Armin Stahl

Similarity Measures for Object-Oriented Case Representations

Object-oriented case representations require approaches for similarity assessment that allow to compare two differently structured objects, in particular, objects belonging to different object classes. Currently, such similarity measures are developed more or less in an ad-hoc fashion. It is mostly unclear, how the structure of an object-oriented case model, e.g., the class hierarchy, influences similarity assessment. Intuitively, it is obvious that the class hierarchy contains knowledge about the similarity of the objects. However, how this knowledge relates to the knowledge that could be represented in similarity measures is not obvious at all. This paper analyzes several situations in which class hierarchies are used in different ways for case modeling and proposes a systematic way of specifying similarity measures for comparing arbitrary objects from the hierarchy. The proposed similarity measures have a clear semantics and are computationally inexpensive to compute at run-time.

Using evolution programs to learn local similarity measures

The definition of similarity measures is one of the most crucial aspects when developing case-based applications. In particular, when employing similarity measures that contain a lot of specific knowledge about the addressed application domain, modelling similarity measures is a complex and time-consuming task. One common element of the similarity representation are local similarity measures used to compute similarities between the values of single attributes. In this paper an approach to learn local similarity measures by employing an evolution program-- a special form of a genetic algorithm--is presented. The goal of the approach is to learn similarity measures that sufficiently approximate the utility of cases for given problem situations in order to obtain reasonable retrieval results.

Learning Feature Weights from Case Order Feedback

Defining adequate similarity measures is one of the most difficult tasks when developing CBR applications. Unfortunately, only a limited number of techniques for supporting this task by using machine learning techniques have been developed up to now. In this paper, a new framework for learning similarity measures is presented. The main advantage of this approach is its generality, because its application is not restricted to classification tasks in contrast to other already known algorithms. A first refinement of the introduced framework for learning feature weights is described and finally some preliminary experimental results are presented.

Applying Recursive CBR for the Custumization of Structured Products in an Electronic Shop

When applying CBR for Electronic Commerce, the adaptation capabilities of CBR can be used for product customization. Most adaptation techniques suffer from the problem that they require a large knowledge acquisition effort which leads to problems in the rapidly changing E-Commerce scenario. In this paper we present a new approach to adaptation that is particularly suited to Electronic Commerce applications. It assumes that products can be structured hierarchically into sub-components. Adaptation is achieved by incrementally replacing unsuitable sub-components through recursively applying CBR to find best-matching alternative sub-components. The presented approach avoids huge portions of the knowledge acquisition effort and is prototypically implemented as an extension of the CBR-Works tool.

Defining Similarity Measures: Top-Down vs. Bottom-Up

Defining similarity measures is a crucial task when developing CBR applications. Particularly, when employing utility-based similarity measures rather than pure distance-based measures one is confronted with a difficult knowledge engineering task. In this paper we point out some problems of the state-of-the-art procedure to defining similarity measures. To overcome these problems we propose an alternative strategy to acquire the necessary domain knowledge based on a Machine Learning approach. To show the feasibility of this strategy several application scenarios are discussed and some results of an experimental evaluation for one of these scenarios are presented.

Exploiting Background Knowledge when Learning Similarity Measures

The definition of similarity measures – one core component of every CBR application – leads to a serious knowledge acquisition problem if domain and application specific requirements have to be considered. To reduce the knowledge acquisition effort, different machine learning techniques have been developed in the past. In this paper, enhancements of our framework for learning knowledge-intensive similarity measures are presented. The described techniques aim to restrict the search space to be considered by the learning algorithm by exploiting available background knowledge. This helps to avoid typical problems of machine learning, such as overfitting the training data.

Intelligent Customer Support for Product Selection with Case-Based Reasoning

Current product-oriented database search facilities are widely used on the Internet but recognized as limited in capability for intelligent sales support. The vision of intelligent knowledgeable virtual sales agents is to incorporate more knowledge about products, customers, and the sales process into an electronic shop. This chapter describes a knowledge-based technology called case-based reasoning (CBR) and shows how it can be adapted and applied for developing intelligent virtual sales agents. To emphasize the advantages for our approach, we implemented several applications, some of which are in daily use.

Collecting Experience on the Systematic Development of CBR Applications Using the INRECA Methodology

This paper presents an overview of the INRECA methodology for building and maintaining CBR applications. This methodology supports the collection and reuse of experience on the systematic development of CBR applications. It is based on the experience factory and the software process modeling approach from software engineering. CBR development experience is documented using software process models and stored in different levels of generality in a three-layered experience base. Up to now, experience from 9 industrial projects enacted by all INRECA II partners has been collected.

Approximation of utility functions by learning similarity measures

Expert systems are often considered to be logical systems producing outputs that can only be correct or incorrect. However, in many application domains results cannot simply be distinguished in this restrictive form. Instead to classify a result as correct or incorrect, here results might be more or less useful for solving a given problem or for satisfying given user demands, respectively. In such a situation, an expert system should be able to estimate the utility of possible outputs a-priori in order to produce reasonable results. In Case-Based Reasoning this is done by using similarity measures which can be seen as an approximation of the domain specific, but a-priori unknown utility function. In this article we present an approach how this approximation of utility functions can be facilitated by employing machine learning techniques.