Mastering CTR/CVR Prediction: Core Techniques and Resources from Recent Competitions

Click‑through‑rate (CTR) and conversion‑rate (CVR) prediction aim to estimate the probability that a user will click, purchase, or otherwise engage with a given item under specific contextual conditions, formally modeling P(click\|context). These metrics are critical for e‑commerce, social platforms, and information‑feed services, where even marginal improvements can generate substantial revenue.

Why CTR/CVR Is a Distinct Research Problem

High‑dimensional sparse features: large user bases (hundreds of millions) and vast product catalogs lead to one‑hot encodings with billions of dimensions.

Numerous categorical and discrete attributes: timestamps, channels, page positions, and other behavioral signals.

Typical Solution Strategies

Two main directions are common:

Rich feature engineering combined with simple models such as Logistic Regression (LR), which demands deep domain expertise.

Reducing manual feature work by leveraging more powerful models that automatically learn interactions, e.g., Facebook’s GBDT+LR, Factorization Machines (FM), Field‑aware FM (FFM), and various deep learning architectures (Deep & Wide, DeepFM, FNN, etc.).

Key Papers and Resources

Below are representative works and practical guides (links are kept as plain URLs for reference):

0. FM/FFM – popularized by NTU participants in the Netflix competition; they embed high‑dimensional sparse data via matrix factorization. See a detailed blog at https://blog.csdn.net/mmc2015/article/details/51760681. Note: libffm may struggle to converge on extremely large, imbalanced datasets such as Kaggle’s TalkingData.

1. FTRL – classic online learning algorithm described in “Ad Click Prediction: A View from the Trenches”. Implementation and discussion can be found at https://www.kaggle.com/titericz/giba-darragh-ftrl-rerevisited.

2. Practical Lessons from Predicting Clicks on Ads at Facebook – combines GBDT leaf indices as features for LR. Example code to extract leaf features with XGBoost: <code>new_feature = xgb.predict(d_test, pred_leaf=True)</code> Further explanations are available in three Chinese blogs linked below.

Relevant blog URLs:

https://breezedeus.github.io/2014/11/19/breezedeus-feature-mining-gbdt.html#fn:fbgbdt

https://blog.csdn.net/dengxing1234/article/details/73739836

https://blog.csdn.net/lilyth_lilyth/article/details/48032119

Neural‑Network Approaches

Embedding‑based models map categorical fields into dense vectors, which are then processed by deep networks for non‑linear classification. An early baseline NN for the TalkingData competition is available at https://www.kaggle.com/baomengjiao/embedding-with-neural-network.

Notable deep learning papers frequently used in top advertising competitions include:

Deep Neural Networks for YouTube Recommendations

Wide & Deep Learning for Recommender Systems

FNN: Deep Learning over Multi‑field Categorical Data

PNN: Product‑based Neural Networks for User Response Prediction

DeepFM: A Factorization‑Machine based Neural Network for CTR Prediction

Competition Resources

Several public CTR/CVR competitions provide datasets and discussion forums useful for practice:

Kaggle Outbrain Click Prediction – https://www.kaggle.com/c/outbrain-click-prediction/discussion

Kaggle Display Advertising Challenge – https://www.kaggle.com/c/criteo-display-ad-challenge

Kaggle Avazu CTR Prediction – https://www.kaggle.com/c/avazu-ctr-prediction/leaderboard

Tencent Social Advertising Competition – http://algo.tpai.qq.com/person/mobile/?from=singlemessage&qz_gdt=cp77gwalayaicijolfwq

Alibaba Tianchi Coupon Usage Prediction – https://tianchi.aliyun.com/competition/introduction.htm?spm=5176.100066.0.0.6f5fd780FIIzQn&raceId=231587

A comprehensive step‑by‑step CTR tutorial (including code and explanations) can be found at http://blog.csdn.net/chengcheng1394/article/details/78940565.