Propagation of trust and distrust

R. GuhaRavi KumarPrabhakar RaghavanAndrew Tomkins

article2004WWW1,624 citations

Develops a matrix-based framework for propagating both positive trust and negative distrust across social networks, proving on large-scale real-world data that even sparse ratings can accurately predict interpersonal trust.

Listen

Online platforms and e-commerce systems face significant challenges in filtering low-quality information, malicious reviews, and fraudulent behavior. While reputation networks can help users assess the trustworthiness of others, real-world trust graphs are highly sparse, meaning most individuals rate only a small number of peers. The article develops and evaluates a computational framework designed to accurately predict pairwise trust and distrust scores across sparse social networks.

To test this framework, the authors conducted an empirical study using real-world data from the review platform Epinions, comprising approximately 841,000 trust and distrust relationships across 131,000 users. The evaluation tested 81 combinations of computational approaches using cross-validation over thousands of masked connections. The framework varied the propagation patterns (direct links, shared citations, reciprocation, and coupled trust), distrust handling mechanisms, numerical iteration strategies, and methods for converting continuous trust scores back into binary decisions.

Key findings show that pairwise trust can be inferred with high accuracy even from sparse network data. The best-performing approach reduced prediction error to 6.4% on random evaluation links and 14.7% on a balanced dataset of equal trust and distrust instances, outperforming the baseline error rate of 50%. Incorporating distrust was critical: treating distrust as propagating a single step—discounting the immediate judgments of distrusted parties rather than chaining distrust across multiple hops—yielded the strongest predictive power. Combining direct connections with indirect network patterns, such as co-citation, substantially outperformed relying on direct trust chains alone. Additionally, local neighborhood rounding methods proved decisive in accurately translating calculated scores into discrete ratings.

These findings indicate that digital marketplaces and community platforms can strengthen content integrity, reduce fraudulent manipulation, and improve personalized recommendations without requiring extensive user rating history. Contrary to common assumptions, modeling distrust requires different mathematical properties than positive trust, as transitive negative chains create distortion. Platforms that actively incorporate distrust signals can significantly improve safety and content filtering compared to systems relying solely on positive feedback.

Organizations implementing web-of-trust frameworks should deploy models that integrate multiple relationship structures, restrict distrust propagation to single-step discounting, and utilize local majority-based rounding. Further research and piloting should explore how these methods perform on other online network topologies, as well as test additive rather than multiplicative models for chaining distrust across larger communities.

IBM Verity Inc.e-commerce recommendation recommendation systems distrust propagation trust propagation trust network

Guha et al (2004).pdf

Paper: The Eigentrust algorithm for reputation management in P2P networks, Sepandar D. Kamvar et al. (2003). EigenTrust established iterative matrix-based propagation of global trust scores across sparse networks, providing the foundational mathematical framework that the source paper adapts and contrasts when handling distrust.
Paper: The link prediction problem for social networks, David Liben-Nowell et al. (2003). This work formalized link prediction and proximity metrics (including co-citation and multi-hop paths) in social graphs, which the source paper directly applies to infer pairwise trust and distrust.
Paper: SimRank: a measure of structural-context similarity, Glen Jeh et al. (2002). SimRank introduced structural context and co-citation similarity propagation over arbitrary graphs, directly informing the propagation patterns evaluated in the source paper.
Paper: Mining knowledge-sharing sites for viral marketing, Matthew Richardson et al. (2002). This study pioneered the extraction and linear modeling of trust networks using the Epinions platform data, establishing the primary empirical testbed used by the source paper.

Paper: Predicting positive and negative links in online social networks, Jure Leskovec et al. (2010). This paper extends the study of signed social networks by evaluating supervised learning over local structural balance and status theories to predict positive and negative edges on Epinions and other graphs.
Paper: A matrix factorization technique with trust propagation for recommendation in social networks, Mohsen Jamali et al. (2010). This research builds upon network-based trust propagation by integrating multi-step trust influence directly into matrix factorization models for recommendation.
Paper: Recommender systems with social regularization, Hao Ma et al. (2011). This work extends trust and social network modeling into collaborative filtering through explicit social regularization constraints on matrix factorization.
Paper: Graph Neural Networks for Social Recommendation, Wenqi Fan et al. (2019). GraphRec modernizes social trust modeling by utilizing graph neural networks to jointly capture user-user trust and user-item interactions on platforms like Epinions.

