Imbalanced Training Data Research Articles

MVSF-AB: Accurate antibody-antigen binding affinity prediction via multi-view sequence feature learning.

Predicting the binding affinity between antigens and antibodies accurately is crucial for assessing therapeutic antibody effectiveness and enhancing antibody engineering and vaccine design. Traditional machine learning methods have been widely used for this purpose, relying on interfacial amino acids' structural information. Nevertheless, due to technological limitations and high costs of acquiring structural data, the structures of most antigens and antibodies are unknown, and sequence-based methods have gained attention. Existing sequence-based approaches designed for protein-protein affinity prediction exhibit a significant drop in performance when applied directly to antibody-antigen affinity prediction due to imbalanced training data and lacking design in the model framework specifically for antibody-antigen, hindering the learning of key features of antibodies and antigens. Therefore, we propose MVSF-AB, a Multi-View Sequence Feature learning for accurate Antibody-antigen Binding affinity prediction. MVSF-AB designs a multi-view method that fuses semantic features and residue features to fully utilize the sequence information of antibody-antigen and predicts the binding affinity. Experimental results demonstrate that MVSF-AB outperforms existing approaches in predicting unobserved natural antibody-antigen affinity and maintains its effectiveness when faced with mutant strains of antibodies. Datasets we used and source code are available on our public GitHub repository https://github.com/TAI-Medical-Lab/MVSF-AB.

SISTEM REKOMENDASI KEBUTUHAN NUTRISI MAKANAN PENDAMPING AIR SUSU IBU (MPASI) MENGGUNAKAN METODE FUZZY TAHANI SEBAGAI UPAYA PENCEGAHAN STUNTING

Stunting is a significant child health issue in Indonesia, where children experience impaired growth, resulting in weight and height below their age average. In 2022, the World Health Organization (WHO) and the Indonesian Nutrition Status Survey (SSGI) reported a stunting prevalence of 21.6%, exceeding the WHO target of 20%. Stunting is caused by chronic malnutrition from pregnancy through the first thousand days of life, a crucial period for brain and cognitive development. Proper complementary feeding (MPASI) after 6 months is essential to meet infants' nutritional needs and prevent developmental delays in children under two. This study aims to develop a mobile web-based system that recommends MPASI according to infants' nutritional needs using the Fuzzy Tahani method. This method was chosen for its ability to handle uncertainty and provide personalized solutions. The developed system achieved 94.5% accuracy, with a precision of 33.1%, recall of 27.4%, and specificity of 85.9%. The system met the research objectives despite lower precision and recall due to imbalanced training data. User Acceptance Testing (UAT) results indicated that 77.73% of users strongly agreed with the system, while 1.86% disagreed. This system is expected to assist in preventing stunting by providing more accurate MPASI recommendations tailored to each infant's needs.

A Dynamic Multi‐Output Convolutional Neural Network for Skin Lesion Classification

ABSTRACTSkin cancer is a pressing global health issue, with high incidence and mortality rates. Convolutional neural network (CNN) models have been proven to be effective in improving performance in skin lesion image classification and reducing the medical burden. However, the inherent class imbalance in training data caused by the difficulty of collecting dermoscopy images leads to categorical overfitting, which still limits the effectiveness of these data‐driven models in recognizing few‐shot categories. To address this challenge, we propose a dynamic multi‐output convolutional neural network (DMO‐CNN) model that incorporates exit nodes into the standard CNN structure and includes feature refinement layers (FRLs) and an adaptive output scheduling (AOS) module. This model improves feature representation ability through multi‐scale sub‐feature maps and reduces the inter‐layer dependencies during backpropagation. The FRLs ensure efficient and low‐loss down‐sampling, while the AOS module utilizes a trainable layer selection mechanism to refocus the model's attention on few‐shot lesion categories. Additionally, a novel correction factor loss is introduced to supervise and promote AOS convergence. Our evaluation of the DMO‐CNN model on the HAM10000 dataset demonstrates its effectiveness in multi‐class skin lesion classification and its superior performance in recognizing few‐shot categories. Despite utilizing a very simple VGG structure as the sole backbone structure, DMO‐CNN achieved impressive performance of 0.885 in BACC and 0.983 in weighted AUC. These results are comparable to those of the ensemble model that won the ISIC 2018 challenge, highlighting the strong potential of DMO‐CNN in dealing with few‐shot skin lesion data.

