In the rapidly evolving landscape of social media, understanding user behavior and predicting engagement rates are crucial for optimizing content strategies and enhancing user experience. The paper represents a comprehensive approach toward classification of Instagram users and the prediction of engagement rates by state-of-the-art machine learning techniques. High variance in engagement rates from fake and inactive users mostly acts as a barrier to the learning process of the model. In this regard, we proposed a new approach to reduce these variations in order to enhance the performance of the model. Two-class classification, followed by a regression analysis. First of all, classification into real and fake accounts, another one is into active and inactive accounts. Further, the prediction of engagement rate was done only on the shortlisted real and active users, and the best performing model—MSE that turned out to be 148.6313—was obtained from XGBoost. To add to this, a performance in the classifier—84.88% accuracy achieved by CatBoost in the real and fake account classification—and a Voting Classifier turning in an accuracy of 96.66% in active and inactive account classification. These results were useful for the multi-stage approach by using the strengths of various machine learning models to obtain strong results in classification and prediction. This study contributed to the growing field of social media analytics by providing insights into user classification and engagement prediction, with potential applications in marketing, content curation, and user retention strategies.