| 1 |
--[[ |
| 2 |
Copyright (c) 2022, Vsevolod Stakhov <[email protected]> |
| 3 |
|
| 4 |
Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 |
you may not use this file except in compliance with the License. |
| 6 |
You may obtain a copy of the License at |
| 7 |
|
| 8 |
http://www.apache.org/licenses/LICENSE-2.0 |
| 9 |
|
| 10 |
Unless required by applicable law or agreed to in writing, software |
| 11 |
distributed under the License is distributed on an "AS IS" BASIS, |
| 12 |
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 |
See the License for the specific language governing permissions and |
| 14 |
limitations under the License. |
| 15 |
]]-- |
| 16 |
|
| 17 |
-- This file contains functions to simplify bayes classifier auto-learning |
| 18 |
|
| 19 |
local lua_util = require "lua_util" |
| 20 |
local lua_verdict = require "lua_verdict" |
| 21 |
local logger = require "rspamd_logger" |
| 22 |
local N = "lua_bayes" |
| 23 |
|
| 24 |
local exports = {} |
| 25 |
|
| 26 |
local function as_set(tbl, case_sensitive) |
| 27 |
if not tbl then |
| 28 |
return nil |
| 29 |
end |
| 30 |
|
| 31 |
local res = {} |
| 32 |
|
| 33 |
for k, v in pairs(tbl) do |
| 34 |
if type(k) == 'number' then |
| 35 |
if type(v) == 'string' then |
| 36 |
res[case_sensitive and v or v:lower()] = true |
| 37 |
else |
| 38 |
res[v] = true |
| 39 |
end |
| 40 |
else |
| 41 |
if type(v) == 'boolean' then |
| 42 |
res[case_sensitive and k or k:lower()] = v |
| 43 |
else |
| 44 |
res[case_sensitive and k or k:lower()] = true |
| 45 |
end |
| 46 |
end |
| 47 |
end |
| 48 |
|
| 49 |
return res |
| 50 |
end |
| 51 |
|
| 52 |
local function merge_options(defaults, module_defaults, overrides) |
| 53 |
local merged = lua_util.override_defaults(defaults, module_defaults or {}) |
| 54 |
|
| 55 |
if overrides then |
| 56 |
merged = lua_util.override_defaults(merged, overrides) |
| 57 |
end |
| 58 |
|
| 59 |
return merged |
| 60 |
end |
| 61 |
|
| 62 |
local function interpret_guard_result(res, reason, extra) |
| 63 |
local ok = res |
| 64 |
local msg = reason |
| 65 |
local meta = extra |
| 66 |
|
| 67 |
if type(res) == 'table' then |
| 68 |
ok = res.ok |
| 69 |
msg = res.reason or res.message or reason |
| 70 |
if res.meta then |
| 71 |
meta = res.meta |
| 72 |
else |
| 73 |
meta = res |
| 74 |
end |
| 75 |
|
| 76 |
if ok == nil then |
| 77 |
if res.allow ~= nil then |
| 78 |
ok = res.allow |
| 79 |
elseif res.block ~= nil then |
| 80 |
ok = not res.block |
| 81 |
end |
| 82 |
end |
| 83 |
|
| 84 |
if ok == nil then |
| 85 |
ok = true |
| 86 |
end |
| 87 |
|
| 88 |
if res.stop ~= nil then |
| 89 |
meta = meta or {} |
| 90 |
meta.stop = res.stop and true or nil |
| 91 |
end |
| 92 |
else |
| 93 |
if ok == nil then |
| 94 |
ok = true |
| 95 |
end |
| 96 |
end |
| 97 |
|
| 98 |
return ok, msg, meta |
| 99 |
end |
| 100 |
|
| 101 |
local function execute_guards(guards, ctx, default_reason) |
| 102 |
if not guards then |
| 103 |
return true |
| 104 |
end |
| 105 |
|
| 106 |
for _, guard in ipairs(guards) do |
| 107 |
