Reduced Feature Space Research Articles

Feature space reduction method for ultrahigh-dimensional, multiclass data: random forest-based multiround screening (RFMS)

In recent years, several screening methods have been published for ultrahigh-dimensional data that contain hundreds of thousands of features, many of which are irrelevant or redundant. However, most of these methods cannot handle data with thousands of classes. Prediction models built to authenticate users based on multichannel biometric data result in this type of problem. In this study, we present a novel method known as random forest-based multiround screening (RFMS) that can be effectively applied under such circumstances. The proposed algorithm divides the feature space into small subsets and executes a series of partial model builds. These partial models are used to implement tournament-based sorting and the selection of features based on their importance. This algorithm successfully filters irrelevant features and also discovers binary and higher-order feature interactions. To benchmark RFMS, a synthetic biometric feature space generator known as BiometricBlender is employed. Based on the results, the RFMS is on par with industry-standard feature screening methods, while simultaneously possessing many advantages over them.

High-Dimensional Feature Space for Diabetes Diagnosis and Identification of Diabetic-Sensitive Features in Ayurvedic Nadi Signals

Nadi-based disease diagnosis is a traditional art in Ayurvedic medicine that is an inquisitive yet not widely comprehended subject. A collection of higher dimensional features from a preprocessed Nadi dataset was extracted and analyzed to diagnose diabetes. The t-distributed Stochastic Neighbor Embedding was used to visualize the higher dimensional feature space in 2-D. The linear dimensionality reduction method of Principal Component Analysis and several linear and nonlinear classifiers were tested on the reduced feature space in identifying diabetes. The key outcomes of this paper are the ability to reduce the feature space by 73.33% while retaining a classification accuracy of 95.4%, identifying age as a compounding factor in diagnosis, and extracting the diabetes-sensitive features with eigenvalue loading.

Deep Neural Networks are Utilized in Energy Management Systems for identifying Theft of Electricity

The installation of connected systems is essential for the identification of renewable energy thieves because it generate enormous volumes of the information, include information about customer behavior, that could potentially be identify electricity theft using machine learning and deep learning techniques. This project describes a method for detecting theft that employs extensive information in the time and frequency domains in a deep neural network-based classification approach. Through data interpolation and synthetic data generation procedures, we solve dataset flaws such as missing data and class imbalance issues. We evaluate and analyze the contribution of features from both the temporal and frequency domains, execute experiments in combined and reduced feature space using principal component analysis, and lastly add a minimal redundancy-based strategy to maximizing significance for determining what's most pertinent features. We increase the detection performance of power theft by optimizing hyper parameters with a Bayesian optimizer and using an adaptive moment estimation optimizer to run tests with varying values of critical parameters to identify the ideal settings that produce the greatest accuracy. Since we train the model with over and under sampling datasets, it provides the equal representation while training the model. Key Words: Cyberbullying, Neural Networks Machine learning, social media. Deep Neural Network, Natural language processing, Artificial Neural Network.

Investigation of process history and underlying phenomena associated with the synthesis of plutonium oxides using Vector Quantizing Variational Autoencoder

Accurate, high-throughput, and unbiased analysis of plutonium oxide particles is necessary for analysis of the underlying phenomena associated with the process parameters involved in its synthesis. Compared to qualitative and taxonomic descriptors, quantitative descriptors of particle morphology using scanning electron microscopy (SEM) have shown success in analyzing process parameters of uranium oxides. Among other candidates, a neural network called Vector Quantized Variational Autoencoder (VQ-VAE) has shown the ability to quantitatively describe particle morphology with >85% accuracy from uranium oxide processing routes. In this study, we utilized VQ-VAE to analyze synthesized plutonium dioxide (PuO2) particles to investigate the underlying phenomena and predict its process parameters. The surface morphology of PuO2 powders calcined from plutonium oxalates precipitated under varying synthesis conditions related to concentrations, temperature, addition and digestion times, precipitant feed, and strike order was performed using SEM. A pipeline was developed to extract and quantify useful images of individual particles with VQ-VAE which was followed by further reduction of the dimensionality of the feature space using a bottlenecking neural network fit to perform multiple classification tasks simultaneously. The reduced feature space could predict process parameters for a single particle with >80% accuracy for some parameters. It also demonstrated potential in successfully grouping particles with similar surface morphology. The clustering and classification results revealed valuable information regarding the chemical process parameters that predominantly influence the PuO2 particle morphology, which includes strike order and oxalic acid feed stock, respectively. Repeating the analysis with multiple particles improved the classification accuracy of each process parameter compared to results yielded using single particle, with statistically significant results being yielded with as few as four particles.

