NEURAL NETWORKS - FROM PREDICTION TO EXPLANATION

摘要

The purpose of this study has been to evaluate if neural nets used in the marketing domain can act not only for prediction but also as explanation. The focus has been to investigate if the powerful but opaque neural nets can be transformed into a representation comprehensible enough to act upon, while keeping a high accuracy on unseen examples. The paper contains a case study where neural nets are first trained to identify weeks with high impact of advertising. The trained networks are then used for "rule extraction" i.e. the knowledge learned by a neural net is transformed into a more comprehensive representation (in this case, decision trees). The study shows that the decision trees generated from the trained network have higher accuracy than decision trees created directly from the data. The study also indicates a need for a process to determine important inputs before using a neural net and shows that reduced input sets may produce more accurate neural nets and more compact decision trees.