Abstract
User targeting in online advertising often suffers from the "cold start" problem, particularly for "long-tail" ads with sparse historical engagement data. This paper presents the Interest-Oriented Network (ION), a bimodal retrieval model that identifies similar ads within the same product category. By leveraging historical click data from retrieved similar ads to augment training sets, ION significantly enhances targeting accuracy for new and niche products. Experimental results on proprietary and public datasets demonstrate superior retrieval performance and a near-doubling of click-through rates in real-world simulations.
Introduction
In digital advertising, effective user targeting—delivering the right ad to the right audience—is paramount. Traditional models rely on historical click-through data, where users who clicked are treated as positive training examples and those who did not as negative. However, this paradigm fails for long-tail ads: new or niche products that lack sufficient exposure to accumulate meaningful engagement data. This data scarcity creates a vicious cycle where inaccurate targeting leads to poor performance, further limiting data collection.
To break this cycle, the authors propose a novel approach: "borrowing" successful clickers from similar ads. If a new ad for green tea lacks data, the model can utilize engagement data from existing, similar green tea ads. The primary challenge lies in accurately and automatically identifying these similarities amidst noisy ad content, which often features distracting imagery and ambiguous text.
The Interest-Oriented Network (ION)
The core of the proposed solution is the Interest-Oriented Network (ION), a bimodal architecture designed to map ads into a unified representation space. Ads belonging to the same product category are projected closer together, enabling efficient retrieval.
Key Architectural Components
- Keyword-Guided Selective Encoding (KSG): Recognizing that not all words in an ad are equally descriptive, ION employs a Transformer encoder paired with a selective gate. This mechanism prioritizes product-relevant keywords (e.g., "Wrist," "Watch") over generic marketing language.
- Semantic-Enhanced Region Extraction: Ad images are frequently cluttered. ION utilizes YOLOv3 for initial region detection but refines this via Semantic Clue Attention (SCA). SCA uses text-derived keywords to re-weight image regions, focusing the model's attention on the actual product rather than background elements.
- Dual-path Fusion Attention (DFA): This layer facilitates cross-modal communication, allowing text and image features to refine one another based on their respective confidence levels, leading to a more robust joint representation.
- Multi-task Training: The model is optimized using a combination of an auxiliary keyword extraction task and a triplet loss function. The triplet loss ensures that ads within the same category are clustered together while distinct categories are separated in the embedding space.
Experimental Evaluation
The researchers evaluated ION using two distinct datasets: the proprietary Mobile Package (MP) from Tencent, featuring over 14,000 ads, and the public MS-COCO benchmark. The MP test set specifically included 40 "unseen" categories to evaluate zero-shot generalization.
Key Results
ION consistently outperformed state-of-the-art multimodal baselines, including ViLBERT and MCAM. On the MP dataset, it achieved a 74.8% Hit Precision@1. More importantly, in simulated user targeting scenarios, augmenting training data with ION-retrieved ads nearly doubled the Click-Through Rate (CTR), increasing it from 7.355‰ to 13.177‰. Furthermore, the model demonstrates high operational efficiency, capable of processing one million ads in under four hours on a single machine.
Discussion and Future Work
A significant insight from this work is the effectiveness of using text semantics as a filter for visual noise. By guiding visual attention with keywords, ION overcomes the common pitfall of models being distracted by prominent but irrelevant image features.
However, limitations remain. The model occasionally exhibits "shape bias," confusing products with similar geometries (e.g., pens and lipsticks). Future research could incorporate finer-grained attributes, such as brand positioning or aesthetic style, to further refine retrieval. Additionally, broader validation on diverse public ad repositories would enhance the generalizability of these findings.
Conclusion
The Interest-Oriented Network (ION) provides a robust solution to the cold-start problem in ad targeting. By intelligently retrieving similar ads to augment training data, ION ensures that potential customers are not missed, even for products with minimal historical data. This approach marks a significant step forward in making digital advertising more efficient and inclusive of long-tail content.
Summary updated