Multiclass classification of environmental chemical stimuli from unbalanced plant electrophysiological data.

Plant electrophysiological response contains useful signature of its environment and health which can be utilized using suitable statistical analysis for developing an inverse model to classify the stimulus applied to the plant. In this paper, we have presented a statistical analysis pipeline to tackle a multiclass environmental stimuli classification problem with unbalanced plant electrophysiological data. The objective here is to classify three different environmental chemical stimuli, using fifteen statistical features, extracted from the plant electrical signals and compare the performance of eight different classification algorithms. A comparison using reduced dimensional projection of the high dimensional features via principal component analysis (PCA) has also been presented. Since the experimental data is highly unbalanced due to varying length of the experiments, we employ a random under-sampling approach for the two majority classes to create an ensemble of confusion matrices to compare the classification performances. Along with this, three other multi-classification performance metrics commonly used for unbalanced data viz. balanced accuracy, F1-score and Matthews correlation coefficient have also been analyzed. From the stacked confusion matrices and the derived performance metrics, we choose the best feature-classifier setting in terms of the classification performances carried out in the original high dimensional vs. the reduced feature space, for this highly unbalanced multiclass problem of plant signal classification due to different chemical stress. Difference in the classification performances in the high vs. reduced dimensions are also quantified using the multivariate analysis of variance (MANOVA) hypothesis testing. Our findings have potential real-world applications in precision agriculture for exploring multiclass classification problems with highly unbalanced datasets, employing a combination of existing machine learning algorithms. This work also advances existing studies on environmental pollution level monitoring using plant electrophysiological data.

Fog-cloud based intrusion detection system using Recurrent Neural Networks and feature selection for IoT networks

Deep learning (DL) techniques are being widely researched for their effectiveness in detecting cyber intrusions against the Internet of Things (IoT). Time sensitive Critical Infrastructures (CIs) that rely on IoT require rapid detection of cyber intrusions close to the constrained devices in order to prevent service delays. Deep learning techniques perform better in detecting attacks compared to shallow machine learning algorithms and can be used for intrusion detection. However, communication overheads due to large volume of IoT data and computation requirements for deep learning models prevents effective application of deep learning models closer to the constrained devices. Existing IDS techniques are either based on shallow learning algorithms or not trained on relevant IoT datasets and furthermore not designed for distributed fog-cloud deployment. To counter these issues, we propose a novel fog-cloud based IoT intrusion detection framework which incorporates a distributed processing by splitting the dataset according to attack class and a feature selection step on time-series IoT data. This is followed by a deep learning Recurrent Neural Network (SimpleRNN and Bi-directional Long Short-Term Memory (LSTM)) for attack detection. The effectiveness of the proposed approach was evaluated using the high-dimensional BoT-IoT dataset which contains large volumes of realistic IoT attack traffic. Results show that feature selection methods significantly reduced the dataset size by 90% under the computation requirements without compromising on the attack detection ability. The models built on reduced dataset achieved higher recall rate compared to models trained on full feature set without loosing class differentiation ability. The SimpleRNN and Bi-LSTM models also did not suffer any underfitting or overfitting with the reduced feature space. The proposed deep learning based IoT intrusion detection framework is suitable for fog-cloud based deployment and can scale well even with large volumes of IoT data.

A Dual-Staged heterogeneous stacked ensemble model for gender recognition using speech signal

Gender classification is one of the most popular topics in the field of machine learning. It can be done with images or sounds. With the aim of developing a low-complexity and highly accurate gender recognition model, this paper proposes a Dual-staged Heterogeneous Stacked Ensemble Model (DH-SEM). In the proposed ensemble model, the classification problem is addressed by the amalgamation of three supervised base learners at stage-1 and random forest as a Meta learner in stage-2. The proposed work uses a speech dataset compiled from various sources (Harvard-Haskins, Carnegie Mellon and McGill University) and consists of statistical features. Rigorous feature engineering has been performed to select the most discriminative information from the dataset. For gender recognition, the DH-SEM method adopts the reduced feature space. The gender recognition accuracy of the proposed DH-SEM model is 99.36%, which is highest as compared to the state-of-the-art methods. The robustness of the proposed technique is also validated by other performance evaluation metrics and receiver operating characteristics.

