The pervasiveness of false information in social media has enabled the fast propagation of rumors driven by mob behavior and viral content. This leads to biased public opinion, weakened journalistic reputation, and deepening of social polarization. For detecting fake news on social media, in this work we have proposed a supervised machine learning based approach. Often we have seen that the news websites/sources which are famous, that is having high credibility, post real news. We have utilized the concept of publisher’s trustworthiness to generate a dataset using web scraping. The base attributes of the comprehensive dataset are scraped based on user interaction data like likes, shares, and publisher information like followers and following. After web scraping, we have performed data annotation that is labelling the datapoint as real or fake using a feature called credibility_score. The credibility score formula is derived based on multiple base features. If credibility score crosses a certain threshold then the datapoint is labelled as real else labelled as fake. Based on these base attributes, 3 derived features like likes_ratio, shares_ratio and comments_ratio are calculated. Supervised machine learning models are evaluated on the entire dataset containing all the base features and the 3 derived features. Decision Tree has achieved the highest accuracy of 99.92% and KNN has achieved the lowest accuracy. Exclusion of the 3 derived features has reduced the model accuracy to approximately 50% for all algorithms and demonstrates their importance. This paper also addresses the accuracy vs. training time trade-off and the relevance of optimized algorithms for real-time applications.