How do you deal with class imbalance or outliers when using cross-entropy loss for classification tasks?

To tackle class imbalance or outliers when using cross-entropy loss, it's essential to bring a human touch by thoughtfully adjusting the dataset. This can involve giving more weight to underrepresented classes, enriching the dataset with more examples of these classes, or even trimming down the overwhelming majority classes. By doing so, we ensure our model learns in a more balanced and equitable manner, reflecting the diversity and nuances of the real world.

💡 Class imbalance impacts classification models, skewing results towards the majority class, neglecting insights. 📊 Class Imbalance When the training data has unequal class representation, models may favor the dominant class, neglecting the minority. This misrepresentation can lead to biased predictions. 📉 Model Bias Imbalanced data leads to a biased model that underperforms on minority classes. The model's accuracy might appear high but fails in recognizing minority class instances. 🔄 Real Distribution Ensuring the training data reflects real-world distribution is crucial. Models trained on imbalanced data often struggle in real-world applications where all classes need equal attention.

Class imbalance is when classes in data are not represented equally. It can bias the model towards the majority class. To handle this, you can use techniques like oversampling the minority class, undersampling the majority class, or using a weighted loss function to give more importance to the minority class.

Understanding Class Imbalance In Classification Tasks. Class imbalance occurs when one class in a classification task has significantly fewer samples than the other classes. This can lead to biased models that are more inclined to predict the majority class, ignoring the minority class altogether. In the context of using cross-entropy loss for classification tasks, this imbalance can result in suboptimal model performance. When dealing with class imbalance, it is essential to address this issue to ensure that the model learns from all classes equally. One common approach is to use techniques such as oversampling or undersampling to balance out the dataset before training the model.

Address class imbalance by using weighted loss functions, oversampling minority classes, or employing data augmentation. Outliers can be managed with regularization techniques, outlier detection methods, or robust loss functions like focal loss to focus on hard-to-classify examples.

Class imbalance occurs when the number of examples in one or more classes (minority classes) is significantly lower than the number of examples in other classes (majority classes). This is common in real-world datasets, especially in areas like: Medical Diagnosis: Rare diseases might have far fewer examples than common conditions. Fraud Detection: Fraudulent transactions are usually much less frequent than legitimate ones. Customer Churn Prediction: Customers who churn are often a minority compared to those who stay.

-Use class weights: Assign higher weights to minority classes in the cross-entropy loss function to balance the impact of each class. -Apply data augmentation: Increase the representation of minority classes by augmenting the training data. -Implement resampling techniques: Either oversample the minority class or undersample the majority class to balance the dataset.

Resampling: Oversample: Create more minority class examples (synthetic/duplicate). Undersample: Reduce majority class examples. Weighted Loss: Assign higher weights to underrepresented classes in the loss function. Algorithm Choice: Use models robust to imbalance, like Random Forest or SMOTEboost. Outliers: Data Cleaning: Identify and remove extreme outliers if they're not representative. Robust Loss Functions: Use Median Absolute Error or Huber loss less sensitive to outliers. Outlier Detection: Train separate outlier detector and exclude detected points during training.

Class imbalance refers to mismatch between the number of samples representing each class of the dataset. This can be tackled by techniques like SMOTE where number of minority class samples are increased to create class balance. Data augmentation is also done by doing minor changes in the minority class samples and adding those samples under the same class. Eg: Adding noise, timeshift to audio signals or image invert, image transform to image dataset.

Addressing class imbalance is essential to ensure fair and accurate predictions across all classes. By implementing techniques such as oversampling, undersampling, or using algorithms like SMOTE, practitioners can rebalance the class distribution and enhance the model's ability to recognize patterns in all classes, ultimately improving the model's performance and predictive accuracy.

💡Addressing class imbalance is crucial for improving model performance and ensuring reliable predictions. 🐾Weighted Loss Weighted cross-entropy. Assigning different weights to each class can help balance the focus of the model on minority classes, improving accuracy. 📊Data Augmentation Data Resampling. Creating synthetic samples or modifying existing ones can balance class distribution. Techniques like oversampling minority classes are effective. 🔄Resampling Techniques Over- and Under-sampling. Adjusting the sample size of each class by either adding more minority samples or reducing majority samples balances class ratios. 📌 Addressing class imbalance ensures fairer, more accurate model predictions and overall better performance.

