{"id":"705647b5-f0f3-4d90-8443-8c2b7af09c5b","arxiv_id":"2505.22153","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PTPM trims a global decision tree into personalized per-sample trees with reinforcement learning and corrects conditional-model bias with inverse propensity scoring, improving watch time prediction.","lead":"Kuaishou researchers built PTPM, a model that predicts watch time in short video feeds by learning a custom decision tree for every user-video pair instead of using one fixed tree. In offline tests and a live A/B experiment it beat the previous TPM model, and it is now deployed for over 400 million daily users.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 17's UCL loss is algebraically inconsistent with the negative log-likelihood it claims, so the unbiased-conditional-learning derivation in Section 4.2 does not support the reported gains.","rationale":"The reader's weakest-assumption pick was the pruning-inheritance assumption in Section 4.1. That is a genuine concern, but it is at least argued for in the text and is indirectly supported by the Table 3 ablation, where adding TSL improves over PTPM without TSL; if inheritance were badly broken, adding TSL would likely hurt. The Eq. 17 problem is more decisive: it is a direct algebraic contradiction in the paper's formal derivation, not a matter of missing evidence or questionable consensus. The displayed equality is false for any path of depth at least two. The unbiasedness proof that follows establishes a different estimator (per-node IPS) than the loss actually written, so the UCL claim is unsupported as stated. Since UCL is one of the two main contributions behind the reported improvements, this is load-bearing for the paper's central claim. I do not recommend REJECT because the paper provides consistent offline and online evidence that some version of the method works, and the loss could be corrected. The reader's CONDITIONAL verdict remains appropriate, with the condition being a corrected derivation and an ablation using the actually derived IPS loss. I partially agree with the reader: they flagged the oracle-versus-estimated propensity issue in the IPS proof, but not this algebraic mismatch, which is the more load-bearing concern.","tokens_in":12551,"tokens_out":12890,"duration_ms":150785,"concrete_test":"Symbolically expand Eq. 17 for a depth-3 path and verify it equals log p_3 - log p_1 rather than log(p_1 p_2 p_3). Then, on KuaiRec, replace Eq. 17 with the per-node IPS-weighted cross-entropy loss L = -sum_j [ I(y_i in parent_j) / P_{j-1} ] log p_j (with P_0 = 1) and rerun the Table 3 ablation. If the UCL row's MAE/XAUC advantage over TPM changes materially or reverses, the published loss, not unbiased IPS, is driving the reported gains; if it matches, the authors should still correct the derivation before claiming unbiasedness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.2, Eq. 17 defines L_ce-UCL first as -log p(y_i in l_k | x_i, T), then as -sum_{j=1..d} log[ p(y in n_j | y in n_{j-1}, x, T) / p(y in n_{j-1} | x) ]. These two expressions are not equal. Writing p_j for the j-th conditional edge probability and P_{j-1} = product_{l<j} p_l for the propensity, the second expression equals sum_j (log p_j - log P_{j-1}). This telescopes to log p_d - sum_{l=1}^{d-2} (d-l-1) log p_l; for d=3 it is log p_3 - log p_1, not log(p_1 p_2 p_3) = log p(leaf | x, T). The unbiasedness proof in Eq. 16 applies to a per-node IPS loss delta(o_i, o_hat_i) / p(parent | x_i), but Eq. 17 is not that loss. Therefore the paper's central claim that UCL 'enables unbiased conditional learning' is not established by the derivation, and the UCL ablation gains may be an artifact of an unstated reweighting. This is load-bearing because UCL is one of the two core modules (with TSL) responsible for the offline and online improvements.