The document describes an experiment to evaluate search engine quality using click logs from real users. Two methods were tested: absolute metrics that measure statistics like abandonment rate for different ranking functions, and interleaved metrics that compare user preferences between ranking function pairs. While absolute metrics did not reliably identify better rankings, interleaved metrics correctly identified the superior function in each pair and differences were statistically significant. However, the study had limitations and many open questions remain about how user behavior relates to search quality.
2. Motivation
• How can we evaluate search engine quality?
Option 1: Ask experts to judge queries & result sets.
For a sample of queries, judges are paid to examine a
sample of documents and mark their relevance. This
standard process gives a reusable dataset.
Option 2: Watch how users act and hope it tells us
something about quality.
For all queries, record how users act and infer the quality
of the search results based on the logs of user actions.
3. Motivation
• How can we evaluate search engine quality?
Option 1: Ask experts to judge queries & result sets.
For a sample of queries, judges are paid to examine a
sample of documents and determine relevance. This
standard process gives a reusable dataset.
Option 2: Watch how users act and hope it tells us
something about quality.
For all queries, record how users act and infer the
quality of the search results based on the logs of user
actions.
• The key question: What is the relationship between user
behaviour and ranking quality?
4. Outline
• Describe a study of evaluation search with clicks
– Control ranking quality, and measure the effect on
user behaviour.
• Evaluation with Absolute Metrics
– Users were shown results from different functions.
– Measure statistics about user responses.
• Evaluation using Paired Comparisons
– Show a combination of results from 2 rankings.
– Infer relative preferences.
• Discuss limitations and open questions
5. Experiment Design
• Start with search ranking function f.
• Intentionally degrade performance in two
steps, making f1 and f2.
• Measure how user behaviour differs between
the ranking functions.
• Interleave results from two rankings and
measure responses.
Setup: f better than f1 better than f2
6. User Study on arXiv.org
– Real users and queries
– Users in natural context
– Degradation types:
ORIG FLAT RAND
• ORIG hand-tuned function
• FLAT ignore meta-data
• RAND randomize top-10
ORIG SWAP2 SWAP4
• ORIG hand-tuned function
• SWAP2 swap 2 pairs
• SWAP4 swap 4 pairs
– How does user behaviour change?
7. Experiment Setup
Phase 1: ORIG-FLAT-RAND
• Each user who comes to the search engine is assigned one of 6 experimental conditions:
– Results generated by ORIG
– Results generated by FLAT
– Results generated by RAND
– Results generated by interleaving ORIG & FLAT
– Results generated by interleaving ORIG & RAND
– Results generated by interleaving FLAT & RAND
Phase 2: ORIG-SWAP2-SWAP4
• Each user who comes to the search engine is assigned one of 6 experimental conditions:
– Results generated by ORIG
– Results generated by SWAP2
– Results generated by SWAP4
– Results generated by interleaving ORIG & SWAP2
– Results generated by interleaving ORIG & SWAP4
– Results generated by interleaving SWAP2 & SWAP4
8. Experiment 1: Absolute Metrics
• Measured eight easily recorded statistics
• As the ranking quality decreases, we can
hypothesize:
Metric Expected change as ranking gets worse
Abandonment Rate Increase (more bad result sets)
Reformulation Rate Increase (more need to reformulate)
Queries per Session Increase (more need to reformulate)
Clicks per Query Decrease (Fewer relevant results)
Max Reciprocal Rank* Decrease (Top results are worse)
Mean Reciprocal Rank* Decrease (More need for many clicks)
Time to First Click* Increase (Good results are lower)
Time to Last Click* Decrease (Fewer relevant results)
(*) Only queries with at least one click count
9. Experiment Statistics
• On average:
– About 700 queries a day
– About 300 distinct IPs
– About 600 clicks on results
• Each experiment phase was run for one month.
• Each experimental condition observed:
– About 3,000 queries
– About 1,000 queries with clicks
– About 600 distinct IPs.
11. Absolute Metrics: Results
• Summarizing the results, out of 6 pairs:
Summary
• Statistical fluctuations after
one month of data make
conclusions hard to draw
• None of the absolute
metrics reliably identify the
better ranking.
12. Experiment 2: Interleaved Metrics
• Paired comparisons in sensory analysis:
– Perceptual qualities are hard to test on absolute
scale (e.g. taste, sound).
– Subjects usually presented with 2+ alternatives.
– Asked to specify which they prefer.
• Can do the same thing with ranking functions:
– Present two rankings, ask which is better.
– But we’d also like evaluation to be transparent.
• So we can do an interleaving experiment.
13. Team Draft Interleaving
• Think of making high school sports teams
– We start with two captains.
– Each has a preference order over players.
– They take turns picking their next player.
• Interleaving Algorithm
– Flip a coin to see which ranking goes first.
– That ranking picks highest ranked available document.
Any clicks on it will be assigned to that ranking.
– The other team picks highest ranked available doc.
– Flip a coin again and continue.
