[ASK] <xDeepFM> help on how to set FEATURE_COUNT values for unseen data for training and for inference.
#1.830 geöffnet am 18. Okt. 2022
Repository-Metriken
- Stars
- (17.706 Stars)
- PR-Merge-Metriken
- (Durchschn. Merge 6T 16h) (10 gemergte PRs in 30 T)
Beschreibung
Hi all, I am trying to understand the FEATURE_COUNT values in xDeepFM model. From the code, I understand that FEATURE_COUNT values is determine by the numbers of features generated when creating the dictionary mapping using LibffmConverter. However, when I try to predict on new dataset, it will throw out of bound or indices not found error. If I am not mistaken, from the paper (xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems), the Feature_COUNT are used to determine the size of the embedding layer according to section 2.1.
If this is the case, do we arbitrary increase the number of FEATURE_COUNT values in order to cater for unseen dataset during inference? But will this increase the size of the embedding layer which might not be used at all during training, in order to cater for inference for unseen data? Or am I mistaken all together, and there is a way to handle unseen data and calculate the FEATURE_COUNT values?