Cover for Propagation of trust and distrust

Abstract

A (directed) network of people connected by ratings or trust scores, and a model for propagating those trust scores, is a fundamental building block in many of today’s most successful e-commerce and recommendation systems. We develop a framework of trust propagation schemes, each of which may be appropriate in certain circumstances, and evaluate the schemes on a large trust network consisting of 800K trust scores expressed among 130K people. We show that a small number of expressed trusts/distrust per individual allows us to predict trust between any two people in the system with high accuracy. Our work appears to be the first to incorporate distrust in a computational trust propagation setting.

Topics

e-commerce recommendation recommendation systems trust relationship

Keywords

distrust propagation trust propagation trust network trust matrix propagation schemes reputation management recommendation systems semantic web

Categories and Subject Descriptors
General Terms
Keywords
1. INTRODUCTION
1.1 Approaches to trust propagation
1.2 Introducing distrust
1.3 Summary of results
2. RELATED WORK
3. ALGORITHMS
3.1 Atomic propagation
3.2 Propagation of trust and distrust
3.2.1 Propagation of distrust
3.2.2 Iterative propagation
3.2.3 Rounding
3.3 On the transitivity of distrust
4. EXPERIMENTAL DATA
4.1 Data source: Epinions
4.2 Trust graph characteristics
5. EXPERIMENTS
5.1 Results
5.1.1 Basis elements
5.1.2 Incorporation of distrust
5.1.3 Rounding
5.1.4 Iteration models
5.1.5 The effect of the number of iterations, K
6. CONCLUSIONS
7. ACKNOWLEDGMENTS
8. REFERENCES

Knowls

Knowl 1 — Atomic Trust Propagation Operators and Combined Matrix Formulation
model/method
In a trust network of $n$ users with trust matrix $T \in [0, 1]^{n \times n}$ and distrust matrix $D \in [0, 1]^{n \times n}$ , inferences can be performed on a belief matrix $B \in \mathbb{R}^{n \times n}$ (where $B$ represents trusts $T$ or combined trust-distrust $T - D$ ). Single-step inferences follow a basis set of four atomic propagation operators, each corresponding to forward and backward path operations on $B$ :
1. Direct Propagation ( $B$ ): If user $i$ trusts user $j$ , and $j$ trusts user $k$ , infer that $i$ trusts $k$ . Multiplying the initial belief matrix $B$ by operator $B$ yields $B \cdot B = B^2$ , which represents all length-2 directed paths.
2. Co-citation ( $B^T B$ ): If user $i_1$ trusts users $j_1$ and $j_2$ , and user $i_2$ trusts $j_2$ , infer that $i_2$ should also trust $j_1$ . Applying operator $B^T B$ yields $B(B^T B) = B B^T B$ , propagating beliefs backward from $j_2$ to $i_1$ and then forward to $j_1$ .
3. Transpose Trust ( $B^T$ ): If user $a$ trusts user $b$ , infer that trusting $b$ implies some level of trust toward $a$ .
4. Trust Coupling ( $B B^T$ ): If users $a$ and $b$ both trust user $c$ , infer that trusting $a$ implies trusting $b$ .
These atomic operators are unified into a single combined atomic propagation matrix $C_{B, \alpha}$ parametrized by a non-negative weight vector $\alpha = (\alpha_1, \alpha_2, \alpha_3, \alpha_4)$ :

