Ranking-Based Case Retrieval with Graph Neural Networks in Process-Oriented Case-Based Reasoning

Abstract

In Process-Oriented Case-Based Reasoning (POCBR), experiential knowledge from previous problem-solving situations is retrieved from a case base to be reused for upcoming problems. The task of retrieval is approached in previous work by using Graph Neural Networks (GNNs) to learn workflow similarities which are, in turn, used to find similar workflows w.r.t. a query workflow. This paper is motivated by the fact that these GNNs are mostly used for predicting the similarity between two workflows (query and case), while the retrieval in CBR is only concerned with the ranking of the most similar workflows from the case base w.r.t. the query. Thus, we propose a novel approach to extend the GNN-based workflow retrieval by a Learning-to-Rank (LTR) component where rankings instead of similarities between cases are predicted. The main contribution of this paper addresses the changes to the GNNs from previous work, such that their model architecture predicts pairwise preferences between cases w.r.t. a query and that they can be trained using labeled preference data. In order to transform these preferences into a case ranking, we also describe rank aggregation methods with different levels of computational complexity. The experimental evaluation compares different models for predicting similarities and rankings in case retrieval scenarios. The results indicate the potential of our ranking-based approach in significantly improving retrieval quality with only small impacts on the performance.