local ok, msg, meta = interpret_guard_result(guard.cb(ctx), default_reason, nil) |
| 108 |
|
| 109 |
if not ok then |
| 110 |
if ctx.result then |
| 111 |
ctx.result.guard = guard.name |
| 112 |
ctx.result.reason = msg or default_reason |
| 113 |
if meta then |
| 114 |
ctx.result.meta = meta |
| 115 |
end |
| 116 |
end |
| 117 |
|
| 118 |
return false, msg, meta |
| 119 |
end |
| 120 |
|
| 121 |
if meta and meta.stop then |
| 122 |
return true, msg, meta |
| 123 |
end |
| 124 |
end |
| 125 |
|
| 126 |
return true |
| 127 |
end |
| 128 |
|
| 129 |
local function register_guard(registry, name, cb, opts) |
| 130 |
if type(name) == 'function' then |
| 131 |
cb = name |
| 132 |
name = nil |
| 133 |
end |
| 134 |
|
| 135 |
if type(cb) ~= 'function' then |
| 136 |
return nil, 'guard callback must be a function' |
| 137 |
end |
| 138 |
|
| 139 |
local guard = { |
| 140 |
name = name or string.format('guard_%d', (#registry) + 1), |
| 141 |
cb = cb, |
| 142 |
priority = opts and opts.priority or 0, |
| 143 |
} |
| 144 |
|
| 145 |
registry[#registry + 1] = guard |
| 146 |
table.sort(registry, function(a, b) |
| 147 |
return (a.priority or 0) > (b.priority or 0) |
| 148 |
end) |
| 149 |
|
| 150 |
return guard.name |
| 151 |
end |
| 152 |
|
| 153 |
local function unregister_guard(registry, name) |
| 154 |
for i = #registry, 1, -1 do |
| 155 |
if registry[i].name == name then |
| 156 |
table.remove(registry, i) |
| 157 |
end |
| 158 |
end |
| 159 |
end |
| 160 |
|
| 161 |
local can_learn_defaults = {} |
| 162 |
local autolearn_defaults = {} |
| 163 |
|
| 164 |
local can_learn_guards = {} |
| 165 |
local autolearn_guards = {} |
| 166 |
|
| 167 |
local default_can_learn_settings = { |
| 168 |
bypass_header = { |
| 169 |
enabled = true, |
| 170 |
header = 'Learn-Type', |
| 171 |
values = {'bulk'}, |
| 172 |
behaviour = 'skip_probability', |
| 173 |
case_sensitive = false, |
| 174 |
}, |
| 175 |
probability_check = { |
| 176 |
enabled = true, |
| 177 |
-- 'can_learn_prob' is written by the C layer (rspamd_stat_classifier_is_skipped) |
| 178 |
-- before invoking this condition. It holds the per-classifier, per-class |
| 179 |
-- probability so each classifier's can_learn decision is independent. |
| 180 |
variable = 'can_learn_prob', |
| 181 |
ctype = 'double', |
| 182 |
-- Unified threshold: >= min_prob means "already in this class, skip learning". |
| 183 |
-- Replaces the old asymmetric spam_min/ham_max pair. |
| 184 |
min_prob = 0.95, |
| 185 |
skip_for_unlearn = false, |
| 186 |
require_value = false, |
| 187 |
}, |
| 188 |
} |
| 189 |
|
| 190 |
local default_autolearn_settings = { |
| 191 |
require_queue_id = true, |
| 192 |
check_local = true, |
| 193 |
check_authed = true, |
| 194 |
verdict_source = { |
| 195 |
name = 'bayes', |
| 196 |
}, |
| 197 |
logging = { |
| 198 |
enabled = true, |
| 199 |
}, |
| 200 |
thresholds = { |
| 201 |
spam = nil, |
| 202 |
ham = nil, |
| 203 |
junk = nil, |
| 204 |
}, |
| 205 |
learn_verdict = false, |
| 206 |
balance = { |
| 207 |
enabled = false, |
| 208 |
min_balance = 0.9, |
| 209 |
spam_key = 'spam_learns', |
| 210 |
