Source-linked AI summary
Warm Up Cold-start Advertisements: Improving CTR Predictions via Learning to Learn ID Embeddings
Feiyang Pan, Shuokai Li, Xiang Ao, Pingzhong Tang, Qing He
TL;DR
ID embeddings improve CTR prediction but require substantial data, making new and small ads difficult to model. Meta-Embedding generates initial embeddings from ad features through gradient-based meta-learning, improving cold-start and warm-up performance across tested models and datasets.
Problem
ID-embedding methods are data demanding and perform poorly for new or small ads with little labeled data, creating a cold-start problem in online advertising.
Method
Meta-Embedding trains a feature-based embedding generator with gradient-based meta-learning and two-phase cold-start and warm-up simulation over previously learned ads.
Results
Meta-Embedding significantly improves cold-start and warm-up performance for six existing CTR prediction methods across three real-world datasets, including over 15% cold-start LogLoss reductions for Wide&Deep, IPNN, and DeepFM.
Takeaways & Limitations
The generated initializer can replace random initialization for new ad IDs and can also be applied to cold-start user and advertiser IDs.
Takeaways & Limitations
The experiments use passive online supervised learning and do not address exploration–exploitation trade-offs or active-learning interviews.
Abstract
from arXiv · showhide
Click-through rate (CTR) prediction has been one of the most central problems in computational advertising. Lately, embedding techniques that produce low-dimensional representations of ad IDs drastically improve CTR prediction accuracies. However, such learning techniques are data demanding and work poorly on new ads with little logging data, which is known as the cold-start problem. In this paper, we aim to improve CTR predictions during both the cold-start phase and the warm-up phase when a new ad is added to the candidate pool. We propose Meta-Embedding, a meta-learning-based approach that learns to generate desirable initial embeddings for new ad IDs. The proposed method trains an embedding generator for new ad IDs by making use of previously learned ads through gradient-based meta-learning. In other words, our method learns how to learn better embeddings. When a new ad comes, the trained generator initializes the embedding of its ID by feeding its contents and attributes. Next, the generated embedding can speed up the model fitting during the warm-up phase when a few labeled examples are available, compared to the existing initialization methods. Experimental results on three real-world datasets showed that Meta-Embedding can significantly improve both the cold-start and warm-up performances for six existing CTR prediction models, ranging from lightweight models such as Factorization Machines to complicated deep models such as PNN and DeepFM. All of the above apply to conversion rate (CVR) predictions as well.
1 INTRODUCTION
Online advertising depends on accurate CTR and CVR estimates for allocating ad slots, but ID embeddings require substantial data and perform poorly for new or small ads. Meta-Embedding learns feature-based initial embeddings to improve both cold-start predictions and subsequent warm-up.
- Motivation: Accurate CTR and CVR estimates help publishers allocate ad slots according to ad value for each user.Ad auctions typically rank ads by bid multiplied by CTR.
- Motivation: Embedding-based CTR models are data demanding, leaving new and small ads with poorly learned ID embeddings.These difficulties constitute the ad cold-start problem.
- Motivation: 5% of ads accounted for over 80% of samples in the KDD Cup 2012 search ads dataset, leaving 95% of small ads with little data.Improving cold-start performance can benefit revenue and small advertisers’ satisfaction.
- Approach: Meta-Embedding generates initial embeddings from ad contents and attributes to support acceptable cold-start predictions and faster warm-up.It trains the generator through a two-phase simulation over previously observed ads using gradient-based meta-learning.
- Approach: The method targets both cold-start and warm-up performance and can be implemented offline or online with static or streaming data.The approach can also cold-start other ID features, including user and advertiser IDs.
- Evaluation: Experiments evaluate Meta-Embedding across six existing CTR prediction methods on three real-world datasets.The reported contributions include improvements for both cold-start new ads and warm-up small-sized ads.
