Steem Sincerity - Update and Community Involvement

in #steemdev7 years ago (edited)





Thanks to everyone for the support and feedback on the Sincerity (anti-spam) API.

Community Response

The community has broadly welcomed this project, but there are also those who feel that the early integration of an imperfect account classifier was handled insensitively. The API provides estimated probability scores rather than absolute classifications, but many people still felt they were unfairly classified because initially, the SteemPlus Chrome browser extension itself used the highest probability as the classification. As a result of feedback, this has now changed to show 'Tell us' in cases of significant doubt.

This has also been a learning experience for me though. I had, perhaps naively, assumed that the community would be more united about what constitutes spam than is actually the case. In the SteemPlus crowdsourced data, there are many conflicting reports where accounts are reported as both 'Human Content Creators' and 'Spammers' for example. With this in mind, it seems inevitable that no spam classifier will please everyone, and that it will not come close to perfectly classifying community sourced training data due to differing perceptions of what spam is.

Progress

Despite this complication, I still believe this is a worthwhile project, and I am currently busy working on improving the account classification algorithm, to increase accuracy and reduce the 'false positives' whereby people are mistakenly identified as spammers. Thanks to the community providing more than 3000 feedback reports through SteemPlus (so far), after some algorithmic validation, I have expanded the training set from 90 to 480 accounts. These will typically be the more challenging accounts, that are 'wrongly' reported by SteemPlus, so they should serve as good training material.

Given the significant differences of opinion about which accounts are spamming however, I have decided to make the training data public, and ask for more direct community involvement about which accounts should be classified as spammers, so at the bottom of this post is the data I'm currently using to train the classifier.

Community Crowdsourcing

It would be great if anyone would like to contribute data to the training set!

If so, please paste your account lists into a comment below in the following form:

Human Content Creators
account1, account2, account3

Spammers
account3, account4

Bots
account5

It is fine to include accounts I'm already using.

If you disagree with any accounts in my current lists below, please comment in this form:

Incorrect Classification
account1, account2, account3, account4

Note that as with the data from SteemPlus, all crowdsourced data will be filtered though an algorithm to decide which accounts are added to the training data. So adding an account here does not guarantee its inclusion.

Discussion and ideas are also welcome in the comments of course :)

Current Training Data

This is mainly from the SteemPlus community feedback data and I have not had time to manually evaluate it, so it does not represent my opinions.

The 274 'human content creator' accounts:
aaronleang, abunagaya, ackhoo, ackza, adsactly, alejandromata, alex-fitness, altobot, amigoponc, andrarchy, andybets, angelggomz, argalf, arorapuneet, arvindkumar, atomcollector, atukh09, austrobot, babysteps, baejaka, bafi, balte, bcuda69, bembelmaniac, berniesanders, binkyprod, brandonp, brimax, bscrypto, buzzard, cardboard, carenina, carlgnash, cattledog, catweasel, chbartist, clarkgold, clevershovel, coin.info, coinbandit2000, contentjunkie, crazybgadventure, crisdevilgamer, cryptogecko, cryptosharon, cyclamen, d00k13, dailypick, dailypro, dailytop10open, dan, dana-edwards, dbooster, deathwing, digitalis, digitokash, djdarkstorm, dnews, doctorrevelator, drakos, drmake, dswigle, eddiespino, ehiboss, elear, eleidap, elgeko, elite.skeete, eliterry12, emergehealthier, espoem, ethandsmith, evildido, exyle, fabinhocrypto, feronio, flauwy, flugschwein, folken, followforupvotes, fr4mer, free999enigma, fxsajol, gbenga, gmichelbkk, grammarnazi, gravitcaper, gray00, greencross, grimmyx, grizzle, grumpycat, guidom, haejin, hakeemshah96, happydaddyfr, hatuvera, heimindanger, hellroute, hethur240, hilladigahackles, hungryhustle, iamankit, indigoocean, infoslink, inventor16, invisusmundi, isi3, itharagaian, j-alhomestudio, jackjohanneshemp, jeffmcmullen, jehovahwitness, jlordc, joannereid, johannfrare, joseph, jpphotography, jrvacation, kadna, kaliju, kamile, kanrat, kastiuz, keter, kingscrown, kryptoe, lanhange, lenasveganliving, lexiconical, lonelywolf, lost108, louis88, lundsten, luzcypher, magicalmoonlight, majharul, marcelgoo, mark-dahl, marsella-2017, masjuan, maverickfoo, meanmommy33, mecurator, megaela, melinda010100, menerva, meno, mfederi, midlet, mohammedpolash, molometer, mountainjewel, mrbean1, mxx, myndnow, nakedverse, neopch, new-york, newenx, nonameslefttouse, obvious, old-guy-photos, oldtimer, oliverschmid, omar-hesham, omersurer, omitaylor, onealfa, oscarps, osmerj, oups, paradise, pars11, patrice, personz, pfunk, photosblog, piaristmonk, pipiczech, pipurilla, polm, pritam20, pyro0816, r-k-m, ragepeanut, rahulsaini, raorac, ravenruis, rebeccaontheroof, reesebrehio, reggaemuffin, reko, rem3600, richardcrill, rival, rivalzzz, roelandp, roxane, rudenc, rusni1122, sahda, sahra-bot, sametceylan, samueldouglas, sargoon, schrosct, scolari-ire, seablue, sebbbl, serylt, seveaux, shadowspub, shahabshah, sharehows, sheorath, sherlockholmes, simondiamond, sircork, sisygoboom, smacommunity, son-of-satire, spawnband, spongechris, ssimkins9, steem-plus, steem.chat, steem.dollar, steemplayroom, steemreports, steevc, stehaller, stellabelle, stoodkev, sukro, superoo7, taphophilia, tarazkp, taskmanager, tattoodjay, teammalaysia, teutonium, thebugiq, themonetaryfew, therealwolf, theturtleproject, timcliff, tm50, tmholdings, tolgahanuzun, toptrendingnews, transisto, travelwithus, trufflepig, tsaaditia30, tts, twoitguys, txatxy, vaansteam, vasil-danev, veerall, veganroma, verhp11, vicrivasr, vladimir-simovic, wissyofenmu, wrath-of-grapes, yann85, yanosh01, yidneth, youarehope, zekans84, zenkly, zonguin

