More graduate defenses and presentations this week

CSE Graduate Defenses
CSE Graduate Defenses

Master's project: “An Integrated Web-Based Platform for Interactive Nitrogen Leaching Exploration”

Babak Samani

Advisor: Hongfeng Yu
Committee members: Haishun Yang, Mohammad Rashedul Hasan, Qiben Yan
Date: April 11, 2018
Time: 10:30 a.m.
Room: 352 Avery Hall

Abstract: While nitrogen (N) is an essential nutrient for corn, its leaching to groundwater is a serious environmental issue and a hazard to public health. N leaching is closely linked to soil texture, crop type, and weather factors, especially rainfall. Prediction of N leaching in cropping systems is critical to the improvement of crop management and the reduction of N leaching. The primary contribution of this project is to develop a web-based platform that provides an end-to-end solution in support of interactive N leaching estimation and exploration across Nebraska corn fields. At the front end, we design a map-based visualization interface that allows a user to interactively specify a region of interest (ROI) and time ranges. Heterogeneous datasets, including weather data, soil textures, and cropland data, within the ROI can be efficiently fetched from our back-end data management system that leverages multiple databases and data store techniques. Maize-N, an N leaching model, takes the fetched data as input and simulates the prediction of N leaching of ROI, which is visualized at the front end. Our design also tackles the security and performance of our system to ensure proper user experience. The platform provides an intuitive and effective tool that can facilitate farmers to appropriately manage the usage of N fertilizer and reduce the N leaching in their fields.



“Performance Evaluation of v-eNodeB using Virtualized Radio Resource Management”

Sai Keerti Teja Boddepalli

Advisor: Jitender Deogun
Committee members: Vinod Variyam, Ashok Saal
Date: April 12, 2018
Time: 11 a.m.
Room: Avery 256C

Abstract: With the demand upsurge for high bandwidth services, continuous increase in the number of cellular subscriptions, adoption of Internet of Things (IoT), and marked growth in Machine-to-Machine (M2M) traffic, there is great stress exerted on cellular network infrastructure. The present wireline and wireless networking technologies are rigid in nature and heavily hardware-dependent, as a result of which the process of infrastructure upgrade to keep up with future demand is cumbersome and expensive.

Software-defined networks (SDN) hold the promise to decrease network rigidity by providing central control and flow abstraction, which in current network setups are hardware-based. The embrace of SDN in traditional cellular networks has led to the implementation of vital network functions in the form of software that are deployed in virtualized environments. This approach to move crucial and hardware intensive network functions to virtual environments is collectively referred to as network function virtualization (NFV). Our work evaluates the cost reduction and energy savings that can be achieved by the application of SDN and NFV technologies in cellular networks.

In this thesis, we implement a virtualized eNodeB component (Radio Resource Management) to add agility to the network setup and improve performance, which we compare with a traditional resource manager. When combined with dynamic network resource allocation techniques proposed in Elastic Handoff, our hardware agnostic approach can achieve a greater reduction in capital and operational expenses through optimal use of network resources and efficient energy utilization.

Our simulation results of the experimental prototype show that the proposed model can deliver up to 33% reduction in energy consumption and revenue increase of as much as 27% by taking advantage of dynamic pricing. In addition, we study the ability of the model to honor service level agreements and propose plans to better handle these agreements under peak network load.




“Cost-Effective Techniques for Continuous Integration Testing”

Jingjing Liang

Advisor: Gregg Rothermel
Committee members: Sebastian Elbaum, Witawas Srisa-An
Date: April 12, 2018
Time: 12:30 p.m.
Room: 103C Avery

Abstract: Continuous integration (CI) development environments allow software engineers to frequently integrate and test their code. While CI environments provide advantages, they also utilize non-trivial amounts of time and resources. To address this issue, researchers have adapted techniques for test case prioritization (TCP) and regression test selection (RTS) to CI environments. In general, RTS techniques select test cases that are important to execute, and TCP techniques arrange test cases in orders that allow faults to be detected earlier in testing, providing faster feedback to develop- ers. In this thesis, we provide new TCP and RTS algorithms that make continuous integration processes more cost-effective.

To date, current TCP techniques under CI environments have operated on test suites, and have not achieved substantial improvements. Moreover, they can be in- appropriate to apply when system build costs are high. In this thesis we explore an alternative: prioritization of commits. We use a lightweight approach based on test suite failure and execution history that is highly efficient; our approach “continu- ously” prioritizes commits that are waiting for execution in response to the arrival of each new commit and the completion of each previously commit scheduled for test- ing. We conduct an empirical study on three datasets, and use the APFDC metric to evaluate this technique. The result shows that, after prioritization, our technique can effectively detect failing commits earlier.

To date, current RTS techniques under CI environment is based on two windows in terms of time. But this technique fails to consider the arrival rate of test suites and only takes the results of test suites execution history into account. In this thesis, we present a Count-Based RTS technique, which is based on the test suite failures and execution history by utilizing two window sizes in terms of number of test suites, and a Transition-Based RTS technique, which adds the test suites’ “pass to fault” transitions for selection prediction in addition to the two window sizes. We again conduct an empirical study on three datasets, and use the percentage of faults and percentage of “pass to fault” transition metrics to evaluate these two techniques. The results show that, after selection, Transition-Based technique detects more faults and more “pass to fault” transitions than the existing techniques.




