Research programme

Cross-Silo Federated Learning under Data Heterogeneity: Aggregation Weighting and Local Adaptation

The Case of Probabilistic Forecasting of Customer Collections for Group Cash Management

Research in progressBy Fitahiantsoa Finaritra RAZAKANARY

Protocol pre-registered before any performance is measured; no results are published at this stage.

  • cross-silo federated learning
  • data heterogeneity (non-IID)
  • aggregation weighting
  • local adaptation
  • negative transfer
  • probabilistic forecasting
  • calibration
  • discrete-time survival
Author
Fitahiantsoa Finaritra RAZAKANARY — AI & Automation Project Lead, Madecasse Group
Affiliation
MSc Big Data and Artificial Intelligence — ESTIA • IT University
Study setting
Madecasse Group — a federation of 8 group companies
Field
Applied machine learning: federated learning, probabilistic forecasting, calibration
Status
Research in progress

Abstract

When several organisations train a model together without pooling their data, the heterogeneity of that data can make collaboration helpful to some participants and harmful to others. This applied research programme aims to isolate the effect of two choices in cross-silo federated learning: how participants are weighted during aggregation (by data volume or uniformly) and how deeply the shared model is adapted locally (recalibration, local head, full network). The case studied is probabilistic forecasting of customer collections, the most uncertain component of short-term cash management, across eight companies of Madecasse Group. A single discrete-time hazard network estimates, for each outstanding invoice, the probability of being paid within 7, 14, 21 and 28 days; these probabilities are aggregated into a company-level forecast with an 80% interval. A factorial design compares local, centralised and federated learning, with or without adaptation, against simple baselines and gradient boosting, under an equal tuning budget. The five hypotheses, the metrics (28-day log-loss, calibration, interval coverage) and the decision criteria are pre-registered before any performance is measured. The research is in progress: no model has been trained yet, and an absence of gain would be a valid result.

In plain words

Each company in the group wants to know which invoices its customers will pay over the next four weeks, and how much cash will reach the bank. Some have a long history, others very little, and their customers do not pay in the same way. Learning together can help the smaller ones, but it can also blur each company’s own habits. Federated learning makes it possible to learn together without pooling the data: a shared model “travels” from company to company; each adjusts it in-house and sends back only its settings, never its invoices or customers. Picture doctors from several hospitals improving a care protocol together by sharing what they have learned, without ever exchanging their patients’ files. The research measures two things: how much weight each company should carry in the average, and how far the shared model should then be readjusted for each one. The evaluation rules are sealed before any result is seen.

Why this work

01

Governance

The companies’ data currently sit in a single database, so federation is not an immediate technical necessity. But the companies are separate legal entities, responsible for their customers’ data, some of whom may be natural persons protected by Malagasy Law No. 2014-038. Pooling detailed data cannot be taken for granted, especially with respect to shareholders or external partners.

02

Architecture

The information system will keep changing. A learning scheme that does not require detailed data to be centralised remains applicable if the databases are split or if outside companies join the arrangement.

03

Method

Comparing local (L), centralised (C) and federated (F) learning measures what the ban on pooling data costs (H5) and what collaboration brings when pooling is not allowed (H1). The centralised model is therefore the natural comparator; concluding that “centralisation with local adaptation is enough” would be a valid result, reported as such.

Research question

To what extent can aggregation and local-adaptation strategies in federated learning improve the accuracy and calibration of probabilistic collection forecasts — the uncertain component of cash management — compared with local or centralised learning, when companies have heterogeneous data volumes and payment behaviours?

Sub-questions

  1. SQ1Does sharing mainly benefit companies with little data, including a company that joins the group?
  2. SQ2Does uniform weighting reduce the dominance of large companies, and at what cost?
  3. SQ3What depth of local adaptation corrects negative transfer without overfitting?
  4. SQ4Does calibration remain adequate within each company?
  5. SQ5What does the federated constraint cost compared with pooling the data?

Hypotheses

H1

Value of sharing

The gain from sharing increases as local history shrinks.

