Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Warm Up Cold-start Advertisements: Improving CTR Predictions via Learning to Learn ID Embeddings (1904.11547v1)

Published 25 Apr 2019 in cs.LG, cs.IR, and stat.ML

Abstract: 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.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Feiyang Pan (13 papers)
  2. Shuokai Li (4 papers)
  3. Xiang Ao (33 papers)
  4. Pingzhong Tang (43 papers)
  5. Qing He (88 papers)
Citations (170)

Summary

Improving Cold-Start CTR Prediction with Meta-Embedding

The optimization of click-through rate (CTR) predictions is integral to maximizing efficiency in online advertising. Yet, a notable challenge lies in the cold-start problem, wherein new advertisements, introduced without substantial historical data, struggle to achieve accurate CTR predictions. The paper by Pan et al. addresses this challenge through a novel meta-learning strategy, termed Meta-Embedding, aimed at enhancing both cold-start and warm-up phases of CTR prediction for new ads.

Core Contributions

The authors introduce Meta-Embedding, a technique leveraging meta-learning to initialize desirable embeddings for new ad IDs, thus addressing the cold-start problem. This approach employs previously learned embeddings via gradient-based meta-learning to effectively "learn how to learn" new embeddings. The proposed method emphasizes two performance aspects: better predictions during the initial cold-start phase and accelerated model fitting during subsequent warm-up phases.

The technique involves simulating these phases using existing data, breaking the problem into a meta-learning framework comprising: a cold-start phase, which assesses initialization effectiveness, and a warm-up phase, measuring improvement after minimal data exposure. Their loss function, designed to balance performance across these phases, promotes both steady improvement in predictions as well as rapid adaptation. The methodology is distinct from existing approaches by targeting CTR prediction with an embedding generator that refines initial model predictions for new ads significantly better than traditional random initialization.

Methodological Advancements

The structure underpinning Meta-Embedding incorporates two elements: a pre-trained base model, which remains static during embedding generator training, and a neural network-based embedding generator that learns ad feature representations and initializes embeddings for new IDs. The authors align this with architectures typical to modern CTR prediction models, such as Factorization Machines, Wide & Deep networks, and DeepFM, ensuring broad applicability.

A key methodological innovation is the generator’s capability to leverage meta-learning insights from established tasks, resulting in embeddings that improve both initial predictive accuracy and learning speed when labeled examples become available. This approach synergizes well with Model-Agnostic Meta-Learning (MAML), adapting its principles to create content-based dynamic embeddings.

Experimental Verification

Experiments were conducted across datasets of varying sizes and complexity — MovieLens-1M, a Tencent CVR dataset, and the KDD Cup 2012 dataset. The Meta-Embedding approach was evaluated on its ability to augment six established CTR models, consistently showing enhanced cold-start and warm-up metrics over baseline approaches. Notably, it yielded considerable improvement for smaller datasets, indicating its potential to offer substantial benefits where data sparsity is more pronounced.

Implications and Future Directions

The practical implications of this paper are significant, offering a meta-learning-based solution that can improve ad targeting efficiency in cold-start situations, thus enhancing advertiser satisfaction and platform performance. Theoretically, this work contributes to the body of literature exploring how learning efficiencies can be transferred across tasks in machine learning and artificial intelligence.

Future research could extend these findings by exploring how similar meta-learning frameworks could adapt to other dynamic features or tasks where labeled data is elusive. Additionally, integrating this approach into active learning frameworks might provide even more robust solutions for CTR and conversion rate predictions.

In conclusion, Meta-Embedding provides a comprehensive solution to the cold-start challenge in CTR prediction, achieving significant performance improvements by learning from the structure of the problem itself. This contribution not only tackles immediate challenges in digital advertising but also opens avenues for future explorations into adaptive learning systems.