Technical Interviewer's Questions

1\ how would you handle imbalanced data in a classification problem? great response "i'd first quantify the imbalance ratio to understand its severity for moderately imbalanced data, i typically use resampling techniques like smote for oversampling the minority class or controlled undersampling of the majority class for highly imbalanced data, i'd consider using specialized algorithms like xgboost with scale pos weight parameter or focal loss functions that down weight easy examples i'd also use stratified sampling in cross validation to maintain class distributions most importantly, i'd move beyond accuracy to more appropriate metrics like precision recall auc, f1 score, or cohen's kappa depending on the business context in production, i might implement ensemble methods combining models trained on different sampling strategies " mediocre response "for imbalanced data, i usually oversample the minority class or undersample the majority class to balance the dataset i'd use metrics like f1 score instead of accuracy since accuracy can be misleading i'd also consider using weighted loss functions to penalize misclassification of the minority class more heavily " poor response "i would use random oversampling to duplicate minority class examples until the classes are balanced then i'd train a standard model like a neural network and evaluate using accuracy if the project deadline is tight, i might just use class weight parameter in scikit learn to save time " 2\ explain the difference between l1 and l2 regularization and when you would use each great response "l1 regularization adds the sum of absolute values of weights to the loss function, which tends to produce sparse models by driving some weights exactly to zero, effectively performing feature selection i'd use l1 when i suspect many features are irrelevant or when model interpretability and computational efficiency are priorities l2 regularization adds the sum of squared weights to the loss function, which results in smaller but non zero weights across all features this helps prevent overfitting by constraining the model's complexity without eliminating features entirely i'd use l2 when most features likely contribute to the prediction, but i want to prevent any single feature from dominating in practice, i often use elastic net, which combines both l1 and l2, giving me the sparsity benefits of l1 and the stability benefits of l2 i'd determine the optimal mix through cross validation based on the specific dataset characteristics " mediocre response "l1 regularization uses the absolute value of weights while l2 uses squared weights l1 can make models sparse by setting some weights to zero, which helps with feature selection l2 keeps weights small but non zero and works better when all features are somewhat relevant i'd try both and select whichever gives better validation performance " poor response "l1 is lasso and l2 is ridge l1 sets some weights to zero and l2 makes weights smaller i usually just use l2 regularization because it's the default in most libraries and generally works well enough if i need to speed up the model, i might try l1 to reduce the number of features " 3\ how would you approach a time series forecasting problem where you have both short term and long term patterns? great response "i'd decompose the time series into its components trend, seasonality, and residuals for capturing both short and long term patterns, i'd implement a hybrid modeling approach for long term trends, i might use methods like sarima or prophet that explicitly model seasonality at different frequencies for short term patterns and complex interactions, i'd complement this with gradient boosting models or rnns/lstms that can capture recent temporal dependencies feature engineering would be critical i'd create lagged features at different time scales relevant to the business context, and include external regressors if available i'd also implement multiple forecast horizons targeted to different business needs, with separate models optimized for short term accuracy versus long term trend capturing for evaluation, i'd use time series cross validation with expanding windows rather than traditional k fold cv, and i'd track different metrics for different forecast horizons mape or smape for longer term forecasts, and mae or rmse for shorter term predictions " mediocre response "i would use models that can handle multiple seasonalities like sarima or prophet for the basic forecasting i'd create lag features at different intervals to capture both patterns for evaluation, i'd use metrics like mae or rmse with time series cross validation i might also try deep learning approaches like lstm if the traditional methods don't perform well " poor response "i'd probably use an lstm neural network since they're good at capturing patterns across different time scales i'd split the data into training and testing sets, with the most recent data as the test set if the results aren't good enough, i might try adding more layers to the network or including more historical data in the training process " 4\ what techniques would you use to handle missing data in a dataset? great response "my approach to missing data follows a systematic process first, i analyze the missingness patterns to determine if data is missing completely at random (mcar), missing at random (mar), or missing not at random (mnar), as this fundamentally affects the appropriate strategy for mcar or mar data where less than 5% is missing, simple imputation methods like mean, median, or mode might suffice for larger amounts of missing data, i'd employ more sophisticated approaches like knn imputation for capturing local structure in the data multiple imputation by chained equations (mice) to account for uncertainty model based imputation using algorithms that handle missing values internally, like xgboost for mnar data, i'd be very cautious about imputation and might consider creating a 'missingness indicator' feature to explicitly model the absence pattern in production systems, i'd implement a consistent pipeline where the imputation strategy is trained on the training data and applied