One-hot Encoding Research Articles

TactONet: Tactile Ordinal Network Based on Unimodal Probability for Object Hardness Classification

Hardness is one of the most critical tactile properties for robots to recognize objects. Machine learning methods have shown superior performance in object hardness classification. However, existing machine learning methods for tactile hardness classification cannot use the ordinal information between hardness classes because the one-hot encoding only cares about the correct class and ignores the inter-class relationship. To solve this problem, we propose to generalize the one-hot encoding using unimodal distributions including the Poisson and binomial distributions for tactile ordinal classification problems, resulting in two tactile ordinal networks (TacONet): TacONet-p and TacONet-b. Furthermore, we collect a tactile hardness dataset on the silicone samples with three different shapes (Shapes A, B, C), and each shape samples have thirteen hardness classes ranging from 0A (Shore A scale) to 60A at 5A intervals. We validate the resulting method for tactile hardness classification using a real robot. Experimental results demonstrate that compared with state-of-the-art methods, the proposed method achieves better classification performance in terms of accuracy and quadratic weighted kappa (QWK) on the tactile hardness dataset, reaching a classification accuracy up to 99.5% and a QWK up to 99.9% on Shape C. <i>Note to Practitioners</i>&#x2014;In the field of robotics tactile recognition, hardness classification is one of the most important and common tasks for robots to accurately recognize objects, particularly when the environment is dark or visual sensors are not working. In this paper, we propose a novel tactile ordinal network for tactile hardness classification tasks. The existing machine learning models for tactile hardness classification are trained by minimizing the cross-entropy loss between predicted vectors and one-hot encoding vectors of true classes, which makes the models only care about the correct classes and ignores the inter-class relationship of hardness classes. In other words, these models have the same probability to misclassify a hardness class with any other hardness class. To tackle this problem, we propose to generalize the one-hot encoding method using a unimodal distribution method to encode the true classes. The unimodal distribution encoding vectors can make the model learn the ordinal information between classes. It is proved that the proposed method is able to effectively improve the classification accuracy and QWK in a tactile hardness classification task.

Pushing ML Predictions Into DBMSs.

In the past decade, many approaches have been suggested to execute ML workloads on a DBMS. However, most of them have looked at in-DBMS ML from a training perspective, whereas ML inference has been largely overlooked. We think that this is an important gap to fill for two main reasons: (1) in the near future, every application will be infused with some sort of ML capability; (2) behind every web page, application, and enterprise there is a DBMS, whereby in-DBMS inference is an appealing solution both for efficiency (e.g., less data movement), performance (e.g., cross-optimizations between relational operators and ML) and governance. In this article, we study whether DBMSs are a good fit for prediction serving. We introduce a technique for translating trained ML pipelines containing both featurizers (e.g., one-hot encoding) and models (e.g., linear and tree-based models) into SQL queries, and we compare in-DBMS performance against popular ML frameworks such as Sklearn and ml.net. Our experiments show that, when pushed inside a DBMS, trained ML pipelines can have performance comparable to ML frameworks in several scenarios, while they perform quite poorly on text featurization and over (even simple) neural networks.

AI-Driven Solutions for Detecting and Mitigating Cyber Threats on Social Media Networks

Cybersecurity has emerged as a vital aspect of organisations’ operation because of the increased usage of technology and the internet, and an increase in callous and legal cybercrimes. Therefore, understanding and managing cyber threats has become an essential component of modern cyber architectures in order to provide protection and sustain technology assets and offerings against ever-involving cyber threats. The purpose of this article is to highlight the importance of vulnerability information analysis and cyberthreats in order to proactively comprehend cyber risks and anomalies and provide suitable mitigation techniques. The current research provides the prospect of an improved approach for responding to cyber threats on social media networks based on AI and machine learning. Leveraging the CIC-IDS2017 dataset. Data using text-based features, like user behaviour and network activity, are then preprocessed and feature-engineered, for instance, by means of one-hot encoding. Classification models such as LSTM, GNB, and LDA were used to group various cyber hazards, such as malware propagation, into distinct groups. Thus, the LSTM reveals its high potential in real-time threat identification in terms of an accuracy of 99.34%, recall of 99%, precision of 99.3%, and an F1-score of 99.34%. However, it reveals that GNB and LDA have lower accuracy and classification measures compared to other algorithms. In case of threat identification, the process of counteraction is also performative immediately as users are informed and accounts are blocked additionally to informing the admins. This framework provides a sound approach to improve the cybersecurity on social media websites.