Exploration of coarse-to-fine FTIR feature extraction for handlebar grip materials: A common path for forensic examination of microtraces

Microtraces are capable of transfer (directly or via some intermediary object) between the items involved in criminal activities. They are inevitably encountered in various crimes and help establish the chain of evidence. Due to the high complexity and diversity of microtraces, it is necessary to build a common approach for different types of microtraces, in order to improve the efficiency of forensic science. Therefore, we chose handlebar grip materials from electric bikes as a typical microtrace, collected 41 highly similar samples with the purpose of discrimination, and achieved the database through Fourier transform infrared spectroscopy (FTIR). Firstly, a forensic strategy for examination of handlebar grip materials was built based on the traces, which was investigated by the manufacturing process and proved by FTIR spectra. Secondly, a coarse-to-fine approach for feature space reduction was performed by a combination of forensic expertise and principal component analysis (PCA). Further, the idea of hypothesis testing was introduced to this project, in order to filter out the possible non-informative variability over the spectral range. Feature annealed independent rules (FAIR) and higher criticism thresholding (HC) were combined with Fisher’s discriminant (LDA) respectively to train the classifier to achieve a coarse-to-fine extraction of features automatically. The performance of FAIR-LDA and HC-LDA was discussed. The results indicated that individual characteristics of handlebar grip materials were presented by weak signal-to-noise ratio in FTIR spectra, which cannot be fully explored by a semi-automatic method based on expert experience. Compared to the coarse extraction of FTIR features by FAIR, HC could filter 99% of invalid features, thus identifying those ‘weak but important’ signals more effectively and significantly. 87.7% of the analysis by HC-LDA can achieve accurate discrimination, which has increased 22.5% and 3.1% from that of LDA and FAIR-LDA respectively. HC integrated with LDA improves the accuracy of individual discrimination by its sparse sensitive discriminant features extraction. The present work demonstrates an accurate method for FTIR feature extraction as well as a scientific basis for forensic examination of microtraces.

A Novel Methodology for Classifying EMG Movements Based on SVM and Genetic Algorithms.

Electromyography (EMG) processing is a fundamental part of medical research. It offers the possibility of developing new devices and techniques for the diagnosis, treatment, care, and rehabilitation of patients, in most cases non-invasively. However, EMG signals are random, non-stationary, and non-linear, making their classification difficult. Due to this, it is of vital importance to define which factors are helpful for the classification process. In order to improve this process, it is possible to apply algorithms capable of identifying which features are most important in the categorization process. Algorithms based on metaheuristic methods have demonstrated an ability to search for suitable subsets of features for optimization problems. Therefore, this work proposes a methodology based on genetic algorithms for feature selection to find the parameter space that offers the slightest classification error in 250 ms signal segments. For classification, a support vector machine is used. For this work, two databases were used, the first corresponding to the right upper extremity and the second formed by movements of the right lower extremity. For both databases, a feature space reduction of over 65% was obtained, with a higher average classification efficiency of 91% for the best subset of parameters. In addition, particle swarm optimization (PSO) was applied based on right upper extremity data, obtaining an 88% average error and a 46% reduction for the best subset of parameters. Finally, a sensitivity analysis was applied to the characteristics selected by PSO and genetic algorithms for the database of the right upper extremity, obtaining that the parameters determined by the genetic algorithms show greater sensitivity for the classification process.

Feature extended energy landscape model for interpreting coercivity mechanism

Coercivity is expressed as a complex correlation between magnetisation and microstructures. However, owing to multiple intrinsic origins, coercivity has not been fully understood in the framework of the conventional Ginzburg–Landau theory. Here, we use machine learning to draw a realistic energy landscape of magnetisation reversal to consider missing parameters in the Ginzburg–Landau theory. The energy landscape in the magnetisation reversal process is visualised as a function of features extracted via machine learning; the correlation between the reduced feature space and hysteresis loop is assigned. Features in the lower dimension dataset strongly correlate with magnetisation and are embedded with morphological information. We analyse the energy landscape for simulated and experimental magnetic domain structures; a similar trend is observed. The landscape map enables visualisation of the energy of the system and coercivity as a function of feature space components.

