Upload RESEARCH_RESOURCES.md

RESEARCH_RESOURCES.md

@@ -0,0 +1,254 @@
+# RTB Bidding Algorithm Comparison — Complete Research Resource List
+
+> Generated: 2026-05-05 | Repository: https://huggingface.co/hamverbot/bidding_algorithms_benchmark
+
+---
+
+## Table of Contents
+
+1. [Bidding Algorithms](#1-bidding-algorithms)
+2. [CTR Prediction Models](#2-ctr-prediction-models)
+3. [Clearing Price / Market Price Prediction](#3-clearing-price--market-price-prediction)
+4. [Datasets](#4-datasets)
+5. [Codebases & Implementations](#5-codebases--implementations)
+6. [Benchmark Leaderboards](#6-benchmark-leaderboards)
+7. [Recommended Architecture](#7-recommended-architecture)
+
+---
+
+## 1. Bidding Algorithms
+
+### 1.1 Lagrangian Dual + Online Gradient Descent (BEST MATCH)
+
+| Property | Detail |
+|----------|--------|
+| **Paper** | "Learning to Bid in Repeated First-Price Auctions with Budgets" |
+| **Authors** | Qian Wang, Zongjun Yang, Xiaotie Deng, Yuqing Kong (2023) |
+| **Venue** | NeurIPS 2023 |
+| **arXiv** | [2304.13477](https://arxiv.org/abs/2304.13477) |
+| **HF Papers** | https://huggingface.co/papers/2304.13477 |
+| **Algorithm** | DualOGD — Lagrangian dual multiplier updated by online error gradient descent |
+| **Auction Type** | First-price (also handles second-price) |
+| **Constraints** | Budget cap: total spend ≤ ρT |
+| **Regret Bound** | Õ(√T) for both full-information and one-sided feedback |
+| **Key Formula** | λ_{t+1} = Proj_{λ>0}(λ_t − ε·(ρ − c̃_t(b_t))) |
+| **Bid Rule** | b_t = argmax_b (r̃_t(v_t, b) − λ_t·c̃_t(b)) |
+| **Prediction Models Needed** | CTR predictor (for v_t), empirical CDF of competing bids (G̃) |
+
+### 1.2 Dual Mirror Descent (Second-Price)
+
+| Property | Detail |
+|----------|--------|
+| **Paper** | "The Best of Many Worlds: Dual Mirror Descent for Online Allocation Problems" |
+| **Authors** | Santiago Balseiro, Haihao Lu, Vahab Mirrokni (2020) |
+| **Venue** | Operations Research (2023) |
+| **arXiv** | [2011.10124](https://arxiv.org/abs/2011.10124) |
+| **Citations** | 135+ |
+| **Algorithm** | Dual mirror descent — generalizes OGD with Bregman divergences |
+| **Auction Type** | Second-price (truthful) |
+| **Bid Rule** | b_t = v_t / (1 + μ_t) |
+| **Dual Update** | μ_{t+1} = Proj(μ_t − η·(ρ − payment_t)) |
+| **Key Insight** | No market price model needed for second-price auctions |
+| **Prediction Models** | CTR predictor only |
+
+### 1.3 Dual Descent with RoS + Budget (Multi-Constraint)
+
+| Property | Detail |
+|----------|--------|
+| **Paper** | "Online Bidding Algorithms for Return-on-Spend Constrained Advertisers" |
+| **Authors** | Zhe Feng, Swati Padmanabhan, Di Wang (2022) |
+| **Venue** | ICML 2022 |
+| **arXiv** | [2208.13713](https://arxiv.org/abs/2208.13713) |
+| **Algorithm** | Two dual variables: λ for RoS, μ for budget |
+| **Bid Rule** | b_t = ((1+λ_t)/(μ_t+λ_t)) · v_t |
+| **Key Insight** | Adaptable for k% spend floor — second dual variable enforces minimum spend |
+
+### 1.4 RLB — Reinforcement Learning Bidding
+
+| Property | Detail |
+|----------|--------|
+| **Paper** | "Real-Time Bidding by Reinforcement Learning in Display Advertising" |
+| **Authors** | Han Cai et al. (2017) |
+| **Venue** | WSDM 2017 |
+| **arXiv** | [1701.02490](https://arxiv.org/abs/1701.02490) |
+| **GitHub** | https://github.com/han-cai/rlb-dp (188 stars) |
+| **Algorithm** | MDP + Dynamic Programming + Neural value function |
