ItemConference Paper

Contextual Embedding Fusion for Quick-Commerce Reviews: Rating, 5-Class Classification, and Sentiment Analysis Across Classical ML, DL, and Transformer Models

Title
Contextual Embedding Fusion for Quick-Commerce Reviews: Rating, 5-Class Classification, and Sentiment Analysis Across Classical ML, DL, and Transformer Models
Creator
Ranka, Monica; Malhotra, Amit; Hada, Bhupendra Singh; Jana, Samiran; Maheshwari, Abhishek; Basha, Md Shaik Amzad
Description
Quick-commerce platforms depend on accurate reading of customer feedback to steer pricing and service quality. We study a dataset of delivery-agent reviews, and build two rating predictors and a text-only classifier suite. On the rating regression task (1-5), we benchmark seven models (Ridge, Lasso, ElasticNet, SVR-RBF, Gradient Boosting, Random Forest, XGBoost). The best test performance is achieved by RandomForestRegressor with MAE = 0.527, RMSE = 0.973, and R2 = 0.555; cross-validation places XGBoost and Gradient Boosting close behind. For 5-class rating classification, we compare seven TF-IDF baselines (LogReg, LinearSVC, two SGD variants, RidgeClassifier, ComplementNB, XGBClassifier) and find SGD (hinge) strongest (Accuracy = 0.817, Macro-F1 = 0.389), yet mid-rating classes (2-4) remain difficult. To address label ordinality, we fine-tune DistilBERT with ordinal heads: CORAL and CORN. These achieve Accuracy = 0.837/0.819 and Macro-F1 = 0.439/0.441, improving Macro-F1 by +5-6 points over TF-IDF baselines, mainly by rescuing classes 2-4. We further map regression outputs to 5 classes via optimized cutpoints, raising Macro-F1 from 0.329 (rounding) to 0.380. Paired, fold-wise comparisons using Wilcoxon signed-rank and bootstrap 95% CIs show consistent gains; with 5 folds the discretized p = 0.0625 indicates all folds favor the better model, suggesting statistical significance with more folds. Overall, ordinal transformers and learned thresholds provide measurable, reproducible improvements for short, noisy e-commerce reviews, delivering higher fidelity on mid-range ratings while preserving high accuracy. 2025 IEEE.
Source
Proceedings of the 2025 International Conference on Artificial Intelligence and Emerging Technology, Global AI Summit 2025;pp.1197-1202
Date
01-01-2025
Publisher
Institute of Electrical and Electronics Engineers Inc.
Subject
Contextual embedding fusion; E-commerce analytics; Quick-commerce; Rating prediction; Sentiment analysis; Transformer models
Coverage
Ranka M., Dayananda Sagar College of Arts, Science and Commerce, Bengaluru, India; Malhotra A., Christ (Deemed to be University), School of Commerce, Finance and Accountancy, Ghaziabad, India; Hada B.S., Christ (Deemed to be University), School of Commerce, Finance and Accountancy, Delhi-NCR, India; Jana S., Soil Institute of Business Design, Haryana, India; Maheshwari A., Christ (Deemed to be University), School of Commerce, Finance and Accountancy, Ghaziabad, India; Basha M.S.A., Christ Academy Institute for Advanced Studies, Gitam School of Business (Deemed to be University), Gandhi Institute of Technology and Managemet, Bangalore, India
Rights
Restricted Access; Hardcopy may be available in the library
Relation
ISBN: 979-833155379-1;
Format
online
Language
English
Type
Conference paper

Citation

Ranka, Monica; Malhotra, Amit; Hada, Bhupendra Singh; Jana, Samiran; Maheshwari, Abhishek; Basha, Md Shaik Amzad, “Contextual Embedding Fusion for Quick-Commerce Reviews: Rating, 5-Class Classification, and Sentiment Analysis Across Classical ML, DL, and Transformer Models,” CHRIST (Deemed To Be University) Institutional Repository, accessed June 19, 2026, https://archives.christuniversity.in/items/show/25754.

Output Formats