Test
“New company” ablation: for three companies, local history is artificially reduced to 1, 2 or 3 months, over the three test folds; the local model L is compared with Fv, Fu and F+A1.
Decision criterion
Supported if, with one month of history, F or F+A achieves a gain over L in at least 2 of the 3 target companies, and that gain decreases as history lengthens.
H2

Weighting

Volume weighting (Fv) favours overall performance; uniform weighting (Fu) favours fairness across companies.

Test
Fv versus Fu: row-weighted mean loss, worst gap relative to the local model, dispersion of scores across companies.
Decision criterion
Trade-off confirmed if Fv is better on the average and Fu better on the worst-off company. A weighting that wins on both counts refutes H2.
H3

Local adaptation

Local adaptation reduces negative transfer; the useful depth of adaptation increases with local data volume.

Test
Number of companies degraded relative to the local model L, for F, F+A0, F+A1 and F+A2.
Decision criterion
Supported if the best adaptation leaves fewer degraded companies than F alone, and no company is degraded by the adaptation itself.
H4

Calibration

A global model that is calibrated on average can be miscalibrated within a given company.

Test
Calibration slope and ECE per company for C, Fv and Fu, then after local recalibration A0; coverage of the 80% intervals.
Decision criterion
Supported if at least one company is out of tolerance while overall calibration is within it, and A0 brings at least 75% of companies back within tolerance (slope 0.8–1.25, ECE ≤ 0.03; proposed tolerances, confirmed at the freeze).
H5

Cost of federation

With identical adaptation, the gap between federated and centralised learning is small.

Test
F+A versus C+A, and F versus C, company by company.
Decision criterion
Supported if the gap is below the equivalence margin ε for at least 6 of the 8 companies; otherwise the cost of federation is quantified and discussed.

Experimental design

Factorial design: only one thing varies at a time (training regime, weighting, adaptation, available history), and everything else is identical across methods — architecture, features and transformations, evaluated population, folds, horizons, tuning budget and seeds. The core comprises thirteen network configurations (L; C and its three adaptations; Fv, Fu and their three adaptations), two simple baselines, two tree-based baselines and the due-date rule. Four experiments make up this core: E1, comparison of all methods over three folds and three seeds (H2, H3, H5); E2, the “new company” ablation (H1); E3, a transfer matrix in which each company’s local model is evaluated on the others; E4, calibration and intervals (H4). An invoice-level cash view (expected collections minus supplier invoices falling due) is planned as an option.

Compared arms

CodeArmDescription
LLocalOne model per company, trained on its own data only, with no sharing at all. Each company’s reference for measuring gains and negative transfer.
CCentralisedAll data pooled to train a single model: the case where pooling is allowed. The natural comparator for federated learning.
FFederated (FedAvg): Fv and FuIn each round, each company fits the shared model on its own premises (1, 2 or 5 epochs), then the coordinator takes the weighted average of the parameters, ws ∝ nsα: α = 1 for Fv (by volume), α = 0 for Fu (uniform). Only parameters, sample counts and aggregated losses leave a company.
C+ACentralised, then adaptedCentralised model subsequently adapted to each company (A0, A1 or A2). Separates the effect of personalisation from that of federation (H5).
F+AFederated, then adaptedFederated model recalibrated (A0), with its head relearned locally (A1) or with the whole network fine-tuned (A2). Measures the useful depth of adaptation (H3, H4).
B0Contractual due datesBusiness reference: the invoice is paid on its due date. Evaluated at company level only, since it does not yield a probability.
B1Historical ratesCollection rates by days-overdue bracket, company-specific (B1-L) or shared (B1-C). Simple but often hard to beat; B1-L is the base for skill scores.
GBDTGradient boostingLightGBM per company (GB-L) or shared with a company indicator (GB-C): a strong tabular baseline, so as not to conclude too quickly that a neural network is needed.
Ext.ExtensionsFederated with α = 0.5, FedProx and Ditto, if the schedule allows; exploratory analyses.

