5th Annual Big Data & Business Analytics Symposium
Research Poster Session
The poster session runs throughout the symposium with formal presentation at 9:30am-10:30am on Friday, March 23rd in Room 285 of the Student Center Building. This session allows our faculty/student research groups to showcase the big data tools, capabilities, and projects to industry attendees. As a bonus the lead student for each poster gets a free registration to the symposium! Did you miss a posters at the Symposium or did you run our of time to read all of the information? You can download our entire 2018 poster collection here.
2018 Poster Session Abstracts:
Obesity Risk Factors Ranking Using Multi-Task Learning
Authors: Lu Wang, Dongxiao Zhu, Elizabeth Towner, Ming Dong (Wayne State University).
Obesity is one of the leading preventable causes of death in the United States (U.S.). Risk factor analysis is a process to identify and understand the risk factors contributing to a particular disease, and is an imperative component in the development of efficient and effective prevention and intervention efforts. We employ the framework of multi-task learning (MTL) to identify a ranked list of obesity risk factors for each subpopulation/participant (task) simultaneously with utilizing appropriate shared information across tasks. By learning multiple related tasks at the same time, MTL provides a paradigm to rank risk factors at multi-levels.
Deep Reinforcement Learning for Optimizing Carpooling Policies
Authors: Ishan Jindal (Wayne State University), Zhiwei (Tony) Qin (DiDi Research America), Xuewen Chen, Matthew Nokleby (Wayne State University), Jieping Ye (DiDi Research America).
We developed a deep reinforcement learning (RL) based system to learn an efficient policy for rideshare carpooling that maximizes the driver's revenue and minimizes traffic congestion using fewer cars to fulfill the same amount of demand. For this purpose, we developed a detailed carpooling simulator using the NYC taxi trip dataset. We observed that in the densely populated regions such as in Downtown Manhattan, it is always favorable to do a carpool that is RL learned policy and fixed policy performed equally well. On the other hand, for the less populated areas such as Uptown Manhattan, RL learned policy significantly outperforms the fixed policy.
Resistance Spot Welding and Big Data Analytics: Opportunities and Challenges
Authors: Saeed Z.Gavidel, Shiyong Lu, Jeremy L. Rickli (Wayne State University).
Resistance Spot Welding (RSW) is a complex process with high-veracity data and highly non-linear relations between process parameters. Shallow predictive models like Random Forests and SVMs are less-efficient in high-veracity scenarios. Deep Neural Nets deployed on Big Data frameworks like Apache Spark, can significantly (average 25%) improve efficiency of RSW predictions.
Timely Detection of Abnormal Inactivity Using Smart Meter Data
Authors: Yanchao Liu, Tingli Hu, Caisheng Wang (Wayne State University).
A probabilistic algorithm is proposed for detecting abnormal inactivities, such as immobilizing medical conditions or sudden death of elderly or disabled occupants, within a single-occupied household based on smart meter readings.
bsnsing: A New R Package for Classification Tree Modeling
Authors: Yanchao Liu (Wayne State University).
The "bsnsing" package provides functions for building a decision tree classifier and making predictions. It is built upon the idea of optimal Boolean sensing and can generate better comprehensibility and predictive accuracy than alternative tree methods.
Machine Learning Models for Rapid Detection of Streptococcus Pyogenes with Raman Spectroscopy
Authors: Brandy Broadbent, Ehsan Majidi, Sally Yurgelevic, Michelle Brusatori, Gregory Auner (Wayne State University).
Clinical detection of Streptococcus pyogenes (strep) require taking a throat swab and culturing it for at least 24 hrs before getting a diagnosis. Raman spectroscopy can shorten this cycle to 30 min. or less. Here we use LDA, QDA, SVM, and random forest to analysis which model best classifies strep.
Practical Economic Model Predictive Control Design through Nonlinear Model Identification
Authors: Laura Giuliani, Helen Durand (Wayne State University).