The 119 'spammer' accounts:
aabisteemvoter, agx, ahlawat, aiqabrago, alexis2, all-aceh, alomgir0101, alomgirhoseain, ambriya11nov, andreacrangel, arielb12, baninduana, benswann, bilal218, bitgeek, bitius, bongje, bradfordtennyson, capari, chadgarber, chanchalroy, cinelonga, coldproject, crispycoinboys, cryptoconfiance, cryptoinside, cryptomario, cryptoriddler, cryptotenx, darryljonesjr, deboas, dogimage, dontryme2, dotwin1981, dreimaldad, dreykan, energypa, fawadsolangi, filo6322, finkployd, forexflo, francosteemvotes, gauravtak, gemce, hagoodman, happydolphin, hiranur, hodgetwins, ili0braz, introbot, irrer-ivan, jahnubis, jeabsywanvisa, jensvoigt, jiyaur, julia6, kate1, kimthewriter, kiporen212, leecamp, lefactuoscope, lefthouse, lianaakobian, liebeilio, love777, lucas3, mattl, mcitron, mianfahad2, michelle2, minecrew, muhammadroni, murattatar, murhadi9, nachon, nicecooking, nicole3, nicole5, oliverstoney, oz27, philou, pippo84, rachel4, rasel49, resteembot, riccardo47, rinis, rkaitra, rollthedice, romyjaykar, rrnayak, sacredwriter, siiiiichfried, sirgatodaniel, skeaa14, soundwavesphoton, speedvoter, steemlota, steemstem-bot, stef77, stmit, streembot, tarikhakan55, thejimmydoreshow, tonimontana, tonkatonka, underpants, unixfriend, vivekkanade, wafrica, wefund, wolf92, yasmin3, yetxuni, yoshiko, yougotresteemed, youneedverse, yundong21, zulacut

The 87 'bot' accounts:
adriatik, aksdwi, allaz, alphaprime, appreciator, arcange, bearwards, bluebot, bodzila, boomerang, boostbot, booster, bottymcbotface, brupvoter, buildawhale, cheetah, childfund, chronocrypto, cryptoempire, cuddlekitten, dailyupvotes, deutschbot, dlivepromoter, dolphinbot, ebargains, edensgarden, emperorofnaps, estabond, estream.studios, fishbaitbot, gaman, honestbot, isotonic, lightningbolt, lost-ninja, lovejuice, lrd, luckyvotes, megabot, minnowbooster, minnowfairy, minnowhelper, minnowsupport, minnowvotes, moneymatchgaming, msp-bidbot, nado.bot, noicebot, oceanwhale, onlyprofitbot, peace-bot, photocontests3, postdoctor, postpromoter, proffit, promobot, pushbot, pushup, pwrup, redlambo, redwhale, resteemable, rocky1, seakraken, singing.beauty, sleeplesswhale, slimwhale, smartsteem, sneaky-ninja, spydo, steembloggers, steemersbot, steemitboard, steeply, sunrawhale, thebot, therising, tomole444, treeplanter, twitterbot, upme, upmyvote, upyou, voterunner, whalebuilder, youtake, zapzap

Sort:  

I love the work that you do. I cam accross you after you commented on my article regarding producing reports for tax purposes.

I was just about to write an article about steemreports and in particular the section referred to here: http://steemreports.com/accounts-reporting-tool/

I noticed that due to no fault of your own it is not working right now. Could you please contact me via Steemit.chat when it is ready to go again?

Thanks for your great work!

Hi, thanks for the support!

I will let you know when the tools are working again.

Ok great. Thanks!

Interesting what you're working on, here. Seems to me like bots would need to be subdivided further-- a manual process. Arcange... sometimes is a phishing warning bot; that's a public service announcement. Cheetah, again serves a service function. Some small bots like cuddlekitten only interact with people who have upvoted/followed them... making them "requested" and probably harmless. Some might be classified as "nuisance," but harmless (steemitboard?). Some are bot-spam.

