Graph Perturbation Effects on Graph Classification Using the GAM Model​

Gunner Lawless: Computer Science, Computer Engineering. University of Arkansas

ABSTRACT:

Deep learning models for graphs have shown promising achievements in recent years for the tasks of node and graph classification. Despite these recent advancements, very little research has been done to challenge the robustness of these models against adversarial attacks. In this work, we test the robustness of the GAM graph classification model against various adversarial perturbations. The GAM model is a deep learning model designed by [2] for graph classification using attention. We begin our experimentation by generating two different types of adversarial data sets: one with randomly generated perturbations and another with perturbations generated using the NETTACK algorithm. The NETTACK algorithm is a poisoning strategy introduced by [1] to lower accuracy of deep learning models designed for node classification. After generating these data sets, we analyze how these perturbed graphs affect the GAM model’s accuracy in a poison attack setting. We further propose ideas as to why neural networks designed for graph classification behave in the way that they do and call for further investigation.​

Dual Classifier Assisted Unsupervised ​Domain Adaptation​

Javier Campos: Computer Engineering. Purdue University Northwest
Alex Chen: Computer Science. Grove City College

ABSTRACT:

Domain adaptation has attracted great attentions to facilitate the sparsely labeled or unlabeled target learning by leveraging previously well-established source domain through knowledge transfer. Recent activities on domain adaptation attempt to build deep architectures to decrease cross-domain divergences by extracting more effective features. However, its generalizability would decrease significantly due to the domain mismatch that enlarges particularly at the top layers. In this work, we develop a novel Dual Classifiers assisted Domain Adaptation framework (DCDA) to solve the domain mismatch across source and target domains. Specifically, we explore the maximize mean discrepancy (MMD) by incorporating the pseudo labels of target samples to measure the domain difference better. Moreover, dual different types of classifiers are jointly trained to optimize the prediction on the target samples to maximally enhance the prediction ability of both classifiers. ​.

Key Phrase Extraction Using NMT and BERT

Matthew Tang: Statistics, Computer Science. University of Illinois Urbana Champaign
Jordyn Blakey: Computer Science. DePauw University

ABSTRACT:

Key phrase extraction is an important task in natural language processing. However, traditional methods may not fulfill this task efficiently. Therefore, deep learning techniques are applied to speed up the process. For example, neural machine translation (NMT) models are used to efficiently work with sequential data like text. However, there has been little work exploring the use of sequential models like NMT for key phrase extraction. One part of this paper examines the usefulness of neural machine translation, a sequence-to-sequence model, for key phrase extraction. First, we created an automated sequential labeling system to extract keywords from tweet sequences, and we examine possible methods to sequentially analyze the key phrases of each tweet to create an overall sentiment polarity for each tweet. Another language representation model we used was BERT: Bidirectional Encoder Representations from Transformer. When using BERT for sentiment analysis, we added a layer to capture the attention of words in determining the polarity of a sentence. By examining these attention values, we could find the important phrases in each sentence. In order to analyze these methods, we worked with a dataset of 600 tweets about the HPV vaccine. In this paper, we hope to explore both methods as a means to extract important phrases from the text.​