Hyperparameter optimized classification pipeline for handling unbalanced urban and rural energy consumption patterns

Energy consumer locations are required for framing effective energy policies. However, due to privacy concerns, it is becoming increasingly difficult to obtain the locational data of the consumers. Machine learning (ML) based classification strategies can be used to find the locational information of the consumers based on their historical energy consumption patterns. The ML methods in this paper are applied to the Residential Energy Consumption Survey 2009 dataset. In this dataset, the number of consumers in the urban area is higher than the rural area, thus making the classification problem unbalanced. The unbalanced classification problem has been solved in original and transformed or reduced feature space using Monte Carlo based under-sampling of the majority class datapoints. The hyperparameters for each classification algorithm family is represented as an optimized pipeline, obtained using the genetic programming (GP) optimizer. The classification performance metrics are then obtained for different algorithm families on the original and transformed feature spaces. Performance comparisons have been reported using univariate and bivariate distributions of the classification metrics viz. accuracy, geometric mean score (GMS), F1 score, precision, area under the curve (AUC) of receiver operator characteristics (ROC). The energy policy aspects for the urban and rural residential consumers based on the classification results have also been discussed.

A High Throughput Experimentation & Material Informatics Platform for the Discovery of Molten Salt Reactor Candidate Structural Materials

ASME Sec(III) Div(5) code-certified structural alloys for Molten (halide) Salt Reactors (MSRs) such as 800H, SS316, and IN617 have significant Cr content. This makes them readily-degradable in molten halide salts due to the thermodynamic favorability of soluble chromium halide formation. It is therefore imperative to discover alloys with corrosion resistance that exceeds those of currently-certified alloys, which also possess the necessary high hardness and irradiation-resistance at reactor operating temperatures as a prerequisite for the licensing and deployment of MSRs.The largely-unexplored quasi-infinite quarternary FeCrMnNi High Entropy Alloy space shows promise to yield desired alloys. We present a high-throughput (HTP) process demonstrating a turnaround time from fabrication to corrosion-testing and analysis of 1 week for 25 samples. 70 alloy samples of 1cm2 were each corrosion-tested on 0.3g salt droplets in isolated corrosion environments. The HTP platform includes the development of an in-situ high-temperature electrochemical sensor system capable of automating the analysis of dissolved corrosion product analytes in the salt.The results of the 70 corrosion tests were used to train a Machine Learning model to predict corrosion performance metrics (eg. Elemental corrosion concentration into salt) based on an input vector parametrizing the physical properties (“descriptors”) of the alloys in the sample set. The model was tested on 20 additional samples and demonstrated conservative predictive capability as well as facilitated the reduction of input feature space from 62 dimensions to 4 dimensions with negligible loss in predictive accuracy.

Mapping spatial distribution of crop residues using PRISMA satellite imaging spectroscopy

ABSTRACT Non-photosynthetic vegetation (NPV) plays a key role in soil conservation, which in turn is important in sustainable agriculture and carbonfarming. For mapping NPV image spectroscopy proved to outperform multispectralsensors. PRISMA (PRecursore IperSpettrale della Missione Applicativa) is theforerunner of a new era of hyperspectral satellite missions, providing theproper spectral resolution for NPV mapping. This study takes advantage fromboth spectroscopy and machine-learning techniques. Exponential GaussianOptimization was used for modelling known absorption bands (cellulose-lignin,pigments, water content and clays), resulting in a reduced feature space, whichis split by a decision tree (DT) for mapping different field conditions (emerging,green and standing dead vegetation, crop residue and bare soil). DT trainingand validation exploited reference data, collected during PRISMA overpasses on a large farmland. Mapping results are accurate both at pixel and parcel level (O.A.> 90%; K > 0.9). Field status and crop rotation trajectories through timeare derived by processing 12 images over 2020 and 2021. Results proved thatPRISMA data are suitable for mapping field conditions at parcel scale with highconfidence level. This is important in the perspective of other hyperspectral missions and is a premise toward quantitative estimates of NPV biophysical variable.