identically to new data i'd also test the sensitivity of my model to different imputation strategies as part of the validation process " mediocre response "i would first check how much data is missing and its distribution for numerical features, i'd use mean or median imputation, or more advanced methods like knn imputation for categorical features, i'd use mode imputation or create a new category for missing values if a feature has too many missing values, i might drop it for time series data, i'd use forward or backward fill methods " poor response "i would drop rows with missing values if i can afford to lose some data otherwise, i'd fill numerical missing values with the mean and categorical missing values with the most common category there are also libraries that handle this automatically, so i'd probably use the simpleimputer from scikit learn since it's quick to implement " 5\ how do you interpret the coefficients in a logistic regression model? great response "in logistic regression, coefficients represent the change in the log odds of the positive class for a one unit increase in the feature, while holding other features constant more practically, we can exponentiate the coefficient to get the odds ratio, which is more intuitive for instance, if a feature has a coefficient of 0 7, exp(0 7) ≈ 2, meaning the odds of the positive class approximately double with a one unit increase in that feature it's important to note that these interpretations assume features are appropriately scaled for standardized features, coefficients can be compared directly to assess relative importance however, this is just a rough indicator since correlation between features can complicate interpretation for non linear relationships or interaction terms, the interpretation becomes more complex in those cases, i prefer to use partial dependence plots or shap values to understand how the feature impacts the prediction across its entire range rather than just relying on the coefficient value " mediocre response "the coefficients in logistic regression show how much the log odds of the target variable changes with a one unit change in the feature positive coefficients increase the probability of the positive class, while negative coefficients decrease it you can exponentiate the coefficients to get odds ratios, which are easier to interpret the magnitude of the coefficient indicates the strength of the feature's influence on the prediction " poor response "logistic regression coefficients tell you how important each feature is for predicting the outcome larger coefficients mean the feature has a stronger effect on the prediction positive coefficients increase the prediction probability and negative ones decrease it you generally want to focus on the features with the largest absolute coefficients " 6\ how do you choose the number of clusters in a clustering algorithm like k means? great response "determining the optimal number of clusters requires both quantitative metrics and domain knowledge i typically use the elbow method by plotting the within cluster sum of squares (wcss) against different k values and looking for the point where diminishing returns set in i complement this with the silhouette score, which measures how similar points are to their assigned cluster versus other clusters, and the davies bouldin index, which evaluates both cluster separation and compactness i also implement stability analysis by running the clustering multiple times with different random initializations or on subsamples of data to see if consistent cluster structures emerge this helps ensure the clusters are robust and not artifacts of the algorithm most importantly, i validate the business relevance of the resulting clusters even statistically optimal clusters are useless if they don't provide actionable insights i'd work with domain experts to evaluate if the identified clusters represent meaningful segments that align with business objectives sometimes a slightly suboptimal number of clusters from a statistical perspective might be more practical from a business perspective " mediocre response "i would use the elbow method by plotting the within cluster sum of squares against the number of clusters and looking for the point where the curve bends i'd also calculate the silhouette score for different numbers of clusters and choose the one with the highest average score sometimes i validate the results with domain experts to ensure the clusters make practical sense " poor response "i usually try different values of k, like 2 to 10, and plot the within cluster sum of squares the point where the curve starts to flatten is generally a good choice for k if that doesn't give a clear answer, i might just choose a value based on what makes sense for the application or what's convenient to work with " 7\ what evaluation metrics would you use for an imbalanced classification problem and why? great response "for imbalanced classification, i'd avoid accuracy since it can be misleading when classes are skewed instead, i'd use a combination of metrics tailored to the business context if false negatives and false positives have similar importance, i'd focus on the f1 score, which balances precision and recall for a comprehensive view across different classification thresholds, i'd use the precision recall auc rather than roc auc, as pr curves are more sensitive to performance on the minority class when costs of errors differ, i'd use weighted metrics like weighted f1 or matthews correlation coefficient, which works well even with severe imbalances for ranking performance, i'd consider the cohen's kappa coefficient, which measures agreement accounting for chance most importantly, i'd translate these technical metrics into business impact measures for example, in fraud detection, i might calculate the potential financial loss prevented versus the operational cost of investigating flagged transactions this helps stakeholders understand model performance in terms directly relevant to business outcomes " mediocre response "for imbalanced datasets, accuracy is misleading because a model could achieve high accuracy by just predicting the majority class i would use metrics