Learning Socially Optimal Information Systems from Egoistic Users
|
|
- Daniel Moody
- 5 years ago
- Views:
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
University of Delaware at Diversity Task of Web Track 2010
University of Delaware at Diversity Task of Web Track 2010 Wei Zheng 1, Xuanhui Wang 2, and Hui Fang 1 1 Department of ECE, University of Delaware 2 Yahoo! Abstract We report our systems and experiments
More informationStructured Learning of Two-Level Dynamic Rankings
Structured earning of Two-evel Dynamic Rankings Karthik Raman Dept. of Computer Science Cornell University Ithaca NY 4850 USA karthik@cs.cornell.edu Thorsten Joachims Dept. of Computer Science Cornell
More informationLearning with Humans in the Loop
Learning with Humans in the Loop CS4780/5780 Machine Learning Fall 2014 Thorsten Joachims Cornell University Optional Reading: - Yisong Yue, J. Broder, R. Kleinberg, T. Joachims, The K-armed Dueling Bandits
More informationLearning to Rank (part 2)
Learning to Rank (part 2) NESCAI 2008 Tutorial Filip Radlinski Cornell University Recap of Part 1 1. Learning to rank is widely used for information retrieval, and by web search engines. 2. There are many
More informationEfficient Diversification of Web Search Results
Efficient Diversification of Web Search Results G. Capannini, F. M. Nardini, R. Perego, and F. Silvestri ISTI-CNR, Pisa, Italy Laboratory Web Search Results Diversification Query: Vinci, what is the user
More informationLearning Ranking Functions with Implicit Feedback
Learning Ranking Functions with Implicit Feedback CS4780 Machine Learning Fall 2011 Pannaga Shivaswamy Cornell University These slides are built on an earlier set of slides by Prof. Joachims. Current Search
More informationA Study of Pattern-based Subtopic Discovery and Integration in the Web Track
A Study of Pattern-based Subtopic Discovery and Integration in the Web Track Wei Zheng and Hui Fang Department of ECE, University of Delaware Abstract We report our systems and experiments in the diversity
More information5. Novelty & Diversity
5. Novelty & Diversity Outline 5.1. Why Novelty & Diversity? 5.2. Probability Ranking Principled Revisited 5.3. Implicit Diversification 5.4. Explicit Diversification 5.5. Evaluating Novelty & Diversity
More informationLearning Ranking Functions with SVMs
Learning Ranking Functions with SVMs CS4780/5780 Machine Learning Fall 2012 Thorsten Joachims Cornell University T. Joachims, Optimizing Search Engines Using Clickthrough Data, Proceedings of the ACM Conference
More informationLarge-Scale Validation and Analysis of Interleaved Search Evaluation
Large-Scale Validation and Analysis of Interleaved Search Evaluation Olivier Chapelle, Thorsten Joachims Filip Radlinski, Yisong Yue Department of Computer Science Cornell University Decide between two
More informationLearning Dense Models of Query Similarity from User Click Logs
Learning Dense Models of Query Similarity from User Click Logs Fabio De Bona, Stefan Riezler*, Keith Hall, Massi Ciaramita, Amac Herdagdelen, Maria Holmqvist Google Research, Zürich *Dept. of Computational
More informationNTU Approaches to Subtopic Mining and Document Ranking at NTCIR-9 Intent Task
NTU Approaches to Subtopic Mining and Document Ranking at NTCIR-9 Intent Task Chieh-Jen Wang, Yung-Wei Lin, *Ming-Feng Tsai and Hsin-Hsi Chen Department of Computer Science and Information Engineering,
More informationInternational Journal of Computer Engineering and Applications, Volume IX, Issue X, Oct. 15 ISSN
DIVERSIFIED DATASET EXPLORATION BASED ON USEFULNESS SCORE Geetanjali Mohite 1, Prof. Gauri Rao 2 1 Student, Department of Computer Engineering, B.V.D.U.C.O.E, Pune, Maharashtra, India 2 Associate Professor,
More informationEvaluating the Accuracy of. Implicit feedback. from Clicks and Query Reformulations in Web Search. Learning with Humans in the Loop
Evaluating the Accuracy of Implicit Feedback from Clicks and Query Reformulations in Web Search Thorsten Joachims, Filip Radlinski, Geri Gay, Laura Granka, Helene Hembrooke, Bing Pang Department of Computer
More informationScalable Diversified Ranking on Large Graphs
