mox/junk.go
Mechiel Lukkien 6aa2139a54
do not use results from junk filter if we have less than 50 positive classifications to base the decision on
useful for new accounts. we don't want to start rejecting incoming messages for
having a score near 0.5 because of too little training material. we err on the
side of allowing messages in. the user will mark them as junk, training the
filter. once enough non-junk has come in, we'll start the actual filtering.

for issue #64 by x8x, and i've also seen this concern on matrix
2025-01-23 22:55:50 +01:00

433 lines
13 KiB
Go

package main
/*
note: these testdata paths are not in the repo, you should gather some of your
own ham/spam emails.
./mox junk train testdata/train/ham testdata/train/spam
./mox junk train -sent-dir testdata/sent testdata/train/ham testdata/train/spam
./mox junk check 'testdata/check/ham/mail1'
./mox junk test testdata/check/ham testdata/check/spam
./mox junk analyze testdata/train/ham testdata/train/spam
./mox junk analyze -top-words 10 -train-ratio 0.5 -spam-threshold 0.85 -max-power 0.01 -sent-dir testdata/sent testdata/train/ham testdata/train/spam
./mox junk play -top-words 10 -train-ratio 0.5 -spam-threshold 0.85 -max-power 0.01 -sent-dir testdata/sent testdata/train/ham testdata/train/spam
*/
import (
"context"
"flag"
"fmt"
"log"
mathrand "math/rand"
"os"
"path/filepath"
"sort"
"time"
"github.com/mjl-/mox/junk"
"github.com/mjl-/mox/message"
"github.com/mjl-/mox/mlog"
"github.com/mjl-/mox/mox-"
)
type junkArgs struct {
params junk.Params
spamThreshold float64
trainRatio float64
seed bool
sentDir string
databasePath, bloomfilterPath string
debug bool
}
func (a junkArgs) SetLogLevel() {
mox.Conf.Log[""] = mlog.LevelInfo
if a.debug {
mox.Conf.Log[""] = mlog.LevelDebug
}
mlog.SetConfig(mox.Conf.Log)
}
func junkFlags(fs *flag.FlagSet) (a junkArgs) {
fs.BoolVar(&a.params.Onegrams, "one-grams", false, "use 1-grams, i.e. single words, for scoring")
fs.BoolVar(&a.params.Twograms, "two-grams", true, "use 2-grams, i.e. word pairs, for scoring")
fs.BoolVar(&a.params.Threegrams, "three-grams", false, "use 3-grams, i.e. word triplets, for scoring")
fs.Float64Var(&a.params.MaxPower, "max-power", 0.05, "maximum word power, e.g. min 0.05/max 0.95")
fs.Float64Var(&a.params.IgnoreWords, "ignore-words", 0.1, "ignore words with ham/spaminess within this distance from 0.5")
fs.IntVar(&a.params.TopWords, "top-words", 10, "number of top spam and number of top ham words from email to use")
fs.IntVar(&a.params.RareWords, "rare-words", 1, "words are rare if encountered this number during training, and skipped for scoring")
fs.BoolVar(&a.debug, "debug", false, "print debug logging when calculating spam probability")
fs.Float64Var(&a.spamThreshold, "spam-threshold", 0.95, "probability where message is seen as spam")
fs.Float64Var(&a.trainRatio, "train-ratio", 0.5, "part of data to use for training versus analyzing (for analyze only)")
fs.StringVar(&a.sentDir, "sent-dir", "", "directory with sent mails, for training")
fs.BoolVar(&a.seed, "seed", false, "seed prng before analysis")
fs.StringVar(&a.databasePath, "dbpath", "filter.db", "database file for ham/spam words")
fs.StringVar(&a.bloomfilterPath, "bloompath", "filter.bloom", "bloom filter for ignoring unique strings")
return
}
func listDir(dir string) (l []string) {
files, err := os.ReadDir(dir)
xcheckf(err, "listing directory %q", dir)
for _, f := range files {
l = append(l, f.Name())
}
return l
}
func must(f *junk.Filter, err error) *junk.Filter {
xcheckf(err, "filter")
return f
}
func cmdJunkTrain(c *cmd) {
c.unlisted = true
c.params = "hamdir spamdir"
c.help = "Train a junk filter with messages from hamdir and spamdir."
