Test Loss behaviour

[rishj97]

When training, the test loss does not decrease (as it should), but the Hit Ratio keeps increasing. Doesn't this suggest that the loss function being used is incorrect? For example here when training DeepICF (with attention, same config params as the README) -

Epoch 0 [153.6s + 600.3s]: HR = 0.5296, NDCG = 0.2971, loss = 0.2525 [54.1s] train_loss = 0.3235 [184.5s]
Epoch 1 [173.1s + 594.7s]: HR = 0.5957, NDCG = 0.3382, loss = 0.2550 [49.2s] train_loss = 0.2908 [152.5s]
Epoch 2 [163.4s + 541.1s]: HR = 0.6270, NDCG = 0.3589, loss = 0.2763 [48.7s] train_loss = 0.2734 [168.5s]
Epoch 3 [165.3s + 540.0s]: HR = 0.6444, NDCG = 0.3757, loss = 0.2468 [49.2s] train_loss = 0.2640 [169.9s]
Epoch 4 [167.5s + 522.3s]: HR = 0.6526, NDCG = 0.3818, loss = 0.2376 [49.0s] train_loss = 0.2592 [162.8s]
Epoch 5 [166.1s + 528.3s]: HR = 0.6588, NDCG = 0.3868, loss = 0.2655 [42.5s] train_loss = 0.2539 [149.1s]
Epoch 6 [163.5s + 491.9s]: HR = 0.6596, NDCG = 0.3893, loss = 0.2553 [44.8s] train_loss = 0.2512 [151.7s]
Epoch 7 [164.6s + 456.7s]: HR = 0.6666, NDCG = 0.3938, loss = 0.2512 [36.5s] train_loss = 0.2488 [125.5s]

And so on... Do you know what might be causing this or if there is a way to make the test loss decrease consistently as well (like the training loss)?

[Honolu]

same question @linzh92