class: center, middle, inverse, title-slide # <font class = "title-panel"> PPOL564 | Data Science 1 | Foundations </font> <font size=6, face="bold"> Week 13 </font> <br> <br> <font size=100, face="bold"> Interpretable Machine Learning </font> ### <font class = "title-footer"> Prof. Eric Dunford ◆ Georgetown University ◆ McCourt School of Public Policy ◆ <a href="mailto:eric.dunford@georgetown.edu" class="email">eric.dunford@georgetown.edu</a></font> --- layout: true <div class="slide-footer"><span> PPOL564 | Data Science 1 | Foundations           Week 13 <!-- Week of the Footer Here -->              Interpretable Machine Learning <!-- Title of the lecture here --> </span></div> --- ## Model Interpretation - Knowing a model is predictive is _necessary_ but rarely _sufficient_ to draw **_substantive insights_**. -- - In the social sciences, we are interested in understanding **_why_** certain features matter in an effort to detect potential **_interventions_**: if we change `\(X\)` will we get a different outcome? -- - Interpretability offers insights into the features the model **_relies on to make its prediction_**. -- - In addition, interpretability is a useful debugging tool for **_detecting bias_** in machine learning models. -- - Model needs to be a fairly **_good approximation of the data generating process_** (i.e. high predictive accuracy) for interpretation to matter --- ## Variable Importance ![:space 2] - Variable/Feature importance is concerned with how much a given model **_relies on a set of variables/features to make accurate predictions_**. - If those variables/features were removed, the model should **_perform worse_** (i.e. diminished predictive capacity). - Determining variable importance helps with **_variable selection_**. - What variables could we drop from the model (not contributing much information)? - What variables should we make sure to always measure and use in the model? --- ## Variable Importance Consider output from a simple multivariate variable OLS regression, what variables seem to matter most? ![:space 3] ``` ## # A tibble: 4 x 5 ## term estimate std.error statistic p.value ## <chr> <dbl> <dbl> <dbl> <dbl> ## 1 (Intercept) -0.00776 0.0314 -0.247 0.805 ## 2 x1 2.00 0.0318 62.9 0 ## 3 x2 -0.00812 0.0312 -0.260 0.795 ## 4 x3 0.0515 0.0321 1.60 0.109 ``` -- ![:space 3] - `x1` is clearly both substantively and statistically significant. We should keep it in the model. --- ## Variable Importance ![:space 2] - Some models offer a natural way of determining importance: - _Regression_: coefficient and test statistic size - _Trees_: split importance ![:space 2] - But other models are more complicated (e.g. support vector machines, KNN, Neural Networks). We call these **_black box_** models because it's difficult to "peer inside" the model to understand how it works. ![:space 2] - We need ways of determining predictive performance that are **_model agnostic_** (i.e. doesn't depend on the type of model you use). --- ### Permutation Importance - Permutation Importance offers a model agnostic way to determine variable importance. - The idea: **_scramble the data_** one variable at a time and see if the predictive performance of the model _decreases_. -- - How it works: + **_Train_** a model + **_Permute_** (i.e. scramble the order) a single variable/feature in the training data. + Use the model to **_predict_** on the data with the permuted variable + See if there is a **_drop in predictive performance_** + **_Repeat_** --- ### The logic of permuting ![:space 5] - Permuting a variable effectively **_breaks the statistical relationship_** between outcome and predictor. - If a **_variable doesn't matter, then permuting it won't change the performance_** (as the model already doesn't rely on this variable ) - We must permute each variable **_multiple times_** as permuting is a random process + We want to ensure a specific importance ordering isn't a results of a single permutation. --- ## Partial Dependence Plots (PDP) ![:space 5] - Variable importance cannot tell us how variables **_relate_** to the outcome. - Partial dependency plots show the **_marginal effect_** one or two features have on the predicted outcome of the model. - A partial dependence plot can show whether the **_relationship_** between the target and a feature is linear, monotonic or more complex. - The partial dependence plot is a **_global method_**: The method considers all instances and gives a statement about the global relationship of a feature with the predicted outcome. --- ## Partial Dependence Plots (PDP) ![:space 5] - The steps: + Train a model + Identify the features that matter most (feature importance) + Manipulate the values of those features (one at a time) and take the average prediction, holding all other features at their observed values. + Plot the values and interpret the curve. --- ## Individual Conditional Expectation Plots (ICE) - Partial dependency offers a plot of the **_average marginal effect_**; however, can obscure a heterogeneous relationship created by **_interactions_**. <img src="interpretable-ml_files/figure-html/unnamed-chunk-2-1.png" style="display: block; margin: auto;" /> --- ## Individual Conditional Expectation Plots (ICE) - Partial dependency offers a plot of the **_average marginal effect_**; however, can obscure a heterogeneous relationship created by **_interactions_**. ![:space 3] - ICE plots plots show the **_marginal effect for each observation_** in the data. - We can observe if there are **_divergence_** or **_convergence_** in the predicted effect cross observations. - The PDP is just the average taken across the different ICE curves.