arXiv cs.AI· Mianwei Zhou, Duo Zhang, Henry Tsai, Yuqi Zheng, Jian Li, Juexin Lin, Sujay Khandagale, Vijay Rajaram, Dong Jiao, Tao Cheng·· 2 小时前AI 评分34
SeqFold：面向用户行为序列骨干的时间戳可寻址压缩状态方法
SeqFold: Timestamp-Addressable Compressed States for User Action Sequence Backbones
AI 导读
arXiv 论文提出 SeqFold，一种将冻结长历史序列骨干的多层用户状态导出为紧凑、时间戳可寻址表示的方法，结合多通道融合、通道级时间衰减与 Lookup-Free Quantization（LFQ）生成离散码。配套流式实现在每个动作后更新融合递归、仅在目标时间戳物化状态。在大规模私有数据集上，SeqFold 将每状态表示存储降低 3,400 倍，同时保留未压缩模型的大部分下游排序性能。
正文 · 原文
Abstract:Existing sequential user modeling methods mainly focus on improving the backbone encoder or learning task-specific user representations for a single prediction objective. We instead study how to export the multi-layer user states of a frozen long-history backbone as compact, timestamp-addressable representations for downstream training backfill and serving. In this setting, the user state at a given timestamp summarizes the user's actions before that time, so downstream systems need only the state aligned to each target timestamp rather than the full hidden-state timeline. To this end, we present \texttt{SeqFold}, a compressed-state interface that combines learned multi-channel fusion, channel-specific temporal decay, and Lookup-Free Quantization (LFQ) to produce compact discrete codes from frozen backbone states. Task-specific decoders then map these shared codes into flexible embedding spaces for downstream models. To make timestamp-addressable export practical at scale, we further develop a fused streaming implementation that updates the fusion recurrence after every action while materializing states only at target timestamps. On large-scale proprietary datasets, SeqFold reduces per-state representation storage by $3{,}400\times$ while preserving most of the downstream ranking performance of the uncompressed model, showing that SeqFold is an effective interface for reusing long-history sequence backbones in downstream models.
| Subjects: | Artificial Intelligence (cs.AI); Social and Information Networks (cs.SI) |
| Cite as: | arXiv:2611.13704 [cs.AI] |
| (or arXiv:2611.13704v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.13704 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Mianwei Zhou [view email]
[v1]
Tue, 10 Nov 2026 07:55:16 UTC (253 KB)
来源：arXiv cs.AI · arxiv.org