Constraint-Induced Symmetric Nonnegative Matrix Factorization for Accurate Community Detection

As a fundamental characteristic of an undirected network, community reveals its networking organization and functional mechanisms, making community detection be a highly-interesting issue in network representation learning. With great interpretability, a symmetric and nonnegative matrix factorization (SNMF)-based approach is frequently adopted to tackle this issue. However, it only adopts a unique feature matrix for describing the symmetry of an undirected network, which unfortunately results in a reduced feature space that evidently impairs its representation learning ability. Motivated by this discovery, this paper proposes a novel Constraint-induced Symmetric Nonnegative Matrix Factorization (C-SNMF) model that adopts three-fold ideas: a) Representing a target undirected network with multiple latent feature matrices, thus preserving its representation learning capacity; b) Incorporating a symmetry-regularizer into its objective function, which preserves the symmetry of the learnt low-rank approximation to the adjacency matrix, thereby making the resultant detector precisely illustrate the target network's symmetry; and c) Introducing a graph-regularizer that preserves local invariance of the network's intrinsic geometry into its learning objective, thus making the achieved detector well-aware of community structure within the target network. Note that the regularization coefficients are selected according to the modularity of the learnt community structure on the training data only, thereby greatly improving the achieved model's practical significance for real applications. Experimental results on six real-world networks demonstrate that the proposed C-SNMF model significantly outperforms the benchmarks and state-of-the-art models in achieving highly-accurate community detection results.

Passive User Authentication Utilizing Behavioral Biometrics for IIoT Systems

Passive authentication is of great importance for security guarantee in industrial Internet of Things (IIoT) systems. Based on the behavioral biometrics from sequential operation actions in IIoT, this article proposes a nonintrusive and passive authentication framework for continuous user authentication against the impersonation attack. We first provide experimental results to demonstrate the discriminability and stability for the intrinsic features of sequential operation actions, and then leverage the Kalman filtering and wavelet techniques for noise elimination and the singular value decomposition method for the dimensionality reduction of feature space. We further exploit the one-class classification technique to formulate the authentication decision process as a hidden Markov model (HMM). Extensive experiments are conducted to illustrate the authentication performance of the passive authentication framework in terms of the false acceptance rate, false rejection rate, and equal-error rate. We also investigate the related authentication efficiency issues in terms of the usability to the operation-action sequence length, the scalability to the number of features and user space, and the sensitivity to the operation action features.

A Reinforcement Learning Application for Optimal Fluid and Vasopressor Interventions in Septic ICU Patients.

Sepsis is one of the leading causes of death in ICU and its timely recognition and management are of primary importance. Resuscitation from hypotension in patients with sepsis is one of the first challenges that require fluid and/or vasopressor administrations. Unfortunately, clinical guidelines provide only indications of the strategy that should be adopted in this critical population but personalized strategies are still missing. In this study, we propose a comparative analysis of reinforcement learning applications on ICU data collected in the electronic health records and publicly available within the MIMIC-III database. The ultimate goal of the study is to estimate the optimal fluid and vasopressor administrations. Results show that, after the use of principal component analysis for reducing feature space dimensionality, model performances increased, thus suggesting that additional preprocessing strategies might be used for both reducing the computational cost and refining model performances. Clinical relevance In a context where clinical guidelines are not able to provide the best treatment strategies at a patient level, reinforcement learning applications trained on retrospectively collected data may be used for developing models able to suggest to clinicians the optimal dosage of fluids and/or vasopressors in order to improve 90-day patients' survival.

Feature Selection Methods Comparison for EEG-based Classifier Constructed Using Discrete Wavelet Transform Features

The paper presents the results of a study in the application of electroencephalography (EEG) for user authentication using discrete wavelet transform. The Leipzig Study for Mind-Body-Emotion Interactions dataset (LEMON) was used. Mean value, standard deviation, and root mean square value are used as features. Feature selection methods based on correlation, on mutual information, and on the χ2 criterion are used for reduce feature space. The SVM model is used for classification. The efficiency of constructed classifier has been tested using cross-validation procedure. Classifier built on feature reduced data via mutual information criteria have improved accuracy (97.4%) with feature space nearly halved (183 features) compared to baseline classifier.

