Learning Socially Optimal Information Systems from Egoistic Users

Transcription

1 Learning Socially Optimal Information Systems from Egoistic Users Karthik Raman Thorsten Joachims Department of Computer Science Cornell University, Ithaca NY karthik Sept 23, 2013 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

2 Example: (Extrinsic) Diversity in Web Search Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

3 Example: (Extrinsic) Diversity in Web Search What did the user mean? Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

4 Example: (Extrinsic) Diversity in Web Search Silvercorp Metals 40% Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

5 Example: (Extrinsic) Diversity in Web Search Silvercorp Metals 40% Support Vector Machines 35% Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

6 Example: (Extrinsic) Diversity in Web Search Silvercorp Metals 40% Support Vector Machines 35% SVM Gift Cards 15% Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

7 Example: (Extrinsic) Diversity in Web Search Silvercorp Metals 40% Support Vector Machines 35% SVM Gift Cards 15% Society of Vascular Medicine 10% Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

8 Example: (Extrinsic) Diversity in Web Search Silvercorp Metals 40% Support Vector Machines 35% SVM Gift Cards 15% Society of Vascular Medicine 10% Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

9 Example: (Extrinsic) Diversity in Web Search Silvercorp Metals 40% Support Vector Machines 35% SVM Gift Cards 15% SOCIALLY BENEFICIAL Society of Vascular Medicine 10% Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

10 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

11 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Ex: News Sites, Product & Media Recommendation, Frontpages. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

12 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Ex: News Sites, Product & Media Recommendation, Frontpages. Diverse User population: N user types. User type i has probability p i. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

13 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Ex: News Sites, Product & Media Recommendation, Frontpages. Diverse User population: N user types. User type i has probability p i. Personal utility for object (e.g. ranking) y: U i (, y). Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

14 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Ex: News Sites, Product & Media Recommendation, Frontpages. Diverse User population: N user types. User type i has probability p i. Personal utility for object (e.g. ranking) y: U i (, y). Social utility is the expected utility (over all users): N U(, y) = E[U i (, y)] = p i U i (, y) i=1 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

15 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Ex: News Sites, Product & Media Recommendation, Frontpages. Diverse User population: N user types. User type i has probability p i. Personal utility for object (e.g. ranking) y: U i (, y). Social utility is the expected utility (over all users): N U(, y) = E[U i (, y)] = p i U i (, y) i=1 Goal: Find Socially Optimal Solution y = argmax y U(, y). Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

16 Socially optimal solutions in Information systems Problem: Common content for users with different tastes. Hedge against uncertainty in user s preferences. Ex: News Sites, Product & Media Recommendation, Frontpages. Diverse User population: N user types. User type i has probability p i. Personal utility for object (e.g. ranking) y: U i (, y). Social utility is the expected utility (over all users): N U(, y) = E[U i (, y)] = p i U i (, y) i=1 Goal: Find Socially Optimal Solution y = argmax y U(, y). Challenge: Learn from egoistic, weak, noisy user feedback. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

17 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

18 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

19 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Not social utility U. As in the case of prior work [RSJ12] for the intrinsic diversity problem. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

20 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Not social utility U. As in the case of prior work [RSJ12] for the intrinsic diversity problem. Need to infer social utility from such conflicting, individual feedback. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

21 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Not social utility U. As in the case of prior work [RSJ12] for the intrinsic diversity problem. Need to infer social utility from such conflicting, individual feedback. User Interest Pref Ranking Company a 1, a 2, a 3,... ML Gift Cards Social Opt Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

22 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Not social utility U. As in the case of prior work [RSJ12] for the intrinsic diversity problem. Need to infer social utility from such conflicting, individual feedback. User Interest Pref Ranking Company a 1, a 2, a 3,... ML b 1, b 2, b 3,... Gift Cards Social Opt Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

23 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Not social utility U. As in the case of prior work [RSJ12] for the intrinsic diversity problem. Need to infer social utility from such conflicting, individual feedback. User Interest Pref Ranking Company a 1, a 2, a 3,... ML b 1, b 2, b 3,... Gift Cards c 1, c 2, c 3,... Social Opt Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

