The Effects of Feedback on Human Behavior in Social Media: An Inverse Reinforcement Learning Model

We introduce and validate a learning model of human behavior change in response to feedback on social media. People who participate in these types of websites, like Wikipedia, Reddit, and others, are learning agents whose choices about how to allocate their effort are dynamic and responsive to how they feel their efforts were received in the past. By explicitly taking into account the reinforcement effects of different types of feedback received on prior contributions, our model is able to significantly outperform all known baselines in predicting future contributions both on synthetic data and on real data collected from the social news site reddit.com. Our model has an intuitive interpretation as users playing mixed strategies in a game-like setting with thousands of other users and thousands of available pure strategies. In this interpretation, our task is then inverse reinforcement learning: recovering users' reward functions based on observed behavior.