Heikki's experiments with ranking Personal notes, likely to be out of date. Things to experiment with, and find out, and mess about Goals: - Understand the ranking - make a better ranking merging algorithm Tue 19-Nov-2013 Started this branch Wed 20-Nov-2013 Make a script that tests ranking against yaz-zserver (as that is the default config). Mostly to have a script to build on later. Thu 21-Nov-2013. Start my own complete config Fri 22-Nov-2013. Adam defined a new sort type, relevance_h, and put it place in the code. Now I have a place to implement my stuff. Relevant places: pazpar2_config.c:1020 - minor session.c:1318 - call relevance_prepare_read also for my type reclists.c:104 - parse params reclists.c:166 - compare function (for quicksort) relevance.c:417 - calculate score (same function as for relevance, but with extra arg for type) The compare function compares positions, when sorting by Metadata_sortkey_position This loops through the records (in the cluster) and finds the smallest rec->pos and then compares those. Next: See if I can implement a round robin. - clients.h declares int clients_count(void) - rec->client is a pointer to the client, but we don't have an ordinal from that - keep an array of structs with the pointer, and locate the client number that way - robin-score = pos * n_clients + client_num relevance_new_rec is called every time a new record pops up. One or more to count_word, exactly one to done_rec. That's where I can compare to the ranking of the previous record. struct_relevance is one structure I have for myself, global (for the user session), so I can keep my stuff in there, possibly an array of things for each target. I should also add stuff directly to the client, and to the record, as I need. Next: Plot the tf/idf scores against round-robin sorted order. Will be messy, but later when we get a target that returns sorted records, it will make sense. Wed 27-Nov Setting up multiple SOLR targets in the same pazpar2 - Add #999 to the z-urls, so pazpar2 won't merge them. Different number for each This URL shows the databases, with their numbers http://lui.indexdata.com/solr/select?q=database:*&facet=true&facet.method=fc&facet.field=author_exact&facet.field=subject_exact&facet.field=date&facet.field=medium_exact&facet.field=database&rows=0&facet.mincount=1 Add this to the target defs After this, it should be possible to get records from different databases, some with many records, some with a few. This is a good testing ground for merging rankings! Test first with a round-robin, and plot the scores. Thu 28-Nov Ok, I can now merge a number of SOLR databases (harvest jobs), and plot their rankings as solr gives them, in the order of different merge strategies Next: Add the normalizing merge strategy. Then plot different strategies against different queries Write a conclusion, and consider this plotting job done