Structural models of Mealy finite state machines detecting faults in control systems

The subject matter of this article is a control system for unmanned aerial vehicles (UAVs) whose mathematical model is a finite state machine (FSM). The goal is to develop FSM structural models that enable (1) detection of multiple faults of FSM elements caused by an electromagnetic pulse or laser beam, and (2) prevent negative impacts on the controlled object. The tasks to be solved are as follows: to develop FSM structural models to detect invalid input vector X for the whole FSM and in each state, to detect invalid output vector Y for the whole FSM, at each transition and in each state, invalid code of the present (current) state, invalid code of the next state, and invalid transition between states; to determine the possible causes of the faults, which can be the failure in the logic Φ of forming the code of the next state, the invalid input vector X, the failure in the feedback circuit, the failure in the logic Ψ of forming the output vector, the failure in the state register R, the failure in the wire between the FSM input and the input of the logic Ψ; development of a combined structural model for the detection of all listed faults with a minimum number of additional combinational circuits, as well as a structural model that combines all additional combinational circuits. The methods used are: the theory of finite state machines, structural models of FSMs, state encoding methods of FSMs, representation methods of FSMs, and Verilog hardware description language. The following results were obtained: (1) the Mealy FSM structural models were developed to detect all the above mentioned faults, (2) the combined FSM structural models were developed, and (3) the possible causes of faults detected by each FSM structural model were identified. Experimental studies have shown that for the presented FSM structural models, the area overhead averages 3-23%, for one-hot encoding of FSM states, and 2-8%, for binary encoding of FSM states. Conclusions. The scientific novelty of the obtained results consists in the following for the first time FSM faults that are not caused by radiation and cosmic rays but by an electromagnetic pulse affecting the control device are considered; the number of faults is not limited for the state codes as well as for the input and output vectors; the faults can be detected not only in the state register R but also in the input vector X, in the logic Φ of generating the next state code, in the logic Ψ of generating the output signals, and in the feedback circuit; the invalid transitions of FSMs and the transitions to invalid states are also detected; the proposed structural models not only detect FSM failures but also prevent their negative impact on the controlled object; combined structural models allow simultaneous detection of faults in all elements of the FSM. Future research will focus on developing structural models for correcting FSM failures.

An integrated data- and theory-driven crash severity model

For crash severity modeling, researchers typically view theory-driven models and data-driven models as different or even conflicting approaches. The reason is that the machine-learning models offer good predictability but weak interpretability, while the latter has robust interpretability but moderate predictability. In order to alleviate the tension between them, this study proposes an integrated data- and theory-driven crash-severity model, known as Embedded Fusion model based on Text Vector Representations (TVR-EF), by leveraging the complementary strengths of both. The model specification consists of two parts. (i) the data-driven component not only mitigate the deficiencies of traditional econometric models, where one-hot encoding is frequently used and makes it impossible to observe semantic relatedness between variable categories, but also enhances the interpretability for the relationship between crash severity and potential influencing factors using the learned embedding weight matrix. (ii) In the theory-driven component, the multinomial logit model is implemented as a 2D-Convolutional Neural Network (2D-CNN) to increase flexibility and decrease dependency on prior knowledge for different crash-severity outcomes. A crash dataset from Guangdong Province, China, is utilized to estimate the TVR-EF model, which is then benchmarked against two traditional econometric models and three widely used machine-learning models. Results indicate that TVR-EF model does not only improve the predictive performance but also makes it easier to interpret.

Puzzle—Integer Linear Programming: Spreadsheet Solver Excellence Without Excel

Because of the limitations on problem size with the built-in optimization solver in Microsoft Excel, other software applications are required to illustrate the power of integer linear programming in spreadsheets. The built-in linear optimization solver in LibreOffice Calc allows much larger problems. In this paper, we build a spreadsheet in LibreOffice Calc to solve sudoku problems and a simple generalization that further illustrates the power of integer linear programming in spreadsheets. The generalization we consider, n-doku, also introduces an interesting puzzle that includes, as a special case, a union of multiple sudoku puzzles. Besides further demonstrating the power of integer linear programming in spreadsheets, we use the n-doku solutions to develop optimization intuition and to, more generally, foster critical thinking. We use this spreadsheet in the classroom to transition from continuous to categorical decisions, emphasizing the idea of one-hot encoding. We also illustrate multiple optima and add preferences to sudoku puzzles as a way introduce multiobjective optimization. Supplemental Material: The online supplement is available at https://doi.org/10.1287/ited.2022.0068 .