+| **Results** | +22% clicks over linear bidding on iPinYou |
+| **Prediction Models** | CTR θ(x) + market price distribution m(δ, x) |
+
+### 1.5 HiBid — Industrial Hierarchical Dual-RL
+
+| Property | Detail |
+|----------|--------|
+| **Paper** | "HiBid: A Cross-Channel Constrained Bidding System" |
+| **arXiv** | [2312.17503](https://arxiv.org/abs/2312.17503) |
+| **Scale** | 64K advertisers, 70M requests/day, 4 channels, Meituan |
+| **Algorithm** | High-level RL budget allocation + Low-level λ-parameterized bidding |
+
+### Unified Dual Multiplier Template
+
+```
+For each auction t:
+1. Observe value v_t (from CTR prediction × click value)
+2. Compute bid: b_t = f(v_t, dual_multiplier_t)
+3. Observe outcome: payment c_t (if won) or 0 (if lost)
+4. Compute gradient: g_t = ρ − c_t
+5. Update multiplier: λ_{t+1} = Proj_{λ≥0}(λ_t − η·g_t)
+```
+
+| Method | Auction | Bid Function f(v, λ) |
+|--------|---------|----------------------|
+| Wang 2023 | First-price | argmax_b (r̃(v,b) − λ·c̃(b)) |
+| Balseiro 2020 | Second-price | v / (1+λ) |
+| Feng 2022 | Second-price | ((1+λ_RoS)/(λ_RoS+λ_budget)) · v |
+
+---
+
+## 2. CTR Prediction Models
+
+### 2.1 FinalMLP (RECOMMENDED)
+
+| Property | Detail |
+|----------|--------|
+| **Paper** | "FinalMLP: An Enhanced Two-Stream MLP Model for CTR Prediction" |
+| **arXiv** | [2304.00902](https://arxiv.org/abs/2304.00902) |
+| **Criteo AUC** | **0.8149** |
+| **Avazu AUC** | **0.7666** |
+| **Architecture** | Two-stream MLP + feature gating + bilinear fusion |
+| **Inference** | <1ms — best for RTB latency constraints |
+
+### 2.2 Other Top Models
+
+| Model | Criteo AUC | Architecture | RTB-Suitable |
+|-------|-----------|-------------|--------------|
+| **FinalMLP** | 0.8149 | Two-stream MLP | ✅ Best |
+| **DCNv2** | 0.8142-0.8144 | CrossNetV2 + DNN | ✅ |
+| **GDCN** | 0.8161* | Gated Cross + DNN | ✅ |
+| **DeepFM** | 0.8138 | FM + DNN | ✅ |
+| **FCN** | New | LCN + ECN (no DNN) | ✅ |
+| DIN | — | Attention (user history) | ❌ Slow |
+| DIEN | — | GRU + attention | ❌ Slow |
+
+*GDCN uses own data split — not directly comparable.
+
+**BARS Meta-Finding (2009.05794):** After 7,000+ experiments, SOTA deep CTR models differ by only 0.1-0.3% AUC. Architecture matters less than data preprocessing, hyperparameter tuning, and feature engineering.
+
+---
+
+## 3. Clearing Price / Market Price Prediction
+
+### 3.1 Non-Parametric Empirical CDF (BASELINE)
+
+| Property | Detail |
+|----------|--------|
+| **Source** | Wang et al. (2023), Algorithm 1 |
+| **Method** | G̃_t(b) = (1/(t-1))∑𝟙{b ≥ d_s} |
+| **Pros** | No training, theoretically sound, handles distribution shift |
+| **Cons** | No context, cold-start |
+
+### 3.2 Deep Censored Learning / Survival Analysis
+
+| Property | Detail |
+|----------|--------|
+| **Library** | **TorchSurv** (Novartis, 200★) [2404.10761] |
+| **URL** | https://github.com/Novartis/torchsurv |
+| **Method** | Neural net with censored survival loss |
+| **Loss** | Win: -log f(price\|x); Loss: -log S(bid\|x) |
+| **Key Insight** | Proper survival framework handles censoring |
+
+### 3.3 Censored Linear Regression (Wu et al. 2015, KDD)
+
+| Property | Detail |
+|----------|--------|
+| **Method** | Tobit-like: log(market_price) = β·x + ε, ε ~ N(0, σ²) |
+| **Pros** | Contextual, simple |
+| **Cons** | Linear — limited capacity |
+
+### Comparison
+
