Bridging Theory and Prediction: A Hybrid Explainable SEM–Machine Learning Approach to Consumer Purchase Intention

Abstract

The growing use of Instagram as a visual and interactive marketing platform has intensified scholarly interest in how social media content shapes consumer purchase intention. However, most prior studies have relied either on theory-driven Structural Equation Modeling (SEM) or data-driven machine learning, with limited integration between causal explanation, predictive evaluation, and model interpretability. This study addresses this methodological gap by proposing a hybrid explainable SEM–machine learning framework that combines PLS-SEM, XGBoost, and SHAP to examine the relationship between social media content, brand image, and purchase intention. Data were collected from 500 Indonesian Instagram users exposed to fashion and lifestyle brand-related content. The PLS-SEM results show that social media content significantly affects brand image (β = 0.581, p < 0.001), while brand image significantly influences purchase intention (β = 0.511, p < 0.001). Brand image also significantly mediates the relationship between social media content and purchase intention, with a significant indirect effect (β = 0.297; 95% BC-CI: 0.241–0.356). In the predictive stage, Linear Regression and tuned XGBoost demonstrated stable generalization, with test R² values of 0.288 and 0.277, respectively, while Random Forest showed overfitting with a negative test R². SHAP analysis revealed that brand image was the strongest predictive feature (mean |SHAP| = 0.302), followed by social media content (0.268), indicating that brand image plays a more prominent role in forecasting purchase intention. The findings contribute theoretically by reinforcing brand image as a key mediating mechanism, methodologically by integrating validated latent constructs into explainable machine learning, and practically by offering digital marketers a dual-lens approach that combines structural explanation with predictive importance.