Default K1 and B parameters for bm25_weight

[fsroque]

Hi, the default parameters from K1 and B are set to 100 and 0.8 respectively in the function bm25_weight defined in nearest_neighbours.py. The literature suggests values between 0 and 3 for K1 and 0 and 1 for B. Is there a specific reason for K1 to be set this high by default? Might work poorly as a first approach with those parameters.

[benfred]

I set K1 that way for the last.fm example where it works pretty well, but I agree it is probably a poor default value in general.

I'm going to leave this open for now. One thing I want to do is get better default values for all parameters (across models) by running experiments on multiple datasets and selecting defaults that usually work well, of which K1 etc will be one of them.

[christopheralex]

Any update on this yet @benfred ? Also can anyone help me with the literature for bm_25 weights ?

[singsinghai]

Any update on this yet @benfred ? Also can anyone help me with the literature for bm_25 weights ?

I think you can relate to this blog to understand bm_25 clear: https://kmwllc.com/index.php/2020/03/20/understanding-tf-idf-and-bm-25/