acceptodds
Under review as a conference paper at ICLR 2027

Towards Temporal Interest Modeling in Recommendation via Reinforcement Learning

Abstract

Reinforcement Learning (RL) methods are promising for recommendation systems due to their ability to optimize long-term user-system interactions. However, current RL recommendation methods treat user interactions as homogeneous decision rounds, ignoring irregular temporal gaps and the resulting interest decay over time. To incorporate temporal interest signals into RL, we study the relationship between temporal gaps and user interest through mutual information, finding power-law decay at the population-level and exponential decay at the individual level. Guided by the empirical observations, we propose a structured decision-process model: Temporal Interest Semi-MDP (TIS-MDP), which incorporates temporal gaps via a survival function to make RL methods temporal-aware. The effectiveness of TIS-MDP is verified across experiments.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Recommendation, Temporal Interest

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jK59OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an approach for incorporating temporal gaps (how to capture a users changing interest over time - TIS-MDP) in recommendation systems. If a user takes an action A and then comes back after a time, how much credit should flow back or be attributed to action A - this is the core contribution. While for a single user the interest decay is exponential, for a population it follows power law. By using mutual information to measure the correlation between historic item categories, and future clicks (CTTC), the authors propose a survival function based on the powerlaw to incorporate the temporal gap into Bellman. They prove the contraction properties and measure the effectiveness of TIS-MDP.

Critical strengths

The temporal gaps in user interactions is a real and under studied problem. How does this affect the effectiveness of of RL? Where does credit assignment fail? The paper motivates this, proposes a rigorous approach for addressing this. The paper grounds the contributions in established approaches (semi-MDP and Bellman, Law of large numbers). That makes it readable and relatable to researches. While the gamma distributed heterogeneity has some limitation in my POV, the observation of law of exponents for individual decay vs power law for population decay is interesting.

Critical weaknesses

One weakness of this paper are the assumptions made in the soundness of the approach, which are even partly addressed in the appendix (e.g. the population-level approximation). It would be even more valuable to perform ablation tests here. Also, it would be very good to show some tradeoffs when using the TIS-MDP in the experiments section, for example, why the improvements are different for the RL backbones (BCQ improves the length, but not the reward; CQL improves the reward, but not the length). Do they even need more data to estimate s(Δ) accurately? Simply mentioning this would already be sufficient.

Questions for authors

Ablation studies on interest continuation using only temporal gap for ablation. How sensitive is CTTC to biases outlined earlier? Since most of section 5 is standard RL material, more experimental work on these gaps would be very interesting to read about.

2 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We thank the reviewer for the constructive comments and for recognizing the motivation of our work. We respond to your questions as follows.

Assumptions. The individual exponential decay and the population power law are not assumed a priori. They form a conjecture drawn from the data, which Appendix B.3 checks in three steps. The exponential fits better for 41 of the top-50 users, the fitted rates are compatible with a Gamma distribution under a K–S test, and the implied Lomax curve matches a direct power-law fit (R2=0.700R^2=0.700 vs. 0.703). We use the population-level approximation because per-user curves are sparse and noisy (Table 4). Theorems 5.3 and C.1 quantify the resulting aggregation and misspecification errors, and the gap-only continuation model is further tested by ablations on the survival form (Table II) and on the decay parameters α\alpha and τ\tau (Tables 9 and 10).

Varying improvements across backbones, and trade-offs. RL optimizes the cumulative reward, and TIS-MDP improves it for all three backbones on KuaiRand (+7.8%, +15.3% and +0.8% over 5 seeds, Table II). Length and per-step reward are the two factors of this quantity (Rcumu≈Len×RavgR_{\mathrm{cumu}}\approx\mathrm{Len}\times R_{\mathrm{avg}}) and are not optimized separately. How the gain splits between them depends on how each backbone constrains its policy (Section 6.2), and this split is stable across seeds. BCQ (TIS) gains mainly in length (5.140 vs. 4.800) and CQL (TIS) mainly in per-step reward (0.320 vs. 0.270), and the exponential form shows the same pattern. If one component matters more for an application, it can be targeted through the reward definition.

The main trade-off of TIS-MDP lies in the strength of the decay. An overly aggressive decay also removes genuine long-term credit, and there is no gain at α=5\alpha=5 (Table 9). In addition, the method brings little benefit when interactions are dense and the gaps carry little temporal information, as on MovieLens (Table 6).

