This research improves the detection of social media abuse by assessing post-comment combinations and integrating emoji traits. Traditional approaches ignore the post-comment context and emoji semantics. We are able to detect subtle signs of spam that text-only methods miss because of these features. Combining various machine learning models, such as Support Vector Machines, Random Forest, and XGBoost, improves the classification accuracy. In order to examine actual conversations taking place on social media, this dataset makes use of spam labels. The engineering of features includes emoji sentiment, frequency, and context. Ensemble approaches seem to be superior to single-model baselines time and time again. Significant improvements in recall and precision were shown by the results. This system is capable of scaling content moderation. Future developments that require deep learning and multimodal data will likewise be made easier by this.