Calculations and Motivating forces for Hearty Positioning

Presentation Transcript

Calculations and Incentives for Robust Ranking Rajat Bhattacharjee Ashish Goel Stanford University Algorithms and impetuses for vigorous positioning . ACM-SIAM Symposium on Discrete Algorithms (SODA), 2007. Motivating force based positioning components . EC Workshop, Economics of Networked Systems, 2006.

Outline Motivation Model Incentive Structure Ranking Algorithm Algorithms and impetuses for strong positioning

Traditional Content era was brought together (book distributers, motion picture creation organizations, daily papers) Content dissemination was liable to publication control (paid experts: commentators, editors) Internet Content era is for the most part decentralized (people make pages, writes) No focal article control on substance dispersion (rather there are positioning and reco. frameworks like google, hurray) Content : then and now Algorithms and motivations for vigorous positioning

Heuristics Race PageRank (utilizes interface structure of the web) Spammers attempt to diversion the framework by making fake connection structures Heuristics race: internet searchers and spammers have executed progressively modern heuristics to neutralize each other New procedures to counter the heuristics [Gyongyi, Garcia-Molina] Detecting PageRank opening up structures  sparsest cut issue (NP-hard) [Zhang et al.] Algorithms and impetuses for hearty positioning

Amplification Ratio [Zhang, Goel, … ] Consider a set S , which is a subset of V In(S) : add up to weight of edges from V-S to S Local(S) : add up to weight of edges from S to S 10 w(S) = Local(S) + In(S) Amp(S) = w(S)/In(S) High Amp(S) → S is unscrupulous Low Amp(S) → S is straightforward Collusion free diagram: where all sets are straightforward S Algorithms and motivators for powerful positioning

Heuristics Race Then why do web indexes work so well? Our conviction: since heuristics are not openly space Is this "the arrangement"? Input/click investigation [Anupam et al.] [Metwally et al.] Suffers from snap spam Problem of elements with little criticism Too numerous site pages, can't put them on top openings to assemble input Algorithms and motivating forces for hearty positioning

Ranking inversion Ranking inversion Entity An is superior to element B, yet B is positioned higher than A Keyword: Search Engine Algorithms and motivators for powerful positioning

Our outcome Theorem we would have jumped at the chance to demonstrate Here is a notoriety framework and it is vigorous, i.e., has no positioning inversions even within the sight of pernicious conduct Theorem we demonstrate Here is a positioning calculation and impetus structure, which when connected together suggest an arbitrage open door for the clients of the framework at whatever point there is a positioning inversion (even within the sight of malevolent conduct) Algorithms and motivating forces for strong positioning

Where is the cash? Illustrations Amazon.com: better suggestions → more buys → more income Netflix: better proposals → expanded consumer loyalty → expanded enrollment → more income Google/Yahoo: better positioning → more eyeballs → more income through promotions Revenue per substance Simple for Amazon.com and Netflix For Google/Yahoo, we can circulate the income from a client on the website pages he takes a gander at (different methodologies conceivable) Algorithms and motivating forces for powerful positioning

My valuable Why share? Since they will take it in any case!!! Calculations and motivations for hearty positioning

Less convincing reasons Difficulty of evoking legitimate criticism is outstanding [Resnick et al.] [Dellarocas] Search motor rankings are self-strengthening [Cho, Roy] Strong motivator for players to diversion the framework Ballot stuffing and abusing in notoriety frameworks [Bhattacharjee, Goel] [Dellarocas] Click spam in web rankings in light of snaps [Anupam et al.] Web structures have been concocted to amusement PageRank [Gyongyi, Garcia-Molina] Problem of new elements How ought to the framework find high caliber, new substances in the framework? By what method ought to the framework find a website page whose significance has all of a sudden changed (might be because of some present occasion)? Calculations and impetuses for hearty positioning

Outline Motivation Model Incentive Structure Ranking Algorithm Algorithms and motivators for powerful positioning

I-U Model Inspect (I) User peruses a piece joined to an output (Google/Yahoo) Looks at a suggestion for a book (Amazon.com) Utilize (U) User goes to the real site page (Google/Yahoo) Buys the book (Amazon.com) Algorithms and motivators for strong positioning

I-U Model Entities Web pages (Google/Yahoo), Books (Amazon.com) Each element i has an innate quality q i (consider it the likelihood that a client would use element i , molded on the way that the substance was reviewed by the client) The qualities q i are obscure, however we wish to rank elements as per their qualities Feedback Tokens (positive and negative) set on an element by clients Ranking is an element of the relative number of tokens got by elements Slots Placeholders for the consequences of an inquiry Algorithms and motivations for vigorous positioning

Sheep and Connoisseurs Sheep can value a fantastic element when appeared But wouldn't go searching for a top notch element Most clients are sheep Connoisseurs will burrow for an excellent element which is not positioned sufficiently high The objective of this plan is to total the data that the authorities have Algorithms and motivations for hearty positioning

User reaction Algorithms and motivations for hearty positioning

I-U Model User reaction to a run of the mill question Chooses to investigate the top j positions User picks j at irregular from an obscure yet altered conveyance Utility era occasion for e i happens if the client uses an element e i (expecting e i is set among the top j openings) Formally Utility era occasion is caught by arbitrary variable G i = I r(i) U i r(i) : rank of element e I r(i) ,U i : autonomous Bernoulli arbitrary factors E[U i ] = q i (obscure) E[I 1 ] ≥ E[I 2 ] ≥ … ≥ E[I k ] (known) Algorithms and motivations for strong positioning

Outline Motivation Model Incentive Structure Ranking Algorithm Algorithms and motivators for powerful positioning

Information Markets View the issue as an information accumulation issue Float shares of elements and let the market choose their esteem (positioning) [Hanson] [Pennock] Rank as per the value set by the market Work best to predict results which are target Elections (Iowa electronic market) Distinguishing elements of the positioning issue Fundamental issue: result is not target Revenue: due to more eyeballs or better quality? Eyeballs thus rely on upon the value set by the market However, an extra lever: the positioning calculation Algorithms and motivating forces for powerful positioning

Game theoretic methodologies Example: [Miller et al.] Framework to manipulate legit criticism Counter absence of target results by contrasting a client's audits with that of his companions Selfish premiums of a client ought to be in accordance with the alluring properties of the framework Doesn't address vindictive clients Benefits from the framework, may originate from outside the framework too Revenue from result of these frameworks may overpower the income from the framework itself Algorithms and motivators for vigorous positioning

Ranking component: diagram Overview: Users put token (positive and negative) on the substances Ranking is figured in view of the quantity of tokens on the elements Whenever an income era occasion happens, the income is shared among the clients Ranking calculation Input: criticism scores of elements Output: probabilistic appropriation over rankings of the elements Ensures that the quantity of examinations an element gets is corresponding to the division of tokens on it Algorithms and impetuses for strong positioning

Incentive structure A token is a three tuple: (p, u, e) p : +1 or - 1 relying upon whether a token is a positive token or a negative token u : client who set the token e : element on which the token was put Net weight of the tokens a client can place is limited, that is | p i | is limited User can't continue putting positive tokens without setting a negative token and the other way around Algorithms and motivating forces for hearty positioning

User account Each client has a record Revenue shares are included or deducted from a client's record Withdrawal is allowed yet stores are not Users can make benefits from the framework but rather not pick up control by paying If a client's share goes negative: expel it from the framework for some pre-characterized time Let  <1 and s>1 be pre-characterized framework parameters The part of income that the framework conveys as impetuses to the clients:  Parameter s will be set later Algorithms and impetuses for strong positioning

8 7 6 5 4 3 2 1 Revenue share Suppose an income era occasion happens for an element e at time t R : income produced For every token i put on element e an i is the net weight (positive - negative) of tokens set on element e before token i was put on e The income imparted by the framework to the client who set token i is corresponding to p i  R/an i s Adds up to at most  R Negative token: the income share is negative, deduct from the client's record Algorithms and motivations for strong positioning

Revenue share Some elements Parameter s controls relative significance of tokens put before Tokens put after token i make little difference to the income share of the client who set token i Hence s is entirely more prominent than 1 Incentive for disclosure of amazing elements Hence the decision of lessening prizes Emphasis is on making the procedure as verifiable as could be expected under the circumstances Resistance to hustling The framework shouldn't permit a rehashed cycle of activities wh