Keyphrase recommendation in e-commerce advertising faces significant challenges, particularly in balancing relevance and effectiveness for sellers and advertisers. The primary issue lies in recommending keyphrases that are relevant to items and represent actual user queries, crucial for targeted advertising. This problem has been approached as an Extreme Multi-Label Classification (XMC) task, utilizing search logs to map items to multiple queries. However, current XMC models exhibit limitations in addressing the full spectrum of keyphrases. They tend to focus on tail keyphrases, which are less frequently searched, while overlooking head keyphrases that drive higher revenue due to their popularity. Also, the training data derived from search logs is heavily skewed, with 90% of items associated with only one query in terms of engagement. This skew introduces bias towards popular items, neglecting the vast majority of inventory that could benefit from advertising. The challenge is further compounded by the biased presentation of items in search results, where ranking significantly influences buyer engagement, potentially misrepresenting the relevance of less popular items to certain queries.
Previous attempts to mitigate keyphrase recommendation challenges have employed various methods, each with its limitations. Open-vocabulary models like GROOV, One2Seq, and One2One often suggest keyphrases outside the label space, reducing their practical applicability. Keyphrase extraction methods, such as keyBERT, treat the problem as a two-step process: generation and ranking. However, this approach is constrained by token adjacency and presence in the item’s text and doesn’t guarantee that suggested keyphrases align with actual buyer search queries. Other deployed models include fastText, a basic linear neural network using word vectors and hierarchical softmax, and Graphite, a state-of-the-art XMC model utilizing bipartite graphs for efficient mapping. Proprietary models like Rules Engine (RE) and Similar Listing (SL) variants have also been implemented, focusing on historical co-occurrences and item similarities respectively. While these methods offer some improvements, they still struggle with comprehensive keyphrase recommendations, especially for new or less popular items, and often fail to balance between head and tail keyphrases effectively.
Researchers from eBay Inc. USA and Pennsylvania State University have introduced GraphEx, a unique graph-based approach to keyphrase recommendation, addressing the limitations of previous methods. This innovative technique extracts token permutations from item titles to suggest relevant keyphrases to sellers. The researchers highlight the inadequacy of traditional metrics like precision and recall in evaluating real-world performance, proposing a more comprehensive set of metrics that assess both keyphrase relevance and potential buyer outreach. GraphEx demonstrates superior performance compared to existing production models at eBay, effectively balancing the dual objectives of relevance and reach. The method is designed for scalability and is capable of handling billions of items while supporting near real-time inferencing in resource-constrained production environments. This approach represents a significant advancement in keyphrase recommendation, offering a more nuanced and practical solution to the challenges faced in e-commerce advertising.
GraphEx employs a unique approach to keyphrase recommendation by formulating it as a permutation problem that matches title strings to a set of predefined keyphrases. The method consists of two main phases: Construction and Inference.
In the Construction phase, GraphEx builds a series of bipartite graphs for each leaf category within a metacategory. These graphs map the relationship between words in keyphrases and the keyphrases themselves. The vertex set of each graph is divided into two subsets: X, containing all unique words from the keyphrases, and Y, containing the unique keyphrases. Edges are created between words and the keyphrases they belong to, with both words and keyphrases represented as non-negative integers for efficient processing.
The Inference phase, though not fully detailed in the provided text, likely involves using these bipartite graphs to generate keyphrase recommendations for new item titles. This approach allows GraphEx to overcome the limitations of token adjacency and presence in item text, potentially leading to more relevant and diverse keyphrase suggestions.
GraphEx’s design enables efficient scaling for billions of items and supports near real-time inferencing in resource-constrained environments, addressing key challenges in large-scale e-commerce platforms.
