Introduction to sensAI: Supervised Learning with VectorModels#
Show code cell content
%load_ext autoreload
%autoreload 2
Show code cell content
import sys
sys.path.append("../src")
import sensai
import numpy as np
Logging#
sensAI will log relevant activies and inform about ongoing processes as well as results via the log. It is therefore highly recommended that logging be enabled when using sensAI.
sensAI provides a logging module which includes Python’s standard logging module and adds some additional functionality. To enable logging, simply use its configureLogging function.
from sensai.util import logging
logging.configure(level=logging.INFO)
To additionally write log output to a file, use the function logging.add_file_logger.
VectorModels#
The central base class for supervised learning problems in sensAI is VectorModel. A VectorModel is any model which operates on data points that can be reprsented as vectors of data. Here, vector is to be understood not in the mathematical sense but in the computer science sense, where a vector is simply an array of (potentially arbitaririly complex) data. (The mathematical equivalent is a tuple.) Models are typically expected to be able to process more than one data point at a time and thus should be able to process a sequence of vectors.
We use pandas DataFrames to represent such sequences of data points. Note that pandas DataFrames are not limited to primitive data types but can hold arbitrary objects in every cell. When dealing with a large number of inputs, DataFrames also provide at least limited meta-information in the form of column names, so we do not lose track of what is contained in which element of a data point (vector).
VectorModel itself is an abstract base class, which provides a lot of useful functionality that all its specialisations inherit (as we will see later, particularly in the more advanced tutorials). The class is specialised in VectorClassificationModel and VectorRegressionModel, which in turn are specialised for various machine learning frameworks (such as sklearn and PyTorch) or can be directly subclassed to create your own model.
In this tutorial, we will be dealing with a classification problem. Therefore, we will apply subclasses of VectorClassificationModel such as SkLearnRandomForestVectorClassificationModel. As an sklearn classification model which uses a well-defined training and inference interface, the implementation of the class is essentially justa few lines of code (given the intermediate abstraction AbstractSkLearnVectorClassificationModel for all classification models that use the sklearn protocol).
Training and Evaluating Models#
First, let us load a dataset which we can experiment. sklearn provides, for example, the Iris classification dataset, where the task is to differentiate three different types of flowers based on measurements of their petals and sepals.
import sklearn.datasets
import pandas as pd
iris_data = sklearn.datasets.load_iris()
iris_input_df = pd.DataFrame(iris_data["data"], columns=iris_data["feature_names"]).reset_index(drop=True)
iris_output_df = pd.DataFrame({"class": [iris_data["target_names"][idx] for idx in iris_data["target"]]}).reset_index(drop=True)
Here’s a sample of the data, combining both the inputs and outputs:
iris_combined_df = pd.concat((iris_input_df, iris_output_df), axis=1)
iris_combined_df.sample(10)
| sepal length (cm) | sepal width (cm) | petal length (cm) | petal width (cm) | class | |
|---|---|---|---|---|---|
| 133 | 6.3 | 2.8 | 5.1 | 1.5 | virginica |
| 112 | 6.8 | 3.0 | 5.5 | 2.1 | virginica |
| 32 | 5.2 | 4.1 | 1.5 | 0.1 | setosa |
| 149 | 5.9 | 3.0 | 5.1 | 1.8 | virginica |
| 6 | 4.6 | 3.4 | 1.4 | 0.3 | setosa |
| 129 | 7.2 | 3.0 | 5.8 | 1.6 | virginica |
| 100 | 6.3 | 3.3 | 6.0 | 2.5 | virginica |
| 103 | 6.3 | 2.9 | 5.6 | 1.8 | virginica |
| 63 | 6.1 | 2.9 | 4.7 | 1.4 | versicolor |
| 91 | 6.1 | 3.0 | 4.6 | 1.4 | versicolor |
When working with sensAI, we typically use DataFrames such as this as the starting point.
We create an instance of InputOutputData from the two data frames.
iris_io_data = sensai.InputOutputData(iris_input_df, iris_output_df)
Low-Level Training and Inference#
We use a DataSplitter (see subclasses) to split the data into a training and test set, specifically a DataSplitterFractional.
data_splitter = sensai.data.DataSplitterFractional(0.8, shuffle=True)
training_io_data, test_io_data = data_splitter.split(iris_io_data)
Now we are ready to train a model. Let us train a random forest classifier, which should work well for this sort of problem. sensAI provides models from various libraries, including scikit-learn, PyTorch, lightgbm, xgboost, catboost, and TensorFlow.
