Item Level Exploration Traffic Allocation in Large-scale Recommendation Systems

TL;DR

The paper tackles the item cold-start problem in large-scale recommender systems by introducing a dedicated exploration framework that allocates impressions to fresh items. It models the discoverability of an item as a function of exploration level via $P(G_i=1|Y_i)$, and determines per-item traffic $X_i$ by targeting a desired discovery probability $CF$ through $P(G_i=1|X_i)=CF$. The approach includes discretized bucketing, cross-entropy training on item metadata and engagement signals, and a region-based traffic allocation across High/Moderate/Low Discoverability items, with continuous retraining to adapt to evolving systems. The method achieves strong predictive performance (AUC $=0.96$, PR-AUC $=0.84$), demonstrates efficient tail-trimming of exploration, and has been deployed in production to enrich the discoverable item corpus and improve long-term recommendation quality.

Abstract

This paper contributes to addressing the item cold start problem in large-scale recommender systems, focusing on how to efficiently gain initial visibility for newly ingested content. We propose an exploration system designed to efficiently allocate impressions to these fresh items. Our approach leverages a learned probabilistic model to predict an item's discoverability, which then informs a scalable and adaptive traffic allocation strategy. This system intelligently distributes exploration budgets, optimizing for the long-term benefit of the recommendation platform. The impact is a demonstrably more efficient cold-start process, leading to a significant increase in the discoverability of new content and ultimately enriching the item corpus available for exploitation, as evidenced by its successful deployment in a large-scale production environment.

Figures (5)

• Figure 1: Model architecture for predicting $P(G_i = 1 | Y_i)$.
• Figure 2: An example model output for an item i
• Figure 3: Item Level Traffic Allocation Approach
• Figure 4: Accuracy, Precision, Recall, and F1 for the model predicting whether an item is discoverable.
• Figure 5: Precision-Recall Curve for the model predicting whether an item is discoverable.