The primary aim of this project is to implement and evaluate GAIN (Generative Adversarial Nets) for the task of missing data imputation. The goal is to demonstrate that GAIN provides superior imputation accuracy compared to traditional methods like Mean Imputation and MICE. This is achieved through an adversarial training process where a Generator network learns to fill in missing values, while a Discriminator network attempts to distinguish the imputed values from the real, observed ones. The entire system is trained via a minimax optimization process until the Generator can effectively "fool" the Discriminator. Once trained, the Generator is used to produce the final, complete dataset.
The architecture of GAIN is built around two multi-layer fully connected neural networks: a Generator and a Discriminator. The process begins by creating a mask vector (M) to identify the locations of missing data. The Generator takes the partially observed data, the mask, and a random noise vector as input to output a completed data vector. The Discriminator then receives this imputed vector along with a "hint" vector, a critical mechanism that provides partial information about the mask. The Discriminator's objective is to analyze the completed vector and predict the original mask, thereby trying to identify which values were originally observed and which were imputed by the Generator.