The limitations and Biases were missing so I thought of adding some limitations and Biases

Limitations and several biases for humans, models, evaluations are involved.


README.md

@@ -121,6 +121,16 @@ Example 4-A multiclass example, using different averages.
 
 
 ## Limitations and Bias
+1. Imbalanced Class Distribution - When the number of positive classes is much smaller than the negative classes, the recall may be inflated. This is because a model that predicts all cases as negative will have high accuracy but a low recall. Therefore it is important to consider other metrics like precision, F1-score, AUC-ROC to evaluate the performance of the models in imbalanced datasets.
+2. Misclassification of negative classes - Recall only measures the ability of the model to identify positive cases correctly, without taking into account the misclassification of negative cases. Therefore, a model with a high recall may still misclassify a significant number of negative cases, which could have serious consequences depending on the application.
+3. Bias towards models that predicts more positives - Recall is biased towards models that predict more positive cases, which may not be desirable in some cases. For example, in a medical diagnosis scenario, a model that predicts a high number of false positives may lead to unnecessary treatments or surgeries, causing harm to patients.
+4. Dependency on the decision threshold - Recall, like other classification metrics, depends on the decision threshold used to classify cases into positive and negative classes. Depending on the problem, different decision thresholds may be more appropriate, and adjusting the threshold may affect the recall score.
+5. Scenario/Context dependency - The importance of recall may vary depending on the context and the consequences of missing positive cases. For example, in a fraud detection system, missing a true positive case may have a higher cost than in other applications, such as email spam filtering.
+6. Biases due to sampling - Recall can be biased by the way the data is sampled. For instance, if the dataset used for evaluation is not representative of the population to which the model will be applied, then the recall score may not generalize well to new data. Sampling bias can also arise when the positive cases are oversampled or undersampled, leading to an artificially high or low recall score.
+7. Biases due to data quality - Recall can also be biased by the quality of the data. If the positive cases are noisy or mislabeled, then the recall score may be lower than expected. Similarly, if the negative cases are not well defined or ambiguous, then the recall score may be inflated.
+8. Biases due to feature selection - Recall can be biased by the features used to train the model. If the features used to represent the positive cases are not sufficiently distinctive, then the recall score may be lower than expected. Conversely, if the features used to represent the negative cases overlap significantly with the positive cases, then the recall score may be inflated.
+9. Biases due to Model Selection - Recall can also be biased by the model selection process. If the model is overfit to the training data, then the recall score may be artificially high. Conversely, if the model is underfit, then the recall score may be lower than expected.
+10. Biases due to Model Bias - Recall can be biased by the way the labels are assigned. For example, if the labels are assigned by a human expert, then the recall score may be influenced by the expert's subjective judgment. Similarly, if the labels are assigned by an automated process, then the recall score may be influenced by the quality of the labeling algorithm.
 
 
 ## Citation(s)