a := junkFlags(c.flag)
args := c.Parse()
if len(args) != 2 {
c.Usage()
}
a.SetLogLevel()
f := must(junk.NewFilter(context.Background(), c.log, a.params, a.databasePath, a.bloomfilterPath))
defer func() {
if err := f.Close(); err != nil {
log.Printf("closing junk filter: %v", err)
}
}()
hamFiles := listDir(args[0])
spamFiles := listDir(args[1])
var sentFiles []string
if a.sentDir != "" {
sentFiles = listDir(a.sentDir)
}
err := f.TrainDirs(args[0], a.sentDir, args[1], hamFiles, sentFiles, spamFiles)
xcheckf(err, "train")
}
func cmdJunkCheck(c *cmd) {
c.unlisted = true
c.params = "mailfile"
c.help = "Check an email message against a junk filter, printing the probability of spam on a scale from 0 to 1."
a := junkFlags(c.flag)
args := c.Parse()
if len(args) != 1 {
c.Usage()
}
a.SetLogLevel()
f := must(junk.OpenFilter(context.Background(), c.log, a.params, a.databasePath, a.bloomfilterPath, false))
defer func() {
if err := f.Close(); err != nil {
log.Printf("closing junk filter: %v", err)
}
}()
result, err := f.ClassifyMessagePath(context.Background(), args[0])
xcheckf(err, "testing mail")
sig := "significant"
if !result.Significant {
sig = "not significant"
}
fmt.Printf("%.6f, %s\n", result.Probability, sig)
}
func cmdJunkTest(c *cmd) {
c.unlisted = true
c.params = "hamdir spamdir"
c.help = "Check a directory with hams and one with spams against the junk filter, and report the success ratio."
a := junkFlags(c.flag)
args := c.Parse()
if len(args) != 2 {
c.Usage()
}
a.SetLogLevel()
f := must(junk.OpenFilter(context.Background(), c.log, a.params, a.databasePath, a.bloomfilterPath, false))
defer func() {
if err := f.Close(); err != nil {
log.Printf("closing junk filter: %v", err)
}
}()
testDir := func(dir string, ham bool) (int, int) {
ok, bad := 0, 0
files, err := os.ReadDir(dir)
xcheckf(err, "readdir %q", dir)
for _, fi := range files {
path := filepath.Join(dir, fi.Name())
result, err := f.ClassifyMessagePath(context.Background(), path)
if err != nil {
log.Printf("classify message %q: %s", path, err)
continue
}
if ham && result.Probability < a.spamThreshold || !ham && result.Probability > a.spamThreshold {
ok++
} else {
bad++
}
if ham && result.Probability > a.spamThreshold {
fmt.Printf("ham %q: %.4f\n", path, result.Probability)
}
if !ham && result.Probability < a.spamThreshold {
fmt.Printf("spam %q: %.4f\n", path, result.Probability)
}
}
return ok, bad
}
nhamok, nhambad := testDir(args[0], true)
nspamok, nspambad := testDir(args[1], false)
fmt.Printf("total ham, ok %d, bad %d\n", nhamok, nhambad)
fmt.Printf("total spam, ok %d, bad %d\n", nspamok, nspambad)
fmt.Printf("specifity (true negatives, hams identified): %.6f\n", float64(nhamok)/(float64(nhamok+nhambad)))
fmt.Printf("sensitivity (true positives, spams identified): %.6f\n", float64(nspamok)/(float64(nspamok+nspambad)))
fmt.Printf("accuracy: %.6f\n", float64(nhamok+nspamok)/float64(nhamok+nhambad+nspamok+nspambad))
}
func cmdJunkAnalyze(c *cmd) {
c.unlisted = true
c.params = "hamdir spamdir"
c.help = `Analyze a directory with ham messages and one with spam messages.
