What is model performance?

Assessing model performance happens during the model evaluation and model monitoring stages of a machine learning pipeline. After artificial intelligence (AI) practitioners work on the initial phases of ML projects, they then evaluate a model’s performance across multiple datasets, tasks and metrics to gauge its effectiveness. Once the model is deployed, machine learning operations (MLOps) teams monitor model performance for continuous improvement.

An AI model’s performance is generally measured using a test set, comparing the model’s outputs against predictions on the baseline test set. Insights gained from evaluating performance help determine if a model is ready for real-world deployment or if it needs tweaking or additional training.

Here are some factors that can impact a machine learning model’s performance:

• Data quality

• Data leakage

• Feature selection

• Model fit

• Model drift

• Bias

A model is only as good as the data used to train it. Model performance falls short when its training data is flawed, containing inaccuracies or inconsistencies like duplicates, missing values and wrong data labels or annotations. A lack of balance—such as having too many values for one scenario over another or a training dataset that’s not sufficient or diverse enough to correctly capture correlations—can also lead to skewed results.

Data leakage in machine learning occurs when a model uses information during training that wouldn’t be available at the time of prediction. This can be caused by data preprocessing errors or contamination due to improper splitting of data into training, validation and test sets. Data leakage causes a predictive model to struggle when generalizing on unseen data, yield inaccurate or unreliable results, or inflate or deflate performance metrics.

Feature selection involves choosing the most relevant features of a dataset to use for model training. Data features influence how machine learning algorithms configure their weights during training, which in turn drives performance. Additionally, reducing the feature space to a selected subset can help improve performance while lowering computational demands. However, picking irrelevant or insignificant features can weaken model performance.

Overfitting happens when an ML model is too complex and fits too closely or even exactly to its training data, so it doesn’t generalize well on new data. Conversely, underfitting occurs when a model is so simple that it fails to capture the underlying patterns in both training and testing data.

Model drift refers to a model’s performance degrading because of changes in data or in the relationships between input and output variables. This decay can negatively impact model performance, leading to faulty decision-making and bad predictions.

Bias in AI can be introduced at any phase of a machine learning workflow, but it’s particularly prevalent in the data processing and model development stages. Data bias occurs when the unrepresentative nature of training and fine-tuning datasets adversely affects model behavior and performance. Meanwhile, algorithmic bias is not caused by the algorithm itself but by how data science teams collect and code training data and how AI programmers design and develop machine learning algorithms. AI bias can lead to inaccurate outputs and potentially harmful outcomes.

It’s important to align metrics with the business goals a model is meant to meet. While each type of machine learning model has its own set of evaluation metrics, many models share a few measures in common:

• Accuracy

• Recall

• Precision

• F1 score

Accuracy is calculated as the number of correct predictions divided by the total number of predictions. This percentage is a very common metric.

Model accuracy and model performance are often likened, but model accuracy is just one part of model performance. And while they’re intertwined, accurate predictions alone can’t provide a holistic view of how well a model performs.

Recall quantifies the number of true positives—the actual correct predictions. It’s also known as the sensitivity rate or true positive rate (TPR).

This metric is critical in healthcare, for example, when diagnosing diseases or detecting cancer. An ML model with high recall can correctly identify positive cases while minimizing false negatives (actual positive cases incorrectly predicted as negative cases).

Precision is the proportion of positive predictions that are actual positives. A machine learning model with high precision can minimize false positives (actual negative cases incorrectly predicted as positive cases).

This metric is crucial in finance, for instance, when detecting fraud. Flagged transactions must indeed be fraudulent (true positives) since flagging legitimate transactions as fraudulent (false positives) can have negative consequences.

The F1 score is the harmonic mean of recall and precision, blending both metrics into a single one. It considers the two measures to be of equal weight to balance out any false positives or false negatives. It’s especially useful for imbalanced datasets, such as when detecting rare diseases, since negative cases far outweigh positive ones.

