User Lifelong Behavior Modeling (ULBM) is becoming a critical area in industrial recommendation systems, where user sequences can reach lengths of 10^5 or more. This survey note synthesizes recent research on how to balance the need for capturing long-term user interests with the computational constraints of real-time serving. Key approaches include hierarchical attention mechanisms, memory networks, and efficient sequence compression techniques. The note emphasizes that effectiveness alone is insufficient; models must also be deployable within strict latency and resource budgets. For engineers, this means understanding the trade-offs between model complexity and inference speed. The survey categorizes methods by their efficiency-accuracy profiles, offering a practical framework for choosing the right approach for a given platform's scale. This is particularly relevant for video platforms like Kuaishou, where daily interactions per user can be in the hundreds. The insights are also applicable to e-commerce and social media recommendation systems.
This paper note covers the emerging field of User Lifelong Behavior Modeling (ULBM), which addresses the challenge of modeling ultra-long user behavior sequences under strict industrial efficiency constraints. It highlights the trade-offs between model effectiveness and computational efficiency, a critical issue for platforms like Kuaishou. The content is valuable for engineers designing scalable recommendation systems.