Recommendation Systems and Web Search

1. Recommendation SystemsandWeb Search

2. Recommendation systems • eBay • Buyers and sellers rate each transaction • Users check the ratings before each transaction • Amazon • System monitors user purchases • System recommends products based on this history • SpamNet (a collaborative filter) • SpamNet filters mail, users identify mistakes • Users gain reputations • Reputations affect weight of recommendations

3. The problem • There are many known attacks • Many villains boost each other’s reputation • One villain acts honest for a while • How can honest users collaborate effectively … in the presence of dishonest, malicious users • What is a good model for recommendation systems? • What are algorithms can we prove something about? • We give fast, robust, practical algorithms.

4. A recommendation model • n players, both honest and dishonest (n honest) • dishonest players can be arbitrarily malicious, collude • m objects, both good and bad (m good) • players probe objects to learn whether good or bad • there is a cost to probing a bad object • a public billboard • players post the results of their probes (the honest do…) • billboard is free: no cost to read or write • We think this is a direct abstraction of eBay

5. A simple game • At each step of the game, one player takes a turn: • consult the billboard • probe an object • post the result on the billboard • Honest players follow the protocol • Dishonest players can collude in arbitrary ways • Goal: help honest users find good objects at minimal cost.

6. Bad ways to play the game • Always try the object with the highest number of positive recommendations • dishonest players recommend bad objects • honest players try all bad objects first • Always try the object with the least number of negative recommendations • dishonest players slander good objects • honest players try all bad objects first • (a popular strategy on eBay, but quite vulnerable) • Simple combinations also fail

7. What about other approaches? • Collaborative filtering [Goldberg, Nichols, Oki, Terry 1998], [Azar, Fiat, Karlin, McSherry, Saia 2001], [Drineas, Kerenidis, Raghavan 2002], [Kleinberg, Sandler 2003] • all players are honest, solutions are centralized • Web-searching algorithms [Brin, Page 1998], [Kleinberg 1999] • compute a “transitive popular vote” of the participants • easily spammed by a clique of colluding crooks • Trust [Kamvar, Schlosser, Garcia-Molina 2003] • use trusted authorities to assign trust to players • do we really care about trust? we care about cost! Simple randomized algorithms can minimize cost.

8. Our results • A simple, efficient algorithm for many contexts • Objects come and go over time • Players access different subsets of objects • Players have different tastes for objects • The most interesting model • Our algorithm can be substantially better than others • Players probe few objects • Only a constant when most players are honest • Corporate web search is the perfect application • Implemented on HP’s internal search engine

9. Good ways to play the game • Exploration rule: • “choose a random object, and probe it” • okay if most objects are good • Exploitation rule: • “choose a random player, and probe an object it liked” • okay if most players are honest • But are there many good objects/honest players?!?

10. The balanced rule • The balanced rule: flip a coin • if heads, follow exploration rule • if tails, follow exploitation rule • How well does it work? • We compute expected cost of probe sequence : cost() = number of bad objects probed by honest players

11. Split  each time a good player finds a good object •  = … a b c de d e d e e … D0 D1 D2 • expected cost of D0 = • each probe hits good object with probability • expected cost of Di = • each probe hits good object with probability • total expected cost is

12. Understanding that number Spreading the good news Finding a good object Probes by a player to learn the good news This is just a log factor away from optimal (lower bounds later)

13. Different models • Dynamic object model • objects come and go over time • Partial access model • players access overlapping subsets of the objects • Taste model • players have different notions of good and bad • In each model, analysis has similar feel…

14. Dynamic object model • Objects come and go with each step: • first the objects change (changes are announced to all) • then some player takes a step • Algorithm for player p: if p has found a good object o then p probes o else p follows the Balanced Rule

15. Competitive analysis • Optimality cost depends on the environment: • player and object schedules, honest players, good objects • adversary à powerful, adaptive, and Byzantine! • Optimal probe sequence opt has simple form: • opt = a a ax x xc c c c d d d e e e … good object no good objects • switches(opt) = number of distinct objects in opt • Our probe sequence  partitioned by the switches: •  = o o o o o o o o o o o o o o o o …

