Thesis title: On-Core Challenges in Industrial Recommender Systems
Industrial-scale recommender systems are a cornerstone of the modern digital economy, demanding solutions that are not only accurate but also scalable, efficient, and robust. This thesis, focused on the practical challenges of deploying these systems, aims to bridge the persistent gap between academic advancements and industrial requirements. This work unfolds the methodologies, innovations, and practical insights crucial for developing graph-based recommender systems that perform effectively under real-world production constraints. The following summary provides an overview of the key contributions and insights of this thesis.
The introduction provides a context for the research, clarifying the motivations and objectives driving the study. It highlights the mismatch between academic benchmarks and industrial realities, such as massive-scale graphs, real-time inference demands, and the critical cold-start problem, setting the stage for a systematic exploration.
Starting our exploration, we first ground the research in a tangible industrial context by introducing TIM-Rec, a novel and unique dataset for multi-item upselling recommendations derived from telecommunication call records. We then turn our attention to one of the most critical industrial challenges: the cold-start problem. We introduce a novel framework with a re-ranking method designed to improve recommendations for new users and items on large graphs and further investigate the crucial trade-off between diversity and coverage to ensure new content is both relevant and discoverable.
We then broaden our investigation to high-value industrial problems, introducing CRAB, a reinforcement learning-based agent for proactive customer churn reduction. This demonstrates the versatility of our data-driven approach in solving critical business challenges beyond traditional recommendation tasks within the same industrial environment.
The core of this thesis directly confronts the challenge of scalability. We propose efficient recommendation methods based on graph coarsening and label propagation, which simplify massive user-item graphs while preserving their essential properties. Furthermore, we challenge the prevailing trend of increasing model complexity by demonstrating that learning an optimal graph structure allows simpler, more scalable models to achieve competitive or even superior performance.
The final part of this thesis summarizes the key findings and contributions to the field of industrial recommender systems. It shows how we achieved the objectives outlined in the introduction, advancing the understanding and practical implementation of scalable, graph-based systems. It also offers insights into future research directions, emphasizing the ongoing need for solutions that balance theoretical soundness with deployability.
In conclusion, this thesis represents a significant step in the ongoing pursuit of practical and effective recommender systems. It stands as a valuable resource for researchers and practitioners striving to build systems that meet the rigorous demands of industrial-scale deployment. With the principles of scalability, efficiency, and real-world applicability, this research contributes to the collective effort of ensuring that recommender systems deliver tangible and reliable value in complex business environments.