UPI Sentinel: A Transaction-Contextual Deep Learning Framework for UPI Fraud Detection

UPI Sentinel: A Transaction-Contextual Deep Learning Framework for UPI Fraud Detection

Ridham Taneja, Rahul Kumar Patel, Navneet Kumar, Paritosh Mukherjee, Arpita Singh

Department of Computer Science and Engineering (AI & ML) Raj Kumar Goel Institute of Technology

Ghaziabad, India

AbstractThe exponential adoption of the Unied Payments Interface (UPI) has revolutionized digital nance, yet it has concurrently created a fertile ground for sophisticated nancial fraud. Although existing fraud detection systems employing traditional Machine Learning (ML) classierssuch as Random Forests and Support Vector Machineshave shown efcacy in identifying static anomalies, they often struggle to capture the complex, sequential dependencies inherent in modern trans-actional fraud (e.g., high-frequency layering or mule account chaining). To address this limitation, this paper proposes a novel Hybrid Deep Learning framework specically adapted for the UPI domain. Our architecture synergizes Convolutional Neural Networks (CNN) to extract granular spatial features from transaction attributes with Long Short-Term Memory (LSTM) networks to model temporal dependencies across transaction sequences. Furthermore, we integrate an Attention Mechanism to dynamically weight the most critical time steps in a transaction history, allowing the model to focus on subtle indicators of fraud that are often diluted in long sequences. The framework is evaluated on the PaySim dataset, selected for its extreme class imbalance and similarity to real-world UPI transaction topologies. Experimental results demonstrate strong discrimina-tive performance, achieving an ROC-AUC of 0.997 and a PR-AUC of 0.872. At an optimized decision threshold, the model attains a fraud detection F1-score of 0.85, with a precision of 0.91 and recall of 0.80, while maintaining a very low false positive rate. These ndings indicate that combining spatio-temporal modeling with attention-based aggregation provides a robust and adaptable solution for fraud risk scoring in large-scale digital payment ecosystems.

Index TermsUPI Fraud Detection, Hybrid Deep Learning,

CNN-LSTM, Attention Mechanism, Imbalanced Classication, Transaction Context

1. Introduction

   The proliferation of mobile internet connectivity has po-sitioned the Unied Payments Interface (UPI) as a leading global infrastructure for real-time peer-to-peer (P2P) pay-ments. According to ofcial statistics released by the Na-tional Payments Corporation of India (NPCI), UPI processed more than 110 billion transactions in the rst half of 2025 alone [1]. Although this rapid adoption has expanded nancial accessibility, it has concurrently increased the attack surface for cybercriminals. As transactions become instantaneous, the window available for fraud detection narrows, necessitating accurate near real-time decision-making mechanisms.

   Financial fraud has evolved from isolated unauthorized ac-cess to coordinated schemes such as Authorized Push Payment

   (APP) fraud, layering, and mule account networks. Traditional fraud detection systems often rely on rule-based engines or static Machine Learning (ML) classiers, including Logistic Regression and Random Forests [2]. Although effective at detecting point anomalies (e.g., unusually large transaction amounts), such approaches typically fail to capture short-term temporal dependencies that characterize sophisticated fraud strategies. Fraudsters frequently execute a sequence of small testing transactions prior to a major withdrawala behavioral pattern that static models often overlook.

   To address these limitations, we propose UPI Sentinel, a hy-brid deep learning framework tailored to the high-throughput UPI ecosystem. Rather than introducing a novel standalone architecture, this study emphasizes the strategic integration of complementary deep learning paradigms to enhance detection performance [3]. Convolutional Neural Networks (CNNs) are employed to extract spatial representations from individual transaction records, while Long Short-Term Memory (LSTM) networks model evolving behavioral sequences. An attention mechanism is further incorporated to rene temporal aggre-gation [4]. By dynamically assigning greater importance to high-risk time stepssuch as sudden bursts of transactional activitythe model ensures that classication decisions are guided by the most informative behavioral signals.

   The primary contributions of this study are summarized as follows:

   • Hybrid Spatio-Temporal Architecture: Integration of a domain-tuned CRNN framework leveraging 1D-CNN layers for spatial feature extraction and BiLSTM layers for temporal modeling tailored to UPI transaction se-quences.

   • Context-Aware Data Pipeline: Implementation of a global sliding window strategy that converts isolated transaction logs into short-term sequential representations for contextual risk assessment.

