While testing metrics for pyspark evaluation, I've noticed that the ranking metrics like NDCG seems to be using binary relevances only, while python evaluation has a parameter to chose between binary, exponential or raw relevances. The snippet below shows that behavior (it will only consider which items are relevant, but not accessing their relevances):
While testing metrics for pyspark evaluation, I've noticed that the ranking metrics like NDCG seems to be using binary relevances only, while python evaluation has a parameter to chose between binary, exponential or raw relevances. The snippet below shows that behavior (it will only consider which items are relevant, but not accessing their relevances):
https://github.com/recommenders-team/recommenders/blob/c2ea583d27bb1a4d58a09a1621d5ce95672ef1dc/recommenders/evaluation/spark_evaluation.py#L292-L295
Is it possible to use exponencial or raw relevances in spark evaluation currently or am I wrong in this analysis?