When a new user joins a recommendation platform, the system knows almost nothing about them. Most research on this cold-start problem asks how accurately the system can predict ratings or clicks after the user has supplied a handful of interactions. That question hides a selection effect. Many new users never supply that handful: they see a few bad items and leave. Benchmarks and training sets built from users with enough history for adaptation contain only the survivors of onboarding. We propose to treat onboarding as a race between two mutually exclusive outcomes. Either the user "warms up", accumulating enough satisfied interactions for collaborative filtering to work, or the user churns first. Competing-risk analysis is the part of survival statistics built for exactly this kind of race. It models the probability that one event happens before another while correctly handling users whose story is still unfinished when the log ends.
The core claim has two parts, and both can be tested. The first is about measurement. Evaluating cold-start methods only on users who survived long enough to be evaluated changes which method looks best, and a competing-risk metric on all new users gives a different and more decision-relevant ranking. The second is about method. A cold-start recommender that ranks items by their predicted effect on the probability of warming up before churning should raise that probability over recommenders trained on click or rating targets, including the strongest content-conditioned and meta-learned baselines. The predictions must be calibrated and must handle unfinished users properly. The work is valuable if either part holds. The first part alone would change how cold-start results are reported.
Core Claim and Formulation
For a new user, index their impressions by step $j = 1, 2, \dots$. After each impression, one of three things happens. The user continues in the feed ($E_j = 0$), ends the session and returns later ($E_j = 1$), or ends the session and does not return within a fixed window, which we call churn ($E_j = 2$). Separately, the impression may or may not be a satisfied interaction (a long view or a like). The user becomes warm once $m$ satisfied interactions have accumulated. Warm-up and churn then compete as absorbing events, and users still active when the log ends are right-censored, meaning their final outcome is unobserved.
A discrete-time competing-risk model predicts, for user context $x$ and candidate item $i$, the per-step outcome probabilities $h_k(j \mid x, i) = P(E_j = k \mid \text{still active at } j, x, i)$. The quantity we optimize is the cumulative incidence of warm-up by horizon $\tau$, $\mathrm{CIF}_{W}(\tau \mid x, \pi)$: the probability that a user with context $x$, shown items by policy $\pi$, reaches the warm state before churning and before $\tau$. The ranking rule is myopic, scoring item $i$ at step $j$ by
$$ s(i) = p_{\text{sat}}(i \mid x) \, V_{m - c_j} - \lambda \, h_2(j \mid x, i), $$
where $p_{\text{sat}}$ is the predicted satisfaction probability, $c_j$ is the number of satisfied interactions so far, $V_{r}$ is a learned estimate of how much one more satisfied interaction raises the warm-up probability when $r$ remain, and $\lambda \geq 0$ trades off immediate churn risk. This is deliberately simple: it is the one-step lookahead of the race objective, not a full reinforcement-learning policy.
The motivation for a competing-risk treatment comes from the clinical literature. There, counting a competing event as ordinary censoring removes subjects from the at-risk group as if they could still experience the target event. This distorts cumulative incidence, as in the secondary-cancer analysis of Risk and Prognosis of Secondary Bladder Cancer After Radiation Therapy for Rectal Cancer, which used the Fine–Gray subdistribution model with death as a competing risk. The lymphoma cohort in Type 2 diabetes mellitus compromises the survival of diffuse large B-cell lymphoma patients treated with (R)-CHOP makes a related point. A survival gap there was driven by non-relapse mortality (22.5% vs 8.4%) while relapse was similar, and only a competing-risk decomposition revealed that. The analogy we test is that a gap in cold-start success between two recommenders may come from churn rather than from how accurately they model preference, and single-outcome metrics cannot tell these apart.
Relation to Prior Work
Survival models have entered recommendation before, always with a single event. Recommendation method for extending subscription periods combined a Cox proportional-hazards retention model with an interest model to extend manga subscriptions. It treated churn as one event, and it handled new users only by steering them toward sequences typical of long-term subscribers. Deep Recurrent Survival Analysis (DRSA) predicts per-interval hazards with a recurrent network, including user return times on a music service. It names competing risks as future work and evaluates only users with histories. Our hazard model is essentially DRSA's discrete-time construction with a multi-outcome head, applied to users with almost no history. We reuse its method and change its regime.