A Smart Farm DNN Survival Model Considering Tomato Farm Effect

Recently, smart farming research based on artificial intelligence (AI) has been widely applied in the field of agriculture to improve crop cultivation and management. Predicting the harvest time (time-to-harvest) of crops is important in smart farming to solve problems such as planning the production schedule of crops and optimizing the yield and quality. This helps farmers plan their labor and resources more efficiently. In this paper, our concern is to predict the time-to-harvest (i.e., survival time) of tomatoes on a smart farm. For this, it is first necessary to develop a deep learning modeling approach that takes into account the farm effect on the tomato plants, as each farm has multiple tomato plant subjects and outcomes on the same farm can be correlated. In this paper, we propose deep neural network (DNN) survival models to account for the farm effect as a fixed effect using one-hot encoding. The tomato data used in our study were collected on a weekly basis using the Internet of Things (IoT). We compare the predictive performance of our proposed method with that of existing DNN and statistical survival modeling methods. The results show that our proposed DNN method outperforms the existing methods in terms of the root mean squared error (RMSE), concordance index (C-index), and Brier score.

Cap-DiBiL: an automated model for crop water requirement prediction and suitable crop recommendation in agriculture

In this technological era, several approaches used to provide the information about suitable crop recommendation means which is crop is suitable for soil. Some of approaches depends on the IoT smart agricultural-devices to gather information from surrounding area. However, several collection of data are used to predict the crops details but it not efficient to provide better performance. Therefore, the proposed model uses various techniques to improve the performance efficiently. Some steps involved in the proposed model as data pre-processing, feature extraction, feature selection, water requirement prediction and recommendation. Initially, the collected IoT data from dataset are pre-processed using data normalization, missing value imputation and one-hot encoding. Then, extract the features from pre-processed data using Gated Residual autoencoder (GRA) model, whereas optimal features are selected using Chaotic Northern Goshawk Optimization (ChaNgo) algorithm. Based on the farmland details, the crop water requirement prediction and suitable crop recommendation due to the market price are carried out using a novel hybrid deep learning model called Channel capsule-assisted stacked dilated Bi-LSTM (Cap-DiBiL). The channel capsule network predicts the crop water requirement and stacked dilated Bi-LSTM is used for suitable crop recommendations such as millets, rice and other crops. Then the proposed model analyses the performance and compares it with several existing techniques to prove the proposed model’s enhancement. The proposed model improved the accuracy as 98.18% for predicting the crop water requirement and crop recommendation. The performance of proposed model for Precision, Recall and F1 score also enhanced as 98.31%, 98.18% and 98.20%.

A Hybridized- Logistic Regression and Deep Learning-based Approaches for Precise Anomaly Detection in Cloud

Anomaly Detection plays a pivot role in determining the abnormal behaviour in the cloud domain. The objective of the manuscript is to present two approaches for Precise Anomaly Detection Approaches by hybridizing RBM with LR and SVM models. The various phases in the present approach are (a) Data collection (b) Pre-processing and normalization; OneHot Encoder for converting categorical values to numerical values followed by encoding the binary features through normalization (c) training the data (d) Building the Feedforward Deep Belief Network (EDBN) using hybridizing Restricted Boltzmann Machine (RBM) with Logistic Regression (LR) and Support Vector Machine (SVM); In the first approach, RBM model is trained through unsupervised pre-training followed by fine-tuning using LR model. In the later approach, RBM model is trained through unsupervised pre-training followed by fine-tuning using SVM model; both the approaches adopt unsupervised pre-training followed by supervised-fine-tuning operations (e) Model Evaluation using the significant parameters such as Precision, Recall, Accuracy, F1-score and Confusion Matrix. The experimental evaluations concluded the effective anomaly detection techniques by integrating the RBM with LR and SVM for capturing the intricate patterns and complex relationships among the data. The proposed approaches paves a path to improved anomaly detection technique, thereby enhancing the security features and anomaly monitoring systems across distinct domains.