16. Competitive analysis • Compare cost of  and opt switch-by-switch cost() · cost(opt) + switches(opt) (cost to satisfy) · cost(opt) + switches(opt) (1/ + n log n) · cost(opt) + switches(opt) (m + n log n) • Lower bound (Yao’s Lemma): cost() ¸ cost(opt) + switches(opt) (m)

17. P O Partial access model • Each player can access only some objects w p x a q y b good object r bad object z • common interest essential for help from neighbors • q gets help from p, but not r • r gets no useful help at all • hard to measure help from neighbors accurately • we bound work of players P with common interest O

18. Similar analysis: exploration+exploitation work until one in P finds O work until all in P learn of O = + Partial access algorithm • Similar algorithm: • choose randomly from neighbors, not all players ¿ m + n log n • Good analysis in this model is very difficult: • we know common interest is essential for others to help • sometimes others don’t help (players might as well sample)even if each player has common interest with many players

19. Simulation 10,000 players, 50% honest10,000 objects, 1% good 1,000 objects/player steps exploration probability • Balanced rule tolerates coin bias quite well • Can we optimize the rule by changing the bias? • many good objects  emphasize exploration • many honest players  emphasize exploitation

20. Differing tastes model • Player tastes differ: good(player) µ objects • Player tastes overlap: good(p1) Å good(p2) ; • Special interest group S=(P,O): “anything in O could satisfy anyone in P”O µ good(p) 8 p 2 P • Players don’t know which SIG they are in. • Players follow the Balanced Rule.

21. Complexity • Theorem: Given a SIG (P,O), the expected total work to satisfy everyone in P is • Pretty good: within a log factor when |P|=O(n)

22. A very popular model • Most algorithms are off-line: prepare in advance • Assume underlying object set structure: The web • PageRank [Google], HITS [Kleinberg] • Assume underlying stochastic model: Analyze past choices • [KRRT’98], [KS’03], [AFKMS’01] • Heuristically identify similar users/objects • GroupLens Project • Drineas, Kerenidis, Raghavan: first on-line algorithm • Algorithm recommends, observers user response

23. DKR’02 • Assumes: • User preferences given by types (up to noise) • O(1) dominant types cover most users • Must know the number of dominant types • Type orthogonality: • No two types like the same object • Type gap: • The dominant types are by far the most popular • The algorithm: • Uses SVD to compute the types, learn user types • Runs in time O(mn)

24. The Balanced Rule • No types: SIGs a much looser characterization • No type orthogonality: SIGs can overlap, even “subtyping” is allowed • No type gap: SIG popularity irrelevant • A simple distributed algorithm • Tolerates malicious players • Shares work evenly if players run at similar rates • Fast: runs in time O(m + n log n)

25. A look at individual work • Consider a synchronous, round-based model • Each player probes an object each round until satisfied • One model of players running at similar speeds • A good model for studying individual work • Theorem: Lower bounds • Theorem: Balanced rule halts in

26. How is constant even possible? honest players, 1 good object all objects Contains the good object probe, probe n objects, n honest probesgood object gets a vote Expect the good objectat most bad objects all objects with 1 vote probe, probe Expect the good objectat most 2 bad objectsso probe them all all objects with votes p n + 1 objects, n honest probesgood object getspn votes probe, probe, probe

27. Candidate sets work well • Theorem: “Distilling candidate sets” is faster: • Constant 1/ rounds if n1- honest players • Even with many dishonest players, expected rounds is only • Remember: Balanced rule was

28. Conclusions • Contributions • a new model for the study of reputation • simple algorithms for collaboration in spite of • asynchronous behavior • changing objects, differing access, differing tastes • arbitrarily malicious behavior from players • Our work generalizes to multiple values • players want a “good enough” object • reduces to our binary values, even for multiple thresholds • players want “the best” or “in the top 10%” • early stopping is no longer possible

29. Future work • Better lower bounds, better analysis… • We never used • the identities of the malicious players • the negative votes on the billboard • the number of good objects or honest players • can we get rid of the log factors if we do? • What if the supply of each object is finite? • What if objects have prices? And the prices affect whether a player wants the object?

