Pith. sign in

REVIEW 3 major objections 6 minor 44 references

STEPS, a self-triggered agentic push system deployed at Douyin, jointly decides when to invoke itself and whether to send a notification, increasing user active days by 0.2843% while cutting push-permission revocations by 1.9089% and comput

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 17:51 UTC pith:BMK7TB6E

load-bearing objection Genuinely new agentic push architecture with real deployment and a 14-day A/B test, but the headline 'significant' gains are asserted without any uncertainty bounds — that needs to be fixed before the claims are credible. the 3 major comments →

arxiv 2608.01949 v1 pith:BMK7TB6E submitted 2026-08-03 cs.IR

A Self-Triggered Agentic Push Recommendation System

Zhao-Yu Zhang , Qingying Chen , Chunyuan Zheng , Jing Zhou , Jian Sun , Siqi Chen , Leiying Chen , Chuan Zhou
show 4 more authors
Huiyou Jiang Xin Tao Haoxuan Li Zhouchen Lin
This is my paper
classification cs.IR
keywords push notification schedulingself-triggered agentic systemdecision transformerordinal regressionreturn-to-govalue-guided learningonline A/B testinguser engagement
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

STEPS reformulates push notification delivery as a self-triggered closed loop: rather than pre-allocating delivery times offline or polling at fixed intervals, the system itself predicts the next moment it should re-examine each user (a time gap generated by a planning agent) and then, at that moment, decides whether to send a notification (an execution agent). The paper claims this joint control of 'whether' and 'when' outperforms both passive paradigms, based on a full deployment at Douyin with over one billion users. Reported gains include a 0.2843% increase in user active days, a 1.9089% reduction in push-permission revocations, and a 79.42% reduction in computational overhead (the last from a lightweight filtering agent that prunes low-value triggers before expensive item-sorting runs). The authors argue the end-to-end return signal, Bellman-updated value estimates, and suppression of overly short inter-push intervals are the mechanisms behind these gains. If correct, this points to a general design: let the system control its own invocation schedule instead of treating triggering as an external clock.

Core claim

The paper's central claim is that the 'whether and when' problem in push recommendation should be solved as a single closed-loop agentic process, not as two separate offline/online stages. In STEPS, a planning agent outputs the next activation gap using gated ordinal regression over 100 equal-frequency time buckets; an execution agent then decides whether to push at that moment using a value-guided decision transformer whose Q-functions are learned through Bellman recursion on binary rewards; and a 3-layer MLP filtering agent, distilled from the execution agent and deliberately excluding item features, blocks low-value triggers before they reach the expensive sorting pipeline. The paper repo

What carries the argument

The central mechanism is the self-triggered loop itself: the system's planning agent (a decision transformer, a sequence model that conditions actions on desired returns) generates the next invocation time Δt via gated ordinal regression, in which the return-to-go (RTG) is upscaled by an MLP and injected into the state embedding by element-wise multiplication, forcing the model to condition on the target rather than ignoring it. The execution agent learns a Q-function through Bellman updating on a binary reward (a user becomes active at a single time M), then reweights its action policy by the advantage of pushing versus not pushing. A distilled 3-layer MLP filtering agent operates on user/e

Load-bearing premise

The load-bearing premise is that the effect of a push on a user is fully captured by a single binary reward—the user becomes active at one specific future moment M (reward 1) or not (reward 0)—with already-active users always getting zero reward; if gradual, timing-dependent, or cumulative effects of push fatigue matter, the learned values and decisions are miscalibrated.

