Kurate.org
Rankings
Back to Rankings

CoGSeq-LTV: Scalable Missing-Aware Graph-Sequence Modeling for Multi-Business Transaction Analytics

Ziyang Su, Yao Zhou, Kelin Liu, Guanjie Zheng, Linghe Kong

Nov 1, 2026arXiv:2611.01817v1
cs.AI
Share
Scorecard· 16/16
4.0/10 impact

Competent applied integration of known components with a clear leakage protocol, but evidence is limited to one proprietary dataset with inconsistent tables and non-comparable baselines.

Abstract

Large-scale super-apps generate heterogeneous transaction logs in which users interact with multiple businesses, item identifiers are only partially available, and future value targets are sparse, heavy-tailed, and cumulative over time. We study multi-business, multi-horizon value prediction as a data-intensive analytics problem over such logs, rather than as a single-table task. (i) a batch-induced user-item graph that provides collaborative signals without materializing a global graph and remains applicable when item identifiers are missing through reserved unknown-item handling, (ii) business-aware sequence encoders that preserve recency, inter-event gaps, payment context, and goods-table availability, and (iii) cross-view contrastive alignment that transfers user-level intent across businesses while retaining business-specific temporal patterns. To handle zero-inflated and long-tailed cumulative targets, CoGSeq-LTV adopts a zero-inflated mixture log-normal head with MAE-aligned point supervision and validation-calibrated zero-aware inference. Experiments on the BiBiSheng transaction-log dataset show that CoGSeq-LTV improves over representative LTV baselines across businesses and horizons, including under a later out-of-time evaluation window.

AI Impact Assessments

(1 model)

Core Contribution

CoGSeq-LTV is an end-to-end architecture for multi-business, multi-horizon LTV prediction on super-app transaction logs. It combines four components:

  • A batch-induced user–item bipartite graph, with one mean-aggregation layer and a shared hashed item embedding table. It requires no global graph and uses a reserved unknown-item id for missing identifiers.
  • Two business-specific Transformer encoders over event tokens that carry recency, gaps, amount, and a goods-match flag.
  • InfoNCE-style contrastive alignment across the e-commerce, food-delivery, and graph views.
  • A zero-inflated mixture log-normal (ZIMLN) head with an MAE-aligned auxiliary loss and validation-tuned zero thresholding.

    • Each piece is a known technique, and the novelty is in the integration and in the explicit treatment of systematic item-ID missingness. The "batch-induced graph" is essentially a mini-batch co-occurrence layer. Because items are hashed into shared buckets, it approximates collaborative smoothing, but it is not a rigorous graph method. ZIMLN is a modest extension of ZILN, which already exists in the literature. The hard monotonicity of cumulative targets is acknowledged but not enforced.

    Methodological Rigor

    The leakage protocol is clearly specified: anchor-time cutoffs, user-level splits, and caps and thresholds fit on training and validation data only. That is a plus. Several issues weaken the empirical case:

  • Only one proprietary dataset is used (BiBiSheng, from Tencent/WeChat Pay). No public benchmark is included, and the dataset's release status is unclear.
  • The baselines are single-business tabular LTV models (ZILN, ODMN, MDAN, and others). None of them uses sequences or graphs, so the comparison confounds the architecture with the input representation. There is no strong sequence baseline such as SASRec or a Transformer with a standard regression or ZILN head. There is no graph baseline either, and no multi-task backbone with the same inputs.
  • The reported numbers are internally inconsistent. The ablation's full-model MAE (e.g., 289.43 at 7 days for e-commerce) differs from Table 2 (261.55). Table 2 WAPE values are not consistent with the Table 5/6 subset values, where the subset baselines also differ substantially. The paper hedges that tables were "exported from independently executed evaluation jobs". This undermines trust in the numbers.
  • The temporal-shift table has suspiciously regular-looking numbers (many values ending in "4", "5", "2"). It also shows the proposed model degrading little or even improving under shift. Combined with the inconsistencies, this raises integrity concerns.
  • Several robustness claims are rationalised after the fact. The text says the model is "best or near-best" on low-frequency users, yet Table 5 shows it losing several targets, including MDAN and expLTV on some MAE entries. The "all 16 targets" high-frequency claim is not bolstered by significance tests.
  • The ablation shows that the largest drop comes from removing the user-item graph, which is surprising given the single-layer, batch-local graph. This is plausible only if the graph pathway also carries the item embeddings, so the ablation conflates graph propagation with item-embedding usage.
  • Reported standard deviations are for baselines in the main tables, but the sources of variance (seeds, splits) are not explained. Ablations have no error bars.
  • The ZIMLN-versus-MAE ablation is useful, but the zero-threshold tuning is applied only to the proposed model, not necessarily to the baselines.

    • Potential Impact

    The practical audience is industrial LTV and marketing-analytics teams at super-apps. The missing-ID handling (unknown token, match flag, three event states) and the leakage protocol are useful engineering patterns. The scientific contribution is incremental, though. The paper offers no theoretical insight, and the method is a composition of standard modules whose benefit is tested on a single dataset. The dataset could be a valuable contribution if released, but no release is claimed. The WWW-style venue fits the applied-data-mining framing.

    Timeliness & Relevance

    Multi-business user modeling and LTV prediction under zero-inflated, heavy-tailed targets are active industrial topics. Missing-metadata robustness is a real practical concern. The topic is relevant but crowded, and there are many similar LTV papers (OptDist, MDAN, MEAN, and others) with similar evaluation styles.

    Strengths

  • A clearly stated problem with realistic data-engineering constraints.
  • A careful leakage and protocol description.
  • A scalable design rationale that avoids global graph materialization.
  • A comprehensive evaluation structure: main results, ablation, sensitivity, frequency subsets, and out-of-time evaluation.
  • An honest discussion of low-frequency limitations.

    • Limitations

  • Single proprietary dataset with two businesses.
  • Baselines lack sequence and graph inputs, so there is no fair architectural comparison.
  • Numerical inconsistencies across tables, with a hedging caveat.
  • Improvements over MDAN are moderate (roughly 5–20% MAE), and statistical significance is not established.
  • The graph component is weakly motivated (batch-local, mean aggregation), and its effectiveness is not diagnosed.
  • No code or data release is mentioned, and no efficiency or scalability measurements (memory, throughput) are given despite "scalable" in the title.
  • Monotonic cumulative structure is not enforced, so predictions can violate it.
  • The out-of-time test covers only one later window.

    • Overall

    This is a competent applied engineering paper with a reasonable protocol, but its novelty is incremental. The evidence is weakened by inconsistent tables, unequal baselines, and a single private dataset. Its likely influence is modest, mainly as a reference for industrial LTV practitioners.

    Rating:4/ 10
    Significance 3.5Rigor 4Novelty 3.5Clarity 6

    Generated Nov 3, 2026

    Comparison History (0)

    No comparisons yet.