DynamiX: Dynamic Resource eXploration for Personalized Ad-Recommendations

AI summary

Overview

Research area: Machine learning for online advertising, specifically sequence-based ad-recommendation, user segmentation, and dynamic feature selection.

Technical level: Advanced. The paper derives a Bayesian/logistic relationship between dwell-time and conversion events and requires familiarity with maximum-likelihood estimation, self-supervised learning, Normalized Entropy (NE), and QPS-based serving metrics.

Scope: The paper describes DynamiX, a framework from Meta's Ads Data and Representation learning group that segments users into active and passive groups at run time and then selectively removes or boosts Event Based Feature (EBF) attributes for each group to improve training and inference throughput and CTR prediction.

What This Paper Is About

Ad-recommendation systems process long sequences of user-ad engagement events, which is computationally expensive and noisy, and they allocate the same training and serving resources to every user regardless of how predictable that user's behavior is. The authors ask whether resources can be reallocated on the fly: pruning low-importance features for users whose behavior is predictable (active users), and boosting features for those users while limiting resource spend on less predictable (passive) users. The goal is to gain compute efficiency in training and inference without losing ad-recommendation accuracy.

Key Contributions

1. A self-supervised method that detects correlations between a user's ad-view (dwell-time) history and their conversion history, allowing run-time segmentation of users into active and passive groups through pre-processing transformations.
2. The DynamiX framework itself, described as a maximum relevance approach that dynamically selects EBFs conditional on the run-time-allocated user groups, enabling selective feature-attribute removal, feature selection, and boosting.
3. A demonstration of dynamic feature removal from 4 of 11 ad-impression EBF attributes for active versus passive user segments, showing training and inference QPS improvements with minimal NE change.
4. A demonstration of selective feature boosting across 3 EBF sources (ad-impression, organic-impression, new-page impression), showing that personalized boosting avoids the large inference QPS regression seen with uniform boosting.

Main Findings

• Training and inference throughput gains from removal: The abstract reports that dynamic resource removal increases training and inference throughput by 1.15% and 1.8%, respectively. In Table 2, attribute removal for active users with no change for passive users produced a Training QPS of +1.1% and Inference QPS of +1.8%, with an NE Gain of 0.006.
• Removal for passive users behaved differently: Attribute removal for passive users with no change for active users produced an NE Gain of 0.012, Training QPS of +0.4%, and Inference QPS of -0.37%.
• Boosting gains: The abstract reports that dynamic feature boosting provides 0.033 NE gains while boosting inference QPS by 4.2% over baseline models. Table 3 shows the segmented run (3 EBF sources for active users, organic-impression EBF for passive users) at NE Gain 0.033, Training QPS -0.96%, and Inference QPS +4.2%.
• Uniform boosting is compute-heavy: Adding attributes to all 3 EBF sources for all users produced NE Gain 0.031, Training QPS -2.13%, and Inference QPS -10.1%, which the authors describe as a greater than 10% inference QPS regression that is infeasible.
• Relative improvement from segmentation: The paper states that dynamic user-segmentation and feature boosting leads to a 14.3% inference QPS improvement over uniform EBF boosting across all users.
• User traffic split: Empirically, 66.67% of ad-traffic users were segmented at any given time as active users, with the remaining 33.33% segmented as passive users.
• Dwell-time before conversion: The histogram in Figure 3 shows that, on average, dwell-time increases in the 5 minutes preceding a conversion event, and that in a 5-minute window prior to an engagement event users typically engage with multiple ads.
• Per-user variation: Different users exhibit different degrees of correlation, and the same user can show positive, low, and negative correlation at different times of the day (three sample users shown in Figure 4).
• Noise interpretation: NE Gains in the range [0, 0.02] and QPS < 1% can be considered experimental noise.
• Passive users are harder to predict: The authors attribute the weaker results on the passive group to low correlation between conversion events and dwell-time, making those engagements less predictable even after feature pruning.

Methodology in Plain English

The method starts from a simple observation: the time a user spends looking at an ad (dwell-time) is related to whether they later convert (click or otherwise engage). The authors treat a user's own past conversion events as a stand-in for training labels, which makes the process self-supervised: no manual labels are needed. They define a window of length s (the forecast horizon) after each impression, and mark whether a conversion occurred inside that window. They then assume the logarithm of dwell-time is normally distributed, separately for impressions that fall inside versus outside a conversion window, and apply Bayes' rule. That derivation yields a logistic sigmoid form, sigma(-w · log(D_t) + b), where the weight w and bias b depend on five distribution parameters: the two means, the two standard deviations, and the prior probability of conversion in a window of size s. The authors simplify by assuming the two standard deviations are equal across users.

