Payment platforms use view-specific models trained on transaction attributes, behavioral histories, and relational structures to rank transactions under limited review budgets. However, existing fusion methods often focus on prediction scores or interactions among model representations, with less attention to how these interactions should guide aggregation using each model's original prediction and input coverage. To address this limitation, we propose InterFuse, a fusion framework that uses a shared fusion state from cross-model interaction to guide transaction-adaptive aggregation. InterFuse encodes each model's representation, prediction score, and input coverage as a model-specific token, using learned states for unavailable outputs. Availability-aware attention updates these tokens and produces a shared fusion state. This state then guides aggregation over available models using their updated representations, original predictions, and input coverage. We evaluate InterFuse on an industrial payment replay and three public financial benchmarks using the same upstream models across fusion methods. InterFuse achieves the highest top-ranked recall and AUC on all four datasets, improving Recall@1% by 1.11-8.09 percentage points over the strongest baselines. Ablations show the benefits of cross-model interaction and adaptive aggregation, while industrial analyses find larger interaction gains when model predictions disagree. Additional analyses examine robustness to missing outputs and the trade-off between ranking gains and fusion latency.
Figures & tables
Figure 1. Risk views and model fusion. InterFuse uses cross-model interaction to guide adaptive aggregation. Three vertically stacked sections. At the top, an illustrative transaction table on the left shows repeated one-dollar charges followed by a 980-dollar payment. A sequence on the upper right shows event order and timing; a graph below it shows shared devices and risk-linked accounts. The middle section illustrates score-level fusion of tabular, sequence, and graph predictions. The bottom section shows InterFuse updating model representations through cross-model interaction and producing a shared fusion state. Updated representations enter adaptive aggregation, while the shared fusion state guides the aggregation weights. Dashed paths preserve direct access to original scores and input coverage. The table values are illustrative, not experimental observations.
Figure 2. InterFuse architecture. Frozen model outputs are encoded as model tokens, updated through availability-aware attention, and adaptively aggregated for risk prediction. Card and graph outputs are unavailable in this example. Four-panel architecture diagram. Five frozen models feed model-identified tokens. Multi-head attention uses an availability mask: the fusion token reads all states, other model tokens cannot read unavailable keys, and self-connections are retained. Available model-specific representations are weighted, summed, and concatenated with the fusion state to predict a risk-ranking score.
Industrial ( Δ pp)
TransXion (%)
IBM CCT (%)
MBD (%)
Method
AUC
AP
R@1
AUC
AP
R@1
AUC
AP
R@1
AUC
AP
R@1
B1
+0.00
+0.00
+0.00
82.47
13.21
39.47
94.06
3.06
34.01
68.20
0.396
3.18
B5
+0.85 ± 0.11
+0.37 ± 0.35
+1.77 ± 0.21
93.69 ± 0.07
18.85 ± 0.41
60.31 ± 0.52
95.33 ± 0.07
4.09 ± 0.25
36.51 ± 0.46
57.62 ± 0.42
0.308 ± 0.009
2.63 ± 0.17
Concat
+1.78 ± 0.17
+2.67 ± 0.22
+1.99 ± 0.44
94.30 ± 0.07
18.76 ± 0.37
62.64 ± 0.62
94.69 ± 0.40
4.69 ± 1.07
36.25 ± 1.75
72.84 ± 0.81
0.647 ± 0.024
6.20 ± 0.14
Sum-pooling
-0.67 ± 0.13
+1.37 ± 0.31
+1.81 ± 0.21
94.60 ± 0.08
19.26 ± 0.65
62.97 ± 0.96
94.62 ± 0.13
4.54 ± 0.28
36.07 ± 0.35
71.55 ± 0.73
0.605 ± 0.032
5.97 ± 0.24
CE
-0.53 ± 0.14
+1.19 ± 0.48
+1.71 ± 0.71
94.83 ± 0.06
18.89 ± 0.82
62.75 ± 1.25
94.85 ± 0.30
4.74 ± 0.61
36.84 ± 1.39
72.38 ± 0.39
0.530 ± 0.013
4.48 ± 0.43
Table 1. Overall ranking performance (mean ± standard deviation over five fusion-training seeds). Industrial reports gains over fixed B1 (pp); public datasets report percentages. B1 is fixed across seeds. R@1 denotes Recall@1%; MBD averages four targets. Bold and underline mark the best and second-best means.
Figure 3. Ranking across review budgets. Shading shows the gap to the best plotted baseline at each budget; labels give 1% gains. Industrial: pp over B1; others: recall (%). A single-column two-by-two layout retains seven method curves per dataset. InterFuse is a solid burgundy line matching Figure 5. Light-blue shading shows where it exceeds the best plotted baseline at each budget; warm shading shows negative gaps. Dotted lines and gain labels mark the one-percent budget. UMI appears on Industrial, TransXion, and IBM CCT, and BALM on MBD.
Variant
Ind.
TX
CCT
MBD
InterFuse (Full / Both)
+5.44
89.69
61.46
10.20
Input information
w/o representation h
+0.15
88.14
56.94
9.33
w/o score s
+2.59
89.02
59.18
9.76
w/o coverage c
+2.71
89.14
59.48
9.88
Cross-model interaction
Table 2. Component analysis, Recall@1%. Industrial: gains over B1 (pp); public datasets: percentages. Full / Both is the complete configuration.