Does TIS-MDP require more data to estimate s(Δ)s(\Delta)? The answer is No. Its two parameters are not estimated from the RL training data. The exponent α=0.185\alpha=0.185 is obtained once on Amazon and kept fixed, and τ\tau is selected from a small set. The value fitted on KuaiRand itself (about 0.25) is not the best-performing one, and the gains hold for all α∈[0.1,1.0]\alpha\in[0.1,1.0] (Table 9). What matters is therefore the decay shape rather than a precisely calibrated rate.

Sensitivity of CTTC to data volume and outlying points. The conclusion does not rest on the sparse and noisier long-gap points. When the fit is restricted to the first five days, where observations are densest, the power law is still closer to linear on the log-log scale than the exponential on the semi-log scale (R2R^2 of 0.973 vs. 0.863, Appendix B.2). The same holds over the full 30 days (0.936 vs. 0.725, Fig. 2), on KuaiRand-1K (0.977 vs. 0.881, Fig. 9) and on Foursquare (0.854 vs. 0.806, Fig. 10).

Sensitivity of CTTC to the biases in Appendix F. The three datasets are affected by different biases. Amazon uses sampled negatives, KuaiRand items are surfaced by a feed recommender, and Foursquare consists of user-initiated check-ins, yet all three show the same decay shape. Moreover, only this shape is used downstream, and the RL results are robust to its exact calibration (Table 9).

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

More experimental analysis. We agree that Section 5.1 is largely standard theory, and it serves only to confirm that the backup is well-posed. Therefore, we add a standard MDP baseline with the gap as input, and a per-backbone comparison with Semi-MDP discounting. These results are summarized in Tables I–III below.

Table I. Synthetic environment: cumulative reward (mean ± std over 5 seeds)

BackboneStandard MDPStandard MDP + gap inputTIS-MDP (exponential)TIS-MDP (power-law)
DQN47.29±0.2349.68±0.2352.67±1.9953.27±2.55
A2C52.81±0.1453.99±3.3856.07±1.0055.47±0.53
PPO51.86±1.5253.12±1.7255.15±0.5955.91±0.67

Table II. KuaiRand (EasyRL4Rec): mean ± std over 5 seeds

BackboneMethodRcumuR_{\mathrm{cumu}}LenRavgR_{\mathrm{avg}}
BCQStandard MDP1.930±0.104.800±0.250.400±0.001
Standard MDP + gap input1.990±0.064.936±0.190.404±0.005
TIS-MDP (exponential)2.065±0.025.106±0.040.409±0.002
TIS-MDP (power-law)2.080±0.065.140±0.270.410±0.009
CQLStandard MDP0.850±0.013.130±0.010.270±0.004
Standard MDP + gap input0.902±0.013.114±0.010.290±0.005
TIS-MDP (exponential)0.993±0.043.086±0.020.325±0.012
TIS-MDP (power-law)0.980±0.053.090±0.000.320±0.016
CRRStandard MDP1.260±0.015.830±0.030.220±0.001
Standard MDP + gap input1.264±0.015.858±0.020.220±0.000
TIS-MDP (exponential)1.269±0.015.893±0.030.220±0.001
TIS-MDP (power-law)1.270±0.025.900±0.060.220±0.000

Table III. KuaiSim cross-session retention: mean ± std over 5 seeds

MethodReturn day ↓Retention ↑
TD32.2842±0.13660.8264±0.0212
RLUR2.2415±0.52170.8032±0.0781
TD3 + gap input2.2287±0.04190.8467±0.0067
TD3 (TIS)2.1454±0.10530.8772±0.0156

We thank reviewer for the time to review our paper and hope that our responses address the reviewer's concerns so reviewer can re-evaluate our work. Please let us know if you have any further questions, and we will do our best to address them.

