RECOMMENDING FOR LONG-TERM MEMBER SATISFACTION AT NETFLIX

Summary from the TLDR Web Dev newsletter

Netflix optimizes recommendations for long-term member satisfaction by going beyond short-term engagement metrics like clicks. It uses a contextual bandit framework where user actions, like completing a season or giving a thumbs-down, are used to define a proxy reward function. This function aims to capture long-term satisfaction and is iteratively refined through a process called reward engineering. To address the challenge of delayed feedback, Netflix predicts missing feedback using observed data and integrates it into the proxy reward function, allowing for better bandit policy updates.