CC-OR-Net: A Unified Framework for LTV Prediction through Structural Decoupling

AI summary

Overview

Research area: Machine learning for user modeling and computational advertising — specifically Customer Lifetime Value (LTV) prediction on large-scale web platforms. Published at the ACM Web Conference 2026 (WWW '26), Dubai, United Arab Emirates (DOI: 10.1145/3774904.3792259).

Technical level: Advanced. The paper assumes familiarity with ordinal regression, zero-inflated and long-tail distributions, cascaded classifiers, residual networks, attention-guided augmentation, and multi-objective loss design.

One-sentence scope: The paper proposes CC-OR-Net (Conditional Cascaded Ordinal-Residual Networks), a unified architecture that structurally guarantees user ranking through a cascaded ordinal decomposition, then recovers fine-grained regression and high-value precision through dedicated residual and augmentation modules, evaluated on three industrial LTV datasets.

What This Paper Is About

LTV prediction data is extremely skewed: most users are non-paying or low-value, non-zero values follow a severe long tail, and a sparse but business-critical "whale" segment generates a large share of revenue but contributes almost nothing to a standard loss function. The paper argues that models face a "LTV prediction trilemma" — they must simultaneously optimize ranking quality, regression accuracy, and high-value precision, and existing approaches (fixed statistical assumptions, or ordered buckets that enforce ordinality only through loss penalties) cannot balance all three. CC-OR-Net's goal is to convert this hard conflict into a manageable trade-off by decoupling ranking from regression at the architectural level, so that ranking is guaranteed by structure rather than learned constraints.

Key Contributions

1. A unified framework that structurally decouples ordinal ranking from regression, featuring a scalable ordinal cascade with O(K) complexity suited to industrial-scale deployment.
2. An attention-guided augmentation strategy that targets learning on rare, high-value users, positioned as an alternative to generic re-sampling methods.
3. The Stratified Value Accuracy (SVA) metric, a business-centric evaluation tool for user stratification tasks based on a value-distribution-driven adaptive threshold.
4. Validation at scale on three industrial datasets drawn from different domains, totaling 322M records (248M + 41M + 33M), compared against ten baseline methods across four categories.

Main Findings

• CC-OR-Net leads on the combined trade-off, not on every single metric. On Domain 1/2/3 it achieves GINI of 0.803/0.490/0.501, Spearman rho of 0.761/0.278/0.442, AMBE of 4.85/0.90/21.18, NMAE of 0.776/0.986/0.982, MAPE of 5.53/1.25/21.06, NRMSE of 2.093/4.834/3.316, and SVA of 67.01%/60.50%/53.89%. The paper reports all improvements as statistically significant (p < 0.05).

• SVA gains over the strongest specialized baselines. MDME reaches SVA of 60.55%/52.78%/51.41%, ExpLTV 57.15%/49.90%/38.10%, and OptDist 55.42%/40.49%/32.87%, versus CC-OR-Net's 67.01%/60.50%/53.89%. ExpLTV records lower AMBE on Domain 1 (5.12 vs. 4.85 is CC-OR-Net's own 4.85 — ExpLTV's 5.12 is higher; MDME's is 5.95) and lower AMBE on Domain 2 (1.15 vs. 0.90), so the paper frames its advantage as holistic rather than single-metric.

• GINI improvements are modest by design. The paper states that for severely imbalanced LTV distributions a high global GINI can mask deficiencies in high-value segments, citing prior industrial-scale work that GINI may not reflect performance differences on imbalanced data.

• Ablation shows each module changes the trade-off differently (Domain 1). The baseline scores GINI 0.671, Spearman rho 0.726, NMAE 1.135, MAPE 11.023, AMBE 5.704, NRMSE 2.999, SVA 62.50%, F1 0.695, Bucket-Acc 0.486. Adding cascaded classification lifts GINI to 0.798 and SVA to 67.17%. Adding the distillation module raises Spearman rho to 0.759 and lowers NMAE to 0.767 and NRMSE to 2.144, while GINI slips to 0.794 and SVA to 66.55% — the paper describes this as distillation slightly diluting focus on the tail. Adding residual learning pushes GINI to 0.805 (the peak) but raises AMBE to 6.463, which the authors call empirical confirmation that globally optimized regression is ill-equipped for whale users. The full model (with high-value augmentation) settles at GINI 0.803, AMBE 4.849, NRMSE 2.093, SVA 67.01%.