$C_{B, \alpha} = \alpha_1 B + \alpha_2 B^T B + \alpha_3 B^T + \alpha_4 B B^T$

where the $(i, j)$ -th entry of $C_{B, \alpha}$ quantifies the inferred flow of belief from user $i$ to user $j$ in a single atomic step.
Knowl 2 — Models for Distrust Propagation
model/method
Given an initial trust matrix $T \in [0, 1]^{n \times n}$ and distrust matrix $D \in [0, 1]^{n \times n}$ , repeated propagation of beliefs over $k$ steps is modeled by powering a combined atomic propagation matrix $C_{B, \alpha} = \alpha_1 B + \alpha_2 B^T B + \alpha_3 B^T + \alpha_4 B B^T$ . The belief matrix $B$ and the $k$ -step propagation matrix $P^{(k)}$ are defined under three distinct models of distrust:
1. Trust Only: Distrust is completely ignored, propagating only positive trust ratings:
$B = T, \quad P^{(k)} = C_{B, \alpha}^k$
1. One-Step Distrust: Distrusted users have their outgoing judgments discounted entirely, so distrust propagates only a single step at the termination of the path, while positive trust propagates repeatedly across $k$ steps:
$B = T, \quad P^{(k)} = C_{B, \alpha}^k \cdot (T - D)$
1. Propagated Distrust: Trust and distrust are treated as opposite ends of a continuous spectrum, propagating together along multi-hop paths:
$B = T - D, \quad P^{(k)} = C_{B, \alpha}^k$
Knowl 3 — Iterative Propagation Aggregation Schemes
model/method
To combine inferences across propagation steps into a final belief matrix $F \in \mathbb{R}^{n \times n}$ (where $F_{ij}$ denotes the computed trust of user $i$ for user $j$ ), two aggregation schemes are defined over the $k$ -step propagation matrices $P^{(k)}$ up to a maximum step horizon $K$ :
1. Eigenvalue Propagation (EIG): Uses the propagation matrix at the terminal step $K$ directly:
$F = P^{(K)}$
1. Weighted Linear Combinations (WLC): Sums propagation matrices across all steps $k \in \{1, \dots, K\}$ weighted by a discount factor $\gamma$ strictly smaller than the largest eigenvalue of the atomic propagation matrix $C_{B, \alpha}$ , penalizing longer inference chains:
$F = \sum_{k=1}^K \gamma^k P^{(k)}$
Knowl 4 — Rounding Methods for Continuous Trust Beliefs
model/method
To map continuous computed belief scores $F_i$ (the $i$ -th row of belief matrix $F$ ) into discrete Boolean classifications ( $+1$ for trust, $-1$ for distrust) for a target user $j$ , three rounding algorithms are used:
1. Global Rounding: User $i$ is predicted to trust $j$ if and only if $F_{ij}$ lies within the top $\tau$ fraction of all values in the vector $F_i$ under standard real ordering, where $\tau \in [0, 1]$ is fixed to the global ratio of expressed trust to distrust edges across the entire network.
2. Local Rounding: User $i$ is predicted to trust $j$ if and only if $F_{ij}$ lies within the top $\tau$ fraction of values in $F_i$ , where $\tau$ is user-specific and equals the fraction of trust judgments among all trust and distrust ratings explicitly expressed by user $i$ .
3. Majority Rounding: Let $J$ be the set of all users for whom user $i$ has explicitly expressed a trust or distrust label. The users in $J \cup \{j\}$ are sorted in ascending order according to their continuous belief scores $F_{i, \cdot}$ . In this 1D ordered sequence, the unrated user $j$ is embedded at a unique position among known $+1$ and $-1$ labels. The predicted rating for $j$ is the majority label of the smallest symmetric neighborhood around $j$ in which a majority is non-tied.
Knowl 5 — Multiplicative versus Additive Distrust Transitivity
model/method
Chaining distrust across multi-step paths yields two philosophically distinct semantics:
1. Multiplicative Distrust Propagation: If user $i$ distrusts user $j$ and user $j$ distrusts user $k$ , multiplying negative values implies that $i$ trusts $k$ ("the enemy of my enemy is my friend"). Standard matrix multiplication naturally implements this model. However, multiplicative propagation can produce pathological behaviors, such as directed negative cycles that cause a user to distrust themselves or generate beliefs that overwhelm explicitly expressed ratings.
2. Additive Distrust Propagation: If user $i$ distrusts user $j$ because $j$ 's judgment is inferior, and $j$ distrusts $k$ , user $i$ should distrust $k$ even more strongly ("do not trust someone who is not trusted by someone you distrust"). Additive propagation can be integrated into matrix multiplication by applying an exponential transformation to the initial belief matrix entries $m_{ij}$ before running iterative propagation:
$m'_{ij} = \begin{cases} \exp(m_{ij}) & \text{if } m_{ij} \neq 0, \\ 0 & \text{otherwise.} \end{cases}$