What would settle it

Record ground-truth multi-day engagement for a large cohort of users with varied push timing, then compute the paper's Bellman Q-values (binary reward at M) and a gradual, dose-dependent reward version (e.g., partial credit for sessions within 24 hours, penalties for frequency). If the binary-reward Q-values do not rank-order long-term active days better than the gradual version—or if an online A/B test with the gradual reward changes UAD/NE in a material direction—the binary reward proxy is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the reported results are correct, pre-planned frequency allocation and fixed-interval polling are dominated in settings where engagement and fatigue are both important, because the self-triggered loop adapts to real-time user state at lower cost.
  • End-to-end return collection means any request blocked by a downstream agent yields zero RTG, so the planning agent is implicitly trained to propose time gaps that survive the full pipeline.
  • Shifting inter-push intervals away from 0–20 minutes toward 3–6 hours is, on the paper's evidence, the primary lever for reducing push-permission revocations.
  • Early value-aware pruning without item features can cut compute by roughly three-quarters while slightly improving the negative-experience metric, suggesting that cheap pre-filters are viable in large-scale recommendation.
  • Bellman-RTG estimation avoids the early spike then decay of observed-RTG training, stabilizing long-term engagement value in online tests.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The binary reward model (single activation moment M) is the paper's simplifying assumption; if real re-engagement is gradual or dose-dependent, the learned Q-values may misestimate long-term effects even if headline metrics improve.
  • The self-triggered paradigm could extend to other proactive channels—email digests, reminders, ads—where the system currently relies on external scheduling; a testable extension is to replace fixed send-time calendars with a learned next-check policy.
  • A direct falsification experiment would compare the Bellman Q-values trained on the binary reward against a reward model with partial credit for multi-day engagement; if the binary version underperforms on a held-out long-term retention measure, the reward proxy is the weak link.
  • The filtering agent's 74.88% request drop at equal final send volume implies many downstream execution calls are near-sure rejections; similar early-pruning layers might generalize to other cascading ranking systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes STEPS, a self-triggered agentic push-notification system deployed at Douyin. STEPS replaces pre-planned frequency allocation and fixed-interval polling with two decision-transformer agents: a planning agent that predicts the time gap until the next system invocation via gated ordinal regression, and an execution agent that decides whether to send a push using a Bellman-updated value function and advantage-weighted policy loss. A lightweight filtering agent prunes low-value requests before the expensive item-sorting pipeline. The paper reports a 14-day online A/B test against the production pre-planned frequency baseline, claiming a +0.2843% increase in User Active Days, a -1.9089% reduction in permission revocations (NE), and a -79.42% reduction in resource consumption. Ablations attribute the largest resource savings to the filtering agent and decompose the UAD/NE gains among planning, execution, and filtering components.

Significance. If the reported deployment and effect sizes are reliable, this is a substantial industrial contribution: it demonstrates a closed-loop, self-triggered alternative to both dominant push-scheduling paradigms, with large-scale online deployment, careful component-level ablations, and a detailed treatment of computational overhead (Tables 2, 6, 7; Figure 8). The gated-RTG injection and ordinal-regression formulation are clearly motivated, and the paper includes useful sensitivity analyses for the ordinal bucket configuration and filtering threshold. However, the central empirical evidence—the online A/B result—currently lacks inferential statistics, and the execution agent's reward model rests on a strong binary-activation assumption that is validated only against a target constructed from the same assumption. The contribution is credible but not yet fully supported as stated.

