DynamiX: Dynamic Resource eXploration for Personalized Ad-Recommendations

TL;DR

Online ad-recommendation systems incur high compute costs and noise when processing full user-engagement histories. Dynamix couples self-supervised dwell-time correlations with a maximum-relevance objective to perform dynamic per-user segmentation (active vs passive) and selective EBF feature removal or boosting, grounded by a Bayes-derived logistic model between dwell-time and conversion probability. The approach leverages production-scale EBFs and demonstrates throughput gains (training and inference QPS) and NE improvements on large-scale data (>40B samples), while highlighting the value of targeted resource allocation over uniform handling. This work advances scalable personalization for sequence-based ads and suggests extensions to richer segmentation and explainable resource-management strategies.

Abstract

For online ad-recommendation systems, processing complete user-ad-engagement histories is both computationally intensive and noise-prone. We introduce Dynamix, a scalable, personalized sequence exploration framework that optimizes event history processing using maximum relevance principles and self-supervised learning through Event Based Features (EBFs). Dynamix categorizes users-engagements at session and surface-levels by leveraging correlations between dwell-times and ad-conversion events. This enables targeted, event-level feature removal and selective feature boosting for certain user-segments, thereby yielding training and inference efficiency wins without sacrificing engaging ad-prediction accuracy. While, dynamic resource removal increases training and inference throughput by 1.15% and 1.8%, respectively, dynamic feature boosting provides 0.033 NE gains while boosting inference QPS by 4.2% over baseline models. These results demonstrate that Dynamix achieves significant cost efficiency and performance improvements in online user-sequence based recommendation models. Self-supervised user-segmentation and resource exploration can further boost complex feature selection strategies while optimizing for workflow and compute resources.

Figures (5)

• Figure 1: System diagrams for personalized-sequence scaling by static and dynamic feature selection methods. Predictability of user-content engagement remains relatively stable for passive-users while active or engaged users can be represented as power-users and averagely engaged users.
• Figure 2: Example of time-windows applied to compare the average dwell-time of conversion events within windows to that outside the windows to capture statistically significant trends in a user’s dwell-time behavior.
• Figure 3: Histogram of log dwell-times in the 5 minutes preceding a conversion event (green) vs. more than 5 minutes before the next conversion event (blue).
• Figure 4: Different dwell-time patterns across individual users
• Figure 5: Convergence patterns for Dynamic resource removal and boosting, thereby leading to dynamic feature selection.