A two-module bias-correction model for sea wave hindcasting based on the long-short term memory neural network

The lack of observations and biases of winds and bathymetry in the coastal region result in significant biases in the numerical modeling of sea waves. In this study, we establish a two-module bias correction model based on the Long-Short Term Memory neural network (denoted as TM-BCM) which accounts for the data imbalance by separately training the neural network for low and high sea states. The in-situ observations in 2020 and 2021 are used to model training and testing, respectively. The test results show that TM-BCM significantly reduces the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) of modeled sea waves by 52.9% and 65.5%, respectively. In addition, TM-BCM outperforms the single-module model (denoted as SM-BCM) that does not account for the imbalance of training data. Our results highlight the great value of deep learning in correcting biases for numerical models and the importance of accounting for data imbalance in training the neural network.

GTR: GAN-Based Trusted Routing Algorithm for Underwater Wireless Sensor Networks.

The transmission environment of underwater wireless sensor networks is open, and important transmission data can be easily intercepted, interfered with, and tampered with by malicious nodes. Malicious nodes can be mixed in the network and are difficult to distinguish, especially in time-varying underwater environments. To address this issue, this article proposes a GAN-based trusted routing algorithm (GTR). GTR defines the trust feature attributes and trust evaluation matrix of underwater network nodes, constructs the trust evaluation model based on a generative adversarial network (GAN), and achieves malicious node detection by establishing a trust feature profile of a trusted node, which improves the detection performance for malicious nodes in underwater networks under unlabeled and imbalanced training data conditions. GTR combines the trust evaluation algorithm with the adaptive routing algorithm based on Q-Learning to provide an optimal trusted data forwarding route for underwater network applications, improving the security, reliability, and efficiency of data forwarding in underwater networks. GTR relies on the trust feature profile of trusted nodes to distinguish malicious nodes and can adaptively select the forwarding route based on the status of trusted candidate next-hop nodes, which enables GTR to better cope with the changing underwater transmission environment and more accurately detect malicious nodes, especially unknown malicious node intrusions, compared to baseline algorithms. Simulation experiments showed that, compared to baseline algorithms, GTR can provide a better malicious node detection performance and data forwarding performance. Under the condition of 15% malicious nodes and 10% unknown malicious nodes mixed in, the detection rate of malicious nodes by the underwater network configured with GTR increased by 5.4%, the error detection rate decreased by 36.4%, the packet delivery rate increased by 11.0%, the energy tax decreased by 11.4%, and the network throughput increased by 20.4%.

Imbalanced spectral data analysis using data augmentation based on the generative adversarial network

Spectroscopic techniques generate one-dimensional spectra with distinct peaks and specific widths in the frequency domain. These features act as unique identities for material characteristics. Deep neural networks (DNNs) has recently been considered a powerful tool for automatically categorizing experimental spectra data by supervised classification to evaluate material characteristics. However, most existing work assumes balanced spectral data among various classes in the training data, contrary to actual experiments, where the spectral data is usually imbalanced. The imbalanced training data deteriorates the supervised classification performance, hindering understanding of the phase behavior, specifically, sol-gel transition (gelation) of soft materials and glycomaterials. To address this issue, this paper applies a novel data augmentation method based on a generative adversarial network (GAN) proposed by the authors in their prior work. To demonstrate the effectiveness of the proposed method, the actual imbalanced spectral data from Pluronic F-127 hydrogel and Alpha-Cyclodextrin hydrogel are used to classify the phases of data. Specifically, our approach improves 8.8%, 6.4%, and 6.2% of the performance of the existing data augmentation methods regarding the classifier’s F-score, Precision, and Recall on average, respectively. Specifically, our method consists of three DNNs: the generator, discriminator, and classifier. The method generates samples that are not only authentic but emphasize the differentiation between material characteristics to provide balanced training data, improving the classification results. Based on these validated results, we expect the method’s broader applications in addressing imbalanced measurement data across diverse domains in materials science and chemical engineering.