major comments (3)
  1. [§4, Table 1] The central empirical claim—that STEPS "significantly" improves UAD and NE—is not supported by the reported statistics. Table 1 gives only relative point changes (+0.2843% UAD, -1.9089% NE, -79.42% resource) with no confidence intervals, standard errors, p-values, per-arm sample sizes, or description of variance-reduction techniques. NE is a rare-event count (permission revocations), so a -1.9089% relative change could be a very small absolute count and could fall within daily noise; the UAD lift is also small in relative terms. Please report absolute counts, 95% confidence intervals, and significance tests, or explicitly state that the reported numbers are descriptive. The word "significantly" in the Abstract and §4 should be removed or justified. The same issue applies to the ablation rows in Table 2, which are used to attribute gains to individual agents.
  2. [§3.3, displayed equations for r+_t and R+_t] The execution agent's Q-values and advantage weights are built on a binary activation reward: r+_t = 1 only at a single time M and zero otherwise, and R+_t = 0 for users already active. This collapses long-term engagement into one conversion moment and assigns zero value to maintaining already-active users, ignoring the gradual, dose-dependent, and timing-dependent effects that are central to push fatigue and re-engagement. The paper's calibration check (Figure 7a) compares predicted Q-values to "ground-truth RTG," but that RTG is constructed from the same binary reward assumption, so it does not validate the proxy against an independent long-term outcome. Please provide a sensitivity analysis under alternative reward definitions (e.g., multi-day activity windows, decayed activity counts) or an offline validation against a held-out long-term metric, and discuss how misspecification of th
  3. [§4.1, Table 1] The fixed-interval baseline comparison may be confounded by resource consumption. The text states that the fixed-interval polling frequency was tuned to align resource consumption "as closely as possible" with the baseline, yet Table 1 reports +6.548% resource consumption for the fixed-interval method, with no measurement units, confidence intervals, or description of the tuning procedure. Since the paper's paradigm-level claim is that self-triggering dominates fixed-interval polling at comparable cost, please report absolute resource units (e.g., QPS, CPU cost, or normalized cost), show the UAD/NE comparisons at matched resource budgets, and state precisely how "resource consumption" is measured.
minor comments (6)
  1. [§3.1–3.3] Several key hyperparameters are never given numeric values: λ_n, λ_1, λ_2, γ, α_r, the filtering threshold, and the hardcoded boundary-control windows. Model architecture details (transformer layers, context length, training steps) are also missing. Please add a hyperparameter table and, where feasible, sensitivity analyses for λ_1, λ_2, γ, and α_r, or state that these are fixed in deployment.
  2. [§3.1/§3.3] The notation for R_t is used inconsistently: R_t is first defined as a sum over j=t..T of (r+_j - λ_n r-_j), and later as R_t = R+_t - λ_n R-_t while L_Q regresses R_{t+1} against Q(s_t,a_t|λ_n). Please align the time indexing and clarify whether R_t in the policy loss is the same object as the Bellman target.
  3. [§4.3, Figure 5] Figure 5a lacks axis labels and units, and the text's "Regression AUC" does not match the y-axis label "Offline RegAUC" in Figure 5b. The correlation-ratio statistic is reported as a point estimate without error bars; a single value near 1.8 does not by itself establish a stable condition-action alignment.
  4. [Tables 3 and 4] Pass-rate changes and inter-push-gap distribution changes are reported as relative percentages without absolute base rates or confidence intervals. Please provide base rates or at least the absolute shares for the reported intervals, since relative changes alone are difficult to interpret.
  5. [§4.4, Figure 6b] The claim that Bellman RTG "exhibits a continuous and steady rise" describes a figure curve without quantitative endpoints or a statistical comparison. Please report the final increments and associated uncertainty, and specify how many days of the 14-day A/B are shown.
  6. [Throughout] Copyediting issues include missing spaces in "whetherandwhento" and "determinewhether", and inline LaTeX-like text in Figure 2 ("withitem featureswithoutitem features"). Please proofread the camera-ready version.

Circularity Check

0 steps flagged

No significant circularity: the reported gains are external A/B measurements; the only self-citation is the baseline, which is not load-bearing.

full rationale

The paper's derivation chain is not circular. STEPS defines a training objective (L_next, L_Q, L_pi) and then measures UAD/NE/resource in online A/B tests against a production baseline; those metrics are external measurements, not outputs of the model equations. The only self-citation is reference [15], used as the pre-planned-frequency production baseline (Section 4.1). This is not load-bearing for the central claim because the comparison is a live randomized experiment, not a deduction from [15]'s equations; even if [15] were replaced by an independent description of the same production system, the result would stand. The ablation table (Table 2) sums exactly to the headline table (Table 1) — an accounting identity, not a prediction. The correlation-ratio analysis (Figure 5a) and pass-rate analysis (Table 3) are sanity checks on the trained model and reflect its training objective, but they are not presented as independent predictions of the headline metrics. The main weakness is statistical: 'significantly' in the abstract is not backed by confidence intervals or p-values in Table 1. That is a missing-support concern, not circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The paper's gains depend heavily on tuned engineering hyperparameters (loss weights, thresholds, bucket count) and on a simplified binary reward model. The scientific claim is not derived from first principles; it is an empirical system report.

