Airline Recommendation prediction Based on User Ratings using Optimized Machine Learning and Deep Learning Models

Abstract

The rich sentiment in passenger reviews is often ignored in Airline recommendation system which yields low interpretability and weak predictive power. In the present study, it is suggested that AIRRecNet (Airline Recommendation Network) be used as a multimodal deep learning framework to be trained to use textual reviews, numeric service rates, and categorical metadata to facilitate contextual airline recommendation. Two variant forms (Full AIRRecNet and AIRRecNet w/o Cross-Attention) exist in which the cross-modal attention is applied to align the features to fit and AIRRecNet w/o Cross-Attention which merely concatenates the features. AIRRecNet will outperform the existing baselines of Logistic Regression, Random Forest, XGBoost, SVM, and MLP, and a text-only DistilBERT, model trained and tested on 8,100 reviewed airline reviews (2016-2024). The Full AIRRecNet is the better (0.9321, 0.9311) in terms of accuracy and F1-score and the ablation variant is the better (0.9862) in terms of the AUC-ROC and a much quicker improvement. The outcomes of the experiment imply that the multimodal fusion is a rational approach to incorporate the latent sentiment-service dynamics that will determine the passenger satisfaction and transform AIRRecNet into a reliable and interpretable airline recommendation analytics framework of tomorrow.