2 BACKGROUND AND FORMULATIONS
CTR prediction models encode ad IDs as low-dimensional embeddings, but unseen IDs lack learned representations. Meta-Embedding generates initial embeddings from ad features for new IDs, then supports updating as labeled data arrives.
- CTR prediction estimates the probability of a binary click label from ad, user, and contextual features.
- One-hot ID encoding is sparse and potentially very large, so embedding matrices provide low-dimensional dense representations through row lookup.The embedding matrix is also called a look-up table.
- The deployment scheme initializes new-ID embeddings before labels, updates them after observing data and prediction loss, and then applies the adapted embeddings.This sequence is intended to support warm-up under the standard updating scheme.
- Offline training updates the base-model parameters and embedding matrix together with stochastic gradient descent to minimize log-loss.
- Unseen ad IDs retain initial embedding values because they have no labeled data, leading to low testing accuracy during cold start.This issue also affects ads with relatively few training samples.
- Meta-Embedding feeds an unseen ad’s contents and attributes into an embedding generator that produces its initial ID embedding.The approach trains the generator using previously learned ads through gradient-based meta-learning.
3 LEARNING TO LEARN THE ID EMBEDDINGS FOR NEW ADS
Meta-Embedding recasts each ad-specific CTR learning problem as a meta-learning task and trains a content-based generator to initialize embeddings for unseen ads. Its unified objective evaluates both initial prediction quality and adaptation after limited labeled data, using gradient-based updates and a frozen base model.
- Meta-learning formulation: CTR prediction is formulated as meta-learning, with each ad ID treated as a task sharing base-model parameters while retaining task-specific embeddings.For unseen IDs, the task-specific embedding is unknown and must be learned from prior ads.
- Meta-Embedding: The pre-trained base-model parameters are frozen during Meta-Embedding training, focusing optimization on learning embeddings for new IDs.This relies on the previously trained base model while adapting the embedding-generation component.
- Meta-Embedding: Meta-Embedding uses a shared function of ad features to generate initial embeddings for new IDs in place of unknown task-specific parameters.The generator produces an embedding-conditioned prediction model without using the embedding matrix directly.
- Cold-start and warm-up phases: The two-phase simulation first evaluates generated embeddings with no adaptation, then performs a gradient update using a second minibatch to measure warm-up performance.Two disjoint minibatches of K samples are used, with K assumed much smaller than half the available samples for an ID.
- Unified optimization objective: The meta-loss combines cold-start loss la and post-update loss lb as lmeta = αla + (1 −α)lb, balancing initial prediction accuracy with fast adaptation.Second derivatives are used when back-propagating through the adaptation step, and stochastic gradient descent updates the generator parameters.
- Architecture and training: Meta-Embedding can be trained offline or online and uses a lightweight neural generator that may reuse frozen base-model feature-embedding layers.Feature embeddings can be aggregated by average pooling, max pooling, or concatenation before a dense output layer.
4 EXPERIMENTS
The experiments evaluate Meta-Embedding across three datasets and six embedding-based CTR models, using cold-start and three-step warm-up tests. Results generally show improved CTR and CVR prediction performance, though gains vary by dataset and model.
- Datasets: Experiments cover MovieLens-1M, Tencent CVR prediction, and KDD Cup 2012 search-ad datasets.MovieLens simulates CTR with binarized ratings; Tencent contains conversion labels; KDD Cup contains search-ad session logs.
- Base models: The evaluation spans six representative models sharing an Embedding & MLP structure, including FM, Wide & Deep, PNN variants, and DeepFM.Embedding vectors are fixed to 256 dimensions across experiments.
- Dataset splits: Ads are split into old and new groups by labeled-instance thresholds, with new ads retaining enough samples for warm-up and testing.Thresholds are 300, 20,000, and 5,000 for MovieLens, Tencent CVR, and KDD Cup, respectively.
- Evaluation protocol: The protocol pre-trains base models, trains Meta-Embedding, initializes new-ad embeddings, evaluates cold-start performance, and applies three sequential warm-up batches.Each warm-up batch contains K instances, and evaluation uses a hold-out set.
