A system often has to act long before it learns whether the act worked: a recommender sees a click in seconds and a purchase in days. With $K$ actions and a delay of $d$ rounds, the best rate known for this setting is $\widetilde{O}(\sqrt{(K+d)T})$ over $T$ rounds, so a longer menu is always more expensive to learn from. It need not be: if the outcome depends on the action only through the state it produced, then one late outcome informs every action that could have produced the observed state, and the price is set by how many genuinely different states the actions produce rather than by how many actions there are. We measure this using an effective dimension $v_t$ between $1$ and the number of states, and prove $\widetilde{O}(\sqrt{(d+1)V\log K})$ for a rotating algorithm and $\widetilde{O}(\sqrt{V^{-}}+\sqrt{dT})$ for the single-copy algorithm used in practice, for any budget fixed in advance; merging similar states lowers the price further, at an explicit bias. Even when given the exact losses from $d$ rounds ago, no algorithm escapes $Ω(\sqrt{dE\min\{1+\log J,T/d\}})$, where $J$ counts the drifting directions and $E$ bounds how far losses move while the learner waits. On generated data, the state channel cuts regret by up to 79 percent against action-level weighting and, on the funnel family, by 32 to 68 percent against a tuned minimax-optimal method.
What is the right delay complexity when a learner can track only $C$ pending feedback items and discarded feedback is permanently lost? Existing one-point bandit convex optimization guarantees in this model pay $\sqrt{Tσ_{\max}}$, where $σ_{\max}$ is the peak backlog, although unlimited tracking admits the sharper $\sqrt{d_{\mathrm{tot}}}$ dependence on total delay. We introduce a scheduler-side conditional-energy interface that separates rate adaptation from the one-point perturbation filtration and handles the dependent importance weights created by randomized admission. Under the same semi-clairvoyant oracle and pathwise hard-capacity contract, this yields an untuned learner whose delay term scales as $O(\sqrt{E_C d_{\mathrm{tot}}})$, with only an explicit restart factor $E_C$; a public constant-factor peak bound removes this factor while $d_{\mathrm{tot}}$ remains unknown. Under strong convexity, the same interface yields the temporal cost $H_A(d)=\sum_t σ_t/(A+t)$. Two delay vectors with identical delay multisets, $d_{\mathrm{tot}}$, $σ_{\max}$, and capacity can nevertheless have polynomially different minimax regret, showing that timing matters under curvature even when aggregate delay summaries agree. Finally, a continuous hard family converts tracking capacity into a zeroth-order query budget and gives a complementary capacity-starvation lower endpoint. The upper bounds require $C\ge \ln T+1$ and do not constitute a complete capacity minimax characterization.
Sang Su Lee, Vineeth Loganathan, Shishir Dash +1cs.LG
Personalizing marketing messages with contextual multi-armed bandits (CMABs) drives real business value, yet the objective that ultimately matters - a downstream conversion - is observed only weeks later, too late to drive online learning. Teams therefore train the bandit on a fast proxy reward, and separately must judge whether a contextual bandit is worth its complexity over sending one best message. Settling both decisions with the usual offline checks - a batch off-policy estimate, a marginal arm-discrimination test, a confidence interval - can mislead systematically under delayed feedback. We give an ordered diagnostic protocol that screens a reward-and-policy candidate on two axes, alignment (does optimizing the reward move the north-star?) and learnability (can the bandit identify the reward-optimal policy?), before trusting any reported lift. We validate it where the truth is known - a public off-policy-evaluation benchmark and a controllable synthetic generator - and illustrate it on a deployed large-marketplace push system (where, with five arms and one split, the evidence is directional rather than powered). Two lessons recur. (N1) A single offline number can mis-rank rewards: a denser reward signal gives the bandit more to learn from, so rewards that look tied in a static estimate pull apart once learning happens online. (N2) If you cannot tell in advance which single message is best, a per-user policy partly just avoids betting on the wrong one - that looks like personalization but is really robustness, so a "personalization premium" is easily overstated. Our contribution is methodological rather than algorithmic: the ordered protocol, the two lessons it surfaces, and the end-to-end experience of applying it to a delayed-feedback CMAB.
Xiaoyi Gu, Julia Tavares, Eder Santana +3cs.IR cs.LG
One of the most challenging problems entertainment live-streaming services face in recommendation systems is that user behaviors are sparse and delayed, and interaction data exhibits bias for different user segments. Unlike e-commerce applications where user actions follow linear sequences, live-streaming viewers engage in multiple concurrent behaviors of watching, chatting, following, and spending, each occurring with varying delays. We address these challenges through three key contributions: 1) a delayed window approach that extends feedback collection beyond immediate responses, 2) a multi-model architecture that combines fresh and delayed signals, and a segment-aware targeting module that optimizes ranking scores differently across user lifecycle stages, and 3) Multi-gate Mixture-of-Experts (MMoE) integration that jointly models correlated targets while reducing model parameters by 41.9% compared to independent models. Online A/B testing demonstrates significant improvements, including a +0.09% increase in Daily Active Viewers (DAV), generating millions more annual active viewer days, and +0.56% increase in highly engaged viewers' capped Average Revenue Per User (ARPU). Viewer-segment targeting achieved an additional +0.15% DAV improvement for newer and less engaged viewers, while MMoE enhancement added +0.08% overall DAV and +0.27% new follows. The proposed system processes ranking requests with low latency, providing a scalable approach for balancing multiple business objectives across diverse user populations. In addition, we tested the multi-model architecture on the Twitch mobile live feed and achieved a +1.12% increase in positive user-channel interactions (clicks, follows, and likes), demonstrating applicability beyond the primary use case.