Knowl 6 — Empirical Prediction Errors for 81 Trust and Distrust Propagation Schemes on Epinions Data

data/table

Cross-validation results across 81 combinations of propagation mechanisms evaluated on 3,250 randomly masked edges from the Epinions trust network. The table reports the prediction error rate $\epsilon$ across all 3,250 trials (where a naive baseline predicting all trust incurs $\epsilon = 0.150$ due to 85.29% trust edge prevalence) and the prediction error $\epsilon_S$ on a balanced subset $S$ of 996 edges (498 trust, 498 distrust; naive baseline error is 0.500). Propagation schemes use weight vectors $e_1 = (1, 0, 0, 0)$ (direct propagation only), $e_2 = (0, 1, 0, 0)$ (co-citation only), and $e^* = (0.4, 0.4, 0.1, 0.1)$ with $K = 20$ propagation steps.

Iteration	$\alpha$	Propagation	Global round.		Local round.		Maj. round.
			$\epsilon$	$\epsilon_S$	$\epsilon$	$\epsilon_S$	$\epsilon$	$\epsilon_S$
EIG	$e_1$	Trust only	0.153	0.500	0.123	0.399	0.077	0.175
		One-step distrust	0.119	0.251	0.108	0.223	0.067	0.162
		Prop. distrust	0.365	0.452	0.368	0.430	0.084	0.206
	$e_2$	Trust only	0.153	0.500	0.114	0.365	0.080	0.190
		One-step distrust	0.097	0.259	0.087	0.234	0.066	0.159
		Prop. distrust	0.149	0.380	0.121	0.279	0.080	0.187
	$e^*$	Trust only	0.153	0.500	0.107	0.336	0.077	0.180
		One-step distrust	0.096	0.253	0.086	0.220	0.064	0.147
		Prop. distrust	0.110	0.284	0.101	0.238	0.079	0.180
	$e_1$	Trust only	0.153	0.500	0.123	0.390	0.189	0.163
		One-step distrust	0.093	0.231	0.083	0.205	0.098	0.205
		Prop. distrust	0.102	0.221	0.098	0.199	0.121	0.295
	$e_2$	Trust only	0.153	0.500	0.113	0.354	0.074	0.174
		One-step distrust	0.088	0.254	0.080	0.231	0.093	0.187
		Prop. distrust	0.126	0.336	0.100	0.252	0.076	0.177
	$e^*$	Trust only	0.153	0.500	0.108	0.340	0.078	0.159
		One-step distrust	0.086	0.247	0.076	0.217	0.092	0.190
		Prop. distrust	0.087	0.237	0.079	0.203	0.074	0.162
	$e_1$	Trust only	0.153	0.500	0.123	0.391	0.132	0.152
		One-step distrust	0.102	0.241	0.092	0.216	0.069	0.171
		Prop. distrust	0.111	0.238	0.106	0.211	0.101	0.227
	$e_2$	Trust only	0.153	0.500	0.113	0.356	0.078	0.184
		One-step distrust	0.092	0.260	0.082	0.235	0.071	0.173
		Prop. distrust	0.134	0.355	0.106	0.261	0.078	0.188
	$e^*$	Trust only	0.153	0.500	0.107	0.337	0.075	0.169
		One-step distrust	0.091	0.253	0.082	0.222	0.072	0.171
		Prop. distrust	0.091	0.254	0.081	0.209	0.078	0.177

The lowest error rates on both the overall sample ( $\epsilon = 6.4\%$ ) and the balanced subset ( $\epsilon_S = 14.7\%$ ) are achieved by the combination of EIG iteration, one-step distrust propagation, combined basis weights $e^* = (0.4, 0.4, 0.1, 0.1)$ , and majority rounding.

