codecombat/scripts/buildSchoolGraph.coffee
2015-12-08 12:12:56 -08:00

230 lines
10 KiB
CoffeeScript

# Organize our users' schoolNames.
database = require '../server/commons/database'
mongoose = require 'mongoose'
log = require 'winston'
async = require 'async'
### SET UP ###
do (setupLodash = this) ->
GLOBAL._ = require 'lodash'
_.str = require 'underscore.string'
_.mixin _.str.exports()
GLOBAL.tv4 = require('tv4').tv4
database.connect()
UserHandler = require '../server/users/user_handler'
User = require '../server/users/User'
#startDate = new Date 2015, 11, 1
startDate = new Date 2015, 11, 8 # Testing
query = dateCreated: {$gt: startDate}, emailLower: {$exists: true}
selection = 'name emailLower schoolName courseInstances clans ageRange dateCreated referrer points lastIP hourOfCode preferredLanguage lastLevel'
User.find(query).select(selection).lean().exec (err, users) ->
usersWithSchools = _.filter users, 'schoolName'
schoolNames = _.uniq (u.schoolName for u in usersWithSchools)
log.info "Found #{usersWithSchools.length} users of #{users.length} users registered after #{startDate} with schools like:\n\t#{schoolNames.slice(0, 10).join('\n\t')}"
# For each user, come up with a confidence that their school is correct.
# For users with low confidence, look for similarities to other users with high confidence.
# If we have enough data, prompt to update the school.
# After each update, recalculate confidence to find the next user with low confidence.
# How do we come up with confidence estimate?
# If there are many students with the same school name, it's either correct or a rename must happen.
# If the school name is unique but similar to a school name with many students, it's probably incorrect.
# But if we determine it is correct, how can we record this fact so it doesn't keep asking?
# How can we infer the school name when we think it's not correct?
# We look for users with confident schoolNames in shared courseInstances.
# ... in shared clans.
# ... with the same lastIP that doesn't cover the lastIP of students from multiple schools.
# If we find a school-district-formatted email domain, we could try to match to other schoolNames in that domain, but I doubt that will be helpful until we have a lot of data and a lot of time to manually look things up.
nextPrompt users
nextPrompt = (users, question) ->
sortUsers users
return console.log('Done.') or process.exit() unless [userToSchool, suggestions] = findUserToSchool users
question ?= formatSuggestions userToSchool, suggestions
prompt question, (answer) ->
return console.log('Bye.') or process.exit() if answer in ['q', 'quit']
answer = answer.trim()
if answer is ''
users = _.without users, userToSchool
else unless _.isNaN(num = parseInt(answer, 10))
schoolName = if num then suggestions[num - 1]?.schoolName else userToSchool.schoolName
return finalizePrompt userToSchool, suggestions, schoolName, users
else if answer.length < 10
console.log "#{answer}? That's kind of short--I don't think school names and locations can be this short. What should it really be?"
return nextPrompt users, "> "
else
return finalizePrompt userToSchool, suggestions, answer, users
nextPrompt users
finalizePrompt = (userToSchool, suggestions, schoolName, users) ->
console.log "Selected schoolName: \"#{schoolName}\""
question = "Also apply this to other users? Ex.: 'all', '0 1 2 5', 'all -3 -4 -5', '0' to just do this one, or blank to retype school name.\n> "
prompt question, (answer) ->
answer = answer.trim()
if answer is ''
console.log "Should just do", userToSchool._id, userToSchool.emailLower, userToSchool.schoolName
targets = [userToSchool]
else if answer is 'all'
targets = [userToSchool].concat (s.user for s in suggestions)
console.log "Doing all #{targets.length} users..."
else if /^all/.test answer
targets = [userToSchool].concat (s.user for s in suggestions)
numbers = _.filter (parseInt(d, 10) for d in answer.split(/ *-/)), (n) -> not _.isNaN n
for number in numbers
skip = if number then suggestions[number - 1].user else userToSchool
targets = _.without targets, skip
console.log "Doing all #{targets.length} users without #{numbers}..."
else
numbers = _.filter (parseInt(d, 10) for d in answer.split(/ +/)), (n) -> not _.isNaN n
targets = ((if number then suggestions[number - 1].user else userToSchool) for number in numbers)
console.log "Doing #{targets.length} users for #{numbers}..."
#User.update {_id: {$in: (_.map targets, '_id')}}, {schoolName: schoolName}, {multi: true}, (err, result) ->
User.update {_id: {$in: []}}, {schoolName: schoolName}, {multi: true}, (err, result) ->
if err
console.error "Ran into error doing the save:", err
return finalizePrompt userToSchool, suggestions, schoolName, users
console.log "Updated users' schoolNames. Result:", result
remainingUsers = _.without users, targets...
nextPrompt remainingUsers
formatUser = (user) ->
# TODO: replace date string with relative time since signup compared to target user
_.values(_.pick(user, ['name', 'emailLower', 'ageRange', 'dateCreated', 'lastLevel', 'points', 'referrer', 'hourOfCode'])).join(' ')
formatSuggestions = (userToSchool, suggestions) ->
suggestionPrompts = ("#{_.str.rpad(i + 1, 3)} #{_.str.rpad(s.schoolName, 50)} #{s.reasons.join(' + ')}\tfrom user: #{formatUser(s.user)}" for s, i in suggestions).join('\n')
"""
What should the school for this user be?
