Fairness of the user recommendation systems for Mastodon

Last updated 3 months ago

Motivation

  • The presence of users is oligopolized by few celebrities in consumer generated media.

  • Decentralization of users is expected.

User recommendation systems for Mastodon

  • List of recommended users

    • Mastodon User Matching (Vaginaplant's)

    • Mastodon Recommended User Search (Cucmberium's)

      • Recommended users

      • Similar users

    • Recommended Followers (Osa's)

    • Potential friendship (Embedded)

  • Who to follow panel

    • Pawoo

    • Pleroma

    • Halcyon

    • Misskey

Other user discovery methods

  • User ranking

    • By their number of followers

    • By their number of boosts

    • By their activity (toots per day)

    • Newcomers

    • Manual

    • Public profile endorsements (Embedded)

  • Timelines

    • Home (boosts)

    • Federated

    • Local

  • User search

  • Toot search

Fairness of user recommendations

  • Fair in this slide: To support decentralization of users.

  • Unfair in this slide: To support oligarchy of users.

Summery of fairness and unfairness

  • Fair

  • Newcomers

  • User search

  • Mastodon User Matching (Vaginaplant's)

  • Similar users (Cucmberium's)

  • Toot search

  • Potential friendship (Embedded)

  • Local timeline

  • Federated timeline

  • Home timeline

  • User ranking by their activity

  • Pawoo's who to follow panel

  • Misskey's who to follow panel

  • User ranking by their number of boosts

  • Public profile endorsements (Embedded)

  • Manual guidebook

  • Recommended users (Cucmberium's)

  • Recommended Followers (Osa's)

  • User ranking by their numbers of followers

  • Unfair

Apparent patterns of unfairness

Oldies effect

Old famous users still famous ever.

Positive feedback

If someone is recommended by the system, it cause the farther recommendation itself.

Shrink wrap effect

If the content is hidden, we evaluate someone by the fame or popularity.

Power user effect

  • The unfaireness of the list of the power users (much toots per day).

  • The culture of an instance is formed by the exclusive communication of some power users. [4, 5]

  • Though the power users contribute their instance's culture and security, they are some kind of celebrities.

Case 1: User ranking by their numbers of followers

Implementation

  • Mastodon Ranking [6] by User Local, Inc.

Specification

  • Sorted by the numbers of followers.

Unfairness

  • Oldies effect: Once someone has earned many followers, the user holds the followers ever, though the user posts no interesting posts. Followers rarely remove that guy.

  • Positive feedback: If we follow the users listed in this kind of ranking, it directly accelerates itself.

  • Shrink wrap effect: The ranking shows screen names, bios and avatars. No contents are shown apparently. We are attracted by the famous names or aesthetic avatars.

Fairness

  • No power user effect

Summary

  • One of the most unfair user discovery methods ever.

Case 2: Recommended Followers

Implementation

  • Recommended Followers [7] by Osa

Specification

  • When you use this recommendation system, it shows the users who are followed by the users who you follow.

Unfairness

  • The same oldies effect, positive feedback and shrink wrap effect as the user ranking by their numbers of followers.

  • Recommends the celebrities in the small society.

Fairness

  • No power user effect

Summary

  • One of the most unfair user discovery methods ever.

Case 3: Recommended users in Mastodon Recommended User Search

Implementation

  • Mastodon Recommended User Search [8] by Cucmberium

Specification

  • The users who your similar users follow.

Unfairness

  • The same oldies effect, positive feedback and shrink wrap effect as the user ranking by their numbers of followers.

Fairness

  • No power user effect

Summary

  • One of the most unfair user discovery methods ever.

Case 4: Manual guidebooks

Implementation

  • Mastodon Famous Accounts [9] by Takumi

Specification

  • A blog post which is manually written.

Unfairness

  • Oldies effect: Manual guidebooks are rarely updated.

  • Shrink wrap effect: The authors of the guidebooks are attracted by famous names or aesthetic avatars.

  • The authors of the guidebooks sometimes just follow some existing authority.

Fairness

  • No positive feedback: Our action does not affects to the list.

  • No power user effect

Summery

  • Super unfair user discovery.

  • Fair minded authors of the guidebooks are expected.

Case 5: Public profile endorsements

Implementation

  • Embedded [10]

Specification

  • The user pins other users on one's profile.

Fairness and Unfairness

  • Almost same as the manual guidebooks.

Summery

  • Super unfair user discovery.

  • Endorsement is worship or flattery. [11]

Case 6: User ranking by their number of boosts

Implementation

  • Mastodon Ranking [6] by User Local, Inc.

Specification

  • Sorted by the number of boosts.

Unfairness

  • Oldies effect: The number of boosts is accumulated ever.

  • Positive feedback: If we follow the user who is boosted well, we boost such kind of user's toots more often.

  • Power user effect: More toots, more boosts.

Fairness

  • No shrink wrap effect: We read the toot before we boost it.