A multi-domain adaptive neural machine translation method based on domain data balancer

Most methods for multi-domain adaptive neural machine translation (NMT) currently rely on mixing data from multiple domains in a single model to achieve multi-domain translation. However, this mixing can lead to imbalanced training data, causing the model to focus on training for the large-scale general domain while ignoring the scarce resources of specific domains, resulting in a decrease in translation performance. In this paper, we propose a multi-domain adaptive NMT method based on Domain Data Balancer (DDB) to address the problems of imbalanced data caused by simple fine-tuning. By adding DDB to the Transformer model, we adaptively learn the sampling distribution of each group of training data, replace the maximum likelihood estimation criterion with empirical risk minimization training, and introduce a reward-based iterative update of the bilevel optimizer based on reinforcement learning. Experimental results show that the proposed method improves the baseline model by an average of 1.55 and 0.14 BLEU (Bilingual Evaluation Understudy) scores respectively in English-German and Chinese-English multi-domain NMT.

Managing Expectations and Imbalanced Training Data in Reactive Force Field Development: An Application to Water Adsorption on Alumina.

ReaxFF is a computationally efficient model for reactive molecular dynamics simulations that has been applied to a wide variety of chemical systems. When ReaxFF parameters are not yet available for a chemistry of interest, they must be (re)optimized, for which one defines a set of training data that the new ReaxFF parameters should reproduce. ReaxFF training sets typically contain diverse properties with different units, some of which are more abundant (by orders of magnitude) than others. To find the best parameters, one conventionally minimizes a weighted sum of squared errors over all of the data in the training set. One of the challenges in such numerical optimizations is to assign weights so that the optimized parameters represent a good compromise among all the requirements defined in the training set. This work introduces a new loss function, called Balanced Loss, and a workflow that replaces weight assignment with a more manageable procedure. The training data are divided into categories with corresponding "tolerances", i.e., acceptable root-mean-square errors for the categories, which define the expectations for the optimized ReaxFF parameters. Through the Log-Sum-Exp form of Balanced Loss, the parameter optimization is also a validation of one's expectations, providing meaningful feedback that can be used to reconfigure the tolerances if needed. The new methodology is demonstrated with a nontrivial parametrization of ReaxFF for water adsorption on alumina. This results in a new force field that reproduces both the rare and frequent properties of a validation set not used for training. We also demonstrate the robustness of the new force field with a molecular dynamics simulation of water desorption from a γ-Al2O3 slab model.

Comparison of CNN and SVM Methods on Web-based Skin Disease Classification Process

Skin, as the outermost layer of the body, is often in contact with bacteria, germs and viruses because of its most external position. According to statistics from the 2009 Indonesian Health Profile, skin illness is the third most common ailment seen in outpatient settings across the country's hospitals. Therefore, maintaining healthy skin is important because it protects the body's internal organs from injury and attack by pathogens. The development of image classification, such as the classification of skin diseases, has become a focus in the health sector. This research analyses the performance of Convolutional Neural Network (CNN) and Support Vector Machine (SVM) in web-based skin disease classification and overcomes the problem of imbalanced training data. With data augmentation and preprocess, this research improves data generalization and compares performance metrics such as Recall, Accuracy, and F1 Score. The results show that the average accuracy of CNN is 83.8%, while SVM reaches 81%. Although both models have high metrics for the normal class, other more complicated classes can only be handled by CNN with a value of more than 0.9. Apart from that, the CNN method also provides a higher Confidence Score than SVM, as well as a faster execution time. In conclusion, the CNN method is superior and recommended for skin disease classification based on web applications based on various performance test results.

Classification of the machine state in turning processes by using the acoustic emission

Processing digital information stands as a crucial foundation of Industry 4.0, facilitating a spectrum of activities from monitoring processes to their understanding and optimization. The application of data processing techniques, including feature extraction and classification, coupled with the identification of the most suitable features for specific purposes, continues to pose a significant challenge in the manufacturing sector. This research investigates the suitability of classification methods for machine and tool state classification by employing acoustic emission (AE) sensors during the dry turning of Ti6Al4V. Features such as quantiles, Fourier coefficients, and mel-frequency cepstral coefficients are extracted from the AE signals to facilitate classification. From this features the 20 best are selected for the classification to reduce the dimension of the feature space and redundancy. Algorithms including decision tree, k-nearest-neighbors (KNN), and quadratic discriminant analysis (QDA) are tested for the classification of machine states. Of these, QDA exhibits the highest accuracy at 98.6 %. Nonetheless, an examination of the confusion matrix reveals that certain classes, influenced by imbalanced training data, exhibit a lower prediction accuracy. In summary, the study affirms the potential of AE sensors for machine state recognition and tool condition monitoring. Although QDA emerges as the most acurate classifier, there remains an avenue for refinement, particularly in training data optimization and decision-making processes, to augment accuracy.

