acceptodds
Under review as a conference paper at ICLR 2027

Learning the Count, Keeping the Unit: Task-Algebra Heads for Relational Regression

Abstract

Relational prediction tasks define their targets through programs: a future sum, a count, or an event condition. These programs can specify how a model should predict, as well as what it should predict. We develop Task-Algebra, a compiler for a restricted aggregation-query fragment that extracts a query-time unit factor, transforms the training target, and reconstructs predictions with that factor kept exact. For a static entity-owned unit, a future sum becomes ; monetary absolute error becomes exactly -weighted count error. A conditional-median approximation result explains why this parameterization can save finite model capacity even when both models receive identical inputs. On four chronological Amazon item-LTV development evaluations with matched 67-feature CatBoost heads, the factorized head reduces mean absolute error from 44.481 to 42.071, a 5.42% reduction, with a category-clustered gain interval of . Larger trees, numerical loss controls, a history-selected population, and continuous decoders reveal how approximation and optimization contribute. The resulting account connects a task's aggregation semantics to an executable prediction head and to the regimes in which keeping a known unit outside the learner is useful.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
relational learning, relational regression, task semantics, inductive bias, target parameterization, tabular learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ppZgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper observes that several RelBench regression targets are defined by SQL of the form "sum a value column over future events joined to the entity", and that when the value column is a static per-entity unit (a price) the target factorises as price × count, with the price known at query time. TASK-ALGEBRA is a small SQLGlot-based compiler that detects this pattern in the task manifest, trains the model on the count (or on a price-scaled residual over an anchor) and multiplies back. Because the monetary absolute error is the price times the count error, a monetary MAE objective becomes a price-weighted count MAE, implemented with sample weights. A proposition shows that under a uniform price and a count whose conditional median does not depend on price, a constant count head is Bayes-optimal for MAE while any piecewise-constant monetary predictor with finitely many pieces has a positive excess error. Experiments on rel-amazon item-ltv with a CatBoost head report a mean MAE reduction of about 2.4 over four development origins (Table 1), followed by controls on tree size, CatBoost's quantile tolerance, a residual MLP, a history-selected population, scale transport and price trimming, and an H&M item-sales experiment.

Critical strengths
  • The observation is correct and cleanly stated: if the target is price × count and the price is known, predicting the count and multiplying back is the better-conditioned problem, and the loss correspondence is exact rather than heuristic.
  • Both heads keep the price as an input feature, which avoids the classic mistake of regressing only on the ratio.
  • The compiler's refusal checks and the replay checks against raw history are sensible engineering, and the scale-transport experiment is a nice sanity check that the factorised head is equivariant to a change of currency.
  • The authors are forthright about limitations (the conclusion itself says that broader baselines, multiple seeds and official validation/test evaluation remain necessary), and the finding that CatBoost's default quantile tolerance does not scale with the target is genuinely useful, even though it undercuts the paper's own story.
Critical weaknesses

The idea is standard offset / exposure modelling presented as new, the theoretical result is a classical approximation rate whose bound is negligible for the trained models, and the empirical claim is made on training-period windows with no official split, no published baseline, one seed, and a set of controls that largely erase the gain. The authors are candid about most of these gaps, which I respect, but candour about a missing experimental programme is not a substitute for it.

1. This is offset / exposure modelling, and the paper does not position it as such. Fitting a rate and multiplying by a known per-row scale is the textbook GLM offset, and it is the standard device in actuarial pricing, where the exposure-weighted deviance is exactly the weighted loss derived here. It has been done with boosted trees for years (Henckaerts et al. 2021; TDboost), every major GBDT library exposes it directly (init_score in LightGBM, baseline in CatBoost, base_margin in XGBoost), the Poisson and Tweedie deviances satisfy the same scaling identity, DeepAR normalises each series by a known item scale and weights accordingly, and the M5 competition was explicitly "predict units, evaluate with dollar-weighted error". None of this is cited. The related-work paragraph on known scales cites two papers, and one of them (CANN) is precisely the anchor-plus-residual head used here. What is actually new is deriving the scale automatically from the task SQL, a parser that currently accepts one aggregation pattern on three hand-written manifests, and the identity-link MAE case, which is one line of algebra. The paper needs to say "offset modelling applied to RelBench" and then argue why that is an ICLR contribution.

2. The proposition is a known approximation rate and does not explain the experiments. That a piecewise-constant function with L pieces approximates a linear one with error of order 1/L is classical, and the squared-error version appears in existing tree lower bounds (Tan et al. 2022). More importantly the bound does not bite: CatBoost quantises each feature into a few hundred bins, so the whole ensemble is piecewise constant in price with a bounded number of pieces, and the resulting bound is a few cents, two orders of magnitude below the gain claimed. There is a deeper problem. Since price is an input feature, the conditional median of the monetary target is simply price times the conditional median of the count, so both heads target the same function and the weighted loss has the same population minimiser as the monetary loss. Any gain is therefore a finite-capacity or optimiser effect, which is consistent with the gain shrinking for the larger ensemble in Table 3. If the story is "trees cannot represent the linear factor", the honest baseline is a tree with linear leaves (LightGBM linear_tree, Cubist), which needs no knowledge of the task SQL. It is not run.