The weight w is fitted per user by maximum likelihood. The normalized version of that estimate defines a correlation statistic, Corr^u_s(D,C), which compares a user's average log dwell-time near conversion events to their average log dwell-time away from them. Thresholding the absolute value of this statistic against a parameter epsilon splits users into passive (U=0) and active (U=1) groups. Because the statistic can shift through the day, the same user can move between groups, so pre-processing functions are implemented alongside the EBF sequence at the source.

The experiments use production EBF definitions. Organic-impression EBFs are generated when user-generated content is displayed with at least 50% visibility and viewed for at least 250 milliseconds, with attributes including content-id, dwell-time, media-type, position, and timestamp. Ad-impression EBFs require greater than 50% visibility and at least 250 milliseconds of viewing, with attributes including semantic-ids, ad-id, and timestamp. New-page impression EBFs are generated when users click or engage with specific pages, with attributes including semantic-ids, media-type, and timestamp. All experiments were trained on over 40 billion user-samples curated over a month of usage in one pass.

Two production data challenges are addressed: impression events and conversion events are logged through different systems, so an impression can be logged up to a minute after the conversion it corresponds to, which is handled by replacing the window C_[t, t+s] with C_[t-60, t+s]; and many impression events carry unrealistic dwell-times ranging from a few milliseconds to several hours, which is handled by de-noising ad-impression events based on dwell-time to remove outliers. In the first experiment, 4 of the 11 attributes of the ad-impression EBF data source were chosen for dynamic removal based on low feature-importance ranking. In the second, attributes were added to 3 EBF sources. Performance is measured with NE Gain (equation 10) and with QPS gains for training and inference. The paper does not report the specific model architecture, dataset composition beyond the 40 billion user-samples, or the numeric value of epsilon.

Why This Matters

Impact on research: The paper reframes user segmentation as a run-time, self-supervised operation rather than a static, long-term profiling step, and shows that the same user can be treated differently at different times of day. It connects dwell-time distributions to conversion prediction through a closed-form logistic relationship, offering a lightweight statistic that can drive feature-selection decisions without labels at inference time. Prior work summarized in Table 1 relies on embeddings and clustering; DynamiX instead uses a per-user correlation statistic.

Real-world applications:

• Ad-recommendation ranking systems that must serve billions of impressions under fixed latency budgets.
• Content feeds that use dwell-time signals to decide how much history to process per user.
• Serving infrastructure where inference QPS is a hard constraint and per-request feature cost must be reduced selectively.
• Notification or page-engagement systems that need to distinguish predictable from unpredictable users without manual labels.

Industry relevance: For large advertising platforms, small throughput changes translate into substantial compute cost. The reported +1.8% inference QPS from removal and +4.2% inference QPS from segmented boosting, versus a -10.1% inference QPS regression from uniform boosting, illustrate that personalization is what makes feature expansion affordable. The paper is authored by researchers at Meta's Ads Data and Representation learning group and uses production data, making the efficiency framing directly relevant to deployed systems.

Future Directions

• Refine the correlation estimation method beyond the current windowed, threshold-based statistic.
• Extend beyond binary active/passive segmentation to multi-user-group partitioning via multi-level thresholding, which the authors caution may produce noisier and less stable resource exploration.
• Preserve optimized and explainable workflows while scaling to more user segments.
• Continue improving computational efficiency and recommendation quality simultaneously, as stated in the paper's closing discussion.

Target Audience

This paper is most useful to machine learning engineers and researchers working on large-scale recommendation and computational advertising systems, particularly those concerned with serving latency, feature-store cost, and sequence-model efficiency. It also suits practitioners interested in self-supervised segmentation and dynamic feature selection, and readers who want a concrete example of translating a statistical relationship (dwell-time versus conversion) into production resource-allocation policy. Readers without a background in probability, maximum-likelihood estimation, and ranking metrics such as Normalized Entropy will find the methodology sections demanding.

Authors’ 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.