LOCAL PREVIEW View on GitHub

Few-Shot Learning Scenarios - MangaAssist

Few-shot learning helps MangaAssist adapt quickly when a new intent, catalog pattern, or support issue appears before enough labeled data exists for full fine-tuning.

When This Topic Matters

Use few-shot learning when:

  • OOD clusters reveal a new user need,
  • a seasonal promotion creates a short-lived workflow,
  • a new manga trend appears suddenly,
  • a new support issue has only 20-100 examples.

Scenario 1 - New Intent From OOD Cluster

Example new intent:

subscription_manga_box

Users ask:

  • "Can I get monthly manga boxes?"
  • "Do you have a subscription for random volumes?"
  • "Can MangaAssist build me a monthly reading bundle?"

Initial data:

Data Count
OOD clustered examples 42
human-labeled positives 35
nearest negative examples 120
synthetic positives after review 80

Few-shot method options:

Method Use when
SetFit small labeled text classification
prototypical classifier quick embedding-space intent addition
prompt-based temporary rule need immediate safe handling
full fine-tune later after 500+ validated examples

Promotion gate:

Metric Gate
new-intent recall >= 80%
false positive rate into new intent <= 5%
old-intent accuracy regression <= 0.3 points
support/business owner approval required

Scenario 2 - Seasonal Promotion Intent

For events like holiday manga bundles, do not permanently reshape the intent taxonomy too early.

Better flow:

  1. temporary few-shot detector,
  2. route to promotion workflow,
  3. monitor volume and conversion,
  4. retire or promote after the season.

Scenario 3 - Rapid Genre Adaptation

If a new anime adaptation causes sudden interest in a niche genre, few-shot examples can update recommendation retrieval:

  • add 20-50 editorial query-title pairs,
  • generate hard negatives from adjacent genres,
  • train a small adapter or prototype layer,
  • expire the boost if engagement falls.

Failure Modes

Failure Detection Fix
prototype too broad many false positives add hard negatives
synthetic examples dominate model learns artificial phrasing cap synthetic share and review
temporary intent never removed taxonomy clutter add expiration review date
old intents regress confusion matrix shifts freeze base or use adapter

Production Log

{
  "event": "few_shot_intent_trial",
  "candidate_intent": "subscription_manga_box",
  "positive_examples": 115,
  "new_intent_recall": 0.83,
  "false_positive_rate": 0.041,
  "decision": "shadow_route"
}

Final Decision

Few-shot learning is the right bridge between OOD discovery and full taxonomy expansion. MangaAssist should use it to test demand quickly, then graduate only durable patterns into the main fine-tuned classifier.