I would have to say some of this would be very hard to fully automate. But still following this with interest.

Thanks for your feedback. A service like this will never be perfect because people don't agree about these definitions, but I am working on improvements to the algorithm to make it more useful to most users.

in the meantime... @cuddlekitten's AI classification as a spammer has risen in ranks (she's now classified as the top-8 spammer on steemit - which is ridiculous!)

and I think I am already feeling the repercussions as @abusereports has started targeting her for auto-flagging! Though, that might be purely coincidental of course.

P.S.: Thanks for mentioning my kitten here, it's a tough day and I appreciate the moral support I get from reading this a lot!

@cuddlekitten doesn't seem to be classified as a spammer now for whatever reason. The classifier hasn't yet changed though.

Yeah, her score sincerity score changed from 1.0 spammer to 0.82 human after she wrote a single post about the issue... at first I thought you had manually changed something, but then I realised it's probably due to the significantly different appearance of the account data (a new post, before there were only comments - a ton of flags, before there were zero).

I've also been able to get her off the auto-flagging by abusereports - I do not know if there was any relation between the two situations though.

I've just conducted an experiment with the other cuddle-bot @steemkitten achieving the same result by making a single "wordy" post under the account: [see here]

Cool, but unfortunately publicising such feats of reclassification makes it more likely I'll need to change the algorithm as spammers learn from you. It's a battle to stay ahead of their game you see.

I guess that's exactly what I am trying to do... convince you to improve that algorithm. I feel it's still very unreliable as easily illustrated with my little experiment... so apart from actual spammer having it easy to trick your algorithm... non-spammers also easily get labelled wrongly...

so yeah... improve that algorithm... heh

I will certainly try and do that. I guess I can't expect too much support from people who are sailing so close to the wind ;)

I am giving all the support that I can.

You must excuse if I do sound a bit salty, but the top-8 spammer classification has brought a lot of negative attention to an otherwise positively received project of mine.

But in all of this, I didn't take any offense and as a matter of fact i am encouraging users to try steem plus and help provide human feedback

If there's anything else I could do to help, I will.

I'll say it again. I explicitly support the steem sincerity project. I believe it is a great addition to the steem ecosystem.

It's just that I experienced a negative backlash from an (imho) false classification of my bots and as such I try to provide feedback and hope to illustrate the urgency of improving this asap because it is already being publicly used.

Incorrect Classification
soundwavesphoton, thejimmydoreshow, tomole444

@soundwavesphoton and @thejimmydoreshow are legitimate human (or "brand"?) accounts.

@tomole444 is a bot, yes, but given the amount of downvotes the community does see him as a spammer imho!

P.S.: Kudos to your combined efforts in making this system more accurate and less misleading... let's just hope the crowd-sourced data doesn't include too many trolls... we all now how quickly an AI starts talking Hitler :P

Thanks for the info - I've updated the training data.

Incorrect Classification
trufflepig

I know firsthand that he is a bot :-D

Thanks for the feedback!

Incorrect Classification
benswann

Thanks. benswann was classified as human by my software, despite being on the spammers training list :)

Oh is Benswann spamming or something? What is the definition of spamming that is being used for this, essentially using a ton of bots for upvoting posts?

https://steemit.com/@benswann

I was just looking over his stuff, looks like he posts a few times a week? Was just curious. thanks.

According to the crowdsourced data, he is considered a spammer, receiving several reports. I think though, that it is data collection process is at fault. Either people don't agree on what spam is, or they're just using the spam labelling function of SteemPlus to try and hide accounts they don't like.

or it's just plain trolling :P

Good to see I'm considered 'human' :) Anything like this is tricky. Maybe Steem Plus could show different levels. e.g. a 100% spammer could be bright red, but if they are nearer 50/50 with human is could be more pink.

Yeah, it's certainly not easy :) @stoodkev has decided to mark the accounts with mixed score with a 'Tell Us' button, hopefully as the accuracy improves, there will be fewer unclassified accounts in his SteemPlus interface, but some people were getting understandably annoyed.

Heyyy, I'm in a list! :D

dances

...and it's the right list! :)

I know quite a few names on the list (including myself :P) and I can confirm that your algorithms seem to be quite accurate, ... although:

Incorrect Classicifation:
lianaakobian, underpants.

These two are definitely human creators.

My classification software classified those two as humans, despite them being part of the spammers training list, which pleases me. The people who assigned them as spammers, must just have done so because they couldn't read German I think.

That pleases me as well! :D

Thanks for that, I'll add your records. These lists are not produced by the software, but are to be used to train the software.

Ah, good to know! :D

I would like to see myself classified as Human, even if I sometimes use bots to help me in my work. Even in this case, there is a lot of work done upstream.

My development classifer puts your account more towards human than bot, but that isn't ready yet.

My development classifer puts your account more towards human than bot

You should create a Cyborg category ;)

I'm not a spammer.

Thanks for confirmation - I have updated the training set.