First estimation of hourly full-coverage ground-level ozone from Fengyun-4A satellite using machine learning

Ground-level ozone (O3), renowned for its adverse impacts on human health and crop production, has garnered significant attention from governmental and public sectors. To address the limitations posed by sparse and uneven ground-level O3 observations, this study proposes an innovative method for hourly full-coverage ground-level O3 estimation using machine learning. Meteorological data from National Centers for Environmental Prediction global forecasting system, satellite data from Fengyun-4 A(FY-4 A) and Ozone Monitoring Instrument, emission inventory from Multi-resolution Emission Inventory for China, and other auxiliary data are utilized as input variables, while ground-based O3 observations serve as the response variable. The method is applied on a monthly basis across China for the year 2022, resulting in the generation of an hourly full-coverage high-resolution (4 km) ground-level O3 estimation, termed ML-derived-O3. Cross-validation results demonstrate the robustness of ML-derived-O3 yielding a coefficient of determination (R 2) of 0.96 (0.91) for sample-based (site-based) evaluations and a root-mean-square error (RMSE) of 9.22 (13.65) µg m−3. However, the date-based evaluation is less satisfactory due to the imbalanced training data, resulting from the pronounced daily variations in ground-level O3 concentrations. Nevertheless, the seasonal and hourly ML-derived-O3 exhibits high prediction accuracy, with R 2 values surpassing 0.95 and RMSE remaining below 7.5 µg m−3. This study marks a significant milestone as the first successful attempt to obtain hourly full-coverage ground-level O3 data across China. The diurnal variation of ML-derived-O3 demonstrates high consistency with ground-based observations, irrespective of clear or cloudy days, effectively capturing ground-level O3 pollution exposure events. This novel estimation method will be employed to establish a long-term high spatial-temporal resolution ground-level O3 dataset, which holds valuable applications for air pollution monitoring and environmental health research in future endeavors.

Joint streaming model for backchannel prediction and automatic speech recognition

AbstractIn human conversations, listeners often utilize brief backchannels such as “uh‐huh” or “yeah.” Timely backchannels are crucial to understanding and increasing trust among conversational partners. In human–machine conversation systems, users can engage in natural conversations when a conversational agent generates backchannels like a human listener. We propose a method that simultaneously predicts backchannels and recognizes speech in real time. We use a streaming transformer and adopt multitask learning for concurrent backchannel prediction and speech recognition. The experimental results demonstrate the superior performance of our method compared with previous works while maintaining a similar single‐task speech recognition performance. Owing to the extremely imbalanced training data distribution, the single‐task backchannel prediction model fails to predict any of the backchannel categories, and the proposed multitask approach substantially enhances the backchannel prediction performance. Notably, in the streaming prediction scenario, the performance of backchannel prediction improves by up to 18.7% compared with existing methods.

Quantitative analysis of [formula omitted] thin film micrographs with machine learning

Isolating the features associated with different materials growth conditions is important to facilitate the tuning of these conditions for effective materials growth and characterization. This study presents machine learning models for classifying atomic force microscopy (AFM) images of thin film MoS2 based on their growth temperatures. By employing nine different algorithms and leveraging transfer learning through a pretrained ResNet model, we identify an effective approach for accurately discerning the characteristics related to growth temperature within the AFM micrographs. Robust models with test accuracies of up to 70% were obtained, with the best performing algorithm being an end-to-end ResNet fine-tuned on our image domain. Class activation maps and occlusion attribution reveal that crystal quality and domain boundaries play crucial roles in classification, with models exhibiting the ability to identify latent features that humans could potentially miss. Overall, the models demonstrated high accuracy in identifying thin films grown at different temperatures despite limited and imbalanced training data as well as variation in growth parameters besides temperature, showing that our models and training protocols are suitable for this and similar predictive tasks for accelerated 2D materials characterization.