211 11th IEEE International Conference on Data Mining Scalable Diversified Ranking on Large Graphs Rong-Hua Li Jeffrey Xu Yu The Chinese University of Hong ong, Hong ong {rhli,yu}@se.cuhk.edu.hk Abstract
More informationPart 5: Structured Support Vector Machines
Part 5: Structured Support Vector Machines Sebastian Nowozin and Christoph H. Lampert Providence, 21st June 2012 1 / 34 Problem (Loss-Minimizing Parameter Learning) Let d(x, y) be the (unknown) true data
More informationLearning Optimal Subsets with Implicit User Preferences
Learning Optimal Subsets with Implicit User Preferences Yunsong Guo and Carla Gomes Department of Computer Science Cornell University {guoys, gomes}@cs.cornell.edu Abstract We study the problem of learning
More informationAn Analysis of NP-Completeness in Novelty and Diversity Ranking
An Analysis of NP-Completeness in Novelty and Diversity Ranking Ben Carterette (carteret@cis.udel.edu) Dept. of Computer and Info. Sciences, University of Delaware, Newark, DE, USA Abstract. A useful ability
More informationAn Investigation of Basic Retrieval Models for the Dynamic Domain Task
An Investigation of Basic Retrieval Models for the Dynamic Domain Task Razieh Rahimi and Grace Hui Yang Department of Computer Science, Georgetown University rr1042@georgetown.edu, huiyang@cs.georgetown.edu
More informationMinimally Invasive Randomization for Collecting Unbiased Preferences from Clickthrough Logs
Minimally Invasive Randomization for Collecting Unbiased Preferences from Clickthrough Logs Filip Radlinski and Thorsten Joachims Department of Computer Science Cornell University, Ithaca, NY {filip,tj}@cs.cornell.edu
More informationCoverage-based search result diversification
Inf Retrieval (2012) 15:433 457 DOI 10.1007/s10791-011-9178-4 Coverage-based search result diversification Wei Zheng Xuanhui Wang Hui Fang Hong Cheng Received: 18 April 2011 / Accepted: 28 October 2011
More informationFast Convergence of Regularized Learning in Games
Fast Convergence of Regularized Learning in Games Vasilis Syrgkanis Alekh Agarwal Haipeng Luo Robert Schapire Microsoft Research NYC Microsoft Research NYC Princeton University Microsoft Research NYC Strategic
More informationLearning to Rank. Tie-Yan Liu. Microsoft Research Asia CCIR 2011, Jinan,
Learning to Rank Tie-Yan Liu Microsoft Research Asia CCIR 2011, Jinan, 2011.10 History of Web Search Search engines powered by link analysis Traditional text retrieval engines 2011/10/22 Tie-Yan Liu @
More informationCS 6740: Advanced Language Technologies April 2, Lecturer: Lillian Lee Scribes: Navin Sivakumar, Lakshmi Ganesh, Taiyang Chen.
CS 6740: Advanced Language Technologies April 2, 2010 Lecture 15: Implicit Relevance Feedback & Clickthrough Data Lecturer: Lillian Lee Scribes: Navin Sivakumar, Lakshmi Ganesh, Taiyang Chen Abstract Explicit
More informationA Survey: Static and Dynamic Ranking
A Survey: Static and Dynamic Ranking Aditi Sharma Amity University Noida, U.P. India Nishtha Adhao Amity University Noida, U.P. India Anju Mishra Amity University Noida, U.P. India ABSTRACT The search
More informationLecture 3: Improving Ranking with
Modeling User Behavior and Interactions Lecture 3: Improving Ranking with Behavior Data 1 * &!" ) & $ & 6+ $ & 6+ + "& & 8 > " + & 2 Lecture 3 Plan $ ( ' # $ #! & $ #% #"! #! -( #", # + 4 / 0 3 21 0 /
More informationModeling Contextual Factors of Click Rates
Modeling Contextual Factors of Click Rates Hila Becker Columbia University 500 W. 120th Street New York, NY 10027 Christopher Meek Microsoft Research 1 Microsoft Way Redmond, WA 98052 David Maxwell Chickering
More informationSubmodularity in Speech/NLP
Submodularity in Speech/NLP Jeffrey A. Bilmes Professor Departments of Electrical Engineering & Computer Science and Engineering University of Washington, Seattle http://melodi.ee.washington.edu/~bilmes
More informationModern Retrieval Evaluations. Hongning Wang
Modern Retrieval Evaluations Hongning Wang CS@UVa What we have known about IR evaluations Three key elements for IR evaluation A document collection A test suite of information needs A set of relevance
More informationCSE 546 Machine Learning, Autumn 2013 Homework 2