like precision, recall, and f1 score that focus on the performance of the minority class the roc auc score is also useful because it evaluates classification performance across different thresholds depending on the application, i might prioritize either precision or recall based on the costs of false positives versus false negatives " poor response "i would use the f1 score instead of accuracy because it works better for imbalanced datasets i'd also look at the confusion matrix to see how many false positives and false negatives there are if needed, i might adjust the classification threshold to get better results on the minority class, focusing on whichever metric the project requirements specify as most important " 8\ explain how gradient descent works and how you would tackle the problem of getting stuck in local minima great response "gradient descent optimizes model parameters by iteratively moving in the direction of steepest descent of the loss function at each step, we calculate the gradient (vector of partial derivatives) of the loss with respect to each parameter, then update parameters in the opposite direction with a step size controlled by the learning rate to address local minima challenges, i use several strategies stochastic gradient descent with momentum, which adds a fraction of the previous update to the current one, helping to maintain velocity through shallow local minima and saddle points learning rate schedules like learning rate decay or adaptive methods such as adam, which combines momentum with per parameter learning rates multiple random initializations to explore different regions of the parameter space for deep learning specifically, architectural choices like proper weight initialization (e g , he or xavier) and batch normalization can create better conditioned optimization landscapes for non convex problems, i might use more advanced techniques like simulated annealing or genetic algorithms that can probabilistically accept uphill moves to escape local minima the specific approach depends on the problem complexity and computational constraints, with simpler problems often needing only basic momentum, while complex deep learning tasks might require combinations of these techniques " mediocre response "gradient descent works by calculating the gradient of the loss function with respect to the model parameters and updating the parameters in the opposite direction of the gradient, scaled by a learning rate this moves the parameters toward the minimum of the loss function to avoid getting stuck in local minima, i would use techniques like using stochastic gradient descent which adds noise to the updates implementing momentum to help overcome small barriers trying different random initializations using adaptive learning rate methods like adam or rmsprop" poor response "gradient descent is an optimization algorithm that updates model weights by moving in the direction that reduces the loss function to handle local minima, i'd use a technique like momentum or adam optimizer, which are available in most libraries if those don't work, i'd try different random initializations or increase the model complexity to find a better solution " 9\ what is the bias variance tradeoff and how do you manage it in practice? great response "the bias variance tradeoff represents the fundamental tension between a model's ability to capture patterns in the training data (low bias) and its ability to generalize to unseen data (low variance) high bias models underfit by oversimplifying, while high variance models overfit by capturing noise as signal in practice, i manage this tradeoff through a systematic approach first, i establish a reliable cross validation strategy appropriate for the data structure (stratified k fold for classification, time based for temporal data) to get stable estimates of generalization error i then implement a progressive modeling strategy, starting with simpler models to establish a baseline and gradually increasing complexity while monitoring both training and validation errors the point where validation error begins to increase while training error continues to decrease signals the onset of overfitting for specific algorithms, i carefully tune regularization parameters (like c in svms or alpha in ridge/lasso regression) through grid search or bayesian optimization in ensemble methods, i control the tradeoff by adjusting the number of weak learners and their individual complexity finally, i consider the application context in some high stakes domains where interpretability is crucial, i might accept slightly higher bias for much lower variance, while in other cases where capturing subtle patterns is essential, i might tolerate slightly higher variance " mediocre response "the bias variance tradeoff is the balance between a model's ability to fit the training data (bias) and its ability to generalize to new data (variance) models with high bias tend to underfit, while models with high variance tend to overfit to manage this tradeoff, i use techniques like cross validation to detect overfitting regularization to reduce model complexity ensemble methods like random forests that can reduce variance grid search to find optimal hyperparameters that balance bias and variance" poor response "bias variance tradeoff means that reducing error from bias often increases error from variance and vice versa to handle this, i usually add regularization to complex models to prevent overfitting if the model is underfitting, i'd add more features or use a more complex model i generally rely on the default parameters in libraries like scikit learn, then adjust them if the validation scores aren't good enough " 10\ how would you deploy a machine learning model to production and what considerations would you keep in mind? great response "deploying ml models to production involves several critical phases, each with important considerations for the model preparation phase, i'd focus on ensuring model reproducibility by versioning data, code, and random seeds optimizing the model for inference speed through quantization or distillation if needed converting the model to a deployment friendly format like onnx