Source-linked AI summary
Deep CTR Prediction in Display Advertising
Junxuan Chen, Baigui Sun, Hao Li, Hongtao Lu, Xian-Sheng Hua
TL;DR
Image-ad CTR prediction requires models that can capture nonlinear visual and contextual effects beyond logistic regression over handcrafted features. The paper introduces DeepCTR, an end-to-end network that learns from raw images and basic features, and reports effective, scalable performance on a large real-world dataset. The authors also show that training can be accelerated and that the model learns meaningful visual signals, while evaluation remains tied to the paper’s commercial advertising setting.
Problem
Logistic regression and handcrafted visual features are limited for extracting complex nonlinear relationships in image-ad CTR prediction.
Method
DeepCTR jointly processes raw image pixels and basic features with convolutional and fully connected networks in an end-to-end model.
Results
The model improves CTR prediction accuracy over baseline approaches, with batch normalization improving training and performance and visual features providing additional information on new ads.
Takeaways & Limitations
DeepCTR provides a unified way to learn visual and contextual CTR features while making large-scale image-ad training suitable for practical systems.
Takeaways & Limitations
The evaluation uses a commercial advertising dataset from one arbitrary week, and a compared method failed to converge on data containing many product categories.
Abstract
from arXiv · showhide
Click through rate (CTR) prediction of image ads is the core task of online display advertising systems, and logistic regression (LR) has been frequently applied as the prediction model. However, LR model lacks the ability of extracting complex and intrinsic nonlinear features from handcrafted high-dimensional image features, which limits its effectiveness. To solve this issue, in this paper, we introduce a novel deep neural network (DNN) based model that directly predicts the CTR of an image ad based on raw image pixels and other basic features in one step. The DNN model employs convolution layers to automatically extract representative visual features from images, and nonlinear CTR features are then learned from visual features and other contextual features by using fully-connected layers. Empirical evaluations on a real world dataset with over 50 million records demonstrate the effectiveness and efficiency of this method.
1. INTRODUCTION
Online display advertising depends on accurate CTR prediction to rank image ads, but conventional models struggle to learn nonlinear visual effects from high-dimensional handcrafted features. The paper proposes an end-to-end DNN that learns from raw image pixels and basic features, with scalable training and evaluation on more than 50 million records.
- CTR prediction is central to display-ad allocation because systems rank available ads by predicted click probability and show the top K to users.
- Image appearance can substantially change CTR even when ads share the same basic contextual features, motivating effective use of visual information.
- Logistic regression handles large-scale data efficiently but cannot extract higher-order nonlinear features without potentially causing feature explosion.
- DeepCTR directly combines raw image pixels and basic features, using convolutional and fully connected layers to learn representative visual and nonlinear cross-feature representations.
- The study introduces efficient training methods, evaluates the model on over 50 million real-world response records, and visualizes saliency maps of image ads.
2. RELATED WORK
Prior CTR systems commonly rely on logistic regression and handcrafted features, while later models add nonlinear modeling or visual representations. The paper positions DeepCTR as an end-to-end approach that jointly learns from raw images and basic features.
- Paired image examples illustrate that low contrast and excessive subjects can correspond to different CTRs under otherwise matched ad conditions.
- Earlier display-ad CTR methods used handcrafted features with logistic regression, while factorization machines, trees, and neural networks addressed nonlinear modeling mainly over basic or handcrafted visual features.
- Image-ad studies showed visual appearance affects user response, but commonly extracted more than 30 handcrafted image features before prediction.
- The closest prior CNN approach omitted basic features during image extraction and used logistic regression to combine image and basic information, limiting nonlinear interaction modeling.
- Deep learning had already improved performance over traditional CTR models using basic features alone, motivating extension to jointly model image and contextual inputs.
3. METHOD OVERVIEW
The method learns an impression-to-CTR embedding directly from raw image pixels and basic features instead of relying on handcrafted visual representations. It uses an integrated deep network to model both modalities and their nonlinear correlations.
- An impression contains an image and a high-dimensional basic feature vector, and the model predicts the probability that the user clicks the image ad.
- DeepCTR learns the embedding function directly from raw image pixels and basic features in one integrated neural network.
4. NETWORK ARCHITECTURE
DeepCTR separates raw-image and basic-feature processing, then combines their learned representations to model CTR with fully connected layers.