free parameters (8)
  • lambda_n = sampled from U(0,1); no fixed value
    Trade-off weight between positive and negative rewards; conditioned on by both agents; adjustable online to trade off business objectives.
  • lambda_1
    Weight for Q-learning loss L_Q in total loss L = L_next + lambda_1 L_Q + lambda_2 L_pi; value not reported.
  • lambda_2
    Weight for policy loss L_pi; value not reported.
  • gamma
    Discount factor in Bellman equation for Q-learning; value not reported.
  • alpha_r
    Scaling hyperparameter in the advantage weight w_t = Sigmoid(alpha_r * (Q(s_t,a_t) - Q(s_t,1-a_t))); value not reported.
  • K (ordinal bucket count) = 100
    Number of equal-frequency ordinal buckets; selected from sensitivity analysis in Figure 5b.
  • Filtering threshold = minute-scale trigger gap
    Fixed threshold on filter score; tuned in Figure 8 to balance UAD and QPS; final choice described as conservative minute-scale.
  • Boundary control windows = minutes, hardcoded
    System-side and user-side suppression windows after recent push or natural activation; described as hardcoded minute-scale thresholds.
axioms (5)
  • domain assumption Decision Transformer conditional sequence modeling can represent near-optimal push policies
    The paper relies on DT as the backbone for both agents; this is an empirical claim from prior work, not proven here.
  • standard math Bellman equation for Q-learning is valid with function approximation on offline data
    Used in Section 3.3 to define R+_t and L_Q; standard but with known instability caveats.
  • domain assumption Offline production logs are representative of online traffic and user behavior
    Training uses six months of logs; the A/B test period may differ; this is assumed throughout.
  • ad hoc to paper Users can be modeled as binary active/inactive with a single activation time M
    Section 3.3 defines r+_t = 1 only at the moment of activation; this is a strong simplification introduced for this paper.
  • domain assumption The Gated-RTG multiplication e_s ⊙ MLP(R_t) forces RTG conditioning without loss of state information
    Proposed in Section 3.2; the paper provides empirical evidence but no theoretical guarantee.

pith-pipeline@v1.3.0-daily-deepseek · 13509 in / 14347 out tokens · 150585 ms · 2026-08-04T17:51:29.908170+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of A Self-Triggered Agentic Push Recommendation System." pith.science (2026). https://pith.science/paper/BMK7TB6E

@misc{pith2026260801949,
  author       = {Pith},
  title        = {Pith review of: A Self-Triggered Agentic Push Recommendation System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BMK7TB6E}},
  note         = {Machine review of arXiv:2608.01949}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Push notification is a critical recommendation scenario on large-scale platforms, allowing the system to proactively reach users outside the application to improve long-term re-engagement. However, designing an optimal push system requires handling a complex action space for the "whether and when" delivery problem under strict system resource constraints. Existing solutions typically fall into two passive paradigms: pre-planned frequency methods that allocate delivery times via offline modeling, limiting real-time adaptability; and fixed-interval triggering methods that periodically poll the system, creating a strict dilemma between excessive computational overhead and diminished optimal timing capture. Furthermore, such multi-stage frameworks severely suffer from local optima. To overcome these limitations, in this paper, we propose STEPS, a proactive, Self-Triggered End-to-end Agentic Push Recommendation System, which is already fully deployed at Douyin with over 1 billion users. STEPS reformulates push recommendation as a self-triggered agentic process in which the system decides not only whether to send a push, but also when to invoke itself again, thereby forming a closed loop that balances real-time effectiveness and efficiency. Specifically, STEPS consists of two decision transformer-based agents: a planning agent that schedules the next system invocation using a gated ordinal regression method, and an execution agent that decides whether to send a push based on trajectory rewards. Furthermore, we introduce a lightweight filtering agent to both control computational overhead and act as a crucial safeguard against unreasonable planning behaviors. Online A/B testing demonstrates that STEPS significantly increases user active days by 0.2843% and reduces the push permission disablement rate by 1.9089%, while the filtering agent reduces computational overhead by 79.42%.

Figures

Figures reproduced from arXiv: 2608.01949 by Chuan Zhou, Chunyuan Zheng, Haoxuan Li, Huiyou Jiang, Jian Sun, Jing Zhou, Leiying Chen, Qingying Chen, Siqi Chen, Xin Tao, Zhao-Yu Zhang, Zhouchen Lin.

