Online Social Media and Event-based Social Networks are exploring a vast volume of unwanted information on the proliferation of user-generated content. Social Media platform is envisioned to provide a member to generate and exchange range with other members in social media for socializing and knowledge sharing. However, it exhibits numerous complications in the network, especially to members' walls, with the propagation of unwanted information from cyber bullies and haters to their information posted as it spreads among various groups and communities. However, many traditional techniques are employed to manage these challenges, but it produces the wrong misinterpretation. To mitigate the challenges mentioned above, a new deep learning architecture named a deep social attention network is designed to detect and prevent the propagation of unwanted content in social media. In this architecture, the dataset is collected from Facebook. It contains various posts for various statuses and is available in a CSV file. CSV file is preprocessed using data normalization, stop removal, and stemming process for removing stop words, numbers, hashtags, and emojis. Preprocessed data is employed in the vector space model to process the data as the feature vector. Further, the feature vector is employed in the BERT model to extract the sentiment of the feature vector as a positive instance or negative instance. The positive polarity of the example of the vector is classified as a regular comment, and the negative polarity of the feature vector model is classified as an unwanted comment using a convolution Neural Network learning classifier considered as deep learning architecture. Negatively Classified content is eliminated and prevented from posting in the user wall. Instead, it generates a warning message to the unwanted content posted to members, and unwanted information is placed in the LTSM model as hidden data against exposing the member and other audiences. Experimental analysis of the proposed approach is compared with conventional methods to evaluate the performance. Further performance of the system is computed using performance metrics such as precision, recall, and fmeasure. The proposed model produces 99% accuracy in identifying unwanted information compared to conventional approaches.

Unwanted Content Classification, Convolution Neural Network, Deep Learning, Sentiment Analysis