Executive Analytics Overview
High-level customer interaction funnels, recommendation health, and offline performance metrics.
Personalized Recommendation Console
Query neural vector similarity, tune multi-objective ranking weights dynamically, and test cold-start fallbacks.
⚙️ Dynamic Ranking Weights & MMR Parameters
Similar Items Explorer (Item-to-Item Vector Search)
Discover product neighbors directly in the 64-dimensional learned metric embedding space.
A/B Testing Simulator & Statistical Lab
Run controlled Monte-Carlo randomized experiment simulations comparing Popularity Control against Two-Tower Treatment.
Compute the exact traffic required before deploying an experiment to prevent underpowered trials.
• 95% Confidence Interval for Absolute Lift: [+0.08%, +0.82%]
• Revenue Per User Impact (Bootstrap):
Two-Tower Deep Neural Architecture
PyTorch implementation with independent User & Item projection layers mapped to a shared 64-dimensional metric space.
| Embedding Dimension | 64 (L2-Normalized) |
| User Tower MLP | Linear(66 -> 128) -> LayerNorm -> ReLU -> Dropout(0.1) -> Linear(128 -> 64) |
| Item Tower MLP | Linear(82 -> 128) -> LayerNorm -> ReLU -> Dropout(0.1) -> Linear(128 -> 64) |
| Loss Function | In-Batch Sampled Softmax (InfoNCE, τ=0.07) |
| Optimizer | AdamW (lr=0.001, weight_decay=1e-5) |
| Checkpoint Status | Saved (best_two_tower_model.pt) |
Data Explorer & Sparsity Analysis
Examining the 2.75M event interaction stream from Retailrocket e-commerce.
| Split Partition | Interactions | Proportion | Leakage Status |
|---|---|---|---|
| Train Set | 2,204,512 | 80.0% | Fitted on Historical |
| Validation Set | 275,564 | 10.0% | Early Stopping Check |
| Test Set | 275,565 | 10.0% | Strict Future Evaluation |
System & Data Quality Diagnostics
Real-time status of FastAPI service, vector retrieval index, and pipeline integrity checks.
| Check | Rule | Status |
|---|---|---|
| Primary Key Integrity | No missing user IDs or item IDs in interaction events | PASSED |
| Chronological Monotonicity | Timestamps are strictly ordered (no forward temporal leakage) | PASSED |
| Vector Embedding Space | All precomputed item embeddings are L2-normalized unit vectors | PASSED |
| Cold-Start Fallback Strategy | Unknown visitor IDs automatically fallback to popularity baseline | PASSED |
| A/B Testing Consistency | SHA-256 hash modulo guarantees sticky user experiment assignment | PASSED |