ir evaluation using rank-biased precision
DESCRIPTION
How IR systems (search engines) are evaluated, in particular under the TREC methodology. The common measure of Mean Average Precision is discussed and compared to the newly proposed (Moffat and Zobel 2008) Rank-Biased Precision. For more discussion, see: http://alteregozi.com/2009/01/18/evaluating-search-engines-relevance/TRANSCRIPT
Measuring Search Engine Quality
Rank-Biased PrecisionAlistair Moffat and Justin Zobel, “Rank-Biased Precision for Measurement of Retrieval Effectiveness”, TOIS vol.27 no. 1, 2008.
Ofer EgoziLARA group, Technion
Introduction to IR Evaluation
Mean Average Precision
Rank-Biased Precision
Analysis of RBP
Outline
Task: given query q, output ranked list of documents◦ Find probability that document d is relevant for q
IR Evaluation
Task: given query q, output ranked list of documents◦ Find probability that document d is relevant for q
Evaluation is difficult◦ No (per query) test data
◦ Queries vary tremendously
◦ Relevance is a vague (human) concept
IR Evaluation
Precision / recall
◦ Precision and recall usually conflict
◦ Single measures proposed (P@X, RR, AP…)
Elementary IR Measures
Drel(q,D)alg(q,D)
Precision: |alg rel|/|alg| Recall: |alg rel|/|rel|
Relevancy requires human judgment◦ Exhaustive judging is not scalable
◦ TREC uses pooling
◦ Shown to miss significant relevant portion…
◦ … but shown to compare cross-system well
◦ Bias against novel approaches
Pooling for Scalable Judging
In real-world, what does recall measure?◦ Recall important only with “perfect” knowledge◦ If I got one result, and there is another I don’t
know of, am I half-satisfied?...◦ …yes, for specific needs (legal, patent) session◦ “Boiling temperature of lead”
How Important is Recall?
In real-world, what does recall measure?◦ Recall important only with “perfect” knowledge◦ If I got one result, and there is another I don’t
know of, am I half-satisfied?...◦ …yes, for specific needs (legal, patent) session◦ “Boiling temperature of lead”
Precision is more user-oriented◦ P@10 measures real user satisfaction◦ Still, P@10=0.3 can mean first three or last
three…
How Important is Recall?
Calculated as ◦ Intuitively: sum all P@X where rel found, divide by
total rel to normalize for summing across queries Example: $$---$----$-----$---
(Mean) Average Precision
Calculated as ◦ Intuitively: sum all P@X where rel found, divide by
total rel to normalize for summing across queries Example: $$---$----$-----$--- Consider: $$---$----$-----$$$$
◦ AP is down to 0.5234, despite P@20 increasing
◦ Finding more rels can harm AP performance!
◦ Similar problems if some are initially unjudged
(Mean) Average Precision
Methodological problem of instability◦ Results may depend on judging extent
◦ More judging can be destabilizing (meaning error margins don’t shrink with reducing uncertainty)
MAP is Unstable
Complex abstraction of user satisfaction◦ “Every time a relevant document is encountered, the user pauses, asks “Over the
documents I have seen so far, on average how satisfied am I?” and writes a number on a piece of paper. Finally, when the user has examined every document in the collection — because this is the only way to be sure that all of the relevant ones have been seen — the user computes the average of the values they have written.”
How can R be truly calculated? Think evaluating a Google query…
MAP is not “Real-Life”
Complex abstraction of user satisfaction◦ “Every time a relevant document is encountered, the user pauses, asks “Over the
documents I have seen so far, on average how satisfied am I?” and writes a number on a piece of paper. Finally, when the user has examined every document in the collection — because this is the only way to be sure that all of the relevant ones have been seen — the user computes the average of the values they have written.”
How can R be truly calculated? Think evaluating a Google query…
Still, MAP is highly popular and useful: ◦ Validated in numerous TREC researches◦ Shown to be stable and robust across query sets (for
deep enough pools)
MAP is not “Real-Life”
Enter RBP…
Induced by a user model
Rank-Biased Precision
Induced by a user model
◦ Each document is observed at probability pi-1
◦ Expected #docs seen:
◦ Total expected utility (ri = known relevance function):
◦ RBP = expected utility rate = utility/effort
Rank-Biased Precision
Values of p reflect user behaviors◦ P=0.95 persistent user (60% chance for 2nd page)
◦ P=0.5 impatient (0.1% chance for 2nd page)
RBP Intuitions
Values of p reflect user behaviors◦ P=0.95 persistent user (60% chance for 2nd page)
◦ P=0.5 impatient (0.1% chance for 2nd page)
◦ P=0 I’m feeling lucky (identical to P@1)
RBP Intuitions
Values of p reflect user behaviors◦ P=0.95 persistent user (60% chance for 2nd page)
◦ P=0.5 impatient (0.1% chance for 2nd page)
◦ P=0 I’m feeling lucky (identical to P@1)
Values of p control contribution of each relevant document◦ But always positive!
RBP Intuitions
Evaluating a new evaluation measure…
RBP Stability
Uncertainty: how many relevant documents? (down the ranking, or even in current depth)
RBP value is inherently lower bound
Error Bounds
Uncertainty: how many relevant documents? (down the ranking, or even in current depth)
RBP value is inherently lower bound Residual uncertainty is easy to calculate –
assume relevant…
Error Bounds
RBP in Comparison
Similarity (correlation) between measures
Detected significance in evaluated systems’ ranking
RBP has significant advantages:◦ Based on a solid and supported user model
◦ Real-life, no unknown factors (R, |D|)
◦ Error bounds for uncertainty
◦ Statistical significance as good as others
But also:◦ Absolute values, not relative to query difficulty
◦ A choice for p must be made
Conclusion