README.md

metadata

title: FomoFeed User Profiler
emoji: 👤
colorFrom: indigo
colorTo: green
sdk: docker
pinned: false

FomoFeed User Profiler AI

AI-powered user profiling using Turkish BERT embeddings and behavioral analysis.

🎯 Purpose

Creates comprehensive user profiles based on content consumption, creation, and engagement patterns.

🧠 Model

• BERT: dbmdz/bert-base-turkish-cased (768-dim embeddings)
• Analysis: Rule-based behavioral classification
• Output: Semantic embeddings + structured profile

📥 Input

{
  "user_id": 12,
  "post_captions": ["Harika bir gün", "Yeni proje"],
  "moment_captions": ["Kahve molası"],
  "liked_tags": ["yazilim", "teknoloji", "kahve"],
  "saved_tags": ["yazilim", "python"],
  "commented_tags": ["teknoloji"],
  "engagement_hours": [9, 12, 19, 20, 21],
  "engagement_types": ["view", "like", "save", "comment"]
}

📤 Output

{
  "user_id": 12,
  "interests": ["yazilim", "teknoloji", "python"],
  "content_preference": {
    "passive_consumer": 0.25,
    "active_engager": 0.60,
    "content_creator": 0.15
  },
  "activity_pattern": {
    "peak_hours": [19, 20, 21],
    "activity_distribution": {
      "morning": 0.2,
      "afternoon": 0.3,
      "evening": 0.4,
      "night": 0.1
    },
    "timezone_pattern": "evening"
  },
  "engagement_style": {
    "style": "active",
    "interaction_rate": 0.45,
    "content_saver": true,
    "commenter": true
  },
  "optimal_hours": [19, 20, 21],
  "confidence": 0.78
}

🚀 Usage

Create Profile

curl -X POST "https://YOUR-SPACE-URL/profile" \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": 12,
    "liked_tags": ["tech", "ai"],
    "engagement_hours": [19, 20, 21],
    "engagement_types": ["like", "save"]
  }'

Get BERT Embedding

curl -X POST "https://YOUR-SPACE-URL/embedding" \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": 12,
    "post_captions": ["AI is amazing"],
    "liked_tags": ["ai", "tech"]
  }'

📊 Profile Dimensions

Interests

Top 10 tags based on weighted engagement (saves > comments > likes)

Content Preference

• Passive Consumer: High view ratio
• Active Engager: High like/save ratio
• Content Creator: Post/moment creation frequency

Activity Pattern

• Peak Hours: Top 3 most active hours
• Distribution: Morning/afternoon/evening/night breakdown
• Timezone Pattern: Primary activity window

Engagement Style

• Lurker: <10% interaction rate
• Casual: 10-30% interaction rate
• Active: 30-60% interaction rate
• Power User: >60% interaction rate

🎯 Use Cases

• Content recommendation
• Feed personalization
• Notification timing
• Creator matching
• Trend prediction

📈 Performance

• Profile creation: <200ms
• BERT embedding: ~500ms
• Batch support: Up to 50 users
• Confidence: 0.0-1.0 (based on data volume)

Built with FastAPI + Transformers + PyTorch