This dissertation develops methods to minimize recommendation error costs when inputs to a rule-based expert system are prone to errors. The problem often arises in web-based applications where data are inherently noisy or provided by users who perceive some benefit from falsifying inputs. Prior studies proposed methods that attempted to minimize the probability of recommendation error, but did not take into account the relative costs of different types of errors. In situations where these differences are significant, an approach that minimizes the expected misclassification error costs has advantages over extant methods that ignore these costs. Building on the existing literature, two new techniques - Cost-Based Input Modification (CBIM) and Cost-Based Knowledge-Base Modification (CBKM) were developed and evaluated. Each method takes as inputs (1) the joint probability distribution of a set of rules, (2) the distortion matrix for input noise as characterized by the probability distribution of the observed input vectors conditioned on their true values, and (3) the misclassification cost for each type of recommendation error. Under CBIM, for any observed input vector v, the recommendation is based on a modified input vector v' such that the expected error costs are minimized. Under CBKM the rule base itself is modified to minimize the expected cost of error. The proposed methods were investigated as follows: as a control, in the special case where the costs associated with different types of errors are identical, the recommendations under these methods were compared for consistency with those obtained under extant methods. Next, the relative advantages of CBIM and CBKM were compared as (1) the noise level changed, and (2) the structure of the cost matrix varied. As expected, CBKM and CBIM outperformed the extant Knowledge Base Modification (KM) and Input Modification (IM) methods over a wide range of input distortion and cost matrices, with some restrictions. Under the control, with constant misclassification costs, the new methods performed equally with the extant methods. As misclassification costs increased, CBKM outperformed KM and CBIM outperformed IM. Using different cost matrices to increase misclassification cost asymmetry and order, CBKM and CBIM performance increased. At very low distortion levels, CBKM and CBIM underperformed as error probability became more significant in each method's estimation. Additionally, CBKM outperformed CBIM over a wide range of input distortion as its technique of modifying an original knowledge base outperformed the technique of modifying inputs to an unmodified decision tree.

No description available.

The model training algorithm is a critical component in the statistical pattern recognition approaches which are based on the Bayes decision theory. Conventional applications of the Bayes decision theory usually assume uniform error cost and result in a ubiquitous use of the maximum a posteriori (MAP) decision policy and the paradigm of distribution estimation as practice in the design of a statistical pattern recognition system. The minimum classification error (MCE) training method is proposed to overcome some substantial limitations for the conventional distribution estimation methods. In this thesis, three aspects of the MCE method are generalized. First, an optimal classifier/recognizer design framework is constructed, aiming at minimizing non-uniform error cost.A generalized training criterion named weighted MCE is proposed for pattern and speech recognition tasks with non-uniform error cost. Second, the MCE method for speech recognition tasks requires appropriate management of multiple recognition hypotheses for each data segment. A modified version of the MCE method with a new approach to selecting and organizing recognition hypotheses is proposed for continuous phoneme recognition. Third, the minimum verification error (MVE) method for detection-based automatic speech recognition (ASR) is studied. The MVE method can be viewed as a special version of the MCE method which aims at minimizing detection/verification errors. We present many experiments on pattern recognition and speech recognition tasks to justify the effectiveness of our generalizations.