Factors and levels

FactorLevels
Training regimelocal (L) · centralised (C) · federated (F)
Aggregation weightingws ∝ nsα, where ns is the company’s number of training rows: α = 1 (Fv, by volume) · α = 0 (Fu, uniform) · α = 0.5 (exploratory)
Local adaptationnone · A0 recalibration (z′k = a zk + bk, a > 0: five numbers, invoice ranking preserved) · A1 local head (last layer relearned, rest frozen) · A2 full network (reduced learning rate, early stopping). Uses the company’s own data only: its last four closed origins, reserved for this purpose for all methods.
Available local history (ablation)full · reduced to 1, 2 or 3 months for three companies, to simulate a company joining the group

Model

Shared model: a discrete-time hazard network, from the family of discrete-time survival models — a compact multilayer perceptron (two hidden layers of 32 to 128 units, ReLU, dropout). Every Sunday (forecast origin), for each outstanding invoice less than one year old, it outputs four weekly hazards λk = σ(zk), which give the probabilities of being paid within 7, 14, 21 and 28 days, F(7k) = 1 − ∏j≤k (1 − λj), monotone by construction. The loss is cross-entropy with fractional labels, masked over horizons not yet closed: a proper scoring rule that handles partial payments without extra assumptions. The network contains neither a company identifier nor batch normalisation, whose company-specific statistics would disturb federated averaging; feature transformations are fixed in advance, with no statistic computed on the union of companies. At company level, the expected amount is the sum of outstanding balances weighted by these probabilities; 2,000 simulations, in which invoices from the same customer share a single draw, yield an 80% interval [q10; q90], corrected by local conformal calibration (CQR-type score). Federation is simulated on a single machine (explicit FedAvg loop in PyTorch, at most 200 rounds, Flower-compatible interface); it provides no formal privacy guarantee.

Metrics

Log-loss (cross-entropy) — primary criterion
With fractional labels, at 28 days (primary) and 7 days (confirmation). The main ranking rests on 28-day log-loss, as a simple average over the eight companies and three folds, reported with seed dispersion, intervals and per-company results.
Brier score and skill score
Squared difference between forecast probability and outcome; the skill score expresses the relative improvement over the B1-L baseline.
Calibration: slope, intercept and ECE
Reliability diagrams, calibration slope and intercept, and ECE over 10 equal-count bins, per company. Proposed tolerances, to be confirmed at the freeze: slope between 0.8 and 1.25; ECE ≤ 0.03.
Discrimination
ROC AUC and PR AUC, as secondary measures.
Aggregated intervals: coverage and width
Coverage of [q10; q90] (nominal 80%), relative width and simulation-based CRPS: a good interval contains the outcome about 8 times out of 10 while staying narrow. Nominal coverage must lie within the 95% binomial confidence interval of observed coverage.
Company-level aggregate errors
MAE, WAPE and signed bias, in amount and as a share of the outstanding balance: a business reading, kept out of the main ranking because it rests on fewer independent observations.
Heterogeneity and transfer
Per-company scores, simple and row-weighted averages, worst-off company and dispersion across companies; paired difference with the local model and counts of gains, degradations and inconclusive results. An average can hide a sacrificed company.
Cost
Number of rounds, volume exchanged, compute time and memory: a method that is barely better but far more expensive is not necessarily preferable.

