Date of Award
Doctor of Philosophy (PhD)
Electrical Engineering and Computer Science
Pramod K. Varshney
Classification systems are ubiquitous, and the design of effective classification algorithms has been an even more active area of research since the emergence of machine learning techniques. Despite the significant efforts devoted to training and feature selection in classification systems, misclassifications do occur and their effects can be critical in various applications. The central goal of this thesis is to analyze classification problems in human-driven and data-driven systems, with potentially unreliable components and design effective strategies to ensure reliable and effective classification algorithms in such systems. The components/agents in the system can be machines and/or humans. The system components can be unreliable due to a variety of reasons such as faulty machines, security attacks causing machines to send falsified information, unskilled human workers sending imperfect information, or human workers providing random responses. This thesis first quantifies the effect of such unreliable agents on the classification performance of the systems and then designs schemes that mitigate misclassifications and their effects by adapting the behavior of the classifier on samples from machines and/or humans and ensure an effective and reliable overall classification.
In the first part of this thesis, we study the case when only humans are present in the systems, and consider crowdsourcing systems. Human workers in crowdsourcing systems observe the data and respond individually by providing label related information to a fusion center in a distributed manner. In such systems, we consider the presence of unskilled human workers where they have a reject option so that they may choose not to provide information regarding the label of the data. To maximize the classification performance at the fusion center, an optimal aggregation rule is proposed to fuse the human workers' responses in a weighted majority voting manner.
Next, the presence of unreliable human workers, referred to as spammers, is considered. Spammers are human workers that provide random guesses regarding the data label information to the fusion center in crowdsourcing systems. The effect of spammers on the overall classification performance is characterized when the spammers can strategically respond to maximize their reward in reward-based crowdsourcing systems. For such systems, an optimal aggregation rule is proposed by adapting the classifier based on the responses from the workers.
The next line of human-driven classification is considered in the context of social networks. The classification problem is studied to classify a human whether he/she is influential or not in propagating information in social networks. Since the knowledge of social network structures is not always available, the influential agent classification problem without knowing the social network structure is studied. A multi-task low rank linear influence model is proposed to exploit the relationships between different information topics. The proposed approach can simultaneously predict the volume of information diffusion for each topic and automatically classify the influential nodes for each topic.
In the third part of the thesis, a data-driven decentralized classification framework is developed where machines interact with each other to perform complex classification tasks. However, the machines in the system can be unreliable due to a variety of reasons such as noise, faults and attacks. Providing erroneous updates leads the classification process in a wrong direction, and degrades the performance of decentralized classification algorithms. First, the effect of erroneous updates on the convergence of the classification algorithm is analyzed, and it is shown that the algorithm linearly converges to a neighborhood of the optimal classification solution. Next, guidelines are provided for network design to achieve faster convergence. Finally, to mitigate the impact of unreliable machines, a robust variant of ADMM is proposed, and its resilience to unreliable machines is shown with an exact convergence to the optimal classification result.
The final part of research in this thesis considers machine-only data-driven classification problems. First, the fundamentals of classification are studied in an information theoretic framework. We investigate the nonparametric classification problem for arbitrary unknown composite distributions in the asymptotic regime where both the sample size and the number of classes grow exponentially large. The notion of discrimination capacity is introduced, which captures the largest exponential growth rate of the number of classes relative to the samples size so that there exists a test with asymptotically vanishing probability of error. Error exponent analysis using the maximum mean discrepancy is provided and the discrimination rate, i.e., lower bound on the discrimination capacity is characterized. Furthermore, an upper bound on the discrimination capacity based on Fano's inequality is developed.
Li, Qunwei, "On Classification in Human-driven and Data-driven Systems" (2018). Dissertations - ALL. 991.