András Benczúr

6603069526

Publications - 8

DISTRIBUTION-FREE DATA UNCERTAINTY FOR NEURAL NETWORK REGRESSION

Péter Kersch Domokos Kelen Ádám Jung András Benczúr

Publication Name: 13th International Conference on Learning Representations Iclr 2025

Publication Date: 2025-01-01

Volume: Unknown

Issue: Unknown

Page Range: 67459-67490

Description:

Quantifying uncertainty is an essential part of predictive modeling, especially in the context of high-stakes decision-making. While classification output includes data uncertainty by design in the form of class probabilities, the regression task generally aims only to predict the expected value of the target variable. Probabilistic extensions often assume parametric distributions around the expected value, optimizing the likelihood over the resulting explicit densities. However, using parametric distributions can limit practical applicability, making it difficult for models to capture skewed, multi-modal, or otherwise complex distributions. In this paper, we propose optimizing a novel nondeterministic neural network regression architecture for loss functions derived from a sample-based approximation of the continuous ranked probability score (CRPS), enabling a truly distribution-free approach by learning to sample from the target's aleatoric distribution, rather than predicting explicit densities. Our approach allows the model to learn well-calibrated, arbitrary uni- and multivariate output distributions. We evaluate the method on a variety of synthetic and real-world tasks, including uni- and multivariate problems, function inverse approximation, and standard regression uncertainty benchmarks. Finally, we make all experiment code publicly available.

Open Access: Yes

DOI: DOI not available

Constructing and sampling partite, 3-uniform hypergraphs with given degree sequence

András Hubai Tamás Róbert Mezei Ferenc Béres András Benczúr István Miklós

Publication Name: Plos One

Publication Date: 2024-05-01

Volume: 19

Issue: 5 May

Page Range: Unknown

Description:

Partite, 3-uniform hypergraphs are 3-uniform hypergraphs in which each hyperedge contains exactly one point from each of the 3 disjoint vertex classes. We consider the degree sequence problem of partite, 3-uniform hypergraphs, that is, to decide if such a hypergraph with prescribed degree sequences exists. We prove that this decision problem is NP-complete in general, and give a polynomial running time algorithm for third almost-regular degree sequences, that is, when each degree in one of the vertex classes is k or k − 1 for some fixed k, and there is no restriction for the other two vertex classes. We also consider the sampling problem, that is, to uniformly sample partite, 3-uniform hypergraphs with prescribed degree sequences. We propose a Parallel Tempering method, where the hypothetical energy of the hypergraphs measures the deviation from the prescribed degree sequence. The method has been implemented and tested on synthetic and real data. It can also be applied for χ2 testing of contingency tables. We have shown that this hypergraph-based χ2 test is more sensitive than the standard χ2 test. The extra sensitivity is especially advantageous on small data sets, where the proposed Parallel Tempering method shows promising performance.

Open Access: Yes

DOI: 10.1371/journal.pone.0303155

Application of spatio-temporal data in site-specific maize yield prediction with machine learning methods

E. Harsányi G. Milics Bálint Daróczy C. Kerepesi A. J. Kovács J. Nagy A. Nyéki M. Neményi András Benczúr

Publication Name: Precision Agriculture

Publication Date: 2021-10-01

Volume: 22

Issue: 5

Page Range: 1397-1415

Description:

In order to meet the requirements of sustainability and to determine yield drivers and limiting factors, it is now more likely that traditional yield modelling will be carried out using artificial intelligence (AI). The aim of this study was to predict maize yields using AI that uses spatio-temporal training data. The paper has advanced a new method of maize yield prediction, which is based on spatio-temporal data mining. To find the best solution, various models were used: counter-propagation artificial neural networks (CP-ANNs), XY-fused Querynetworks (XY-Fs), supervised Kohonen networks (SKNs), neural networks with Rectangular Linear Activations (ReLU), extreme gradient boosting (XGBoost), support-vector machine (SVM), and different subsets of the independent variables in five vegetation periods. Input variables for modelling included: soil parameters (pH, P2O5, K2O, Zn, clay content, ECa, draught force, Cone index), micro-relief averages, and meteorological parameters for the 63 treatment units in a 15.3 ha research field. The best performing method (XGBoost) reached 92.1% and 95.3% accuracy on the training and the test sets. Additionally, a novel method was introduced to treat individual units in a lattice system. The lattice-based smoothing performed an additional increase in Area under the curve (AUC) to 97.5% over the individual predictions of the XGBoost model. The models were developed using 48 different subsets of variables to determine which variables consistently contributed to prediction accuracy. By comparing the resulting models, it was shown that the best regression model was Extreme Gradient Boosting Trees, with 92.1% accuracy (on the training set). In addition, the method calculates the influence of the spatial distribution of site-specific soil fertility on maize grain yields. This paper provides a new method of spatio-temporal data analyses, taking the most important influencing factors on maize yields into account.

Open Access: Yes

DOI: 10.1007/s11119-021-09833-8

SZTAKI @ ImageCLEFmed 2020 Tuberculosis Task

Bálint Daróczy Bence Lestyan András Benczúr

Publication Name: Ceur Workshop Proceedings

Publication Date: 2020-01-01

Volume: 2696

Issue: Unknown

Page Range: Unknown

Description:

In this paper we describe our submission to the ImageCLEFmed 2020 Tuberculosis task and discuss additional results on the training set with various neural networks. After some centralization and normalization we independently categorized the 2D slices with convolutional neural networks (traditional and residual feed-forward networks) and we aggregated the individual predictions based on the positions of the lung and the slices. Our additional experiments with various aggregation methods indicate that individual slices do not necessary contain enough information about such complex structures.

