Publications

Citation counts are on my Google Scholar profile.

2024

P. Loyola, K. Hasegawa, A. Hoyos-Idrobo, K. Ono, T. Suzumura, Y. Hirate, M. Yamaoka. Annealing machine-assisted learning of graph neural network for combinatorial optimization. NeurIPS 2024 Workshop on Machine Learning with New Compute Paradigms, 2024.
Partial solutions from an annealing machine, computed on a compressed version of the graph, guide a graph neural network so that it can solve combinatorial problems larger than the machine itself can handle.

combinatorial optimizationgraph neural networksannealing machines
BibTeX
```
@inproceedings{loyola2024annealing,
  author        = {Loyola, Pablo and Hasegawa, Kento and Hoyos-Idrobo, Andrés and Ono, Kazuo and Suzumura, Toyotaro and Hirate, Yu and Yamaoka, Masanao},
  title         = {Annealing Machine-assisted Learning of Graph Neural Network for Combinatorial Optimization},
  booktitle     = {NeurIPS 2024 Workshop on Machine Learning with New Compute Paradigms},
  year          = {2024},
  eprint        = {2501.05845},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}
```

A. Abraham, A. Hoyos-Idrobo. Improving bias correction standards by quantifying its effects on treatment outcomes. ECML-PKDD 2024 Workshops, Communications in Computer and Information Science, 189–204, 2024. [arXiv]

A2A scores propensity score matching methods on artificial tasks with known outcomes, reducing treatment-effect estimation error by up to 50% and its variability by up to 90%.

causal inferencepropensity score matchingtreatment effectsselection bias

BibTeX

@inproceedings{abraham2026improving,
  author    = {Abraham, Alexandre and Hoyos-Idrobo, Andrés},
  title     = {Improving Bias Correction Standards by Quantifying Its Effects on Treatment Outcomes},
  booktitle = {Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2024 Workshops), Communications in Computer and Information Science},
  pages     = {189--204},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  doi       = {10.1007/978-3-032-25305-7_14}
}

2023

P. Loyola, E. Marrese-Taylor, A. Hoyos-Idrobo. Perceptual structure in the absence of grounding: the impact of abstractedness and subjectivity in color language for LLMs. Findings of the Association for Computational Linguistics: EMNLP 2023, 1536–1542, 2023. [arXiv]
Language models' embeddings of color descriptions align with perceptual color space for simple color names, but much less for subjective or abstract descriptions, suggesting these need grounding beyond text.

large language modelsgroundingcolor languagerepresentation alignment
BibTeX
```
@inproceedings{loyola2023perceptual,
  author    = {Loyola, Pablo and Marrese-Taylor, Edison and Hoyos-Idrobo, Andrés},
  title     = {Perceptual Structure in the absence of grounding: the impact of abstractedness and subjectivity in color language for {LLMs}},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2023},
  pages     = {1536--1542},
  year      = {2023},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/2023.findings-emnlp.102}
}
```
A. Hoyos-Idrobo. Learning to re-rank with constrained meta-optimal transport. Proceedings of the 46th International ACM SIGIR Conference (SIGIR 2023), 48–57, 2023. [arXiv]
A single network, shared across queries, predicts fair stochastic ranking policies for any query, unseen ones included, replacing slow per-query optimization while matching its fairness and ranking quality.

learning to rankoptimal transportfairness of exposurere-ranking
BibTeX
```
@inproceedings{hoyosidrobo2023learning,
  author    = {Hoyos-Idrobo, Andrés},
  title     = {Learning to Re-rank with Constrained Meta-Optimal Transport},
  booktitle = {Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  pages     = {48--57},
  year      = {2023},
  publisher = {ACM},
  doi       = {10.1145/3539618.3591714}
}
```

2022

P. Goswami, A. Hoyos-Idrobo, T. Deschamps, N. Feuillet. From theory to practice: testing harvesting-based propensity estimators on counterfactual learning-to-rank from implicit feedback for e-commerce search. SIGIR Workshop on eCommerce (SIGIR eCom 2022), 2022.
On Rakuten Ichiba product search, correcting click data for position bias only helps when paired with the right optimizer, and harvesting-based propensity estimators are unreliable where data is scarce.

