CoGSeq-LTV: Scalable Missing-Aware Graph-Sequence Modeling for Multi-Business Transaction Analytics
Authors: Ziyang Su, Yao Zhou, Kelin Liu, Guanjie Zheng, Linghe Kong
Organizations: School of Computer Science, Shanghai Jiao Tong University, Shanghai, China · Tencent WeChat Pay Research and Development Department, Shenzhen, China
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.
Figures & tables
Figure 1. The overall architecture of CoGSeq-LTV. A block diagram showing a shared user representation, a batch-induced user-item graph encoder, two business-specific transaction sequence encoders, cross-business contrastive alignment, and a zero-inflated mixture log-normal prediction head.
Statistic
Value
Statistic
Value
#Users
4,850,141
#Total records
41,842,772
#In-domain items
1,159,378
#In-domain records
19,067,648
#E-commerce (ch. 237)
10,452,694
#Food delivery (ch. 689)
31,390,078
Channel
#Records (%)
Domain
Coverage
2 (JD)
6,597,985 (15.8%)
In
100%
3 (VIP)
399,327 (1.0%)
In
100%
Table 1. Statistics of the BiBiSheng dataset. “In-domain”: accessible goods_id ; “Out-domain”: goods_id missing.
Horizon
Metric
Model
ZILN ( Wang et al., 2019 )
ODMN ( Li et al., 2022 )
perCLTV ( Zhao et al., 2023 )
OptDist ( Weng et al., 2024 )
expLTV ( Zhang et al., 2023 )
MEAN ( Liu et al., 2025 )
MDAN ( Liu et al., 2024 )
CoGSeq-LTV
7
MAE
683.47 ± 42.15
371.04 ± 58.32
614.25 ± 61.47
542.17 ± 45.81
512.39 ± 41.26
324.96 ± 38.54
305.28 ± 37.19
261.55 ± 39.43
WAPE
1.872 ± 0.022
1.165 ± 0.035
1.628 ± 0.038
1.618 ± 0.024
1.548 ± 0.021
1.082 ± 0.019
0.992 ± 0.018
0.944 ± 0.020
14
MAE
3346.55 ± 115.42
1588.39 ± 152.63
2776.92 ± 168.91
2315.42 ± 121.37
2197.84 ± 118.54
1503.27 ± 105.28
1342.71 ± 102.16
1040.59 ± 107.85
WAPE
1.894 ± 0.032
1.229 ± 0.048
1.589 ± 0.051
1.722 ± 0.035
1.635 ± 0.031
1.097 ± 0.028
1.020 ± 0.027
0.948 ± 0.030
30
MAE
5432.74 ± 135.82
2763.75 ± 185.47
4912.48 ± 192.36
3964.81 ± 142.19
3768.15 ± 138.65
2436.52 ± 125.43
2279.56 ± 122.81
1897.07 ± 127.15
Table 2. Performance comparison on E-commerce business (MAE and WAPE, lower is better).
Horizon
Metric
Model
ZILN ( Wang et al., 2019 )
ODMN ( Li et al., 2022 )
perCLTV ( Zhao et al., 2023 )
OptDist ( Weng et al., 2024 )
expLTV ( Zhang et al., 2023 )
MEAN ( Liu et al., 2025 )
MDAN ( Liu et al., 2024 )
CoGSeq-LTV
7
MAE
711.34 ± 14.52
458.27 ± 28.45
651.68 ± 35.12
671.39 ± 12.84
634.81 ± 11.36
415.73 ± 8.45
381.42 ± 6.12
365.61 ± 4.69
WAPE
1.895 ± 0.012
1.138 ± 0.025
1.662 ± 0.031
1.654 ± 0.015
1.582 ± 0.011
1.087 ± 0.009
0.982 ± 0.007
0.948 ± 0.004
14
MAE
1715.62 ± 32.15
1032.51 ± 65.34
1553.17 ± 82.19
1402.53 ± 28.41
1328.46 ± 25.63
962.18 ± 21.05
879.15 ± 18.34
826.87 ± 19.81
WAPE
1.972 ± 0.018
1.170 ± 0.038
1.648 ± 0.042
1.687 ± 0.016
1.611 ± 0.014
1.041 ± 0.012
0.968 ± 0.011
0.926 ± 0.010
30
MAE
3823.18 ± 68.42
2315.92 ± 145.67
3279.34 ± 186.23
3468.15 ± 55.18
3291.37 ± 52.34
2087.45 ± 42.15
1885.07 ± 38.62
1793.41 ± 39.96
Table 3. Performance comparison on Food Delivery business (MAE and WAPE, lower is better).