24 Challenge of egoistic feedback Challenge: Learn from egoistic, weak, noisy user feedback. User i s feedback reflects them behaving as per personal utility U i. Not social utility U. As in the case of prior work [RSJ12] for the intrinsic diversity problem. Need to infer social utility from such conflicting, individual feedback. User Interest Pref Ranking Company a 1, a 2, a 3,... ML b 1, b 2, b 3,... Gift Cards c 1, c 2, c 3,... Social Opt a 1, b 1, c 1,... Even if social optimal is presented, users may indicate preferences for other rankings. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

25 Related work [CG98, ZCL03, CK06]: Address Extrinsic Diversity. Do not use learning. [YJ08, SMO10, RJS11]: Use learning for diversity. Require relevance labels for all user-document pairs. [RKJ08]: Uses online learning: Array of (decoupled) MA Bandits. Learns very slowly. Does not generalize across queries. [SRG13]: Couples the arms together. Does not generalize across queries. Hard-coded notion of diversity. [YG12]: Generalizes across queries. Requires cardinal utilities. [RSJ12]: Learns from user preferences. Requires all users directly optimize social utility U. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

26 Preferential Feedback What feedback do we obtain from users? Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

27 Preferential Feedback What feedback do we obtain from users? Implicit feedback (e.g. clicks) is timely and easily available. Presented (y) Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

28 Preferential Feedback What feedback do we obtain from users? Implicit feedback (e.g. clicks) is timely and easily available. Presented (y) User feedback does not reflect cardinal utilities. Shown in user studies [JGP + 07]. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

29 Preferential Feedback What feedback do we obtain from users? Implicit feedback (e.g. clicks) is timely and easily available. Presented (y) User feedback does not reflect cardinal utilities. Shown in user studies [JGP + 07]. KEY: Treat user feedback as preferences. Feedback (y ) Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

30 Preferential Feedback What feedback do we obtain from users? Implicit feedback (e.g. clicks) is timely and easily available. Presented (y) User feedback does not reflect cardinal utilities. Shown in user studies [JGP + 07]. KEY: Treat user feedback as preferences. How do we learn from such preferential feedback? Feedback (y ) Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

31 Learning from Preferences: Coactive Learning [SJ12, RJSS13] Learning model Repeat forever: System receives context x t. System makes prediction y t. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

32 Learning from Preferences: Coactive Learning [SJ12, RJSS13] Learning model Repeat forever: System receives context x t. System makes prediction y t. e.g. : Search Engine User Query Ranking Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

33 Learning from Preferences: Coactive Learning [SJ12, RJSS13] Learning model Repeat forever: System receives context x t. System makes prediction y t. Regret = Regret + U(xt, yt ) U(x t, y t ) e.g. : Search Engine User Query Ranking Social utility Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

34 Learning from Preferences: Coactive Learning [SJ12, RJSS13] Learning model Repeat forever: System receives context x t. System makes prediction y t. Regret = Regret + U(xt, y t ) U(x t, y t ) System gets (preference) feedback: U i (x t, ȳ t ) α,δ U i (x t, y t ) e.g. : Search Engine User Query Ranking Social utility Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

35 Learning from Preferences: Coactive Learning [SJ12, RJSS13] Learning model Repeat forever: System receives context x t. System makes prediction y t. Regret = Regret + U(xt, y t ) U(x t, y t ) System gets (preference) feedback: U i (x t, ȳ t ) α,δ U i (x t, y t ) e.g. : Search Engine User Query Ranking Social utility Feedback received in terms of personal utilities U i. But regret is in terms of social utility U. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

36 Learning from Preferences: Coactive Learning [SJ12, RJSS13] Learning model Repeat forever: System receives context x t. System makes prediction y t. Regret = Regret + U(xt, y t ) U(x t, y t ) System gets (preference) feedback: U i (x t, ȳ t ) α,δ U i (x t, y t ) e.g. : Search Engine User Query Ranking Social utility Feedback received in terms of personal utilities U i. But regret is in terms of social utility U. How does we model utilities? Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

37 Modeling User Utility: Submodularity Assume personal utilities are submodular. Diminishing returns: Marginal benefit of additional document on ML diminishes if 10 docs already shown vs only 1 previous doc. D1 D2 D4 D3 Computing ranking Submodular maximization Use simple, efficient greedy algorithm. Approximation guarantee of 1 2 (under partition matroid constraint). Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