• The augmentation module is a targeted corrector on the top bucket. Table 4 reports the top bucket moving from AMBE 6.463 to 4.849 (-25.0%), NRMSE from 2.475 to 2.132 (-13.9%), and F1 from 0.637 to 0.674 (+5.8%), at the cost of a "negligible dip" in global ranking.

• Distillation reduces distribution mismatch. The paper reports a 12.7% reduction in Chi-squared on Domain 1 from the distillation loss regularizing batch-level bucket distribution against the empirical one.

• Two-stage separation alone is insufficient. Two-stage XGB shows high AMBE on the whale segment (10.32 on Domain 1, 4.38 on Domain 2, 71.10 on Domain 3), which the paper cites as evidence that simply separating zero/non-zero prediction does not solve high-value bias.

• Recall@k business analysis is described but not quantified in the available content. The paper introduces Recall@5000 on Domain 1 and poses the question of what fraction of true whales are captured within a fixed budget; the truncated text ends before the results are reported.

Methodology in Plain English

The approach rests on reframing one hard regression problem as a chain of easier binary questions. The team sets four value buckets (K = 4) using fixed, data-driven quantile thresholds. Bucket 1 isolates zero-value users (values within [-10⁻⁶, 10⁻⁶]); the non-zero users are split into three further buckets using non-zero quantiles (e.g., 50th, 75th) to create balanced populations.

A shared encoder turns heterogeneous features (categorical, continuous, sequential) into one representation used by every downstream module. The centerpiece is the conditional cascaded ordinal decomposition: K-1 = 4-1 = 3 binary classifiers, each asking "is this user's value above threshold tau_k, given they already passed the previous threshold?" Chaining these conditional probabilities yields a valid K-way distribution whose components sum to 1 (proven via a telescoping series in Appendix A). Because bucket k can only be reached by passing all earlier thresholds, ordinality is guaranteed by the architecture rather than by a penalty term.

Training uses teacher forcing, where each classifier predicts marginal probabilities on the full dataset, while inference walks the cascade sequentially with conditional probabilities. To close the train-inference gap, the authors add a distillation loss that matches the predicted batch-level bucket distribution to the empirical one, plus fixed data-driven quantile thresholds. A stop-gradient operation sits between the cascade and distillation modules to prevent gradient conflicts.

For fine-grained regression, a Feature Alignment Mechanism enriches the shared representation with both soft (cascade probabilities) and hard (predicted bucket embedding) context, then filters it through a GLU-style gate: h_aligned = g ⊙ c. The Intra-Bucket Residual Learning module uses a dual-block residual architecture with skip connections, predicting a normalized value in [-1, 1] via tanh and denormalizing with bucket-specific half-range and center: v_final = v_norm × r_b + c_b. Zero-value samples get dynamic label smoothing — their normalized target of 0 is replaced by a uniform sample from [-θ, θ], where θ decays linearly from 1.0 to 0.1 during training.

For whales, an attention-guided augmentation module operates only on the predicted highest-value segment. It computes attention weights from the feature vector, its summary statistics (mean, standard deviation, max, min), and the bucket embedding, then adds noise scaled by those weights: h_aug = h + ε ⊙ w with noise drawn from N(0, σ²_noise) and σ_noise = 0.1. The paper describes this as feature-level regularization similar to structured adversarial training.

The loss is a weighted sum with γ = 0.8: L_i = γ·L_main,i + (1-γ)·I(i ∈ H)·L_high_value,i. The main loss combines cascade, residual, and distillation terms. The residual term uses a value-weighted MSE with a sigmoid weighting term at β = 0.5 to up-weight higher-value samples within each bucket. The high-value loss combines a relative regression error (denominator clipped at 1.0 for stability) with a Focal Loss-inspired confidence term at β_focal = 2. The confidence head is training-time only and is not used during inference.

Why This Matters

Impact on research: The paper's core claim is a design-principle shift — ordinal consistency should be enforced by architecture, not by loss constraints. If that holds, it reframes how researchers approach any ordered prediction problem with skewed distributions, and the SVA metric offers a business-aligned alternative to GINI and NMAE for evaluating stratification. The fixed, structurally-defined pipeline is also positioned explicitly against soft-routing multi-expert systems like MMOE and ExpLTV, arguing that learnable routing is unstable when high-value data is sparse.