“A Performance Investigation of Neural Network Based Channel Estimation in 2x2 and 4x4 MIMO Communications”

Kanghyeok Yang

Advisor: Mehmet Can Vuran
Committee members: Stephen Scott, Qiben Yan
Date: April 13, 2018
Time: 3 p.m.
Room: Schorr 211

Abstract: With increasing needs of fast and reliable communication between devices, wireless communication techniques are rapidly evolving to meet such needs. Multiple input and output (MIMO) systems are one of the key techniques that utilize multiple antennas for high-throughput communication. However, increasing the number of antennas in wireless communication also adds to the complexity of channel estimation which is essential to accurately decode the transmitted data. Thereby, a development of accurate and efficient channel estimation process is necessary for wireless communication. In this manner, this project proposes machine learning based channel estimation approaches that are envisioned to enhance the channel estimation performance in the high-noise environment during a MIMO communication. A 2x2 and 4x4 MIMO communication with space-time block coding model (STBC) was simulated in MATLAB and different neural network based regression algorithms were tested through simulations. The analysis results demonstrate that a generalized regression neural network (GRNN) model achieves higher estimation performance compared to the existing techniques especially when the neural network model is trained in the low signal to noise ratio (SNR) regime. The results of this study present an opportunity for achieving better performance in channel estimation through machine and deep learning algorithms. Also, the proposed method contributes to developing scalable channel estimation models for MIMO communications.
Keywords: Channel Estimation, MIMO, Machine Learning, Generalized Regression Neural Network, Deep Neural Network




“Assessing the Quality and Stability of Recommender Systems”

David Shriver

Advisor: Sebastian Elbaum
Committee members: Matt Dwyer, Gregg Rothermel
Date: April 16, 2018
Time: 2:30 p.m.
Room: 103C Avery

Abstract: Recommender systems help users to find products they may like when lacking personal experience or facing an overwhelmingly large set of items. However, assessing the quality and stability of recommender systems can present challenges for developers. First, traditional accuracy metrics, such as precision and recall, for validating the quality of recommendations, offer only a coarse, one-dimensional view of the system performance. Second, assessing the stability of a recommender systems requires generating new data and retraining a system, which is expensive.

In this work, we present two new approaches for assessing the quality and stability of recommender systems to address these challenges. We first present a general and extensible approach for assessing the quality of the behavior of a recommender system using logical property templates. The approach is general in that it defines recommendation systems in terms of sets of rankings, ratings, users, and items on which property templates are defined. It is extensible in that these property templates define a space of properties that can be instantiated and parameterized to characterize a recommendation system. We study the application of the approach to several recommendation systems. Our findings demonstrate the potential of these properties, illustrating the insights they can provide about the different algorithms and evolving datasets.

We also present an approach for influence-guided fuzz testing of recommender system stability. We infer influence models for aspects of a dataset, such as users or items, from the recommendations produced by a recommender system and its training data. We define dataset fuzzing heuristics that use these influence models for generating modifications to an original dataset and we present a test oracle based on a threshold of acceptable instability. We implement our approach and evaluate it on several recommender algorithms using the MovieLens dataset and we find that influence-guided fuzzing can effectively find small sets of modifications that cause significantly more instability than random approaches.




“Collective Factorization for Multi-relational Yelp Dataset”

Yang Liu

Advisor: Lisong Xu
Committee members: Peter Revesz, Hongfeng Yu
Date: April 18, 2018
Time: 2:30 p.m.
Room: 211 Schorr

Abstract: Unknown values prediction in relational database systems is an active topic in data mining area. The collective factorization is proved to be a successful approach to discover hidden factors from different entities which affect the predicted unknown values. We focus on the Yelp open source which contains 6 relational entities and different types of data. In our work, we extended the traditional collective matrix factorization model which only support binary relational data to incorporate multi-relational ones in the Yelp dataset. Moreover, the modified Latent Dirichlet Allocation (LDA) approach is utilized to extract representational topics for text preprocessing. Furthermore, the filtered topics are included in the collective factorization model to enable text information in the analysis. The evaluation shows that the model greatly improve the accuracy in unknown values prediction.




“Collective Factorization for Multi-relational Yelp Dataset”

Yang Liu

Advisor: Lisong Xu
Committee members: Peter Revesz, Hongfeng Yu
Date: April 18, 2018
Time: 2:30 p.m.
Room: 211 Schorr

Abstract: Unknown values prediction in relational database systems is an active topic in data mining area. The collective factorization is proved to be a successful approach to discover hidden factors from different entities which affect the predicted unknown values. We focus on the Yelp open source which contains 6 relational entities and different types of data. In our work, we extended the traditional collective matrix factorization model which only support binary relational data to incorporate multi-relational ones in the Yelp dataset. Moreover, the modified Latent Dirichlet Allocation (LDA) approach is utilized to extract representational topics for text preprocessing. Furthermore, the filtered topics are included in the collective factorization model to enable text information in the analysis. The evaluation shows that the model greatly improve the accuracy in unknown values prediction.