or tensorflow savedmodel for deployment infrastructure, i'd consider containerization with docker to ensure consistent environments implementing a/b testing capabilities to safely validate new models designing for appropriate serving architecture based on latency requirements (batch vs real time) ensuring scalability through kubernetes or cloud native services for production monitoring, i'd establish performance dashboards tracking both technical metrics (latency, throughput) and business kpis drift detection for data and concept drift that could degrade model performance automated alerting systems when metrics cross predefined thresholds feedback loops to capture ground truth for continuous model improvement from a devops perspective, i'd implement ci/cd pipelines to automate testing and deployment canary deployments to minimize risk when rolling out updates rollback mechanisms for quick recovery from problematic deployments security and compliance considerations would include model access controls and authentication privacy preservation techniques if handling sensitive data documentation for regulatory compliance where applicable" mediocre response "to deploy a machine learning model to production, i would serialize the model and save it in a format like pickle or joblib wrap the model in an api using flask or fastapi containerize the application with docker for consistency set up monitoring to track performance metrics implement logging to catch errors and unexpected behavior create a ci/cd pipeline for updates i'd consider factors like scalability, latency requirements, and how to handle drift in the input data over time i'd also need to think about how to retrain the model when performance degrades " poor response "i would export the trained model to a file format like pickle, then create a simple api endpoint using a framework like flask that loads the model and makes predictions for deployment, i'd upload this to a cloud server and set up some basic logging if the performance starts to drop, i'd retrain the model with new data and update the deployed version " 11\ explain the concept of ensemble methods and when you would use them great response "ensemble methods combine multiple models to produce stronger predictions than any individual model could achieve alone they work by leveraging the diversity of errors across different models where one model fails, another might succeed i use different ensemble approaches based on the specific problem characteristics for bagging methods like random forests, i'd use them when i'm concerned about variance and overfitting, particularly with complex models like decision trees by training models on different bootstrap samples and averaging predictions, bagging reduces variance while maintaining the same bias boosting methods like xgboost or adaboost work sequentially, with each model correcting errors made by previous ones i'd use these when i need to squeeze out maximum performance and have sufficient data to prevent overfitting they're particularly effective for problems with complex decision boundaries stacking involves training a meta model on the outputs of base models i'd implement stacking when i have models with complementary strengths like combining the structural understanding of tree based models with the smooth decision boundaries of neural networks ensemble diversity is key to their success, so i ensure diversity through techniques like feature subsampling, parameter variation, algorithm diversity, and different data preprocessing approaches for production systems, i balance performance gains against increased complexity, resource requirements, and potential maintenance challenges " mediocre response "ensemble methods combine multiple models to improve prediction performance the main types are bagging (like random forests), boosting (like xgboost), and stacking (combining predictions from different models) i would use ensemble methods when individual models are unstable or have high variance i need to improve model performance beyond what a single model can achieve i have sufficient computational resources for training and inference the problem is complex and benefits from different modeling approaches" poor response "ensemble methods combine multiple models together to get better results than a single model the most common ones are random forests and xgboost i would use them when i need better accuracy and have enough computing power available they usually perform better than simpler models like linear regression or decision trees, so i'd try them if the simple models aren't meeting the requirements " 12\ how would you handle feature selection for a high dimensional dataset? great response "for high dimensional feature selection, i employ a multi stage approach that combines statistical methods, model based techniques, and domain knowledge first, i'd conduct preliminary dimensionality reduction through removing features with near zero variance eliminating highly correlated features (correlation threshold > 0 95) using domain expertise to filter obviously irrelevant features next, i'd apply more sophisticated techniques depending on computational constraints filter methods like mutual information or anova f tests provide quick initial screening wrapper methods like recursive feature elimination with cross validation (rfecv) for more thorough selection embedded methods like l1 regularization (lasso) or tree based feature importance for extremely high dimensions (10,000+), i'd consider random feature selection with ensemble methods auto encoders for non linear dimensionality reduction modern techniques like boruta that leverage random forest permutation importance throughout this process, i'd implement proper cross validation strategies to prevent data leakage, particularly separating feature selection from model evaluation i'd also validate stability of selected features through techniques like bootstrap sampling to ensure robustness finally, i'd evaluate the tradeoff between model performance, interpretability, and computational efficiency at different feature subset sizes to determine the optimal dimension for the specific business context " mediocre