Prediction understanding¶

Setup¶

In [ ]:

pip install ydf -U

pip install ydf -U

In [1]:

import ydf
import pandas as pd

import ydf import pandas as pd

What is prediction understanding?¶

Prediction understanding aims explain a single model prediction.

Prediction understanding is different from model understanding which aims to explain a model as a whole. To explain models, see the model understanding tutorial instead.

In YDF, prediction understanding is simply done with model.analyze_prediction().

Information: Counter-factual analysis is complementary and more complex way to understand predictions. See the counterfactual notebook for details.

Gathering dataset and training model¶

In [5]:

ds_path = "https://raw.githubusercontent.com/google/yggdrasil-decision-forests/main/yggdrasil_decision_forests/test_data/dataset"
train_ds = pd.read_csv(f"{ds_path}/adult_train.csv")
test_ds = pd.read_csv(f"{ds_path}/adult_test.csv")

# Print the first 5 training examples
train_ds.head(5)

# Train a model
model = ydf.GradientBoostedTreesLearner(label="income").train(train_ds)

ds_path = "https://raw.githubusercontent.com/google/yggdrasil-decision-forests/main/yggdrasil_decision_forests/test_data/dataset" train_ds = pd.read_csv(f"{ds_path}/adult_train.csv") test_ds = pd.read_csv(f"{ds_path}/adult_test.csv") # Print the first 5 training examples train_ds.head(5) # Train a model model = ydf.GradientBoostedTreesLearner(label="income").train(train_ds)

Train model on 22792 examples
Model trained in 0:00:01.120417

Prediction analysis¶

In contrast to model analysis (model.analyze), which examines a model as a whole, prediction analysis (model.analyze_prediction) explains a single prediction of the model. The next example explains the model's prediction on the first test example.

In [6]:

model.analyze_prediction(test_ds.iloc[:1])

model.analyze_prediction(test_ds.iloc[:1])

Out[6]:

Feature VariationSHAP values

Feature	>50K
capital_gain	-0.987081 ----------------|
marital_status	-0.703905 -----------|
education	0.497574 |++++++++
relationship	-0.457539 -------|
age	0.361084 |++++++
workclass	-0.308857 -----|
fnlwgt	-0.201710 ---|
sex	0.180184 |+++
occupation	-0.136641 --|
education_num	0.118709 |++
hours_per_week	-0.094915 --|
native_country	0.037439 |+
capital_loss	-0.011167 |
race	0.007206 |

Counterfactual examples¶

Counterfactual examples are the training examples that are most similar to a prediction according to a model. Counterfactual examples can be used to explain the model's prediction by examining the similarities and differences between their features.

For more information, see the standalone counterfactual notebook.