GraphEx demonstrates superior performance compared to other models in keyphrase recommendation across multiple metrics and categories. The evaluation focuses on the relevance, popularity (head vs. tail), and diversity of recommended keyphrases. In terms of Relevant Proportion (RP) and Head Proportion (HP), GraphEx shows a balanced performance. While some models like RE and RE-trank have higher RP due to their limited predictions, GraphEx outperforms most models in HP, especially in larger categories. GraphEx consistently outperforms other models in Relative Relevant Ratio (RRR) and Relative Head Ratio (RHR), indicating its ability to recommend more relevant and popular keyphrases.
GraphEx excels in recommending diverse head keyphrases, outperforming other models by factors ranging from 1.11x to 23.9x across different categories. This diversity is crucial for increasing potential buyer engagement. GraphEx’s execution performance shows impressive results. It achieves up to 17x speedup compared to fastText and 13x speedup compared to Graphite in the largest category (CAT_1) for inference latency. GraphEx also requires the least storage space for its models, even after constructing graphs for multiple leaf categories. Training time for GraphEx is significantly shorter, taking less than 1 minute across all categories, compared to hours or days for other models.
GraphEx’s engineering architecture for serving keyphrase recommendations to sellers on eBay’s platform demonstrates its efficiency and scalability in real-world applications. The system is designed to handle both batch and near real-time (NRT) inference, catering to different scenarios of item updates and additions. The batch inference process is conducted in two parts: a comprehensive run for all items on eBay, and a daily differential update for new or revised items. This approach ensures that the system maintains up-to-date recommendations while optimizing resource usage. The NRT inference, crucial for newly created or revised items, is implemented using Python code hosted on eBay’s internal ML inference service, Darwin.
GraphEx’s performance in batch inference is particularly noteworthy. Running on eBay’s machine learning platform Krylov, it processes 200 million items in just 1.5 hours, a significant improvement over fastText and Graphite, which take 1.75 and 1.5 days, respectively. This efficiency allows for daily model refreshes, enabling GraphEx to adapt quickly to new keywords and trends. The architecture utilizes eBay’s existing infrastructure, including Spark for data processing and a Key-Value store (NuKV) for serving recommendations. This integration allows GraphEx to scale effectively, handling billions of items and hundreds of billions of keywords across eBay’s platform. GraphEx’s quick training time, comparable to Graphite but vastly superior to fastText, enables daily model updates. This frequent refresh cycle ensures that the system can rapidly incorporate new keywords and trends, maintaining relevance in the dynamic e-commerce environment.
GraphEx represents a significant advancement in keyphrase recommendation for e-commerce advertising. This robust graph-based extraction method effectively addresses the challenges of mapping item titles to relevant keyphrases without being constrained by the item’s vocabulary or token order. Its design is particularly tailored for the online advertising sector in e-commerce platforms.
Key strengths of GraphEx include:
1. Improved relevance: It generates more item-relevant keyphrases, enhancing the accuracy of recommendations.
2. Focus on head keyphrases: By targeting popular keyphrases preferred by advertisers, GraphEx helps drive more sales.
3. Scalability: Successfully implemented at eBay, it handles billions of items daily, demonstrating its ability to operate at scale.
4. Comprehensive evaluation: The researchers employed a combination of metrics and AI evaluations, acknowledging the limitations of traditional metrics in accurately comparing model performance.
5. Superior performance: When evaluated against existing production models at eBay, GraphEx demonstrated superior results across various metrics.
6. Efficient cold start recommendations: It offers the most profitable keyphrase suggestions for new items or advertisers.
7. Low latency: GraphEx achieves the lowest inference latency in eBay’s current system, enabling quick real-time recommendations.
8. Frequent updates: The model allows for daily refreshes, ensuring it remains responsive to the rapidly changing query space in e-commerce.
Briefly, GraphEx addresses critical challenges in keyphrase recommendation for e-commerce advertising, offering a solution that balances relevance, popularity, and efficiency while demonstrating superior performance in a large-scale, real-world application.
Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. If you like our work, you will love our newsletter..
Don’t Forget to join our 50k+ ML SubReddit