Xiangdong An

Revising Markov boundary for multiagent probabilistic inference

Multiply sectioned Bayesian networks (MSBNs) extend Bayesian networks (BNs) to graphical models that provide a coherent framework for probabilistic inference in cooperative multiagent distributed interpretation systems. Observation plays an important role in the inference with graphical models. Since observation of each observable variable has a cost, it would be helpful if we can find the most relevant variables to observe. In a probabilistic model, a Markov boundary of a variable provides a minimal set of variables that shields the variable from the influence of all other variables. However, the concept cannot be used directly for observation. First, it is generally intractable to verify conditional independencies in a probabilistic model. Second, the Markov boundary members may not be observable. Third, it is defined only for a single variable. Finally, it is not unique. By revising the concept to address these issues, we introduce the concept of observable Markov boundary of a set of nodes defined on d-separation of graphical models. The observable Markov boundary captures all relevant variables to observe for probabilistic inference with graphical models. In an MSBN, the observable Markov boundary of a set of nodes may span across all Bayesian subnets. We present an algorithm for cooperative computation of the observable Markov boundary of a set of nodes in an MSBN without revealing subnet structures.

Finding best evidence for evidence-based best practice recommendations in health care: the initial decision support system design

A major problem for Canadian health organizations is finding best evidence for evidence-based best practice recommendations. Medications are not always effectively used and misuse may harm patients. Drugs are the fastest-growing element of Canadian health care spending, second only to hospital spending. Three hundred million prescriptions are filled annually. Prescription drugs accounted for 5.8% of total health care spending in 1980 and close to 18% today. A primary long-term goal of this research is to develop a decision support system for evidence-based management, quality control and best practice recommendations for medical prescriptions. Our results will improve accessibility and management of information by: (1) building an prototype for adaptive information extraction, text and data mining from (online) documents to find evidence on which to base best practices; and (2) employing multiply sectioned Bayesian networks (MSBNs) to infer a probabilistic interpretation to validate evidence for recommendations; MSBNs provide this structure. Best practices to improve drug-related health outcomes; patients’ quality of life; and cost-effective use of medications by changing knowledge and behavior. This research will support next generation eHealth decision support systems, which routinely find and verify evidence from multiple sources, leading to cost-effective use of drugs, improve patients’ quality of life and optimize drug-related health outcomes.

Simulation of Graphical Models for Multiagent Probabilistic Inference

Multiply-sectioned Bayesian networks (MSBNs) extend Bayesian networks to graphical models for multiagent probabilistic reasoning. The empirical study of algorithms for manipulations of MSBNs (e.g., verification, compilation, and inference) requires experimental MSBNs. As engineering MSBNs in large problem domains requires significant knowledge and engineering effort, the authors explore automatic simulation of MSBNs. Due to the large domain over which an MSBN is defined and a set of constraints to be satisfied, a generate-and-test approach toward simulation has a high rate of failure. The authors present an alternative approach that treats the simulation process as a sequence of decisions. They constrain the space of each decision so that backtracking is minimized and the outcome is always a legal MSBN. A suite of algorithms that implements this approach is presented, and experimental results are shown.

Uncertain inference control in privacy protection

Context management is the key enabler for emerging context-aware applications, and it includes context acquisition, understanding and exchanging. Context exchanging should be made privacy-conscious. We can specify privacy preferences to limit the disclosure of sensitive contexts, but the sensitive contexts could still be derived from those insensitive. To date, there have been very few inference control mechanisms for context management, especially when the environments are uncertain. In this paper, we present an inference control method for private context protection in uncertain environments.

A Bayesian Network Approach to Detecting Privacy Intrusion

Personal information privacy could be compromised during information collection, transmission, and handling. In information handling, privacy could be violated by both the inside and the outside intruders. Though, within an organization, private data are generally protected by the organizations privacy policies and the corresponding platforms for privacy practices, private data could still be misused intentionally or unintentionally by individuals who have legitimate access to them in the organization. In this paper, we propose a Bayesian network-based method for insider privacy intrusion detection in database systems

Lexical acquisition and clustering of word senses to conceptual lexicon construction

We describe a mechanism and an algorithm to support construction of a large complex conceptual lexicon from an existing alphabetical lexicon. As part of this research, we define lexical models to present words and lexicons. Given the fact that an alphabetical lexicon contains lexical information about words which are organized by their spelling, constructing a conceptual lexicon requires an identification of lexical concepts and their relationships. Lexical acquisition and word-sense clustering are introduced to identify the lexical concepts and to discover the conceptual relationships. The result of this research is a set of candidate concepts which can be treated as initial concepts for the conceptual lexicon construction.

