PPDL: A Real-world Industrial User Retention Ratio Forecasting Framework Integrating Physical Priors with Deep Learning
Zibo Zhao, Zhengxiong Guan, Chaoli Zhang, Linyuan Geng, Xuanbing Zhu, Zhonglong Zheng, Fan Wu
cs.LG
Quick verdict
Weibull-prior DL framework for 335-day retention forecasting reports strong industrial gains; proprietary data, thin baselines and weak stats limit it.
Summary
This paper proposes PPDL, a forecasting framework for channel-level user retention on a large short-video platform: given 30 days of observed retention for a paid-acquisition cohort, it predicts the remaining 335 days so that LT365 (cumulative 365-day retention) can be estimated early for budget allocation. PPDL decomposes each curve into trend and residual. The trend is a Weibull survival curve whose shape, scale, and floor parameters are initialized by Theil-Sen regression on the log-linearized form and refined by a gated MLP conditioned on covariates (channel IDs, Gregorian and lunar calendar features, event indicators). The residual is forecast by a TSMixer backbone augmented with channel, phase, and joint auxiliary embeddings, trained with a multiscale trend-penalized MSE loss. On three proprietary industrial datasets (30+ channels per app, hundreds of millions of users), PPDL attains an average weighted cumulative MAPEa of 5.07%, versus 6.67% for its TSMixer backbone and 6.33% for the online statistical model. Ablations attribute -19.23% (decomposition), -9.97% (embeddings), and only -2.23% (trend loss, exactly zero on App2) relative reductions. Main limitations: proprietary data with no code release, thin baselines, no variance or significance reporting, and PPDL on iTransformer/TFT backbones (7.70%/6.43% average) does not clearly beat plain TSMixer (6.67%).
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- PPDL is the first deep-learning time-series framework for channel-level user retention ratio forecasting (stated in Secs. I and V).
- Modeling the trend as a Weibull survival curve with Theil-Sen initialization and gated-MLP refinement injects a decay-then-saturation prior that compensates for the 30-day look-back window.
- On three proprietary apps, PPDL reduces average weighted cumulative MAPEa to 5.07% versus 6.67% for the TSMixer backbone and 6.33% for the deployed online statistical model (Tables I/VII).
- Ablations attribute the largest share of improvement to trend-residual decomposition (19.23% average relative reduction), followed by auxiliary embeddings (9.97%) and the multiscale trend loss (2.23%, exactly zero on App2).
- The framework generalizes across backbones, improving each backbone relative to itself (Table VIII), though PPDL(I) at 7.70% still trails plain TSMixer at 6.67%.
Strengths
Well-motivated industrial task with clear business framing
Predicting LT365 at day 30 (Sec. I) is a concrete, high-value decision problem; the MAPEa metric (Eq. 10), activation-volume weighted and computed on cumulative day-30-365 retention, directly matches the budget-allocation use case.
Sensible two-stage trend estimator
Sec. III-B's Theil-Sen initialization on the log-linearized Weibull form, followed by bounded gated-MLP corrections with projection, is a defensible way to stabilize (k, λ, f) estimation from only 30 noisy points; covariates (dual calendars, events, channel IDs, Sec. II) are domain-appropriate.
Unusually candid ablation suite
Consistent gains across apps and backbones relative to each model's own baseline
Weaknesses
No uncertainty quantification anywhere
Thin baselines; the decisive comparison is missing
Inconsistent relative-reduction arithmetic
Overstated theoretical framing and priority claim
Not reproducible as published
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
PPDL targets a concrete industrial task: for each paid-acquisition channel of a short-video platform, forecast a user cohort's retention curve for days 31-365 from only 30 days of observation, so that LT365 (defined in Sec. I as the cumulative sum LTn = Σ R(t)) can be estimated early enough to steer budget allocation. The paper claims three contributions: (1) a trend-residual decomposition whose trend is a Weibull survival curve with MLP-refined parameters; (2) auxiliary channel/phase/joint embeddings over a TSMixer backbone; (3) a multiscale trend-penalized loss. It also claims — twice (Sec. I, Sec. V) — to be the first deep-learning time-series approach to channel-level retention forecasting, and reports that PPDL beats both its backbone and a deployed statistical model on three proprietary apps (Table I: MAPEa 5.07% vs 6.67% and 6.33%).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the PPDL framework?
PPDL is a forecasting framework from Zhejiang Normal University, Douyin Group, and Shanghai Jiao Tong University researchers that predicts, for each paid user-acquisition channel of a short-video app, the user retention curve for days 31-365 using only the first 30 days of data. It combines a Weibull-distribution trend model (parameters refined by a gated MLP), channel/phase/joint auxiliary embeddings on a TSMixer backbone, and a multiscale trend-penalized loss function.
How much more accurate is PPDL than existing methods?
On three industrial datasets from a large short-video platform (30+ channels per app), PPDL achieves an average weighted cumulative-retention MAPE of 5.07%, versus 6.67% for the TSMixer backbone alone and 6.33% for the previously deployed online statistical model. Gains are largest on the biggest app (App1: 3.6% vs 5.7%).
Is the PPDL code or dataset publicly available?
No. All experiments use proprietary production data from a short-video platform (three apps with hundreds of millions of users), and the paper mentions no code release or public benchmark, so the reported numbers cannot currently be independently reproduced.
Why does the paper use a Weibull distribution for retention curves?
Retention curves decay sharply and then plateau. The authors argue via the Fisher-Tippett-Gnedenko extreme-value theorem that retention duration — viewed as the minimum of many competing risks — converges to a Weibull form; they acknowledge the assumptions may not strictly hold, but find the Weibull shape empirically captures the decay-then-saturation trend, with an MLP refining its parameters from data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are present only as captions in the extraction (images not rendered); the visual claims about curve-tracking behavior in Sec. IV-F could not be verified against the actual plots.
- Equations are OCR-mangled (e.g., Eq. 2 renders with stray symbols), but all parameter definitions were recoverable from surrounding prose, so the assessment is unaffected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.