A novel attention-based feature learning and optimal deep learning approach for network intrusion detection

Rapid technological advances and network progress has occurred in recent decades, as has the global growth of services via the Internet. Consequently, piracy has become more prevalent, and many modern systems have been infiltrated, making it vital to build information security tools to identify new threats. An intrusion detection system (IDS) is a critical information security technology that detects network fluctuations with the help of machine learning (ML) and deep learning (DL) approaches. However, conventional techniques could be more effective in dealing with advanced attacks. So, this paper proposes an efficient DL approach for network intrusion detection (NID) using an optimal weight-based deep neural network (OWDNN). The network traffic data was initially collected from three openly available datasets: NSL-KDD, CSE-CIC-IDS2018 and UNSW-NB15. Then preprocessing was carried out on the collected data based on missing values imputation, one-hot encoding, and normalization. After that, the data under-sampling process is performed using the butterfly-optimized k-means clustering (BOKMC) algorithm to balance the unbalanced dataset. The relevant features from the balanced dataset are selected using inception version 3 with multi-head attention (IV3MHA) mechanism to reduce the computation burden of the classifier. After that, the dimensionality of the selected feature is reduced based on principal component analysis (PCA). Finally, the classification is done using OWDNN, which classifies the network traffic as normal and anomalous. Experiments on NSL-KDD, CSE-CIC-IDS2018 and UNSW-NB15 datasets show that the OWDNN performs better than the other ID methods.

Development of an Improved One-Hot Encoding Method for Bubbly Flow Image Prediction Generation under Continuous Superficial Gas Velocities

Development of an Improved One-Hot Encoding Method for Bubbly Flow Image Prediction Generation under Continuous Superficial Gas Velocities

A Machine Learning Approach for Detecting Network Threats

In the context of the rapid expansion of computer networks and the increasing reliance on digital communication, ensuring the security and integrity of network systems has become a paramount concern. Detecting network threats play a vital role in safeguarding networks by identifying potential malicious activities and unauthorized access attempts. In this research paper, we present a comprehensive study that explores the efficacy of different machine learning classifiers for network threat detection. The study utilizes a publicly available dataset containing network traffic data, encompassing various network protocols and attack. The dataset is pre-processed to convert categorical variables into numerical form using one-hot encoding. Four popular classifiers are employed in this study: Support Vector Machine (SVM), Decision Tree Classifier (DTC), K-Nearest Neighbors (KNN), and Bernoulli Naive Bayes (BNB). The classifiers are trained on the pre-processed training data, and their performance is evaluated using accuracy metrics, classification reports, and confusion matrices. Results offer insightful information on the weaknesses and weaknesses of each classifier for detecting network anomaly. The findings demonstrate that the DTC classifier exhibits high accuracy and robustness in detecting network anomalies. The KNN classifier also achieves competitive results but may suffer from scalability issues with large datasets. The SVM classifier demonstrates satisfactory performance, while the BNB classifier, which converts numeric features to binary form, exhibits relatively lower accuracy. Keywords - Network threat Detection, Machine Learning Classifiers, Network Security, Anomaly Detection, Cyber Threats.

NeuroCNN_GNB: an ensemble model to predict neuropeptides based on a convolution neural network and Gaussian naive Bayes.

Neuropeptides contain more chemical information than other classical neurotransmitters and have multiple receptor recognition sites. These characteristics allow neuropeptides to have a correspondingly higher selectivity for nerve receptors and fewer side effects. Traditional experimental methods, such as mass spectrometry and liquid chromatography technology, still need the support of a complete neuropeptide precursor database and the basic characteristics of neuropeptides. Incomplete neuropeptide precursor and information databases will lead to false-positives or reduce the sensitivity of recognition. In recent years, studies have proven that machine learning methods can rapidly and effectively predict neuropeptides. In this work, we have made a systematic attempt to create an ensemble tool based on four convolution neural network models. These baseline models were separately trained on one-hot encoding, AAIndex, G-gap dipeptide encoding and word2vec and integrated using Gaussian Naive Bayes (NB) to construct our predictor designated NeuroCNN_GNB. Both 5-fold cross-validation tests using benchmark datasets and independent tests showed that NeuroCNN_GNB outperformed other state-of-the-art methods. Furthermore, this novel framework provides essential interpretations that aid the understanding of model success by leveraging the powerful Shapley Additive exPlanation (SHAP) algorithm, thereby highlighting the most important features relevant for predicting neuropeptides.