In this case, let us use the random forest implementation from sklearn, which is provided via the wrapper class SkLearnRandomForestVectorClassificationModel.
sensAI’s VectorModel classes (specialised for classification and regression) provide a common interface with a lot of useful functionality, which we will see later.
random_forest_model = sensai.sklearn.classification.SkLearnRandomForestVectorClassificationModel(
min_samples_leaf=2).with_name("RandomForest")
The class suppports all the parameters supported by the original sklearn model. In this case, we only set the minimum number of samples that must end up in each leaf.
We train the model using the fitInputOutputData method; we could also use the fit method, which is analogous to the sklearn interface and takes two arguments (input, output).
random_forest_model.fit_input_output_data(training_io_data)
random_forest_model
INFO 2024-08-13 22:11:24,221 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:24,222 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:24,318 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
SkLearnRandomForestVectorClassificationModel[id=139958354930416, featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
We can now apply the trained model and predict the outputs for the test set we reserved.
predicted_outputs_df = random_forest_model.predict(test_io_data.inputs)
predicted_outputs_df.head(5)
| class | |
|---|---|
| 91 | versicolor |
| 41 | setosa |
| 58 | versicolor |
| 90 | versicolor |
| 48 | setosa |
Let’s compare some of the predictions to the ground truth.
pd.concat((predicted_outputs_df.rename(columns={"class": "predictedClass"}), test_io_data.outputs), axis=1).sample(10)
| predictedClass | class | |
|---|---|---|
| 107 | virginica | virginica |
| 87 | versicolor | versicolor |
| 149 | virginica | virginica |
| 57 | versicolor | versicolor |
| 37 | setosa | setosa |
| 90 | versicolor | versicolor |
| 99 | versicolor | versicolor |
| 124 | virginica | virginica |
| 88 | versicolor | versicolor |
| 20 | setosa | setosa |
Using the ground truth and predicted values, we could now compute the metrics we’re interested in. We could, for example, use the metrics implemented in sklearn to analyse the result. Yet sensAI already provides abstractions that facilitate the generation of metrics and the collection of results. Read on!
Using Evaluators#
sensAI provides evaluator abstractions which facilitate the training and evaluation of models.
For a classification problem, we instantiate a VectorClassificationModelEvaluator. An evaluator serves to evaluate one or more models based on the same data, so we construct it with the data and instructions on how to handle/split the data for evaluation.
evaluator_params = sensai.evaluation.ClassificationEvaluatorParams(data_splitter=data_splitter, compute_probabilities=True)
evaluator = sensai.evaluation.VectorClassificationModelEvaluator(iris_io_data, params=evaluator_params)
We can use this evaluator to evaluate one or more models. Let us evaluate the random forest model from above.
evaluator.fit_model(random_forest_model)
eval_data = evaluator.eval_model(random_forest_model)
INFO 2024-08-13 22:11:24,414 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:24,415 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:24,509 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
The evaluation data holds, in particular, an EvalStats object, which can provide data on the quality of the results. Depending on the type of problem, many metrics will already be computed by default.
eval_stats = eval_data.get_eval_stats()
eval_stats
ClassificationEvalStats[id=139957312899632, accuracy=0.9666666666666667, balancedAccuracy=0.9722222222222222, geoMeanTrueClassProb=0.8890980234758366, N=30]
We can get the metrics in a dictionary as follows:
eval_stats.metrics_dict()
{'accuracy': 0.9666666666666667,
'balancedAccuracy': 0.9722222222222222,
'geoMeanTrueClassProb': 0.8890980234758366}
We can compute additional metrics by passing a metric to the compute_metric_value method, but we could also have added additional metrics to the evaluator_params above and have the metric included in all results.
Let’s see how frequently the true class is among the top two most probable classes.
eval_stats.compute_metric_value(sensai.eval_stats_classification.ClassificationMetricTopNAccuracy(2))
1.0
The EvalStats object can also be used to generate plots, such as a confusion matrix or a precision-recall plot for binary classification.
eval_stats.plot_confusion_matrix(normalize=True);
Using the Fully-Integrated Evaluation Utilities#
sensAI’s evaluation utilities take things one step further and assist you in out all the evaluation steps and results computations in a single call.
