Simple metrics are producing different values for the same recommendations

[nikosT]

Describe the bug
Simple metrics ItemCoverage, UserCoverage, NumRetrieved are producing different values for the same recs/ folder between the procedure of calculating the model and evaluating it and the procedure of loading the model and evaluating it.

To Reproduce
Steps to reproduce the behavior:

1. Run a model e.g. itemKNN which will produce the recommendations in the results/recs/ dir
2. Keep evaluation of the running model
3. Load the results/recs/ dir as a RecommendationFolder
4. Run evaluation only
5. Compare the two evaluations

In both cases, the input dataset has a strategy: fixed. The train.tsv and test.tsv files have previously been produced from a random 0.2 splitting procedure and used as they are for both cases.

System details (please complete the following information):

• OS: Debian
• Python Version 3.8
• Version of the Libraries installed with conda elliot_env as described in doc