Figure 1
Figure 1. Figure 1: Push notification examples. and user activity, but poorly timed ones lead to push notification disablement because of user fatigue [4, 25, 28]. Solving the "whether and when" problem is highly non-trivial, since push delivery is a continuous and sequenced optimization problem [1, 17]. To find the optimal delivery moment, a naive ap￾proach would be to evaluate every user every second. However, in large-scal… view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework. methods in large-scale industrial systems can sometimes present challenges, as these approaches may be susceptible to unstable bootstrapping and sensitivity to distribution shifts in offline logs. To address this, recent generative RL methods offer an alterna￾tive by formulating sequential decision-making as a conditional sequence modeling problem. Decision transformer (DT) [8] pre￾dicts… view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of planning agent. 3.2 Planning Agent The planning agent acts as the proactive engine of our agentic frame￾work. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of the execution agent. we seamlessly integrate these event-driven triggers with the agent￾generated timestamps to ensure a responsive and comprehensive user experience. Second, our platform handles millions of queries per second (QPS). Because our system discards fixed budget limits and pre￾defined delivery windows, there is an inherent risk of anomalous temporal planning (e.g., generating in… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the effectiveness of RTG-action cor [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Analysis across push time and RTG calculation. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Analysis of the learned 𝑄 value. and steady rise, stabilizing consistently above the baseline. Since both evaluated methods employ the Gated-RTG mechanism, this sustained long-term improvement can be primarily attributed to the Bellman RTG formulation, demonstrating its effectiveness in optimizing cumulative trajectory value. Alleviation of Extreme Sending Behaviors. In addition to opti￾mizing long-term re… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 3 linked inside Pith

  1. [1]

    Sami Abboud, Eleanor Hanna, Olivier Jeunen, Vineesha Raheja, and Schaun Wheeler. 2025. Agentic Personalisation of Cross-Channel Marketing Experiences. InACM Conference on Recommender Systems

    2025

  2. [2]

    Axel Abels, Diederik Roijers, Tom Lenaerts, Ann Nowé, and Denis Steckelmacher

  3. [3]

    Michal Aharon, Yohay Kaplan, Rina Levy, Oren Somekh, Ayelet Blanc, Neetai Eshel, Avi Shahar, Assaf Singer, and Alex Zlotnik. 2019. Soft Frequency Capping for Improved Ad Click Prediction in Yahoo Gemini Native. InACM International Conference on Information and Knowledge Management

    2019

  4. [4]

    Renee Barnes, Rory Mulcahy, and Aimee Riedel. 2025. Push Notifications and News Snacking: The Impact of Mobile News Alert Framing on Reader Engage- ment.New Media & Society27, 3 (2025), 1486–1506

    2025

  5. [5]

    Stephen Bonner and Flavian Vasile. 2018. Causal Embeddings for Recommenda- tion. InACM Conference on Recommender Systems

    2018

  6. [6]

    Paul-Christian Bürkner and Matti Vuorre. 2019. Ordinal Regression Models in Psychology: A Tutorial.Advances in Methods and Practices in Psychological Science2, 1 (2019), 77–101

    2019

  7. [7]

    Jiaju Chen, Wenjie Wang, Chongming Gao, Peng Wu, Jianxiong Wei, and Qing- song Hua. 2024. Treatment Effect Estimation for User Interest Exploration on Recommender Systems. InInternational SIGIR Conference on Research and Devel- opment in Information Retrieval

    2024

  8. [8]

    Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. 2021. Decision Transformer: Reinforcement Learning Via Sequence Modeling. InAdvances in Neural Information Processing Systems

    2021

  9. [9]

    Jingtong Gao, Yewen Li, Shuai Mao, Peng Jiang, Nan Jiang, Yejing Wang, Qing- peng Cai, Fei Pan, Peng Jiang, Kun Gai, Bo An, and Xiangyu Zhao. 2025. Gen- erative Auto-Bidding With Value-Guided Explorations. InInternational SIGIR Conference on Research and Development in Information Retrieval

    2025

  10. [10]

    Rupesh Gupta, Guanfeng Liang, Hsiao-Ping Tseng, Ravi Kiran Holur Vijay, Xi- aoyu Chen, and Rómer Rosales. 2016. Email Volume Optimization at LinkedIn. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2016