You can perform evaluations based on a single split or cross-validation. We simply declare the necessary parameters for both types of computations (or the one type we seek to carry out).
evaluatorParams = sensai.evaluation.ClassificationEvaluatorParams(
data_splitter=data_splitter, compute_probabilities=True,
additional_metrics=[sensai.eval_stats_classification.ClassificationMetricTopNAccuracy(2)])
cross_validator_params = sensai.evaluation.crossval.VectorModelCrossValidatorParams(folds=10,
evaluator_params=evaluator_params)
eval_util = sensai.evaluation.ClassificationModelEvaluation(iris_io_data,
evaluator_params=evaluatorParams, cross_validator_params=cross_validator_params)
In practice, we will usually want to save evaluation results. The evaluation methods of eval_util take a parameter result_writer which allows us to define where results shall be written. Within this notebook, we shall simply inspect the resulting metrics in the log that is printed, and we shall configure plots to be shown directly.
Simple Evaluation#
We can perform the same evaluation as above (which uses a single split) like so:
eval_util.perform_simple_evaluation(random_forest_model, show_plots=True)
INFO 2024-08-13 22:11:24,947 sensai.evaluation.eval_util:perform_simple_evaluation:286 - Evaluating SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)] via <sensai.evaluation.evaluator.VectorClassificationModelEvaluator object at 0x7f4a59cf30a0>
INFO 2024-08-13 22:11:24,947 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:24,949 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:25,060 sensai.vector_model:fit:400 - Fitting completed in 0.11 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:25,072 sensai.evaluation.eval_util:gather_results:294 - Evaluation results for class: ClassificationEvalStats[accuracy=0.9666666666666667, balancedAccuracy=0.9722222222222222, geoMeanTrueClassProb=0.8890980234758366, top2Accuracy=1.0, N=30]
<sensai.evaluation.evaluator.VectorClassificationModelEvaluationData at 0x7f4a59e92fd0>
Customising the Set of Plots#
If we decide that we don’t really want to have the normalised confusion matrix, we can disable it for any further experiments.
eval_util.eval_stats_plot_collector.get_enabled_plots()
['confusion-matrix-rel',
'confusion-matrix-abs',
'precision-recall',
'threshold-precision-recall',
'threshold-counts']
Some of these are only active for binary classification. The one we don’t want is “confusion-matrix-rel”.
eval_util.eval_stats_plot_collector.disable_plots("confusion-matrix-rel")
We could also define our own plot class (by creating a new subclass of ClassificationEvalStatsPlot) and add it to the plot collector in order to have the plot auto-generated whenever we apply one of eval_util’s methods.
Cross-Validation#
We can similarly run cross-validation and produce the respective evaluation metrics with a single call.
eval_util.perform_cross_validation(random_forest_model, show_plots=True)
INFO 2024-08-13 22:11:25,673 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 1/10 ...
INFO 2024-08-13 22:11:25,673 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:25,674 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:25,770 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:25,784 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 1/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9553599022560102, N=15]
INFO 2024-08-13 22:11:25,784 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 2/10 ...
INFO 2024-08-13 22:11:25,785 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:25,786 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:25,880 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:25,894 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 2/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9734228107223192, N=15]
INFO 2024-08-13 22:11:25,895 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 3/10 ...
INFO 2024-08-13 22:11:25,896 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:25,897 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:25,993 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,005 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 3/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.976458969438705, N=15]
INFO 2024-08-13 22:11:26,006 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 4/10 ...
INFO 2024-08-13 22:11:26,006 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,007 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,102 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,114 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 4/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9333333333333332, geoMeanTrueClassProb=0.9411926691126593, N=15]
INFO 2024-08-13 22:11:26,115 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 5/10 ...
INFO 2024-08-13 22:11:26,116 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,117 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,212 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,223 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 5/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9298964965102043, N=15]
INFO 2024-08-13 22:11:26,224 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 6/10 ...
INFO 2024-08-13 22:11:26,225 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,226 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,319 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,331 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 6/10: ClassificationEvalStats[accuracy=0.8666666666666667, balancedAccuracy=0.8888888888888888, geoMeanTrueClassProb=0.6379522029160006, N=15]
INFO 2024-08-13 22:11:26,332 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 7/10 ...
INFO 2024-08-13 22:11:26,332 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,334 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,428 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,440 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 7/10: ClassificationEvalStats[accuracy=0.8666666666666667, balancedAccuracy=0.8611111111111112, geoMeanTrueClassProb=0.8099785263057822, N=15]
INFO 2024-08-13 22:11:26,440 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 8/10 ...