ham_key = 'ham_learns', |
| 211 |
value_type = 'int64', |
| 212 |
}, |
| 213 |
} |
| 214 |
|
| 215 |
--- Allows global overrides for can_learn defaults (e.g. from config) |
| 216 |
-- @param opts table of default overrides |
| 217 |
exports.configure_can_learn = function(opts) |
| 218 |
if opts then |
| 219 |
can_learn_defaults = lua_util.override_defaults(can_learn_defaults, opts) |
| 220 |
end |
| 221 |
end |
| 222 |
|
| 223 |
--- Allows global overrides for autolearn defaults (e.g. from config) |
| 224 |
-- @param opts table of default overrides |
| 225 |
exports.configure_autolearn = function(opts) |
| 226 |
if opts then |
| 227 |
autolearn_defaults = lua_util.override_defaults(autolearn_defaults, opts) |
| 228 |
end |
| 229 |
end |
| 230 |
|
| 231 |
--- Register an extra guard for can_learn checks |
| 232 |
-- @param name string or callback (name optional) |
| 233 |
-- @param cb guard callback |
| 234 |
-- @param opts optional table with priority |
| 235 |
exports.register_can_learn_guard = function(name, cb, opts) |
| 236 |
return register_guard(can_learn_guards, name, cb, opts) |
| 237 |
end |
| 238 |
|
| 239 |
--- Remove a previously registered can_learn guard by name |
| 240 |
exports.unregister_can_learn_guard = function(name) |
| 241 |
unregister_guard(can_learn_guards, name) |
| 242 |
end |
| 243 |
|
| 244 |
--- Register an extra guard for autolearn decisions |
| 245 |
-- @param name string or callback (name optional) |
| 246 |
-- @param cb guard callback |
| 247 |
-- @param opts optional table with priority |
| 248 |
exports.register_autolearn_guard = function(name, cb, opts) |
| 249 |
return register_guard(autolearn_guards, name, cb, opts) |
| 250 |
end |
| 251 |
|
| 252 |
--- Remove a previously registered autolearn guard by name |
| 253 |
exports.unregister_autolearn_guard = function(name) |
| 254 |
unregister_guard(autolearn_guards, name) |
| 255 |
end |
| 256 |
|
| 257 |
|
| 258 |
--- Determines if a message can be learned by Bayes |
| 259 |
-- @param task rspamd_task |
| 260 |
-- @param is_spam boolean indicates target class |
| 261 |
-- @param is_unlearn boolean indicates unlearn operation |
| 262 |
-- @param overrides optional per-call overrides |
| 263 |
exports.can_learn = function(task, is_spam, is_unlearn, overrides) |
| 264 |
local opts = merge_options(default_can_learn_settings, can_learn_defaults, overrides) |
| 265 |
|
| 266 |
if opts.bypass_header and opts.bypass_header.values then |
| 267 |
opts.bypass_header._set = as_set(opts.bypass_header.values, opts.bypass_header.case_sensitive) |
| 268 |
end |
| 269 |
|
| 270 |
local ctx = { |
| 271 |
task = task, |
| 272 |
is_spam = is_spam, |
| 273 |
is_unlearn = is_unlearn, |
| 274 |
options = opts, |
| 275 |
state = {}, |
| 276 |
result = {}, |
| 277 |
} |
| 278 |
|
| 279 |
if overrides and overrides.guards then |
| 280 |
local ok, msg, meta = execute_guards(overrides.guards, ctx, 'blocked by can_learn guard') |
| 281 |
|
| 282 |
if not ok then |
| 283 |
return false, msg, ctx.result |
| 284 |
end |
| 285 |
|
| 286 |
if meta and meta.stop then |