A part of the messages is used for training, and remaining for testing. The
messages are shuffled, with optional random seed.`
a := junkFlags(c.flag)
args := c.Parse()
if len(args) != 2 {
c.Usage()
}
a.SetLogLevel()
f := must(junk.NewFilter(context.Background(), c.log, a.params, a.databasePath, a.bloomfilterPath))
defer func() {
if err := f.Close(); err != nil {
log.Printf("closing junk filter: %v", err)
}
}()
hamDir := args[0]
spamDir := args[1]
hamFiles := listDir(hamDir)
spamFiles := listDir(spamDir)
var seed int64
if a.seed {
seed = time.Now().UnixMilli()
}
// Still at math/rand (v1 instead of v2) for potential comparison to earlier test results.
rand := mathrand.New(mathrand.NewSource(seed))
shuffle := func(l []string) {
count := len(l)
for i := range l {
n := rand.Intn(count)
l[i], l[n] = l[n], l[i]
}
}
shuffle(hamFiles)
shuffle(spamFiles)
ntrainham := int(a.trainRatio * float64(len(hamFiles)))
ntrainspam := int(a.trainRatio * float64(len(spamFiles)))
trainHam := hamFiles[:ntrainham]
trainSpam := spamFiles[:ntrainspam]
testHam := hamFiles[ntrainham:]
testSpam := spamFiles[ntrainspam:]
var trainSent []string
if a.sentDir != "" {
trainSent = listDir(a.sentDir)
}
err := f.TrainDirs(hamDir, a.sentDir, spamDir, trainHam, trainSent, trainSpam)
xcheckf(err, "train")
testDir := func(dir string, files []string, ham bool) (ok, bad, malformed int) {
for _, name := range files {
path := filepath.Join(dir, name)
result, err := f.ClassifyMessagePath(context.Background(), path)
if err != nil {
// log.Infof("%s: %s", path, err)
malformed++
continue
}
if ham && result.Probability < a.spamThreshold || !ham && result.Probability > a.spamThreshold {
ok++
} else {
bad++
}
if ham && result.Probability > a.spamThreshold {
fmt.Printf("ham %q: %.4f\n", path, result.Probability)
}
if !ham && result.Probability < a.spamThreshold {
fmt.Printf("spam %q: %.4f\n", path, result.Probability)
}
}
return
}
nhamok, nhambad, nmalformedham := testDir(args[0], testHam, true)
nspamok, nspambad, nmalformedspam := testDir(args[1], testSpam, false)
fmt.Printf("training done, nham %d, nsent %d, nspam %d\n", ntrainham, len(trainSent), ntrainspam)
fmt.Printf("total ham, ok %d, bad %d, malformed %d\n", nhamok, nhambad, nmalformedham)
fmt.Printf("total spam, ok %d, bad %d, malformed %d\n", nspamok, nspambad, nmalformedspam)
fmt.Printf("specifity (true negatives, hams identified): %.6f\n", float64(nhamok)/(float64(nhamok+nhambad)))
fmt.Printf("sensitivity (true positives, spams identified): %.6f\n", float64(nspamok)/(float64(nspamok+nspambad)))
fmt.Printf("accuracy: %.6f\n", float64(nhamok+nspamok)/float64(nhamok+nhambad+nspamok+nspambad))
}
func cmdJunkPlay(c *cmd) {
c.unlisted = true
c.params = "hamdir spamdir"
c.help = "Play messages from ham and spam directory according to their time of arrival and report on junk filter performance."