Reviewer_Zf2fOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates user interest decay under irregular temporal gaps in recommendation systems and devises a structured decision-process model. Using the CTTC (Category-wise Target-aware Temporal-gap Correlation) mutual information metric, the analysis indicates that user interest follows a power-law decay at the population level and an exponential decay at the individual level, and Gamma-distributed heterogeneity accounts for this macro–micro discrepancy. This paper proposes TIS-MDP (Temporal Interest Semi-MDP) framework, which constructs a temporally-aware effective discount through a survival function and features plug-and-play compatibility with various RL algorithms. Meanwhile, the paper proves the contraction property and fixed-point uniqueness of the Bellman operator, and derives an error bound as well as a Law of Large Numbers justification for the population-level model. The method is evaluated in three settings: synthetic experiments, offline RL-based recommendation, and cross-session user retention optimization, evaluated with multiple RL backbone algorithms.

Critical strengths
  1. Unlike other RL recommendation works that rely on heuristically prescribed temporal decay mechanisms, this paper quantitatively reveals the power-law interest decay at the population level and exponential decay at the individual level through mutual information analysis. It explains the underlying mechanism via Gamma-distributed heterogeneity, delivering reusable empirical conclusions and theoretical interpretations that form a solid theoretical foundation.
  2. TIS-MDP adopts a plug-in design that is well compatible with a wide range of RL algorithms. It features low engineering deployment cost and strong practical application value.
Critical weaknesses
  1. The evaluation compares only with temporally agnostic RL baselines and does not include alternative temporal-discounting mechanisms. Consequently, it cannot isolate whether the observed gains arise from the TIS-MDP design or from the inclusion of temporal information per se, which limits the evidence for the method’s specific contribution.
  2. A second concern is that the paper treats CTTC decay as a measure of interest decay and interprets the survival function as a continuation probability, without operational validation of either mapping. The proposed interest chain is a latent, unlabeled variable whose interpretation is not calibrated against observed user behavior. The argument therefore remains largely heuristic, and the paper does not provide empirical evidence that would support the proposed causal interpretation.
Questions for authors
  1. The evaluation includes only temporally agnostic RL baselines. It omits both the common baseline that supplies temporal gaps as state features and alternative temporal-discounting algorithms. The authors should therefore isolate the effect of the TIS-MDP design from that of temporal information, for example by adding the state-feature baseline and alternative discounting mechanisms.
  2. The paper treats CTTC decay as a measure of interest decay and the survival function as a continuation probability, but neither mapping is validated causally, and the latent interest chain is not calibrated against observed user behavior. The authors should also address potential confounders such as exposure policies and demand saturation and provide evidence for the validity of these assumptions.
1 comment

We thank the reviewer for the careful reading and the constructive comments. We respond to your questions as follows.

Isolating the TIS-MDP design from temporal information. We add the two comparisons requested, and the results are reported in Tables I–III.

(a) Time as a feature. We add a standard-MDP baseline that takes the temporal gap as an input and keeps the constant-discount backup. Since Δt\Delta_t is realized only after ata_t is taken, the feature is the elapsed time since the previous interaction, Δt−1\Delta_{t-1}. This feature improves the standard MDP, but it recovers less than half of the gain of TIS-MDP in the experiment, and TIS-MDP stays ahead of it for all six backbones and on KuaiSim. We think the reason is that, although a gap feature enriches state, the backup still treats a ten-second and a ten-day transition identically. TIS-MDP instead changes how credit is propagated.

(b) Alternative temporal discounting. The alternatives are different shapes of a gap-dependent discount γeff(Δ)\gamma_{\mathrm{eff}}(\Delta): constant (standard MDP), two fixed levels by session boundary (RLUR), exponential (standard Semi-MDP) and power-law (ours).

Table I. Synthetic environment: cumulative reward (mean ± std over 5 seeds)

BackboneStandard MDPStandard MDP + gap inputTIS-MDP (exponential)TIS-MDP (power-law)
DQN47.29±0.2349.68±0.2352.67±1.9953.27±2.55
A2C52.81±0.1453.99±3.3856.07±1.0055.47±0.53
PPO51.86±1.5253.12±1.7255.15±0.5955.91±0.67