Knowl 7 — Empirical Performance Findings across Trust Propagation Dimensions
empirical result
Systematic evaluation of 81 combinations on the Epinions web of trust reveals four main empirical conclusions:
1. Impact of Distrust: Incorporating distrust consistently improves prediction accuracy over trust-only propagation across almost all settings. Under EIG iteration, one-step distrust propagation outperforms both trust-only and propagated distrust across all basis vectors and rounding methods.
2. Superiority of Majority Rounding: Majority rounding decisively outperforms local rounding, which in turn outperforms global rounding across all iteration models and propagation bases. Preserving local neighbor clustering in 1D sorted score space is critical for accurate Boolean decision-making.
3. Effectiveness of Co-Citation and Combined Bases: Co-citation alone ( $\alpha = e_2$ ) achieves strong predictive accuracy despite not relying on direct path transitivity. The combined basis vector $e^* = (0.4, 0.4, 0.1, 0.1)$ yields the lowest error overall, providing robustness across different graph sub-topologies.
4. Pathology of Pure Direct Distrust Propagation: Pure direct propagation ( $\alpha = e_1$ ) paired with propagated distrust often degrades accuracy substantially compared to one-step distrust due to pathological multiplicative sign-alternation over multi-step paths.
Knowl 8 — Structural and Connectivity Properties of the Epinions Trust Network
empirical result
Analysis of the full Epinions trust dataset establishes the following structural properties:
- Scale and Label Balance: The graph contains $n = 131,829$ nodes and $841,372$ directed edges. Of these, 85.29% represent trust ( $+1.0$ ) and 14.71% represent distrust ( $-1.0$ ).
- Degree Distributions: Both indegree and outdegree follow a power-law distribution with an exponent of approximately $-1.7$ , which is flatter than typical Web hyperlink graph exponents (which are usually below $-2.0$ ).
- Bow-Tie Topology: The directed network exhibits a pronounced bow-tie structure:
  - Strongly Connected Component (SCC): 41,441 nodes (the second largest SCC contains only 15 nodes).
  - IN Component (nodes with directed paths into the SCC): 39,888 nodes.
  - OUT Component (nodes reachable from the SCC): 30,823 nodes.
  - Giant Undirected Component: When edge directionality is ignored, 119,130 nodes form a single connected component.
- Trust Subgraph Invariance: Degree distributions and bow-tie connectivity remain largely unchanged when restricting the network strictly to the positive trust subgraph.

Knowl 9 — Impact of Propagation Step Count on Prediction Error

data/table

Evaluation of prediction error rates $\epsilon$ (full sample) and $\epsilon_S$ (balanced sample $S$ ) across iteration horizons $K \in \{1, \dots, 7\}$ using EIG iteration ( $\gamma = 0.9$ , 1,000 samples) and cluster/majority rounding:

Iter. $K$	Trust only, $\alpha = e_1$		*One-step distrust, $\alpha = e^$**		*Prop. distrust, $\alpha = e^$**
	$\epsilon$	$\epsilon_S$	$\epsilon$	$\epsilon_S$	$\epsilon$	$\epsilon_S$
1	0.120	0.300	0.096	0.209	0.080	0.209
2	0.189	0.216	0.086	0.197	0.082	0.191
3	0.177	0.184	0.088	0.203	0.074	0.184
4	0.157	0.153	0.091	0.206	0.084	0.188
5	0.150	0.156	0.086	0.200	0.082	0.197
6	0.141	0.153	0.086	0.203	0.080	0.197
7	0.135	0.156	0.082	0.197	0.081	0.194

For direct propagation alone ( $\alpha = e_1$ ), increasing iterations from $K=1$ to $K=7$ substantially reduces the balanced error $\epsilon_S$ from $0.300$ to $0.156$ , showing that multi-step reach is necessary to traverse directed paths. For combined propagation mechanisms ( $\alpha = e^*$ ), performance converges by step 2 or 3 (reaching $\epsilon_S \approx 0.19 - 0.20$ ), because co-citation and coupling operators bridge user pairs via short 2-step paths without requiring deep iteration.