   • Attention-Based Temporal Aggregation: Incorporation of an attention mechanism to dynamically weight infor-mative transaction steps, isolating high-risk patterns while suppressing benign transactional noise.

   • Probabilistic Risk Scoring: Utilization of a sigmoid-based continuous risk scoring mechanism enabling exi-ble decision thresholds for deployment scenarios such as transaction blocking or step-up authentication.

   • Commercial Viability Analysis: Comprehensive eval-uation on the highly imbalanced PaySim dataset with post-training threshold calibration, demonstrating prac-tical applicability by minimizing false positives while maintaining strong fraud detection sensitivity in real-world deployment settings.

2. Related Work

   1. Evolution from Static Classiers to Deep Learning

      Financial fraud detection initially relied on Machine Learn-ing (ML) classiers such as Logistic Regression, Support Vec-tor Machines (SVM)and Random Forests [2], [5]. Although computationally efcient and interpretable, these models ana-lyze transactions independently and are therefore effective at identifying isolated anomalies (e.g., extreme monetary values). However, they struggle to capture coordinated fraud behaviors that unfold across short transaction sequences. This challenge is further intensied by the severe class imbalance typical of nancial datasets, where minority fraud cases are easily overshadowed by legitimate activity [6].

      To address these limitations, deep learning approaches were introduced to enhance representation learning from trans-actional data [7]. Fully connected Deep Neural Networks (DNNs) demonstrated improved modeling of non-linear fea-ture interactions compared to traditional ML methods. Nev-ertheless, feedforward architectures remain inherently static and do not explicitly model the temporal evolution of user behavior. This limitation motivated the adoption of sequential modeling techniques capable of capturing behavioral progres-sion across consecutive transactions.

   2. Sequential Modeling and Attention Mechanisms

   Sequential deep learning architectures treat fraud detection as a temporal modeling task. Recurrent Neural Networks (RNNs) marked an early step in this direction but exhib-ited limitations in preserving long-range contextual depen-dencies. Long Short-Term Memory (LSTM) networks were subsequently adopted to better retain relevant historical infor-mation across transaction streams [8]. Bidirectional LSTs (BiLSTMs) further extend this capability by incorporating contextual information from both past and future positions within a transaction window [9].

   Although recurrent architectures improve temporal repre-sentation, they typically aggregate sequence information uni-formly, which may dilute critical fraud indicators within noisy histories. Attention mechanisms mitigate this issue by assign-ing adaptive importance to informative time steps, allowing models to emphasize high-risk behavioral segments [10]. Hy-brid frameworks combining convolutional layers for localized feature extraction with BiLSTM and attention components have demonstrated promising results in sequential learning applications [11].

   Despite these advancements, existing approaches frequently rely on rigid user-level grouping or long transaction histories, which may be impractical in privacy-sensitive or high-volume

   environments. Moreover, limited work jointly addresses short-term sequential modeling and extreme class imbalance without explicit user proling. This study addresses these gaps by introducing a transaction-contextual framework that integrates CRNN architectures with attention-based aggregation under a global sliding window formulation.

3. Methodology

   1. Dataset Description

      This study utilizes PaySim, a synthetic mobile money dataset introduced by Lopez-Rojas et al. [12]. The dataset comprises 6,362,620 transactions, of which 8,213 (0.13%) are labeled as fraudulent. Each record includes key attributes such as transaction amount, transaction type (e.g., Transfer, Cash-Out), temporal step index, and pre- and post-transaction balances for both sender and receiver accounts. The binary label isFraud indicates whether a transaction is fraudulent or legitimate. Owing to its severe class imbalance and sequential structure, PaySim provides a practical benchmark for evalu-ating context-aware fraud detection methods under realistic constraints [6].

   2. Dataset Preprocessing and Feature Encoding

      To prepare the dataset for the model, a set of preprocessing steps are performed to ensure data consistency, numerical compatibility, and stable learning behavior [2]. The categorical transaction type attribute is transformed into numerical form using label encoding, enabling its integration into neural net-work models while preserving distinctions between different transaction categories [5].