Summary

  • Super unfair user discovery.

  • Better than user discovery methods depends on the number of followers.

Case 7: Misskey's who to follow panel

Implementation

  • Misskey [12, 13]

Specification

  • Posting in the last 7 days.

  • Local user.

  • Many followers.

Unfairness

  • The same positive feedback and shrink wrap effect as the user ranking by their numbers of followers.

Fairness

  • Less oldies effect: Requires posting in the last 7 days.

  • No power user effect

Summary

  • Remarkably unfair user discovery.

  • Better than oldies effected user discovery methods.

Case 8: Pawoo's who to follow panel

Implementation

  • Pawoo [14, 15]

Specification

  • Following in pixiv.

  • Popular (many favorited, many followers, many pictures) users.

  • Active (many recent toots) users.

Unfairness

  • Many favorited or many followers criteria causes oldies effect, positive feedback and shrink wrap effect.

  • Many pictures criteria causes oldies effect.

  • Many favorited or many pictures criteria cause power user effect.

Fairness

  • Less oldies effect: Requires recent posts.

  • Self effort criteria: Requires updating many pictures.

Summary

  • Remarkably unfair user discovery.

  • Better than oldies effected user discovery methods.

Case 9: User ranking by their activity

Implementation

  • Powerful Mastodon/Pleroma Users [16] by Vaginaplant

Specification

  • Sorted by toots per day.

Unfairness

  • Power user effect: Some power users control their instance's culture and security.

  • Shrinkwrap effect: Only avatars and screen names are displayed.

Fairness

  • No oldies effect: Recent toots affect.

  • No positive feedback: Our action dose not affects.

Summery

  • Still unfair user discovery.

  • Better than positive feedbacked or oldies effected user discoverty methods.

Case 10: Home timeline

Implementstion

  • Embedded

Specification

  • Following users' toots.

  • Following users' boosts.

Unfairness

  • Positive feedback: More followers, more boosts.

  • Power user effect: More toots, more boosts.

Fairness

  • No oldies effect: Fresh toots tend to be boosted.

  • No shrinkwrap effect: Content of toots are visible.

Summery

  • Still unfair user discovery.

  • Power user effect remarkably brings unfairness.

Case 11: Federated timeline

Implementstion

  • Embedded

Specification

  • Local users' toots.

  • Toots by the remote users who are followed by the local users.

Unfairness

  • Positive feedback: More followers, more boosts.

  • Power user effect: More toots, more boosts.

Fairness

  • No oldies effect: Fresh toots tend to be boosted.

  • No shrinkwrap effect: Content of toots are visible.

Summery

  • Still unfair user discovery.

  • Mixed feature of the home and local timelines.

Case 12: Local timeline

Implementstion

  • Embedded

Specification

  • Local users' toots

Unfairness

  • Power user effect: More toots, more boosts.

Fairness

  • No oldies effect: Fresh toots tend to be boosted.

  • No positive feedback: Any local user can posts to the local timeline.

  • No shrinkwrap effect: Content of toots are visible.

Summery

  • Still unfair user discovery.

  • Power user effect remarkably brings unfairness.

Case 13: Potential friendship

Implementstion

  • Embedded [17]

Specification

  • The user engages (replies, favourites and boosts) other user, the potential friendship counter is bumped.

Fairness and Unfairness

  • Mixed feature of the home, fedarated, and local timelines. [18]

Summery

  • Still unfair user discovery.

Case 14: Toot search

Implementation

  • Mastodon Search Portal [19]

  • Tootdon [20]

  • Mastodon Realtime Search [21]

  • Tootsearch [22]

  • Mastodon Search by Google [23]

Specification

  • Search toots by keyword.

Unfairness

  • Power user effect: More toots, more hits.

Fairness

  • No oldies effect

  • No positive feedback

  • No shrinkwrap effect

Summery

  • Still unfair user discovery.

  • Not suit for the user discovery use. Explicit keyword is mandatory.

Case 15: Similar users in Mastodon Recommended User Search

Implementation

  • Mastodon Recommended User Search [8] by Cucmberium

Specification

  • The users who follow who you follow.

Unfairness

  • Shrinkwrap effect

Fairness

  • No oldies effect

  • No positive feedback

  • No power user effect

Summary

  • Neutral for fairness or unfairness.

  • Small cliques or clusters are formed.

Case 16: Mastodon User Matching

Implementation

  • Mastodon User Matching [24] by Vaginaplant

  • Who to follow panel in Pleroma [25]

  • Who to follow panel in Halcyon [26]

Specification

  • The users who have the vocabulary which similar to you.

  • Vocabulary is abstracted from one's screen name, bio and toots.

Unfairness

  • Shrinkwrap effect

Fairness

  • No oldies effect

  • No positive feedback

  • No power user effect

Summery

  • Neutral for fairness or unfairness.

  • Some users may be recommended beyond the social graph.

