Michiel van Wezel

Improved customer choice predictions using ensemble methods

In this paper various ensemble learning methods from machine learning and statistics are considered and applied to the customer choice modeling problem. The application of ensemble learning usually improves the prediction quality of flexible models like decision trees and thus leads to improved predictions. We give experimental results for two real-life marketing datasets using decision trees, ensemble versions of decision trees and the logistic regression model, which is a standard approach for this problem. The ensemble models are found to improve upon individual decision trees and outperform logistic regression. Next, an additive decomposition of the prediction error of a model, the bias/variance decomposition, is considered. A model with a high bias lacks the flexibility to fit the data well. A high variance indicates that a model is instable with respect to different datasets. Decision trees have a high variance component and a low bias component in the prediction error, whereas logistic regression has a high bias component and a low variance component. It is shown that ensemble methods aim at minimizing the variance component in the prediction error while leaving the bias component unaltered. Bias/variance decompositions for all models for both customer choice datasets are given to illustrate these concepts.

Application of reinforcement learning to the game of Othello

Operations research and management science are often confronted with sequential decision making problems with large state spaces. Standard methods that are used for solving such complex problems are associated with some difficulties. As we discuss in this article, these methods are plagued by the so-called curse of dimensionality and the curse of modelling. In this article, we discuss reinforcement learning, a machine learning technique for solving sequential decision making problems with large state spaces. We describe how reinforcement learning can be combined with a function approximation method to avoid both the curse of dimensionality and the curse of modelling. To illustrate the usefulness of this approach, we apply it to a problem with a huge state space-learning to play the game of Othello. We describe experiments in which reinforcement learning agents learn to play the game of Othello without the use of any knowledge provided by human experts. It turns out that the reinforcement learning agents learn to play the game of Othello better than players that use basic strategies.

A graphical shopping interface based on product attributes

Most recommender systems present recommended products in lists to the user. By doing so, much information is lost about the mutual similarity between recommended products. We propose to represent the mutual similarities of the recommended products in a two dimensional map, where similar products are located close to each other and dissimilar products far apart. As a dissimilarity measure we use an adaptation of Gowers similarity coefficient based on the attributes of a product. Two recommender systems are developed that use this approach. The first, the graphical recommender system, uses a description given by the user in terms of product attributes of an ideal product. The second system, the graphical shopping interface, allows the user to navigate towards the product she wants. We show a prototype application of both systems to MP3-players.

Two algorithms for generating structured and unstructured monotone ordinal data sets

Monotone constraints are very common while dealing with multi-attribute ordinal problems. Grinding wheels hardness selection, timely replacements of costly laser sensors in silicon wafer manufacturing, and the selection of the right personnel for sensitive production facilities, are just a few examples of ordinal problems where monotonicity makes sense. In order to evaluate the performance of various ordinal classifiers one needs both artificially generated as well as real world data sets. Two algorithms are presented for generating monotone ordinal data sets. The first can be used for generating random monotone ordinal data sets without an underlying structure. The second algorithm, which is the main contribution of this paper, describes for the first time how structured monotone data sets can be generated.

Feature Selection for Neural Networks through Functional Links Found by Evolutionary Computation

In this paper we describe different ways to select and transform features using evolutionary computation. The features are intended to serve as inputs to a feedforward network. The first way is the selection of features using a standard genetic algorithm, and the solution found specifies whether a certain feature should be present or not. We show that for the prediction of unemployment rates in various European countries, this is a succesfull approach. In fact, this kind of selection of features is a special case of so-called functional links. Functional links transform the input pattern space to a new pattern space. As functional links one can use polynomials, or more general functions. Both can be found using evolutionary computation. Polynomial functional links are found by evolving a coding of the powers of the polynomial. For symbolic functions we can use genetic programming. Genetic programming finds the symbolic functions that are to be applied to the inputs. We compare the workings of the latter two methods on two artificial datasets, and on a real-world medical image dataset.

Genetic Improvement of Railway Timetables

We consider how to improve railway timetables. As case we take the Rompnet of the dutch railways, a highly constrained problem, containing both hard and soft constraints. We show how to cast the constraints into the format of the Genocop system. Every train should run every hour at the same time and hence the constraints should be interpreted modulo sixty. This gives a non-convex search space of integer vectors. The Genocop system is designed for convex search spaces, but we show how to adapt the Genocop operators to deal with this nonconvex search space. The results of two experiments using the adapted operators are very encouraging.

Competitive Neural Networks for Customer Choice Models

In this paper we propose and examine two different models for customer choices in for instance a wholesale department, given the actual sales. Both customers and products are modeled by points in a k-dimensional real vector space. Two possible strategies are discussed: in one model the customer buys the nearest option from categories of products, in the other he/she buys all products within a certain radius of his/her position. Now we deal with the following problem: given only the sales list, how can we retrieve the relative positions corresponding to customers and products? In particular we are interested in the dimension k of the space: we are looking for low dimensional solutions with a good “fit” to the real sales list. Theoretical complexity of these problems is addressed: they are very hard to solve exactly; special cases are shown to be NP-complete. We use competitive neural network techniques for both artificial and real life data, and report the results.

Choosing attribute weights for item dissimilarity using clikstream data with an application to a product catalog map

In content- and knowledge-based recommender systems often a measure of (dis)similarity between items is used. Frequently, this measure is based on the attributes of the items. However, which attributes are important for the users of the system remains an important question to answer. In this paper, we present an approach to determine attribute weights in a dissimilarity measure using clickstream data of an e-commerce website. Counted is how many times products are sold and based on this a Poisson regression model is estimated. Estimates of this model are then used to determine the attribute weights in the dissimilarity measure. We show an application of this approach on a product catalog of MP3 players provided by Compare Group, owner of the Dutch price comparison site http://www.vergelijk.nl, and show how the dissimilarity measure can be used to improve 2D product catalog visualizations.

Online shopping using a two dimensional product map

In this paper, we propose a user interface for online shopping that uses a two dimensional product map to present products. This map is created using multidimensional scaling (MDS). Dissimilarities between products are computed using an adapted version of Gowers coefficient of similarity based on the attributes of the product. The user can zoom in and out by drawing rectangles. We show an application of this user interface to MP3 players and give an interpretation of the product map.

Nonmetric Multidimensional Scaling with Neural Networks

In this paper we present a neural network for nonmetric multidimensional scaling. In our approach, the monotone transformation that is a part of every nonmetric scaling algorithm is performed by a special feedforward neural network with a modified backpropagation algorithm. Contrary to traditional methods, we thus explicitly model the monotone transformation by a special purpose neural network. The architecture of the new network and the derivation of the learning rule are given, as well as some experimental results. The experimental results are positive.