Multi-objective cluster based bidding algorithm for E-commerce search engine marketing system.

Area of Science:

• E-commerce
• Machine Learning
• Computational Advertising

Background:

• Search engine marketing (SEM) is crucial for e-commerce success.
• High feature sparsity and large bid request volumes pose challenges for industrial-level bidding systems.
• Existing systems struggle with accurate ad performance prediction and computational burden.

Purpose of the Study:

• To introduce an end-to-end multi-objective bidding system for Walmart's e-commerce SEM.
• To address the challenges of feature sparsity and computational load in large-scale bidding.
• To optimize for multiple business metrics simultaneously.

Main Methods:

• Developed an optimization model targeting a mixture of SEM metrics.
• Utilized Transformer models to extract ad vector representations.
• Employed clustering techniques to leverage geometric relationships for collaborative bidding predictions, mitigating sparsity issues.

Main Results:

• Successfully implemented a system handling tens of millions of bids daily.
• Addressed ad performance feature sparsity through collaborative predictions.
• Demonstrated production efficiency through theoretical and numerical analyses.

Conclusions:

• The proposed multi-objective bidding system is an efficient solution for large-scale e-commerce SEM.
• Transformer embeddings and clustering effectively handle sparse features and improve prediction accuracy.
• The system provides a scalable and robust approach to modern industrial-level bidding.