Open Access: Yes

DOI: DOI not available

Node embeddings in dynamic graphs

Róbert Pálovics Domokos Kelen Ferenc Béres András Benczúr

Publication Name: Applied Network Science

Publication Date: 2019-12-01

Volume: 4

Issue: 1

Page Range: Unknown

Description:

In this paper, we present algorithms that learn and update temporal node embeddings on the fly for tracking and measuring node similarity over time in graph streams. Recently, several representation learning methods have been proposed that are capable of embedding nodes in a vector space in a way that captures the network structure. Most of the known techniques extract embeddings from static graph snapshots. By contrast, modeling the dynamics of the nodes in temporal networks requires evolving node representations. In order to update node representations that reflect the temporal changes in the local graph structure, we rely on ideas for data stream algorithms. For example, we assess neighborhood overlap by a MinHash fingerprint-based algorithm. To evaluate our methods, in addition to the standard link prediction task, we provide dynamic ground truth data for the quantitative evaluation of similarity search by using online updated node embeddings. In our experiments, we constructed tennis tournament Twitter mention graphs as edge streams and compiled dynamic ground truth by using tournament schedule as external source. Our new algorithms outperformed snapshot-based batch methods for both link prediction and similarity search.

Open Access: Yes

DOI: 10.1007/s41109-019-0169-5

Session recommendation via recurrent neural networks over fisher embedding vectors

Bálint Daróczy Frederick Ayala-Gómez Domokos Kelen Anna Ország András Benczúr

Publication Name: Sensors Switzerland

Publication Date: 2019-08-02

Volume: 19

Issue: 16

Page Range: Unknown

Description:

Recommendation services bear great importance in e-commerce, shopping, tourism, and social media, as they aid the user in navigating through the items that are most relevant to their needs. In order to build recommender systems, organizations log the item consumption in their user sessions by using different sensors. For instance, Web sites use Web data loggers, museums and shopping centers rely on user in-door positioning systems to register user movement, and Location-Based Social Networks use Global Positioning System for out-door user tracking. Most organizations do not have a detailed history of previous activities or purchases by the user. Hence, in most cases recommenders propose items that are similar to the most recent ones viewed in the current user session. The corresponding task is called session based, and when only the last item is considered, it is referred to as item-to-item recommendation. A natural way of building next-item recommendations relies on item-to-item similarities and item-to-item transitions in the form of “people who viewed this, also viewed” lists. Such methods, however, depend on local information for the given item pairs, which can result in unstable results for items with short transaction history, especially in connection with the cold-start items that recently appeared and had no time yet to accumulate a sufficient number of transactions. In this paper, we give new algorithms by defining a global probabilistic similarity model of all the items based on Random Fields. We give a generative model for the item interactions based on arbitrary distance measures over the items, including explicit, implicit ratings and external metadata to estimate and predict item-to-item transition probabilities. We exploit our new model in two different item similarity algorithms, as well as a feature representation in a recurrent neural network based recommender. Our experiments on various publicly available data sets show that our new model outperforms simple similarity baseline methods and combines well with recent item-to-item and deep learning recommenders under several different performance metrics.

Open Access: Yes

DOI: 10.3390/s19163498

Maize yield prediction based on artificial intelligence using spatio-temporal data

G. Milics Bálint Daróczy C. Kerepesi A. J. Kovács A. Nyéki M. Neményi András Benczúr

Publication Name: Precision Agriculture 2019 Papers Presented at the 12th European Conference on Precision Agriculture Ecpa 2019

Publication Date: 2019-01-01

Volume: Unknown

Issue: Unknown

Page Range: 1011-1017

Description:

The aim of this study was to predict maize yield by artificial intelligence using spatio-temporal training data. Counter-propagation artificial neural networks (CP-ANNs), XY-fused networks (XY-Fs), supervised Kohonen networks (SKNs), extreme gradient boosting (XGBoost) and support-vector machine (SVM) were used for predicting maize yield in 5 vegetation periods. Input variables for modelling were: soil parameters (pH, P2O5, K2O, Zn, Clay content, ECa, draught force, Cone index), and micro-relief averages and meteorological parameters for the 63 treatment units. The best performing method (XGBoost) attained 92.1 and 95.3% of accuracy on the training and the test set.

Open Access: Yes

DOI: 10.3920/978-90-8686-888-9_124

The applicability of on-line contextual calibration to a neural network based monocular collision avoidance system on a UAV

Bálint Daróczy Antal Hiba Koppány Pázmán Rita Aleksziev Péter Bauer András Benczúr Ákos Zarándy

Publication Name: IFAC Papersonline

Publication Date: 2019-01-01

Volume: 52

Issue: 11

Page Range: 7-12

Description:

Contextual calibration for object detection is a technique where a pretrained network collects attractive false positives during a calibration phase and use this calibration data for further training. This paper investigates the applicability of this method to a vision based onboard sense and avoid system, which requires intruder aircraft detection in camera images. Various landscape and sky backgrounds were generated by Unreal4 3D engine for calibration tests. Contextual calibration is a promising candidate for handling extreme situations which are not covered well in the training data.

Open Access: Yes

DOI: 10.1016/j.ifacol.2019.09.110