- Results: Overall, Meta-Embedding significantly improves cold-start CTR and CVR predictions for almost all tested base models and datasets.Log-loss and AUC are evaluated on hold-out sets, with results averaged over three runs.
- Results: On MovieLens, Meta-Embedding generally improves AUC and LogLoss over random initialization, including LogLoss reductions above 15% for Wide&Deep, IPNN, and DeepFM during cold start.After three warm-up updates, Meta-Embedding still outperforms the baseline; improvements are also reported on the two larger datasets.
- Results: Relative improvements are more significant on MovieLens than on larger datasets, while some cases show mixed performance across models and warm-up stages.For Tencent CVR with Wide & Deep, the baseline can surpass Meta-Embedding after the third update; FM on MovieLens shows similar AUC but lower LogLoss.
A. Methods addressing the cold-start problem.
Cold-start methods address the difficulty of making predictions for new ads with little or no data. Existing approaches either gather information actively or use side information and incremental learning, but their applicability is limited across settings.
- Active methods: One class of methods addresses cold start through decision-making strategies such as contextual bandits or interviews collecting information about cold items or users.These approaches actively seek information rather than relying solely on supervised prediction.
- Supervised methods: Another class treats cold start within online supervised learning by using user attributes, item attributes, or relational data.The paper views these approaches as base-cold baselines without ad IDs.
- Incremental learning: Incremental-learning methods can accelerate fitting with small amounts of data but cannot be directly applied when new ads have no samples.They are also mostly designed for matrix factorization rather than general CTR prediction.
- Related approaches: Dropout-Net addresses missing inputs for deep collaborative filtering but is excluded because this paper targets general CTR prediction rather than collaborative filtering.It is treated as a training method for pre-training base models.
B. Meta-Learning.
Meta-learning transfers prior experience across tasks to enable fast adaptation. For cold-start advertising, the paper adapts this idea to learning ID embeddings because standard MAML does not scale directly to millions of ad-specific tasks.
- Meta-learning: Meta-learning learns how to learn new tasks by using prior experience with related tasks.The paradigm has been applied in recommendation, natural language processing, and computer vision.
- Problem framing: The paper frames warming up cold-start ads as a few-shot learning and fast-adaptation problem.Its approach follows the spirit of MAML, which learns shared parameters across tasks.
- Adaptation to advertising: MAML cannot be applied directly to CTR prediction because it learns one model per task, which is unacceptable for millions of ad tasks.Meta-Embedding instead learns a content-based generator for ID embeddings.
- Comparison: Unlike related item-cold-start work, Meta-Embedding does not model user activity and focuses on learning ID embeddings for ads.The distinction is between user-history representations and ad-embedding initialization.
- Terminology: The paper’s Meta-Embedding differs from NLP methods with the same name, which aggregate pre-trained word vectors for new corpora.The authors characterize the NLP usage as specific to that domain and lacking the same underlying idea.
6 CONCLUSION AND DISCUSSION
Meta-Embedding uses meta-learning to initialize new-ad ID embeddings, improving CTR prediction in both cold-start and warm-up phases. The paper focuses on ad-ID embedding learning while identifying extensions to broader online-learning tasks.
- Meta-Embedding improves cold-start and warm-up CTR prediction by replacing random initialization for new-ad ID embeddings.The method was evaluated across three real-world datasets and six state-of-the-art CTR models.
- The approach trains an embedding generator through a two-phase simulation over previously learned ads and a unified loss.The generator initializes embeddings for new ads during testing.
- The paper focuses on learning ad-ID embeddings as a representation-learning component for online CTR and CVR prediction.The discussion positions this as a connection between representation learning and learning to learn.
- Future work could extend learning-to-learn beyond ad-ID embeddings to evolving feature-label distributions and task-specific hyperparameter tuning.