Reasoning about obfuscated private information: who have lied and how to lie

In ubiquitous environments, context sharing among agents should be made privacy-conscious. Privacy preferences are generally specified to govern the context exchanging among agents. Besides who has rights to see what information, a users privacy preference could also designate who has rights to have what obfuscated information. By obfuscation, people could present their private information in a coarser granularity, or simply in a falsified manner, depending on the specific situations. Nevertheless, people cannot randomly obfuscate their private information because by reasoning the recipients could detect the obfuscation. In this paper, we present a Bayesian network-based method to reason about the obfuscation. On the one hand, it can be used to find if the received information has been obfuscated, and if so, what the true information could be; on the other hand, it can be used to help the obfuscators reasonably obfuscate their private information.

Iterative Multiagent Probabilistic Inference

Multiply sectioned Bayesian networks (MSBNs) support multiagent probabilistic inference in distributed large problem domains, where agents are organized in a tree structure (called hypertree). In earlier work, agents need to follow an order of the depth-first traversal of the hypertree to update their belief. Hence, agents need some synchronization with each other and belief updating can only be done in a limited parallel. Especially, belief updating will fail if any communication channels have problems. In this paper, we present an iterative method where multiple agents asynchronously perform belief updating in a complete parallel. Compared to the previous work, the iterative method is simple, self- adaptive and robust.

Dynamic inference control in privacy preference enforcement

In pervasive (ubiquitous) environments, context-aware agents are used to obtain, understand, and share local contexts with each other so that the environments could be integrated seamlessly. Context sharing among agents should be made privacy-conscious. Privacy preferences are generally specified to regulate the exchange of the contexts, where who have rights under what conditions to have what contexts are designated. However, released contexts could be used to infer those unreleased. In particular, different contexts released could endanger the security of different contexts unreleased. The existing privacy preference specification platforms do not have a mechanism to prevent inference. To date, there have been very few inference control mechanisms specifically tailored to context management in pervasive (ubiquitous) environments. A Bayesian network based mechanism has been proposed to prevent privacy-sensitive contexts from being inferred from those to be released. Nevertheless, contexts in pervasive (ubiquitous) environments could change from time to time and are history dependent. In this paper, we propose to use dynamic Bayesian networks to track the most updated beliefs of the adversaries about the dynamic domains in order to evaluate which contexts in the domains could be released safely in various situations.

A statistical model for predicting power demand peaks in power systems

The total commodity cost for electricity also includes the cost of building new electricity infrastructure and the expenses of providing conservation and demand response programs. This is called the Global Adjustment (GA) by the Independent Electricity System Operator (IESO), a corporate entity in Ontario, Canada working at the heart of Ontarios power system to ensure there is enough power to meet the provinces energy needs in real-time and to plan and secure energy for the future. In Ontario, approximately 300 Class A customers representing Ontarios largest electricity consumers pay GA based on how much they contribute to the 5 highest peak hour demands in a fiscal year. Accurately predicting the top 5 peak hours in a fiscal year may help such a customer minimize its energy consumption in such periods and save it tens of millions of dollars in adjustment cost. In the meantime, this will reduce the size of demand peaks and the need of new electricity infrastructure for exceptionally high peak demands. This paper proposes to learn a statistical model for predicting in real-time the top 5 peak demand hours in a fiscal year, where feature selection is discussed. Preliminary experimental studies indicate the proposed model can effectively help locate the potential peak hours. This work is conducted for an application project, so we also discuss the implementation and deployment details of this model, where a client-server architecture with Ajax is adopted to ensure the updated peak hour information is delivered to all customers in real-time.