Skip to content

Latest commit

 

History

History
81 lines (52 loc) · 7.04 KB

S02. The Trust Network.md

File metadata and controls

81 lines (52 loc) · 7.04 KB

The Trust Network

2.1 Please

  • Take a few moments to read the opening README.md file which provides a brief overview of this project.
  • Interact with this and other articles I'll post here. I'm looking forward to hearing your thoughts, questions, and concerns.

Discussion Link: https://hachyderm.io/@davidshq/109829713731777957

2.2 The Manipulation Challenge

Search Engines are complicated beasts that tackle difficult problems. It is no small feat to index a representative portion of the web, maintain the freshness of that index, provide relevant rankings, and timely responses to user queries (to name only a few of the challenges).

Thankfully a number of these challenges have proven, satisfactory solutions. One area that continues to be a problem, however, is the manipulation of search results. While this challenge is not unique to human augmented search engines it has been a significant factor in their historical downfall.

While there is no single solution to results manipulation I believe that human curation with a trust graph can play a significant role in mitigating this problem.

2.3 The Trust Graph

Phoebe provides each (non-anonymous) user with an authority score. This score starts quite low but can be increased over time through several methods:

  1. Trust Grants - A trusted user explicitly grants trust to another user.
  2. Consistent Quality Contributions - The user makes substantive contributions over time that are consistently well received by the community and correlate well with Phoebe's automated ranking methods.
  3. Identity Verification - The user provides various forms of proof that they are a unique individual.[1]
  4. Credential Verification - The user provides proof of their expertise in a given subject.[2]

2.4 Preventing Gaming of Trust

While a network of trust provides some inherent protection against manipulation it is not invulnerable. Particularly:

  • An individual (or entity) might open multiple accounts and grant trust to themselves, with a large enough number of trust grants this could result in a high authority score.
  • Individuals might agree to grant trust to one another in exchange for reciprocal trust grants.
  • Genuine accounts might sell trust grants to less reputable accounts.

To prevent this type of manipulation from occurring several measures can be taken - the primary one involves rolling penalties, which I've outlined below.

Definitions: "Grantor" is the account that is granting trust to another account. "Grantee" is the recipient of a grantor's trust.

  • When an individual account (Grantor) trusts another account (Grantee), the Grantor account's reputation becomes linked to the Grantee's reputation.
  • If the Grantor trusts a spammy Grantee, the Grantor will receive a partial penalty to their reputation when the Grantor misbehaves.
  • This penalty increases exponentially with additional Grantee misbehavior. That is, if a Grantee is generally making quality contributions but makes one or two malicious contributions, the Grantor's account will receive a minor penalty. However, if the Grantee's behavior is consistently and severely malicious, the Grantor's account will receive more severe penalties to their authority.

In addition, this rolling penalty is not limited to first-level trust grants. Accounts that are one or two removed from a malicious Grantee will receive a penalty as well, although the penalty will be less severe.

For example:

JohnDoe (grantor) trusts JaneDoe (grantee). JaneDoe (grantor) trusts IAmSpammy (grantee). IAmSpammy (grantee) creates bad faith content. Penalties might be as follows: IAmSpammy (-10), JaneDoe (-5), JohnDoe (-2).

This doesn't make the trust graph impervious to manipulation but it does make it significantly more difficult to game. In order to gain authority one has to (a) provide value and (b) build relationships. To build relationships one must be trustworthy to an extent that the Grantor is willing to risk their own reputation if the Grantee misbehaves.

2.5 Compromised Accounts

It is inevitable that some trustworthy accounts will be compromised (e.g. hacked) but the trust network should be able to quickly detect this anomalous behavior, freeze the account, and hopefully recover the account for the legitimate owner.

2.6 Ensuring Marginalized Voices Are Heard

I recognize that some of the ways that trust are granted may be easier for some to accumulate than others and that this could result in specific "groups" of people having easier access to authority than others. Phoebe is committed to ensuring that marginalized voices are heard and that the trust network is not used to silence them.

Phoebe will intentionally counter this tendency by:

  1. Actively seeking out trustworthy members of marginalized communities and providing them with an additional authority grant to help offset such imbalances.
  2. Accept applications from individuals seeking an authority grant to help offset such imbalances for a specific community.
  3. Intentionally integrating multiple perspectives into the results. Individuals may submit an application for any topic to include an alternative perspective section in the results.
  4. Seek out assistance from recognized experts in the field of diversity and inclusion to help ensure that Phoebe is doing everything that can be done to ensure that marginalized voices are heard.

2.7 Credit Where Credit Is Due

I want to acknowledge my significant indebtedness to the marvelous innovations in automatic trust calculations pioneered by StackOverflow and Discourse.[3]

Discussion Link: https://hachyderm.io/@davidshq/109829713731777957

2.7 Footnotes

[1] Keybase was a good example of the variety of methods one could use to verify identity.
[2] Examples of proof might include authorship of books, academic credentials, career history as well as more unusual methods such as geographic location (e.g., one might be granted a higher authority score if providing information on a specific locale in which one resides).
[3] It is worth noting that trustworthiness does not necessarily translate into kindness. StackOverflow had a bit of a reckoning on this front and took a number of steps to improve the tone of the community. See for example the following articles:
1. StackOverflow Isn't Very Welcoming. It's Time for That to Change. (4/2018)
2. Rolling out the Welcome Wagon: June Update (6/2018)
3. Iterating on Inclusion (10/2019)
4. The Unfriendly Robot: Automatically flagging unwelcoming comments (4/2020)

Phoebe will implement measures to ensure that the community is welcoming and inclusive while also allowing maximal freedom of expression.