response "for high dimensional data, i would use a combination of techniques first, i'd remove highly correlated features using correlation analysis i'd use filter methods like chi square or anova f test for initial screening then i'd apply wrapper methods like recursive feature elimination or embedded methods like l1 regularization i might also use tree based feature importance from algorithms like random forest or xgboost pca could help reduce dimensionality while preserving variance i would validate the selected features using cross validation to ensure they generalize well to unseen data " poor response "i would use feature importance from a random forest or gradient boosting model to rank features, then select the top n features based on importance scores if that's too slow, i might just use correlation with the target variable to pick features pca is another option to reduce dimensions, though it makes the features less interpretable the simplest approach is to just use a model with built in feature selection like lasso " 13\ what are the differences between bagging and boosting techniques? great response "bagging and boosting represent fundamentally different strategies for ensemble learning, with distinct characteristics across multiple dimensions in terms of training methodology, bagging (bootstrap aggregating) trains models in parallel on random subsets of data sampled with replacement each model is independent and unaware of others boosting, however, trains models sequentially, with each model explicitly designed to correct errors made by previous models, creating a dependency chain for variance bias management, bagging primarily reduces variance by averaging out noise across multiple models while maintaining similar bias this makes it excellent for high variance models like unpruned decision trees boosting primarily reduces bias by focusing subsequent models on difficult examples, making it effective for simple models with high bias regarding overfitting tendencies, bagging is naturally resistant to overfitting due to its averaging approach and independent model training boosting, being an iterative error correcting process, is more prone to overfitting, especially with noisy data, requiring careful regularization and early stopping in practical implementations, random forest exemplifies bagging with additional feature randomization, resulting in robust performance across many domains with minimal tuning adaboost, gradient boosting, and xgboost represent boosting approaches with increasingly sophisticated optimizations for performance and overfitting control the choice between them depends on data characteristics, model complexity needs, and computational constraints i often use random forests when stability and interpretability are priorities, and boosting methods when maximizing predictive performance is the primary goal " mediocre response "bagging and boosting are both ensemble techniques but work differently bagging (bootstrap aggregating) trains models in parallel on random subsets of data each model is independent reduces variance but doesn't affect bias much examples include random forests more resistant to overfitting boosting trains models sequentially each model tries to correct mistakes made by previous models reduces bias and variance examples include adaboost, gradient boosting, xgboost more prone to overfitting on noisy data" poor response "bagging combines multiple models trained on different subsets of data and averages their predictions random forest is the most common example boosting builds models sequentially where each new model focuses on the mistakes of previous models xgboost is a popular boosting algorithm boosting usually gives better performance but is more likely to overfit, while bagging is more stable but might not achieve the same level of accuracy " 14\ how would you address the cold start problem in a recommendation system? great response "the cold start problem in recommendation systems occurs when we lack sufficient interaction data for new users or items i tackle this through a multi faceted approach for new users, i implement content based initial recommendations using explicitly gathered preferences during onboarding demographic based recommendations leveraging similarities to existing user segments popularity based recommendations with strategic diversity to probe user interests active learning approaches that present diverse item sets to quickly learn preferences for new items, i utilize content based methods using item metadata and natural language processing of descriptions item similarity models based on product attributes rather than user interactions controlled exposure strategies within recommendation slots dedicated to new items transfer learning from similar items with established histories hybrid approaches are particularly effective, so i'd combine multiple strategies into an ensemble that weights different signals appropriately as interaction data accumulates i'd implement exploration exploitation frameworks like multi armed bandits that balance showing reliable recommendations with gathering information about new preferences in production, i'd build an analytics framework to measure how quickly the system overcomes cold start issues, tracking metrics like time to first meaningful engagement for new users and exposure to conversion rates for new items this data would inform continuous refinement of the cold start strategy " mediocre response "the cold start problem occurs when a recommendation system doesn't have enough data for new users or items to address this, i would for new users ask users for their preferences during registration use demographic information to match them with similar users recommend popular items initially implement a hybrid approach combining content and collaborative filtering for new items use content based features of the items expose new items to diverse users to gather initial feedback implement a/b testing to evaluate different strategies i would gradually transition to collaborative filtering as more interaction data becomes available " poor response "for the cold start problem, i would simply recommend the most popular