Reproducibility and Improvement Analysis on Graph Recurrent Attention Networks

Dan Liu

Abstract: In this report, three different modifications are applied to the GRAN model for the ablation studies. To examine the changes in model performance, different model components will be removed and an extra layer will be added in this experiment. In terms of removing the model components, the size of the GRU layers will be reduced from seven to two and the number of the hidden units (e.g. 512,256 and 128) will be reduced to 50. The extra angular layer is introduced to Bernoulli mixture model, so as to fulfill the improvement task. There are four groups of experiments running 5,000 iterations to study the impact of the modifications. The baseline model is the original model with the default configurations. The experiments compare the performance between the modified models and the original model. The results show that the reduction methods achieve similar performance on the lobster and point cloud dataset, while the addition of the angular layer improves models' performance.

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=SJxYOVSgUB