Domain knowledge is expertise about the subject matter and phenomena represented by a dataset. It turns business understanding into modeling decisions rather than treating features as interchangeable columns. section 7.4
Feature engineering is a primary way to apply domain knowledge: encode how the problem works, create useful predictors, and represent information in a form the model can learn. Deep learning can discover representations automatically, but domain-specific insights may still improve particular applications. section 7.4 section 1.3
Domain knowledge supports practical choices that statistical analysis alone may not settle:
- Identify a likely causal feature among duplicated or highly collinear features, then remove redundant alternatives. section 2.2.2
- Choose which feature to retain when the underlying cause is not recorded, using theoretical deductions, causal analysis, target association, and data quality. section 2.2.2
- Set missing-value defaults according to the meaning of each field instead of applying one universal replacement rule. section 9.2.1
- Shape exploratory analysis, validation, pipelines, baselines, and model tuning around the problem’s targets and predictors. section 7.4
Real-estate missing values
For Bathrooms, use the column median because residential properties rarely have no washrooms. For Car Parks, use zero because many properties have no parking spaces. For categorical fields, use explicit unknown values such as unknown_furnishing, unknown_property, and unknown_location. section 9.2.1
Listing 9.4 Defining the default replacement values for missing values
misc_col_dict: # default values to replace missing values for general columns Bathrooms: median #1 Car Parks: 0 #2 Furnishing: unknown_furnishing #3 Property Type: unknown_property Location: unknown_location
Inspect the configuration pattern that records domain-informed defaults for general columns before applying them in a cleanup function. section 9.2.1
