Recommendations in CPQ software : Designing and implementing a recommender system for CPQ software

Abstract

Configure, Price, Quote (CPQ) systems are specialized sales software designed to support and automate the quotation, pricing and configuration of customizable products and services. Recommender systems (RS) are software systems and techniques that provide suggestions of items or selections to users.

This thesis presents the design and implementation of a generalizable RS for Summium CPQ. The work examines how historical configuration data can be used to generate practical and explainable recommendations considering the stakeholder requirements and technical limitations of an existing production CPQ system.

Based on a stakeholder analysis and a literature review, a case-based recommendation (CBR) approach was selected for the initial implementation. Historical order configurations are parsed and represented as per-model binary vectors. A central contribution of this thesis is a decomposition method for cardinality-based feature model (CBFM) configurations, where configurations with submodels and variable multiplicity are represented as linked per-model vectors. This representation is then used to generate recommendations for complex configurations.

The recommender consists of an offline phase for vectorization and indexing, and an online phase for real-time recommendation generation. Similarity search is performed with k-nearest-neighbour (KNN) search using Jaccard similarity in PostgreSQL with PGVector and HNSW indexing.

Two recommendation scenarios are implemented: (1) offer-view recommendations of fully configured historical products and (2) configuration-view recommendations of feature selections based on a partial configuration.

The implementation provides a practical starting point for recommendations in Summium CPQ, but it has clear limitations. The approach is best suited to boolean or discrete categorical selections. Comprehensive user- and system-centric evaluation is left for future work. Future work may also include more advanced recommendation methods, supplier-specific adaptations, and tighter integration with constraint solving.