Source-linked AI summary
Chain Dimension Scaling: Communication-Efficient Loop-Style Scaling for Industrial Short-Video Ranking
Zhimeng Jiang, Yuening Li, Ruixiao Sun, Yuanzhen Lin, Kun Jin, Chuan He, Liang Liu
TL;DR
Industrial recommendation rankers scale width and depth to improve ranking quality, but this brings parameter and distributed-training costs. CDS reuses feature-interaction parameters across iterations to add a chain-length scaling axis. In production, it reduces parameters and training time in the compression regime while quality rises with chain length under a fixed budget.
Problem
Industrial rankers scale width and depth to improve ranking quality, but these strategies bring substantial training and serving costs.
Source: §1 · p. 1Method
CDS repeatedly applies a feature-interaction block with shared parameters, using chain length as a scaling dimension orthogonal to width, depth, and data.
Source: §1 · p. 2Results
26.6% fewer trainable parameters and 6.4% higher training throughput accompanied less than a 0.03% offline ranking-quality degradation; quality also improves monotonically with chain length at a fixed parameter budget.
Source: §1 · p. 2Takeaways & Limitations
The efficiency gain mainly comes from reduced parameter synchronization and optimizer-update costs, rather than lower FLOPs.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Industrial recommendation rankers keep scaling in width and depth to billions of parameters, but this increasingly incurs costs in parameter growth, distributed communication overhead, and training inefficiency. We introduce \underline{C}hain \underline{D}imension \underline{S}caling (CDS), a scaling strategy orthogonal to width and depth. CDS repeatedly applies a single feature-interaction block with parameters shared across iterations, so the number of iterations (the chain length) becomes a new scaling dimension that adds no parameters. A lightweight intra-chain residual connection stabilizes optimization and improves quality. CDS enables two regimes. In the \emph{compression regime}, shallow blocks with a longer chain match deeper baselines with fewer parameters and faster training; in the \emph{quality-scaling regime}, increasing chain length under a fixed parameter budget yields consistent offline and online gains, suggesting the chain dimension adds modeling capacity beyond parameter reuse. We evaluate CDS in a large-scale multi-task short-video ranking system in production. In the compression regime, CDS stays within $0.03\%$ of baseline offline quality while reducing trainable parameters by 26.6\% and improving training throughput by 6.4\%, with neutral-to-positive online participation. In the quality-scaling regime, quality improves monotonically with chain length. Wall-clock analysis shows the efficiency gain comes mainly not from reduced FLOPs but from lower cross-device synchronization overhead and reduced optimizer update cost. These results position parameter-shared iterative computation as a practical, communication-efficient scaling axis for industrial ranking systems.
1 Introduction
Industrial rankers scale through wider and deeper models, but parameter growth and distributed-training costs motivate increasing computation depth without adding unique parameterized layers. CDS shares feature-interaction parameters across iterations, enabling compression and quality-scaling regimes alongside communication-efficiency gains.
Motivation
- The production ranker’s feature-interaction module dominates parameter count and training cost, while independently parameterized layers make distributed communication a bottleneck.The paper identifies all-gather and gradient communication overhead as growing with model depth.Source: §1 · pp. 1–2
Approach
- CDS repeatedly applies a feature-interaction block with shared parameters, making chain length a scaling axis orthogonal to width, depth, and data.Increasing chain length does not increase trainable parameter count.Source: §1 · p. 2
Operating regimes
- A shallow backbone with a longer chain can match baseline quality with fewer parameters and faster training, while fixed-budget chain growth improves ranking quality.The latter regime indicates modeling capacity beyond simple parameter reuse.Source: §1 · p. 2
Efficiency analysis
- Efficiency gains mainly reflect fewer parameter synchronization operations and smaller optimizer updates, not reduced forward FLOPs.The reduced synchronization comes from having fewer unique parameterized layers.Source: §1 · p. 2
Results
- 26.6% fewer trainable parameters and 6.4% higher training throughput came with less than a 0.03% offline ranking-quality degradation and neutral-to-positive online engagement.In the fixed-budget quality-scaling regime, ranking quality improves monotonically with chain length.Source: §1 · p. 2
2 Preliminaries
The preliminaries describe the ranker’s feature-token inputs and multi-task outputs, then locate its main parameter and training-cost bottleneck in the stacked feature-interaction module.
2.1 Notations
- Feature groups are projected into a shared-dimensional token sequence; the final interaction-block output feeds a shared projection and task-specific prediction towers.The described feature groups include dense and sparse features and short- and long-term user histories.Source: §2.1 · p. 3
2.2 Where the Cost Concentrates
- The stacked feature-interaction module accounts for most trainable parameters and training-step time, and its independently parameterized layers make it the main communication bottleneck.Under sharded data-parallel training, each layer adds parameter all-gather and gradient communication operations.Source: §2.2 · p. 3
3 Method
CDS adds chain length as a scaling axis by repeatedly applying shared feature-interaction blocks, addressing the parameter and communication costs of independently parameterized depth. Its intra-iteration residual stabilizes longer chains, while compression and quality-scaling regimes use chain length for different goals.