      Numerical features, including transaction amount and ac-count balance attributes, are examined for missing values. Any missing entries are replaced with zero, which is appropriate for the PaySim dataset as such values typically correspond to inactive or non-participating accounts rather than corrupted records [12]. To improve numerical stability and accelerate convergence during training, all numerical features are stan-dardized using z-score normalization [13]

   3. Transaction Sequence Construction

      To incorporate short-term transactional context into the fraud detection process, transactions are transformed into xed-length sequences using a global sliding window protocol. Prior work has explored sequential transaction modeling using sliding window mechanisms to capture short-term tem-poral dependencies [14], [15]. Unlike user-grouped sequence construction, which requires persistent user identiers and sufcient transaction history per account, a global sliding window enables contextual modeling over the complete trans-action stream without discarding data or introducing user-level

      dependencies.

      Rather than treating each transaction independently, this approach enables the model to observe a small temporal neighborhood of transactions and capture contextual patterns

      occurring over consecutive transaction steps. First, all transac-tions are ordered according to their temporal step to preserve temporal ordering within each constructed sequence.

      This process is illustrated in Fig.1, where overlapping windows generate sequential samples, and each sequence is labeled using the nal transaction in the window.

      Fig. 1. Global sliding window protocol for transaction sequence construction.

   4. Class Imbalance Handling Using SMOTE

      Financial transaction datasets are inherently imbalanced, with fraudulent cases representing only a small fraction of total transactions. Such imbalance can bias models toward the majority class, reducing fraud detection sensitivity.

      The dataset is partitioned into training and testing sets using an 80:20 split while preserving the original class distribution. To prevent information leakage, the Synthetic Minority Over-sampling Technique (SMOTE) is applied exclusively to the training data [16], while the test set remains unchanged.

      SMOTE generates synthetic minority samples through in-terpolation between existing fraud instances. In this study, oversampling is performed after sequence construction to preserve the temporal structure of transaction windows. This strategy improves minority-class representation during training while ensuring unbiased evaluation on the original distribution.

   5. Proposed CRNN Architecture

      To model short-term transactional patterns within con-structed transaction sequences, this study adopts a Con-volutional Recurrent Neural Network (CRNN) architecture augmented with an attention mechanism. The architecture combines convolutional layers for local feature extraction, recurrent layers for temporal dependency modeling, and an attention module to emphasize informative transaction steps. This integrated design enables effective learning from sequen-tial transaction data while maintaining robustness to noise and variability commonly present in nancial transactions. The architecture follows the general CRNN paradigm that combines convolutional and recurrent layers for sequential data modeling [11], and incorporates an attention mechanism to emphasize informative temporal patterns [14]. The proposed architecture is illustrated in Fig. 2.

      Fig. 2. Proposed CRNN architecture for transaction-level fraud detection

      1. Convolutional Feature Extraction: The architecture be-gins with a one-dimensional convolutional layer designed to extract local patterns from transaction sequences. Given an input sequence of transactions, the convolutional layer operates along the temporal dimension, enabling the model to learn short-range dependencies and interactions between consecu-tive transactions. Specically, the Conv1D layer applies 64 lters with a kernel size of 2, allowing the model to capture ne-grained interactions between adjacent transaction steps [7]. A Rectied Linear Unit (ReLU) activation function is used to introduce non-linearity and enhance the models ability to learn complex transactional patterns [17].

         To improve generalization and mitigate overtting, a dropout layer with a rate of 0.2 is applied immediately after the convolutional operation [18]. The resulting feature maps preserve the sequential structure of the input while providing enriched representations that are subsequently passed to the recurrent component for temporal modeling.

      2. Bidirectional LSTM for Temporal Modeling: Although convolutional layers capture local transactional patterns, mod-eling longer-range temporal dependencies requires a recurrent mechanism. To achieve this, the architecture employs a Bidi-rectional Long Short-Term Memory (BiLSTM) layer that pro-cesses the convolutional feature sequences in both forward and backward temporal directions [9]. This bidirectional structure enables the model to capture dependencies that may occur both before and after a given transaction within a sequence.

         The BiLSTM layer consists of 64 hidden units in each di-rection and is congured to return the full sequence of hidden states. This allows subsequent layers to access temporal infor-mation at each trnsaction step rather than a single aggregated representation. LSTM networks are particularly well-suited for sequential nancial data due to their ability to mitigate vanishing gradient issues and retain relevant information over time [19].