Business
Horizon
Metric
Model/Variants
w/o User-Item Graph
w/o Contrastive Loss
w/o ZIMLN Module
w/o MAE-aligned
CoGSeq-LTV
E-commerce
7
MAE
425.34
312.45
335.18
352.67
289.43
WAPE
1.2541
0.9912
1.0456
1.1123
0.9583
14
MAE
1815.47
1368.74
1425.36
1512.89
1299.69
WAPE
1.2845
1.0125
1.0678
1.1567
0.9689
30
MAE
3138.41
2315.42
2458.91
2615.34
2199.11
Table 4. Ablation experiment results on E-commerce and Food Delivery businesses. Variants are compared within the same ablation run group.
Figure 2. Parameter sensitivity results of mixture distribution number K .
Figure 3. Parameter sensitivity results of contrastive loss weight λcl .
Figure 4. Case study result of CoGSeq-LTV model.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Business
Horizon
Metric
Model
ZILN ( Wang et al., 2019 )
ODMN ( Li et al., 2022 )
perCLTV ( Zhao et al., 2023 )
OptDist ( Weng et al., 2024 )
expLTV ( Zhang et al., 2023 )
MEAN ( Liu et al., 2025 )
MDAN ( Liu et al., 2024 )
CoGSeq-LTV
E-commerce
7
MAE
743.80
371.04
665.44
536.84
281.92
322.59
279.35
274.68
WAPE
2.1753
1.2750
1.8642
1.7711
1.1518
1.1920
1.1469
1.1481
14
MAE
3694.30
1572.14
3008.71
2290.44
1244.80
1470.11
1233.92
1224.73
WAPE
2.2583
1.3417
1.8621
1.8933
1.1627
1.2114
1.1694
1.1586
30
MAE
5965.29
2747.23
5275.38
3914.75
2111.36
2384.61
2105.97
2108.64
Appendix
Table 5. Results on the low-frequency user subset.
Business
Horizon
Metric
Model
ZILN ( Wang et al., 2019 )
ODMN ( Li et al., 2022 )
perCLTV ( Zhao et al., 2023 )
OptDist ( Weng et al., 2024 )
expLTV ( Zhang et al., 2023 )
MEAN ( Liu et al., 2025 )
MDAN ( Liu et al., 2024 )
CoGSeq-LTV
E-commerce
7
MAE
807.32
417.75
738.48
619.84
331.26
365.04
326.88
321.74
WAPE
2.1671
1.2729
1.8512
1.7610
1.0614
1.1793
1.0558
1.0496
14
MAE
4004.68
1778.64
3355.18
2640.71
1468.30
1671.70
1449.62
1436.18
WAPE
2.2413
1.3445
1.8421
1.8895
1.0705
1.2029
1.0667
1.0601
30
MAE
6498.13
3103.75
5911.41
4473.83
2472.15
2724.86
2490.36
2458.27
Appendix
Table 6. Results on the high-frequency user subset.
Business
Horizon
Metric
Model
ZILN ( Wang et al., 2019 )
ODMN ( Li et al., 2022 )
perCLTV ( Zhao et al., 2023 )
OptDist ( Weng et al., 2024 )
expLTV ( Zhang et al., 2023 )
MEAN ( Liu et al., 2025 )
MDAN ( Liu et al., 2024 )
CoGSeq-LTV
E-commerce
7
MAE
765.23
385.12
685.41
561.38
532.14
338.45
345.67
295.12
WAPE
2.0541
1.1895
1.7542
1.6451
1.5782
1.1024
1.0543
0.9652
14
MAE
3785.14
1635.21
3105.42
2385.16
2265.41
1545.62
1485.67
1315.48
WAPE
2.1245
1.2564
1.7456
1.7652
1.6745
1.1245
1.0856
0.9754
30
MAE
6124.53
2845.61
5456.21
4085.16
3885.42
2512.38
2545.12
2235.64
Appendix
Table 7. Robustness study: performance under temporal distribution shift (training on Jul–Oct, testing on Feb).