Abstract

The task of mapping a domain-specific problem space to an adequate set of data mining (DM) methods is a crucial step in data science projects. While there have been several efforts for automated method selection in general, only few approaches consider the particularities of problem contexts expressed in domain-specific language. Therefore, we propose the concept of a text-based recommender system (TBRS) which takes problem descriptions articulated in domain language as inputs and then recommends the best suitable class of DM methods. Following a design science research methodology, the current focus is on the initial steps of motivating the problem and conducting a requirements analysis. In particular, we outline the problem setting using an exemplary scenario from industrial practice and derive requirements towards an adequate solution artifact. Subsequently, we discuss potential TBRS methods with regard to requirement fulfillment while organizing both methods and requirements in a structured framework. Finally, we conclude the paper, discuss limitations and draw an outlook.

Share

COinS
 

Towards a Text-based Recommender System for Data Mining Method Selection

The task of mapping a domain-specific problem space to an adequate set of data mining (DM) methods is a crucial step in data science projects. While there have been several efforts for automated method selection in general, only few approaches consider the particularities of problem contexts expressed in domain-specific language. Therefore, we propose the concept of a text-based recommender system (TBRS) which takes problem descriptions articulated in domain language as inputs and then recommends the best suitable class of DM methods. Following a design science research methodology, the current focus is on the initial steps of motivating the problem and conducting a requirements analysis. In particular, we outline the problem setting using an exemplary scenario from industrial practice and derive requirements towards an adequate solution artifact. Subsequently, we discuss potential TBRS methods with regard to requirement fulfillment while organizing both methods and requirements in a structured framework. Finally, we conclude the paper, discuss limitations and draw an outlook.