Dynamic Basket Recommendations in a Changing Market through an AI and similarity-based approach

In an era of rapid transformation within the automotive industry, the demand for personalized and adaptive vehicle recommendations is growing. This thesis presents the development of a Dynamic Basket Recommendation System tailored for the automotive market, leveraging cutting-edge technologies in Language Models (LMs) and Multi-Agent Systems (MAS), and exploiting a traditional similarity-based approach. The system is designed to ingest real-time data and respond to evolving user preferences and market offerings, addressing the challenges posed by frequent updates in vehicle models, trims, and features. The foundation of the system lies in the integration of GPT-based models, specifically AzureChatOpenAI, orchestrated through LangChain and LangGraph, two useful Python libraries that enable modular agent-based architectures. The project was developed in collaboration with JATO Dynamics, a global leader in automotive data intelligence. Using JATO’s structured dataset Carspecs, the system recommends vehicles based on either user-defined preferences or improvements over an existing car. The thesis explores two primary use cases: (1) Generic Search, where users describe desired vehicle features, and (2) Improvement Request, where users provide a reference vehicle and optionally specify desired enhancements. The proposed framework consists of three core components: a Chatbot for user interaction, an Orchestrator for agent/tool selection, and a Basket Generator Agent that utilizes tools like GenericSearchTool and SimilarCarsTool. The system employs some methods from LangGraph to integrate LLMs with callable tools, using GPT-4o-mini with moderate creativity settings. The dataset used is filtered to include UK vehicles from 2024 onward, and it is enriched by merging it with a second dataset containing monetary values for equipment features. These features are grouped into six categories (Comfort, Safety, Infotainment, Connected Car, Design, Dynamics) and normalized to ensure distributional consistency. Several tools are developed to process and analyze the data: SimilarCarsTool, GenericSearchTool, CarImproverTool, OrdinalCollectorTool, MapperTool, and AdjusterTool. These tools form a pipeline that parses user input, encodes features, and computes similarity using cosine similarity. Validation is performed by comparing the system’s recommendations with those generated by Microsoft Copilot and by consulting JATO researchers. Visual justification is provided through Principal Component Analysis (PCA) plots, showing proximity among recommended vehicles. The thesis concludes that combining AI-driven models with similarity-based algorithms offers a strategic advantage. While similarity methods are fast and interpretable, AI models excel in capturing complex, non-linear relationships and contextual information. The hybrid approach enhances both performance and explainability.