| 287 |
return true, nil, ctx.result |
| 288 |
end |
| 289 |
end |
| 290 |
|
| 291 |
local ok, msg = execute_guards(can_learn_guards, ctx, 'blocked by can_learn guard') |
| 292 |
|
| 293 |
if not ok then |
| 294 |
return false, msg, ctx.result |
| 295 |
end |
| 296 |
|
| 297 |
local probability_opts = opts.probability_check |
| 298 |
local skip_probability = false |
| 299 |
|
| 300 |
if opts.bypass_header and opts.bypass_header.enabled ~= false then |
| 301 |
local header_name = opts.bypass_header.header or 'Learn-Type' |
| 302 |
local header_value = task:get_request_header(header_name) |
| 303 |
|
| 304 |
if header_value then |
| 305 |
header_value = tostring(header_value) |
| 306 |
if opts.bypass_header.case_sensitive ~= true then |
| 307 |
header_value = header_value:lower() |
| 308 |
end |
| 309 |
|
| 310 |
local matched |
| 311 |
if opts.bypass_header._set then |
| 312 |
matched = opts.bypass_header._set[header_value] |
| 313 |
else |
| 314 |
matched = false |
| 315 |
end |
| 316 |
|
| 317 |
if matched then |
| 318 |
if opts.bypass_header.behaviour == 'allow' then |
| 319 |
ctx.result.reason = 'bypass header matched' |
| 320 |
return true, nil, ctx.result |
| 321 |
elseif opts.bypass_header.behaviour == 'deny' then |
| 322 |
ctx.result.reason = opts.bypass_header.reason or 'bypass header denies learning' |
| 323 |
ctx.result.guard = 'bypass_header' |
| 324 |
return false, ctx.result.reason, ctx.result |
| 325 |
else |
| 326 |
skip_probability = true |
| 327 |
end |
| 328 |
end |
| 329 |
end |
| 330 |
end |
| 331 |
|
| 332 |
if ctx.state.skip_probability ~= nil then |
| 333 |
skip_probability = ctx.state.skip_probability |
| 334 |
end |
| 335 |
|
| 336 |
if probability_opts and probability_opts.enabled ~= false and not skip_probability then |
| 337 |
if is_unlearn and probability_opts.skip_for_unlearn then |
| 338 |
ctx.result.reason = 'probability check skipped for unlearn' |
| 339 |
else |
| 340 |
local prob |
| 341 |
|
| 342 |
if probability_opts.resolver and type(probability_opts.resolver) == 'function' then |
| 343 |
prob = probability_opts.resolver(ctx) |
| 344 |
else |
| 345 |
prob = task:get_mempool():get_variable(probability_opts.variable or 'bayes_prob', |
| 346 |
probability_opts.ctype or 'double') |
| 347 |
end |
| 348 |
|
| 349 |
ctx.result.probability = prob |
| 350 |
|
| 351 |
if prob == nil then |
| 352 |
if probability_opts.require_value then |
| 353 |
local reason = probability_opts.missing_reason or 'probability value is missing' |
| 354 |
ctx.result.guard = 'probability_check' |
| 355 |
ctx.result.reason = reason |
| 356 |
return false, reason, ctx.result |
| 357 |
end |
| 358 |
else |
| 359 |
local in_class |
| 360 |
local guard_msg |
| 361 |
|
| 362 |
if probability_opts.check and type(probability_opts.check) == 'function' then |
| 363 |
in_class, guard_msg = probability_opts.check(ctx, prob) |
| 364 |
else |
| 365 |
-- Unified check: high probability means the message is already confidently |
| 366 |
-- in the target class (works for both binary spam/ham and multiclass). |