items to new users until we collect enough data about them for new items, i'd add them to recommendations for some users to start gathering data another approach would be to ask users for their preferences directly or use their demographic information to make initial recommendations once we have more interaction data, the regular recommendation algorithm would take over " 15\ explain how you would detect and handle outliers in your dataset great response "my approach to outlier detection combines statistical methods, visualization techniques, and domain knowledge in a workflow that fits the specific data context first, i'd start with exploratory visualization using box plots, histograms, and scatter plots to get an intuitive understanding of the data distribution and identify potential anomalies for univariate outlier detection, i'd employ multiple statistical methods in parallel z score or modified z score (using median and mad) for roughly normal distributions iqr method (flagging points beyond 1 5×iqr from quartiles) for skewed distributions percentile based approaches for heavy tailed distributions for multivariate outliers, which are harder to detect visually, i'd implement mahalanobis distance for correlated variables with roughly multivariate normal distribution isolation forest or local outlier factor for complex data without distributional assumptions dbscan clustering to identify points in low density regions the handling strategy would depend on the root cause analysis for measurement errors, i'd correct if possible or remove if necessary for valid but extreme values, i'd consider transformation (log, box cox) to reduce influence for influential observations, i'd run models with and without them to quantify impact for intentional fraud/anomalies, i'd preserve them if anomaly detection is the goal most importantly, i'd consult domain experts to validate whether statistical outliers represent genuine anomalies in the business context, as what's statistically unusual may be businessly important " mediocre response "to detect outliers, i would use a combination of statistical methods like z score or iqr (interquartile range) visualization techniques like box plots and scatter plots machine learning algorithms like isolation forest or local outlier factor for multivariate outliers once identified, i would handle outliers by investigating their source to determine if they're errors or genuine extreme values removing them if they're clearly errors or if they'd significantly bias the model capping them at a threshold (winsorization) to reduce their influence using robust algorithms that are less sensitive to outliers the specific approach would depend on the context and the amount of data available " poor response "i would identify outliers using the iqr method, which considers values beyond 1 5 times the interquartile range as outliers i could also use z scores and flag values that are more than 3 standard deviations from the mean once identified, i'd usually remove these outliers from the dataset since they can negatively impact model performance if removing them reduces the dataset too much, i might cap them at a certain value instead " 16\ what is the difference between a parametric and non parametric model? give examples of each great response "parametric and non parametric models differ fundamentally in their assumptions about data structure and how they approach learning from data parametric models assume a fixed functional form for the relationship between features and target, specified by a finite number of parameters the complexity of these models is bounded regardless of data size examples include linear and logistic regression, where we estimate coefficients for each feature neural networks with fixed architecture, where we learn weights and biases naive bayes, which estimates class conditional probabilities parametric statistical distributions like gaussian or poisson the key advantage of parametric models is interpretability and computational efficiency, but they risk underfitting if their functional form doesn't match the true data distribution non parametric models make minimal assumptions about the underlying function and let the complexity grow with data size examples include k nearest neighbors, which makes predictions based on similar training examples decision trees and their ensembles (random forests, gradient boosting), which partition the feature space adaptively kernel methods like svms with non linear kernels gaussian processes, which define distributions over functions non parametric models offer flexibility to capture complex patterns but risk overfitting with insufficient regularization and typically require more data to generalize well in practice, i often use parametric models when i have strong prior knowledge about the relationship form or need highly interpretable results, and non parametric models when exploring complex, unknown relationships where flexibility is paramount " mediocre response "parametric models assume a specific functional form for the relationship between variables and estimate a fixed number of parameters regardless of the data size examples include linear regression, logistic regression, and naive bayes they're generally faster and more interpretable but can be limited in their flexibility non parametric models don't assume a specific functional form and let the complexity grow with the amount of data examples include decision trees, random forests, k nearest neighbors, and kernel svm they're more flexible for capturing complex relationships but may require more data and be more prone to overfitting the choice between them depends on the complexity of the relationship you're trying to model and how much data you have available " poor response "parametric models have a fixed number of parameters regardless of how much data you have, like linear regression where you just need to find the coefficients non parametric models don't have a fixed structure and can grow more complex as you get more data, like decision trees or k nearest neighbors parametric models are usually simpler and faster, but non parametric models can capture more complex relationships "