Protocol

  1. Rolling evaluation: development and tuning on a development period (internal two-fold temporal cross-validation), then testing on three successive monthly folds. Before each fold, models are retrained on all closed origins with frozen settings, and the evaluation is run only once.
  2. Unit of analysis: for a given company, horizon and pair of methods, the mean loss difference computed on exactly the same invoices and weekly origins (paired comparison).
  3. 95% confidence intervals from a customer-cluster bootstrap (10,000 replicates, percentile interval): whole customers are resampled with replacement, because invoices from the same customer are not independent. As a sensitivity check: block bootstrap over four consecutive origins.
  4. Equivalence margin εs: the larger of 1% of the company’s local-model loss and twice the median between-seed standard deviation observed during development. Both the computation rule and the value are fixed before the test is opened.
  5. Decision rule: a gain if the interval is entirely negative, the effect exceeds ε and the sign is the same in at least two of the three folds; a degradation (negative transfer) in the symmetric case; otherwise, equivalence or an inconclusive result.
  6. Multiple comparisons: one pre-declared primary contrast per hypothesis, at 28 days, with a Holm correction across companies; any other comparison is exploratory.
  7. Fair comparison: the same tuning budget for every method family (20 configurations drawn at random from a shared search space); three seeds per stochastic method, raised to five if between-seed spread exceeds ε in more than two companies; the last four closed origins are reserved for early stopping, adaptation and calibration, for all methods.
  8. Small samples: a company whose interval is too wide is reported with its sample sizes, without an individual conclusion.
  9. Planned sensitivity analyses: strict definition of a collection (direct bank receipts only), full portfolio of all ages, test window extended with a reduced maturity margin; if the schedule allows: information as recorded at forecast time, alternative weightings, independence between a customer’s invoices in the simulation, inclusion of the company held out for robustness testing.
  10. Scope: conclusions apply to the companies and period evaluated. Organisational proximity within a single group and a short shared history, without a full annual cycle, allow neither generalisation to other firms nor identification of annual seasonality.

Integrity and ethics

Research integrity

  • Pre-registration: the protocol, configurations, metrics, decision criteria and evaluation code are frozen before any performance of a trained model is measured, and filed with a Git tag and a SHA-256 hash.
  • Sealed test: no performance measure is computed on the test window until hyperparameters and the margin ε are frozen; every run is logged.
  • Deviation tracking: any later change produces a new dated version of the protocol, stating the reason and the results before and after. A method modified after seeing the test is never presented as validated by that same test, and any added analysis is declared exploratory.
  • Frozen data: the study relies on a single data snapshot, with no new extraction; its SHA-256 hashes are checked at every load, so that an authorised third party can redo every computation identically.
  • Reproducibility: pinned dependencies, fixed seeds and an experiment log (configuration, hash, seed, duration, results). Re-running a reference configuration must recover the data, folds and scores to within 10−6, barring documented randomness; every figure in the thesis is linked to an identified run.
  • No leakage: features at an origin use only events dated on or before that origin, which a perturbation test on future data verifies. A maturity rule evaluates only the weeks whose payments have had time to be recorded.
  • Software quality: automated tests of the critical rules (balance reconstruction, definition of a collection, absence of leakage, monotone probabilities, isolation between companies), with at least 80% of the data-preparation code covered.

Ethics and data

  • Legal framework: Malagasy Law No. 2014-038 of 9 January 2015 on the protection of personal data; customer data may concern natural persons.
  • Authorisation: written consent from the sponsor (group management) for academic use of the data; a confidentiality clause if the thesis is to circulate beyond the examination board.
  • Minimisation: only the necessary accounting tables are read; no names, contact details or attachments are used; customer identifiers are pseudonymised (keyed HMAC, with the key held by the company).
  • Data location: detailed data stay on the company’s infrastructure, outside the code repository and the thesis. In the federated setting, each company accesses only its own folder and the coordinator receives only parameters, sample counts and aggregated losses.
  • Dissemination: results are published only as per-company aggregates, under anonymised company codes; no personally identifying data.
  • Acknowledged limitation: the federated simulation provides no formal privacy guarantee (neither differential privacy nor secure aggregation) and is not presented as a compliance measure.

Expected contributions

  • Provide a documented, pre-registered comparative protocol, applied to real heterogeneous data, that isolates the effect of aggregation weighting and that of local adaptation in cross-silo federated learning.
  • Measure, company by company rather than as a single average, the gains, degradations (negative transfer) and inconclusive results of each strategy.
  • Estimate what the federated constraint costs compared with pooling the data (H5), and what sharing brings to a company with a short history (H1).
  • Analyse per-company calibration and the reliability of aggregated intervals, to check whether calibration that is adequate on average masks local deviations (H4).
  • Deliver a reproducible prototype — an executable pipeline from the data snapshot to the result tables, automated tests, an experiment log — and a hand-over guide.
  • Draft a recommendations note for the group: use of the forecast in cash management, data-freshness requirements, choice between a centralised and a federated architecture.
  • No minimum gain is set in advance: a null result, a gain limited to some companies, or the conclusion that centralisation with local adaptation is enough, will be reported as such.