         By incorporating bidirectional processing, the model gains a more comprehensive understanding of transaction dynamics within each sequence, which is especially benecial in fraud detection scenarios where contextual cues may span multiple neighboring transactions [8]. The sequence-aware representa-tions produced by the BiLSTM layer are then forwarded to the attention mechanism for selective temporal aggregation.

      3. Attention Mechanism: Although recurrent layers model temporal dependencies across transaction sequences, not all transactions within a sequence contribute equally to fraud detection. To address this, an attention mechanism is integrated after the BiLSTM layer to enable selective emphasis on informative transaction steps. The attention layer operates on the sequence of hidden states produced by the BiLSTM and learns a set of importance weights over time. These weights indicate the relative contribution of each transaction in the sequence toward the nal prediction. A weighted aggregation of the hidden states is then performed to produce a xed-length context vector, which summarizes the most relevant temporal information for classication.

         By dynamically focusing on critical transactionssuch as sudden balance changes or unusually large transfersthe attention mechanism enhances interpretability and allows the model to prioritize salient temporal patterns. This approach has been shown to improve performance in sequence model-ing tasks by enabling adaptive feature weighting rather than uniform temporal aggregation [14], [10].

      4. Output Layer and Prediction: The context vector pro-duced by the attention mechanism is passed through a fully connected layer with 128 neurons and a ReLU activation function to learn higher-level abstractions from the aggregated temporal representation. To further enhance generalization and reduce overtting, a dropout layer with a rate of 0.3 is applied before the nal classication stage.

   The output layer consists of a single neuron with a sigmoid activation function, which maps the learned representation to a probability score indicating the likelihood of a transaction being fraudulent. This probabilistic output enables exible decision-making through threshold adjustment, allowing the model to balance precision and recall according to deployment requirements [20].

4. Experiments and Results

   1. Experimental Setup

      The proposed model was implemented using Python 3 with the TensorFlow/Keras deep learning framework and the Imbalanced-Learn library for class imbalance handling. All experiments were conducted in a Google Colab environment utilizing an NVIDIA Tesla T4 GPU for accelerated training.

      1. Dataset and Partitioning: The PaySim dataset [12], after preprocessing and transformation into sequential data using the global sliding window protocol, was used for all experiments. The dataset was divided using a stratied split, allocating 80% of the data for training and 20% for testing, ensuring that the original class distribution was preserved in both sets.

         To address the severe class imbalance inherent in fraud detection, the Synthetic Minority Over-sampling Technique (SMOTE) was applied only to the training data after sequence construction, thereby preventing data leakage [16]. The test set was kept unchanged and used solely for evaluation.

         To ensure computational efciency while maintaining sta-tistical relevance, model training was performed on a stratied and balanced subset of 100,000 samples drawn from the resampled training data.

      2. Hyperparameter Conguration: The model was trained using the Adam optimizer with a learning rate of 0.001, and binary cross-entropy was employed as the loss function. In addition to SMOTE, class-weighted training was utilized to further mitigate residual class imbalance during model opti-mization. The key hyperparameters used in the experiments are summarized in Table I.

         TABLE I Hyperparameter Configuration

         Parameter	Value
         Optimizer	Adam (Learning Rate = 0.001)
         Loss Function	Binary Cross-Entropy
         Batch Size	64
         Maximum Epochs	10
         Early Stopping	Patience = 3 (Validation Loss)
         Conv1D Filters	64 (Kernel Size = 2)
         BiLSTM Units	64
         Dropout Rates	0.2 (after CNN), 0.3 (after Dense)

   2. Evaluation Metrics and Threshold Tuning

      The model achieved an overall accuracy of 99.89% on the test set. Although this gure appears exceptionally high, it was deemed insufcient and misleading for performance evaluation due to the extreme class imbalance (0.13% fraud prevalence). In a dataset where 99.87% of transactions are legitimate, a trivial model that predicts Normal for every single transaction would still achieve 99.87% accuracy while failing to detect a single fraud case.

      Consequently, accuracy was discarded as a primary metric. Instead, performance was evaluated using F1-Score , Recall, Precision, and the Area Under the Curve (AUC) for both ROC and Precision-Recall (PR) characteristics [6].

      To minimize false positives while maintaining high fraud detection capability, we applied threshold tuning. The decision threshold was shifted from the default 0.5 to 0.9975, based on the maximization of the F1-score [20]. This adjustment resulted in a model that is highly conservative in agging transactions, thereby preserving the user experience for legit-imate customers.