38 Modeling User Utility: Submodularity Assume personal utilities are submodular. Diminishing returns: Marginal benefit of additional document on ML diminishes if 10 docs already shown vs only 1 previous doc. D1 D2 D4 D3 Computing ranking Submodular maximization Use simple, efficient greedy algorithm. Approximation guarantee of 1 2 (under partition matroid constraint). How does this lead to diversity? Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

39 Diversity via Submodularity: An example Posn Doc machine learning metal silver MAX of Col Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Marginal Benefit Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

40 Diversity via Submodularity: An example Posn Doc machine learning metal silver MAX of Col Doc Marginal Benefit d 1 3*5 + 3*7 36 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

41 Diversity via Submodularity: An example Posn Doc machine learning metal silver MAX of Col Doc Marginal Benefit d 1 3*5 + 3*7 36 d 2 5*5 + 2*7 39 d 3 2*5 + 5*7 45 d 4 2*5 + 3*7 31 d 5 3*4 + 5*6 42 d 6 6*4 + 2*6 36 d 7 1*5 + 4*4 + 2*6 33 d 8 1*5 + 3*4 + 1*6 23 Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

42 Diversity via Submodularity: An example Posn Doc machine learning metal silver 1 d MAX of Col Doc Marginal Benefit d 1 3*5 + 3*7 36 d 2 5*5 + 2*7 39 d 3 2*5 + 5*7 45 d 4 2*5 + 3*7 31 d 5 3*4 + 5*6 42 d 6 6*4 + 2*6 36 d 7 1*5 + 4*4 + 2*6 33 d 8 1*5 + 3*4 + 1*6 23 Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

43 Diversity via Submodularity: An example Posn Doc machine learning metal silver 1 d MAX of Col Doc Marginal Benefit d 1 (3-2)*5 5 d 2 (5-2)*5 15 d d d 5 3*4 + 5*6 42 d 6 6*4 + 2*6 36 d 7 4*4 + 2*6 28 d 8 3*4 + 1*6 18 Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

44 Diversity via Submodularity: An example Posn Doc machine learning metal silver 1 d d MAX of Col Doc Marginal Benefit d 1 (3-2)*5 5 d 2 (5-2)*5 15 d d d 5 3*4 + 5*6 42 d 6 6*4 + 2*6 36 d 7 4*4 + 2*6 28 d 8 3*4 + 1*6 18 Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

45 Diversity via Submodularity: An example Posn Doc machine learning metal silver 1 d d d MAX of Col Doc Marginal Benefit d 1 (3-2)*5 5 d 2 (5-2)*5 15 d d d d 6 (6-3)*4 12 d 7 (4-3)*4 4 d Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

46 Diversity via Submodularity: An example Posn Doc machine learning metal silver 1 d d d d MAX of Col Doc Marginal Benefit d d d d d d 6 (6-3)*4 12 d d Doc d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 Words ma:3 le:3 ma:5 le:2 ma:2 le:5 ma:2 le:3 me:3 si:5 me:6 si:2 me:4 si:2 ma:1 me:3 si:1 ma:1 Word Weight machine 5 learning 7 metal 4 silver 6 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

47 More General Submodular Utility Can we use other submodular functions? Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

48 More General Submodular Utility Can we use other submodular functions? Yes. Given ranking/set y = (d i1,..., d in ), aggregate features as: φ j F (x, y) = F (γ 1φ j (x, d i1 ), γ 2 φ j (x, d i2 ),......, γ n φ j (x, d in )) φ j (x, d i ) is j th feature of d i. F is a submodular function (modeling decision). γ 1 γ 2... γ n 0 are position-discount factors Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

49 More General Submodular Utility Can we use other submodular functions? Yes. Given ranking/set y = (d i1,..., d in ), aggregate features as: φ j F (x, y) = F (γ 1φ j (x, d i1 ), γ 2 φ j (x, d i2 ),......, γ n φ j (x, d in )) φ j (x, d i ) is j th feature of d i. F is a submodular function (modeling decision). γ 1 γ 2... γ n 0 are position-discount factors Utility modeled as linear in aggregate features: U(x, y) = w T φ F (x, y) Submodular aggregation leads to diversity. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

