Predicting Social Media Engagement using Machine Learning

Scientific Impact Assessment

Core Contribution. This paper applies a machine learning pipeline to predict social media engagement for image-based Facebook posts, using a dataset of ~32,456 posts from 162 U.S. furniture brands. The stated contributions are twofold: (1) jointly modeling visual, textual, and temporal features (extracted via off-the-shelf tools: spaCy, TextBlob/VADER/RoBERTa, OpenCV, YOLO, CLIP) within a single gradient-boosting framework, and (2) introducing "brand-relative deviation" features (z-scores measuring how far a post departs from a brand's own historical norms), which emerge as among the strongest predictors. The work is explicitly framed as an in-depth single-industry case study rather than a general theory of engagement drivers, and it delivers practitioner-oriented recommendations for social media managers.

Methodological Rigor. This is the paper's strongest dimension relative to its modest ambitions. The authors show more methodological care than typical applied engagement-prediction work: brand-relative and z-score features are constructed using training-set statistics only and applied to held-out data; feature selection is performed within the training fold to avoid leakage; five-fold CV confirms stability (CV R² vs. test R² differ by only 0.001). Notably, they include a brand-only baseline (R²=0.688) to disentangle brand identity from content, showing content features add ~0.13 R². They further stress-test the deviation effects against seasonality (three separate checks), age effects, and brand-size confounding, and report a paired t-test establishing LightGBM's advantage over XGBoost (p=0.010). They are appropriately careful to frame findings as associational, not causal. These controls anticipate several obvious alternative explanations and lend credibility to the core claims.

Potential Impact. The impact is likely to be modest and confined to the marketing analytics / social media analytics subfield. The core methods (LightGBM, SHAP, CLIP, YOLO) are all standard, and the empirical setting is narrow (one industry, one platform). The findings — hashtags help, colorful/multi-image posts perform better, consistency with brand norms matters, brand identity dominates engagement — are largely consistent with or incrementally refine prior work (e.g., Gkikas et al., Li & Xie). The brand-relative deviation feature construction is the most transferable idea and could be reused by others studying branded content, but it is a feature-engineering refinement rather than a conceptual breakthrough. Practitioner value is real but framed cautiously ("suggestions worth testing").

Timeliness & Relevance. Social media engagement prediction is an active and commercially relevant area, and the integration of modern vision-language embeddings (CLIP) with tabular boosting is current. However, the paper does not address a pressing methodological bottleneck; it consolidates existing tools rather than pushing a frontier.

Strengths.

Limitations.

Additional Observations. The R²=0.8178 headline figure is inflated in perceived usefulness by the brand-identity component; the more honest and interesting result is the ~0.13 incremental R² from content. The paper is commendably transparent about this. Resource requirements are low (scraping + single-GPU-scale training), making the work easy to extend but also unremarkable in scale. The paper reads as a competent, well-executed applied study — likely to be cited occasionally within marketing analytics but unlikely to shift practices broadly or become foundational.

Overall, this is a solid, methodologically careful applied case study with limited novelty and narrow scope. Its rigor exceeds its ambition; its impact will be incremental.