NDCG-score in top-k recommendations

Question

Rating prediction taks in Recommender Systems. I have prepared a custom data set, and I applied models to predict rating which user gave to the item. I keep getting the results, where NDCG-5 is higher than NDCG-3. Is it really possible? Attached, I send file with the results per user.

I tried two libraries to calclulate NDCG with python:

1. sklearn.metrics.ndcg_score
2. tfr.keras.metrics.NDCGMetric

I expected that NDCG-5 will be lower. Instead, it is higher!

Answer 1

I figured out that the NDCG for top-5 can be (indeed) be higher than NDCG for top-3. Some models provide a higher ranking for top-3, but lower for top-5. The example is shown below in the code: