04-epilogue.Rmd

# Epilogue

Let's summarize what has happened in the previous sections.

* Section \@ref(useCaseApartmetns) shows two models with equal performance for `apartments` dataset.
* Section \@ref(modelPerformance) shows that in general the random forest model has smaller residuals than the linear model but there is a small fraction of very large residuals.
* Section \@ref(outlierDetection) shows that the random forest model under-predicts expensive apartments. It is not a model that we would like to employ.
* Section \@ref(featureImportance) shows that `construction_year` is  important for the random forest model. 
* Section \@ref(variableResponse) shows that the relation between `construction_year` and the price of square meter is non linear. 

In this section we showed how to improve the basic linear model by feature engineering of `construction_year`. Findings from the random forest models will help to create a new feature for the linear model.

```{r final_model, message=FALSE, warning=FALSE, fig.height=2, fig.cap="Distribution of residuals for the new improved linear model"}
library("DALEX")

apartments_lm_model_improved <- lm(m2.price ~ I(construction.year < 1935 | construction.year > 1995) + surface + floor + 
                         no.rooms + district, data = apartments)

explainer_lm_improved <- explain(apartments_lm_model_improved, 
                          data = apartmentsTest[,2:6], y = apartmentsTest$m2.price)

mp_lm_improved <- model_performance(explainer_lm_improved)
plot(mp_lm_improved, geom = "boxplot")
```

In conclusion, the results presented above prove that the `apartments_lm_model_improved` model is much better than the two initial models introduced in Chapter \@ref(modelUnderstanding).

In this use-case we showed that explainers implemented in DALEX help to better understand the model and that this knowledge may be used to create a better final model.

Find more examples, vignietts and cheatsheets at DALEX website https://github.com/pbiecek/DALEX.