Positioning

The project is applied research in machine learning. It assumes neither the invention of a new algorithm nor the superiority of federated learning: no gain, or a gain limited to some companies, is a usable result provided the experimentation is rigorous. The author is also the group’s AI and automation project lead. This practitioner-researcher position gives access to the field and to a real decision — whether or not to pool learning across companies — but exposes the work to confirmation bias, the tendency to see what one hopes to see. This risk is handled explicitly: the protocol, metrics, decision criteria and evaluation code are frozen and timestamped before any performance of a trained model is measured; the comparators include strong baselines (gradient boosting, adapted centralised model) under an equal tuning budget; the test folds are opened only once settings are frozen; any analysis added afterwards is declared exploratory. As with an exam, the questions and marking scheme are sealed in an envelope before any paper is graded; here, the envelope is a cryptographic hash.

Glossary

Silo (cross-silo federated learning)
An organisation that keeps its data on its own premises. “Cross-silo” federated learning brings together a few organisations (hospitals, banks, companies), each holding a lot of data, as opposed to the “cross-device” setting (millions of phones). Here, each group company is a silo.
Heterogeneity (non-IID)
The participants’ data do not follow the same distribution: volumes, customers and payment habits differ from one company to another. This is the central difficulty of federated learning.
Federated averaging (FedAvg)
In each round, every silo fits the shared model on its own data; the coordinator takes the weighted average of the parameters received (weights ws ∝ nsα) and sends it back to all. The data never leave their silo.
Local adaptation (personalisation)
Fitting the shared model to a single company’s data, from lightest to strongest: recalibration of probabilities (A0), relearning the last layer (A1), careful fine-tuning of the whole network (A2).
Negative transfer
The case where learning with the other companies yields worse forecasts than learning alone, with one’s own local model.
Discrete hazard
The probability that an invoice is paid during a given week, given that it has not been paid before. Chained together, weekly hazards give a cumulative probability of payment that can only increase with the horizon.
Log-loss (cross-entropy)
An error score for probabilistic forecasts: the lower it is, the better the forecasts. It mainly penalises overconfidence and cannot be gamed: it is a proper scoring rule.
Calibration
Agreement between stated probabilities and observed frequencies: invoices forecast as “70% likely to be paid” should be paid about 7 times out of 10. The calibration slope equals 1 when probabilities have the right spread; below 1, the model is overconfident.
ECE (expected calibration error)
The average gap between stated probability and observed frequency, computed over equal-count bins of forecasts; 0.03 corresponds to 3 percentage points.
Interval coverage
The share of cases in which the amount actually collected falls within the stated interval. An “80%” interval should contain the outcome about 8 weeks out of 10, while staying as narrow as possible.
Equivalence margin ε
The smallest difference deemed useful in practice, set in advance. Below it, two methods are considered equivalent, even if a tiny difference exists.
Pre-registration
A dated freeze of the protocol, metrics, decision criteria and evaluation code before any performance is measured, evidenced by a Git tag and a SHA-256 hash: the rules can no longer be adjusted after seeing the results.