Economic model predictive control (EMPC) is an optimization-based control design that can enhance the profitability of industrial chemical processes by determining control actions that operate a process in an economically-optimal fashion, subject to a process model and constraints. Unlike traditional model predictive control designs that force a process to operate at steady-state, EMPC may operate a process in a time-varying fashion and therefore requires a sufficiently accurate process model for the purpose of developing an appropriate objective function and constraints for the controller, and for allowing more accurate predictions of the process state within the controller when it is selecting optimal control actions. To address these challenges, we develop an EMPC design that can operate the process in a manner that generates desired data to be used in seeking to develop and verify more physically-based process models, and using these models to then update the control design on-line to enhance its ability to achieve operating goals.
Rapid Detection of Clostridium Difficile Toxins Using Raman Spectroscopy
Authors: S. Kiran Koya, Jonathan V. Martin, Sally Yurgelevic, Michelle Brusatori, Changhe Huang, David M. Liberati, Gregory W. Auner, Lawrence N. Diebel (Wayne State University).
Clostridium difficile infection (CDI) is due to the effects of toxins, toxin A & toxin B on the host. Current technologies have low sensitivity to detect C. difficile toxemia in serum. Raman spectroscopy (RS) was performed on drop of toxin-spiked serum and control serum. Spectra were analyzed by Support Vector Classification (C-SVC) method. Model performance was assessed by cross-validation and bootstrap methods. At 0.1 pg/ml conc. toxin spiked serum was distinguished from control serum 100% with cross-validation error rate below 8%.
Using Common Value Auction in Cultural Algorithm to Enhance Robustness and Resilience of Social Knowledge Distribution Systems
Authors: Anas Al-Tirawi, Robert Reynolds (Wayne State University).
In Cultural Systems, there are many ways to collect and distribute problem-solving knowledge within social networks. Such mechanisms include games, auctions, and various voting mechanisms. Here, a new auction mechanism, Common Value auctions, is presented. In this poster, we describe how Common Value auctions are used to distribute problem-solving knowledge within a given model of social systems.
Reduction of Uncertainty in ED Patient Disposition Decision for Early Resource Allocation in Inpatient Units
Authors: Seung Yup Lee, Ratna Babu Chinnam, Evrim Dalkiran (Wayne State University), Seth Krupp, Michael Nauss (Henry Ford Health System).
Health information technology (HIT) has widely been adopted in health care facilities including emergency departments (ED). While there are many research works to predict future information of ED patients, there is a lack of study that takes the operationalization of future information into consideration when models are built. In this study, we investigate the patterns of uncertainty reduction in the disposition decisions of ED patients. The results provide a number of interesting insights for operationalizing outcomes of predictions.
Market Basket Analysis Using Large-Scale Graph-Based Methods
Authors: Elham Nosrat, Ratna Babu Chinnam (Wayne State University).
Market basket analysis is the science of discovering customer purchase behavior in order to design marketing strategies. Our proposed solution is based on the graph of transactions, which updates dynamically in time. The output can be fed directly to other subsystems such as assortment planning, inventory planning, pricing and promotions, etc. to take suitable actions.
Improvement to the Prediction of Fuel Cost Distribution Using ARIMA Model
Authors: Zhongyang Zhoa, Chang Fu, Caisheng Wang, Carol J. Miller (Wayne State University).
Availability of a validated, realistic fuel cost model is a prerequisite to the development and validation of new optimization methods and control tools. This research uses an autoregressive integrated moving average (ARIMA) model with historical fuel cost data in development of a three-step-ahead fuel cost distribution prediction.
Detecting Qualitative Changes in Biological Systems
Authors: Cristina Mitrea (Wayne State University), Aliccia Bollig-Fischer (Barbara Ann Karmanos Cancer Institute), Călin Voichița (Advaita Bioinformatics), Michele Donato (Stanford University), Roberto Romero (National Institutes of Health), Sorin Drăghici (Wayne State University).
Currently, most diseases are diagnosed only after significant disease-associated changes have occurred. This is particularly true in the case of complex conditions such as cancer. We propose a qualitative change detection (QCD) approach able to identify when systemic qualitative changes in biological systems happen, thus opening the possibility of therapeutic interventions before the occurrence of symptoms. QCD takes as input large datasets of sequential measurements as described by time series (or progressive disease stages) together with known interactions described by biological networks and applies an impact analysis approach to identify the time interval in which the system transitions to a different qualitative state.
