MB5: Use-Inspired, Data-Driven Modeling in Various Fields

Room: Old Main Academic Center 3110

Webex Link

OrganizerJonathan Barlow, Mississippi State University

Long Tian, NSPARC, Mississippi State University

Time: 10:00 am - 10:25 am (CST)

Title: Workforce Pattern in Mississippi and Next Career Recommendation

Abstract: To connect unemployed people with job opening in job market is a big challenge in this post-pandemic workforce market. With help of artificial intelligent and big data, we have addressed this issue to create a deep learning model that provide realistic job recommendation for unemployed people based on the employed history of each individual. At first, the transfer learning model is applied for the job title and O*NET Standard Occupational Classification (OSOC) code matching based on data set from Workforce Innovation and Opportunity Act (WIOA) system in Mississippi Department of Employment Security (MDES), where OSOC is a standard occupational classification-based system that used by U.S. federal agencies to classify workers into occupational categories. Then Long Short-Term Memory (LSTM) model is created for career pathway prediction, which generate top three recommendation job OSOC based on the individual’s employ history. The accuracy of created model is 72.8% when individual’s education history records are included.

             Tony Luczak, NSPARC, Mississippi State University 

Time: 10:25 am - 10:50 am (CST)

Title: Survey on Using Computer Vision to Push 2D Video to 3D Space

Abstract: The use of two-dimensional (2D) video for golf instruction has become ubiquitous due to smart camera devices. It is commonplace for instructors to draw lines and circles to represent where golfers should position their body and club to change movement trajectories in an attempt to change swing patterns. However, there are challenges to this type of approach including: (a) how 2D reference symbols relate to three-dimensional (3D) space, (b) skill information is presented post performance, and (c) relying on a subject matter expert increases the cost of learning and training. To overcome these challenges, the use of artificial intelligence (AI) and computer vision(CV)techniques and methods can mathematically infer 2D objects to 3D space allowing for the capture and tracking of the objects, provide real-time feedback on movement performance, and create an AI agent to assist with learning and training. This survey will explore the requirements and constraints involved in 2D to 3D video tracking. Understanding theCV requirements for depth perception and tracking while knowing the constraints based on the availability of cameras and other sensors can help future research in the development of expert models of human movement performance and AI self-learning and training systems.

            Yonghua Yan, Jackson State University

Time: 10:50 am - 11:15 am (CST)

Title: Study on the Effect of MVG on High-Speed Flows with Different Mach Numbers

Abstract: Large Eddy Simulations (LES) of Micro Vortex Generator (MVG) controlled flows were conducted in this study. The effect of (MVG) on the high-speed flow field with differentMach numbers (from Ma 1.5 to Ma 5.0) was investigated. The downstream flow structures, especially the large vortex structure generated in the upper boundary layer that plays a very important role in Shock Boundary Layer Interaction (SBLI), were analyzed in detail. It is found that with the increase of Ma, the ring-like vortices are harder to detect. At a higher Mach number, the momentum deficit is located closer to the lower boundary layer. Due to the lower position, more interaction between the ring-like vortices the lower vortex structures in the boundary layer was observed, resulting in greater deformation of the ring-like vortices. The back flow zone behind the MVG increases with increasing Ma number, and the flow separation is more complicated.

This work is in collaboration with Caixia Chen.

             Tianyu Li, NSPARC, Mississippi State University

Time: 11:15 am - 11:40 am (CST)

Title: Unemployment Insurance Predictive Analytics with Unbalanced Data

Abstract: Prediction at the time of admission into unemployment insurance (UI) can greatly enhance the quality of UI and help with accurate planning for UI benefits. The objective is to build data-driven models using Machine Learning (ML) and Artificial Intelligence (AI) that can predict the individual’s length of UI benefits based on various dimensions of a large-scale data set, including applicant, employment, and economic indicators. The prediction model is built by a series of steps, including feature transformation, feature selection, and ensemble of classifiers. The existing classifiers have limitations in coping with imbalanced training datasets as they are sensitive to the proportions of the different classes. As a consequence, the UI predictive model tends to favor the class with the “majority class”, which may lead to misleading accuracy. This study aims to provide solutions to the imbalanced data problem, which is the situation where the number of observations is not the same for all the classes in a classification dataset. A correct classification of a "minority” class was believed to be helpful to dimension reduction and accuracy improvement for further exploration, which will eventually enableUI providers to predict the distribution of UI benefits durations for different job types.

             Patrick Nelsen, NSPARC, Mississippi State University

Time: 11:40 am - 12:05 pm (CST)

Title: Markerless Human Pose Estimation to Improve Human Pose Estimation: A Work In Progress

Abstract: Here we report our work-in-progress to advance 2-dimensional (2D) and 3-dimensional (3D) human pose estimation (HPE) research to analyze golf swings. There are four main components of this research: (a) data, (b) 2D HPE, (c) 3D HPE, and (d) an iPadOS application. We created the first high-frame rate contextual human pose golf data set, which contains 1,350 videos of golf swings with 3D motion capture. Along with the motion capture data, we designed and enacted a super key frame methodology for humans to label each of the 650,000 frames with a human pose skeleton of the golfer to create a full video, 2D pose, and 3D pose dataset of golf data. State-of-the-art neural network models in both 2D and 3D HPE were trained and fine-tuned on our golf dataset. These models, once optimized for golf HPE, will be integrated into an iPadOS application that will allow for the user to record, analyze, and view their own golf swings in 3D. This paper presents the progress made thus far, current problems and pitfalls, and the pathway forwards to continue HPE research in golf.

This work is in collaboration with Tony Luczack, Kait Jackson, Long Tian, John Ball, Martin Duclos, Michael Taquino, and Steven M Grace.