14. Team Draft Interleaving Phase
Phase 3: ORIG-FLAT-RAND and ORIG-SWAP2-SWAP4
• Each user who comes to the search engine is
assigned one of 6 experimental conditions:
– Results generated by team-draft: ORIG & FLAT
– Results generated by team-draft: ORIG & RAND
– Results generated by team-draft: FLAT & RAND
– Results generated by team-draft: ORIG & SWAP2
– Results generated by team-draft: ORIG & SWAP4
– Results generated by team-draft: SWAP2 & SWAP4
15. Team Draft Interleaving
Ranking A
1. Napa Valley – The authority for lodging...
www.napavalley.com
2. Napa Valley Wineries - Plan your wine...
www.napavalley.com/wineries
3. Napa Valley College
www.napavalley.edu/homex.asp
4. Been There | Tips | Napa Valley
www.ivebeenthere.co.uk/tips/16681
5. Napa Valley Wineries and Wine
www.napavintners.com
6. Napa Country, California – Wikipedia
en.wikipedia.org/wiki/Napa_Valley
Ranking B
1. Napa Country, California – Wikipedia
en.wikipedia.org/wiki/Napa_Valley
2. Napa Valley – The authority for lodging...
www.napavalley.com
3. Napa: The Story of an American Eden...
books.google.co.uk/books?isbn=...
4. Napa Valley Hotels – Bed and Breakfast...
www.napalinks.com
5. NapaValley.org
www.napavalley.org
6. The Napa Valley Marathon
www.napavalleymarathon.org
Presented Ranking
1. Napa Valley – The authority for lodging...
www.napavalley.com
2. Napa Country, California – Wikipedia
en.wikipedia.org/wiki/Napa_Valley
3. Napa: The Story of an American Eden...
books.google.co.uk/books?isbn=...
4. Napa Valley Wineries – Plan your wine...
www.napavalley.com/wineries
5. Napa Valley Hotels – Bed and Breakfast...
www.napalinks.com
6. Napa Balley College
www.napavalley.edu/homex.asp
7 NapaValley.org
www.napavalley.org
AB
16. Team Draft Interleaving
Ranking A
1. Napa Valley – The authority for lodging...
www.napavalley.com
2. Napa Valley Wineries - Plan your wine...
www.napavalley.com/wineries
3. Napa Valley College
www.napavalley.edu/homex.asp
4. Been There | Tips | Napa Valley
www.ivebeenthere.co.uk/tips/16681
5. Napa Valley Wineries and Wine
www.napavintners.com
6. Napa Country, California – Wikipedia
en.wikipedia.org/wiki/Napa_Valley
Ranking B
1. Napa Country, California – Wikipedia
en.wikipedia.org/wiki/Napa_Valley
2. Napa Valley – The authority for lodging...
www.napavalley.com
3. Napa: The Story of an American Eden...
books.google.co.uk/books?isbn=...
4. Napa Valley Hotels – Bed and Breakfast...
www.napalinks.com
5. NapaValley.org
www.napavalley.org
6. The Napa Valley Marathon
www.napavalleymarathon.org
Presented Ranking
1. Napa Valley – The authority for lodging...
www.napavalley.com
2. Napa Country, California – Wikipedia
en.wikipedia.org/wiki/Napa_Valley
3. Napa: The Story of an American Eden...
books.google.co.uk/books?isbn=...
4. Napa Valley Wineries – Plan your wine...
www.napavalley.com/wineries
5. Napa Valley Hotels – Bed and Breakfast...
www.napalinks.com
6. Napa Balley College
www.napavalley.edu/homex.asp
7 NapaValley.org
www.napavalley.org
Tie!
18. Interleaving Results
• The conclusion is consistent and stronger:
(Absolute Metrics)
Summary
• Paired comparison tests
always correctly identified
the better ranking.
• Most of the differences are
statistically significant.
19. Summary of Experiment
• Constructed two triplets of ranking functions.
• Tested on real users.
• Absolute metrics didn’t change as we expected.
– Changes weren’t always monotonic.
• Interleaved gave more significant results, and
was more reliable.
– But cannot be run “after the fact” from logs.
• But there are many caveats to think about...
20. Discussion: Users & Queries
• We were only able to explore a few aspects of
the problem:
– The users are not “typical” web users.
– The type of queries is not typical.
– Results could be different in other settings:
enterprise search, general web search, personalized
search, desktop search, mobile search...
– It would be interesting to conduct similar
experiments in some of these other settings.
21. Discussion: User Interactions
• All click evaluations rely on clicks being useful.
• Presentation should not bias toward either ranking
function.
– If naively interleave two rankings with different snippet
engines, could bias users.
– But what if, say, url length just differs?
• Answer may be in the snippet (“instant answers”).
– In that case there may be no click.
– Other effects (e.g. temporal, mouse movement, browser
buttons) may give more information, but harder to log.
22. Discussion: Click Metrics
• The metrics we used were fairly simple
– What if “clicked followed by back within 5 seconds”
didn’t count?
– If we got much more data, absolute metrics could
also become reliable.
– More sophisticated absolute metrics may be more
powerful or reliable.
– More sophisticated interleaved metrics may also be.
23. Discussion: Log Reusability
• Say somebody else comes up with a new ranking
function. Are our logs useful to them?
– For absolute metrics
• Would provide baseline performance numbers.
• But temporal effects, etc, may affect evaluation.
– For paired comparison test:
• Hard to know what the user would have clicked given a
different input, so probably not
24. Conclusions
• We’d like to evaluate rankings by observing
real users: reflects real needs, cheaper, faster.
• This can be done using absolute measures, or
designing a paired comparison experiment.
• In this particular setting, the paired
comparison was more reliable and sensitive.
• There are many open questions about when
paired comparison is indeed better.