TL;DR
This paper introduces DeepMCP, a novel framework for click-through rate (CTR) prediction that goes beyond traditional feature-CTR modeling. By incorporating user-ad and ad-ad relationships through dedicated subnets, DeepMCP learns richer feature representations, leading to improved prediction accuracy.
What Problem Does This Paper Address?
Click-through rate (CTR) prediction is crucial for online advertising, as it directly impacts ad ranking and revenue. Existing methods primarily focus on modeling the relationship between features and the CTR itself. However, these approaches often struggle with data sparsity and fail to capture more nuanced relationships between entities like users and ads, or even between ads themselves within a user's interaction sequence. This limits their ability to learn truly informative and robust feature representations.
Methods
DeepMCP is a deep learning framework designed to address these limitations. It comprises three interconnected subnets, all sharing a common embedding matrix:
- Prediction Subnet: This is a standard deep neural network (DNN) that models the direct feature-CTR relationship. It takes all available features (user, query, ad, and other) and predicts the CTR. This subnet can be replaced with other state-of-the-art CTR prediction models.
- Matching Subnet: This subnet models the user-ad relationship, aiming to learn whether an ad is relevant to a user's interests. It takes user and query features as input for one part and ad features for another, projecting them into a common latent space and learning a matching score. This is inspired by semantic matching models.
- Correlation Subnet: This subnet models the ad-ad relationship, specifically considering ads that appear within a user's click sequence. It leverages a skip-gram-like approach, similar to word embeddings, to learn representations of ads that are correlated in a temporal context.
During training, all three subnets are jointly optimized using a combined loss function. Crucially, during inference (testing), only the prediction subnet is active, making the process efficient.
Datasets
The paper evaluates DeepMCP on two large-scale datasets:
- Avito advertising dataset: Logs from avito.ru, a Russian classifieds website, covering ad impressions and clicks.
- Company advertising dataset: Ad impression and click logs from a commercial advertising system at Alibaba.
Key Insights and Ideas
- Beyond Feature-CTR: The core idea is that modeling relationships between entities (user-ad, ad-ad) provides richer feature representations than solely focusing on the feature-CTR link.
- Synergistic Learning: Jointly training the prediction, matching, and correlation subnets allows them to benefit from each other. The matching and correlation subnets learn representations that are useful for prediction, while the prediction task guides the learning of these auxiliary relationships.
- User-Ad Matching is Key: The experiments suggest that the user-ad matching subnet contributes more significantly to performance improvement than the ad-ad correlation subnet, likely because it directly involves both the user and the ad.
- Efficient Inference: The architecture allows for a streamlined testing phase, where only the prediction subnet is used, making it practical for real-world applications.
Key Results
DeepMCP consistently outperforms several state-of-the-art CTR prediction models, including LR, FM, DNN, Wide&Deep, and DeepFM, on both the Avito and Company datasets. For instance, on the Avito dataset, DeepMCP achieved an AUC of 0.7927 and Logloss of 0.05518, outperforming the next best model (DeepMP) by a small margin. The paper also shows that DeepMCP's performance is sensitive to the balancing parameters ($\alpha$ and $\beta$) for the matching and correlation subnets, with optimal values leading to significant gains.
Limitations and Future Work
- Hyperparameter Tuning: The performance of DeepMCP is sensitive to the balancing parameters $\alpha$ and $\beta$. Finding optimal values requires careful tuning, and excessively large values can even degrade performance.
- Complexity: While inference is efficient, the training process involves multiple subnets, potentially increasing computational cost.
- Future Work: The authors suggest exploring more sophisticated ways to model user-ad and ad-ad relationships, potentially incorporating other types of auxiliary tasks or leveraging more advanced deep learning architectures. Further investigation into the interpretability of the learned representations from the matching and correlation subnets could also be valuable.
Summary updated