| 367 |
-- can_learn_prob is set per-classifier by C so there is no cross-contamination. |
| 368 |
in_class = prob >= (probability_opts.min_prob or |
| 369 |
probability_opts.spam_min or 0.95) |
| 370 |
end |
| 371 |
|
| 372 |
if in_class then |
| 373 |
-- class name is written by C before invoking this condition |
| 374 |
local cl = task:get_mempool():get_variable('can_learn_class') or |
| 375 |
(is_spam and 'spam' or 'ham') |
| 376 |
local reason |
| 377 |
|
| 378 |
if probability_opts.message_formatter and type(probability_opts.message_formatter) == 'function' then |
| 379 |
reason = probability_opts.message_formatter(ctx, prob, cl) or guard_msg |
| 380 |
end |
| 381 |
|
| 382 |
reason = reason or guard_msg or |
| 383 |
string.format('already in class %s; probability %.2f%%', cl, prob * 100.0) |
| 384 |
|
| 385 |
ctx.result.guard = 'probability_check' |
| 386 |
ctx.result.reason = reason |
| 387 |
|
| 388 |
return false, reason, ctx.result |
| 389 |
end |
| 390 |
end |
| 391 |
end |
| 392 |
end |
| 393 |
|
| 394 |
ctx.result.guard = ctx.result.guard or 'can_learn' |
| 395 |
|
| 396 |
return true, nil, ctx.result |
| 397 |
end |
| 398 |
|
| 399 |
--- Decide if a message should be auto-learned and return class |
| 400 |
-- @param task rspamd_task |
| 401 |
-- @param conf classifier autolearn configuration |
| 402 |
-- @param overrides optional per-call overrides |
| 403 |
exports.autolearn = function(task, conf, overrides) |
| 404 |
local opts = merge_options(default_autolearn_settings, autolearn_defaults, overrides) |
| 405 |
opts = merge_options(opts, {}, conf) |
| 406 |
|
| 407 |
if opts.check_balance ~= nil then |
| 408 |
opts.balance = opts.balance or {} |
| 409 |
if opts.balance.enabled == nil then |
| 410 |
opts.balance.enabled = opts.check_balance and true or false |
| 411 |
end |
| 412 |
end |
| 413 |
|
| 414 |
if opts.min_balance ~= nil then |
| 415 |
opts.balance = opts.balance or {} |
| 416 |
if opts.balance.min_balance == nil then |
| 417 |
opts.balance.min_balance = opts.min_balance |
| 418 |
end |
| 419 |
end |
| 420 |
|
| 421 |
local external_options = opts.options |
| 422 |
if external_options ~= nil then |
| 423 |
opts.options = nil |
| 424 |
|
| 425 |
if type(external_options) == 'function' then |
| 426 |
local ok, res = pcall(external_options, task, opts) |
| 427 |
|
| 428 |
if ok and type(res) == 'table' then |
| 429 |
opts = merge_options(opts, {}, res) |
| 430 |
else |
| 431 |
lua_util.debugm(N, task, 'autolearn options callback failed: %s', res) |
| 432 |
end |
| 433 |
elseif type(external_options) == 'table' then |
| 434 |
opts = merge_options(opts, {}, external_options) |
| 435 |
else |
| 436 |
lua_util.debugm(N, task, 'autolearn options must be a table or function, got %s', |
| 437 |
type(external_options)) |
| 438 |
end |
| 439 |
end |
| 440 |
|
| 441 |
local ctx = { |
| 442 |
task = task, |
| 443 |
conf = opts, |
| 444 |
state = {}, |
| 445 |
result = {}, |
| 446 |
} |
| 447 |
|
| 448 |
if overrides and overrides.guards then |
| 449 |
local ok, msg, meta = execute_guards(overrides.guards, ctx, 'blocked by autolearn guard') |