Case 17: User search

Implementation

  • Mastodon/Pleroma User Search [27] by Vaginaplant

Specification

  • Search users by user name, screen name and bio.

Unfairness

  • Shrinkwrap effect

Fairness

  • No oldies effect

  • No positive feedback

  • No power user effect

Summery

  • Neutral for fairness or unfairness.

  • Not suit for the user discovery use. Explicit keyword is mandatory.

Case 18: Newcomers

Implementation

  • Mastodon/Pleroma Newcomers [28] by Vaginaplant

Specification

  • List of newcomers

Unfairness

  • Shrinkwrap effect

Fairness

  • Reverse oldies effect

  • No positive feedback

  • No power user effect

Summery

  • Fair user discovery.

Discussions

Fairness and convenience are trade-off?

No

  • Too many user recommendation systems are unfair AND inconvenient!

  • All of the oldies effected user discovery methods are shit.

Yes

  • It is too hard to be fair and convenient.

  • Google or other web rulers provide their positive feedbacked recommendations.

Future works

  • Objective and/or quantitative proof.

    • Quantitative metric for the decentralization of users.

    • Quantitative metric for fairness of the user discovery method.

Conclusion

  • Fairness of the user recommendation is a mandatory feature for the consumer generated media.

References

[1] Hakaba Hitoyo, 脱中央集権のためのデザイン: セレブのためのインターネットを99 %の手に取り戻す, https://hakabahitoyo.wordpress.com/2017/12/24/decentralization-by-design/

[2] Hakaba Hitoyo, マストドンのユーザーレコメンデーションがブームに?, https://hakabahitoyo.wordpress.com/2018/03/29/mastodon-user-recommendation-2018/

[3] Hakaba Hitoyo, マストドンのユーザーレコメンデーションについて追記, https://hakabahitoyo.wordpress.com/2018/04/18/misskey-halcyon-recommendation/

[4] Jakob Nielsen, The 90-9-1 Rule for Participation Inequality in Social Media and Online Communities, https://www.nngroup.com/articles/participation-inequality/

[5] ニールセン, 参加の仕方は一様ではない: もっと大勢のユーザに書き込んでもらうには, https://u-site.jp/alertbox/20061009_participation_inequality

[6] User Local, Inc., Mastodon Ranking, http://mastodon.userlocal.jp/

[7] Osa, Recommended Followers, https://followlink.osa-p.net/recommend.html

[8] Cucmberium, Mastodonおすすめユーザー検索, https://nocona.cucmber.net/eryngium/

[9] Takumi, マストドン (Mastodon) ユーザなら必ずフォローしたい! アカウント一覧, https://takulog.info/mastodon-famous-accounts/

[10] Gargron, Public profile endorsements (accounts picked by profile owner) #8146, https://github.com/tootsuite/mastodon/pull/8146

[11] Hakaba Hitoyo, Endorsement is some kind of worship or flattery, https://github.com/tootsuite/mastodon/pull/8146#issuecomment-411702072

[12] Syuilo, Misskey, https://misskey.xyz/

[13] https://github.com/syuilo/misskey/blob/master/src/server/api/endpoints/users/recommendation.ts

[14] pixiv Inc., Pawoo, https://pawoo.net

[15] https://github.com/pixiv/mastodon/blob/pawoo/app/models/suggested_account_query.rb

[16] Hakaba Hitoyo, Powerful Mastodon/Pleroma Users, http://vinayaka.distsn.org/user-speed.html

[17] Gargron, Re-add follow recommendations API #7918, https://github.com/tootsuite/mastodon/pull/7918

[18] Hakaba Hitoyo, ユーザーレコメンデーションがマストドン本体に導入された: フェアネスの評価は「現状維持」, https://hakabahitoyo.wordpress.com/2018/07/03/gargrons-official-user-recommendation/

[19] マストドン検索ポータル, http://mastodonsearch.jp/

[20] MobiRocket Inc., 株式会社つくりと, Tootdon, http://tootdon.club/

[21] User Local, Inc., マストドンリアルタイム検索, http://realtime.userlocal.jp/

[22] Kirino Minato, Tootsearch, https://tootsearch.chotto.moe/

[23] Osa, マストドン検索 by google, https://mastosearch.osa-p.net/

[24] Hakaba Hitoyo, Mastodon User Matching, https://vinayaka.distsn.org

[25] lain, Pleroma, https://pleroma.social

[26] Neetshin, Niklas Poslovski, Halcyon, https://github.com/halcyon-suite/halcyon

[27] Hakaba Hitoyo, Mastodon/Pleroma User Search, http://vinayaka.distsn.org/user-search.html

[28] Hakaba Hitoyo, Mastodon/Pleroma Newcomers, http://vinayaka.distsn.org/user-new.html

[29] Hakaba Hitoyo, User Recommendation Systems for Mastodon and their Fairness, https://hakabahitoyo.gitlab.io/slides/recommendations

Contents