IoT Based Indoor and Outdoor Localization Framework with WI-FI Fingerprinting Based on Scalable Resnet Models

A scalable indoor localization technique is a vital technology for future large-scale location-aware services covering a complex of multi-story buildings. Our research on the usage of ResNet for scalable building/floor categorization and floor-level position estimation based on Wi-Fi fingerprinting is presented in this publication. Building and floor-level coordinates are estimated using our new ResNet architecture, which utilizes a stacked autoencoder to reduce feature space and a feed-forward classifier to classify multiple labels of building/floor/location. This architecture is the foundation for our multi-building and multi-floor indoor localization system based on Wi-Fi fingerprinting. On the Jaume I University (UJI) campus in Spain, we test the accuracy of building/floor estimation and floor-level coordinates estimation for three buildings with four or five stories each. ResNet-based indoor localization using a single ResNet has been proven to be feasible, with results that are close to the state of the art. One ResNet is all that is needed in order for the proposed indoor localization system based on Wi-Fi fingerprinting to function at levels close to the current state of the art, allowing it to be implemented with less complexity and less energy consumption on mobile devices.

Malaria parasite classification framework using a novel channel squeezed and boosted CNN.

Malaria is a life-threatening infection that infects the red blood cells and gradually grows throughout the body. The plasmodium parasite is transmitted by a female Anopheles mosquito bite and severely affects numerous individuals within the world every year. Therefore, early detection tests are required to identify parasite-infected cells. The proposed technique exploits the learning capability of deep convolutional neural network (CNN) to distinguish the parasite-infected patients from healthy individuals using thin blood smear. In this regard, the detection is accomplished using a novel STM-SB-RENet block-based CNN that employs the idea of split-transform-merge (STM) and channel squeezing-boosting (SB) in a modified fashion. In this connection, a new convolutional block-based STM is developed, which systematically implements region and edge operations to explore the parasitic infection pattern of malaria related to region homogeneity, structural obstruction and boundary-defining features. Moreover, the diverse boosted feature maps are achieved by incorporating the new channel SB and transfer learning (TL) idea in each STM block at abstract, intermediate and target levels to capture minor contrast and texture variation between parasite-infected and normal artifacts. The malaria input images for the proposed models are initially transformed using discrete wavelet transform to generate enhanced and reduced feature space. The proposed architectures are validated using hold-out cross-validation on the National Institute of Health Malaria dataset. The proposed methods outperform training from scratch and TL-based fine-tuned existing techniques. The considerable performance (accuracy: 97.98%, sensitivity: 0.988, F-score: 0.980 and area under the curve: 0.996) of STM-SB-RENet suggests that it can be utilized to screen malaria-parasite-infected patients. Graphical Abstract.

Non-linear Embedding Methods for Identifying Similar Brain Activity in 1 Million iEEG Records Captured From 256 RNS System Patients.

Finding electrophysiological features that are similar across patients with epilepsy may facilitate identifying treatment options for one patient that worked in patients with similar brain activity patterns. Three non-linear iEEG (intracranial electroencephalogram) embedding methods of finding similar cross-patient iEEG records in a large iEEG dataset were developed and compared. About 1 million iEEG records from 256 patients with drug-resistant focal onset seizures who were treated in prospective trials of the RNS System were used for analyses. Data from 200, 25, and 31 patients were randomly selected to be in the train, validation, and test datasets. In method 1, ResNet50 convolutional neural network (CNN) model pre-trained on the ImageNet dataset was used for extracting feature maps from spectrogram images (ImageNet-ResNet) of iEEG records. In method 2, ResNet50 custom trained on an iEEG classification task using ~138,000 manually labeled iEEG records was used as the feature extractor (ESC-ResNet). Feature maps were passed through dimensionality reduction and k nearest neighbors were found in the reduced feature space. In method 3, a 256 dimensional iEEG embedding space was learned via contrastive learning by training a ResNet50 model with triplet training sets generated using within-patient iEEG clustering (CL-ResNet). All three methods had comparable performance when identifying iEEG records from the search dataset similar to test iEEG records of baseline (non-seizure) and interictal spiking activity. Epileptic interictal spikes are represented by vertical (broadband) edges in spectrogram images, and hence even generic features extracted using models trained on everyday images appear to be sufficient to represent iEEG records with similar levels of interictal spiking activity in close proximity. In the case of electrographic seizures, however, the ESC-ResNet model, identified cross-patient iEEG records with electrographic seizure morphology features that were most similar to the test iEEG records. For nuanced electrographic seizure iEEG representation learning, domain specific model training with manually generated labels had the advantage. Finally, representative iEEG records were selected from every patient using an unsupervised clustering method which effectively reduced the number of iEEG records in the search dataset from ~750,000 to 2,148, thus substantially reducing the time required for finding similar cross-patient iEEG records.