Real-world applications:

• Marketing budget allocation and targeting, where identifying true whales within a fixed budget determines ROI.
• User segmentation and stratification for campaigns that need reliable zero / low-value / high-value separation.
• ROI optimization and personalized user experiences on large web platforms, which the paper names as the motivating context.
• Extensions to value settings with negative outcomes — the authors note returns, service costs, platform subsidies, and marketing activities can be handled with dedicated negative and zero buckets and an adjusted regression head (e.g., removing the final ReLU), which they state they validated experimentally.

Industry relevance: The work comes from Meituan and is evaluated on 322M records across three domains, with batching of 100,000 samples on 14-core CPUs and O(K) complexity. The paper repeatedly emphasizes industrial-scale deployment and the accuracy-latency-cost trade-off, including a three-way analysis (Figure 3) plotting SVA against AMBE with inference latency as bubble size, showing that the augmentation module spends a minor latency increase for a 25% AMBE reduction.

Future Directions

1. Publish the Recall@k business results. The paper introduces Recall@5000 on Domain 1 and frames the budget-efficiency question, but the truncated content does not report the numbers — completing this analysis would substantiate the commercial claims in the abstract.

2. Test larger values of K. The authors state K = 4 was chosen to balance granularity with computational efficiency, and that K can be increased for finer-grained predictions, claiming higher K yields finer stratification without compromising performance or stability. That claim is asserted for K = 4 only and remains to be verified at larger K.

3. Formalize the negative-value extension. The framework is described as readily extensible to negative LTV (returns, service costs, subsidies, marketing activities) and the authors report validating it, but state it is beyond the paper's core scope and do not discuss it in detail.

4. Compare against the architectures that were only surveyed. NGBoost, ZIGP, DeepGBM, and MDLUR are discussed in Related Work and described as offering non-competitive trade-offs in preliminary experiments, but they were omitted from the main tables; publishing those comparisons would clarify the boundary of the method's advantage.

5. Reconcile the results with GINI-based leaderboards. The paper acknowledges that some specialized baselines are stronger on individual metrics such as AMBE, and that GINI can mislead on imbalanced data. Independent evaluation of whether SVA and the reported trade-off generalize beyond these three industrial datasets would be a natural next step.

Target Audience

Researchers and practitioners working on customer lifetime value prediction, computational advertising, and large-scale user modeling on web platforms. It is most useful for readers already comfortable with ordinal regression, long-tail learning, and multi-task loss design, and for applied scientists at industrial platforms who need a deployable architecture rather than a purely theoretical result. Readers focused on the business evaluation side — marketing analytics and ROI measurement — will find the SVA metric and the Recall@k framing the most directly transferable parts. The paper is not beginner-friendly: the methodology section relies on probability chaining, residual block design, and specific loss formulations without introductory exposition.

Authors’ abstract

Customer Lifetime Value (LTV) prediction, a central problem in modern marketing, is characterized by a unique zero-inflated and long-tail data distribution. This distribution presents two fundamental challenges: (1) the vast majority of low-to-medium value users numerically overwhelm the small but critically important segment of high-value "whale" users, and (2) significant value heterogeneity exists even within the low-to-medium value user base. Common approaches either rely on rigid statistical assumptions or attempt to decouple ranking and regression using ordered buckets; however, they often enforce ordinality through loss-based constraints rather than inherent architectural design, failing to balance global accuracy with high-value precision. To address this gap, we propose \textbf{C}onditional \textbf{C}ascaded \textbf{O}rdinal-\textbf{R}esidual Networks \textbf{(CC-OR-Net)}, a novel unified framework that achieves a more robust decoupling through \textbf{structural decomposition}, where ranking is architecturally guaranteed. CC-OR-Net integrates three specialized components: a \textit{structural ordinal decomposition module} for robust ranking, an \textit{intra-bucket residual module} for fine-grained regression, and a \textit{targeted high-value augmentation module} for precision on top-tier users. Evaluated on real-world datasets with over 300M users, CC-OR-Net achieves a superior trade-off across all key business metrics, outperforming state-of-the-art methods in creating a holistic and commercially valuable LTV prediction solution.