counterfactual learning to rankposition biase-commerce searchimplicit feedback
BibTeX
```
@inproceedings{goswami2022theory,
  author    = {Goswami, Parantapa and Hoyos-Idrobo, Andrés and Deschamps, Théo and Feuillet, Nicolas},
  title     = {From Theory to Practice: Testing Harvesting-Based Propensity Estimators on Counterfactual Learning-To-Rank from Implicit Feedback for E-commerce Search},
  booktitle = {SIGIR Workshop on eCommerce (SIGIR eCom)},
  year      = {2022},
  url       = {https://sigir-ecom.github.io/ecom22Papers/paper_1119.pdf}
}
```

2020

A. Hoyos-Idrobo. Aligning hyperbolic representations: an optimal transport-based approach. arXiv preprint arXiv:2012.01089, 2020.
Optimal transport between embeddings in the Poincaré ball extends Euclidean domain-adaptation methods to align hierarchical representations, such as ontologies or cross-lingual vocabularies.

optimal transporthyperbolic embeddingsdomain adaptationalignment
BibTeX
```
@misc{hoyosidrobo2020aligning,
  author        = {Hoyos-Idrobo, Andrés},
  title         = {Aligning Hyperbolic Representations: an Optimal Transport-based approach},
  year          = {2020},
  eprint        = {2012.01089},
  archivePrefix = {arXiv},
  primaryClass  = {stat.ML}
}
```
A. Hoyos-Idrobo. Approximate Birkhoff-von-Neumann decomposition: a differentiable approach. OpenReview preprint, 2020.
A differentiable approximation of the Birkhoff-von-Neumann decomposition writes a doubly stochastic matrix as a mixture of permutations that can be learned with gradient-based methods.

Birkhoff-von-Neumann decompositionpermutationsdifferentiable optimizationranking
BibTeX
```
@misc{hoyosidrobo2020approximate,
  author = {Hoyos-Idrobo, Andrés},
  title  = {Approximate {Birkhoff}-von-Neumann decomposition: a differentiable approach},
  year   = {2020},
  url    = {https://openreview.net/forum?id=IpsTSvfIB6},
  note   = {OpenReview preprint}
}
```

2019

A. Hoyos-Idrobo. Local Bures-Wasserstein transport: a practical and fast mapping approximation. arXiv preprint arXiv:1906.08227, 2019.
Matching the Gaussian components of two densities and using the closed-form Bures-Wasserstein map between each pair gives a transport map that applies to new samples and is about 80 times faster to compute than existing estimators.

optimal transportWasserstein barycenterGaussian mixturestransport maps
BibTeX
```
@misc{hoyosidrobo2019local,
  author        = {Hoyos-Idrobo, Andrés},
  title         = {Local {Bures}-{Wasserstein} Transport: A Practical and Fast Mapping Approximation},
  year          = {2019},
  eprint        = {1906.08227},
  archivePrefix = {arXiv},
  primaryClass  = {stat.ML}
}
```

2018

A. Hoyos-Idrobo, G. Varoquaux, Y. Schwartz, B. Thirion. FReM – scalable and stable decoding with fast regularized ensemble of models. NeuroImage 180, 160–172, 2018. [nilearn]
Combining fast voxel clustering with an ensemble of cross-validated models gives brain decoders with more stable maps and predictions, at a fraction of the cost of spatially regularized methods.

brain decodingfMRIensemble learningclustering
BibTeX
```
@article{hoyosidrobo2018frem,
  author    = {Hoyos-Idrobo, Andrés and Varoquaux, Gaël and Schwartz, Yannick and Thirion, Bertrand},
  title     = {{FReM} – Scalable and stable decoding with fast regularized ensemble of models},
  journal   = {NeuroImage},
  volume    = {180},
  pages     = {160--172},
  year      = {2018},
  publisher = {Elsevier BV},
  doi       = {10.1016/j.neuroimage.2017.10.005}
}
```