| 450 |
|
| 451 |
if not ok then |
| 452 |
return nil, msg, ctx.result |
| 453 |
end |
| 454 |
|
| 455 |
if meta and meta.stop then |
| 456 |
return ctx.result.decision, msg, ctx.result |
| 457 |
end |
| 458 |
end |
| 459 |
|
| 460 |
local ok, msg = execute_guards(autolearn_guards, ctx, 'blocked by autolearn guard') |
| 461 |
|
| 462 |
if not ok then |
| 463 |
return nil, msg, ctx.result |
| 464 |
end |
| 465 |
|
| 466 |
if opts.require_queue_id and not task:get_queue_id() then |
| 467 |
lua_util.debugm(N, task, 'no need to autolearn - queue id is missing') |
| 468 |
ctx.result.reason = 'queue id is missing' |
| 469 |
return nil, ctx.result.reason, ctx.result |
| 470 |
end |
| 471 |
|
| 472 |
local skip_conf = {opts.check_local, opts.check_authed} |
| 473 |
if lua_util.is_skip_local_or_authed(task, skip_conf) then |
| 474 |
lua_util.debugm(N, task, 'skip autolearn for local or authed network') |
| 475 |
ctx.result.reason = 'local or authed network' |
| 476 |
return nil, ctx.result.reason, ctx.result |
| 477 |
end |
| 478 |
|
| 479 |
local verdict_source = opts.verdict_source or {} |
| 480 |
local verdict, score |
| 481 |
|
| 482 |
if verdict_source.extractor and type(verdict_source.extractor) == 'function' then |
| 483 |
verdict, score = verdict_source.extractor(ctx) |
| 484 |
else |
| 485 |
verdict, score = lua_verdict.get_specific_verdict(verdict_source.name or 'bayes', task) |
| 486 |
end |
| 487 |
|
| 488 |
ctx.result.verdict = verdict |
| 489 |
ctx.result.score = score |
| 490 |
|
| 491 |
if verdict == 'passthrough' then |
| 492 |
lua_util.debugm(N, task, 'no need to autolearn - verdict: %s', verdict) |
| 493 |
ctx.result.reason = 'verdict passthrough' |
| 494 |
return nil, ctx.result.reason, ctx.result |
| 495 |
end |
| 496 |
|
| 497 |
local learn_spam, learn_ham = false, false |
| 498 |
|
| 499 |
local thresholds = opts.thresholds or {} |
| 500 |
thresholds.spam = thresholds.spam or opts.spam_threshold |
| 501 |
thresholds.ham = thresholds.ham or opts.ham_threshold |
| 502 |
thresholds.junk = thresholds.junk or opts.junk_threshold |
| 503 |
|
| 504 |
local log_opts = opts.logging or {} |
| 505 |
local function log_can_autolearn(verdict_name, score_value, threshold) |
| 506 |
if log_opts.enabled == false then |
| 507 |
return |
| 508 |
end |
| 509 |
|
| 510 |
local from = task:get_from('smtp') |
| 511 |
local mime_rcpts = 'undef' |
| 512 |
local mr = task:get_recipients('mime') |
| 513 |
if mr then |
| 514 |
local r_addrs = {} |
| 515 |
for _, r in ipairs(mr) do |
| 516 |
r_addrs[#r_addrs + 1] = r.addr |
| 517 |
end |
| 518 |
if #r_addrs > 0 then |
| 519 |
mime_rcpts = table.concat(r_addrs, ',') |
| 520 |
end |
| 521 |
end |
| 522 |
|
| 523 |
logger.info(task, 'id: %s, from: <%s>: can autolearn %s: score %s %s %s, mime_rcpts: <%s>', |
| 524 |
task:get_header('Message-Id') or '<undef>', |
| 525 |
from and from[1].addr or 'undef', |
| 526 |
verdict_name, |
| 527 |
string.format('%.2f', score_value or 0), |
| 528 |
verdict_name == 'ham' and '<=' or verdict_name == 'spam' and '>=' or '/', |
