Big tech ad networks rely on tracking and personal data to target users. PrivAds is a new kind of ad platform: it learns what users like—without ever storing their identity, browsing history, or demographics. We use only click feedback and page context to serve relevant, privacy-respecting ads.

• No user tracking or segments: We use user embeddings, not personas or segments that could be traced back.
• Multimodal ad understanding: Ads are processed with a vision-language model (Jina CLIP v2) to extract both visual and textual signals.
• Context-aware serving: Ad selection considers both user embedding and the current page context.
• Custom and dynamic ads: The system can generate or enhance ads on the fly, tailored to user interests and page content.

Ad (text + image) → Jina CLIP v2 (frozen) → z_ad (2048D) → Projector (MLP) → p_ad (512D)
                                                        ↓
User clicks → Co-click graph → Margin/Contrastive Loss → user_embeddings (512D)
                                                        ↓
Scoring: cos(user_emb, p_ad)

• Frozen VLM: Jina CLIP v2 is never fine-tuned, ensuring robust, general ad representations.
• Learned Projector: A 2-layer MLP maps ad embeddings into the user space.
• User Embeddings: Learned via margin-based contrastive loss on a co-click graph (users who clicked the same ads).
• All embeddings L2-normalized for cosine similarity.

1. pip install -r requirements.txt
2. python main.py
   • Loads Jina CLIP v2
   • Generates synthetic or loads real click data
   • Trains user embeddings and projector
   • Saves models and embeddings
3. Outputs in models/ and data/ (see below)

models/
  ├── user_embeddings.npy      # (n_users, 512)
  ├── global_mean.npy          # (512,)
  └── projector.pt             # Projector weights
data/
  ├── ad_embeddings_raw.npz    # (n_ads, 2048)
  └── ad_projected.npz         # (n_ads, 512)

• Contrastive Learning: Margin-based loss encourages user embeddings to be closer to ads they clicked than to negatives, by a margin.
• Co-click Graph: Users are connected if they clicked the same ad; this graph is the basis for contrastive training.
• Synthetic & Real Data: Swap between synthetic and real click data with a single line of code.
• Evaluation: Metrics include Recall@100 and AUC for retrieval quality.
• No PII, no history: Only abstract vectors are stored; no user or behavioral data is ever saved.

cd src
python ad_encoder.py         # Test ad encoder
python projector.py          # Test projector
python click_data.py         # Test click data
python train_user_embeddings.py
python train_projector.py

• No user metadata, no tracking, no segments
• Embeddings are abstract and cannot be reversed to user data
• GDPR-compliant: delete a user by removing their embedding

• Thompson Sampling for exploration
• Real-time serving API
• Feedback loop for online learning
• Real ad data with images/videos
• Advanced evaluation (NDCG, etc.)
• Fast ANN retrieval (FAISS)

https://privads-demo.onrender.com/

Apache 2.0