CSE 546 Machine Learning, Autumn 2013 Homework 2 Due: Monday, October 28, beginning of class 1 Boosting [30 Points] We learned about boosting in lecture and the topic is covered in Murphy 16.4. On page
More informationInteractive Exploratory Search for Multi Page Search Results
Interactive Exploratory Search for Multi Page Search Results Xiaoran Jin and Marc Sloan and Jun Wang Department of Computer Science University College London ucabxij@ucl.ac.uk, M.Sloan, J.Wang}@cs.ucl.ac.uk
More informationA Multiple-Play Bandit Algorithm Applied to Recommender Systems
Proceedings of the Twenty-Eighth International Florida Artificial Intelligence Research Society Conference A Multiple-Play Bandit Algorithm Applied to Recommender Systems Jonathan Louëdec Institut de Mathématiques
More informationMicrosoft Research Asia at the NTCIR-9 Intent Task
Microsoft Research Asia at the NTCIR-9 Intent Task Jialong Han 1,4, Qinglei Wang 2, Naoki Orii 3, Zhicheng Dou 4, Tetsuya Sakai 4, and Ruihua Song 4 4 Microsoft Research Asia; 1 Renmin University of China;
More informationA Class of Submodular Functions for Document Summarization
A Class of Submodular Functions for Document Summarization Hui Lin, Jeff Bilmes University of Washington, Seattle Dept. of Electrical Engineering June 20, 2011 Lin and Bilmes Submodular Summarization June
More informationNoisy Submodular Maximization via Adaptive Sampling with Applications to Crowdsourced Image Collection Summarization
Noisy Submodular Maximization via Adaptive Sampling with Applications to Crowdsourced Image Collection Summarization Adish Singla ETH Zurich adish.singla@inf.ethz.ch Sebastian Tschiatschek ETH Zurich sebastian.tschiatschek@inf.ethz.ch
More informationMicrosoft Research Asia at the Web Track of TREC 2009
Microsoft Research Asia at the Web Track of TREC 2009 Zhicheng Dou, Kun Chen, Ruihua Song, Yunxiao Ma, Shuming Shi, and Ji-Rong Wen Microsoft Research Asia, Xi an Jiongtong University {zhichdou, rsong,
More informationECS289: Scalable Machine Learning
ECS289: Scalable Machine Learning Cho-Jui Hsieh UC Davis Sept 22, 2016 Course Information Website: http://www.stat.ucdavis.edu/~chohsieh/teaching/ ECS289G_Fall2016/main.html My office: Mathematical Sciences
More informationPromoting Ranking Diversity for Biomedical Information Retrieval based on LDA
Promoting Ranking Diversity for Biomedical Information Retrieval based on LDA Yan Chen, Xiaoshi Yin, Zhoujun Li, Xiaohua Hu and Jimmy Huang State Key Laboratory of Software Development Environment, Beihang
More informationRobust Relevance-Based Language Models
Robust Relevance-Based Language Models Xiaoyan Li Department of Computer Science, Mount Holyoke College 50 College Street, South Hadley, MA 01075, USA Email: xli@mtholyoke.edu ABSTRACT We propose a new
More informationPart 5: Structured Support Vector Machines
Part 5: Structured Support Vector Machines Sebastian Noozin and Christoph H. Lampert Colorado Springs, 25th June 2011 1 / 56 Problem (Loss-Minimizing Parameter Learning) Let d(x, y) be the (unknon) true
More informationImproving Aggregated Search Coherence
Improving Aggregated Search Coherence Jaime Arguello University of North Carolina at Chapel Hill jarguello@unc.edu Abstract. Aggregated search is that task of blending results from different search services,
More informationSubmodular Optimization in Computational Sustainability. Andreas Krause
Submodular Optimization in Computational Sustainability Andreas Krause Master Class at CompSust 2012 Combinatorial optimization in computational sustainability Many applications in computational sustainability
More informationLearning to Rank. from heuristics to theoretic approaches. Hongning Wang
Learning to Rank from heuristics to theoretic approaches Hongning Wang Congratulations Job Offer from Bing Core Ranking team Design the ranking module for Bing.com CS 6501: Information Retrieval 2 How
More informationExploiting the Diversity of User Preferences for Recommendation
Exploiting the Diversity of User Preferences for Recommendation Saúl Vargas and Pablo Castells Escuela Politécnica Superior Universidad Autónoma de Madrid 28049 Madrid, Spain {saul.vargas,pablo.castells}@uam.es
More informationMarkov Decision Processes (MDPs) (cont.)
