporal data. It exploits local and global node correlations by improving the pre-defined graph Laplacian and automatically generates the graph adjacency matrix based on a data-driven method. In addition, a dynamic multi-task learning method is introduced to optimize the model training progress by dynamically adjusting the loss weights assigned to each task. It systematically monitors the sample-level prediction errors, task-level weight parameter changing rate, and iteration-level total loss to adjust the weight balance among tasks. The proposed framework has been evaluated on various datasets, including disaster event videos, social media, traffic flow, and other public datasets."],"dc:identifier.uri":["https://hdl.handle.net/10355/95107"],"dc:title":["Multimedia Big Data Analytics and Fusion for Data Science"],"thesis:degree_discipline":["Computer Science (UMKC)","Electrical and Computer Engineering (UMKC)"],"thesis:degree_level":["Doctoral"],"thesis:institution_name":["University of Missouri--Kansas City","Ph.D. (Doctor of Philosophy)"]},"updated_at":"2026-07-24T05:16:57Z"}