Improved prediction of extreme ENSO events using an artificial neural network with weighted loss functions

The El Niño–Southern Oscillation (ENSO) causes a wide array of abnormal climates and extreme events, including severe droughts and floods, which have a major impact on humanity. With the development of artificial neural network techniques, various attempts are being made to predict ENSO more accurately. However, there are still limitations in accurately predicting ENSO beyond 6 months, especially for abnormal years with less frequent but greater impact, such as strong El Niño or La Niña, mainly due to insufficient and imbalanced training data. Here, we propose a new weighted loss function to improve ENSO prediction for abnormal years, in which the original (vanilla) loss function is multiplied by the weight function that relatively reduces the weight of high-frequency normal events. The new method applied to recurrent neural networks shows significant improvement in ENSO predictions for all lead times from 1 month to 12 months compared to using the vanilla loss function; in particular, the longer the prediction lead time, the greater the prediction improvement. This method can be applied to a variety of other extreme weather and climate events of low frequency but high impact.

Diverse Data Generation for Retinal Layer Segmentation With Potential Structure Modeling.

Accurate retinal layer segmentation on optical coherence tomography (OCT) images is hampered by the challenges of collecting OCT images with diverse pathological characterization and balanced distribution. Current generative models can produce high-realistic images and corresponding labels without quantitative limitations by fitting distributions of real collected data. Nevertheless, the diversity of their generated data is still limited due to the inherent imbalance of training data. To address these issues, we propose an image-label pair generation framework that generates diverse and balanced potential data from imbalanced real samples. Specifically, the framework first generates diverse layer masks, and then generates plausible OCT images corresponding to these layer masks using two customized diffusion probabilistic models respectively. To learn from imbalanced data and facilitate balanced generation, we introduce pathological-related conditions to guide the generation processes. To enhance the diversity of the generated image-label pairs, we propose a potential structure modeling technique that transfers the knowledge of diverse sub-structures from lowly- or non-pathological samples to highly pathological samples. We conducted extensive experiments on two public datasets for retinal layer segmentation. Firstly, our method generates OCT images with higher image quality and diversity compared to other generative methods. Furthermore, based on the extensive training with the generated OCT images, downstream retinal layer segmentation tasks demonstrate improved results. The code is publicly available at: https://github.com/nicetomeetu21/GenPSM.

Detecting anomalous electricity consumption with transformer and synthesized anomalies

Non-technical losses are consistently a troubling issue for power suppliers. With the application and popularization of smart grid and advanced measurement systems, it has become possible to use data-driven methods to detect anomalous electricity consumption to reduce non-technical losses. A range of machine learning models have been utilized for detecting anomalous electricity consumption and have achieved promising results. However, with the evolution of techniques like electricity theft, coupled with the exponential increase in electricity consumption data, new challenges are constantly being posed for anomalous electricity consumption detection. We propose a Transformer-based method for detecting anomalous electricity consumption. The Transformer is composed of multi-head attention, layer normalization, point-wise feed-forward network, etc., which can effectively handle electricity consumption time-series data. Meanwhile, to alleviate the problem of imbalanced training data between anomalous and normal electricity consumption, we propose a method for synthesizing anomalies. The experimental results demonstrate that our proposed Transformer-based method outperforms the state-of-the-art methods in detecting anomalous electricity consumption, achieving a precision of 93.9%, a recall of 96.3%, an F1-score of 0.951, and an accuracy of 95.6% on a dataset released by the State Grid Corporation of China.

Defect identification method for ultrasonic inspection of pipeline welds based on cross-modal zero-shot learning

Ultrasonic inspection of pipeline welds still uses the traditional visual inspection signal method to identify pipeline defects. The identification of defects relies entirely on the subjective judgment of practitioners and is highly dependent on their level of experience. Deep learning models have achieved very good results in classification tasks, but they rely on a large number of annotated data samples for each category. However, it is difficult to collect a large number of samples with different defects and annotate them for the classification of pipe welding defects. Based on the idea of zero-shot learning (ZSL), which makes full use of experts’ semantic descriptions of defect categories, artificial semantic features are integrated cross-modally with ultrasonic inspection signal features. In this way, a common semantic space containing seen and unseen classes is constructed to achieve the detection of various defects. Meanwhile, to alleviate the problem of extreme imbalance of training data between the seen and unseen classes in ZSL model training, a ZSL model Feature-GAN-ZSL (FGZ) fused with a generative adversarial network (GAN) is proposed. The model utilizes a Feature-GAN network to generate unseen class features during training and adds a classifier to enhance the generation of features with stronger discriminative power. In the experiments, sample data for porosity, incomplete penetration, and cracks were used as visible classes, and samples for incomplete fusion and slag entrapment were used as unseen classes. Five state-of-the-art models in the ZSL domain were compared. The results show that the FGZ model has a good ability to recognize various defects, not only the types of defects that participated in the training but also the defects that did not participate in the training. This plays a perfect role in dealing with various pipeline welding defects.

