To improve models by improving its data, one should correctly evaluate the model to determine if it needs Data-Centric AI approaches and how well they do, and to investigate the model’s shortcomings.
Take into consideration that this note will use a multi-class classification model.
Reporting model performance
The loss function is as fundamental in training models as it is in evaluating them. The loss may be a function of either:
-
The predicted class y for a given sample x, like accuracy, balanced accuracy, precision, and recall.
-
The predicted probability of each class for a given sample x, like log loss, AUROC, calibration error.
Each metric shows a different angle from the same scenario, and should be evaluated depending on the context on which the model will be applied to.
Take, for example, a Fraud vs. NotFraud classification. Without checking for class imbalance, you train a model and the accuracy is 99%. One could say that the model performs very well, but maybe if you look further, you can discover the balanced accuracy, which shows the model is not that accurate after all. This can happen because of the Class Imbalance of the problem.
Don’t focus on only one metric. Check other metrics based on the context which the model will be applied and the data you’re working with. Try to represent sub-populated classes in your evaluation. A Confusion Matrix might help.
Data Leakage
To recap, data leakage is basically when information that the model should not have access to is fed into the training data.
For example, an email spam classification. You might use features such as the frequency of certain words and the length of the email, but you cannot use a feature indicating whether the sender is known to be associated with spam. The model will learn to associate this feature to the training data, but when used in a production scenario, won’t have access to this feature and won’t generalize well, even though it can show great performance in evaluation.
Another pitfall is not using truly held-out data. It’s very easy to overfit the model by using the same data for validation and hyperparameter optimization to also evaluate the model as if it were held out. The sample for evaluation should only be used for computing scores, not for modelling decisions.
A great way to deal with this is to split the data in three different parts: train, validation and test. Click this link for reference.
Underperforming data and imbalanced target
For a given model, it’s possible for the loss of a specific subpopulation of the data to be significantly higher than the rest of the dataset. This phenomena is known as Underperforming Subpopulations, where a specific sample of the dataset can be a challenge to work with, being less frequent and less accurate with respect to the features itself, not the target. Evaluating these are a bit different and need its own methods.
Another face of the same coin is Class Imbalance, where the issue is the target of the training data, which can present very concentrated distribution, leaving some classes with almost no data points.
Inspecting isolated data points
There are a lot of reasons on why a classifier might output a bad prediction for a given example. Some of them are:
-
The given label is incorrect, and the model actually predicted right.
This can happen if the dataset is not accurate and clean enough. One way to remedy this is by using Confident Learning for label correction.
-
The example does not belong to any of the classes or is fundamentally not predictable.
For example, a blurry image or data of something the model doesn’t know yet. If it doesn’t take part in the deployment context, delete from the model or create an “Other” category for aggregating such classes.
-
The example doesn’t conform to the training data distribution.
It may be an outlier, or noisy data, as there’s nothing similar in the dataset. You can either use a Outlier Removal technique, or collect and engineer additional features for this particular case. Also, consider a preprocessing operation that normalize the data.
-
The model may be suboptimal or underfitting.
To diagnose this, up-weight the example or duplicate it many times in the dataset. If the model still can’t capture it, it’s not data related. Try to mess around with the modelling itself.
-
The dataset contains similar examples with different labels.
It’s possible that may exist other data points nearly identical to the one being wrongfully predicted, but containing different labels. In this scenario, there’s little you can do to improve performance, besides feature engineering and calibration techniques.
Influence of individual data points
Certain data points might exert a significant influence on the model by being unique examples. This influence can be quantified by the Influence Function I(x), which can measure the change in the model’s prediction vs. the change in its loss (i.e., its performance). Influence reveals which data points impacts the most. If you were to correct a label from a very influential data point, it would produce a much better result than to correct a low influence data point.
Leave-One-Out Influence — LOO
After training and evaluating the model with all the data, omit a specific data point and check how the predictions, accuracy, and parameters changed as a result. One data point alone may not have a huge impact on the model, but together with its subgroups, can define an entire class.
Data Shapley
Another form of influence is the LOO influence of a given data point in any subset that contains it. Averaging this quantity over all subsets exposes a value that better represents the influence. This approach is called Data Shapley.
e.g, if there are two identical data points in a dataset where omitting both severely harms model accuracy, LOO influence may still conclude that neither is too important (unlike the Data Shapely value).
Monte-Carlo influence for classification
For an arbitrary classifier, it’s possible to approximate influence via Monte-Carlo sampling:
-
Subsample T different data subsets Dt from the original training dataset (without replacement).
-
Train a separate copy of your model on each subset Dt and report its accuracy on held-out validation data.
-
To assess the value of a given data point, compare the average accuracy of models for those subsets that contained the data point vs. those that did not.
Accuracy here could be any kind of loss function.
Closed-form Computation of Influence
For Linear Regression with Least Ordinary Squares—MSE, the LOO Influence I(x) can be easily calculated via a formula known as Cook’s Distance, which shows the influence of each data point on the fitted response values. This happens because the parameters of a linear regression model are, mathematically, a closed form function of the data itself.
Reference
-
Data-Centric Evaluation of ML Models — MIT — Introduction to Data-Centric AI — Course.
-
Why Is My Classifier Discriminatory? — Cornell University — Paper — 2018.
-
Boosting Model Performance Through Error Analysis — Landing AI — Blog post.
-
Using SHAP Values to Explain How Your Machine Learning Model Works — Towards Data Science — Medium post.