Three lines of work model engagement beyond clicks without a competing-risk structure. Reweighting Clicks with Dwell Time in Recommendation uses fixed dwell-time thresholds to define valid reads, and its production deployment showed that reading time can fall while clicks rise. KuaiSim simulates session exit with a patience counter and models the delay before a user returns with a geometric distribution. KuaiRand supplies the randomly exposed short-video logs that make unbiased per-impression estimation possible. We use the last two as evaluation infrastructure. We replace the heuristic exit and threshold machinery of the first two with a fitted, censoring-aware hazard model.
On the cold-start side, three methods supply our backbones and main competitors. DropoutNet randomly zeroes the learned preference vectors during training, so the network learns to predict from content features when those vectors are missing. MAMO uses memory modules to personalize the starting parameters of a meta-learned recommender, adapting from 5 to 15 interactions. BC-ICL turns frozen tabular foundation models into contextual bandits by feeding them resampled interaction histories. All three optimize immediate targets: rating error, reconstruction of a warm model's scores, or scalar reward. MAMO's adaptation window also presupposes a user who stays for several interactions. We keep a DropoutNet-style content-conditioned encoder and change only what it is trained to predict and how the predictions are used.
Two competing-risk papers constrain the design. On the calibration of survival models with competing risks shows that good discrimination often hides badly calibrated absolute risks for rare outcomes. It proposes CR D-calibration and a marginal calibration check against the Aalen–Johansen estimator, the standard non-parametric estimate of cumulative incidence under competing risks, along with post-hoc recalibration. Churn at a given step is exactly such a rare outcome, and our score uses absolute probabilities, so we adopt that recalibration. The same paper warns that its calibration metric assumes horizons long enough for incidence to saturate, which our short onboarding horizon violates. We therefore rely mainly on the Aalen–Johansen comparison at fixed horizons. Discrete Survival Knowledge Distillation for Competing Risks Analysis shows that hazards learned by a model on data-rich subjects can be distilled into a model for data-poor ones, but only when the variance reduction outweighs the teacher's bias. We use this as an optional ingredient: a teacher trained on warm users' later sessions, distilled into the cold-start hazard head. We expect it to be a bias-variance question and test it as such.
Research Program
The first component is a measurement study on KuaiRand-27K and KuaiRand-Pure. We define new users as those whose first logged day falls in a late window, reconstruct sessions from timestamps, and label each impression with satisfaction and exit type. For each user we compute whether and when warm-up or churn occurred, or whether the user was censored. We report what fraction of new users would be excluded by a minimum-interaction filter of the kind cold-start benchmarks use, and how different the excluded users are. We then estimate cumulative incidence of warm-up two ways: with the Aalen–Johansen estimator, and with one minus Kaplan–Meier treating churn as censoring. The gap between them measures how much the single-event shortcut distorts the target in this domain.
The second component is the model. A DropoutNet-style encoder takes item content features and whatever interaction history the user has, possibly none. On top of it sits a discrete-time multi-outcome hazard head and a satisfaction head. It is trained with the censored likelihood that DRSA uses, extended to several outcomes. Every impression a user actually reached counts in the at-risk group, including those of users who later churned, and items the user never reached count as unobserved rather than negative. The model is then recalibrated per outcome with the method of Alberge et al., the value estimates $V_r$ are fit from warm-up rates on training users, and items are ranked with $s(i)$.
The third component is policy evaluation. Per-step hazard and satisfaction predictions are scored on the randomly exposed impressions, which are unbiased with respect to the logging policy. Sequential consequences, meaning the actual warm-up probability under a policy, need a simulator. We use KuaiSim with its native patience-based exit module. Fitting a hazard-based exit module to held-out data and simulating with it would be partly circular, so we treat it only as a robustness check and state that limitation.
Baselines
The naive reference is a popularity recommender evaluated on the same held-out new users. Any method that fails to raise warm-up incidence above popularity has not solved cold start in the sense we care about.
The strongest existing cold-start competitors are DropoutNet and MAMO, trained on the same KuaiRand training users and features with an equal hyperparameter budget of 50 search trials each. MAMO is given demographic fields only where KuaiRand provides them. DropoutNet reported a 6% recall@100 gain over a content baseline on CiteULike, and MAMO reported a mean absolute error of 0.8894 versus MeLU's 0.9299 on MovieLens-1M for cold users and cold items. Neither result is on KuaiRand or on a retention metric, and we found no reported cold-start result on Lacuna for KuaiRand, so both must be re-run. BC-ICL is included on a reduced candidate set because its repeated forward passes do not scale to the full catalog. Its reward is immediate satisfaction, which makes it a strong immediate-reward explorer.