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PTPM, an extension of the tree-based progressive regression model (TPM) for watch-time prediction in short-video recommendation. PTPM has two main additions: a Tree Structure Learning (TSL) module that learns a personalized, sample-dependent tree by pruning a global full binary tree via a REINFORCE-style self-critical objective, and an Unbiased Conditional Learning (UCL) module that claims to remove the sample-selection bias induced by training each node classifier only on samples that pass through its parent. Offline experiments on KuaiRec, CIKM16, and WeChat21 report improved MAE and XAUC over TPM and other baselines, ablations attribute gains to both modules, and a live A/B test on Kuaishou reports a +0.428% watch-time lift with negligible latency overhead. The paper concludes that PTPM has been fully deployed and serves over 400 million daily users.","tokens_in":12847,"tokens_out":5118,"duration_ms":56713,"significance":"If the technical claims hold, PTPM is a practically relevant contribution: it is, to my knowledge, the first end-to-end learned discretization for watch-time prediction, and the pruning-based joint training of tree structure and classifiers is an elegant way to avoid bilevel optimization. The paper also provides useful evidence beyond offline benchmarks: ablations, a live A/B test, and deployment-scale performance numbers. The strength of the empirical demonstration, however, depends critically on the correctness of the UCL derivation and on the inherited-parameter assumption in TSL. The current manuscript has a load-bearing algebraic inconsistency in Eq. (17) and an unverified assumption in Eq. (6); until these are resolved, the central claims of unbiased conditional learning and of parameter-free structure inheritance are not established.","major_comments":[{"comment":"The definition of L_ce-UCL is algebraically inconsistent. The first expression is -log p(y_i in l_k | x_i, T), which equals -sum_{j=1..d} log p_j, where p_j = p(y_i in n_{phi(j)} | y_i in n_{phi(j-1)}, x_i, T). The second expression is -sum_j log[ p_j / p(y_i in n_{phi(j-1)} | x_i) ]. Using the paper's own Eq. (14), p(y_i in n_{phi(j-1)} | x_i) = product_{l<=j-2} p_l, so the sum telescopes to -log p_d + log p_1 (for d=3 it is -log p_3 + log p_1), not -log(p_1 p_2 p_3). Thus Eq. (17) does not define the negative log-likelihood of the leaf, and the unbiasedness proof in Eq. (16) applies to a per-node IPS loss while Eq. (17) is a different quantity. This is load-bearing because UCL is one of the two core modules; the UCL ablation gains in Table 3 may be due to an unstated reweighting rather than to debiasing. The authors need to provide a correct derivation of the loss used in Algorithm 1, or remove the unbiasedness claim and reframe the module as a heuristic reweighting.","section":"Sec. 4.2, Eq. (17)"},{"comment":"The key assumption that inherited parameters theta*(T) = theta*(T_g) * (1 - a_T) 'remain well-adapted, and no additional fine-tuning is required' is asserted without direct evidence. This is load-bearing for the TSL module: if pruned-subtree classifiers are miscalibrated or poorly adapted, the personalized trees would produce degraded predictions, and the efficiency claim (no inner optimization, fast convergence in Fig. 6) would be weakened. The paper should isolate this assumption, e.g., by comparing (i) pruning with inherited parameters, (ii) pruning followed by fine-tuning on the pruned subtree, and (iii) training the pruned structure from scratch, on at least one offline dataset.","section":"Sec. 4.1, Eq. (6) and following paragraph"},{"comment":"The unbiasedness proof in Eq. (16) assumes that the propensity p(y_i in n_{phi(j-1)} | x_i) is the true conditional distribution, but the implementation uses the model's own predicted probabilities from Eq. (14), which are products of learned classifier outputs. IPS is unbiased only when propensities are known or consistently estimated under suitable overlap; using self-estimated propensities as if they were fixed oracle values is not justified, and the fact that these propensities are functions of the same parameters theta being optimized makes the 'unbiased' claim circular. The authors should state explicitly what conditions are needed for Eq. (16) and provide either a corrected analysis or empirical validation (e.g., with known synthetic propensities) that the claimed bias reduction actually occurs.","section":"Sec. 4.2, Eq. (14)-(16)"},{"comment":"The paper reports that UCL improves AUC and prior-to-posterior calibration, but given the Eq. (17) issue, these results are not sufficient to establish that the improvement comes from unbiased conditional learning. The experiments should be repeated with the corrected loss (once derived) and, ideally, with a baseline that uses the same per-node weights but without the log-probability structure, to separate the effect of the reweighting from the effect of the claimed unbiased estimator.","section":"Sec. 5.2.2, Table 3 and Fig. 7"}],"minor_comments":[{"comment":"There is a typo in 'REINFORECE'; it should be 'REINFORCE'.","section":"Sec. 