diff options
| author | Vsevolod Stakhov <vsevolod@highsecure.ru> | 2015-05-05 18:18:25 +0100 |
|---|---|---|
| committer | Vsevolod Stakhov <vsevolod@highsecure.ru> | 2015-05-05 18:18:25 +0100 |
| commit | 8d19881b31eaf63b1356fffbadb66215f8c7d0ad (patch) | |
| tree | c09d74c4fb2bf089c7272180c73314d369a321f3 /src/libstat/classifiers | |
| parent | 851a2f6af747768d96b264ee1eac820e66b2cfae (diff) | |
| download | rspamd-8d19881b31eaf63b1356fffbadb66215f8c7d0ad.tar.gz rspamd-8d19881b31eaf63b1356fffbadb66215f8c7d0ad.zip | |
Take OSB feature multiplier into account.
Diffstat (limited to 'src/libstat/classifiers')
| -rw-r--r-- | src/libstat/classifiers/bayes.c | 14 |
1 files changed, 11 insertions, 3 deletions
diff --git a/src/libstat/classifiers/bayes.c b/src/libstat/classifiers/bayes.c index acd04f602..78d112dd0 100644 --- a/src/libstat/classifiers/bayes.c +++ b/src/libstat/classifiers/bayes.c @@ -79,6 +79,9 @@ inv_chi_square (gdouble value, gint freedom_deg) return MIN (1.0, sum); } +static const double feature_weight[] = { 0, 3125, 256, 27, 4, 1 }; + +#define PROB_COMBINE(prob, cnt, weight, assumed) (((weight) * (assumed) + (cnt) * (prob)) / ((weight) + (cnt))) /* * In this callback we calculate local probabilities for tokens */ @@ -90,7 +93,8 @@ bayes_classify_callback (gpointer key, gpointer value, gpointer data) guint i; struct rspamd_token_result *res; guint64 spam_count = 0, ham_count = 0, total_count = 0; - double spam_prob, spam_freq, ham_freq, bayes_spam_prob; + double spam_prob, spam_freq, ham_freq, bayes_spam_prob, bayes_ham_prob, + ham_prob, fw, w; for (i = rt->start_pos; i < rt->end_pos; i++) { res = &g_array_index (node->results, struct rspamd_token_result, i); @@ -112,9 +116,13 @@ bayes_classify_callback (gpointer key, gpointer value, gpointer data) spam_freq = ((double)spam_count / MAX (1., (double)rt->total_spam)); ham_freq = ((double)ham_count / MAX (1., (double)rt->total_ham)); spam_prob = spam_freq / (spam_freq + ham_freq); - bayes_spam_prob = (0.5 + spam_prob * total_count) / (1. + total_count); + ham_prob = ham_freq / (spam_freq + ham_freq); + fw = feature_weight[node->window_idx % G_N_ELEMENTS (feature_weight)]; + w = (fw * total_count) / (4.0 * (1.0 + fw * total_count)); + bayes_spam_prob = PROB_COMBINE (spam_prob, total_count, w, 0.5); + bayes_ham_prob = PROB_COMBINE (ham_prob, total_count, w, 0.5); rt->spam_prob += log (bayes_spam_prob); - rt->ham_prob += log (1. - bayes_spam_prob); + rt->ham_prob += log (bayes_ham_prob); res->cl_runtime->processed_tokens ++; } |