Long-horizon conversion prediction under delayed feedback creates a two-clock, two-window learning problem in online advertising. A short base observation window releases recent clicks on the click clock before their outcomes mature, whereas conversions continue to arrive on the conversion clock throughout a longer target conversion window. The click clock provides timely but partially observed status supervision. The conversion clock reveals long-tail delays, but the delay composition within an arrival-time slice is weighted by historical click cohorts with different traffic volumes and target-window conversion rates. We present TWICE, a framework that factorizes long-horizon post-click conversion rate (CVR) into a target-window conversion probability and a grouped elapsed-delay cumulative distribution function (CDF). The two clocks provide complementary supervision. Click-clock records train the target-window CVR head through a current-status likelihood over the base observation window. Newly arrived conversions train the delay model on the conversion clock. To account for the cohort mixture, TWICE uses fixed click-time predicted CVR (pCVR) mass as cohort exposure in an arrival-conditioned likelihood. This accounts for differences in cohort traffic and conversion propensity. The resulting aggregate records are self-contained. A single learned CDF produces monotone predictions for all requested horizons up to the target conversion window. Serving requires neither historical lookup nor convolution. Experiments on a public benchmark and an industrial advertising dataset demonstrate the effectiveness of TWICE. In an online A/B test in Kwai's advertising system, TWICE increased expected revenue, revenue, and conversions by 2.486%, 1.858%, and 2.061%, respectively. It was subsequently deployed to full traffic.
State Space Models (SSMs) have emerged as a powerful paradigm for efficient long-sequence modeling, offering parallel training and fast linear-time recurrent inference. However, like other recurrent architectures, SSMs must compress an unbounded history into a fixed-size state, which limits context retention and makes precise retrieval over long-range context inherently difficult. To overcome this limitation, we propose Delay State Space Models (DSSMs), a delay differential equation (DDE)-inspired extension of diagonal SSMs that augments discrete SSM recurrences with explicit delayed-state feedback. Making explicit delayed feedback practical requires new stability parameterization, history management, and FFT-training tools. We address these challenges with a practical discretization and parameterization grounded in a simple delay-independent stability condition. To bypass direct time-domain kernel construction, we derive the DSSM transfer function and compute kernels in the frequency domain, using a kernel contour shift to suppress aliasing and recover accurate FFT training. Empirically, DSSMs substantially improve targeted delayed-retrieval tasks while outperforming S4D on most standard sequence metrics and remaining close on the others.
Reinforcement learning in real world environments often suffers from severe performance degradation due to delayed feedback. Existing approaches typically mitigate performance degradation caused by observation delays by constructing augmented states or predicting the true states. However, these methods often overlook the inherent discrepancy between delayed state and true states induced by stochastic MDP. We theoretically prove the existence of such a discrepancy and show that it leads to the degradation of the optimal policy. To address this challenge, we propose Diffusion Guided Uncertainty Aware Delayed Policy Optimization (DUPO). Our method explicitly models the relationship between delayed state message and the current state using a diffusion model, and leverages the resulting discrepancy estimates to weight delayed policies. Extensive experiments on continuous robotic control tasks with multiple stochastic delays demonstrate that DUPO consistently outperforms existing methods and remains effective even under long and random delay scenarios.
We study stochastic linear bandits with delayed feedback under several delay models and establish near-optimal regret guarantees. Our results identify when delayed linear bandits exhibit the same qualitative behavior as multi-armed bandits (MAB), and when the linear structure creates fundamentally new challenges. Specifically, (1) for \emph{loss-independent delays}, where the delay does not depend on the realized loss (but potentially depends on the arm), we show that delays incur only an additive regret penalty. Under stochastic delays, this penalty scales with the expected delay, while under adversarial delays, it scales with the maximum number of outstanding observations. Notably, both delay penalties are dimension-free, improving upon the state-of-the-art results; (2) for \emph{loss-dependent delays}, we show that linear bandits are substantially harder than MAB: unlike in MAB, we prove matching (up to log factors) upper and lower bounds in linear bandits, whose delay penalty depends on the square root of the dimension. (3) for the \emph{delay-as-payoff model}, a special case of loss-dependent delay, we show that the optimal MAB guarantee, which depends only on the delay of the optimal arm, is also unattainable in linear bandits. Together, these results provide a sharp characterization of how delayed feedback interacts with linear generalization.
Alexander Ryabchenko, Idan Attias, Daniel M. Roycs.LG stat.ML
Online learning with delayed feedback typically assumes that the learner can track all pending rounds until their feedback arrives. In practice, tracking resources are finite, and feedback from untracked rounds is permanently lost. In this paper, we study delayed online convex optimization (OCO) under a hard capacity constraint, where at most $C$ pending rounds can be tracked at any time. To model delay information, we introduce a semi-clairvoyant model that refines the clairvoyant assumption from prior work: rather than requiring delays to be known at prediction time, the learner observes delay expirations online, consistent with the classical unconstrained delayed setting. Our approach proceeds via a reduction to a novel ``delayed and weighted'' OCO problem, using a scheduler that randomizes tracking decisions and importance-weights the resulting observations. For this base problem, we propose and analyze Delayed-Weighted FTRL and its bandit analogue, establishing regret bounds that explicitly characterize the interaction between time-varying weights and delayed feedback. Combining these base learners with our schedulers yields the first regret guarantees for capacity-constrained OCO under convex and strongly convex losses, for both first-order and bandit feedback. For first-order feedback, capacity $C = Ω(\log T)$ suffices to recover standard delayed OCO rates up to logarithmic factors. For bandit feedback, the regret rates are modulated by powers of $(1 + σ_{\text{max}}/C)$, where $σ_{\text{max}}$ is the maximum number of pending observations at any time. This allows the regret bound to degrade gracefully when $C < σ_{\text{max}}$, while remaining sublinear.