  11. [11]

    Chengkai Huang, Junda Wu, Yu Xia, Zixu Yu, Ruhan Wang, Tong Yu, Ruiyi Zhang, Ryan A Rossi, Branislav Kveton, Dongruo Zhou, et al . 2025. Towards agentic recommender systems in the era of multimodal large language models. arXiv:2503.16734(2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Eugene Ie, Vihan Jain, Jing Wang, Sanmit Narvekar, Ritesh Agarwal, Rui Wu, Heng-Tze Cheng, Tushar Chandra, and Craig Boutilier. 2019. SlateQ: A Tractable Decomposition for Reinforcement Learning With Recommendation Sets. InIn- ternational Joint Conferences on Artificial Intelligence

    2019

  13. [13]

    Michael Janner, Qiyang Li, and Sergey Levine. 2021. Offline Reinforcement Learn- ing as One Big Sequence Modeling Problem. InAdvances in Neural Information Processing Systems

    2021

  14. [14]

    Datla, Sadid A

    Kathy Lee, Ashequl Qadir, Vivek V. Datla, Sadid A. Hasan, Joey Liu, Aaditya Prakash, and Oladimeji Farri. 2016. Assorted Textual Features and Dynamic Push Strategies for Real-Time Tweet Notification. InText REtrieval Conference

    2016

  15. [15]

    Xiang Li, Zhao-Yu Zhang, Chunyuan Zheng, Qingying Chen, Huiyou Jiang, Haoxuan Li, and Zhouchen Lin. 2026. User Activity Modeling Under Inflated Distribution. InInternational SIGIR Conference on Research and Development in Information Retrieval

    2026

  16. [16]

    Babak Loni, Anne Schuth, Lucas de Haas, Jeroen Jansze, Vasco Visser, and Marlies van der Wees. 2019. Personalized Push Notifications for News Recommenda- tion. InWorkshop on Online Recommender Systems and User Modeling at ACM Conference on Recommender Systems

    2019

  17. [17]

    Bayram Berkay Mumcu and Ayça Çebi. 2025. You Have a Notification: The Role of Push Notifications in Shaping Students’ Engagement, Self-Regulation and Academic Procrastination.International Journal of Educational Technology in Higher Education22, 1 (2025), 36

    2025

  18. [18]

    Patra, Haotian Shen, David Liu, Yiwen Yuan, Gokulraj Mohanasundaram, Fedor Borisyuk, and Prakruthi Prabhakar

    Borja Ocejo, Ruofan Wang, Ke Liu, Rohit K. Patra, Haotian Shen, David Liu, Yiwen Yuan, Gokulraj Mohanasundaram, Fedor Borisyuk, and Prakruthi Prabhakar. 2025. Generative Sequential Notification Optimization Via Multi-Objective Decision Transformers. arXiv:2509.02458

    Pith/arXiv arXiv 2025

  19. [19]

    Tadashi Okoshi, Kota Tsubouchi, and Hideyuki Tokuda. 2019. Real-World Product Deployment of Adaptive Push Notification Scheduling on Smartphones. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2019

  20. [20]

    Xuan-Lam Pham, Thi-Huyen Nguyen, Wu-Yuin Hwang, and Gwo-Dong Chen

  21. [21]

    Martin Pielot, Amalia Vradi, and Souneil Park. 2018. Dismissed!: A Detailed Ex- ploration of How Mobile Phone Users Handle Push Notifications. InInternational Conference on Human-Computer Interaction with Mobile Devices and Services

    2018

  22. [22]

    Prakruthi Prabhakar, Yiping Yuan, Guangyu Yang, Wensheng Sun, and Ajith Muralidharan. 2022. Multi-Objective Optimization of Notifications Using Offline Reinforcement Learning. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2022

  23. [23]

    Yuta Saito, Suguru Yaginuma, Yuta Nishino, Hayato Sakata, and Kazuhide Nakata

  24. [24]

    David Silver and Richard S Sutton. 2025. Welcome to the era of experience.Google AI(2025)

    2025

  25. [25]

