“Mitigating YouTube Recommendation Polarity Using BERT and K-Means Clustering”, Zoheb Ahmad, Danish Quasim, Shreeya Garg, Vinod Kumar2023-03-15 (retrieval AI, Transformer, politics)⁠:

YouTube’s recommendation system is famous for its success in maintaining high retention rates. The cause of its success is its ability to learn and predict an individual user’s preferences appropriately. An unintended consequence, however, is that users get stuck in what is known as their own “echo chambers” when dealing with and feeding users back their preferences. These echo chambers can cause increasing perspective bias within users, making it difficult for users to understand differing opinions.

This work aims to prepare a model that counteracts YouTube’s recommendation system by forcefully exposing users to content from varying viewpoints. The SSKA pipeline (Suno Sabki, Karo Apni) is a complementary deep learning model that involves Natural Language Processing (NLP) and K-Means clustering. It uses modern software libraries such as the YouTube API (Application Programming Interface) for data collection and was trained and tested on a varied set of users.

The results prove that the model is successful in decreasing the bias recommendation by exposing users to the content of varying opinions and helping them break away from their echo chambers. The proposed methodology of explicitly exposing users to the content of varying opinions can positively impact local societies and the global community.

[Keywords: Bidirectional Encoder Representations from Transformers (BERT), K-Means Clustering, encoder, Google Universal Encoder (GUE), polarity co-efficient, natural language processing (NLP)]