A. Hoyos-Idrobo, G. Varoquaux, J. Kahn, B. Thirion. Recursive nearest agglomeration (ReNA): fast clustering for approximation of structured signals. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(3), 669–681, 2018. [arXiv] [code] [nilearn]

Repeatedly merging each cluster with its nearest neighbor gives a linear-time clustering that compresses images as well as quadratic-time methods while also removing noise.

clusteringdimension reductionneuroimagingscalability

BibTeX

@article{hoyosidrobo2019recursive,
  author    = {Hoyos-Idrobo, Andrés and Varoquaux, Gael and Kahn, Jonas and Thirion, Bertrand},
  title     = {Recursive Nearest Agglomeration ({ReNA}): Fast Clustering for Approximation of Structured Signals},
  journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume    = {41},
  number    = {3},
  pages     = {669--681},
  year      = {2019},
  publisher = {Institute of Electrical and Electronics Engineers (IEEE)},
  doi       = {10.1109/tpami.2018.2815524}
}

2017

M. Eickenberg, A. Frau-Pascual, A. Hoyos-Idrobo. Gaussian processes for HRF estimation for BOLD fMRI. arXiv preprint arXiv:1708.09762, 2017.
Modeling the hemodynamic response as a Gaussian process inside the fMRI general linear model estimates its shape with uncertainty and stays accurate when it departs from the standard response.

fMRIhemodynamic response functionGaussian processes
BibTeX
```
@misc{eickenberg2017gaussian,
  author        = {Eickenberg, Michael and Frau-Pascual, Aina and Hoyos-Idrobo, Andrés},
  title         = {{Gaussian} Processes for {HRF} estimation for {BOLD} {fMRI}},
  year          = {2017},
  eprint        = {1708.09762},
  archivePrefix = {arXiv},
  primaryClass  = {stat.AP}
}
```
A. Hoyos-Idrobo, G. Varoquaux, B. Thirion. Towards a faster randomized parcellation based inference. International Workshop on Pattern Recognition in Neuroimaging (PRNI 2017), 2017. [HAL]
Building the random parcellations with the fast ReNA clustering makes randomized parcellation-based inference much faster while keeping its recovery performance.

neuroimagingstatistical inferenceparcellationclustering
BibTeX
```
@inproceedings{hoyosidrobo2017faster,
  author    = {Hoyos-Idrobo, Andrés and Varoquaux, Gael and Thirion, Bertrand},
  title     = {Towards a faster randomized parcellation based inference},
  booktitle = {2017 International Workshop on Pattern Recognition in Neuroimaging (PRNI)},
  pages     = {1--4},
  year      = {2017},
  publisher = {IEEE},
  doi       = {10.1109/prni.2017.7981503}
}
```

G. Varoquaux, P. Raamana, D. Engemann, A. Hoyos-Idrobo, Y. Schwartz, B. Thirion. Assessing and tuning brain decoders: cross-validation, caveats, and guidelines. NeuroImage 145, 166–179, 2017. [arXiv]

Cross-validation of brain decoders has large error bars (about 10%), and repeated random splits should replace the unstable leave-one-out strategy.

cross-validationbrain decodingmodel evaluationneuroimaging

BibTeX

@article{varoquaux2017assessing,
  author    = {Varoquaux, Gaël and Raamana, Pradeep Reddy and Engemann, Denis A. and Hoyos-Idrobo, Andrés and Schwartz, Yannick and Thirion, Bertrand},
  title     = {Assessing and tuning brain decoders: Cross-validation, caveats, and guidelines},
  journal   = {NeuroImage},
  volume    = {145},
  pages     = {166--179},
  year      = {2017},
  publisher = {Elsevier BV},
  doi       = {10.1016/j.neuroimage.2016.10.038}
}

2016

A. Hoyos-Idrobo, G. Varoquaux, B. Thirion. Fast brain decoding with random sampling and random projections. International Workshop on Pattern Recognition in Neuroimaging (PRNI 2016), 2016. [HAL]
Randomly sampling voxels before decoding cuts computation time while keeping weight maps consistent with those learned on the full data.