References

  1. Andreux, M., Manoel, A., Menuet, R., Saillard, C. & Simpson, C. (2020). Federated Survival Analysis with Discrete-Time Cox Models. arXiv:2006.08997. arxiv.org/abs/2006.08997 (new tab)
  2. Angelopoulos, A. N. & Bates, S. (2023). Conformal Prediction: A Gentle Introduction. Foundations and Trends in Machine Learning, 16(4), 494–591. arxiv.org/abs/2107.07511 (new tab)
  3. Arivazhagan, M. G., Aggarwal, V., Singh, A. K. & Choudhary, S. (2019). Federated Learning with Personalization Layers. arXiv:1912.00818. arxiv.org/abs/1912.00818 (new tab)
  4. Bergmeir, C. & Benítez, J. M. (2012). On the use of cross-validation for time series predictor evaluation. Information Sciences, 191, 192–213.
  5. Beutel, D. J. et al. (2020). Flower: A Friendly Federated Learning Research Framework. arXiv:2007.14390. arxiv.org/abs/2007.14390 (new tab)
  6. Bouthillier, X. et al. (2021). Accounting for Variance in Machine Learning Benchmarks. MLSys. arxiv.org/abs/2103.03098 (new tab)
  7. Cheng, G., Chadha, K. & Duchi, J. (2023). Federated Asymptotics: a model to compare federated learning algorithms. AISTATS, PMLR 206.
  8. Collins, L., Hassani, H., Mokhtari, A. & Shakkottai, S. (2021). Exploiting Shared Representations for Personalized Federated Learning. ICML, PMLR 139, 2089–2099. arxiv.org/abs/2102.07078 (new tab)
  9. Gensheimer, M. F. & Narasimhan, B. (2019). A scalable discrete-time survival model for neural networks. PeerJ, 7, e6257. doi.org/10.7717/peerj.6257 (new tab)
  10. Gneiting, T., Balabdaoui, F. & Raftery, A. E. (2007). Probabilistic forecasts, calibration and sharpness. Journal of the Royal Statistical Society B, 69(2), 243–268. doi.org/10.1111/j.1467-9868.2007.00587.x (new tab)
  11. Gneiting, T. & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477), 359–378. doi.org/10.1198/016214506000001437 (new tab)
  12. Grinsztajn, L., Oyallon, E. & Varoquaux, G. (2022). Why do tree-based models still outperform deep learning on typical tabular data? NeurIPS, Datasets and Benchmarks Track. arxiv.org/abs/2207.08815 (new tab)
  13. Guo, C., Pleiss, G., Sun, Y. & Weinberger, K. Q. (2017). On Calibration of Modern Neural Networks. ICML, PMLR 70, 1321–1330. arxiv.org/abs/1706.04599 (new tab)
  14. Humbert, P., Le Bars, B., Bellet, A. & Arlot, S. (2023). One-Shot Federated Conformal Prediction. ICML, PMLR 202.
  15. Hyndman, R. J. & Athanasopoulos, G. (2021). Forecasting: Principles and Practice, 3rd ed. OTexts.
  16. Kairouz, P., McMahan, H. B. et al. (2021). Advances and Open Problems in Federated Learning. Foundations and Trends in Machine Learning, 14(1–2), 1–210. arxiv.org/abs/1912.04977 (new tab)
  17. Karimireddy, S. P., Kale, S., Mohri, M., Reddi, S., Stich, S. & Suresh, A. T. (2020). SCAFFOLD: Stochastic Controlled Averaging for Federated Learning. ICML, PMLR 119. arxiv.org/abs/1910.06378 (new tab)
  18. Kaufman, S., Rosset, S., Perlich, C. & Stitelman, O. (2012). Leakage in data mining: Formulation, detection, and avoidance. ACM Transactions on Knowledge Discovery from Data, 6(4), article 15.
  19. Ke, G. et al. (2017). LightGBM: A Highly Efficient Gradient Boosting Decision Tree. NeurIPS.
  20. Künsch, H. R. (1989). The jackknife and the bootstrap for general stationary observations. The Annals of Statistics, 17(3), 1217–1241.
  21. Kvamme, H., Borgan, Ø. & Scheel, I. (2019). Time-to-Event Prediction with Neural Networks and Cox Regression. Journal of Machine Learning Research, 20(129), 1–30. arxiv.org/abs/1907.00825 (new tab)
  22. Li, Q., Diao, Y., Chen, Q. & He, B. (2022). Federated Learning on Non-IID Data Silos: An Experimental Study. IEEE ICDE, 965–978. arxiv.org/abs/2102.02079 (new tab)