0 #{_.str.rpad(userToSchool.schoolName, 50)} #{formatUser(userToSchool)}
Suggestions:
#{suggestionPrompts}
Choose a number, type a name, enter to skip, or q to quit.
> """
findUserToSchool = (users) ->
# TODO: don't show users where everyone in the suggestion already has the same school (because we have already done this group)
[bestTarget, bestTargetSuggestions, mostReasons] = [null, [], 0]
for field, groups of topGroups
largestGroup = groups[0]
target = userCategories[field][largestGroup][0]
suggestions = findSuggestions target
reasons = _.reduce suggestions, ((sum, suggestion) -> sum + (if suggestion.schoolName then suggestion.reasons.length else 0)), 0
if reasons > mostReasons
bestTarget = target
bestTargetSuggestions = suggestions
mostReasons = reasons
return [bestTarget, bestTargetSuggestions]
findSuggestions = (target) ->
suggestions = []
if target.lastIP
for otherUser in userCategories.lastIP[target.lastIP] when otherUser isnt target
suggestions.push schoolName: otherUser.schoolName, reasons: ["IP match"], user: otherUser
for leagueType in ['courseInstances', 'clans']
if target[leagueType]?.length
for league in target[leagueType]
for otherUser in userCategories[leagueType][league] when otherUser isnt target
reason = "#{_.str.humanize(leagueType)} match"
if existingSuggestion = _.find(suggestions, user: otherUser)
existingSuggestion.reasons.push reason
else
suggestions.push schoolName: otherUser.schoolName, reasons: [reason], user: otherUser
if target.schoolName?.length > 5
nameMatches = []
for otherSchoolName in topGroups.schoolName
score = stringScore otherSchoolName, target.schoolName, 0.8
continue if score < 0.25
nameMatches.push schoolName: otherSchoolName, score: score
nameMatches = (match.schoolName for match in (_.sortBy nameMatches, (match) -> -match.score))
for match in nameMatches.slice(0, 10)
reason = "Name match"
for otherUser in userCategories.schoolName[match] when otherUser isnt target
if existingSuggestion = _.find(suggestions, user: otherUser)
existingSuggestion.reasons.push reason
else
suggestions.push schoolName: match, reasons: [reason], user: otherUser
return _.uniq suggestions, 'user'
userCategories = {}
topGroups = {}
usersCategorized = {}
sortUsers = (users) ->
users = _.sortBy users, (u) -> -u.points
users = _.sortBy users, ['schoolName', 'lastIP']
# TODO: also match users by shared school email domains when we can identify those
for field in ['courseInstances', 'lastIP', 'schoolName', 'clans']
userCategories[field] = categorizeUsers users, field
topGroups[field] = _.sortBy _.keys(userCategories[field]), (key) -> -userCategories[field][key].length
topGroups[field] = (group for group in topGroups[field] when 2 < userCategories[field][group].length < (if field is 'clans' then 30 else 5000))
categorizeUsers = (users, field) ->
categories = {}
for user in users when value = user[field]
values = if _.isArray(value) then value else [value]
for value in values when value
continue if value.trim and not value.trim()
categories[value] ?= []
categories[value].push user
categories
# https://github.com/joshaven/string_score
stringScore = (_a, word, fuzziness) ->
return 1 if word is _a
return 0 if word is ""
runningScore = 0
string = _a
lString = string.toLowerCase()
strLength = string.length
lWord = word.toLowerCase()
wordLength = word.length
startAt = 0
fuzzies = 1
if fuzziness
fuzzyFactor = 1 - fuzziness
if fuzziness
for i in [0...wordLength]
idxOf = lString.indexOf lWord[i], startAt
if idxOf is -1
fuzzies += fuzzyFactor
else
if startAt is idxOf
charScore = 0.7
else
charScore = 0.1
charScore += 0.8 if string[idxOf - 1] is ' '
charScore += 0.1 if string[idxOf] is word[i]
runningScore += charScore
startAt = idxOf + 1
else
for i in [0...wordLength]
idxOf = lString.indexOf lWord[i], startAt
return 0 if idxOf is -1
if startAt is idxOf
charScore = 0.7
else
charScore = 0.1
charScore += 0.8 if string[idxOf - 1] is word[i]
runningScore += charScore
startAt = idxOf + 1
finalScore = 0.5 * (runningScore / strLength + runningScore / wordLength) / fuzzies
finalScore += 0.15 if lWord[0] is lString[0] and finalScore < 0.85
finalScore
prompt = (question, callback) ->
process.stdin.resume()
process.stdout.write question
process.stdin.once 'data', (data) ->
callback data.toString().trim()