Table II. KuaiRand (EasyRL4Rec): mean ± std over 5 seeds

BackboneMethodRcumuR_{\mathrm{cumu}}LenRavgR_{\mathrm{avg}}
BCQStandard MDP1.930±0.104.800±0.250.400±0.001
Standard MDP + gap input1.990±0.064.936±0.190.404±0.005
TIS-MDP (exponential)2.065±0.025.106±0.040.409±0.002
TIS-MDP (power-law)2.080±0.065.140±0.270.410±0.009
CQLStandard MDP0.850±0.013.130±0.010.270±0.004
Standard MDP + gap input0.902±0.013.114±0.010.290±0.005
TIS-MDP (exponential)0.993±0.043.086±0.020.325±0.012
TIS-MDP (power-law)0.980±0.053.090±0.000.320±0.016
CRRStandard MDP1.260±0.015.830±0.030.220±0.001
Standard MDP + gap input1.264±0.015.858±0.020.220±0.000
TIS-MDP (exponential)1.269±0.015.893±0.030.220±0.001
TIS-MDP (power-law)1.270±0.025.900±0.060.220±0.000

Table III. KuaiSim cross-session retention: mean ± std over 5 seeds

MethodReturn day ↓Retention ↑
TD32.2842±0.13660.8264±0.0212
RLUR2.2415±0.52170.8032±0.0781
TD3 + gap input2.2287±0.04190.8467±0.0067
TD3 (TIS)2.1454±0.10530.8772±0.0156

The gap-input baseline receives the same temporal information as TIS-MDP, so the remaining difference between them comes from how the gap enters the Bellman backup rather than from the temporal information itself. Among the discounting mechanisms, the exponential and power-law forms perform similarly and both clearly above the standard MDP, while the session-level discount of RLUR stays below TD3 (TIS) on KuaiSim.

CTTC, interest decay and the continuation probability. CTTC is a statistic of how much a user's behavior Δ\Delta days ago still tells us about whether the user will respond to the current item. We do not use its exact values and do not interpret it causally. We only use how fast it decreases as Δ\Delta grows, and this speed of decay tells us how much credit an earlier recommendation should receive. The survival function s(Δ)s(\Delta) turns this decay into a weight in the TD target. It can be read as the chance that the user is still following the same interest after a gap of Δ\Delta (Eq. (8)). We never observe the exact moment when a user's interest changes, just as a platform never observes the exact moment when a user stops using it. In both cases, the longer the user has been away, the less likely the user is still in the same state. So s(Δ)s(\Delta) smoothly lowers the credit passed across a long gap instead of cutting it off at a fixed point.

For the confounders, the same decay shape appears in three very different datasets, namely Amazon purchases, KuaiRand short videos and Foursquare check-ins. Their items reach users in different ways, since KuaiRand videos are pushed by a recommender while Foursquare are made by users themselves. Their demand also behaves differently, since a user who has bought an item rarely buys it again soon, while watching videos can repeat freely. The shared shape is therefore unlikely to come from how items are shown or from demand being used up. Moreover, the decay parameter estimated once on Amazon (α=0.185\alpha=0.185) is used unchanged on KuaiRand and KuaiSim, and the gains hold for all α∈[0.1,1.0]\alpha\in[0.1,1.0] (Table 9).

We thank reviewer for the time to review our paper, and hope that our responses address the reviewer's concerns so the reviewer can re-evaluate our work. Please let us know if you have any further questions, and we will do our best to address them.

Reviewer_RPWiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces Temporal Interest Semi-MDP (TIS-MDP), a framework designed to fix a fundamental credit-assignment flaw in reinforcement learning (RL) recommendation systems: standard RL treats every interaction step equally, incorrectly crediting past recommendations for unrelated actions taken weeks later.

Critical strengths

Rather than choosing a power-law discount arbitrarily, the authors provide a closed-form probabilistic derivation showing how individual exponential forgetting integrates into a population-level Lomax/power-law distribution via the Laplace transform.

Traditional RL for recommendations either treats interactions as immediate and homogeneous (standard infinite-horizon MDPs) or forces hard terminal cutoffs at session boundaries (episodic MDPs). By introducing a temporal-aware survival discount factor this approach enables a continuous, soft transition: it allows long-term credit to propagate smoothly across brief breaks while truncating credit propagation when the time gap indicates the user's interest chain has likely shifted.

The method operates as a modular, drop-in replacement for the standard temporal-difference. It requires no specialized network architectures or modifications to action spaces, allowing it to seamlessly augment virtually any standard RL backbone

The survival function relies solely on elapsed real time Δ\Delta across the entire population, ignoring the user's specific context, history, and the item's identity. In practice, return intervals and session breaks are heavily confounded by external factors—such as diurnal rhythms, workday routines, ...