  23. Li, T., Hu, S., Beirami, A. & Smith, V. (2021). Ditto: Fair and Robust Federated Learning Through Personalization. ICML, PMLR 139, 6357–6368. arxiv.org/abs/2012.04221 (new tab)
  24. Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A. & Smith, V. (2020). Federated Optimization in Heterogeneous Networks. MLSys. arxiv.org/abs/1812.06127 (new tab)
  25. Li, T., Sanjabi, M., Beirami, A. & Smith, V. (2020). Fair Resource Allocation in Federated Learning. ICLR. arxiv.org/abs/1905.10497 (new tab)
  26. Li, X., Jiang, M., Zhang, X., Kamp, M. & Dou, Q. (2021). FedBN: Federated Learning on Non-IID Features via Local Batch Normalization. ICLR. arxiv.org/abs/2102.07623 (new tab)
  27. Lu, C., Yu, Y., Karimireddy, S. P., Jordan, M. I. & Raskar, R. (2023). Federated Conformal Predictors for Distributed Uncertainty Quantification. ICML, PMLR 202.
  28. McMahan, B., Moore, E., Ramage, D., Hampson, S. & Agüera y Arcas, B. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS, PMLR 54, 1273–1282. arxiv.org/abs/1602.05629 (new tab)
  29. Mohri, M., Sivek, G. & Suresh, A. T. (2019). Agnostic Federated Learning. ICML, PMLR 97, 4615–4625. arxiv.org/abs/1902.00146 (new tab)
  30. Ogier du Terrail, J. et al. (2022). FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings. NeurIPS, Datasets and Benchmarks Track. arxiv.org/abs/2210.04620 (new tab)
  31. République de Madagascar. Loi n° 2014-038 du 9 janvier 2015 sur la protection des données à caractère personnel.
  32. Romano, Y., Patterson, E. & Candès, E. (2019). Conformalized Quantile Regression. NeurIPS. arxiv.org/abs/1905.03222 (new tab)
  33. Saini, S., Manai, G., van den Boom, W., De Iorio, M. & Qian, F. (2024). Invoice level forecasting with discrete survival methods for effective forecasting of account receivables in supply chain. Discover Analytics, 2, article 5.
  34. Tan, A. Z., Yu, H., Cui, L. & Yang, Q. (2023). Towards Personalized Federated Learning. IEEE Transactions on Neural Networks and Learning Systems, 34(12), 9587–9603. arxiv.org/abs/2103.00710 (new tab)
  35. Tashman, L. J. (2000). Out-of-sample tests of forecasting accuracy: an analysis and review. International Journal of Forecasting, 16(4), 437–450.
  36. Tutz, G. & Schmid, M. (2016). Modeling Discrete Time-to-Event Data. Springer.
  37. Wang, J., Charles, Z., Xu, Z. et al. (2021). A Field Guide to Federated Optimization. arXiv:2107.06917. arxiv.org/abs/2107.06917 (new tab)
  38. Wang, Z., Dai, Z., Póczos, B. & Carbonell, J. (2019). Characterizing and Avoiding Negative Transfer. CVPR, 11293–11302. arxiv.org/abs/1811.09751 (new tab)
  39. Weytjens, H., Lohmann, E. & Kleinsteuber, M. (2021). Cash flow prediction: MLP and LSTM compared to ARIMA and Prophet. Electronic Commerce Research, 21, 371–391.
  40. Yu, T., Bagdasaryan, E. & Shmatikov, V. (2020). Salvaging Federated Learning by Local Adaptation. arXiv:2002.04758. arxiv.org/abs/2002.04758 (new tab)
  41. Zeng, S., Melville, P., Lang, C. A., Boier-Martin, I. & Murphy, C. (2008). Using Predictive Analysis to Improve Invoice-to-Cash Collection. KDD, 1043–1050.

Cite this work

Work in progress, unpublished: please cite this page and the date you accessed it.

BibTeX
@mastersthesis{razakanary-fedcash,
  author = {Razakanary, Fitahiantsoa Finaritra},
  title  = {Apprentissage fédéré entre silos hétérogènes : pondération de l'agrégation et adaptation locale},
  school = {ESTIA --- IT University},
  type   = {Mémoire de MSc Big Data et Intelligence Artificielle},
  note   = {Recherche en cours. Programme FedCash},
}