Knowl 10 — Sparse Matrix-Vector Evaluation Algorithm for Pairwise Trust Prediction
algorithm
Because full matrix-matrix multiplication on an $n \times n$ graph ( $n = 131,829$ ) is computationally prohibitive, predicting whether user $i$ trusts user $j$ is computed by evaluating only the $i$ -th row vector $F_i$ via iterative matrix-vector multiplications.
Input: Trust matrix $T \in \{0, 1\}^{n \times n}$ , distrust matrix $D \in \{0, 1\}^{n \times n}$ , weight vector $\alpha = (\alpha_1, \alpha_2, \alpha_3, \alpha_4)$ , iteration depth $K$ , iteration type (EIG or WLC with discount $\gamma$ ), distrust mode (Trust only, One-step, or Propagated), source user $i$ , target user $j$
Output: Predicted discrete label $\hat{y}_{ij} \in \{+1, -1\}$
if distrust mode is "Propagated" then
     $B = T - D$
else
     $B = T$
Initialize row vector $v^{(0)} = e_i$ (where $e_i$ is the $i$ -th standard basis vector in $\mathbb{R}^{1 \times n}$ )
if iteration type is EIG then
    for $k = 1$ to $K$ do
         $v^{(k)} = \alpha_1 (v^{(k-1)} B) + \alpha_2 ((v^{(k-1)} B^T) B) + \alpha_3 (v^{(k-1)} B^T) + \alpha_4 ((v^{(k-1)} B) B^T)$
    if distrust mode is "One-step" then
         $F_i = v^{(K)} (T - D)$
    else
         $F_i = v^{(K)}$
else if iteration type is WLC then
     $F_i = 0$
    for $k = 1$ to $K$ do
         $v^{(k)} = \alpha_1 (v^{(k-1)} B) + \alpha_2 ((v^{(k-1)} B^T) B) + \alpha_3 (v^{(k-1)} B^T) + \alpha_4 ((v^{(k-1)} B) B^T)$
        if distrust mode is "One-step" then
             $u^{(k)} = v^{(k)} (T - D)$
        else
             $u^{(k)} = v^{(k)}$
         $F_i = F_i + \gamma^k u^{(k)}$
Apply rounding (Global, Local, or Majority) to $F_i$ to determine whether user $i$ trusts or distrusts user $j$
return predicted label $\hat{y}_{ij}$

Coverage note — None was omitted; all contributed models, distrust propagation modes, aggregation and rounding algorithms, structural data characterizations, and empirical findings are represented.

References

1.G. Ackerlof. The market for lemons: Quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84:488–500, 1970.
2.A. Armstrong and J. Hagel III. The real value of online communities. Harvard Business Review, pages 134–141, 1996.
3.C. Avery, P. Resnick, and R. Zeckhauser. The market for evaluations. The American Economic Review, 89:564–584, 1999.
4.S. Ba and P. Pavlou. Evidence of the effect of trust building technology in electronic markets: Price premiums and buyer behavior. MIS Quarterly, 26(3):243–268, 2002.
5.S. Ba, A. B. Whinston, and H. Zhang. Building trust in online auction markets through an economic incentive mechanism. Decision Support Systems, 35(3):273–286, 2002.
6.T. Beth, M. Borcherding, and B. Klein. Valuation of trust in open networks. In 3rd European Symposium on Research in Computer Security, pages 3–19, 1994.
7.A. Z. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins, and J. Wiener. Graph structure in the web. WWW9/Computer Networks, 33(1–6):309–320, 2000.
8.M. Burrows, M. Abadi, and R. Needham. A logic of authentication. ACM Transactions on Computer Systems, 8(1):18–36, 1990.
9.J. Coleman. Foundations of Social Theory. Harvard University Press, 1990.
10.U. Frendrup, H. Hüttel, and J. N. Jensen. Modal logics for cryptographic processes. Electronic Notes in Theoretical Computer Science, 68(1), 2002.
11.M. Gladwell. The Tipping Point, How Little Things Can Make a Big Difference. Little Brown, 2000.
12.D. Houser and J. Wooders. Reputation in auctions: Theory, and evidence from eBay. Technical report, University of Arizona, 2000.
13.D. Kahneman, P. Slovic, and A. Tversky. Judgment Under Uncertainty: Heuristics and Biases. Cambridge University Press, 1982.
14.S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina. The eigentrust algorithm for reputation management in P2P networks. In Proceedings of the 12th International World Wide Web Conference, pages 640–651, 2003.
15.J. M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5):604–632, 1999.
16.P. Kollock. The production of trust in online markets. In E. J. Lawler amd M. Macy, S. Thyne, and H. A. Walker, editors, Advances in Group Processes, volume 16, pages 99–123. JAI Press, 1999.
17.C. G. McDonald and V. C. Slawson Jr. Reputation in an internet auction model. Technical report, University of Missouri-Columbia, 2000.
18.B. Misztal. Trust in Modern Societies: The Search for the Bases of Social Order. Polity Press, 1996.
19.P. Resnick and R. Zeckhauser. Trust among strangers in internet transactions: Empirical analysis of eBay's reputation system. Technical report, University of Michigan, 2001.
20.P. Resnick, R. Zeckhauser, E. Friedman, and K. Kuwabara. Reputation systems. Communications of the ACM, 43:45–8, 2000.
21.M. Richardson, R. Agrawal, and P. Domingos. Trust management for the semantic web. In Proceedings of the Second International Semantic Web Conference, pages 351–368, 2003.
22.J. M. Snyder. Online auction fraud: Are the auction houses doing all they should or could to stop online fraud. Federal Communications Law Journal, 52:453–472, 2000.
23.P. Sztompka. Trust: A Sociological Theory. Cambridge University Press, 1999.
24.The MIT PGP Team.
25.B. Yu and M. P. Singh. A social mechanism of reputation management in electronic communities. In Cooperative Information Agents, pages 154–165, 2000.