4.1.2"},{"comment":"Line 5 says 'Compute L_ce-UCL, the log-likelihood of (x_i,y_i) belonging to path phi, and E(y_i|x_i,T_g) and Var(y_i|x_i,T_g), as Eq. 17, 4.' This is unclear: Eq. 4 defines the TPM objective, not the UCL loss; please separate the references to Eq. 17 and the components of Eq. 4.","section":"Sec. 4.3, Algorithm 1"},{"comment":"The notation 'I[y_i in n_{phi(j)}]' is used both as a conditioning event and as an indicator in Eq. (16); please define the subscript on the expectation (over what distribution) and clarify that the propensity is evaluated at x_i.","section":"Sec. 4.2, Eq. (13)"},{"comment":"The table caption states 'Statistically significant results (p-value < 0.05) are highlighted in bold', but the body of the table does not show confidence intervals or p-values; please report the actual p-values or confidence intervals for the watch-time metric and the constraints.","section":"Sec. 5.3, Table 4"},{"comment":"The abbreviation 'SSB' is used in the sentence 'the debiasing of SSB improves the generalization of classifiers' without being defined; it should be spelled out (likely 'sample selection bias').","section":"Sec. 5.2.2"},{"comment":"The left panel labels both bars as 'TPM' which appears to be a typo; one bar should presumably be labeled 'PTPM' to match the text about parameter count and flops.","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the genuinely new thing: PTPM is the first watch-time model I know that learns the discretization tree end-to-end via pruning, instead of fixing a full tree (TPM) or grid-searching binning heuristics (CREAD). The pruning formulation is practical, and the paper does the work: three public datasets, ablations, parameter sensitivity, an efficiency plot, and a live A/B test with a significant +0.428% watch-time lift at Kuaishou scale. That is a credible industrial contribution, and the claim of serving 400M users is plausible given the authors' affiliation and the A/B setup.\n\nThe soft spots are real, and one is serious. Section 4.2's derivation of the 'unbiased conditional learning' loss, Eq. 17, is algebraically wrong. The first equality claims L = -log p(leaf | x, T), but the second expression expands it as a sum over j of log p_j - log P_{j-1}, where p_j is the conditional edge probability and P_{j-1} is the product of earlier p's. Those are not equal. For depth 3 you get log p_3 - log p_1, not log(p_1 p_2 p_3). The IPS unbiasedness proof in Eq. 16 applies to a per-node loss delta / p(parent), not to this pathwise expression. So the claim that UCL 'enables unbiased conditional learning' is not established by the math in the paper. The ablation still shows gains, but the mechanism is unclear; it could be an unstated reweighting that happens to help, not IPS debiasing. This is load-bearing because UCL is one of two core modules.\n\nTwo more moderate issues. The inherited-parameter assumption in Section 4.1 — that classifiers trained on the global tree remain well-adapted after pruning without fine-tuning — is asserted but never tested in isolation. I'd like to see an experiment comparing pruned-inherited versus pruned-fine-tuned. Also, the IPS proof assumes oracle propensities, while the implementation uses the model's own estimates; that is a known gap, but it should be acknowledged.\n\nThe paper is worth sending to review. The tree-learning idea is novel, the empirical evidence is strong and consistent, and the flaws are fixable in revision — but the UCL derivation must be corrected or the claim narrowed.","headline":"Genuine extension of TPM with strong empirical results, but the UCL unbiasedness derivation has a real algebraic error that needs to be addressed before the debiasing claim can be trusted.","tokens_in":13386,"tokens_out":3372,"would_cite":true,"duration_ms":36163,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Pruning a global decision tree per user-item pair yields better watch-time prediction and a 0.428% online watch-time gain.","keywords":["watch time prediction","tree-based progressive regression","personalized tree structure","ordinal regression","sample selection bias","inverse propensity scoring","recommender systems","short video recommendation"],"falsifier":"Train two versions of PTPM on the same data: the paper's inherited-parameter pruning versus re-training the classifiers of every sampled pruned subtree from scratch for the same number of steps. If the retrained subtree beats the inherited-parameter version by a margin