3. The evaluation never touches the official validation or test split and never compares to a published RelBench number. All four origins in Table 1 are the last four official train timestamps, and Appendix B says the official validation/test outcomes are not used in this draft. The same windows serve for development and for reporting. The paper cites RelBench, RDL, KumoRFM-2 and others but reports none of their item-ltv or item-sales numbers, and does not run the LightGBM baseline that ships with the benchmark. Those numbers exist, and the recent systems put the item-ltv test bar at roughly 41–42, so a 42 on training-period windows cannot be placed against anything. RDL's own validation-to-test gap on this task is about twice the headline gain, so I have no idea whether the gain survives a period shift. RelBench scores the masked test labels through its own evaluator, so not using the frozen split is a choice, not a constraint. For an empirical paper on a public benchmark this is disqualifying on its own.

4. The headline gain is fragile under the paper's own controls, and the official roster helps the factor head. With the quantile tolerance set to zero (the numerically clean setting) the October gain flips sign and the two-origin mean has an interval covering zero; trimming the top 1% of prices removes most of the gain; for the residual MLP the factor head is worse; on the history-selected population, the only roster that can actually be built at prediction time, a plain direct MLP beats the factor tree at both origins; and on H&M unit normalisation adds nothing over the ordinary residual baseline. Read together, the gain is real only for CatBoost at default tolerance on the full price tail. Separately, the official item-ltv roster only contains products with at least one future review, so every label is at least the price. The count head only has to learn "at least one", which a constant already does, while the direct tree has to learn a floor that varies with price. Part of the gain is therefore an artefact of a roster that requires future knowledge. I would want the gain broken down by price decile before believing anything about the mechanism.

Questions for authors
  1. What are the official RelBench validation and test MAEs of both heads on item-ltv and item-sales, in one table with the published LightGBM, RDL and KumoRFM-2 numbers?
  2. Please report Tables 1 and 3 with five seeds per arm, with the tolerance set to zero as the primary configuration, matched rows in both arms, and per-head tuning on a held-out origin.
  3. How does the method differ from CatBoost with a Poisson/Tweedie objective and log-price as baseline, or from a linear-leaf tree, scored by monetary MAE? Please run both on the same features.
  4. Since price is an input feature, do you agree that the weighted count loss and the monetary loss have the same population minimiser? If so, what is the proposition explaining, and what is its numerical value in dollars for the trained ensemble?
Reviewer_FKRtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper’s main idea is to separate out learning the unknown component of a prediction while keeping known components (either fixed/or available fairly directly as features) factorized out. Using item lifetime-value forecasting, it shows that when price is fixed and known, a monetary target can be predicted by forecasting quantity and multiplying by price; its Amazon benchmark uses future review counts rather than actual unit sales. Across four chronological Amazon evaluations, this reduced CatBoost’s mean absolute error from 44.481 to 42.071—a 5.42% improvement. However, neural-network results were mixed, and on H&M’s variable-price task, unit normalization showed no reliable additional benefit over a residual model anchored to historical sales amounts. Evidence remains concentrated on development evaluations, without multiple-seed validation or official held-out testing.

Critical strengths

Simplifies certain revenue prediction/forecasting tasks to their core learnable/predictable pieces such as unit sales while keeping know factors out of prediction (such as unit price). Its a practical approach, and speaks more towards effectively setting up forecasting ml tasks.

Critical weaknesses

The novelty of the work appears limited beyond automating an established factorization. Section 2.1, Eq. (2), derives Y=sN for a declared static price, and Section 2.2, Eq. (4), gives the corresponding loss weights. Section 5 acknowledges these established modeling ideas. Proposition 1 formalizes an advantage over restricted piecewise-constant predictors, but does not address changing prices.

• Existing treatment: Section 2.1 allows an estimated unit for event-varying prices, and Section 3, Eq. (7), identifies a unit-drift error term. However, Section 2.3 and Appendix E rely on declared metadata and an explicitly supplied historical estimator; they do not demonstrate detection of emerging instability or adaptive head selection. The H&M experiment in Section 4.4 shows essentially no additional benefit from normalization over the residual head.

• Direction for stronger novelty: Detect price instability from historical observations and adapt between fixed-unit count prediction, time-varying unit modeling, and direct amount prediction. Demonstrating gains under temporal price changes would strengthen the contribution beyond the current query-to-head compilation.

Questions for authors

expressed in review comments

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

This paper proposes an approach called Task-Algebra, which is a compiler that factors a static unit ss out of a relational SUM target (Y=sNY = sN), and trains a head to predict the count N and reconstructs Y^=s∗g(X)\hat{Y} = s * g(X). The authors show an MAE gain with a CatBoost count head compared to the alternative on an Amazon item-LTV dataset. Overall, the paper's observation is valid but the overall novelty and depth is limited and makes it unsuitable for publication at this time.

Critical strengths
  • The math and key idea is sound -- the authors show that regressing count rather than price and having to learn multiplication is indeed better than the alternative.

  • Figure 1 is clear and illustrates the key idea well.

Critical weaknesses
  • The novelty of the work is fairly thin and the authors themselves point out the idea adopted here is commonly used in frequency-severity pricing.

  • The δ=0\delta=0 setting for CatBoost includes a confidence interval encompassing 0 gain, which weakens the core claim as very relevant to an implementation detail. The main results are shown on one dataset and seed without official validation and test splits.

  • There are no baselines or reproducibility used as a comparison point. There is ample space left in the manuscript and missing these strongly undercuts the contribution.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.