Constraint-based recommendation systems are used in many different domains like notebooks, cars, and mobile phones. Such systems describe product domains in sets of product and customer variables, their domains, and constraints which define the relationship between the variables. Maintenance is a crucial task in constraint-based recommendation systems, because it is timeconsuming and error-prone. We implemented a new application called 'iCone' (intelligent environment for the development and maintenance of configuration knowledge bases) to support knowledge engineers and their maintenance tasks. We present intelligent techniques like recommendation, anomaly management, dependency detection, and metrics to support knowledge engineers when maintaining constraint-based systems.
Many different product domains like, computers, cars, or mobile phones use constraint-based recommendation systems as a model for modern e-commerce services. The model can be represented as a constraint satisfaction problem (CSP) and consists of products, questions for the customer, and constraints which are describing the relationships between products and questions.
In this paper we give an overview of the iCone interface which is a web-tool for the development and maintenance of constraint-based systems.2. This system can be used to create and maintain constraint-based systems like knowledge-based recommendation or knowledge-based configuration. The advantage of this system is a strong intelligent support of knowledge engineers when maintaining a constraint-based system which will be described in this paper. 2 iCone iCone is a java-based web-application with a SQLite data base. SQL will be used to do consistency checks and to 1The authors are ordered in reverse alphabetical order.
2http://ase-projects-studies.ist.tugraz. at:8080/iCone/index.jsp save and load the knowledge bases. The application has four main packages called ’KnowledgeBase’ (with the main class ’KnowledgeBase’), ’Algorithm’ (calculating anomalies), ’Anomaly’ (e.g., explaining anomalies), and ’Solver’ (preparing consistency checks). For the visualization, the users need to have a modern internet browser with active JavaScript. All screens consists of login, navigation, recommendation and notification, and main area. In the main area, the user sees lists of knowledge bases, products, questions, constraints, results of anomaly checks or a preview. Figure 1 shows a preview of the constraint-based recommendation system for mobile phones. In the left area the user can answer some questions. Based on constraints which are defining the relationship between questions and products, the system recommends a set of products for the user.
The main object of iCone is the Knowledge base. It contains all products, product variables, questions, and constraints. Furthermore, iCone deals with the analysis package, which detects anomalies in the knowledge base, generates recommendations for knowledge engineers, approximates dependencies between variables in the knowledge base and generates metrics to evaluate the knowledge base. Next, we give an overview about the four main supporting techniques.
We implemented four different recommendation techniques.3 user-independent recommendation techniques like most viewed recommendation and recently added items. user-dependent recommendation techniques like collaborative filtering (find peer users) and content-based filtering (find similar items)
In our implementation we consider three different types of anomalies:4 conflict: A conflict is a set of constraints which can not be fulfilled. Conflicts can be resolved by sets of diagnoses. redundancy: A set of constraints can be denoted as redundant, if the removal of this set does not change the behavior of the knowledge base.
well-formedness violation: Well-formedness violations don’t change the behavior of the knowledge base but make it difficult to maintain a knowledge base.
In our implementation we have visualizations for constraint dependencies and another one for variable dependencies.5
ship between products, question, and constraints based on content-based recommendation.
lated (if possible) or approximated (in big knowledge bases). For the approximation we use Gibb’s sampling.
To get an overview of the quality of the knowledge base, the iCone interface offers an overview of several metrics.6 Goal question metrics (GQM): We indicated three goals, five questions and 16 metrics.
input variables (from the software engineering to the knowledge engineering domain) and present the results for each knowledge base in our application.
In this paper we presented recommendation techniques, anomaly management, dependency detection, and metric calculation, which are novel techniques for the maintenance of constraint-based systems.
3
Summary In this paper we gave an overview of our iCone system. With iCone you can create constraint-based configuration, knowledge-based recommendation, and MAUT-based knowledge bases. You can create, read, update, and delete products, product variables, questions, and constraints.
Further enhancements are possible in the context of supporting collaboration (e.g. when two or more knowledge engineers are working on the same knowledge base) and in the support of test case generation.