Class imbalance is a common problem in machine learning where the total number of a class of data (minority) is far less than the total number of another class of data (majority). This can lead to misleading classification accuracy. To handle this, you can use techniques like Resampling, Weighted Cross-Entropy Loss and Data Augmentation.

Techniques For Handling Class Imbalance In Cross-Entropy Loss. One common issue in classification tasks is class imbalance, where one class has significantly fewer samples than the others. This can lead to biased models that perform poorly on minority classes. When using cross-entropy loss for classification, there are several techniques to handle class imbalance. One approach is to use weighted cross-entropy loss, where the loss function assigns higher weights to minority class samples. This helps the model pay more attention to these samples during training and improve its ability to correctly classify them. Another technique is oversampling or undersampling, where you either duplicate minority class samples.

Oversampling: Random Oversampling: Duplicate instances from the minority class to balance the dataset. SMOTE (Synthetic Minority Over-sampling Technique): Creates synthetic samples for the minority class by interpolating between existing minority instances. ADASYN: Adaptively generates synthetic samples based on the density distribution of the minority class. Undersampling: Random Undersampling: Remove instances from the majority class randomly to balance the dataset. Tomek Links: Removes majority class instances that are close to minority class instances, improving the decision boundary. NearMiss: Selects majority class instances based on their distance to the nearest minority class instances.

-Use weighted cross-entropy loss: Assign different weights to classes to balance their impact in the loss function. -Apply data augmentation: Create synthetic samples or modify existing ones to increase the representation of minority classes. -Employ resampling techniques: Oversample minority classes or undersample majority classes to balance the dataset.

To handle class imbalance, you can try various techniques. One way is to balance the dataset by either oversampling the minority class (making more copies) or undersampling the majority class (removing some instances). Another approach is to use algorithms that are inherently robust to class imbalance, such as ensemble methods like Random Forest or boosting algorithms like XGBoost. Additionally, adjusting the class weights during model training can help the algorithm pay more attention to the minority class. Lastly, using evaluation metrics like F1-score or area under the ROC curve (AUC-ROC) can provide a better understanding of model performance in imbalanced datasets.

Weighted cross-entropy loss assigns different weights to classes based on their frequency or significance, allowing the model to focus more on underrepresented classes and reduce prediction errors. Data augmentation and resampling techniques offer another effective solution by creating balanced datasets through synthetic sample generation or modification. Implementing these strategies empowers machine learning models to make more informed and equitable predictions, enhancing performance and reliability in real-world applications.

If you're facing class imbalance, don't despair. Embrace the power of weighted cross-entropy loss and let it be your ally in achieving fair and reliable machine learning models. By ensuring that all classes receive the attention they deserve, you can overcome class imbalance and unlock the true potential of your models. Take control of class imbalance. Implement weighted cross-entropy loss and witness the transformation in your model's performance. Let's make fairness and balance the cornerstones of your machine learning journey.

Existem várias técnicas para lidar com o desequilíbrio de classes em conjuntos de dados de treinamento, algumas delas são: Reamostragem, geração de dados sintéticos, ponderação de classes, utilização de algoritmos específicos, utilização de métricas de avaliação adequadas. A escolha da técnica adequada depende do contexto específico do problema e da natureza dos dados. Experimentar diferentes abordagens e avaliar o desempenho do modelo é fundamental para encontrar a melhor solução.

To add what I wrote in the past, I do still believe Weighted cross-entropy loss offers a method to assign different weights to classes based on their frequency or significance, allowing the model to prioritize minority classes and reduce prediction errors. Data augmentation and resampling techniques provide alternative approaches to balancing class distribution by creating new samples or modifying existing ones. By generating synthetic data or adjusting sample frequencies, developers can alleviate class imbalance issues and improve model performance.

Strategies For Addressing Outliers In Classification Tasks. One common challenge when using cross-entropy loss for classification tasks is dealing with outliers. Outliers are data points that deviate significantly from the rest of the data and can have a disproportionate impact on the model's performance. These outliers can skew the decision boundary of the classifier and lead to poor generalization to unseen data. There are several strategies for addressing outliers in classification tasks. One approach is to preprocess the data by removing or down-weighting outliers before training the model. This can be done by using techniques such as clipping, winsorizing, or robust scaling to limit the influence of extreme values.