Markov Decision Processes (MDPs) (cont.) Machine Learning 070/578 Carlos Guestrin Carnegie Mellon University November 29 th, 2007 Markov Decision Process (MDP) Representation State space: Joint state x
More informationSupport Vector Machine Learning for Interdependent and Structured Output Spaces
Support Vector Machine Learning for Interdependent and Structured Output Spaces I. Tsochantaridis, T. Hofmann, T. Joachims, and Y. Altun, ICML, 2004. And also I. Tsochantaridis, T. Joachims, T. Hofmann,
More informationRandomized Composable Core-sets for Distributed Optimization Vahab Mirrokni
Randomized Composable Core-sets for Distributed Optimization Vahab Mirrokni Mainly based on joint work with: Algorithms Research Group, Google Research, New York Hossein Bateni, Aditya Bhaskara, Hossein
More informationThe Offset Tree for Learning with Partial Labels
The Offset Tree for Learning with Partial Labels Alina Beygelzimer IBM Research John Langford Yahoo! Research June 30, 2009 KDD 2009 1 A user with some hidden interests make a query on Yahoo. 2 Yahoo chooses
More informationLetter Pair Similarity Classification and URL Ranking Based on Feedback Approach
Letter Pair Similarity Classification and URL Ranking Based on Feedback Approach P.T.Shijili 1 P.G Student, Department of CSE, Dr.Nallini Institute of Engineering & Technology, Dharapuram, Tamilnadu, India
More informationCombine the PA Algorithm with a Proximal Classifier
Combine the Passive and Aggressive Algorithm with a Proximal Classifier Yuh-Jye Lee Joint work with Y.-C. Tseng Dept. of Computer Science & Information Engineering TaiwanTech. Dept. of Statistics@NCKU
More informationRisk Minimization and Language Modeling in Text Retrieval Thesis Summary
Risk Minimization and Language Modeling in Text Retrieval Thesis Summary ChengXiang Zhai Language Technologies Institute School of Computer Science Carnegie Mellon University July 21, 2002 Abstract This
More informationarxiv: v2 [cs.ai] 1 Dec 2015
Noisy Submodular Maximization via Adaptive Sampling with Applications to Crowdsourced Image Collection Summarization Adish Singla ETH Zurich adish.singla@inf.ethz.ch Sebastian Tschiatschek ETH Zurich sebastian.tschiatschek@inf.ethz.ch
More informationKernels and representation
Kernels and representation Corso di AA, anno 2017/18, Padova Fabio Aiolli 20 Dicembre 2017 Fabio Aiolli Kernels and representation 20 Dicembre 2017 1 / 19 (Hierarchical) Representation Learning Hierarchical
More informationThe 1st International Workshop on Diversity in Document Retrieval
WORKSHOP REPORT The 1st International Workshop on Diversity in Document Retrieval Craig Macdonald School of Computing Science University of Glasgow UK craig.macdonald@glasgow.ac.uk Jun Wang Department
More informationEvaluating Search Engines by Modeling the Relationship Between Relevance and Clicks
Evaluating Search Engines by Modeling the Relationship Between Relevance and Clicks Ben Carterette Center for Intelligent Information Retrieval University of Massachusetts Amherst Amherst, MA 01003 carteret@cs.umass.edu
More informationmnir: Diversifying Search Results based on a Mixture of Novelty, Intention and Relevance
mnir: Diversifying Search Results based on a Mixture of Novelty, Intention and Relevance Reza Taghizadeh Hemayati, Laleh Jafarian Dehordi, Weiyi Meng Department of Computer Science, Binghamton University
More informationOnline Learning of Assignments
Online Learning of Assignments Matthew Streeter Google, Inc. Pittsburgh, PA 15213 mstreeter@google.com Daniel Golovin Carnegie Mellon University Pittsburgh, PA 15213 dgolovin@cs.cmu.edu Andreas Krause
