Conn-Eval : A Connectionist Method for Evaluating Multiple Attribute Items

The task of evaluating and ranking multiple attribute items is relevant in many dierent aspects of e-commerce including RFQ, negotiations, personalized catalogues, proling and customer modeling. Usually a parametric utility function is considered which is a linear weighted sum of the individualattribute utilities, and the weights of the individual attributes are estimated by minimizing the discrepancy between the predicted order and the true order. However, the individual attribute utilities may not be linearly independent and they may not be known a priori. In this paper, we propose a nonlinear model for ranking the multiple attribute items and their subsequent evaluation without assuming any independence between the attributes and any prior knowledge about individual attribute utility functions. A neural network (connectionist model) has been used at the core of the algorithm to learn and rank the items. Since the desired utility value for a bid is unknown, the usual techniques of function approximation cannot be employed in this paradigm, and a new objective function (error measure) is dened in this context. New rules are proposed for automatic selection of learning rate and prescriptions are made for selection of the architecture of the neural network for this task. New query-based sampling technique is also provided to improve the performance of the method. Experimental results illustrate the effectiveness of the method for ranking items with complex utility functions.