a := junkFlags(c.flag)
args := c.Parse()
if len(args) != 2 {
c.Usage()
}
a.SetLogLevel()
f := must(junk.NewFilter(context.Background(), c.log, a.params, a.databasePath, a.bloomfilterPath))
defer func() {
if err := f.Close(); err != nil {
log.Printf("closing junk filter: %v", err)
}
}()
// We'll go through all emails to find their dates.
type msg struct {
dir, filename string
ham, sent bool
t time.Time
}
var msgs []msg
var nbad, nnodate, nham, nspam, nsent int
scanDir := func(dir string, ham, sent bool) {
for _, name := range listDir(dir) {
path := filepath.Join(dir, name)
mf, err := os.Open(path)
xcheckf(err, "open %q", path)
fi, err := mf.Stat()
xcheckf(err, "stat %q", path)
p, err := message.EnsurePart(c.log.Logger, false, mf, fi.Size())
if err != nil {
nbad++
if err := mf.Close(); err != nil {
log.Printf("closing message file: %v", err)
}
continue
}
if p.Envelope.Date.IsZero() {
nnodate++
if err := mf.Close(); err != nil {
log.Printf("closing message file: %v", err)
}
continue
}
if err := mf.Close(); err != nil {
log.Printf("closing message file: %v", err)
}
msgs = append(msgs, msg{dir, name, ham, sent, p.Envelope.Date})
if sent {
nsent++
} else if ham {
nham++
} else {
nspam++
}
}
}
hamDir := args[0]
spamDir := args[1]
scanDir(hamDir, true, false)
scanDir(spamDir, false, false)
if a.sentDir != "" {
scanDir(a.sentDir, true, true)
}
// Sort the messages, earliest first.
sort.Slice(msgs, func(i, j int) bool {
return msgs[i].t.Before(msgs[j].t)
})
// Play all messages as if they are coming in. We predict their spaminess, check if
// we are right. And we train the system with the result.
var nhamok, nhambad, nspamok, nspambad int
play := func(msg msg) {
var words map[string]struct{}
path := filepath.Join(msg.dir, msg.filename)
if !msg.sent {
result, err := f.ClassifyMessagePath(context.Background(), path)
if err != nil {
nbad++
return
}
if msg.ham {
if result.Probability < a.spamThreshold {
nhamok++
} else {
nhambad++
}
} else {
if result.Probability > a.spamThreshold {
nspamok++
} else {
nspambad++
}
}
} else {
mf, err := os.Open(path)
xcheckf(err, "open %q", path)
defer func() {
if err := mf.Close(); err != nil {
log.Printf("closing message file: %v", err)
}
}()
fi, err := mf.Stat()
xcheckf(err, "stat %q", path)
p, err := message.EnsurePart(c.log.Logger, false, mf, fi.Size())
if err != nil {
log.Printf("bad sent message %q: %s", path, err)
return
}
words, err = f.ParseMessage(p)
if err != nil {
log.Printf("bad sent message %q: %s", path, err)
return
}
}
if err := f.Train(context.Background(), msg.ham, words); err != nil {
log.Printf("train: %s", err)
}
}
for _, m := range msgs {
play(m)
}
err := f.Save()
xcheckf(err, "saving filter")
fmt.Printf("completed, nham %d, nsent %d, nspam %d, nbad %d, nwithoutdate %d\n", nham, nsent, nspam, nbad, nnodate)
fmt.Printf("total ham, ok %d, bad %d\n", nhamok, nhambad)
fmt.Printf("total spam, ok %d, bad %d\n", nspamok, nspambad)
fmt.Printf("specifity (true negatives, hams identified): %.6f\n", float64(nhamok)/(float64(nhamok+nhambad)))
fmt.Printf("sensitivity (true positives, spams identified): %.6f\n", float64(nspamok)/(float64(nspamok+nspambad)))
fmt.Printf("accuracy: %.6f\n", float64(nhamok+nspamok)/float64(nhamok+nhambad+nspamok+nspambad))
}