Vietnamese Name Gender Classification

1. Overview

• Logistic Regression (LR)
• Linear SVM
  This model is designed to be:
• Lightweight – classical ML, no GPUs required
• Easy to integrate – scikit-learn pipelines saved as .joblib
• Practical – includes a clear production decision rule and an ambiguity flag
• Accuracy Model: ~96.5%

2. Training data

• Source: collection of Vietnamese full names with gender labels on a public website
• Labels: Giới tính ∈ {Nam, Nữ} (binary male/female)
• Size:
  • Training: ~50,000 labeled names
  • Test: ~15,000 names (held-out evaluation set)

2.1. Preprocessing

• Normalize whitespace (strip, collapse multiple spaces).
• Lowercase the text, but keep Vietnamese accents (diacritics).
• Create an additional variant name_wo_surname by:
  • Splitting on spaces.
  • Dropping the first token (surname).
  • Keeping middle + given name. Example:
• "Nguyễn Hoàng Phúc" →
  • full_name_clean = "nguyễn hoàng phúc"
  • name_wo_surname = "hoàng phúc" Both forms are used for training to make the model more robust to how names are provided.

2.2. Evaluation

• Accuracy: ~96.5%
• Proportion of names marked ambiguous: ~2.6% | Class | Precision | Recall | F1-score | Support | |-------------|-----------|--------|----------|---------| | Nam | 0.956 | 0.964 | 0.960 | 6637 | | Nữ | 0.972 | 0.965 | 0.968 | 8409 | | Accuracy| | | 0.965| 15046 | | Macro avg | 0.964 | 0.965 | 0.964 | 15046 | | Weighted avg | 0.965 | 0.965 | 0.965 | 15046 |

• Nam: 6,401 correctly predicted vs. 236 misclassified as “Nữ”.
• Nữ: 8,111 correctly predicted vs. 298 misclassified as “Nam”. Number of names marked ambiguous by PROD: 395 / 15046 (2.63%)

3. Model architecture

3.1. Features

• TfidfVectorizer with:
  • analyzer="char_wb" (character n-grams within word boundaries)
  • ngram_range=(2, 6) (bigrams to 6-grams)
  • min_df=2 (ignore n-grams that appear in fewer than 2 documents) This allows the model to learn patterns from substrings of names, which is very effective for Vietnamese given names.

3.2. Base models

• model_lr.joblib
  • TfidfVectorizer → LogisticRegression(max_iter=1000, n_jobs=-1)
  • Supports predict_proba for probability estimates.
• model_svm.joblib
  • TfidfVectorizer → LinearSVC
  • Strong linear classifier, but outputs only class labels (no probabilities).

3.3. Production ensemble logic (simplified)

1. Preprocessing
   • full_name_clean = normalize_name_keep_accent(full_name_raw)
   • name_wo_surname = drop_surname(full_name_clean)
2. SVM predictions
   • SVM_FULL = model_svm.predict([full_name_clean])
   • SVM_WO = model_svm.predict([name_wo_surname])
3. LR combined probabilities
   • Get probabilities from LR on both text variants:
     • p_Nam_full, p_Nu_full = LR(full_name_clean)
     • p_Nam_wo, p_Nu_wo = LR(name_wo_surname)
   • Average the Nam probability:
     • p_Nam_avg = 0.5 * (p_Nam_full + p_Nam_wo)
     • p_Nu_avg = 1.0 - p_Nam_avg
   • LR_label = "Nam" if p_Nam_avg >= p_Nu_avg else "Nữ"
   • LR_conf = max(p_Nam_avg, p_Nu_avg)
4. Decision rule

• If SVM_FULL == SVM_WO (SVMs agree):
  • final_label = SVM_FULL
  • Mark as ambiguous if LR_conf < lr_conf_clear (default 0.60).
• If SVM_FULL != SVM_WO (SVMs disagree, hard case):
  • If LR_conf >= lr_conf_as_referee (default 0.70), follow LR:
    final_label = LR_label
  • Otherwise, default to SVM_WO:
    final_label = SVM_WO
  • Mark as ambiguous = True. The final output includes:
• final_label – production gender prediction ("Nam" or "Nữ")
• LR_conf – LR combined confidence
• is_ambiguous – whether this name is considered ambiguous by the ensemble

• Analytics and statistical exploration of Vietnamese name data.
• Educational demos of classical ML methods on Vietnamese text.
• Lightweight baseline for experiments involving Vietnamese full names.

• The model predicts only binary gender (Nam / Nữ) from naming patterns.
• It does not capture the full diversity of gender identities or self-determined gender.
• It may be inaccurate or biased for:
  • Rare or gender-neutral names
  • Non-Vietnamese names, abbreviations, nicknames
  • Names from under-represented regions or time periods in the training data