comparable to PTPM's stated gains over TPM, then the no-fine-tuning assumption is the true source of the improvement. A second, cheaper check is to measure per-leaf calibration on out-of-parent-interval samples after pruning, comparing inherited versus retrained classifiers.","tokens_in":12355,"feed_emoji":"🌳","tokens_out":6364,"duration_ms":59979,"temperature":0.7,"pith_summary":"The paper proposes PTPM, an extension of the Tree-based Progressive Regression Model (TPM) for predicting how long a user will watch a recommended short video. TPM decomposes watch-time prediction into a fixed full binary tree of ordinal classification tasks, but that fixed equal-width discretization does not match real watch-time distributions, which vary across users and items. PTPM learns a personalized tree for each user-item pair by pruning a global tree, and adds an unbiased conditional learning module that corrects a sample-selection bias introduced by the tree's conditional classifiers. Offline, it improves MAE and XAUC over TPM and strong baselines on three datasets; online, it lifts watch time by 0.428% with negligible latency cost in a production A/B test. If these results hold, watch-time prediction can be made sample-adaptive without sacrificing industrial latency constraints.","feed_headline":"Per-request tree pruning lifts watch-time prediction 0.428%","feed_subtitle":"A pruning-based tree with debiased classifiers beats fixed-tree TPM offline and is already serving production traffic.","key_machinery":"The central object is a global full binary tree $T_g$ over the normalized watch-time interval $[0,1]$, together with a pruning model that maps user-item features to Bernoulli probabilities over non-leaf nodes. Sampling a pruning vector $a_T$ produces a sample-specific subtree $T = T_g \\circledast a_T$, and the inherited classifier parameters $\\theta^*(T)=\\theta^*(T_g)\\odot(1-a_T)$ are reused without fine-tuning because every subtree shares the same path semantics. The pruning policy is trained with a self-critical REINFORCE objective contrasting the pruned tree's reward $R_T-R_{T_g}$, where $R=\\mathrm{XAUC}-\\mathrm{MSE}$. For unbiased learning, the paper derives an inverse-propensity-scoring estimator: each conditional log-likelihood term along the path is weighted by the inverse propensity $1/p(\\hat{y}_i \\in n_{\\phi_{l_k}}(j-1)|x_i)$, with the propensity computed as the product of ancestor classifier outputs. These mechanisms make the tree structure personalized and end-to-end trainable at latency comparable to TPM.","core_discovery":"On its own terms, the paper claims that the tree structure in ordinal watch-time regression is a first-class learnable object, not a fixed hyperparameter. PTPM starts from a globally trained full binary tree and, for each user-item pair, samples a pruning mask with a policy network trained by REINFORCE to maximize the reward of the pruned subtree relative to the global tree, where the reward is XAUC minus MSE. The paper also claims TPM's conditional modeling suffers from selection bias, since each internal classifier is trained only on samples whose watch time falls into its parent interval, and it derives an inverse-propensity-scoring correction whose propensity is obtained from the same path probabilities. The combination yields consistent offline gains on KuaiRec, CIKM16, and WeChat21, and a +0.428% watch-time lift in an online A/B test with essentially unchanged latency, on a production service reported to serve over 400 million users daily.","pith_inferences":["Beyond the paper: the same pruning-plus-inherited-parameters trick could apply to any hierarchical ordinal decomposition, such as CREAD-style threshold classifiers, turning static binning into per-request binning at near-zero added cost.","Beyond the paper: because the IPS propensity is built from the tree's own path probabilities, any conditional classifier tree can in principle be debiased by reweighting with ancestor probabilities, not just watch-time trees.","Beyond the paper: the average learned tree depth converges to about 4.6 for a global depth of 6, which hints that the benefit comes mainly from reshaping rather than deepening; a testable extension is learning depth per cohort.","Beyond the paper: the reward $R_T = \\mathrm{XAUC} - \\mathrm{MSE}$ is one scalar choice; replacing it with ranking-aware or business-aware rewards could yield trees tuned for engagement trade-offs, which the paper does not test."],"forward_implications":["If PTPM's