More informationLearning optimally diverse rankings over large document collections
Aleksandrs Slivkins Microsoft Research, 1065 La Avenida, Mountain View CA, USA Filip Radlinski Microsoft Research, 7 J.J. Thomson Ave., Cambridge UK Sreenivas Gollapudi Microsoft Research, 1065 La Avenida,
More informationCombining Implicit and Explicit Topic Representations for Result Diversification
Combining Implicit and Explicit Topic Representations for Result Diversification Jiyin He J.He@cwi.nl Vera Hollink V.Hollink@cwi.nl Centrum Wiskunde en Informatica Science Park 123, 1098XG Amsterdam, the
More informationTREC OpenSearch Planning Session
TREC OpenSearch Planning Session Anne Schuth, Krisztian Balog TREC OpenSearch OpenSearch is a new evaluation paradigm for IR. The experimentation platform is an existing search engine. Researchers have
More informationEvaluating search engines CE-324: Modern Information Retrieval Sharif University of Technology
Evaluating search engines CE-324: Modern Information Retrieval Sharif University of Technology M. Soleymani Fall 2016 Most slides have been adapted from: Profs. Manning, Nayak & Raghavan (CS-276, Stanford)
More informationVisual Recognition: Examples of Graphical Models
Visual Recognition: Examples of Graphical Models Raquel Urtasun TTI Chicago March 6, 2012 Raquel Urtasun (TTI-C) Visual Recognition March 6, 2012 1 / 64 Graphical models Applications Representation Inference
More information1 Overview. 2 Applications of submodular maximization. AM 221: Advanced Optimization Spring 2016
AM : Advanced Optimization Spring 06 Prof. Yaron Singer Lecture 0 April th Overview Last time we saw the problem of Combinatorial Auctions and framed it as a submodular maximization problem under a partition
More informationA Dynamic Bayesian Network Click Model for Web Search Ranking
A Dynamic Bayesian Network Click Model for Web Search Ranking Olivier Chapelle and Anne Ya Zhang Apr 22, 2009 18th International World Wide Web Conference Introduction Motivation Clicks provide valuable
More informationNon-exhaustive, Overlapping k-means
Non-exhaustive, Overlapping k-means J. J. Whang, I. S. Dhilon, and D. F. Gleich Teresa Lebair University of Maryland, Baltimore County October 29th, 2015 Teresa Lebair UMBC 1/38 Outline Introduction NEO-K-Means
More informationMore Data, Less Work: Runtime as a decreasing function of data set size. Nati Srebro. Toyota Technological Institute Chicago
More Data, Less Work: Runtime as a decreasing function of data set size Nati Srebro Toyota Technological Institute Chicago Outline we are here SVM speculations, other problems Clustering wild speculations,
More informationSearch Result Diversification
Foundations and Trends R in Information Retrieval Vol. 9, No. 1 (2015) 1 90 c 2015 R. L. T. Santos, C. Macdonald and I. Ounis DOI: 10.1561/1500000040 Search Result Diversification Rodrygo L. T. Santos
More informationarxiv: v1 [cs.ir] 11 Oct 2018
Offline Comparison of Ranking Functions using Randomized Data Aman Agarwal, Xuanhui Wang, Cheng Li, Michael Bendersky, Marc Najork Google Mountain View, CA {agaman,xuanhui,chgli,bemike,najork}@google.com
More informationSupport Vector Machines and their Applications
Purushottam Kar Department of Computer Science and Engineering, Indian Institute of Technology Kanpur. Summer School on Expert Systems And Their Applications, Indian Institute of Information Technology
More informationComparisons of Sequence Labeling Algorithms and Extensions
