The document presents a framework for optimizing query recommendations to gently guide users towards more rewarding queries or search sessions. It formulates the problem of recommending the next query as an optimization problem that aims to maximize the expected reward of the user's search path. It proposes greedy heuristics to solve the single-step and multi-step recommendation versions of this problem and evaluates their performance on query logs. The results show the heuristics can increase expected reward compared to baselines without decreasing relevance of recommendations.
5. As you set
out for Ithaca,
hope your road
is a long one,
full of adventure,
full of discovery.
K. Kavafis
6. Problem
Given a set of possible user histories
6 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
7. Problem
Given a value for different states
7 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
8. Problem
Given a value for different states
8 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
9. Problem
“Nudge” the users in a certain direction
-
-
+
+
9 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
10. Problem
“Nudge” the users in a certain direction
10 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
11. Problem definition
Given a set of possible user sessions
Given a certain value for different states
“Nudge” the users in a certain direction
Objectives
– 1. collect a large reward along the way -or-
– 2. end the session at a rewarding action
11 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
12. Constraints
• We are not almighty
Source: not-of-this-earth.com
12 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
13. Constraints
• We are not almighty
– We can only suggest, not order
13 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
14. Constraints
• We are not almighty
– We can only suggest, not order
• We are not all-knowing
Source: Wikimedia commons.
14 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
15. Constraints
• We are not almighty
– We can only suggest, not order
• We are not all-knowing
– We do not know how the users will react
15 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
17. Setting
• Talk: query recommendation
• Paper: general framework
– E.g.: optimizing links in web-sites
17 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
20. Query recommendations
• Reformulation probabilities
– P(q,q') original
0.3 central park map
0.2 central park hotel
central park
0.2 central park new york
0.3
central park zoo
20 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
21. Query recommendations
• Reformulation probabilities
– P(q,q') original
– P'(q,q') perturbed = P(q,q') + ρ(Q,q,q')
-0.1 0.3 central park map
-0.1 0.2 central park hotel
central park
+0.1 0.2 central park new york
+0.1 0.3
central park zoo
21 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
22. Query recommendations
• Reformulation probabilities
– P(q,q') original
– P'(q,q') perturbed = P(q,q') + ρ(Q,q,q')
0.2 central park map
0.1 central park hotel
central park
0.3 central park new york
0.4
central park zoo
22 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
23. Example values w(·)
• Search engine results page
– Quality of search results
• General page
– Dwell time
– User ratings
23 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
24. Objective functions U(·)
Kavafian Machiavellian
“a road full of adventure” “ends justify means”
U(<q1,q2,...,qt>) = Σw(qi) U(<q1,q2,...,qt>) = w(qt)
24 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
25. The Kavafian objective
Useful when users want to explore, or be entertained
25 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
26. The Kavafian objective
Useful when users want to explore, or be entertained
+ “funny video”
-
26 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
27. Optimization problem
• Given:
– Original transition probabilities P
– Starting node q
– Node values w
– Perturbation function ρ
• Add up to k links (per node), maximize
expected utility of paths starting at q
27 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
31. Multi-step recommendation
• Recommend at each step in the future
-
- +
+
q - + t
-
+ -
+
31 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
32. Multi-step case
• Multi- and Single-step
recommendation are NP-complete
– Reduction from MAXIMUM-COVER
• Heuristic for multi-step problem:
– at each node, assume rest of the graph is
unperturbed when computing utility of
adding an edge
32 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
33. Single-step case
• Greedy heuristic for “Machiavellian”
objective: find (qi, qj) maximizing
ρij(E[U(path(qj))] – wi)
• Repeat k times
33 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
34. Observation
• Greedy heuristic achieves utility at least (1-x) of
the optimum, with x << 1 in cases of practical
interest
– It is possible to construct pathological
instances s.t. that greedy performs poorly
– x depends on termination probability at qi
and probabilities of following
recommendations
34 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
36. Large-scale experiment
We observe perturbed probabilities P'(q,q'),
unless we disable search assist to see P(q,q')
36 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
37. Empirical observations
• Notation:
– τq, τ'q session-end probabilities
• Recommendations decrease termination
probability:
• Average τq≈0.90
• Average τ'q≈0.84
37 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
39. Empirical observations
• Recommendations decrease
termination probability, τq≈0.90 τ'q≈0.84
• Decrease is almost entirely due to
more clicks on recommendations
39 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
41. Empirical observations
• Recommendations decrease
termination probability from ≈0.90 to
≈0.84
• Decrease is almost entirely due to
more clicks on recommendations
• ρ is difficult to estimate
41 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
45. The rest are in general
P'(q,q'): with recommendations difficult to predict
P(q,q'): without recommendations
46. So what do we do?
• We approximate ρ by a linear function on
– P(q,q')
– Textual similarity of q and q'
– Terminal probability of q
• We have a low accuracy on this prediction
– r ≈ 0.5
• We use as weights the CTR on results
46 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
48. Evaluation results
Baselines:
– Prefer queries with large w
– Prefer queries with large ρ
– Prefer queries with large ρw
Greedy heuristic performs better for both
utility functions
48 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
50. Evaluation results
Greedy heuristic performs well
What about relevance?
420 queries assessed by 3 judges
There were no significant changes in
relevance between systems
50 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10
51. Conclusions
• General framework for
“nudging” users in a
certain direction
• Open algorithmic and
practical questions
• In the paper: related work
51 A. Anagnostopoulos, L. Becchetti, C. Castillo, A. Gionis WSDM'10