New subject: Search Satisfaction/Success/Whatever Metrics and The Relevance Forge

4 Mar 2016

This thread started off list, but I'm hoping all of you watching along can
help us along to brainstorm and improve search satisfaction. Note that
these aren't all my thoughts, they are a conglomeration of thoughts (many
copy/pasted from off-list emails) from Trey, David, Mikhail and I. That's
also why this might not all read like one person wrote it.

A few weeks ago I attended ElasticON and there was a good presentation
about search satisfaction by Paul Nelson.  One of the things he thought was
incredibly important, that we had already been thinking about but hadn't
moved forward enough on, was generating an Engine Score. This week Paul
held an online webinar where he gave the same presentation but without such
strict time constraints which Trey attended. You can find my summary of
this presentation in last weeks email to this list, 'ElasticON notes'

Some things of note:

   - He doesn't like the idea of golden corpora—but his idea is different
   from Trey's. He imagines a hand-selected set of "important" queries that
   find "important" documents. I don't like that either (at least not by
   itself). I always imagine a random selection of queries for a golden corpus.
   - He lumps BM25 in with TF/IDF and calls them ancient and unmotivated
   and from the 80s and 90s. David's convinced us that BM25 is a good thing to
   pursue. Of course, part of Search Technologies' purpose is to drum up
   business, so they can't say, "hey just use this in Elastic Search" or
   they'd be out of business.
   - He explains the mysterious K factor that got all this started in the
   first place. It controls how much weight changes far down the results list
   carry. It sounds like he might tune K based on the number of results for
   every query, but my question about that wasn't answered. In the demo, he's
   only pulling 25 results, which Erik's click-through data shows is probably
   enough.
   - He mentions that 25,000 "clicks" is a good enough sized set for
   measuring a score (and having random noise come out in the wash). Not clear
   if he meant 25K clicks, or 25K user sessions, since it was in the Q&A.

David and Trey talked about this some, and Trey think's the idea of Paul's
metric (Σ power(FACTOR, position) * isRelevant[user,
searchResult[Q,position].DocID]) has a lot of appeal. It's based on clicks
and user sessions, so we'd have to be able to capture all the relevant
information and make it available somewhere to replay in Relevance Forge
for assessment. We currently have a reasonable amount of clickthrough data
collected from 0.5% of desktop search sessions that we can use for this
task. There are some complications though because this is PII data and so
has to be treated carefully.

Mikhail's goal for our user satisfaction metric is to have a function that
maps features including dwell time to user satisfaction ratio. (e.g., 10s =
20% likely to be satisfied, 10m = 94% likely to be satisfied, etc.). The
predictive model is going to include a variety of features of varying
predictive power, such as dwell time, clickthrough rate, engagement
(scrolling), etc. One problem with the user satisfaction metric is that it
isn't replayable. We can't re-run the queries in vitro and get data on what
users think of the new results. However it does play into Nelson's idea,
discussed in the paper and maybe in the video, of gradable relevance.
Assigning a user satisfaction score to a given result would allow us to
weight various clicks in his metric rather than treating them all as equal
(though that works, too, if it's all you have).

We need to  build a system that we are able to tune in an effective way.
As pointed by Trey cirrus does not allow us to tune the core similarity
function params. David tend's to think that we need to replace our core
similarity function with a new one that is suited for optimizations and
BM25 allows it, there are certainly others and we could build our own. But
the problem will be:

How to tune these parameters in an effective way, with BM25 we will have 7
fields with 2 analyzers : 14 internal lucene fields. BM25 allows to tune 3
params : weight, k1, and b for each field
- weight is likely to range between 0 and 1 with maybe 2 digits precision
steps
- k1 from 1 to 2
- b from 0 to 1
And I'm not talking about the query independent factors like popularity,
pagerank & co that we may want to add. It's clear that we will have to
tackle hard search performance problems...

David tend's to think that we need to apply an optimization algorithm that
will search for optimal combination according to an objective. David
doesn't think we can run such optimization plan with A/B testing, it's why
we need a way to replay a set of queries and compute various search engine
scores.
We don't know what's the best approach here:
- extract the metrics from the search satisfaction schema that do not
require user intervention (click and result position).
- build our own set of queries with the tool Erik is building (temporary
location: http://portal.wmflabs.org/search/index.php)
-- Erik thinks we should do both, as they will give us completely different
sets of information. The metrics about what our users are doing is a great
source of information provides a good signal. The tool Erik is building
comes at the problem from a different direction, sourcing search results
from wiki/google/bing/ddg and getting humans to rate which results are
relevant/not relevant on a scale of 1 to 4. This can be used with other
algorithms to generate an independent score. Essentially I think the best
Relevance Forge will output a multi-dimensional engine score and not just a
single number.
-- We should set up records of how this engine score changes over days,
months, and longer, so we can see a rate of improvement (or lack thereof.
But hopefully improvement :)

And in the end will this (BM25 and/or searching with weights per field)
work?
- not sure, maybe the text features we have today are not relevant and we
need to spend more time on extracting relevant text features from the
mediawiki content model (https://phabricator.wikimedia.org/T128076)
  but we should be able to say : this field has no or only bad impact
impact.

The big picture would be:
- Refactor cirrus in a way that everything is suited for optimization
- search engine score: the objective (Erik added it as goal)
- Optimization algorithm to search/tune the system params. Trey has prior
experience working within optimization frameworks. Mikhail also has
relevant machine learning experience.
- A/B testing with advanced metrics to confirm that the optimization found
good combination

With a framework like that we could spend more time on big impact text
features (wikitext, synonyms, spelling correction ...).
But yes it's a huge engineering task with a lot of challenges :/ It's also