Predicting Consumer Purchase Intention for Pre-Prepared Meals Based on Random Forest and Explainable AI (SHAP): A Study in Jilin Province, China

TL;DR

This study uses machine learning to understand what drives consumers to buy pre-prepared meals in Jilin Province, China, and how these factors interact.

Contribution

The study introduces an integrated machine learning framework combining Random Forest and SHAP for interpretable predictions of consumer behavior.

Findings

Recommendation willingness is the top predictor of purchase intention, contributing over 72% of predictive importance.

Convenience and channel accessibility are foundational enablers of consumer decisions.

Recommendation willingness shows an S-shaped nonlinear threshold, with a shift to 'relatively willing' being a key marketing opportunity.

Abstract

The pre-prepared meal industry is a vital engine for food sector upgrading in China. This study investigates the key drivers of consumer purchasing decisions and identifies strategic pathways to support high-quality industry development. Grounded in behavioral decision theory and the stimulus–organism–response framework, we propose two central research questions: (1) What are the dominant determinants of consumer purchase intention for pre-prepared meals? and (2) How do these determinants interact in nonlinear and asymmetric ways to shape final decisions? To address these questions, we analyzed 805 valid questionnaires collected in Jilin Province using an integrated machine learning framework. Data quality and validity were ensured through baseline balance tests, and sample imbalance was corrected using the SMOTE–Tomek algorithm. Six models, including Random Forest (RF) and XGBoost, were optimized via Gaussian process-based Bayesian optimization. The RF model achieved optimal performance on the test set, with an F1 score of 0.907, an AUC of 0.928, and a prediction accuracy of 0.876. To enhance model interpretability, Mean Decrease Impurity (MDI) was integrated with the SHAP framework. Our findings reveal that: (1) purchase decisions are predominantly willingness-driven, with behavioral tendency—especially recommendation willingness—accounting for over 72% of predictive importance; (2) rational considerations, such as convenience and channel accessibility, serve as foundational enablers; and (3) recommendation willingness exhibits a significant S-shaped nonlinear threshold, where a shift to “relatively willing” marks a critical marketing intervention window. SHAP force plot analysis further uncovers an asymmetric decision logic: high willingness can compensate for perceived product shortcomings, whereas the absence of core intention functions as a non-compensatory barrier. Theoretically, these findings synthesize machine learning outputs with classical behavioral models (e.g., the Theory of Planned Behavior and Prospect Theory) by empirically quantifying bounded rationality and nonlinear activation mechanisms. These findings suggest that enterprises should transition from traffic-centric to retention-oriented strategies by leveraging word-of-mouth and proximity-based channels. Moreover, establishing a collaborative governance system is essential to mitigate risk perception and ensure long-term industry prosperity.

Figures