Product Review Categorization Using Ripper Rule Induction Algorithm in Digital Feedback Systems
Annotatsiya
The problem with purely predictive accuracy is that it doesn't provide any level of auditability. The routing of a customer review to a business function (e.g. the shipping team, product quality team, customer service team or pricing team) requires an operational need that predictive accuracy can't solve: it needs to be auditable to the staff to whom it's being sent, meaning a review that's been routed to the wrong team must be traced back to the correct and explainable reason and not accepted as an opaque model output. Cohen's (1995) rule induction algorithm, RIPPER, incrementally constructs a human-readable set of IF-THEN rules via grow, prune and optimise steps, and provides a categorisation process with a traceable explicit inspectable decision made at each step. This paper applies RIPPER to six-category review classification (Positive/No-Issue, Product Quality, Shipping/Delivery, Customer Service, Sizing/Fit, Pricing/Value) across 200,000 reviews from three e-commerce marketplaces (Marketplace A: 86,400; B: 68,400; C: 45,200), extracted into 1,200 candidate boolean features (lexical n-grams, sentiment indicators, star rating, length bins) and a held-out test set of 40,000 reviews (20%). RIPPER's grow-prune-optimise cycle ($\mathbf{k}=\mathbf{2}$optimisation passes; 2/3-1/3 growing-pruning split per rule) converges to a final rule set of 84 rules using 48 of the 1,200 candidate features, achieving 80.4% accuracy and macro-$\mathbf{F} \mathbf{1} \boldsymbol{=} \mathbf{0. 7 6 4}$-outperforming Naive Bayes (68.4%; +17.5%) and a C4.5 decision tree$(76.4 {\%}; +5.2 {\%})$, and trailing a 500-tree Random Forest (84.8%; −5.5%) by a margin the deployment evidence below suggests is operationally acceptable given the corresponding interpretability gain. Deployed as an automated review-triage system, RIPPER's auto-routing reduces mean time-to-route from 4.8 hours (manual triage) to 14.4 minutes (95.0% reduction) and reduces triage staffing requirements by 87.5% (4.8 to 0.6 FTE-equivalent hours per day), while the rule set's inspectability enabled a six-month continuous-refinement pilot in which staff-identified corrections to 18 of the 84 rules raised live accuracy from 80.4% to 84.2%-a refinement pathway materially more difficult to execute against a 500-tree black-box ensemble.
Hali tarjima qilinmagan