50 Social Perceptron for Ranking 1 Initialize weight vector w Given context x t compute y t argmax y w t φ(x t, y). 3 Observe user clicks D. 4 Construct preference feedback ȳ t PrefFeedback(y t, D). 5 w t+1 w t + φ(x t, ȳ t ) φ(x t, y t ) 6 Clip: w j t+1 max( wj t+1, 0) 7 Repeat from step 2. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

51 Social Perceptron for Ranking 1 Initialize weight vector w Given context x t compute y t argmax y wt φ(x t, y). Using greedy algorithm. 3 Observe user clicks D. 4 Construct preference feedback ȳ t PrefFeedback(y t, D). 5 w t+1 w t + φ(x t, ȳ t ) φ(x t, y t ) 6 Clip: w j t+1 max( wj t+1, 0) 7 Repeat from step 2. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

52 Social Perceptron for Ranking 1 Initialize weight vector w Given context x t compute y t argmax y wt φ(x t, y). Using greedy algorithm. 3 Observe user clicks D. 4 Construct preference feedback ȳ t PrefFeedback(y t, D). Pairwise feedback. 5 w t+1 w t + φ(x t, ȳ t ) φ(x t, y t ) 6 Clip: w j t+1 max( wj t+1, 0) 7 Repeat from step 2. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

53 Social Perceptron for Ranking 1 Initialize weight vector w Given context x t compute y t argmax y w t φ(x t, y). Using greedy algorithm. 3 Observe user clicks D. 4 Construct preference feedback ȳ t PrefFeedback(y t, D). Pairwise feedback. 5 w t+1 w t + φ(x t, ȳ t ) φ(x t, y t ) Perceptron update. 6 Clip: w j t+1 max( wj t+1, 0) To ensure submodularity. 7 Repeat from step 2. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

54 Regret Bound Definition User feedback is expected α i, δ i -informative if ξ t R is chosen s.t. : ( ) Eȳt [U i (x t, ȳ t )] (1 + δ i )U i (x t, y t ) + α i U i (x t, yt,i ) U i (x t, y t ) ξ t. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

55 Regret Bound Definition User feedback is expected α i, δ i -informative if ξ t R is chosen s.t. : ( ) Eȳt [U i (x t, ȳ t )] (1 + δ i )U i (x t, y t ) + α i U i (x t, yt,i ) U i (x t, y t ) ξ t. Theorem For any w R m and φ(x, y) l2 R the average regret of the SoPer-R algorithm can be upper bounded as: REG T 1 T 1 E i [p i ξ t ] + R w 15R w + ηt 2η η 2T. with: δ i t=0 ( Γ F 1 p ) i, η = min p i α i. p i i Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

56 Regret Bound: Analysis Theorem Average regret of the SoPer-R algorithm can be upper bounded as: REG T 1 T 1 E i [p i ξ t ] + R w 15R w + ηt 2η η 2T. with: δ i t=0 ( Γ F 1 p ) i, η = min p i α i. p i i Does not depend on number of dimensions. Only on feature ball radius R. Decays gracefully with noisy feedback (the α i s and η). Need not converge to optimal. Partly due to NP-hardness of submodular maximization. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

57 Social Perceptron for Sets SoPer-S Algorithm for predicting diverse sets. See paper for more details. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

58 Experimental Setup Used standard TREC 6-8 Interactive search-diversification dataset: Each query has 7-56 user types. Setup as in previous work [BJYB11, RJS11]. Simulated user behavior: Users scan rankings top to bottom. Click on first document relevant to them (with small error chance). Utility function: Normalized DCG-coverage function (i.e. F (x 1,..., x n ) = max γ i x i ) upto rank 5. i Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

59 Learning to Diversify: Single Query 0.88 Normalized List Utility Random Ranked Bandit SoPer-R Number of Iterations Per Query Improved learning for single-query diversification. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

60 Learning to Diversify: Cross-Query 0.68 Normalized List Utility Random StructPerc SoPer-R Number of Iterations Per Query StructPerc is (rough) skyline: Uses optimal for training. First method to learn cross-query diversity from implicit feedback. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