Citation

MLA

Guha, R., et al. “Propagation of Trust and Distrust”. Proceedings of the 13th International Conference on World Wide Web, 2004, pp. 403–12, https://doi.org/10.1145/988672.988727.

APA

Guha, R., Kumar, R., Raghavan, P., & Tomkins, A. (2004). Propagation of trust and distrust. Proceedings of the 13th International Conference on World Wide Web, 403–412. https://doi.org/10.1145/988672.988727

Chicago

Guha, R., R. Kumar, P. Raghavan, and A. Tomkins. 2004. “Propagation of Trust and Distrust”. Proceedings of the 13th International Conference on World Wide Web, 403–12. https://doi.org/10.1145/988672.988727.

Harvard

Guha, R. et al. (2004) “Propagation of trust and distrust”, Proceedings of the 13th international conference on World Wide Web. ACM, pp. 403–412. Available at: https://doi.org/10.1145/988672.988727.

Vancouver

1. Guha R, Kumar R, Raghavan P, Tomkins A (2004) Propagation of trust and distrust. In: Proceedings of the 13th international conference on World Wide Web. ACM, pp 403–412

BibTeX

@inproceedings{Guha_2004, series={WWW04}, title={Propagation of trust and distrust}, url={http://dx.doi.org/10.1145/988672.988727}, DOI={10.1145/988672.988727}, booktitle={Proceedings of the 13th international conference on World Wide Web}, publisher={ACM}, author={Guha, R. and Kumar, Ravi and Raghavan, Prabhakar and Tomkins, Andrew}, year={2004}, month=May, pages={403–412}, collection={WWW04} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Propagation of trust and distrust

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Atomic Trust Propagation Operators and Combined Matrix Formulation

Knowl 2 — Models for Distrust Propagation

Knowl 3 — Iterative Propagation Aggregation Schemes

Knowl 4 — Rounding Methods for Continuous Trust Beliefs

Knowl 5 — Multiplicative versus Additive Distrust Transitivity

Knowl 6 — Empirical Prediction Errors for 81 Trust and Distrust Propagation Schemes on Epinions Data

Knowl 7 — Empirical Performance Findings across Trust Propagation Dimensions

Knowl 8 — Structural and Connectivity Properties of the Epinions Trust Network

Knowl 9 — Impact of Propagation Step Count on Prediction Error

Knowl 10 — Sparse Matrix-Vector Evaluation Algorithm for Pairwise Trust Prediction

References

Citation

BibTeX

Access the Paper