Observational Data-Driven Modeling and Optimization of Manufacturing Processes
Authors: Najibe Sadati, Ratna Babu Chinnam, Milad Zafar Nezhad (Wayne State University).
The dramatic increase of observational data across industries provides unparalleled opportunities for data-driven decision making and management, including the manufacturing industry. In the context of production, data-driven approaches can exploit observational data to model, control and improve process performance. When supplied by observational data with adequate coverage to inform the true process performance dynamics, they can overcome the cost associated with intrusive controlled designed experiments and can be applied for both process monitoring and improvement. We propose a novel integrated approach that uses observational data for identifying significant control variables while simultaneously facilitating process parameter design. We evaluate our method using data from synthetic experiments and also apply it to a real-world case setting from a tire manufacturing company.
Representation Learning with Autoencoders for Modeling Length of Stay in ICUs
Authors: Najibe Sadati, Dongxiao Zhu, Milad Zafar Nezhad, Ratna Babu Chinnam (Wayne State University).
Increasing volume of Electronic Health Records (EHR) in recent years provides huge opportunities for data scientists to collaborate in different aspects of healthcare by applying advanced analytics to these EHR clinical data. A key requirement, however, is obtaining meaningful insights from high dimensional, sparse and complex biological/ biomedicine data. Data mining and machine learning approaches can address this challenge by performing feature representation (feature engineering or unsupervised learning) in order to build more reliable and informative features from personalized clinical data and subsequently training the supervised learning using this represented features. In this study, we propose a prediction framework based on deep learning for feature engineering in a healthcare setting. We perform our approach on two EHR datasets obtained from e-ICU collaborative research database to train a predictive model of length of stay (LOS) in intensive care units for neuro-patients and cardiac-patients.
Characterizing Customer Choice Modeling for Introducing Niche Products
Elham Nosrat, Ratna Babu Chinnam, Evrim Dalkiran (Wayne State University).
In today's agile market, products emerge and die faster than ever. A company's ability to decide which products to keep offering, which products to retire, and which new products to introduce is crucial to its profitability and competitiveness. This research proposes a method to analyze the set of products currently offered by a firm (store, retailer, or manufacturer) and estimate the attractiveness of each of these products and any potential new products considering different customers' behaviors. Specifically, we develop a model to split customers into several segments, wherein customers have similar preferences towards the products. For each customer segment, we then construct another model to obtain the preference ranking of products (current and new) based on their attributes. Attribute-based analysis allows us to compare existing products with potential ones in a quantitative manner.
A Sample Average Approximation-Based Parallel Algorithm for Application Placement in Edge Computing Systems
Authors: Hossein Badri, Tayebeh Bahreini, Daniel Grosu, Kai Yang (Wayne State University).
We address the application placement problem in Mobile Edge Computing (MEC) systems. We develop a multi-stage stochastic programming model and design a novel parallel greedy algorithm based on the Sample Average Approximation method. We evaluate the performance of the proposed algorithm by conducting an extensive experimental analysis using data extracted from a real-world dataset.
Driver Age Estimation through Deep Learning for Autonomous Safety
Authors: Shixing Chen and Ming Dong (Wayne State University).
We present a novel ranking-based Convolutional Neural Network architecture for driver age estimation. Furthermore, we introduce a transfer learning framework to solve the domain discrepancy caused by images from various sources.
Evaluating the Performance of Coordinated Signal Timing: A Comparison of Common Data Types with Automated Vehicle Location Data
Authors: Stephen Remias, Jonathan Waddell, Jenna Kirsch (Wayne State University).