   3. Confusion Matrix Analysis

      The classication performance at the optimal threshold ( = 0.9975) is visualized in Fig. 4. The confusion matrix reveals that the model successfully identied 1,253 fraudulent transactions (True Positives) out of a total of 1,643 cases.

      Notably, the model demonstrated exceptional specicity, misclassifying only 68 legitimate transactions as fraud (False

      Fig. 3. Training and validation loss across epochs for the proposed hybrid model..

      Positives) out of over 1.27 million legitimate samples. This low False Positive rate is critical in nancial fraud detection systems, as excessive false alarms can lead to customer dis-satisfaction and operational bottlenecks. The low number of False Negatives (390) indicates that while the model is highly precise, it still captures the vast majority of fraudulent activity.

      Fig. 4. Confusion matrix at threshold = 0.9975, illustrating the classica-tion outcomes of the proposed hybrid model.

   4. Receiver Operating Characteristic (ROC) Curve

      The models ability to discriminate between fraudulent and non-fraudulent classes is demonstrated in Fig. 5. The Receiver Operating Characteristic (ROC) curve exhibits a steep initial ascent, indicating a high True Positive Rate (Sensitivity) even at very low False Positive Rates.

      The Area Under the Curve (ROC-AUC) was calculated to be 0.9972. This near-perfect score conrms that the classier maintains robust separability between classes across various decision thresholds, validating the efcacy of the selected feature set and model architecture.

      Fig. 5. Receiver Operating Characteristic (ROC) curve with AUC = 0.9972 for the proposed hybrid model.

   5. Precision-Recall (PR) Curve Analysis

   Given the class imbalance, the Precison-Recall (PR) curve serves as a more reliable performance indicator than the ROC curve. As shown in Fig. 6, the model maintains high precision across a substantial range of recall values.

   The Area Under the Precision-Recall Curve (PR-AUC) achieved was 0.8718. Unlike the ROC metric, which can be overly optimistic in imbalanced scenarios, the PR-AUC specically highlights the models success in the minority class (Fraud). A score of 0.87 represents a strong trade-off, ensuring that when the model predicts fraud, it is highly likely to be correct, without missing a signicant portion of actual fraud cases.A consolidated summary of the models performance at the optimized threshold is presented in Table II.

   Fig. 6. Precision-Recall curve with PR-AUC = 0.8718 for the proposed hybrid model.

   TABLE II

   Summary of Model Performance

   Metric	Value
   Optimal Threshold ( )	0.9975
   ROC-AUC	0.9972
   PR-AUC	0.8718
   False Positives (FP)	68
   True Positives (TP)	1,253
   False Negatives (FN)	390

5. Conclusion and Future Work

This study demonstrates the effective application of a hybrid deep learning framework for transaction-contextual fraud detection using the PaySim mobile money simulation dataset. By integrating convolutional, bidirectional recurrent, and attention-based components within a unied pipeline, the proposed approach captures short-term transactional de-pendencies while maintaining robustness under severe class imbalance.

To address the low fraud prevalence (0.13%), the Synthetic Minority Over-sampling Technique (SMOTE) was applied ex-clusively during training to enhance minority-class representa-tion without introducing evaluation bias. Combined with post-training threshold optimization ( = 0.9975), this strategy en-ables the model to balance fraud sensitivity and false positive control effectively. The nal system achieved an ROC-AUC of 0.9972 and a False Positive Rate of 0.005%, demonstrating strong discriminative capability in highly imbalanced nancial settings.

While PaySim provides a realistic approximation of mobile money transactions, certain attributes (e.g., post-transaction balances) may reect information not immediately available at real-time decision points. Nevertheless, the proposed frame-work does not depend on any single feature; instead, it learns contextual patterns across multiple transactional attributes and short temporal windows. This design supports robust pre-diction even when individual features are delayed, partially observable, or noisy.

While the current framework demonstrates strong perfor-mance, several avenues exist for further enhancement:

• Integration of Transformer-based architectures to enhance long-range dependency modeling.

• Deployment-oriented latency evaluation in real-time streaming environments.

• Incorporation of explainability techniques (e.g., SHAP) to improve interpretability and regulatory transparency.