Figure 4. Full’s recall gains over (a) w/o coverage and Static across coverage groups and (b) C1 and Static across disagreement groups. Boxes show medians and interquartile ranges; whiskers follow the 1.5-IQR rule. (c) Full versus C1 review outcomes in the high-disagreement group. Violin and box plots of component recall gains across coverage and prediction-disagreement groups, with a table comparing recovered positives, reviewed negatives, and total reviews in the high-disagreement group.
Method
Params.
p50
p95
Δ R@1
(K)
(ms)
(ms)
(pp)
Concat
454.66
1.147
1.323
+1.99
Sum-pooling
442.26
1.087
1.167
+1.81
CE
410.11
11.696
11.989
+1.71
MMT
508.42
4.454
4.622
+1.96
UMI
294.43
1.314
1.334
+3.72
Table 3. Industrial fusion cost and Recall@1% gains over B1. Batch-1 latency excludes upstream encoders.
Figure 5. Industrial output removal. (a) Complete removal of each optional model. (b) Increasing joint removal. Values are Recall@1% gains over clean B1 (pp). A five-method by four-model matrix of complete single-model removal results, followed by five joint-removal curves with labeled endpoints.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Dataset
Task / unit
Train
Validation
Test
Positive rate
View-specific models
Industrial
Payment risk / event
5,832,600
1,392,465
1,366,237
0.576%
Tabular; user, card, device sequences; graph
TransXion
AML / event
1,817,503
605,834
605,833
0.149%
Tabular; sender, receiver sequences; graph
IBM CCT
Card fraud / event
19,512,454
2,436,496
2,437,950
0.111%
Tabular; user, card sequences; graph
MBD (trx/geo)
Product purchase / client-month
801,792
200,448
200,448
0.040–0.245%
Profile; transaction and geolocation summaries
Appendix
Table 4. Tasks, data splits, and view-specific models. Public Train counts include model and fusion fitting. Positive rates refer to Test; the range for MBD covers its four product-purchase targets.
Rank aggregation combines ranked lists or pairwise preferences into a consensus ordering for tasks such as search, recommendation, and model evaluation.
Pathology foundation models (PFMs) provide strong tile-level representations via self-supervised pre-training on large-scale pathology images. Yet, PFMs are developed under diverse and often opaque data, architecture, and objective choices, inducing latent representational biases that limit robustness and obscure what each model specialises in. We present AdaFusion, a lightweight adaptive fusion framework that integrates complementary signals from multiple frozen PFMs through (1) low-dimensional feature compression and (2) a sample-conditioned gating module that reweights model-wise (and optionally channel-wise) contributions. Beyond improving predictive accuracy, AdaFusion provides contribution-driven interpretation that offers evidence consistent with model-specific preferences and synergistic interactions across tissue phenotypes. We evaluate AdaFusion on three public benchmarks spanning treatment response prediction, prostate cancer grading, and spatial gene expression inference. AdaFusion consistently outperforms individual PFMs and other fusion baselines, while providing interpretable tissue visualisation which aligns model preferences with morphological patterns. Code is available at: https://github.com/xyx-98/PathoOracle.
Yuxiang Xiao, Yang Hu, Bin Li +5
School of Computer Science and Engineering, South China University of Technology, China · School of Computing and Mathematical Sciences, University of Leicester, UK · Leicester Cancer Research Centre, University of Leicester, UK +4
Credit-card fraud detection is difficult because fraudulent transactions are rare, costly, and unevenly distributed. Strong gradient-boosted tree models already perform well on structured transaction data, so the value of another fusion method is not obvious. This paper examines whether Combinatorial Fusion Analysis (CFA), which searches over model subsets and rank-score fusion rules, can still add value on the IEEE-CIS Fraud Detection benchmark. Using a leakage-free 60/20/20 train/validation/test protocol, we evaluate 480 fusion configurations built from seven base classifiers. The best test-set result comes from diversity-weighted score fusion of Random Forest, XGBoost, and LightGBM (DEF WtScore), with AUC-ROC = 0.9405, AUPRC = 0.6699, and F1 = 0.6373. Bootstrap confidence intervals from 1,000 resamples show that the gains over the strongest single model exclude zero for all three metrics. CFA matches soft voting on AUC-ROC, improves AUPRC and F1, and outperforms stacking in this setting. A CTGAN augmentation experiment gives a negative result: synthetic fraud samples degrade both individual models and CFA. Overall, CFA is most useful here not as a way to combine every classifier, but as a validation-stage method for choosing a small, complementary subset and assigning diversity-aware weights.
Xiao Han, Chenyu Wu
Goizueta Business School · Pratt School of Engineering
Model merging has shown that multitask models can be created by directly combining the parameters of different models that are each specialized on tasks of interest. However, models trained independently on distinct tasks often exhibit interference that degrades the merged model's performance. To solve this problem, we formally define the notion of Cross-Task Interference as the drift in the representation of the merged model relative to its constituent models. Reducing cross-task interference is key to improving merging performance. To address this issue, we propose our method, Resolving Interference (RI), a light-weight adaptation framework which disentangles expert models to be functionally orthogonal to the space of other tasks, thereby reducing cross-task interference. RI does this whilst using only unlabeled auxiliary data as input (i.e., no task-data is needed), allowing it to be applied in data-scarce scenarios. RI consistently improves the performance of state-of-the-art merging methods by up to 3.8% and generalization to unseen domains by up to 2.3%. We also find RI to be robust to the source of auxiliary input while being significantly less sensitive to tuning of merging hyperparameters. Our codebase is available at: https://github.com/pramesh39/resolving_interference
Pratik Ramesh, George Stoica, Arun Iyer +2
Georgia Tech · University of Washington · Microsoft Research +3