claims hold, watch-time models can switch from fixed discretization to sample-adaptive discretization without extra inference latency, since only classifier selection changes.","The UCL correction implies that TPM's conditional classifiers are systematically miscalibrated in tail intervals; debiasing them improves both distributional calibration and classifier AUC.","The pruning formulation reduces bilevel tree-structure search to a one-pass policy-gradient problem, so deeper global trees provide a larger search space and better learned trees, up to diminishing returns.","The online +0.428% watch-time lift suggests that ranking quality improves when prediction granularity adapts to context, and that this gain carries over to the deployed serving system."],"supporting_citations":[{"why":"Defines TPM, the base paradigm this paper extends; supplies the tree decomposition, loss terms, and the fixed-tree baseline.","marker":"[10]"},{"why":"CREAD, the prior discretization-aware baseline that PTPM must beat and whose static binning motivates sample-adaptive learning.","marker":"[12]"},{"why":"Introduces D2Q and the XAUC evaluation metric used throughout the paper; a key baseline.","marker":"[16]"},{"why":"DML baseline for duration-mapped watch-time regression; one of the state-of-the-art comparisons.","marker":"[17]"},{"why":"WLR baseline; supplies the weighted-logistic approach that PTPM compares against as an earlier industrial standard.","marker":"[5]"},{"why":"KuaiRec, the public dataset used for offline evaluation.","marker":"[7]"},{"why":"REINFORCE, the policy-gradient algorithm used to train the pruning model.","marker":"[14]"},{"why":"Supplies the inverse-propensity-scoring background for the unbiased conditional learning module.","marker":"[8]"}],"fun_headline_variants":["Personalized tree pruning boosts watch-time prediction","Learnable tree structure improves short-video watch time","Debiased tree model lifts watch-time accuracy 0.428%","Pruning the tree per request: better watch-time prediction","Adaptive tree structure for watch-time: 0.428% online gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a classifier trained on the global tree remains correct on a pruned subtree of it, so the inherited parameters need no fine-tuning; Section 4.1 asserts this but gives no experiment that isolates it, and if it fails, the sampled pruned trees would misestimate leaf probabilities and both the efficiency and accuracy gains would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Personalized tree pruning boosts watch-time prediction","Learnable tree structure improves short-video watch time","Debiased tree model lifts watch-time accuracy 0.428%","Pruning the tree per request: better watch-time prediction","Adaptive tree structure for watch-time: 0.428% online gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000406,"raw_usage":{"total_tokens":2163,"prompt_tokens":1047,"completion_tokens":1116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1032}},"tokens_in":663,"tokens_out":1116,"duration_ms":8057,"temperature":1.0,"reasoning_tokens":1032,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:14:29.366865+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two versions of PTPM on the same data: the paper's inherited-parameter pruning versus re-training the classifiers of every sampled pruned subtree from scratch for the same number of steps. If the retrained subtree beats the inherited-parameter version by a margin comparable to PTPM's stated gains over TPM, then the no-fine-tuning assumption is the true source of the improvement. A second, cheaper check is to measure per-leaf calibration on out-of-parent-interval samples after pruning, comparing inherited versus retrained classifiers.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CREAD, the prior discretization-aware baseline that PTPM must beat and whose static binning motivates sample-adaptive learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces D2Q and the XAUC evaluation metric used throughout the paper; a key baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DML baseline for duration-mapped watch-time regression; one of the state-of-the-art comparisons."},{"cited_title":"Williams","cited_arxiv_id":null,"evidence_quote":"REINFORCE, the policy-gradient algorithm used to train the pruning model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the inverse-propensity-scoring background for the unbiased conditional learning module."}],"review_version":1}