In practical deployments, access to richer contextual sig-nalssuch as device-level attributes, behavioral velocity indi-cators, and network-level risk featuresmay further enhance predictive performance. Integrating such signals alongside complementary techniques (e.g., graph-based risk modeling) represents a promising direction for improving scalability and adaptability in large-scale nancial systems.

References

1. National Payments Corporation of India, UPI product statistics, 2025, accessed: 2025-01-23. [Online]. Available: https://www.npci.org.in/what-we-do/upi/product-statistics

2. J. Han, M. Kamber, and J. Pei, Data Mining: Concepts and Techniques, 3rd ed. Waltham, MA, USA: Morgan Kaufmann, 2011.

3. S. Jagadeesan, K. S. Arjun, G. Dhanika, G. Karthikeyan, and K. Deepika, UPI fraud detection using machine learning, in Challenges in Informa-tion, Communication and Computing Technology, 1st ed., V. Sharmila et al., Eds. London, U.K.: CRC Press, 2024, pp. 755760.

4. I. Akour, N. Mohamed, and S. Salloum, Hybrid CNN-LSTM with attention mechanism for robust credit card fraud detection, IEEE Access, vol. 13, pp. 112, 2025.

5. A. Ge´ron, Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 2nd ed. Sebastopol, CA, USA: OReilly Media, 2019.

6. A. Dal Pozzolo, O. Caelen, R. A. Johnson, and G. Bontempi, Cali-brating probability with undersampling for unbalanced classication, in Proceedings of the IEEE Symposium Series on Computational Intel-ligence (SSCI), Cape Town, South Africa, Dec. 2015, pp. 159166.

7. Y. LeCun, Y. Bengio, and G. Hinton, Deep learning, Nature, vol. 521, no. 7553, pp. 436444, May 2015.

8. F. A. Gers, J. Schmidhuber, and F. Cummins, Learning to forget: Continual prediction with LSTM, Neural Computation, vol. 12, no. 10,

   pp. 24512471, Oct. 2000.

9. M. Schuster and K. K. Paliwal, Bidirectional recurrent neural net-works, IEEE Transactions on Signal Processing, vol. 45, no. 11, pp. 26732681, Nov. 1997.

10. Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, and E. Hovy, Hierarchical attention networks for document classication, in Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), San Diego, CA, USA, Jun. 2016, pp. 14801489.

11. B. Shi, X. Bai, and C. Yao, An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 11, pp. 22982304, Nov. 2017.

12. E. A. Lopez-Rojas, A. Elmir, and S. Axelsson, PaySim: A nancial mobile money simulator for fraud detection, in Proceedings of the 28th European Modeling and Simulation Symposium (EMSS), Larnaca, Cyprus, 2016.

13. S. Patro and K. K. Sahu, Normalization: A preprocessing stage, arXiv preprint arXiv:1503.06462, 2015. [Online]. Available: https://arxiv.org/abs/1503.06462

14. D. Bahdanau, K. Cho, and Y. Bengio, Neural machine translation by jointly learning to align and translate, in Proceedings of the 3rd International Conference on Learning Representations (ICLR), San Diego, CA, 2015.

15. G. Lai, W.-C. Chang, Y. Yang, and H. Liu, Modeling long- and short-term temporal patterns with deep neural networks, in Proceedings of the 41st International ACM SIGIR Conference on Research & Development in Information Retrieval (SIGIR), Ann Arbor, MI, USA, 2018, pp. 95104.

16. N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, SMOTE: Synthetic minority over-sampling technique, Journal of Articial Intelligence Research, vol. 16, pp. 321357, Jun. 2002.

17. V. Nair and G. E. Hinton, Rectied linear units improve restricted boltz-mann machines, in Proceedings of the 27th International Conference on Machine Learning (ICML), Haifa, Israel, Jun. 2010, pp. 807814.

18. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut-dinov, Dropout: A simple way to prevent neural networks from overt-ting, Journal of Machine Learning Research, vol. 15, no. 1, pp. 19291958, Jun. 2014.

19. Y. Bengio, P. Simard, an P. Frasconi, Learning long-term dependen-cies with gradient descent is difcult, IEEE Transactions on Neural Networks, vol. 5, no. 2, pp. 157166, Mar. 1994.

20. T. Saito and M. Rehmsmeier, The precision-recall plot is more informa-tive than the ROC plot when evaluating binary classiers on imbalanced datasets, PLoS ONE, vol. 10, no. 3, p. e0118432, Mar. 2015.