30. Many interesting problems remain… But now, an application to Intranet search…

31. xSearch: Improving Intranet search • Internet search engines are wonderful • Intranet search is disappointing • Some simple ideas to improve intranet search • Heuristic for deducing user recommendations • Algorithms for reordering search results • Implemented on the @HP search engine • Getting 10% of the queries • Collecting data on performance • Results are preliminary, but look promising…

32. Who is “lampman”? @hp says:

33. Who is “lampman”? We say:

34. What is going on? • Searching corporate webs should be so easy • Google is phenomenal on the external web • Why not just run Google on the internal web? • Because it doesn’t work, and not just at HP…

35. IBM vs Google et al (2003) • Identified the top 200 and median 150 queries • Found the “right” answer to each query by hand • Ran Google over the IBM internal web • Now how does Google do on these queries? • Popular queries: only 57% succeed • Median queries: only 46% succeed • Weak notion of success: “right answer in top 20 results” • Compare that to your normal Google experience! • “Right answer in top 5 results 85% of the time”

36. Link structure is crucial • Internet search algorithms depend on link structure: • Good pages point to great pages, etc. • Important pages are at the “center of the web” • Link structure has a strongly connected component: otherreachable incoming outgoing strongly connected component pages pages pages 30% of the Internet,10% of IBM The meaty part of the Intranet is just a third that of the Internet!

37. Corporate webs are different • More structure • Small groups build large parts of web (silos) • Periodically reviewed and branded • No outgoing links from many pages • Documents intended to be informative not “interesting” • Pages generated automatically from databases • No reward for creating/updating/advertising pages • Do you advertise your project page? • Do you advertise your track club page?

38. Corporate queries are different • Queries are focused and have few “right” answers • What is “vacation” looking for? • Inside HP: • what are the holidays and how to report vacation time • one or two pages on the HR web. • Outside HP: • interesting vacation spots or vacation deals • any number of answers pages would be satisfactory

39. What does @hp do now? • Based on the Verity Ultraseek search engine • Classical information retrieval on steroids • Many tunable parameters and knobs • Manually constructs “best bets” for popular queries • Works as well as any other intranet search engine

40. Our approach: user collaboration query query’ @hp user interface our reordering heuristics @hp search engine hits’ hits How to collect user feedback?How to use user feedback?

41. Collecting feedback: last clicks • The “last click” heuristic: • Assume the last link the user clicks is the right link • Assume users stop hunting when they find the right page • Easy, unobtrusive implementation: • Tag each query by a user with a session id • Log each link clicked on with the session id • Periodically scan the log to compute the “last clicks” • Effective heuristic: • Agreement: 44% of last clicks go to same page • Quality: 72% of last clicked pages are good pages

42. Last clicks example: “mis” “Mellon Investor Services” or “Manufacturing Information Systems” Surprisingly, the @hp top result was never the last click

43. Using feedback: ranking algorithms • Statistical ordering • Interpret last clicks as recommendation • Rank by popularity (most recommended first, etc) • Robust: highly stable in presence of spam • Move-to-front ordering • Each time a page is recommended (last-clicked),move it to the top of the list. • Fast: once the good page is found, it moves to the top (tsunami) • Probabilistic ordering • Choose pages one after another • Probability of selection = frequency of endorsement • Best of both worlds, and many theoretical results • Plus: all new pages have some chance of being seen!

44. An experimental implementation • Working with @hp search (Anne Murray Allen): • Ricky Ralston • Chris Jackson • Richard Boucher • Now part of the @hp search engine: • Getting 10% of queries sent to @hp search • A few months of experience shows promising results…

45. Example: “paycheck” at @hp three copies of slides on redeeming eawards a movie!! notice from 1999

46. Example: “paycheck” with move2front how to access(was #6) how to set up(was #7) May 2004 upgrade notice(was #41) 401(k) catchup(almost no votes)

47. Example: “active gold” at @hp hydra migration can’t read sports award can’t read dead link

48. Example: “active gold” with statistical home page(was #8)(10 of 12 votes)

49. Example: “payroll” at @hp

50. Example: “payroll” with statistical US main payroll(was #6) payroll forms(was #17) the best bet(was #16) payroll phone numbers(was #33)