The strongest survival competitor is DRSA-style single-event modeling with churn as the event and everything else censored, using the same encoder and ranking rule. DRSA reports a C-index of 0.911 against DeepHit's 0.858 on BIDDING, though that result comes from a non-cold setting. The dwell-time competitor applies Xie et al.'s thresholds to define satisfaction and trains a two-tower multi-task model. The simulator anchor is KuaiSim's reported 0.7234 AUC for user behavior prediction on KuaiRand-Pure, which tells us how much to trust simulated policy comparisons.
Three ablations isolate the key ingredients. The first collapses exit types into one outcome. The second trains only on users who survived five or more interactions, which isolates the survivorship effect. The third removes recalibration.
Beating the field means a statistically significant gain in warm-up cumulative incidence at the fixed horizon over DropoutNet, MAMO, and the single-event model. It must hold under KuaiSim and under the robustness simulator, with calibration against Aalen–Johansen no worse than the uncalibrated model.
Why not just…?
Why not just train a multi-task network with a classification head per outcome? This is the most serious threat. A per-step softmax over continue, exit-and-return, and churn, trained on every reached impression, is mathematically the discrete-time competing-risk model. If that is what a reviewer means, we agree, and the contribution becomes the objective, the evaluation, and the censoring discipline rather than the architecture. The version that differs is the usual one, with user-level labels ("did this user churn?"). That version must drop or mislabel users still active at the log's end, and we include it as a baseline. If it matches our model on warm-up incidence, we will report that the gain comes entirely from the race objective and the all-new-users training set. The Xie et al. finding that a more complex multi-gate network brought no gain over simple towers suggests the head design matters less than the target.
Why not just use single-event survival with competing exits censored, as Iwata et al. did? The Aalen–Johansen versus Kaplan–Meier gap from the first component answers this directly. If the gap is negligible on KuaiRand because churn is rare in the horizon, the single-event model will tie ours and we will say so. If the gap is large, the single-event model ranks items with a distorted estimate of warm-up probability, and the ablation that collapses exit types should show the cost in simulated warm-up.
Why not just use dwell-time thresholds to define good interactions? Thresholds are cheap and fit to item-level dwell histories, which cold items lack. Xie et al.'s own deployment showed dwell time falling while clicks rose, so a threshold-defined label does not by itself track longer engagement. The dwell-time baseline tests whether a well-chosen satisfaction label, without any churn term, recovers most of the gain. If it does, the value lies in the churn penalty $\lambda$, which we will measure by sweeping $\lambda$ to zero.
Why not just run a cold-start bandit like BC-ICL with a better reward? Plugging our predicted warm-up probability into BC-ICL as its reward is a legitimate hybrid, and we run it. BC-ICL's resampling captures uncertainty that our myopic score ignores, so the hybrid may win. That would still support the claim, because the gain would come from the competing-risk target. It would also show that exploration and the race objective combine.
Decisive Checks
The measurement claim fails if method rankings on held-out KuaiRand new users are the same under survivor-only recall/NDCG and under all-user warm-up incidence. The method claim fails if the survivorship ablation and the user-level classification baseline match the full model in simulated warm-up incidence, since that would mean the competing-risk machinery adds nothing beyond using all the data. Calibration is checked per outcome against Aalen–Johansen curves at fixed horizons, and time-dependent C-index and integrated Brier score are reported for the hazard heads. The distillation variant is kept only if it improves integrated Brier score for users with fewer than five interactions without hurting calibration, as the bias-variance analysis of DiSKD predicts it may or may not.
References
- Risk and Prognosis of Secondary Bladder Cancer After Radiation Therapy for Rectal Cancer: A Large Population-Based Cohort Study
- Type 2 diabetes mellitus compromises the survival of diffuse large B-cell lymphoma patients treated with (R)-CHOP – the PLRG report
- Recommendation method for extending subscription periods
- Deep Recurrent Survival Analysis
- Reweighting Clicks with Dwell Time in Recommendation
- KuaiSim: A Comprehensive Simulator for Recommender Systems
- KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos
- DropoutNet: Addressing Cold Start in Recommender Systems
- MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation
- Bootstrap-Conditioned Action Selection with Tabular Foundation Models
- On the calibration of survival models with competing risks
- Discrete Survival Knowledge Distillation for Competing Risks Analysis