Emotional Text Detection dengan Long Short Term Memory (LSTM)

Emotional Text Detection is a technique in natural language processing that aims to identify the emotions contained in conversations or text messages. The LSTM (Long Short-Term Memory) method is one of the techniques used in natural language processing to model and predict sequential data. In this study, we propose the use of the LSTM method for emotion detection in conversation. The dataset used is a conversational dataset that contains positive, negative, and neutral emotions. We process datasets using data pre-processing techniques such as tokenization, data cleansing and one-hot encoding. Then, we train the LSTM model on the processed dataset and obtain evaluation results using accuracy metrics. The experimental results show that the LSTM model can be used to detect emotions in conversation with a good degree of accuracy. In addition, we also conducted an analysis on the prediction results of the model and showed that the LSTM model can correctly identify emotions. In conclusion, the LSTM method can be used to detect emotions in conversation with a good degree of accuracy. This method can be used to improve user experience in chat applications and increase the effectiveness of human and machine interactions.

S3-VAE: A novel Supervised-Source-Separation Variational AutoEncoder algorithm to discriminate tumor cell lines in time-lapse microscopy images

The derivation of input–output relationships in deep learning architectures is mostly a black-box process, in which uninformative or confounding factors might bias the classification results, with no clue for the user. However, the analysis of living cells requires extracting latent space representations with interpretable meaning, which can be investigated not only for classification but mainly for understanding purposes. Using the properties of variational autoencoders in deriving unsupervised regular space representations, we propose a novel Supervised-Source-Separation Variational Autoencoder (S3-VAE) algorithm capable of guiding the encoding process of images toward a more effective class-separability. The scope has been reached, under the assumption of full-covariance Gaussian posterior, by introducing a term in the cost function that forces class-dependent sources of variations toward a one-hot encoding. The proposed approach has been designed for an automatic platform for single-cell biological investigations based on time-lapse microscopy, in which a dedicated video processing pipeline has been designed, as an additional contribution, to reduce the effects of confounding factors on the S3-VAE representations. The results obtained in a classification scenario with artificially generated phantom cells and with experimental data of human prostate cells, in non-neoplastic, neoplastic and metastatic neoplastic conditions, show that S3-VAE-based latent representations allow for investigation and recognition of the morphological differences among classes. Comparative analysis with other benchmark deep learning architectures demonstrates the effectiveness of the proposed approach and shows that S3-VAE is capable of achieving high performance in discriminating cell lines based on phenotypic variations.

Load Forecasting Based on LVMD-DBFCM Load Curve Clustering and the CNN-IVIA-BLSTM Model

Power load forecasting plays an important role in power systems, and the accuracy of load forecasting is of vital importance to power system planning as well as economic efficiency. Power load data are nonsmooth, nonlinear time-series and “noisy” data. Traditional load forecasting has low accuracy and curves not fitting the load variation. It is not well predicted by a single forecasting model. In this paper, we propose a novel model based on the combination of data mining and deep learning to improve the prediction accuracy. First, data preprocessing is performed. Second, identification and correction of anomalous data, normalization of continuous sequences, and one-hot encoding of discrete sequences are performed. The load data are decomposed and denoised using the double decomposition modal (LVMD) strategy, the load curves are clustered using the double weighted fuzzy C-means (DBFCM) algorithm, and the typical curves obtained are used as load patterns. In addition, data feature analysis is performed. A convolutional neural network (CNN) is used to extract data features. A bidirectional long short-term memory (BLSTM) network is used for prediction, in which the number of hidden layer neurons, the number of training epochs, the learning rate, the regularization coefficient, and other relevant parameters in the BLSTM network are optimized using the influenza virus immunity optimization algorithm (IVIA). Finally, the historical data of City H from 1 January 2016 to 31 December 2018, are used for load forecasting. The experimental results show that the novel model based on LVMD-DBFCM load c1urve clustering combined with CNN-IVIA-BLSTM proposed in this paper has an error of only 2% for electric load forecasting.

One-Hot Graph Encoder Embedding.

In this paper we propose a lightning fast graph embedding method called one-hot graph encoder embedding. It has a linear computational complexity and the capacity to process billions of edges within minutes on standard PC - making it an ideal candidate for huge graph processing. It is applicable to either adjacency matrix or graph Laplacian, and can be viewed as a transformation of the spectral embedding. Under random graph models, the graph encoder embedding is approximately normally distributed per vertex, and asymptotically converges to its mean. We showcase three applications: vertex classification, vertex clustering, and graph bootstrap. In every case, the graph encoder embedding exhibits unrivalled computational advantages.

