Oct 3, 2020
Recent advances in machine learning have demonstrated the potential of complex models with high-dimensional hypothesis space in prediction-based tasks. By contrast, explanatory models, which are intended to describe the mechanism of a phenomenon, usually avoid such complexity. Take economic models for social networks as an example. Explanatory models in their literature are parsimonious and tractable, but typically do not account for heterogeneity in individual preferences and attributes. Hence, when used to predict individual decisions, these models may not achieve satisfactory performance.
To improve the predictability of individual-level decisions, my research advocates equipping explanatory models with machine learning techniques. Take my work on social network formation as an example . Prior economic literature has established game theoretical models describing how individuals rationally choose neighbors (e.g., friends) on social networks . To represent heterogeneous individual attributes, I incorporated node embedding techniques, i.e., using a latent vector to represent each individual, into these conventional economic models. The efficacy of our model is proven by improved prediction performance in tasks such as emerging network links and personal attributes.
A few computational social scientists have explored how to equip explanatory models with machine learning [3-6]. I will continue to contribute to this area, especially by social network models. Potential useful machine learning models are multi-agent reinforcement learning and graph neural networks.
For example, in most social network models, individuals are assumed to act under certain simple rules (maximizing a simple utility function). Multi-agent reinforcement learning can instead simulate more complex individual decisions. Importantly, rather than using a multi-agent system for prediction-based tasks, I will leverage the technique to examine how the different social network structure affects the social behavior of this multi-agent system. The results of multi-agent systems can be generalized to human systems. This proposed study is closely related to the emerging field of human-AI interactions .References: