Mechanisms for Immunizing Online Reputation Systems Against Unfair Rater Behavior


Having recognized the problem of unfair ratings as a real and important one, this section proposes a number of mechanisms for eliminating or significantly reducing its adverse effects on the reliability of online reputation systems.

Avoiding Negative Unfair Ratings Using Controlled Anonymity

The main argument of this section is that the anonymity regime of an online community can influence the kinds of reputation system attacks that are possible. A slightly surprising result is the realization that a fully transparent marketplace, where everybody knows everybody else's true identity, incurs more dangers of reputation system fraud than a marketplace where the true identities of traders are carefully concealed from each other, but are known to a trusted third entity (usually the market-maker).

Bad mouthing and negative discrimination are based on the ability to pick a few specific "victims" and give them unfairly poor ratings or provide them with poor service respectively. Usually, victims are selected based on some real-life attributes of their associated principal entities (for example, because they are our competitors or because of religious or racial prejudices). This adverse selection process can be avoided if the community conceals the true identities of the buyers and sellers from each other.

In such a "controlled anonymity" scheme, the marketplace knows the true identity of all market participants by applying some effective authentication process before it allows access to any agent (Hutt, Bosworth, & Hoyt, 1995). In addition, it keeps track of all transactions and ratings. The marketplace publishes the estimated reputation of buyers and sellers, but keeps their identities concealed from each other (or assigns them pseudonyms that change from one transaction to the next, in order to make identity detection very difficult). In that way, buyers and sellers make their decisions solely based on the offered terms of trade as well as the published reputations. Because they can no longer identify their "victims," bad mouthing and negative discrimination can be avoided.

It is interesting to observe that, while, in most cases, the anonymity of online communities has been viewed as a source of additional risks (Kollock, 1999; Friedman & Resnick, 2001), here we have an example of a situation where some controlled degree of anonymity can be used to eliminate some transaction risks.

Concealing the identities of buyers and sellers is not possible in all domains. For example, concealing the identity of sellers is not possible in restaurant and hotel ratings (although concealing the identity of buyers is). In other domains, it may require the creative intervention of the marketplace. For example, in a marketplace of electronic component distributors, it may require the marketplace to act as an intermediary shipping hub that will help erase information about the seller's address.

If concealing the identities of both parties from each other is not possible, then it may still be useful to conceal the identity of one party only. More specifically, concealing the identity of buyers but not sellers avoids negative discrimination against buyers but does not avoid bad mouthing of sellers. In an analogous manner, concealing the identity of sellers but not buyers avoids bad-mouthing but not negative discrimination. These results are summarized in Figure 3.

Anonymity Regime

Classes of possible unfair behavior

Buyer's identity known to seller

Seller's identity known to buyer

Bad-mouthing possible

Negative discrimination possible

Ballot-stuffing possible

Positive discrimination possible

Yes

Yes

Yes

No

No

Yes

No

No


Figure 3: Effectiveness of Controlled Anonymity in Preventing Certain Classes of Unfair Behavior

Generally speaking, concealing the identities of buyers is usually easier than concealing the identities of sellers (a similar point is made in Cranor & Resnick, 1999). This means that negative discrimination is easier to avoid than bad mouthing. Further-more, concealing the identities of sellers before a service is performed is usually easier than afterwards. In domains with this property, controlled anonymity can be used at the seller selection stage in order to protect sellers from being intentionally picked for subsequent bad mouthing. For example, in the above-mentioned marketplace of electronic component distributors, one could conceal the identities of sellers until after the closing of a deal. Assuming that the number of distributors for a given component type is relatively large, this strategy would make it difficult for malevolent buyers to intentionally pick specific distributors for subsequent bad mouthing.

It is important to note at this point that even when identities of buyers and sellers are concealed, buyers and sellers who have an incentive to signal their identities to each other can always find clever ways to do so. For example, sellers involved in a ballot-stuffing scheme can use a particular pattern in the amounts that they bid (e.g., amounts ending in .33) in order to signal their presence to their conspirators. Therefore, while controlled anonymity can avoid bad mouthing and negative discrimination, it cannot avoid ballot stuffing and positive discrimination. The following two sections propose some filtering mechanisms, which are applicable in the cases of ballot stuffing as well.

Reducing the Effect of Unfair Ratings Using Median Filtering

In the second section of this chapter, I based the calculation of reputation bias on the assumption that MREs are based on the sample mean of the nearest-neighbor set. In this section I will demonstrate that the effect of unfair ratings can be significantly reduced if, instead of the sample mean, the calculation of MREs is based on the sample median.

The field of robust statistics has devoted considerable attention to the problem of finding estimators of "location" (mean value), which are robust in the presence of contaminated samples (Huber, 1981). Nevertheless, most of that literature treats contamination as "innocent" noise and does not address the problem of malicious raters who, based on their knowledge of the estimator used, strategically distribute unfair ratings in order to maximize the achievable bias. To the knowledge of the author, the analysis presented in this section is novel.

  • Definition: The sample median of n ordered observations Y1 Y2 ... Yn is the middle observation Yk where k = (n+1)/2 if n is odd. When n is even, then is considered to be any value between the two middle observations Yk and Yk+1 where k = n/2, although it is most often taken to be their average.

In the absence of unfair ratings (i.e., when δ = 0), I previously assumed that . It is well known (Hojo, 1931) that as the size n of the sample increases, the median of a sample drawn from a normal distribution converges rapidly to a normal distribution with a mean equal to the median of the parent distribution. In normal distributions, the median is equal to the mean. Therefore, in situations where there are no unfair raters, the use of the sample median results in unbiased fair MREs:

(6)

Let us now assume that unfair raters know that MREs are based on the sample median. They will strategically try to introduce unfair ratings whose values will maximize the absolute bias between the sample median of the fair set and the sample median of the contaminated set. More specifically, "ballot stuffers" will try to maximize that bias while "bad mouthers" will try to minimize it. In the following analysis I consider the case of ballot stuffing. The case of bad mouthing is symmetric, with the signs reversed.

  • Proposition 1: Assume that the nearest neighbor set consists of nf = (1 - δ) n fair ratings and nu=δ n unfair ratings, where 0 δ < 0.5 and n are sufficiently large. If MREs are based on the sample median and fair ratings are drawn from a normal distribution with standard deviation s, then the maximum MRE bias achievable by a strategic "ballot-stuffer" is asymptotically equal to:

    (7) click to expand

    where φ-1(q) is the inverse standard normal CDF.

  • Proof: See Appendix.

Figure 4 shows some of the values of E[Bmax]for various values of δ and σ in the special case where ratings range from 0 to 9. The maximum bias increases with the percentage of unfair ratings and is directly proportional to the standard deviation of the fair ratings. As before, I have highlighted maximum average biases of 5% of the rating range or more. Figure 4 shows that the use of the sample median as a basis of calculating MREs manages to reduce the maximum average bias to below 5% of the rating range for unfair rater ratios of up to 30% to 40% and a wide range of fair rating standard deviations.

Percentage of unfair ratings

Standard Deviation of Fair Ratings

0.25

0.50

0.75

1.00

Reputation Bias

9%

0.03

0.06

0.09

0.13

18%

0.07

0.14

0.21

0.28

27%

0.12

0.24

0.37

0.49

36%

0.20

0.40

0.59

0.79

45%

0.35

0.69

1.04

1.38


Figure 4: Asymptotic Upper Bounds of Average Reputation Bias when MREs are Based on the Median of the Ratings Set (Ratings Range from 0–9); Highlighted Cells Indicate Biases Above 5% of the Ratings Range

Using Frequency Filtering to Eliminate Unfair Ratings Flooding

Equations (5) and (7) confirm the intuitive fact that the reputation bias due to unfair ratings increases with the ratio δ of unfair raters in a given sample. In settings where a seller's quality attributes may vary over time, calculation of reputation should be based on recent ratings only using time discounting or a time-window approach. In those cases, as demonstrated earlier, by "flooding" the system with ratings, a relatively small number of unfair raters can manage to increase the ratio δ of unfair ratings in any given time window above 50% and completely compromise the reliability of the system.

This section proposes an approach for immunizing a reputation system against unfair ratings flooding. The main idea is to filter raters in the nearest-neighbor set based on their ratings submission frequency.

Description of Frequency Filtering

Step 1: Frequency filtering depends on estimating the average frequency of ratings submitted by each buyer for a given seller. Since this frequency is a time-varying quantity (sellers can become more or less popular with the passage of time), it too needs to be estimated using a time-window approach. More specifically:

  1. Calculate the set Fs(t) of buyer-specific average rating submission frequencies for each buyer b that has submitted ratings for seller s during the rating submission frequency calculation time window Wf = [t-E, t]. More precisely:

    click to expand

  2. Set the cutoff frequency to be equal to the kth order statistic of the set Fs(t) where k= (1- D)n, n is the number of elements of Fs(t), and D is a conservative estimate of the fraction of unfair raters in the total buyer population for seller s. For example, if we assume that there are no more than 10% unfair raters among all the buyers for seller s, then D = 0.1. Assuming further that n = 100, i.e., that the set Fs(t) contains average rating submission frequencies from 100 buyers, then the cutoff frequency would be equal to the 90th smallest frequency (the 10th largest frequency) present in the set Fs(t).

    The width E of the ratings submission frequency calculation time window Wf should be large enough to contain at least a few ratings from all buyers for a given seller.

Step 2: During the calculation of an MRE for seller s, eliminate all raters b in the nearest-neighbor set for whom In other words, eliminate all buyers whose average ratings submission frequency for seller s is above the cutoff frequency.

Analysis of Frequency Filtering

Frequency filtering provides effective protection against unfair ratings flooding by guaranteeing that the ratio of unfair raters in the MRE calculation set cannot be more than twice as large as the ratio of unfair raters in the total buyer population.

As before, I will assume that the entire buyer population is n, unfair raters are δ n<<n and the width of the reputation estimation time window is a relatively small W (so that, each rating within W typically comes from a different rater). The following proposition then holds:

  • Proposition 2: Assume that the frequency of fair ratings is uniformly distributed. Then, after applying frequency filtering to the nearest-neighbor set of raters, the ratio of unfair raters δ in the total population of buyers and the ratio δ' of unfair ratings remaining in the nearest-neighbor set satisfy the inequality:

    (8)

  • Proof: See Appendix.

Equation (8) shows that, no matter how hard unfair raters may try to "flood" the system with ratings, the presence of frequency filtering guarantees that they cannot inflate their presence in the final MRE calculation set by more than a factor of two.

In most online communities, the exact ratio δ of unfair raters will not be known exactly. In such cases, if we have a belief that δ < 0.1, then setting D= 0.1 has been experimentally proven to result in inflation ratios, which also fall within the bounds of equation (8).

A more realistic assumption about fair ratings frequencies is that they follow a lognormal distribution. This assumption is consistent with the findings of researchers in marketing (Lawrence, 1980). In this case, the expression for the final ratio δ' cannot be given in closed form. However, a numerical solution yields results, which approximate very closely those obtained analytically for uniformly distributed fair rating frequencies (Figure 5).

click to expand
Figure 5: Maximum Unfair Ratings Inflation Factors Achievable Through Flooding when Frequency Filtering is Used (δ = D = 0.1); Frequency Spread Indicates the Difference Between the Maximum and Minimum Rating Submission Frequencies

One possible criticism of the frequency filtering approach is that it potentially eliminates those fair buyers who transact most frequently with a given seller. In fact, in the absence of unfair raters, all raters who would be filtered out based on their high ratings submission frequency would be fair raters. Nevertheless, I do not believe that this property constitutes a weakness of the approach. I argue that the "best customers" of a given seller often receive preferential treatment, which is in a way a form of positive discrimination on behalf of the seller. Therefore, I believe that the potential elimination of such raters from the final reputation estimate in fact benefits the construction of more unbiased estimates for the benefit of first-time prospective buyers.

Issues in Communities Where Buyer Identity is Not Authenticated

The effectiveness of frequency filtering relies on the assumption that a given principal can only have one buyer agent acting on its behalf in a given marketplace. The technique is also valid in situations where principals have multiple buyer agents with authenticated identifiers. In that case, frequency filtering works if we consider all agents of a given principal as a single buyer for frequency filtering purposes.

In non-authenticated online communities (communities where "pseudonyms" are "cheap," to use the term of Friedman and Resnick) with time-windowed reputation estimation, unfair buyers can still manage to "flood" the system with unfair ratings by creating a large number of pseudonymously known buyer agents acting on their behalf. In that case the total ratio d of unfair agents relative to the entire buyer population can be made arbitrarily high. If each of the unfair agents takes care of submitting unfair ratings for seller s with frequency , because δ will be high, even in the presence of frequency filtering, unfair raters can still manage to severely contaminate a seller's fair reputation.

Further research is needed to develop immunization techniques that are effective in communities where the "true" identity of buyer agents cannot be authenticated. In the meantime, the observations of this section make a strong argument for using some reasonably effective authentication regime for buyers in online communities where trust is based on online reputation. For example, the community can require that all newly registering buyers supply a valid credit card for authentication purposes, or it can insert cookies into buyer computers so that attempts to assume different "false" identities from the same computer fail.




Social and Economic Transformation in the Digital Era
Social and Economic Transformation in the Digital Era
ISBN: 1591402670
EAN: 2147483647
Year: 2003
Pages: 198

flylib.com © 2008-2017.
If you may any questions please contact us: flylib@qtcs.net