Many AI frameworks, such as the Python-based PyTorch, scikit-learn and TensorFlow, offer built-in functions for calculating accuracy, recall, precision and the F1 score. They also provide visualizations of model predictions as a confusion matrix—a table representing both predicted and actual values, with boxes depicting the number of true positives, false positives, true negatives and false negatives.

Classification models sort data points into predefined groups called classes. Here are some metrics specific to classification models:

• ROC curve: A receiver operating characteristic (ROC) curve visualizes the proportion of true positives to true negatives. The chart plots the true positive rate against the true negative rate for each threshold used in model classification. The area under curve (AUC) statistic arises from the ROC curve and measures how likely a randomly selected positive has a higher confidence score than a random negative. AUC-ROC is a helpful metric for tasks involving binary classification (sorting data into two exclusive classes).

• Logarithmic loss: Log loss appraises the confidence of a model’s classifications, penalizing confident incorrect classifications more heavily than less confident ones. This is particularly useful when tackling probabilistic outputs, as models learn to be confident about correct classifications and uncertain about incorrect ones. Lower logarithmic loss values denote better performance.

Regression models are employed for predictions involving continuous values, such as retail sales estimates and stock price forecasts. Since these algorithms deal with quantifiable concepts, their metrics measure errors in predictions:

• Mean absolute error (MAE) is computed as the sum of the absolute value of all errors divided by the sample size. It measures the average absolute difference between the predicted value and actual value.

• Mean squared error (MSE) is calculated as the average of the squared differences between the predicted value and the true value across all training samples. Squaring the error punishes large mistakes and incentivizes the model to reduce them.

• Root mean squared error (RMSE) is the square root of the MSE. Squaring errors before averaging them punishes larger mistakes even more heavily, again encouraging models to minimize them.

These metrics evaluate the performance of natural language processing (NLP) models. They’re also used as benchmarks for large language models (LLMs).

Here are some quantitative NLP model measures:

• Perplexity measures how good a model is at prediction. The lower an LLM’s perplexity score, the better it is at comprehending a task.

• Bilingual evaluation understudy (BLEU) evaluates machine translation by computing the matching n-grams (a sequence of n-adjacent text symbols) between an LLM’s predicted translation and a human-produced translation.

• Recall-oriented understudy for gisting evaluation (ROUGE) appraises text summarization and has several types. ROUGE-N, for instance, does similar calculations as BLEU for summaries, while ROUGE-L computes the longest common subsequence between the predicted summary and the human-produced summary.

Qualitative metrics encompass measures such as coherence, relevance and semantic meaning and usually involve human assessors examining and scoring models. A balance of both quantitative and qualitative metrics can make for a more nuanced evaluation.

Computer vision models, particularly those for instance segmentation and object detection, are evaluated using these two common performance measures:

• Intersection over union (IoU) computes the ratio of the area of intersection over the area of union. Intersection covers the overlapping sections between a bounding box that demarcates a detected object as predicted by a model and the actual object. Union denotes the total area of both the bounding box and the actual object. Computer vision models use IoU to assess the preciseness of localizing detected objects.

• Mean average precision (mAP) calculates the mean of all average precision scores across object classes. Computer vision models use IoU to assess prediction and detection accuracy.

Most techniques for optimizing machine learning performance are implemented during model development, training and evaluation. Once a model is deployed in the real world, however, its performance must be constantly tracked. Model monitoring informs decisions on how to improve performance over time. 

Refining ML model performance entails one or more of these techniques:

• Data preprocessing

• Preventing data leakage

• Choosing the right features

• Hyperparameter tuning

• Ensemble learning

• Transfer learning

• Attaining optimal model fit

• Protecting against model drift

• Addressing bias

Many AI frameworks have prebuilt features that support most of these techniques.

Establishing and maintaining rigorous data preprocessing or data preparation procedures can help avoid data quality issues. While data cleaning, denoising and data normalization are mainstays of data preprocessing, data scientists can also use data automation tools and even AI-powered tools to save time and effort and prevent human errors. For insufficient or imbalanced datasets, synthetic data can fill in the gaps.