Outliers are data points that significantly differ from others in the dataset. They can skew statistical measures and data distribution, leading to misleading results. In classification tasks using cross-entropy loss, outliers can cause the model to overfit to these anomalies. Techniques like robust loss functions, outlier detection, or data preprocessing can help manage outliers.

Outliers are data points that deviate significantly from the overall pattern of your dataset. They can be either: Univariate Outliers: Abnormal values within a single feature (e.g., a person's age of 200 in a dataset about adults). Multivariate Outliers: Combinations of values across multiple features that are unusual (e.g., a house with 10 bedrooms but only 1 bathroom). Outliers can arise due to: Measurement errors: Mistakes during data collection or entry. Data processing errors: Incorrect transformations or calculations. Genuine rare events: Legitimate observations that are simply uncommon.

-Use robust loss functions: Implement loss functions that are less sensitive to outliers, such as Huber loss or focal loss. -Apply outlier detection and removal: Identify and remove or correct outliers before training to reduce their impact. -Employ regularization techniques: Use methods like dropout or weight regularization to mitigate the influence of noisy data.

Class Weights: Assign higher weights to minority classes in the loss function to balance their impact during training. Focal Loss: Replace cross-entropy with focal loss, which down-weights easy examples and focuses on hard-to-classify instances, mitigating imbalance effects. Data Resampling: Use techniques like SMOTE (oversampling minority classes) or undersampling majority classes to balance the dataset. Outlier Detection: Implement outlier detection and removal or robust scaling techniques to minimize their influence on model training.

An outlier is an extremely high or extremely low data point relative to the nearest data point and the rest of the neighboring co-existing values in a data graph or dataset you're working with. They can arise due to errors in data collection, measurement variability, or genuinely unusual phenomena. Outliers can skew statistical analyses and machine learning models, leading to inaccurate results.

Outliers are data points that are significantly different from other data points in a dataset. They are unusual or rare observations that don't follow the pattern of the majority of the data. For example, in a dataset of student ages, if most students are between 18 and 22 years old but there's one student who is 50 years old, that student would be considered an outlier. Outliers can affect statistical analyses and machine learning models, so it's important to identify and handle them properly. Techniques like visualization, statistical tests, or machine learning algorithms can help in detecting and managing outliers.

Addressing outliers is crucial to enhancing model performance and accuracy. Implementing robust techniques such as data preprocessing, outlier detection, or model regularization can help mitigate the influence of outliers, improving the model's resilience to noise and enhancing its ability to make reliable predictions. By effectively managing outliers, practitioners can bolster the robustness and effectiveness of machine learning models in various applications.

One approach is to preprocess the data and remove or adjust outliers based on domain knowledge or statistical methods. Another technique is to use robust loss functions that are less sensitive to extreme values, such as the Huber loss or the modified Huber loss. By mitigating the influence of outliers, you ensure that your model's training is more stable and reliable. This allows for better convergence and improved generalization, leading to more accurate predictions. Don't let outliers derail your model's performance. Take proactive steps to handle them effectively and enhance the stability and reliability of your machine learning models. Embrace robust techniques and unlock the true potential of your data analysis.

Outliers são pontos de dados que se desviam significativamente dos demais, causados por erros, ruídos ou eventos raros. Eles podem aumentar e instabilizar a perda de entropia cruzada, dificultando a convergência e generalização do modelo. Para corrigir os outliers, podemos: 1) Usar métodos estatísticos ou visualização para detectar e excluir outliers. 2) Aplicar transformações logarítmicas ou escalonamento para reduzir o impacto dos outliers. 3) Usar algoritmos ou métodos de regularização que sejam menos sensíveis a outliers. 4) Incluir valores mais apropriados ou substituir outliers por valores medianos ou médias.

Identifying Outliers In The Dataset. Identifying outliers in the dataset is crucial when dealing with class imbalance and using cross-entropy loss for classification tasks. Outliers are data points that deviate significantly from the rest of the dataset and can have a significant impact on the performance of the model. Identifying Outliers In The Dataset Identifying outliers in the dataset is crucial when dealing with class imbalance and using cross-entropy loss for classification tasks. Outliers are data points that deviate significantly from the rest of the dataset and can have a significant impact on the performance of the model.

The best method for outlier detection often depends on your specific dataset and problem: Small Datasets: Visual inspection and statistical methods like Z-score or IQR might be sufficient. Large, High-Dimensional Datasets: Machine learning-based methods like Isolation Forest or LOF can be more effective. Specific Domain Knowledge: If you have domain expertise, you might be able to manually define thresholds or rules for outlier detection. Domain Relevance: Outliers should be considered in the context of your problem domain. A data point that is statistically unusual might not necessarily be an error or irrelevant for your analysis.