Nam Nguyen Yunsong Guo Department of Computer Science, Cornell University, Ithaca, NY 14853, USA NHNGUYEN@CS.CORNELL.EDU GUOYS@CS.CORNELL.EDU Abstract In this paper, we survey the current state-ofart models
More informationDiversification of Query Interpretations and Search Results
Diversification of Query Interpretations and Search Results Advanced Methods of IR Elena Demidova Materials used in the slides: Charles L.A. Clarke, Maheedhar Kolla, Gordon V. Cormack, Olga Vechtomova,
More informationECS289: Scalable Machine Learning
ECS289: Scalable Machine Learning Cho-Jui Hsieh UC Davis Oct 4, 2016 Outline Multi-core v.s. multi-processor Parallel Gradient Descent Parallel Stochastic Gradient Parallel Coordinate Descent Parallel
More informationProteins, Particles, and Pseudo-Max- Marginals: A Submodular Approach Jason Pacheco Erik Sudderth
Proteins, Particles, and Pseudo-Max- Marginals: A Submodular Approach Jason Pacheco Erik Sudderth Department of Computer Science Brown University, Providence RI Protein Side Chain Prediction Estimate side
More informationClickthrough Log Analysis by Collaborative Ranking
Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence (AAAI-10) Clickthrough Log Analysis by Collaborative Ranking Bin Cao 1, Dou Shen 2, Kuansan Wang 3, Qiang Yang 1 1 Hong Kong
More informationOptimal Separating Hyperplane and the Support Vector Machine. Volker Tresp Summer 2018
Optimal Separating Hyperplane and the Support Vector Machine Volker Tresp Summer 2018 1 (Vapnik s) Optimal Separating Hyperplane Let s consider a linear classifier with y i { 1, 1} If classes are linearly
More informationAdaptive Search Engines Learning Ranking Functions with SVMs
Adaptive Search Engines Learning Ranking Functions with SVMs CS478/578 Machine Learning Fall 24 Thorsten Joachims Cornell University T. Joachims, Optimizing Search Engines Using Clickthrough Data, Proceedings
More informationDiversity Maximization Under Matroid Constraints
Diversity Maximization Under Matroid Constraints Zeinab Abbassi Department of Computer Science Columbia University zeinab@cs.olumbia.edu Vahab S. Mirrokni Google Research, New York mirrokni@google.com
More informationDistributed Submodular Maximization in Massive Datasets. Alina Ene. Joint work with Rafael Barbosa, Huy L. Nguyen, Justin Ward
Distributed Submodular Maximization in Massive Datasets Alina Ene Joint work with Rafael Barbosa, Huy L. Nguyen, Justin Ward Combinatorial Optimization Given A set of objects V A function f on subsets
More informationUsing Historical Click Data to Increase Interleaving Sensitivity
Using Historical Click Data to Increase Interleaving Sensitivity Eugene Kharitonov, Craig Macdonald, Pavel Serdyukov, Iadh Ounis Yandex, Moscow, Russia School of Computing Science, University of Glasgow,
More informationINTERSOCIAL: Unleashing the Power of Social Networks for Regional SMEs
INTERSOCIAL I1 1.2, Subsidy Contract No. , MIS Nr 902010 European Territorial Cooperation Programme Greece-Italy 2007-2013 INTERSOCIAL: Unleashing the Power of Social Networks for Regional SMEs
More informationLearning CRFs for Image Parsing with Adaptive Subgradient Descent
Learning CRFs for Image Parsing with Adaptive Subgradient Descent Honghui Zhang Jingdong Wang Ping Tan Jinglu Wang Long Quan The Hong Kong University of Science and Technology Microsoft Research National
More informationAddressing the Challenges of Underspecification in Web Search. Michael Welch
Addressing the Challenges of Underspecification in Web Search Michael Welch mjwelch@cs.ucla.edu Why study Web search?!! Search engines have enormous reach!! Nearly 1 billion queries globally each day!!
