TL;DR
This paper tackles the challenge of generating effective keywords for sponsored search campaigns, especially for new or multi-product campaigns. They propose a generative seq2seq model combined with a trie-based search mechanism to create relevant and diverse keywords from product information, significantly boosting ad exposure.
What Problem Does This Paper Address?
Sponsored search is a crucial advertising channel, but its effectiveness hinges on carefully selected keywords. Advertisers manually choose keywords to match shopper queries, but this is time-consuming and requires deep domain knowledge. Insufficient or low-quality keywords lead to missed opportunities and poor customer experiences. The problem is exacerbated by the trend of multi-product ad campaigns and the difficulty of generating good keywords for "cold" campaigns with no historical performance data. Existing methods often struggle with the long-tail nature of queries and the lack of context in short product descriptions.
Methods
The core of the proposed solution is an end-to-end machine learning framework for keyword generation. It leverages a seq2seq model, specifically using Gated Recurrent Units (GRUs) for both the encoder and decoder. The model is trained on product titles to generate relevant keywords.
A key innovation is the integration of a trie-based search mechanism during the decoding phase. Instead of a standard beam search that can lead to redundant or irrelevant keywords, this trie constrains the generation process. It ensures that generated keywords exist within a canonical query space derived from historical search logs, effectively filtering out invalid or less useful sequences. This approach is applied to both single-product and multi-product campaigns. For multi-product campaigns, they explore two strategies: Generation by Keyword Aggregation (G-KA), where keywords are generated per product and then aggregated, and Generation by Hidden State Mixing (G-HSM), where product encodings are averaged before decoding.
Datasets
The paper uses organic search click data for training, which includes pairs of queries and clicked products. This data is chosen to guide the model in generating keywords that are more aligned with actual shopper behavior than those from existing ad systems. For testing, they use "cold" campaigns with limited impressions from January to March 2021, ensuring no overlap with the training data.
Key Insights and Ideas
- Trie-based decoding is crucial for quality: Simply using a generative model isn't enough. Constraining the output space with a trie built on valid historical queries significantly improves the relevance and validity of generated keywords.
- Product metadata is sufficient for cold campaigns: The method's reliance on product titles makes it applicable to new campaigns without historical performance data, addressing a significant gap in current advertising tools.
- Multi-product campaigns can be handled effectively: The G-HSM approach, which mixes latent representations of products, shows superiority over simply aggregating keywords from individual products, suggesting a more holistic understanding of campaign intent.
- Generative keywords can drive real-world impact: The online A/B testing demonstrates that augmented keywords, when integrated thoughtfully, can lead to tangible increases in ad impressions.
Key Results
- Significant Exposure Boost: G-KA and G-HSM generated keywords led to a substantial increase in ad exposure, with G-KA showing a +1665% and G-HSM a +2194% improvement compared to baseline methods.
- Improved Exposure per Keyword: The "Exposure/Count" metric also saw significant gains, indicating that the generated keywords are not only numerous but also effective in driving traffic.
- Online A/B Testing Success: The proposed keyword augmentation methods increased ad impressions by 5.3% in an online A/B test, demonstrating their practical value.
- G-HSM outperforms G-KA: G-HSM showed better keyword relevancy and validity compared to G-KA, suggesting that averaging latent representations is a more effective strategy for multi-product campaigns.
- Relevance Trade-off: While generative methods achieved higher exposure, their relevance rate was slightly lower than standard baselines, a trade-off the authors acknowledge.
Limitations and Future Work
The authors acknowledge that the relevance rate for generated keywords is lower than standard baselines. They suggest future work could explore more structured decoding strategies combined with the trie to improve generation quality. Additionally, they aim to incorporate more factors into keyword generation, such as long-tail keywords and keyword competitiveness.
Summary updated