About the Paper
Market basket recommendation systems face a persistent challenge known as the cold start problem, in which newly introduced items or users lack the historical interaction data that conventional algorithms depend on. This paper presents a comparative study of ten recommendation architectures spanning three families: Bayesian network models, graph-based retrieval-augmented generation (GraphRAG) systems with varying degrees of knowledge integration, and traditional baselines. All architectures are evaluated under warm and cold start conditions on the Instacart grocery dataset, comprising over 32 million order-product records.
We introduce an LLM-augmented GraphRAG architecture employing a dedicated cold start handler that leverages GPT-4 semantic reasoning to maintain recommendation quality when historical data is absent. Experiments on a 500,000-record sample with 2,000 evaluation baskets reveal a nuanced performance landscape, and a final ablation isolates the cold start handler's contribution across structurally distinct deployment scenarios.
Key Insight
Graph-based representations provide an inherent buffering effect against cold start degradation that purely probabilistic models lack. When transaction data is absent, external knowledge (via GPT-4 semantic matching or a popularity prior) can bridge the gap and rescue a system from total collapse.
Three Families of Architectures
Ten systems, three paradigms, one apples-to-apples comparison
Methodological Contribution
Rather than benchmarking a single new model, the study systematically maps the warm-vs-cold trade-off space across ten architectures, separating the contribution of graph structure, LLM reasoning, and fallback strategy as independent design dimensions.
Key Results
Warm performance, cold collapse, and rescue strategies
- Warm scenario: Bayesian_V3 achieves the strongest F1@10 of 0.0528 among probabilistic models, while GraphRAG_V3 leads among graph-based methods at 0.0569.
- Cold item scenario (10% product holdout): GraphRAG_V2 proves the most resilient, with only 12.8% degradation. By contrast, GraphRAG_V3 and GraphRAG_LLM without a cold start handler suffer a complete 100% collapse.
- Cold start handler: Restores GraphRAG_LLM to F1@10 = 0.0262 with the popularity strategy and 0.0208 with LLM-based semantic matching -- demonstrating that external knowledge can bridge the gap left by absent transaction data.
- Popularity beats semantics in grocery: In this domain, a simple popularity fallback outperforms LLM-based semantic matching as a cold start strategy.
- Statistical significance: Findings confirmed via paired t-tests and bootstrap confidence intervals across 160 cold item evaluation baskets.
Why It Matters
In grocery retail, the cold start problem is not an edge case but a routine operational reality: new products are introduced weekly, seasonal items appear and disappear, and new customers begin shopping without prior transaction history. This work advances the field by:
- Providing a systematic empirical comparison of ten recommender architectures under realistic warm and cold start conditions on a 32M-record retail dataset
- Demonstrating that global graph pathways confer inherent cold start resilience that purely probabilistic models cannot match
- Introducing an LLM-augmented GraphRAG with a dedicated cold start handler that recovers performance when historical data is absent
- Identifying a cascade deployment strategy that optimises coverage across both warm and cold scenarios
Impact
These results give practitioners a principled basis for selecting recommender architectures and fallback strategies based on the warm/cold mixture of their deployment, rather than optimising for a single average-case metric.
Open Science & Collaboration
We are committed to open, reproducible research. The source code for all ten architectures, the cold start handler, and the evaluation harness will be released upon acceptance. If you are interested in collaborating, extending the framework, or accessing pre-publication artifacts, please reach out.
Full Paper & Code
Manuscript currently under peer review. The full paper, source code, and processed datasets will be released upon acceptance.