An adaptive Physics-based feature engineering approach for Machine Learning-assisted alloy discovery

This study investigated the importance of integrating a physics-based perspective in feature engineering for machine learning applications in material science problems. Specifically, we studied the encoding of the variable of temper designation, which contains critical alloy manufacturing information and is commonly included as an important feature for predicting alloy properties in machine learning models. Popular encoding methods such as one-hot encoding or ordinal encoding neglect the physics-based mechanism of temper designations by considering them either totally independent or sequentially ordinal. Following the underlying physical mechanism of the temper designation variable, we propose an adaptive encoding method for temper designations by first decomposing them into categorical and numerical subunits that can be more properly encoded by one-hot encoding and ordinal encoding respectively. The proposed adaptive encoding method is investigated on two independent aluminum alloy datasets consisting of various mechanical and technological properties. Our investigations showed that the proposed adaptive encoding method outperforms traditional encoding methods in the prediction of both mechanical and technological properties. As a general encoding method, this adaptive encoding method can be applied to a variety of decomposable variables to help advance machine-learning-assisted alloy design.

Categorical Feature Encoding Techniques for Improved Classifier Performance when Dealing with Imbalanced Data of Fraudulent Transactions

Fraudulent transaction data tend to have several categorical features with high cardinality. It makes data preprocessing complicated if categories in such features do not have an order or meaningful mapping to numerical values. Even though many encoding techniques exist, their impact on highly imbalanced massive data sets is not thoroughly evaluated.&#x0D; Two transaction datasets with an imbalance lower than 1\% of frauds have been used in our study. Six encoding methods were employed, which belong to either target-agnostic or target-based groups. The experimental procedure has involved the use of several machine-learning techniques, such as ensemble learning, along with both linear and non-linear learning approaches.&#x0D; Our study emphasizes the significance of carefully selecting an appropriate encoding approach for imbalanced datasets and machine learning algorithms. Using target-based encoding techniques can enhance model performance significantly. Among the various encoding methods assessed, the James-Stein and Weight of Evidence (WOE) encoders were the most effective, whereas the CatBoost encoder may not be optimal for imbalanced datasets. Moreover, it is crucial to bear in mind the curse of dimensionality when employing encoding techniques like hashing and One-Hot encoding.

Protein Fitness Prediction Is Impacted by the Interplay of Language Models, Ensemble Learning, and Sampling Methods.

Advances in machine learning (ML) and the availability of protein sequences via high-throughput sequencing techniques have transformed the ability to design novel diagnostic and therapeutic proteins. ML allows protein engineers to capture complex trends hidden within protein sequences that would otherwise be difficult to identify in the context of the immense and rugged protein fitness landscape. Despite this potential, there persists a need for guidance during the training and evaluation of ML methods over sequencing data. Two key challenges for training discriminative models and evaluating their performance include handling severely imbalanced datasets (e.g., few high-fitness proteins among an abundance of non-functional proteins) and selecting appropriate protein sequence representations (numerical encodings). Here, we present a framework for applying ML over assay-labeled datasets to elucidate the capacity of sampling techniques and protein encoding methods to improve binding affinity and thermal stability prediction tasks. For protein sequence representations, we incorporate two widely used methods (One-Hot encoding and physiochemical encoding) and two language-based methods (next-token prediction, UniRep; masked-token prediction, ESM). Elaboration on performance is provided over protein fitness, protein size, and sampling techniques. In addition, an ensemble of protein representation methods is generated to discover the contribution of distinct representations and improve the final prediction score. We then implement multiple criteria decision analysis (MCDA; TOPSIS with entropy weighting), using multiple metrics well-suited for imbalanced data, to ensure statistical rigor in ranking our methods. Within the context of these datasets, the synthetic minority oversampling technique (SMOTE) outperformed undersampling while encoding sequences with One-Hot, UniRep, and ESM representations. Moreover, ensemble learning increased the predictive performance of the affinity-based dataset by 4% compared to the best single-encoding candidate (F1-score = 97%), while ESM alone was rigorous enough in stability prediction (F1-score = 92%).