Outliers can be detected using statistical methods, Like Z-score: Calculate the number of standard deviations a data point lies from the mean. Values above a certain threshold (e.g., 2 or 3) indicate potential outliers. Interquartile range (IQR): Divide the data into four quartiles. Outliers fall outside the range Q1 - 1.5IQR and Q3 + 1.5IQR. Autoencoders: Train an autoencoder neural network on normal data and then identify outliers as instances with high reconstruction errors.

By setting thresholds based on standard deviations, such as removing data points with z-scores exceeding 3 or falling below -3, outliers can be isolated and potentially removed from the dataset. Additionally, machine learning techniques like clustering, isolation forest, and autoencoders offer advanced outlier detection capabilities by analyzing data point similarity or reconstruction errors. Leveraging these methods enables practitioners to pinpoint and label outliers accurately, enhancing the model's robustness and predictive accuracy.

Podemos detectar os outliers usando os seguintes métodos (usar uma combinação desses métodos pode proporcionar uma detecção mais robusta e precisa): 1) Visualização de dados: Boxplot e Scatter plot. 2) Métodos estatísticos: Desvio padrão e quartis e intervalo interquartil; 3) Modelos baseados em densidade: DBSCAN e Isolation Forest. 4) Modelos baseados em proximidade: K-Nearest Neighbors (KNN) e Local Outlier Factor (LOF). 5) Testes estatísticos: Grubbs' test e Dixon's Q test.

To my updated understanding, now I can say that leveraging statistical techniques like z-scores, box plots, and interquartile ranges enables the identification and removal of outliers based on their deviation from the mean or median. Setting thresholds, such as removing data points with z-scores exceeding 3 or falling below -3, facilitates the elimination of outliers that significantly impact model performance. Additionally, employing machine learning methods like clustering, isolation forest, or autoencoders allows for outlier detection based on data point similarity or reconstruction error.

There are a lot of statistical techniques for outliers, which many people have discussed about. I would like to mention about some machine learning techniques that can identify outliers. One class SVM and KNN are some great supervised learning machine learning algorithms to detect outliers. Apart from this some density based techniques like DBSCAN clustering and LOF (Local Outlier Factor) can also aid well in outlier detection. When visualized, the data points are clumped together, being dense, whereas the outliers are scattered and remain loose. This is the main principle to identify outliers.

To detect outliers, you can use various methods. One way is to visually inspect the data using scatter plots or box plots to identify points that are far from the main cluster. Another method is to calculate statistical measures like the mean and standard deviation, and then identify points that are several standard deviations away from the mean. You can also use techniques like the Interquartile Range (IQR) to detect outliers based on the distribution of the data. Additionally, machine learning algorithms like isolation forests or k-nearest neighbors can be used to automatically detect outliers in the data.

1. The simplest and best way is to visualize the dataset, scatter or line plot, and find the points that you think are far away. 2. Statistical methods like z-scores and quartile ranges can also be used. 3. Thresholding is also implemented to find outliers.

Removal: If outliers are due to errors in data collection or preprocessing, removing them can improve accuracy and generalization of your model. However, exercise caution and only remove outliers if you're confident they are erroneous and not representative of rare but valid events Correction: If outliers are due to measurement errors, try to correct values if possible Transformation: Apply transformations like logarithmic, square root, or Box-Cox transformations to reduce impact of outliers Imputation: Replace outliers with plausible values, such as mean, or values predicted by a regression model. Leave as is: Consider whether removing or modifying outliers would distort the underlying data distribution and potentially introduce bias.

To deal with outliers, you can consider several approaches. One option is to remove them from the dataset if they are due to errors or anomalies. Alternatively, you can transform the data using techniques like log transformation to reduce the impact of outliers on statistical analyses. Another approach is to cap or clip the outliers by setting a threshold beyond which values are considered as outliers and replaced with a predefined value. Finally, you can use robust statistical methods or algorithms that are less sensitive to outliers, such as decision trees or support vector machines, to handle outlier influence in predictive models.

