How to read the FactorizedTopK metric?

[zhifeng-huang]

factorized_top_k/top_1_categorical_accuracy: 0.0025 - factorized_top_k/top_5_categorical_accuracy: 0.0200 - factorized_top_k/top_10_categorical_accuracy: 0.0379 - factorized_top_k/top_50_categorical_accuracy: 0.1261 - factorized_top_k/top_100_categorical_accuracy: 0.1962

How should we read these numbers? What are the numbers to be considered as a good model? What are the numbers for the state-of-art model?

Thank you!

[EdwardALockhart]

All depends how these compare against a baseline approach.

Some ideas are: Recommending the n most popular items overall Recommending the n most popular items for that user Recommending the n most recent items for that user

See what scores you get and compare to how the model does, then you’ll know if it is any good.

[mustfkeskin]

These metrics are calculated as batches. Instead, it would be more accurate to calculate recall@k after model training is completed.

Metrics may vary depending on the problem and data. Therefore, it would be better to find and read the articles that are most suitable for you.