Performance measures are essential for managing transportation systems, including signalized corridors. Coordination is an essential element of signal timing, enabling reliable progression of traffic along corridors. Improved progression leads to less user delay, which leads to user cost savings and lower vehicle emissions. This project presents a comparative study of signal coordination assessment using four different technologies. These technologies include detector-based high-resolution controller data, Bluetooth/Wi-Fi sensors, segment-based probe vehicle data, and automated vehicle location data consisting of GPS-based vehicle trajectories, representing the data anticipated from emerging connected vehicle technologies.
Predictive Deep Network with Leveraging Clinical Measure as Auxiliary Task
Authors: Xiangrui Li, Dongxiao Zhu, Phillip Levy (Wayne State University).
Auxiliary-task-augmented-network (ATAN) is proposed to build a predictive model. ATAN leverages measures that are clinically related to the primary target as auxiliary predictive tasks under the framework of multi-task learning as regularization. We apply ATAN in a clinical dataset of hypertension to demonstrate its effectiveness.
Visualization Tools for Electricity Emission Intensity Data
Authors: Amir Kamjou, Carol Miller (Wayne State University).
Dashboards are providing powerful means to monitor conditions at a glance. The aim of this project was to provide a visualized action tool for the energy used and emissions released due to electricity consumption at the Wayne State University Campus by providing a single page dashboard which is developed to show a graphical presentation of historical and real-time data.
Data Integration for Water Network Models to Simulate Historical Operation
Authors: Amir Kamjou, Shawn McElmurry, Carol Miller (Wayne State University).
The aim of this project is to come up with a description for water quality changes in the City of Flint's water network system over time. Water quality can be defined in many ways but we are trying to describe age first and then Chlorine residual and other reactions.
DATAVIEW: Big Data Infrastructure for Resistance Spot Welding Domain
Authors: Changxin Bai, Saeed Z.Gavidel, Shiyong Lu, Jeremy L.Rickli (Wayne State University).
Resistance Spot Welding (RSW) applications can benefit from predictive analytics and high performance computing resources such as Amazon clouds. DATAVIEW, a big data infrastructure, can process high veracity RSW data more efficiently with respect to traditional pipelines constructed from R. Results indicate that DATAVIEW outperforms R from the perspective of training and testing times and resource usage.
Huron-to-Erie Water Quality Data Platform
Authors: Lanyu Xu, Carol J. Miller (Wayne State University).
To address the challenges of creation of a sustainable urban environment, we introduce a mass-oriented, user-friendly and cloud-based data platform to provide integrative water quality data in one of the most critical urban corridors of the Great Lakes system. Several example applications are provided of platform use for temporal and spatial characterization of intake water source quality and urban beach health.
Prioritizing Locations to Mitigate Crashes by the Use of Michigan Crash Data
Authors: Ilyas Ustun (Wayne State University).
Traffic crashes have a significant impact on the economy both in the form of property damage and also in the form of lost time. The congestion likely to happen in busy areas will cause waste of gas and extra emissions of harmful gases. The worst are fatalities which no one wants to happen. Identifying the crash-prone locations will help traffic safety, transportation planning, and law enforcement to prioritize their efforts and resources to minimize the risk of accidents. Michigan Crash Data provides a rich means for this purpose. This poster presents the results of such an effort. Crash-prone locations have been identified along the factors that are correlated. It is believed that by making use of data, number of crashes can be minimized and their effects mitigated.
Automotive Dealership Management: Deriving Tailored Recommendations Using Big Data
Author: Haidar Almohri and Ratna Babu Chinnam (Wayne State University), Mark Colosimo, David Creech, Daniel Mathias (Urban Science)
Deep Neural Architecture for Multi-Modal Retrieval
Authors: Saeid Balaneshin, Alexander Kotov (Wayne State University)
We present an algorithm for joint text and image retrieval in response to queries combining textual or visual modalities. We propose a deep neural architecture combining Skip-gram method as well as convolutional and recurrent neural networks to represent textual and visual modalities of both the queries and retrieval units in a joint embedding space.
See Posters from the 2018 Symposium below:
For questions regarding the Poster Session, please contact student coordinator: Seung-Yup Lee. The Symposium organizing committee prints and mounts posters for students thanks to the generous support of our sponsors. The template for the posters can be found here.