Trust propagation

Step 1 in the pipeline

Trust propagation is tasked to combine pretrusts and vouches to derive trust scores for the different users.

TrustPropagation ¶

Bases: ABC

Base class for Trust Propagation algorithms

call ¶

__call__(users: DataFrame, vouches: DataFrame) -> DataFrame

Propagates trust through vouch network

Parameters:

Name	Type	Description	Default
`users`	`DataFrame`	with columns user_id (int, index) is_pretrusted (bool)	required
`vouches`	`DataFrame`	with columns voucher (str) vouchee (str) vouch (float)	required

Returns:

Name	Type	Description
`users`	`DataFrame`	with columns user_id (int, index) is_pretrusted (bool) trust_score (float)

to_json ¶

to_json()

NoTrustPropagation ¶

NoTrustPropagation(pretrust_value: float = 0.8)

Bases: TrustPropagation

Implements no trust propagation. The result trust_score is directly based on the "pretrusted" status of each user.

Parameters:

Name	Type	Description	Default
`pretrust_value`	`float`	trust score to assign to pretrusted users	`0.8`

pretrust_value ¶

pretrust_value = pretrust_value

call ¶

__call__(users: DataFrame, vouches: DataFrame) -> DataFrame

to_json ¶

to_json()

LipschiTrust ¶

LipschiTrust(
    pretrust_value: float = 0.8,
    decay: float = 0.8,
    sink_vouch: float = 5.0,
    error: float = 1e-08,
)

Bases: TrustPropagation

A robustified variant of PageRank. In this algorithm, we leverage pre-trust (e.g., based on email domains) and vouching to securely assign trust scores to a wider set of contributors. The algorithm inputs pre-trust status and a vouching directed graph.

Parameters:

Name	Type	Description	Default
`pretrust_value`	`float`	The pretrust of a pretrusted user. (\(Trust^{pre}_{\checkmark}\) in paper) The algorithm guarantees that every pre-trusted user is given a trust score which is at least `pretrust_value`. Moreover, all users' trust score will be at most 1.	`0.8`
`decay`	`float`	The decay of trusts in voucher's vouchees. (\(\beta\) in paper) When considering a random walker on the vouch network, (1 - `decay`) is the probability that the random walker resets its walk at each iteration. LipschiTrust essentially robustifies the random walk, by frequently preventing the walker from visiting too frequently visited contributors, thereby bounding the maximal influence of such contributors.	`0.8`
`sink_vouch`	`float`	used to incentivize vouching (\(V^{sink}_{\checkmark}\) in paper) In our model we assume that each participating contributor implicitly vouches for a sink. The sink counts for `sink_vouch` vouchees. As a result, when a contributor with less than `sink_vouch` vouchees vouches for more vouchees, the amount of trust scores the contributor assigns grows almost linearly, thereby not penalizing previously vouched contributors. Vouching is thereby not (too) disincentivized.	`5.0`
`error`	`float`	Positive, is an upper bound on error, in L1 norm. \(\epsilon_{LipschiTrust}\) in paper	`1e-08`

pretrust_value ¶

pretrust_value = pretrust_value

decay ¶

decay = decay

sink_vouch ¶

sink_vouch = sink_vouch

error ¶

error = error

call ¶

__call__(users: DataFrame, vouches: DataFrame) -> DataFrame

to_json ¶

to_json()

NoopTrust ¶

Bases: TrustPropagation

Noop implementation for trust propagation: trust values are simply read from the input users.

call ¶

__call__(
    users: DataFrame, _vouches: DataFrame
) -> DataFrame

to_json ¶

to_json()

TrustAll ¶

Bases: TrustPropagation

A naive implementation that assigns an equal amount of trust to all users.

call ¶

__call__(users: DataFrame, vouches: DataFrame)

Trust propagation