Once outliers are detected, there are several ways to handle them: a) Model-Based Methods: Use robust models that are less influenced by outliers, such as robust regression techniques. b) Transformation: Apply transformations such as log transformation to make the data less sensitive to outliers. c) Winsorization: Cap the extreme values by replacing them with the nearest "normal" value. d) Removal: Remove the outliers from the dataset if they are due to errors or anomalies.

I think that by incorporating regularization methods, practitioners can enhance the model's ability to generalize well to unseen data and minimize the impact of outliers on model predictions. Striking a balance between outlier management strategies and regularization techniques empowers machine learning models to deliver reliable and accurate results across a variety of applications.

Para lidar com outliers podemos adotar várias técnicas. A remoção de outliers pode ser feita manualmente/automaticamente usando métodos estatísticos ou algoritmos específicos. A transformação de dados, como a aplicação de escalonamento logarítmico, raiz quadrada ou z-score, e a normalização para um intervalo fixo, como [0, 1], pode reduzir o impacto dos outliers. Outra técnica é a imputação de valores, substituindo os outliers pela média ou mediana dos dados ou utilizando interpolação. Métodos robustos, como modelos de regressão robusta (RANSAC ou M-estimators) e técnicas de regularização L1 (Lasso), são menos sensíveis a outliers.

Em problemas de machine learning, o uso de funções de perda robustas, como Huber loss, e algoritmos baseados em árvores, como Random Forests ou Gradient Boosting, pode ajudar a mitigar o impacto dos outliers. Além disso, é importante investigar as causas dos outliers para determinar se são erros de medição, entradas incorretas ou dados válidos que refletem eventos raros, e consultar especialistas no domínio para decidir a melhor abordagem.

Remove outliers: Remove outliers that are likely to introduce noise or bias into the model. Cap outliers: Replace outlier values with the maximum or minimum values observed within the dataset. Truncate outliers: Set outlier values to a specific threshold, such as the mean or median. Impute outliers: Replace outlier values with imputed values based on statistical methods or machine learning algorithms. Model-Based Approaches: Use robust statistical models that are less sensitive to outliers, such as robust regression techniques or robust clustering algorithms. These models are designed to handle outliers more effectively than traditional methods.

1. Remove them if they are not that important - Outliers are considered noise and noise can be important to generate new samples. 2. Use thresholding 3. Visualize and remove them manually

Remove them by thresholding if they don't affect the dataset too much. If they are conditional outliers, you might want to keep them. Regularization techniques and dropouts can be used to lower the effect of outliers on the model.

- Use techniques such as SMOTE (Synthetic Minority Over-sampling Technique) or ADASYN for addressing class imbalance by generating synthetic examples in the training dataset. 🔄 - Handling outliers is crucial for improving model accuracy. 📏 - I found helpful re-weighting the loss function to give more importance to minority classes or harder-to-classify examples, which can directly influence the model's learning focus during training. 🎯 - For evaluating the model, using metrics like precision, recall, F1-score, and ROC-AUC provides a comprehensive view of its performance, especially in imbalanced datasets. 📊

Para avaliar seu modelo, use métricas como precisão, recall, F1 score, curva ROC e AUC para medir o desempenho em classificação. Para regressão, considere RMSE ou MAE. Utilize técnicas de validação como validação cruzada k-fold para dividir seus dados em conjuntos de treinamento, validação e teste, garantindo uma estimativa precisa do erro e da confiança do modelo.

1. Validation Set: Split data into training and validation sets. Train the model on the training set and evaluate its performance on the validation set. This provides an estimate of how well the model generalizes to unseen data. 2. Cross-Validation: Divide the data into k folds, train the model on k-1 folds, and evaluate on the remaining fold iteratively. Average the performance metrics across all folds to get a robust estimate of model performance. 3. Performance Metrics: Use appropriate metrics like accuracy, precision, recall, F1-score, and ROC-AUC for classification tasks, or MSE, MAE, and R-squared for regression tasks. Choose metrics based on the specific requirements of the problem.

adjust the class weights within the loss function, assigning higher weights to minority classes to balance their contribution to the loss. Another technique is resampling the dataset to ensure a more equitable distribution of samples across classes, oversampling the minority class or undersampling the majority class, apply outlier detection and removal methods to mitigate the impact of outliers on model performance, techniques like data augmentation or ensemble learning helps improve the robustness of the model and its ability to generalize to outlier-prone datasets, a combination of strategies can help class imbalance and outliers, enhancing the performance and reliability of classification models trained using cross-entropy loss.

