It would be great if we could know which training or validation era is ‘close’ to the live era to enhance our model performance in the tournament. This knowledge may help us avoid a burning era by focusing our modeling on improving its validation performance on that ‘close’ era.
Here I use a simple deep metric learning approach to do that in the following kaggle notebook.
Siamese Networks are mainly used in combination with a Contrastive Loss function aka Pairwise ranking Loss. The loss function is mainly used to learn embeddings (feature vectors) in a way that the metric distance between two examples from the same class is small and that between different classes is large in a metric space.
Triplet network is also a Symmetric neural network architecture but consists of three identical subnetworks that share the same sets of parameters.