61 Robustness User Fncn SoPer-R Function Rand MAX SQRT MAX.630 ± ± ±.006 SQRT.656 ± ± ±.007 Robust to difference between submodular functions used in User s utility and Algorithm s utility. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

62 Robustness User Fncn SoPer-R Function Rand MAX SQRT MAX.630 ± ± ±.006 SQRT.656 ± ± ±.007 Robust to difference between submodular functions used in User s utility and Algorithm s utility. Random No Noise Noise.557 ± ± ±.007 Works even if user feedback is noisy Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

63 Conclusions Proposed online-learning algorithms for aggregating conflicting user preferences of a diverse population. Utilizes the coactive learning model. Modeled user utilities as submodular. Provided regret bounds for algorithms. Works well empirically and is robust. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

64 THANKS QUESTIONS? Poster #15 Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

65 C. Brandt, T. Joachims, Yisong Yue, and J. Bank. Dynamic ranked retrieval. In WSDM, Jaime Carbonell and Jade Goldstein. The use of MMR, diversity-based reranking for reordering documents and producing summaries. In SIGIR, pages , Harr Chen and David R. Karger. Less is more: probabilistic models for retrieving fewer relevant documents. In SIGIR, pages , T. Joachims, L. Granka, Bing Pan, H. Hembrooke, F. Radlinski, and G. Gay. Evaluating the accuracy of implicit feedback from clicks and query reformulations in web search. TOIS, 25(2), April Karthik Raman, Thorsten Joachims, and Pannagadatta Shivaswamy. Structured learning of two-level dynamic rankings. In CIKM, pages , Karthik Raman, Thorsten Joachims, Pannaga Shivaswamy, and Tobias Schnabel. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

66 Stable coactive learning via perturbation. In ICML, pages , Filip Radlinski, Robert Kleinberg, and Thorsten Joachims. Learning diverse rankings with multi-armed bandits. In ICML, pages , Karthik Raman, Pannaga Shivaswamy, and Thorsten Joachims. Online learning to diversify from implicit feedback. In KDD, pages , Pannaga Shivaswamy and Thorsten Joachims. Online structured prediction via coactive learning. In ICML, Rodrygo L.T. Santos, Craig Macdonald, and Iadh Ounis. Selectively diversifying web search results. In CIKM, pages , Aleksandrs Slivkins, Filip Radlinski, and Sreenivas Gollapudi. Ranked bandits in metric spaces: learning optimally diverse rankings over large document collections. JMLR, 14: , Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

67 Yisong Yue and Carlos Guestrin. Linear submodular bandits and their application to diversified retrieval. In NIPS, pages , Yisong Yue and Thorsten Joachims. Predicting diverse subsets using structural SVMs. In ICML, pages , Cheng Xiang Zhai, William W. Cohen, and John Lafferty. Beyond independent relevance: methods and evaluation metrics for subtopic retrieval. In SIGIR, pages 10 17, Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

68 Experimental Setup Details TREC 6-8 Interactive diversification dataset: Contains 17 queries. Each has 7-56 user types. Binary relevance labels. Similar results observed for WEB diversification dataset. Setup details: Re-ranking documents relevant to 1 user. Probability of user type # of documents relevant to user. DCG-position discounting: γ i = 1 log 2 (1 + i). Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20

69 Regret Bound Definition User feedback is expected α i, δ i -informative for user with personal utility function U i, if ξ t R is chosen s.t. for some given α i [0, 1] and δ i > 0: ( ) Eȳt [U i (x t, ȳ t )] (1 + δ i )U i (x t, y t ) + α i U i (x t, yt,i ) U i (x t, y t ) ξ t. Theorem For any w R m and φ(x, y) l2 R the average regret of the SoPer-R algorithm can be upper bounded as: REG T 1 T 1 E i [p i ξ t ] + βr w 2 4 β + 2 R w ηt η η. T with: δ i t=0 ( Γ F 1 p ) i, η = min p i α i and β = (1 β gr ) = 1 p i i 2. Karthik Raman (Cornell University) Learning Socially Optimal Systems Sept 23, / 20