TNN-IDS: Transformer neural network-based intrusion detection system for MQTT-enabled IoT Networks

The Internet of Things (IoT) is a global network that connects a large number of smart devices. MQTT is a de facto standard, lightweight, and reliable protocol for machine-to-machine communication, widely adopted in IoT networks. Various smart devices within these networks are employed to handle sensitive information. However, the scale and openness of IoT networks make them highly vulnerable to security breaches and attacks, such as eavesdropping, weak authentication, and malicious payloads. Hence, there is a need for advanced machine learning (ML) and deep learning (DL)-based intrusion detection systems (IDS). Existing ML-based IoT-IDSs face several limitations in effectively detecting malicious activities, mainly due to imbalanced training data. To address this, this study introduces a transformer neural network-based intrusion detection system (TNN-IDS) specifically designed for MQTT-enabled IoT networks. The proposed approach aims to enhance the detection of malicious activities within these networks. The TNN-IDS leverages the parallel processing capability of the Transformer Neural Network, which accelerates the learning process and results in improved detection of malicious attacks. To evaluate the performance of the proposed system, it was compared with various IDSs based on ML and DL approaches. The experimental results demonstrate that the proposed TNN-IDS outperforms other systems in terms of detecting malicious activity. The TNN-IDS achieved optimum accuracies reaching 99.9% in detecting malicious activities.

Plasma‐Sheet Bubble Identification Using Multivariate Time Series Classification

AbstractPlasma‐sheet bubbles play a major role in the earthward transport of magnetotail particles. The most remarkable feature of bubbles is their fast bulk flow velocities, along with reduced plasma density and pressure accompanied by magnetic field dipolarization. These bubbles can be identified based on in situ observations, but subjective ambiguity necessitates human verification, due to confusion with other phenomena mostly associated with magnetic reconnection and plasma waves. In this study, we aim to employ machine learning (ML) techniques to detect bubbles automatically and to create a tool that can be utilized by individuals without specialized subject expertise. To identify bubbles, we combine three distinct techniques: MINImally RandOm Convolutional KErnel Transform (MINIROCKET), 1D convolution neural network, and Residual Network (ResNet). The imbalanced training data set consists of bubble and non‐bubble events with a ratio of 1:40 from 2007 to 2020. The results indicate that the accuracy of all three models is approximately 99%, and their precision, recall, and F2 score are all above 80% for both the validation and test datasets. The three methods are combined with the intersection set as the minimum set of predictions and the union set as the maximum set. The union set can accurately identify 66.7% of bubbles. The combined method reduces the number of false negatives significantly. In the prediction of bubbles in observations made in the year 2021 using a union set, the bubbles obtained by the model are comparable to those discovered using traditional criteria and manual inspections.

Speaker Recognition Based on the Joint Loss Function

The statistical pyramid dense time-delay neural network (SPD-TDNN) model makes it difficult to deal with the imbalance of training data, poses a high risk of overfitting, and has weak generalization ability. To solve these problems, we propose a method based on the joint loss function and improved statistical pyramid dense time-delay neural network (JLF-ISPD-TDNN), which improves on the SPD-TDNN model and uses the joint loss function method to combine the advantages of the cross-entropy loss function and the comparative learning of the loss function. By minimizing the distance between speech embeddings from the same speaker and maximizing the distance between speech embeddings from different speakers, the model could achieve enhanced generalization performance and more robust speaker feature representation. We evaluated the proposed method’s performance using the evaluation indexes of the equal error rate (EER) and minimum cost function (minDCF). The experimental results show that the EEE and minDCF on the Aishell-1 dataset reached 1.02% and 0.1221%, respectively. Therefore, using the joint loss function in the improved SPD-TDNN model can significantly enhance the model’s speaker recognition performance.