In the realm of e-commerce advertising, the optimization of keyphrase recommendations is a challenge that requires a delicate balance between relevance and effectiveness for sellers and advertisers. The task of Extreme Multi-Label Classification (XMC) focuses on mapping items to multiple queries, utilizing search logs to guide the process. However, existing XMC models exhibit limitations in effectively addressing both head and tail keyphrases and struggle with biased training data derived from search logs.
Numerous methods have been attempted to address these challenges, each with its own set of limitations. Open-vocabulary models often lack practical applicability, while keyphrase extraction methods struggle to guarantee alignment with actual buyer search queries. Other deployed models like fastText, Graphite, Rules Engine (RE), and Similar Listing (SL) variants offer some improvements but still fall short when it comes to comprehensive keyphrase recommendations.
To combat these challenges, researchers at eBay Inc. USA and Pennsylvania State University have introduced GraphEx—a graph-based approach that effectively recommends keyphrases by extracting token permutations from item titles. This innovative technique overcomes the inadequacies of traditional metrics by proposing a new comprehensive set that evaluates both keyphrase relevance and potential buyer outreach.
GraphEx employs a unique construction phase involving bipartite graphs for leaf categories within a metacategory as well as an inference phase using these graphs to generate keyphrase recommendations for new item titles. This approach allows GraphEx to overcome prior limitations regarding token adjacency in item text.
The performance evaluation demonstrates that GraphEx excels in recommending diverse head keyphrases across different categories while outperforming other models in terms of speedup during inference latency and storage space requirements.
Furthermore, GraphEx’s engineering architecture enables efficient scaling for billions of items across eBay’s platform while successfully catering to scenarios of batch updates through daily differential updates for new or revised items as well as near real-time inference using Python code hosted on eBay’s internal ML inference service, Darwin.
Notably, GraphEx boasts strengths including improved relevance delivery which enhances accuracy levels significantly during ad application processes; superior performance demonstrated across various metrics; low latency results contributing notably quick real-time operations; efficient cold start recommendation capabilities benefiting new advertisers or items; scalability showcased by its ability to handle billions of items daily; frequent model refreshes ensuring responsiveness amid rapid query space changes in e-commerce advertising sector – thereby highlighting itself as an effective solution geared towards balancing relevance with popularity efficiently within large-scale scenarios facing e-commerce industries.
Ultimately representing significant advancements addressing prominent challenges prevailing within e-commerce advertising sectors—GraphEx offers an intricate yet practical solution finely balanced between efficiency & effectiveness—setting a benchmark amidst current production models at eBay through its robust engineering design offerings: please refer credit goes here if needed!