Critical weaknesses

The synthetic environment was explicitly hard-coded with the paper's exact survival formula Evaluating an algorithm in a toy world whose ground-truth transition dynamics were specifically programmed to follow the exact mathematical formula of the proposed algorithm only proves the code doesn't crash; it provides zero evidence of real-world applicability or robustness.

To justify adding a power-law discounted Semi-MDP framework, the paper should compare against common-sense heuristics for example, Episode Resets or time-as feature like Passing the elapsed time Δ\Delta as an input feature into the state/critic network, which lets standard RL learn temporal dependencies without modifying the Bellman operator.. a systematic benchmark against standard Semi-MDP formulations across the main tasks is missing.

Evaluation is shallow- Tables 1, 2, 3, 6, and 7 present single-point estimates without standard deviations or multi-seed confidence intervals. the overlap between the exponential and power-law results is large enough that the difference is not statistically significant

Questions for authors

how doe sthe Error Bars / Statistical Significance features in main tables?

The core benefit of the method hinges entirely on having informative, multi-day temporal gaps between interactions. On platforms or datasets characterized by dense, within-session activity—such as MovieLens, where roughly 98% of interactions happen on the same day—the survival factor collapses to a constant, In these regimes, the formulation degenerates back into standard, uncalibrated MDP discounting, yielding essentially zero empirical benefit- how do we verify this is not the case?

using a temporal-gap-aware Bellman backup requires recommender architectures to explicitly record, store, synchronize, and update precise inter-session timestamps across asynchronous user journeys. For massive, low-latency recommendation platforms, building and maintaining this tracking infrastructure across distributed data pipelines adds notable engineering complexity for modest performance gains.- please comment

2 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We thank the reviewer for the constructive comments and for recognizing the contribution of our work. We respond to your questions as follows.

Synthetic experiment. This is a controlled experiment: the interest-break mechanism is known by construction, so the effect of the temporal-aware backup can be isolated, which is impossible on real logs where breaks are unobserved. The outcome is not built in: agents never observe Δt\Delta_t when acting (the gap enters only the TD target), and performance is the cumulative reward collected in the environment, not the TIS objective. Evidence of real-world applicability comes from KuaiRand and KuaiSim, that the two large scale datasets are collected from the real-world recommendation system in industry.

Comparison with temporal baselines. We cover the three types below; see Tables I–III.

(a) Time as a feature. We add a standard-MDP baseline that takes the temporal gap as an input and keeps the constant-discount backup. Since Δt\Delta_t is realized only after ata_t is taken, the feature is the elapsed time since the previous interaction, Δt−1\Delta_{t-1}. This feature improves the standard MDP, but it recovers less than half of the gain of TIS-MDP in the experiment, and TIS-MDP stays ahead of it for all six backbones and on KuaiSim. We think the reason is that, although a gap feature enriches state, the backup still treats a ten-second and a ten-day transition identically. TIS-MDP instead changes how credit is propagated.

(b) Standard Semi-MDP. With s(Δ)=e−λΔs(\Delta)=e^{-\lambda\Delta}, the effective discount γbase e−λΔ\gamma_{\mathrm{base}}\,e^{-\lambda\Delta} is the Semi-MDP discount of Eq. (2) up to a constant factor. It is compared in Table 8 (Appendix D.3) and is now reported per backbone with error bars.

(c) Episode resets. A hard reset is the step survival function s(Δ)=I[Δ≤Δ0]s(\Delta)=\mathbb{I}[\Delta\le\Delta_0], which is the hard-termination limit of TIS-MDP. RLUR is the baseline of this session-boundary type in our experiments, as it applies one fixed discount to every cross-session transition. Over 5 seeds, TD3 (TIS) improves over RLUR in both return day (2.1454 vs. 2.2415) and retention (0.8772 vs. 0.8032), with a much smaller spread across seeds (Table III).

