Recommendation Classification Using Maximum Entropy Stacking Ensemble Model on Book Review Data
Abstrak
Book reviews on Goodreads serve as critical informational resources for prospective readers, yet their length and subjectivity make manually extracting recommendation intent inefficient. This study evaluates an out-of-fold probability stacking ensemble combining Naive Bayes and XGBoost as base learners with Maximum Entropy as a meta-learner, for binary recommendation classification of 900 long-form Goodreads reviews across six titles, manually annotated into recommended (82.1%) and not recommended (17.9%) categories. TF-IDF vectorization was fitted strictly within training partitions to prevent leakage, and class-weighting addressed the imbalance. Benchmarked against standalone Naive Bayes, XGBoost, and Maximum Entropy baselines and a majority-class classifier under an 80:20 split, the stacking ensemble did not achieve the highest raw accuracy (73.89%, versus 82.78% for XGBoost and 77.78% for Maximum Entropy) but achieved the highest macro-F1 (66%) and the highest recall on the minority "not recommended" class (0.75), correctly identifying 24 of 32 negative reviews. Notably, XGBoost's accuracy exceeded the majority-class floor (82.22%) by less than one point while detecting almost none of the minority class (recall 0.19), whereas the stacking ensemble was the only model to exceed this floor's macro-F1 by more than 20 points. Meta-learner coefficients indicated reliance primarily on Naive Bayes' outputs (64.9% of weight). These results suggest the stacking architecture's contribution lies in more balanced classification under substantial class imbalance rather than in maximizing accuracy, though this trade-off requires further validation through repeated resampling and evaluation across additional splits, genres, and platforms.
Identifier
- OpenAlex: https://openalex.org/W7202372567
- DOI: https://doi.org/10.32877/bt.v9i1.4132