Giới thiệu
Bài 12 xây dựng style profile, Bài 13 xây dựng behavioral learning. Bài này kết hợp tất cả thành Recommendation System — AI biết user thích gì và tạo design phù hợp.
1. Recommendation Architecture
User Request: "Generate a cool design"
│
├── Style Profile (Bài 12)
│ ├── style_vector: [0.12, -0.34, ...]
│ ├── aesthetics: {cyberpunk: 0.7, neon: 0.5}
│ └── colors: [#FF00FF, #00FFFF, #0A0A0A]
│
├── Behavioral Data (Bài 13)
│ ├── prompt_themes: {gaming: 0.4, cyberpunk: 0.3}
│ ├── preferred_layouts: [center_chest, full_front]
│ └── engagement_patterns
│
├── Collaborative Filtering
│ └── similar_users_liked: [design_A, design_B, ...]
│
└── Context
├── time_of_day
├── trending_now
└── season
│
▼
┌──────────────────────────────┐
│ Personalization Engine │
│ │
│ 1. Enhance prompt │
│ 2. Adjust generation params │
│ 3. Apply style conditioning │
│ 4. Post-filter results │
└──────────────────────────────┘
│
▼
Personalized Designs
2. Personalized Generation
class PersonalizedGenerator:
"""Generate design có personalization"""
def __init__(self):
self.generator = DesignGenerationService()
self.profiler = StyleProfiler()
self.behavior = BehaviorCollector()
async def generate_personalized(
self,
user_id: str,
prompt: str,
num_variations: int = 4,
) -> list[Image.Image]:
# 1. Get user profile
profile = await self.get_profile(user_id)
level = self._get_personalization_level(profile)
if level == "none":
# New user, no data → generic generation
return await self.generator.generate(prompt)
# 2. Enhance prompt with preferences
enhanced = self._personalize_prompt(prompt, profile)
# 3. Adjust generation parameters
gen_params = self._personalize_params(profile)
# 4. Generate with IP-Adapter style conditioning
designs = await self._generate_with_style(
enhanced, profile, gen_params, num_variations
)
# 5. Re-rank results by user preference
ranked = self._rank_by_preference(designs, profile)
return ranked
def _get_personalization_level(
self, profile: StyleProfile | None
) -> str:
if profile is None:
return "none"
if profile.confidence < 0.3:
return "light" # Chỉ enhance prompt nhẹ
if profile.confidence < 0.7:
return "medium" # Enhance + IP-Adapter nhẹ
return "full" # Full personalization
def _personalize_prompt(
self,
prompt: str,
profile: StyleProfile,
) -> str:
"""Thêm style keywords dựa trên preference"""
# Top aesthetic
top_aesthetic = list(profile.aesthetics.keys())[0]
# Top colors
color_names = [c.get("name", c["hex"])
for c in profile.color_preferences[:2]]
additions = []
if top_aesthetic and profile.confidence > 0.3:
additions.append(f"{top_aesthetic} style")
if color_names and profile.confidence > 0.5:
additions.append(
f"color palette: {', '.join(color_names)}"
)
if additions:
return f"{prompt}, {', '.join(additions)}"
return prompt
3. Cold Start Problem
class ColdStartHandler:
"""Xử lý user mới chưa có dữ liệu"""
async def handle_new_user(
self, user_id: str, prompt: str
) -> list[Image.Image]:
# Strategy 1: Popular designs
trending = await self.get_trending_designs()
# Strategy 2: Diverse variations
# Tạo 4 variations với 4 styles khác nhau
styles = ["cyberpunk", "minimal", "streetwear", "vintage"]
variations = []
for style in styles:
styled_prompt = f"{prompt}, {style} style"
design = await self.generator.generate(
styled_prompt, num_variations=1
)
variations.extend(design)
return variations
def progressive_personalization(
self, interaction_count: int
) -> dict:
"""Personalization tăng dần theo số interactions"""
if interaction_count < 5:
return {
"strategy": "exploration",
"diversity": 0.9, # Rất đa dạng
"personalization_weight": 0.1,
}
elif interaction_count < 20:
return {
"strategy": "balanced",
"diversity": 0.6,
"personalization_weight": 0.4,
}
elif interaction_count < 50:
return {
"strategy": "personalized",
"diversity": 0.3,
"personalization_weight": 0.7,
}
else:
return {
"strategy": "highly_personalized",
"diversity": 0.2,
"personalization_weight": 0.8,
}
4. Design Discovery & Feed
class DesignFeed:
"""Personalized design feed cho marketplace"""
async def get_feed(
self, user_id: str, page: int = 0, limit: int = 20
) -> list[DesignCard]:
profile = await self.get_profile(user_id)
# Mix strategies
feed = []
# 40% — personalized (dựa trên style profile)
personalized = await self._get_similar_to_profile(
profile, limit=int(limit * 0.4)
)
feed.extend(personalized)
# 30% — collaborative filtering
cf_results = await self._get_cf_recommendations(
user_id, limit=int(limit * 0.3)
)
feed.extend(cf_results)
# 20% — trending
trending = await self._get_trending(
limit=int(limit * 0.2)
)
feed.extend(trending)
# 10% — exploration (designs hoàn toàn mới)
explore = await self._get_random_quality(
limit=int(limit * 0.1)
)
feed.extend(explore)
# Deduplicate & shuffle
feed = self._deduplicate(feed)
feed = self._interleave(feed) # Mix strategies
return feed[:limit]
Tổng kết
AI Recommendation System:
- Kết hợp signals — style profile + behavior + collaborative filtering
- Personalized generation — enhance prompt & params theo preference
- Cold start — exploration-first cho user mới, progressive personalization
- Design feed — 40% personalized + 30% CF + 20% trending + 10% exploration
Bài tiếp theo: AI Size Recommendation — dự đoán size áo từ body measurements.