3 Method
- CDS introduces chain length as a scaling dimension orthogonal to width, depth, and data, and repeatedly applies shared-parameter feature-interaction blocks.The method section proceeds from the cost motivation to the forward pass, intra-iteration residual, and two operating regimes.Source: §3 · p. 4, §3.2 · p. 4
3.1 From Cost Structure to a New Axis
- Existing width, depth, and data scaling increases parameter count and distributed communication cost together, motivating a way to increase effective computation depth without adding independently parameterized layers.The stacked feature-interaction module is the main bottleneck for parameter footprint and distributed training cost.Source: §3.1 · p. 4
3.2 Chain Dimension Scaling
- CDS replaces independently parameterized depth with L′ shared-parameter blocks, each iterated C times, so effective depth is L′·C while trainable parameters depend only on L′.Under sharded data-parallel training, communication overhead scales with unique parameterized layers; increasing C at fixed L′ adds no additional parameter synchronization or gradient all-reduce operations.Source: §3.2 · p. 4
3.3 Stabilizing Long Chains: The Intra-Iteration Residual
- The intra-iteration residual adds an identity path at each iteration to improve gradient propagation and stabilize training at longer chain lengths.Among tested residual designs, the intra-iteration residual provides the largest improvement in optimization stability and ranking quality; inter-iteration residual alone has limited quality benefit and adds training overhead, while combining both adds no further performance gain.Source: §3.3 · p. 5
3.4 Two Operating Regimes Along the Chain Dimension
- CDS uses a shallower parameterized depth and longer chain to preserve effective depth while reducing trainable parameters, or increases chain length at fixed parameter budget to test quality gains.The compression regime targets matched quality with lower training cost; the quality-scaling regime tests whether chain length adds modeling capacity beyond parameter sharing or regularization.Source: §3.4 · p. 6
4 Experiments
CDS was evaluated offline and in 10-day production A/B tests on a large-scale, multi-task short-video ranker. It supports both compression with near-baseline quality and quality scaling at fixed parameter count, while its training-efficiency gains mainly reflect reduced communication and optimizer overhead.
4.1 Experiment Setup and Evaluation Metrics
- Experiments used a production short-form video recommendation platform, offline AUC, and 10-day online A/B tests on real traffic.The system jointly optimizes engagement and satisfaction tasks, including tail tasks with positive-label rates around 10^-4 to 10^-5.Source: §4.1 · p. 6
4.2 RQ1: Compression Regime and Online A/B Validation
- With the intra-iteration residual, CDS reduced offline AUC by only 1 bp while maintaining neutral-to-positive online performance; under aggressive compression (L′ = L/6, C = 6), it cut trainable parameters 44.4% and improved training throughput 12.3%.This setting incurred a 9 bps offline AUC drop and a −0.04% online estimate (95% CI [−0.15%, +0.07%]).Source: §4.2 · p. 7
4.3 RQ2: The Efficiency of Chain Dimension Scaling
- +3.0 bps AUC and +0.07% online engagement over the standard L = 3 model came from increasing chain length to C = 2 with zero additional parameters.Against the L = 6 model, CDS (L = 3, C = 2) used roughly 150M fewer parameters, stayed within 1 bp offline AUC (+10.0 bps vs +11.0 bps), and achieved +0.11% vs +0.05% online engagement.Source: Figure 4 · p. 7, §4.3 · p. 8
4.4 RQ3: Which Residual Design Matters?
- Without residuals, CDS lost about 13 and 11 bps on two offline tasks; the intra-iteration residual brought results within 1 bp on the first and slightly above baseline on the second.The intra-residual variant had a +0.06% online engagement gain; adding an inter-iteration residual produced nearly identical offline and online performance.Source: §4.4 · p. 8, Figure 5 · p. 8
4.5 RQ4: Where the Efficiency Gain Comes From
- CDS reduced per-step training time by approximately 4.14 ms, mainly through lower communication overhead rather than lower computation.Communication fell by 2.41 ms per step; optimizer time fell by 0.56 ms, while Conv Fusion time increased by 2.92 ms from iterative computation and residual operations.Source: §4.5 · pp. 9–10, Figure 6 · p. 9
5 Related Work
CDS builds on prior work in industrial ranking and iterative weight sharing, but centers its contribution on compression and distributed-training communication costs. The paper distinguishes its production multi-task setting and two-regime framing from prior loop-style approaches.
Scaling Industrial Ranking Models
- CDS adds chain length as a parameter-free scaling axis to industrial ranking's width, depth, and data scaling.Prior ranking systems report quality gains alongside increased parameter counts and training time; CDS is designed to compose with those existing axes.Source: §5 · p. 10
Iterative Computation and Weight Sharing
- Iterative weight sharing spans language modeling and recommendation; LoopCTR [23] applies recursive sharing to CTR prediction using shared loop blocks and process supervision.The cited work reports strong results on public benchmarks.Source: §5 · p. 10
CDS Distinctions
- CDS is evaluated in a live-traffic, multi-task production ranker and frames chain length for both compression and quality scaling.Its wall-clock analysis attributes speedup to lower per-layer all-gather and reduce-scatter under sharded data-parallel training [26] [31], rather than reduced FLOPs.Source: §5 · pp. 10–11
6 Conclusion
CDS uses shared-parameter iterative feature-interaction blocks and an intra-chain residual to add effective computation depth without adding parameters. In a production short-video ranker, it improved efficiency in compression and quality with longer chains, with gains attributed mainly to reduced communication and optimizer overhead.
6 Conclusion
- CDS reduced trainable parameters by 26.6% and improved training throughput by 6.4% in the compression regime while maintaining ranking quality.At a fixed parameter budget, increasing chain length consistently improved ranking quality; efficiency gains mainly came from lower communication and optimizer overhead under sharded data-parallel training.Source: §6 · p. 11