| 529 |
threshold, |
| 530 |
mime_rcpts) |
| 531 |
end |
| 532 |
|
| 533 |
if thresholds.spam and thresholds.ham then |
| 534 |
if verdict == 'spam' then |
| 535 |
if score and score >= thresholds.spam then |
| 536 |
log_can_autolearn(verdict, score, thresholds.spam) |
| 537 |
learn_spam = true |
| 538 |
end |
| 539 |
elseif verdict == 'junk' then |
| 540 |
if thresholds.junk and score and score >= thresholds.junk then |
| 541 |
log_can_autolearn(verdict, score, thresholds.junk) |
| 542 |
learn_spam = true |
| 543 |
end |
| 544 |
elseif verdict == 'ham' then |
| 545 |
if score and score <= thresholds.ham then |
| 546 |
log_can_autolearn(verdict, score, thresholds.ham) |
| 547 |
learn_ham = true |
| 548 |
end |
| 549 |
end |
| 550 |
elseif opts.learn_verdict then |
| 551 |
if verdict == 'spam' or verdict == 'junk' then |
| 552 |
learn_spam = true |
| 553 |
elseif verdict == 'ham' then |
| 554 |
learn_ham = true |
| 555 |
end |
| 556 |
elseif opts.evaluate and type(opts.evaluate) == 'function' then |
| 557 |
local decision = opts.evaluate(ctx) |
| 558 |
|
| 559 |
if decision == 'spam' then |
| 560 |
learn_spam = true |
| 561 |
elseif decision == 'ham' then |
| 562 |
learn_ham = true |
| 563 |
end |
| 564 |
end |
| 565 |
|
| 566 |
if opts.balance and opts.balance.enabled then |
| 567 |
local balance_opts = opts.balance |
| 568 |
local spam_learns = task:get_mempool():get_variable(balance_opts.spam_key or 'spam_learns', balance_opts.value_type or 'int64') or 0 |
| 569 |
local ham_learns = task:get_mempool():get_variable(balance_opts.ham_key or 'ham_learns', balance_opts.value_type or 'int64') or 0 |
| 570 |
|
| 571 |
local min_balance = balance_opts.min_balance or 0.9 |
| 572 |
|
| 573 |
if spam_learns > 0 or ham_learns > 0 then |
| 574 |
local max_ratio = 1.0 / min_balance |
| 575 |
local spam_learns_ratio = spam_learns / (ham_learns + 1) |
| 576 |
if spam_learns_ratio > max_ratio and learn_spam then |
| 577 |
lua_util.debugm(N, task, |
| 578 |
'skip learning spam, balance is not satisfied: %s < %s; %s spam learns; %s ham learns', |
| 579 |
spam_learns_ratio, min_balance, spam_learns, ham_learns) |
| 580 |
learn_spam = false |
| 581 |
ctx.result.reason = 'spam balance check failed' |
| 582 |
end |
| 583 |
|
| 584 |
local ham_learns_ratio = ham_learns / (spam_learns + 1) |
| 585 |
if ham_learns_ratio > max_ratio and learn_ham then |
| 586 |
lua_util.debugm(N, task, |
| 587 |
'skip learning ham, balance is not satisfied: %s < %s; %s spam learns; %s ham learns', |
| 588 |
ham_learns_ratio, min_balance, spam_learns, ham_learns) |
| 589 |
learn_ham = false |
| 590 |
ctx.result.reason = 'ham balance check failed' |
| 591 |
end |
| 592 |
end |
| 593 |
end |
| 594 |
|
| 595 |
if learn_spam then |
| 596 |
ctx.result.decision = 'spam' |
| 597 |
return 'spam', nil, ctx.result |
| 598 |
elseif learn_ham then |
| 599 |
ctx.result.decision = 'ham' |
| 600 |
return 'ham', nil, ctx.result |
| 601 |
end |
| 602 |
|
| 603 |
return nil, ctx.result.reason, ctx.result |
| 604 |
end |
| 605 |
|
| 606 |
return exports |