Abstract The ability to efficiently resolve conflicts in interactive constraint-based applications is critical for user experience and system reliability. Conflict resolution can be regarded as a specific type of explanation, often denoted as diagnosis. Existing work on integrating machine learning with diagnostic reasoning emphasizes on the combination of hitting set approaches with probabilistic reasoning and memory-based machine learning. An alternative to such two-phase diagnosis approaches is direct diagnosis, which focuses on determining diagnoses without predetermining conflicts. In this article, we utilize diagnosis knowledge from the past to improve diagnosis efficiency while also maintaining user-defined preference criteria. Our approach integrates model-based collaborative filtering (feed-forward neural networks) and other machine learning approaches (e.g., logistic regression and random forest) with direct model-based diagnosis ( FastDiag ). The re-ordering of constraints as input to the diagnosis algorithm increases the efficiency of diagnostic reasoning for determining preference-preserving diagnoses. Through experiments on real-world configuration knowledge bases ( B2C , BusyBox , EA and Linux kernel ), we demonstrate significant runtime improvements and high accuracy in diagnosis prediction. With this, we also contribute to the growing body of literature on combining machine learning and constraint-based reasoning.

Conflict detection is relevant in various application scenarios, ranging from interactive decision-making to the diagnosis of faulty knowledge bases. Conflicts can be regarded as sets of constraints that cause an inconsistency. In many scenarios (e.g., constraint-based configuration), conflicts are repeatedly determined for the same or similar sets of constraints. This misses out on the valuable opportunity for leveraging knowledge reuse and related potential performance improvements, which are extremely important, specifically interactive constraint-based applications. In this paper, we show how to integrate knowledge reuse concepts into non-instructive conflict detection. We introduce the InformedQX algorithm, which is a reuse-aware variant of QuickXPlain. The results of a related performance analysis with the Linux-2.6.3.33 configuration knowledge base show significant improvements in terms of runtime performance compared to QuickXPlain.

Abstract Constraint solving is applied in different application contexts. Examples thereof are the configuration of complex products and services, the determination of production schedules, and the determination of recommendations in online sales scenarios. Constraint solvers apply, for example, search heuristics to assure adequate runtime performance and prediction quality. Several approaches have already been developed showing that machine learning (ML) can be used to optimize search processes in constraint solving. In this article, we provide an overview of the state of the art in applying ML approaches to constraint solving problems including constraint satisfaction, SAT solving, answer set programming (ASP) and applications thereof such as configuration, constraint-based recommendation, and model-based diagnosis. We compare and discuss the advantages and disadvantages of these approaches and point out relevant directions for future work.

Abstract User preferences are a crucial input needed by recommender systems to determine relevant items. In single-shot recommendation scenarios such as content-based filtering and collaborative filtering, user preferences are represented, for example, askeywords,categories, anditem ratings. In conversational recommendation approaches such as constraint-based and critiquing-based recommendation, user preferences are often represented on the semantic level in terms ofitem attribute valuesandcritiques. In this article, we provide an overview of preference representations used in different types of recommender systems. In this context, we take into account the fact thatpreferences aren’t stablebut are ratherconstructedwithin the scope of a recommendation process. In which way preferences are determined and adapted is influenced by various factors such aspersonality traits,emotional states, andcognitive biases. We summarize preference construction related research and also discuss aspects of counteracting cognitive biases.