2nd/10_Wiki/Topic_Programming/AI_and_ML/맞춤형 팩 (Personalized Packs).md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-맞춤형-팩-personalized-packs	맞춤형 팩 (Personalized Packs)	10_Wiki/Topics	verified	self	Personalized Packs Dynamic Bundles Player-Tailored Offers	none	A	0.9	applied	monetization mobile-game personalization ml		2026-05-10	pending	language framework python contextual-bandit / XGBoost

맞춤형 팩 (Personalized Packs)

매 한 줄

"매 player의 progression / collection gap / spend tier에 fit한 bundle을 ML로 generate". 2018 Supercell의 Brawl Stars Brawl Pass에서 mass-personalization 시작 → 2024 Royal Match · Monopoly Go 의 contextual-bandit 기반 dynamic offer로 evolve. 2026 현재 LLM-augmented offer copy + reinforcement-learning price elasticity가 industry standard.

매 핵심

매 Personalization Signal

• Collection gap: 매 missing card / character / skin → highest "completion utility".
• Progression stall: 매 stuck level → relevant booster / energy bundle.
• Spend tier: 매 LTV percentile (whale / dolphin / minnow / non-payer).
• Churn risk: 매 7-day rolling DAU drop → retention offer.
• Session context: 매 just-failed stage → instant-relief bundle.

매 Bundle Composition Heuristic

• Anchor (core item): 매 player가 가장 원하는 single SKU — collection gap based.
• Filler (utility): 매 gold / energy / consumables — perceived value 부풀리기.
• Discount %: 매 30~80% — perceived savings vs. actual margin.
• Time pressure: 매 24~72hr countdown — scarcity-driven conversion.

매 응용

1. Monopoly Go: 매 dice + sticker pack 동적 가격.
2. Royal Match: 매 stuck-level relief bundle.
3. Marvel Snap: 매 collection-gap-aware bundle (spotlight key).
4. Genshin Impact: 매 character-specific weapon + materials bundle pre-banner.

💻 패턴

Contextual Bandit Offer Selection

import numpy as np
from sklearn.linear_model import SGDRegressor

class OfferBandit:
    def __init__(self, n_arms: int, ctx_dim: int, alpha: float = 0.1):
        self.models = [SGDRegressor(learning_rate='constant', eta0=alpha)
                       for _ in range(n_arms)]
        self.ctx_dim = ctx_dim
        for m in self.models:
            m.partial_fit([np.zeros(ctx_dim)], [0])

    def select(self, ctx: np.ndarray, eps: float = 0.1) -> int:
        if np.random.rand() < eps:
            return np.random.randint(len(self.models))
        scores = [m.predict([ctx])[0] for m in self.models]
        return int(np.argmax(scores))

    def update(self, arm: int, ctx: np.ndarray, reward: float):
        self.models[arm].partial_fit([ctx], [reward])

Collection Gap Score

def gap_score(player_inv: set[str], target_set: set[str],
              rarity_weight: dict[str, float]) -> dict[str, float]:
    missing = target_set - player_inv
    return {sku: rarity_weight.get(sku, 1.0) for sku in missing}

def top_anchor(scores: dict[str, float], k: int = 1) -> list[str]:
    return sorted(scores, key=scores.get, reverse=True)[:k]

Price Elasticity Estimator

import numpy as np
from scipy.optimize import minimize_scalar

def expected_revenue(price: float, base_demand: float, elasticity: float) -> float:
    qty = base_demand * (price ** elasticity)  # elasticity < 0
    return price * qty

def optimal_price(base_demand: float, elasticity: float,
                  bounds: tuple = (0.99, 99.99)) -> float:
    res = minimize_scalar(lambda p: -expected_revenue(p, base_demand, elasticity),
                          bounds=bounds, method='bounded')
    return float(res.x)

Bundle Builder

from dataclasses import dataclass

@dataclass
class Bundle:
    anchor: str
    fillers: list[str]
    price_usd: float
    discount_pct: int
    expires_in_hours: int

def build_bundle(player_id: str, anchor_sku: str, ltv_tier: str) -> Bundle:
    tier_config = {
        'whale': (49.99, 60, 24),
        'dolphin': (19.99, 65, 48),
        'minnow': (4.99, 70, 72),
        'non_payer': (0.99, 80, 168),
    }
    price, discount, hours = tier_config[ltv_tier]
    fillers = recommend_fillers(player_id, count=3)
    return Bundle(anchor_sku, fillers, price, discount, hours)

LLM Offer Copy

import anthropic

def generate_copy(bundle: Bundle, player_lang: str = "ko") -> dict:
    client = anthropic.Anthropic()
    msg = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=300,
        system=f"You write mobile-game offer copy in {player_lang}. "
               f"3 outputs: title (max 20ch), subtitle (max 40ch), CTA (max 10ch).",
        messages=[{"role": "user", "content": str(bundle)}],
    )
    return parse_copy(msg.content[0].text)

Frequency Cap & Fatigue

from datetime import datetime, timedelta

def can_show_offer(player_id: str, store: dict) -> bool:
    last = store.get(player_id, {}).get('last_offer_ts')
    if not last: return True
    return datetime.utcnow() - last >= timedelta(hours=6)

매 결정 기준

상황	Approach
Whale (top 1%)	$49.99~$99.99 high-value bundle, low frequency
Dolphin (top 10%)	$9.99~$19.99 staircase progression
Minnow	$0.99~$4.99 starter / IAP-onramp
Non-payer (D7+)	$0.99 introductory + double-currency
Churn risk	retention bundle + 80% discount

기본값: contextual bandit + LTV tier × collection-gap anchor + 6hr frequency cap.

🔗 Graph

• 부모: Personalization
• 변형: Staircase_Monetization_Model · Gacha

🤖 LLM 활용

언제: offer copy generation, A/B variant ideation, anchor SKU rationale explanation. 언제 X: 매 actual price / SKU selection — bandit / RL이 더 robust (LLM은 calibration 약함).

❌ 안티패턴

• Whale-only optimization: 매 minnow / non-payer cohort revenue ignore — long-tail 손실.
• Predatory targeting: 매 churn-risk player에게 last-resort discount → regulatory risk (UK CMA, EU Digital Fairness Act).
• Static bundles: 매 player segment 동일 offer → CTR 50%↓.
• No frequency cap: 매 offer fatigue → uninstall spike.

🧪 검증 / 중복

• Verified (deconstructoroffun.com 2024 case studies, GDC Monetization Summit 2025).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — personalized pack 5-signal model + bandit + price elasticity 정리