Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TE64

We present a novel quadratic programming-based approach to classify bags. Our algorithm imposes no additional constraints on relating instance labels to bag labels and applicable to many learning problems such as image classification, molecule activity prediction and text mining. We demonstrate the computational efficacy and classification success of our approach on a wide range of real world datasets. 3 - Features Level Opinion Mining from Informal Text Corpus Using Machine Learning Techniques Prabin Kumar Panigrahi, Indian Institute of Management Indore, Rau-Pithampur Road, Indore, 453556, India, Nishikant Bele Due to Internet, torrent amount of informal text is generated. People express their views, emotion, feeling, and opinion on blogs, reviews, and social sites. This paper explores the use of machine learning technique at feature level sentiment categorization of Hindi blogs reviews at unigram, bigram, trigram and n gram level. We used six types of machine learning techniques to study whether unigram, bigram, trigram, and n-gram can be used for sentiment mining at the feature level. Our study shows that bi-gram with SVM outperformed the other methods. 4 - Robust Bayesian Level Set Estimation via Gaussian Processes This paper focuses on the problem of determining as large a region as possible where a function exceeds a given threshold with high probability. We assume that we only have access to a noise-corrupted version of the function and that function evaluations are costly. To select the next query point, we propose maximizing the expected area of the domain identified as above the threshold as predicted by a Gaussian process, robustified by a variance term. We also give asymptotic guarantees on the exploration effect of the algorithm, regardless of the prior misspecification. We show by various numerical examples that our approach also outperforms existing techniques in the literature in practice. Joint Session DM/Practice Curated: Intriguing Tweaks in Data Science II Sponsored: Data Mining Sponsored Session Chair: Derya Dinler, Middle East Technical University, ODTU Endystri Myhendisligi, Ankara, Turkey 06800 1 - Sparse Component Analysis of Locally Dominant Source Syed Mujahid, Asst. Professor, KFUPM, Dhahran, Saudi Arabia In this talk, the linear case of Blind Signal Separation (BSS) problem is presented. Conventionally, the BSS problem is solved via the Independent Component Analysis (ICA) methods, which requires the statistically independence assumption. A different class of methods that are based on the notion of sparsity, are called as Sparse Component Analysis (SCA) methods. A novel LP model that can solve a specific class of SCA problems will be presented in this talk. Numerical results will be presented to illustrate the usability of the proposed model. 2 - Laplacian Regularized Gaussian Processes for Modeling Expensive Black-Box Functions Rajitha Meka, University of Texas-San Antonio, One UTSA Circle, San Antonio, TX, 78256, United States, Adel Alaeddini In an increasing number of cases involving estimation of complex functions in the real world, one is often confronted with situations where there are several factors to be evaluated but experiments are prohibitively complex and/or expensive. Gaussian process (GP) is a well-known non-parametric regression method to fit nonlinear functions. We propose Laplacian regularized Gaussian processes with bilateral kernel to make use of both measured and unmeasured points for efficient modeling of expensive black-box functions. 3 - Probabilistic Distance Clustering on Networks Cem Iyigun, Associate Professor, Middle East Technical University, Inonu Blvd, Endustri Myhendisligi, Ankara, 06800, Turkey, Derya Ipek Eroglu In this study, a soft clustering problem on networks is investigated. It is assumed that cluster centers are located not only on vertices, but also on the edges of the network. Two different soft assignment schemes are studied where different membership functions are considered for the assignments. Structural properties for the clustering problem have been derived for different problem settings and the relation to the p-median problem has been shown under specific conditions. As a solution approach, a metaheuristic algorithm using the properties of the problem has been proposed. Computational experiments have been conducted with different problem instances. Junzi Zhang, Stanford University, Palo Alto, CA, 94304, United States, Andrea Zanette, Mykel John Kochenderfer n TE66 West Bldg 105A

n TE64 West Bldg 104A Dynamic Control, System Dynamics, and Neural Networks Sponsored: Data Mining Sponsored Session Chair: Davood Hajinezhad, Iowa State University, Ames, IA, 50010, United States 1 - Learning the Best Metaheuristic Parameters Set for Car-passenger Matching Problem in on Demand Mobility Services Arslan Ali Syed, Bayerische Motoren Werke, Munich, 81375, Germany, Univ. der Bundeswehr, Munich, Germany, Klaus Bogenberger Metaheuristics provide good suboptimal solution to an optimization problem in a reasonable time, but extensive effort is required to select a good set of parameters to obtain reasonable performance. In this work we propose that a neural network could be trained that takes specific features of the problem instance and outputs the best parameters for a specific metaheuristic algorithm. We extract various features for car-passenger matching problem to train a neural network that depending on problem instance returns the best metaheuristic parameters set. 2 - Distributed Reinforcement Control Learning Using Function Approximation Davood Hajinezhad, Duke University, 4225 Larchmont Road, 835, Durham, NC, 27707, United States, Michael Zavlanos In this research we consider the Reinforcement Learning (RL) problem in a distributed network, and propose a primal-dual optimization algorithm to obtain the parameters of a possibly nonlinear function such as neural networks which approximates the action-value function of the underlying RL problem. Utilizing this we are able to achieve the optimal policy for all agents in the network. The proposed algorithm is incremental and online with provable convergence to the set of stationary solutions. 3 - Evaluating Recurrent Neural Networks as a Tool for Process Theory Testing Nicholas Evangelopoulos, University of North Texas, Denton, TX, United States, Anna Sidorova Past events are often the best predictors of the future. Yet, the ability of management scholars to develop process theories is limited by the paucity of statistical tools for testing them. Here we examine the feasibility of Recurrent Neural Networks (RNN), a class of neural networks used for predicting future values based on sequences of historical data, as tools for process theory testing. Using data from SourceForge.net we evaluate selected propositions of the Event System Theory as they apply to the Free/Libre Open Source Software domain using RNN. Methodological recommendations and challenges associated with using RNN for process theory testing are discussed. Joint Session DM/Practice Curated: Data Science and Deep Learning IV Sponsored: Data Mining Sponsored Session Chair: Junzi Zhang, Stanford University, Stanford, CA, 94305, United States 1 - Predicting Information Diffusion Probabilities in Social Networks Based on Neural Networks Zhecheng Qiang, PhD Student, University of Central Florida, 4000 Central Florida Blvd., Orlando, FL, 32826, United States, Alexander Semenov, Qipeng Zheng Predicting information diffusion through social networks plays an important role in human activities analysis and influence maximization realization. Our paper focuses on predicting information diffusion probability in social networks based on artificial neural networks. In this paper, we analyze and study the fundamental factors that might affect the diffusion process. Then we implement deep neural network and generalized regression neural network to predict the diffusion probability. We evaluate our models on several real data sets and it shows our models are effective and outperform the state-of-the-art methods. 2 - Multiple Instance Learning via Quadratic Programming Emel Seyma Kucukasci, Istanbul Commerce University, Kucukyali E5 Kavsagi Inonu Cad No 4, Istanbul, 34840, Turkey Emel Seyma Kucukasci, Bogazici University, Istanbul, Turkey, Mustafa Gokce Baydogan, Z. Caner Taskin In multiple instance learning (MIL), objects are represented by a bag of instances and the class labels are known for the bags, but not for the individual instances. n TE65 West Bldg 104B

402