Document Type : Original/Review Paper

Authors

• mohammad khaki ¹
• fereshte dehghani ²

¹ School of Computer Engineering, Iran University of Science and Technology (IUST), Tehran, Iran

² Department of Electrical and Computer Engineering, University of Kashan, Kashan, Iran

Abstract

The exponential growth of online food platforms has transformed restaurant discovery, yet traditional recommendation systems often struggle with the "cold-start problem" and the inability to capture latent synergies between restaurant attributes. This study proposes a multi-stage machine learning framework designed to uncover hidden patterns within Zomato restaurant data to enhance rating prediction accuracy. To overcome the limitations of raw, uncurated datasets, we implement a hybrid methodology integrating unsupervised latent structure discovery with supervised ensemble classification. Specifically, a dual-clustering strategy (K-means and Hierarchical clustering) is employed to synthesize novel latent features that represent complex relationships between service quality, price, and user preferences. To ensure model robustness, we utilize the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance and Sequential Feature Selection (SFS) to optimize the input space. Experimental results demonstrate that the proposed framework significantly outperforms traditional baseline models, with the Random Forest classifier achieving a 90\% accuracy and a 37\% absolute improvement in F1-score. Granular error analysis reveals that most misclassifications are confined to adjacent rating categories, indicating a high degree of ordinal consistency. These findings underscore the effectiveness of leveraging latent structures to build scalable and interpretable recommendation engines, particularly in data-sparse environments.

Keywords

• Recommendation Systems
• Machine Learning
• Customer Rating Prediction
• Latent Structure Discovery
• Zomato Dataset

Main Subjects

• H.3. Artificial Intelligence