- 4. NETWORK ARCHITECTURE: DeepCTR uses separate Convnet and Basicnet sub-networks before combining their outputs to capture correlations between image and basic features.The design avoids directly concatenating different domains or training fully separate networks.
- 4. NETWORK ARCHITECTURE: The Convnet processes raw image u through convolutional layers, while Basicnet reduces the dimensionality of basic features v.The Convnet architecture contains 17 convolution layers, beginning with 5 × 5 kernels and followed by four groups of 3 × 3 layers.
- 4. NETWORK ARCHITECTURE: Their outputs are concatenated and passed through two fully connected layers in Combnet, whose final real-valued output supports Logloss-based CTR prediction.Combnet receives the joint representation and produces the value z before Logloss is computed.
5. SPEED UP TRAINING
DeepCTR makes large-scale training feasible through sparse computation and image-grouped sampling, reducing processing costs while stabilizing and accelerating optimization.
- 5. SPEED UP TRAINING: The sparse layer and grouped sampling make DeepCTR training suitable for a real online system despite more than 50 million samples and nearly 200,000 basic features.The paper adopts both techniques to target training in less than one day.
- 5. SPEED UP TRAINING: Sparse fully connected layers reduce basic-feature forward and backward complexity to O(nd′), where d′ is the number of nonzero features and d′ ≪ d.CSR representation enables sparse operations, and backward updates touch only weights connected to nonzero dimensions.
- 5. SPEED UP TRAINING: The approach addresses small Convnet-limited batches, whose unstable gradients and inefficient GPU parallelism make convergence difficult on noisy CTR data.Sampling shared-image records allows a larger effective batch for Basicnet while retaining a smaller image-processing batch for Convnet.
- 5. SPEED UP TRAINING: Grouping many basic-feature records by shared image enables batches containing kn samples while using n image computations, substantially increasing effective batch size.The method samples n images and k associated basic-feature vectors per image, copies each image representation k times, and processes them jointly.
- 5. SPEED UP TRAINING: Grouped sampling reduces epoch iterations to several thousand, accelerates training, and produces more stable batch gradients that ease convergence.The paper also evaluates whether grouping affects performance relative to thorough shuffling.
6. EXPERIMENT
Experiments on a commercial dataset evaluate DeepCTR against baseline models, network choices, training techniques, and visual explanations. DeepCTR improves CTR prediction, generalizes to unseen images, and benefits from batch normalization, dropout, and efficient sampling.
- Results and Discussion: The ensemble of three DeepCTR models achieves the best AUC and Logloss, while visual features outperform handcrafted image features and basic-feature baselines.Relative metrics are reported against lr basic; the stated AUC gain corresponds to a 1–2% CTR increase in online experiments.
- Results and Discussion: On 3,090 previously unseen images, DeepCTR wins by a large margin and visual features provide supplementary information when ad-group features are unavailable.The result is attributed to visual features identifying ads with similar characteristics rather than memorizing image IDs.
- Results and Discussion: Batch normalization speeds training and improves performance because Basicnet and Convnet outputs otherwise differ substantially in scale and variance.The mismatch makes the subsequent fully connected layer harder to converge.
- Results and Discussion: Dropout lowers testing Logloss despite requiring more time to reach the lowest test loss, indicating overfitting remains without it.This effect is observed in the final two fully connected layers.
- Results and Discussion: Sparse fully connected layers reduce computation and memory overhead, enabling larger batches that speed training and ease convergence.The data-sampling method preserves performance while costing less training time, and the system fits on a NVIDIA TESLA k20m GPU with 5 GB memory.
- Results and Discussion: Saliency maps show that both high-level objects and low-level properties such as texture, characters, and background affect image-ad CTR.The visualizations are presented as an intuitive aid for ad design choices.
7. CONCLUSIONS
The paper proposes DeepCTR, an end-to-end deep network for predicting image-ad CTR from automatically extracted visual features and basic features. Evaluated on a 50-million-record real-world dataset, DeepCTR demonstrates effectiveness and efficiency.
- DeepCTR combines Convnet, Basicnet, and Combnet to extract image features, reduce basic-feature dimensionality, and learn nonlinear CTR features.Sparse fully connected layers and data sampling techniques speed training.
- The model predicts image-ad CTR directly from image and contextual information in an integrated end-to-end network.
- DeepCTR demonstrates effective and efficient CTR prediction on a 50 million-record real-world dataset.