    Natalie Jomini Stroud, Cynthia Peacock, and Alexander L Curry. 2020. The Effects of Mobile Push Notifications on News Consumption and Learning. InMobile News. Routledge, 32–48

    2020

  26. [26]

    Konglong Tang, Yong Wang, Hao Liu, Yanxiu Sheng, Xi Wang, and Zhiqiang Wei. 2013. Design and Implementation of Push Notification System Based on the MQTT Protocol. InInternational Conference on Information Science and Computer Applications

    2013

  27. [27]

    Wenjie Wang, Fuli Feng, Xiangnan He, Xiang Wang, and Tat-Seng Chua. 2021. Deconfounded Recommendation for Alleviating Bias Amplification. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2021

  28. [28]

    Dawn Wheatley and Raul Ferrer-Conill. 2021. The Temporal Nature of Mobile Push Notification Alerts: A Study of European News Outlets’ Dissemination Patterns.Digital Journalism9, 6 (2021), 694–714

    2021

  29. [29]

    Atilla Wohllebe. 2020. Consumer Acceptance of App Push Notifications: System- atic Review on the Influence of Frequency.International Journal of Interactive Mobile Technologies14, 13 (2020), 36–47

    2020

  30. [30]

    Atilla Wohllebe, Dirk Siegfried Hübner, Uwe Radtke, and Szilárd Podruzsik

  31. [31]

    Taku Yamagata, Ahmed Khalil, and Raul Santos-Rodriguez. 2023. Q-Learning De- cision Transformer: Leveraging Dynamic Programming for Conditional Sequence Modelling in Offline RL. InInternational Conference on Machine Learning

    2023

  32. [32]

    Kevin P Yancey and Burr Settles. 2020. A Sleeping, Recovering Bandit Algorithm for Optimizing Recurring Notifications. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2020

  33. [33]

    Yiping Yuan, Ajith Muralidharan, Preetam Nandy, Miao Cheng, and Prakruthi Prabhakar. 2022. Offline Reinforcement Learning for Mobile Notifications. In ACM International Conference on Information and Knowledge Management

    2022

  34. [34]

    Yiping Yuan, Jing Zhang, Shaunak Chatterjee, Shipeng Yu, and Romer Rosales

  35. [35]

    Yuguang Yue, Yuanpu Xie, Huasen Wu, Haofeng Jia, Shaodan Zhai, Wenzhe Shi, and Jonathan J Hunt. 2022. Learning to Rank For Push Notifications Using Pairwise Expected Regret.arXiv preprint arXiv:2201.07681(2022)

    Pith/arXiv arXiv 2022

  36. [36]

    Bo Zhao, Koichiro Narita, Burkay Orten, and John Egan. 2018. Notification Vol- ume Control and Optimization System at Pinterest. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2018

  37. [37]

    Huasha Zhao, Luo Si, Xiaogang Li, and Qiong Zhang. 2017. Recommending Complementary Products in E-Commerce Push Notifications With a Mixture Model Approach. InInternational SIGIR Conference on Research and Development in Information Retrieval

    2017

  38. [38]

    Lixin Zou, Long Xia, Zhuoye Ding, Jiaxing Song, Weidong Liu, and Dawei Yin

  39. [39]

    InACM International Conference on Web Search and Data Mining

    A State Transition Model for Mobile Notifications Via Survival Analysis. InACM International Conference on Web Search and Data Mining

  40. [44]

    InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Reinforcement Learning to Optimize Long-Term User Engagement in Recommender Systems. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

  41. [2016]

    InIEEE International Conference on Advanced Learning Technologies

    Effects of Push Notifications on Learner Engagement in a Mobile Learning App. InIEEE International Conference on Advanced Learning Technologies

  42. [2019]

    InInter- national Conference on Machine Learning

    Dynamic weights in multi-objective deep reinforcement learning. InInter- national Conference on Machine Learning

  43. [2020]

    InACM International Conference on Web Search and Data Mining

    Unbiased Recommender Learning From Missing-Not-at-Random Implicit Feedback. InACM International Conference on Web Search and Data Mining

  44. [2021]

    Mobile Apps in Retail: Effect of Push Notification Frequency on App User Behavior.Innovative Marketing17, 2 (2021), 102–111

    2021