Model Predictions With Local Explanations

Although many models tend to operate as proverbial black boxes, understanding the rationale behind the model’s predictions helps consumers of the output decide whether or not to trust the model.

Having the ability to interpret your model predictions in order to explain them to knowledgable stakeholders and lay folks alike is a crucial aspect of building that trust. It can also help you understand and assess, at the model developer level, key feature impacts on your model's predictions, so you can achieve ever higher levels of trust resulting from consistently reliable explanations about model behavior and output.

Before diving deeper, it's important to appreciate the difference between interpretion and explanation.

A interpretation is the mapping of an abstract concept — a predicted class in a classification problem, for instance — into a domain that a human consumer of the model's output can easily (hopefully) understand, bearing in mind that there are non-interpretable domains — embeddings and undocumented input features, among others — that aren't quite so easy for some consumers to get their minds around.

An explanation is the collection of features belonging to the interpretable domain that contribute to producing a decision — a classification or regression result. For this reason, an explanation is typically computed at a finer grain than an interpretation.

It almost always boils down to the difference between "what?" and "why?" The interpretation is the what. The explanation is the why.

For now, we'll limit the discussion to two flows in a local compute context: Basic Flow and PySpark Flow. Local compute refers to using TruEra's Python SDK to ingest a model into TruEra from your local machine, rather than from a remote system. PySpark is the Python API for Apache Spark.

For those wishing to go deeper, each of the following links shares a respective tutorial:

• Basic Flow
• PySpark Flow

Click one of the links above or click Next below to continue.