+| Method | Contextual? | Handles Censoring? | Training? | Complexity |
+|--------|-------------|-------------------|-----------|------------|
+| Empirical CDF | ❌ | N/A | None | Minimal |
+| Censored Linear | ✅ | ✅ | Light | Low |
+| Deep Survival | ✅ | ✅ | Neural net | Medium |
+| Win Prob NN | ✅ | ❌ | Neural net | Low |
+
+---
+
+## 4. Datasets
+
+### CTR Prediction (Verified on HF Hub)
+
+| Dataset | HF Path | Size | Verified |
+|---------|---------|------|----------|
+| Criteo_x4 | reczoo/Criteo_x4 | 45.8M rows, 5.6GB | ✅ |
+| Avazu_x4 | reczoo/Avazu_x4 | 40.4M rows, 1.8GB | ✅ |
+
+### RTB Bidding (External Only)
+
+| Dataset | Source | Availability |
+|---------|--------|-------------|
+| iPinYou | data.computational-advertising.org | External download |
+| YOYI | Various mirrors | External download |
+
+---
+
+## 5. Codebases
+
+| Library | URL | Purpose |
+|---------|-----|---------|
+| **FuxiCTR** | https://github.com/reczoo/FuxiCTR | 40+ CTR models, config-driven |
+| **DeepCTR-Torch** | https://github.com/shenweichen/DeepCTR-Torch | 20+ CTR models, simple API |
+| **TorchSurv** | https://github.com/Novartis/torchsurv | Deep survival for clearing price |
+| **BARS** | https://github.com/openbenchmark/BARS | Standardized CTR benchmark |
+| **rlb-dp** | https://github.com/han-cai/rlb-dp | RL for RTB |
+| **budget_constrained_bidding** | https://github.com/dingmu365/budget_constrained_bidding | Budget-constrained algorithms |
+
+---
+
+## 6. Benchmark Leaderboards
+
+| Leaderboard | URL |
+|-------------|-----|
+| BARS CTR Criteo_x4 | https://openbenchmark.github.io/BARS/CTR/leaderboard/criteo_x4.html |
+| BARS CTR Avazu | https://openbenchmark.github.io/BARS/CTR/leaderboard/avazu_x4.html |
+
+---
+
+## 7. Recommended Architecture
+
+```
+┌─────────────────────────────────────────────┐
+│            BIDDING ALGORITHM                  │
+│  Dual OGD: λ_{t+1} = Proj(λ_t - ε·(ρ - c̃)) │
+│  Two-sided: μ (cap) + ν (floor)              │
+├─────────────────────────────────────────────┤
+│          PREDICTION MODELS                    │
+│  ┌──────────────┐  ┌────────────────────┐    │
+│  │ FinalMLP     │  │ Empirical CDF /     │    │
+│  │ v_t=pCTR×V   │  │ TorchSurv           │    │
+│  └──────────────┘  └────────────────────┘    │
+├─────────────────────────────────────────────┤
+│              DATASETS                         │
+│  Criteo_x4 + synthetic auction simulation    │
+└─────────────────────────────────────────────┘
+```
+
+## Paper Index
+
+| # | Paper | arXiv | Year | Citations |
+|---|-------|-------|------|-----------|
+| 1 | Wang et al. — First-Price Auctions with Budgets | 2304.13477 | 2023 | Growing |
+| 2 | Balseiro et al. — Dual Mirror Descent | 2011.10124 | 2020 | 135+ |
+| 3 | Feng et al. — RoS Constrained Bidding | 2208.13713 | 2022 | 38+ |
+| 4 | Cai et al. — RLB | 1701.02490 | 2017 | 300+ |
+| 5 | Wang et al. — HiBid | 2312.17503 | 2023 | New |
+| 6 | — Contextual First-Price (Quantile) | 2603.07207 | 2026 | New |
+| 7 | Mao et al. — FinalMLP | 2304.00902 | 2023 | Growing |
+| 8 | Wang et al. — GDCN | 2311.04635 | 2023 | Growing |
+| 9 | Wang et al. — DCN V2 | 2008.13535 | 2021 | 500+ |
+| 10 | Guo et al. — DeepFM | — | 2017 | 3000+ |
+| 11 | Zhu et al. — BARS-CTR | 2009.05794 | 2021 | 100+ |
+| 12 | Wu et al. — Censored Price Prediction | — | 2015 | 101 |
+| 13 | — TorchSurv | 2404.10761 | 2024 | New |
+| 14 | — Robust Budget Pacing | 2302.02006 | 2023 | Growing |