Careful data handling is key to prevent data leakage. Data must be properly split into training, validation and test sets, with preprocessing done separately for each set.

Cross-validation can also help. Cross-validation splits data into multiple subsets and uses different ones for training and validation in a defined number of iterations.

Feature selection can be challenging and requires domain expertise to pinpoint the most essential and influential features. It’s important to understand the significance of each feature and examine the correlation between features and the target variable (the dependent variable that a model is tasked with predicting).

Feature selection methods for supervised learning include wrapper methods and embedded methods. Wrapper methods train a machine learning algorithm with different subsets of features, adding or removing them and testing the results at each iteration to determine the feature set that leads to optimal model performance. Embedded methods integrate feature selection into model training, identifying underperforming features and eliminating them from future iterations.

With unsupervised learning, models figure out data features, patterns and relationships on their own. Feature selection methods for unsupervised learning include principal component analysis (PCA), independent component analysis (ICA) and autoencoders.

Hyperparameter tuning, also known as hyperparameter optimization or model tuning, identifies, selects and optimizes a deep learning model’s hyperparameters to obtain the best training performance. Hyperparameters govern a model’s learning process, and finding the right combination and configuration of hyperparameters can strengthen model performance in the real world.

Common hyperparameter tuning methods include grid search, random search, Bayesian optimization and hyperband. Data scientists can also implement automated methods to algorithmically discover the optimal hyperparameters that fit their use case.

Ensemble learning combines multiple models to enhance predictive performance, with the assumption that a collective or ensemble of models can produce better predictions than a single model alone.

Here are some popular ensemble learning techniques:

• Bagging, also called bootstrap aggregation, trains models in parallel and independent of each other. It then takes the average (for regression tasks) or majority (for classification problems) of the predictions to compute a more accurate estimate.

• Boosting trains models sequentially, correcting past mistakes in each iteration. It gives more weight to erroneous or misclassified instances in succeeding models, thereby focusing on challenging data points and enhancing performance along the way.

• Stacking trains models from the same dataset but applies a different training algorithm for each one. It then uses the compiled or stacked predictions to train a final model.

Transfer learning takes the knowledge gained by a pretrained model on an initial task or dataset and applies it to a new but related target task or dataset. Repurposing a pretrained model for a different task boosts that model’s generalization capabilities, helping to optimize performance.

Managing overfitting and underfitting is a central challenge in machine learning. An optimally fit model accurately recognizes patterns in data without being too sensitive to random fluctuations or noise.

Techniques to avoid overfitting and underfitting include finding the right training duration to give models just enough time to learn, data augmentation to expand the training set and regularization to reduce variance in a model by applying a penalty to input parameters with larger coefficients.

Drift detection, a core aspect of model monitoring and observability, can help protect against model drift. For instance, AI drift detectors automatically recognize when a model’s accuracy decreases or drifts below a predefined threshold, while monitoring tools continually observe drift scenarios.

Once drift is detected, ML models can be updated in real time or retrained using a new dataset containing more recent and relevant samples.

Mitigating AI bias begins with AI governance, which encompasses guardrails, processes and standards that help ensure AI systems and tools are ethical and safe. Here are some responsible AI practices that can guard against bias:

• Diversify data sources and include data representative of a wide variety of conditions, contexts and demographics.

• Cultivate diverse teams to promote inclusive AI design and development.

• Employ explainable AI techniques for transparency, such as Local Interpretable Model-Agnostic Explanations (LIME) to explain the prediction of classifiers by ML algorithms and Shapley Additive Explanations (SHAP) to explain the output of any ML model.

• Incorporate fairness metrics into the development process and use algorithmic fairness tools and frameworks.

• Conduct regular audits to assess data and algorithms for biases.

• Implement continuous performance monitoring for deployed ML models to swiftly detect and correct bias in outcomes.