More informationSemi-supervised learning and active learning
Semi-supervised learning and active learning Le Song Machine Learning II: Advanced Topics CSE 8803ML, Spring 2012 Combining classifiers Ensemble learning: a machine learning paradigm where multiple learners
More informationWebSci and Learning to Rank for IR
WebSci and Learning to Rank for IR Ernesto Diaz-Aviles L3S Research Center. Hannover, Germany diaz@l3s.de Ernesto Diaz-Aviles www.l3s.de 1/16 Motivation: Information Explosion Ernesto Diaz-Aviles
More informationA Survey on Postive and Unlabelled Learning
A Survey on Postive and Unlabelled Learning Gang Li Computer & Information Sciences University of Delaware ligang@udel.edu Abstract In this paper we survey the main algorithms used in positive and unlabeled
More informationMining the Search Trails of Surfing Crowds: Identifying Relevant Websites from User Activity Data
Mining the Search Trails of Surfing Crowds: Identifying Relevant Websites from User Activity Data Misha Bilenko and Ryen White presented by Matt Richardson Microsoft Research Search = Modeling User Behavior
More informationCoverage Approximation Algorithms
DATA MINING LECTURE 12 Coverage Approximation Algorithms Example Promotion campaign on a social network We have a social network as a graph. People are more likely to buy a product if they have a friend
More informationEstimating Credibility of User Clicks with Mouse Movement and Eye-tracking Information
Estimating Credibility of User Clicks with Mouse Movement and Eye-tracking Information Jiaxin Mao, Yiqun Liu, Min Zhang, Shaoping Ma Department of Computer Science and Technology, Tsinghua University Background
More informationFast Learning of Document Ranking Functions with the Committee Perceptron
Carnegie Mellon University Research Showcase @ CMU Computer Science Department School of Computer Science -008 Fast Learning of Document Ranking Functions with the Committee Perceptron Jonathan L. Elsas
More informationRepresentative & Informative Query Selection for Learning to Rank using Submodular Functions
Representative & Informative Query Selection for Learning to Rank using Submodular Functions Rishabh Mehrotra Dept of Computer Science University College London, UK r.mehrotra@cs.ucl.ac.uk Emine Yilmaz
More informationLearning Ranking Functions with SVMs
Learning Ranking Functions with SVMs CS4780/5780 Machine Learning Fall 2014 Thorsten Joachims Cornell University T. Joachims, Optimizing Search Engines Using Clickthrough Data, Proceedings of the ACM Conference
More informationAdvanced Topics in Information Retrieval. Learning to Rank. ATIR July 14, 2016
Advanced Topics in Information Retrieval Learning to Rank Vinay Setty vsetty@mpi-inf.mpg.de Jannik Strötgen jannik.stroetgen@mpi-inf.mpg.de ATIR July 14, 2016 Before we start oral exams July 28, the full
More informationLecture 9: Support Vector Machines
Lecture 9: Support Vector Machines William Webber (william@williamwebber.com) COMP90042, 2014, Semester 1, Lecture 8 What we ll learn in this lecture Support Vector Machines (SVMs) a highly robust and
More informationImproving Difficult Queries by Leveraging Clusters in Term Graph
Improving Difficult Queries by Leveraging Clusters in Term Graph Rajul Anand and Alexander Kotov Department of Computer Science, Wayne State University, Detroit MI 48226, USA {rajulanand,kotov}@wayne.edu
More informationLearning Hierarchies at Two-class Complexity
Learning Hierarchies at Two-class Complexity Sandor Szedmak ss03v@ecs.soton.ac.uk Craig Saunders cjs@ecs.soton.ac.uk John Shawe-Taylor jst@ecs.soton.ac.uk ISIS Group, Electronics and Computer Science University
More informationICTNET at Web Track 2010 Diversity Task
ICTNET at Web Track 2010 Diversity Task Yuanhai Xue 1,2, Zeying Peng 1,2, Xiaoming Yu 1, Yue Liu 1, Hongbo Xu 1, Xueqi Cheng 1 1. Institute of Computing Technology, Chinese Academy of Sciences, Beijing,
More informationEffective Latent Space Graph-based Re-ranking Model with Global Consistency
Effective Latent Space Graph-based Re-ranking Model with Global Consistency Feb. 12, 2009 1 Outline Introduction Related work Methodology Graph-based re-ranking model Learning a latent space graph A case
More informationarxiv: v2 [cs.cv] 1 Oct 2013
Structured learning of sum-of-submodular higher order energy functions Alexander Fix Thorsten Joachims Sam Park Ramin Zabih arxiv:309.752v2 [cs.cv] Oct 203 Abstract Submodular functions can be exactly
More informationOptimizing Search Engines using Click-through Data
Optimizing Search Engines using Click-through Data By Sameep - 100050003 Rahee - 100050028 Anil - 100050082 1 Overview Web Search Engines : Creating a good information retrieval system Previous Approaches
More informationCOMS 4771 Support Vector Machines. Nakul Verma
COMS 4771 Support Vector Machines Nakul Verma Last time Decision boundaries for classification Linear decision boundary (linear classification) The Perceptron algorithm Mistake bound for the perceptron
More information