Accuracy should never be considered while evaluating models trained on class imbalanced datasets. Let me quote an example. Consider a dataset comprising of 95 positive and 5 negative samples. My model says the output is positive irrespective of the input. If we use accuracy as a metric for evaluation, then the model scores 95% accuracy. But we can't use this model because it fails to identify the another class. Hence we shouldn't rely on accuracy for evaluating class imbalance conditions.

Precision and Recall are important metrics to be considered for the evaluation of machine learning models in the case of class imbalance. Relying on accuracy in this case is useless. The reason is that the model predicting the samples of the majority of the class leads to high accuracy, but not precision and recall, which consider the minority class also.

When dealing with class imbalance or outliers, traditional accuracy metrics can be misleading. Instead, focus on metrics like F1-score, Precision-Recall AUC, and G-mean, which better reflect model performance on imbalanced datasets. Additionally, Cohen’s Kappa and Matthews Correlation Coefficient (MCC) provide more robust evaluations by considering both true and false positives/negatives, offering a balanced view of model performance across all classes. For outlier-sensitive tasks, consider implementing robust cross-validation techniques that systematically evaluate model stability and performance across different data partitions, ensuring resilience to both class imbalance and outliers.

To ensure robustness, I use stratified k-fold cross-validation, maintaining class distribution across folds. For deeper insights, I also leverage SHAP values to interpret model predictions, enhancing transparency in how the model handles class imbalance and outliers.

Explore transfer learning from pre-trained models on large & diverse datasets to leverage the knowledge learned from related tasks & domains. By fine-tuning pre-trained models on the imbalanced dataset, transfer learning enables the model to benefit from features learned from abundant classes while adapting to the specific characteristics of the target task. Personally, I've applied transfer learning techniques with models like VGG or ResNet to address class imbalance in image classification tasks, using features learned from ImageNet pre-trained models to improve performance on minority classes. Investigate anomaly detection algorithms such as isolation forest to identify instances belonging to minority classes or potential outliers.

-Assign different weights to the classes based on their frequency in the dataset. This gives more importance to minority classes, thereby addressing the imbalance issue. You can adjust the class weights inversely proportional to their frequencies. - Oversample the minority class(es) or undersample the majority class(es) to balance the class distribution. Techniques like SMOTE (Synthetic Minority Over-sampling Technique) can be used to generate synthetic samples for the minority class. Conversely, random undersampling can be used to reduce the number of samples in the majority class. -Augment the minority class(es) by applying transformations or adding noise to existing samples. This helps in increasing the diversity of the minority class.

I would adopt a common method known as class weighting, where weights are assigned to different classes based on their frequency in the dataset, thus prioritizing minority classes. Oversampling techniques like SMOTE generate synthetic samples for minority classes, achieving a balanced class distribution. Undersampling randomly removes instances from the majority class to balance proportions. Alternatively, anomaly detection methods identify and handle outliers separately from regular instances. Moreover, utilizing proper evaluation metrics like F1-score or area under the ROC curve (AUC) provides a comprehensive assessment of model performance, considering both precision and recall.

Consider advanced strategies like weighted or focal loss to address imbalance by prioritizing hard-to-classify examples. Use techniques such as SMOTE for oversampling the minority class or ensemble methods for robustness. For outliers, apply robust scaling, utilize anomaly detection algorithms, or consider data augmentation. Adopt robust loss functions that are less sensitive to outliers. Evaluate models using metrics like F1-score or AUPRC instead of accuracy and employ stratified cross-validation. Combining multiple approaches often yields the best results. Regularly monitor and adjust model performance based on new data to ensure continued effectiveness.

Class weighting & resampling help. Class weighting assigns higher weights to minority classes or lower weights to majority classes during training, SMOTE is a good example of resampling, where oversampling the minority class or undersampling the majority class creates a more balanced training dataset. Outlier detection/management techniques, such as robust statistics & anomaly detection algorithms, help identify & handle the impact of outliers on model training.

Before going into the data balancing or outlier detection, be very accurate with the problem statement and what solution is proposed. What is the expected data distribution in production and what is the expected behavior of the solution? Once you have defined those, you know how to build your data to be representative, evaluate your model truthfully and getting a solution that brings value. For example, 0.05% of the data could be the most important defect to detect, or it could be noise to be ignored - only by knowing what that is, you know how to treat your imbalanced data and evaluate your model.