Table I. Synthetic environment: cumulative reward (mean ± std over 5 seeds)

BackboneStandard MDPStandard MDP + gap inputTIS-MDP (exponential)TIS-MDP (power-law)
DQN47.29±0.2349.68±0.2352.67±1.9953.27±2.55
A2C52.81±0.1453.99±3.3856.07±1.0055.47±0.53
PPO51.86±1.5253.12±1.7255.15±0.5955.91±0.67

Table II. KuaiRand (EasyRL4Rec): mean ± std over 5 seeds

BackboneMethodRcumuR_{\mathrm{cumu}}LenRavgR_{\mathrm{avg}}
BCQStandard MDP1.930±0.104.800±0.250.400±0.001
Standard MDP + gap input1.990±0.064.936±0.190.404±0.005
TIS-MDP (exponential)2.065±0.025.106±0.040.409±0.002
TIS-MDP (power-law)2.080±0.065.140±0.270.410±0.009
CQLStandard MDP0.850±0.013.130±0.010.270±0.004
Standard MDP + gap input0.902±0.013.114±0.010.290±0.005
TIS-MDP (exponential)0.993±0.043.086±0.020.325±0.012
TIS-MDP (power-law)0.980±0.053.090±0.000.320±0.016
CRRStandard MDP1.260±0.015.830±0.030.220±0.001
Standard MDP + gap input1.264±0.015.858±0.020.220±0.000
TIS-MDP (exponential)1.269±0.015.893±0.030.220±0.001
TIS-MDP (power-law)1.270±0.025.900±0.060.220±0.000

Table III. KuaiSim cross-session retention: mean ± std over 5 seeds

MethodReturn day ↓Retention ↑
TD32.2842±0.13660.8264±0.0212
RLUR2.2415±0.52170.8032±0.0781
TD3 + gap input2.2287±0.04190.8467±0.0067
TD3 (TIS)2.1454±0.10530.8772±0.0156
AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Error bars and statistical significance. We reran the synthetic, KuaiRand and KuaiSim experiments with 5 seeds and report mean ± std in Tables I–III. TIS-MDP improves significantly over the standard MDP for all three backbones in the experiments. We also agree with the reviewer on the two survival forms. They are not significantly different in any setting, so we do not claim a downstream advantage of the power law over the exponential. We keep the power law as the default because it is the form supported by the population-level data in Section 4.

Dense-interaction regimes. This behavior is by design. As Δ→0\Delta\to0, s(Δ)→1s(\Delta)\to1 and the TIS backup reduces exactly to the standard MDP backup, so the method does not harm the backbone when the data carry little temporal information. MovieLens illustrates this case. About 98% of its interactions happen on the same day, so s(Δ)s(\Delta) is close to one for almost all transitions and the gains in Table 6 are small. Whether a platform is in this regime can therefore be checked before training from the distribution of s(Δ)s(\Delta) over the logged transitions. Recommendation and advertising platforms where users return after days are the regime that TIS-MDP targets, and there the gains are clear (Tables II and III).

Engineering cost of timestamps. Real-world recommender systems usually log a timestamp with every interaction, so the extra cost is small. TIS-MDP only needs the gap between two consecutive interactions, which is one scalar per transition and is used only in the TD target at training time.

We thank reviewer for the time to review our paper. Please let us know if you have any further questions, and we will do our best to address them.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your reviews. As the discussion period is ending, we summarize how we addressed the main concerns in the responses:

  • Isolating the TIS-MDP design from temporal information (RPWi, Zf2f, jK59). We added a “standard MDP + gap input” baseline and per-backbone comparisons with Semi-MDP (exponential) discounting and RLUR. The gap-input baseline recovers less than half of TIS-MDP’s gain, and TIS-MDP stays ahead of it on all six backbones.
  • Statistical reliability (RPWi). We reran the synthetic, KuaiRand and KuaiSim experiments with 5 seeds and report mean ± std in Tables I–III.
  • Validity of the CTTC and survival-function assumptions (Zf2f, jK59). The same decay shape appears on three datasets with very different exposure mechanisms and demand patterns. The decay parameter fitted once on Amazon (α = 0.185) is used unchanged downstream, and the gains hold for all α ∈ [0.1, 1.0] (Table 9).
  • Limitations (RPWi, jK59). We clarified the dense-interaction regime (e.g., MovieLens), the trade-off of overly aggressive decay, and why gains split differently across backbones.

We believe our responses and additional experiments address the reviewers' concerns. As the discussion period is coming to an end, we would be glad to continue the discussion if any questions remain.

Best regards,

Authors of submission 39371

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.