brain decodingrandom projectionsrandom samplingdimension reduction
BibTeX
```
@inproceedings{hoyosidrobo2016fast,
  author    = {Hoyos-Idrobo, Andrés and Varoquaux, Gael and Thirion, Bertrand},
  title     = {Fast brain decoding with random sampling and random projections},
  booktitle = {2016 International Workshop on Pattern Recognition in Neuroimaging (PRNI)},
  pages     = {1--4},
  year      = {2016},
  publisher = {IEEE},
  doi       = {10.1109/prni.2016.7552350}
}
```

2015

A. Hoyos-Idrobo, Y. Schwartz, G. Varoquaux, B. Thirion. Improving sparse recovery on structured images with bagged clustering. International Workshop on Pattern Recognition in Neuroimaging (PRNI 2015), 73–76, 2015. [HAL]

Averaging sparse models fitted on several clusterings of resampled data stabilizes the brain regions they select.

sparse modelsbaggingclusteringneuroimaging

BibTeX

@inproceedings{hoyosidrobo2015improving,
  author    = {Hoyos-Idrobo, Andrés and Schwartz, Yannick and Varoquaux, Gael and Thirion, Bertrand},
  title     = {Improving Sparse Recovery on Structured Images with Bagged Clustering},
  booktitle = {2015 International Workshop on Pattern Recognition in NeuroImaging},
  pages     = {73--76},
  year      = {2015},
  publisher = {IEEE},
  doi       = {10.1109/prni.2015.30}
}

B. Thirion, A. Hoyos-Idrobo, J. Kahn, G. Varoquaux. Fast clustering for scalable statistical analysis on structured images. ICML Workshop on Statistics, Machine Learning and Neuroscience (Stamlins 2015), 2015.
Compressing brain images with a linear-time clustering of neighboring voxels, rather than random projections, preserves their spatial structure and removes noise, making large-scale analyses both faster and more accurate.

clusteringdimension reductionneuroimagingbig data
BibTeX
```
@inproceedings{thirion2015fast,
  author        = {Thirion, Bertrand and Hoyos-Idrobo, Andrés and Kahn, Jonas and Varoquaux, Gael},
  title         = {Fast clustering for scalable statistical analysis on structured images},
  booktitle     = {ICML Workshop on Statistics, Machine Learning and Neuroscience (Stamlins)},
  year          = {2015},
  eprint        = {1511.04898},
  archivePrefix = {arXiv},
  primaryClass  = {stat.ML}
}
```

2014

A. Hoyos-Idrobo, P. Weiss, A. Massire, A. Amadon, N. Boulant. On variant strategies to solve the magnitude least squares optimization problem in parallel transmission pulse design and under strict SAR and power constraints. IEEE Transactions on Medical Imaging 33(3), 739–748, 2014. [arXiv]

Constrained optimization methods design parallel-transmission MRI pulses at 7 T that strictly respect patient-safety (SAR) and power limits, within seconds to a minute.

MRIparallel transmissionRF pulse designconstrained optimization

BibTeX

@article{hoyosidrobo2014variant,
  author    = {Hoyos-Idrobo, Andrés and Weiss, P. and Massire, A. and Amadon, A. and Boulant, N.},
  title     = {On Variant Strategies to Solve the Magnitude Least Squares Optimization Problem in Parallel Transmission Pulse Design and Under Strict {SAR} and Power Constraints},
  journal   = {IEEE Transactions on Medical Imaging},
  volume    = {33},
  number    = {3},
  pages     = {739--748},
  year      = {2014},
  publisher = {Institute of Electrical and Electronics Engineers (IEEE)},
  doi       = {10.1109/tmi.2013.2295465}
}

PhD thesis

A. Hoyos-Idrobo. Ensembles of models in fMRI: stable learning in large-scale settings. Université Paris-Saclay, 2017.

BibTeX

@phdthesis{hoyosidrobo2017ensembles,
  author = {Hoyos-Idrobo, Andrés},
  title  = {Ensembles of models in {fMRI}: stable learning in large-scale settings},
  school = {Université Paris-Saclay},
  year   = {2017},
  url    = {https://theses.hal.science/tel-01526693}
}