INFO 2024-08-13 22:11:26,441 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,442 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,538 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,550 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 8/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9657414983648998, N=15]
INFO 2024-08-13 22:11:26,550 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 9/10 ...
INFO 2024-08-13 22:11:26,551 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,552 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,646 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,659 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 9/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9560552426688266, N=15]
INFO 2024-08-13 22:11:26,660 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 10/10 ...
INFO 2024-08-13 22:11:26,660 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:26,662 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:26,755 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:26,767 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 10/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9523809523809524, geoMeanTrueClassProb=0.841736920530503, N=15]
INFO 2024-08-13 22:11:26,781 sensai.evaluation.eval_util:perform_cross_validation:349 - Cross-validation results:
mean[accuracy] std[accuracy] mean[balancedAccuracy] std[balancedAccuracy] mean[geoMeanTrueClassProb] std[geoMeanTrueClassProb]
class 0.96 0.053333 0.963571 0.050077 0.89878 0.10223
<sensai.evaluation.crossval.VectorClassificationModelCrossValidationData at 0x7f4a59862af0>
As you can see, the plot we disabled earlier is no longer being generated.
Comparing Models#
A most common use case is to compare the performance of several models. The evaluation utility makes it very simple to compare any number of models.
Let’s say we want to compare the random forest we have been using thus far to a simple decision tree.
results = eval_util.compare_models([
random_forest_model,
sensai.sklearn.classification.SkLearnDecisionTreeVectorClassificationModel(min_samples_leaf=2).with_name("DecisionTree")],
use_cross_validation=True)
INFO 2024-08-13 22:11:27,098 sensai.evaluation.eval_util:compare_models:398 - Evaluating model 1/2 named 'RandomForest' ...
INFO 2024-08-13 22:11:27,103 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 1/10 ...
INFO 2024-08-13 22:11:27,104 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,105 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,200 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,214 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 1/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9553599022560102, N=15]
INFO 2024-08-13 22:11:27,215 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 2/10 ...
INFO 2024-08-13 22:11:27,215 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,216 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,312 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,324 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 2/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9734228107223192, N=15]
INFO 2024-08-13 22:11:27,324 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 3/10 ...
INFO 2024-08-13 22:11:27,325 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,326 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,422 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,434 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 3/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.976458969438705, N=15]
INFO 2024-08-13 22:11:27,435 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 4/10 ...
INFO 2024-08-13 22:11:27,435 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,436 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,531 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,546 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 4/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9333333333333332, geoMeanTrueClassProb=0.9411926691126593, N=15]
INFO 2024-08-13 22:11:27,546 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 5/10 ...
INFO 2024-08-13 22:11:27,547 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,548 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,643 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,655 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 5/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9298964965102043, N=15]
INFO 2024-08-13 22:11:27,656 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 6/10 ...
INFO 2024-08-13 22:11:27,656 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,658 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,753 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,764 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 6/10: ClassificationEvalStats[accuracy=0.8666666666666667, balancedAccuracy=0.8888888888888888, geoMeanTrueClassProb=0.6379522029160006, N=15]
INFO 2024-08-13 22:11:27,765 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 7/10 ...
INFO 2024-08-13 22:11:27,766 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,767 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,862 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,876 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 7/10: ClassificationEvalStats[accuracy=0.8666666666666667, balancedAccuracy=0.8611111111111112, geoMeanTrueClassProb=0.8099785263057822, N=15]
INFO 2024-08-13 22:11:27,876 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 8/10 ...
INFO 2024-08-13 22:11:27,877 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,878 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:27,976 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:27,987 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 8/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9657414983648998, N=15]
INFO 2024-08-13 22:11:27,987 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 9/10 ...
INFO 2024-08-13 22:11:27,988 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:27,989 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:28,085 sensai.vector_model:fit:400 - Fitting completed in 0.10 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,098 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 9/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=0.9560552426688266, N=15]
INFO 2024-08-13 22:11:28,098 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 10/10 ...
INFO 2024-08-13 22:11:28,099 sensai.vector_model:fit:371 - Fitting SkLearnRandomForestVectorClassificationModel instance
INFO 2024-08-13 22:11:28,100 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type RandomForestClassifier
INFO 2024-08-13 22:11:28,194 sensai.vector_model:fit:400 - Fitting completed in 0.09 seconds: SkLearnRandomForestVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=RandomForest, model=RandomForestClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,207 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 10/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9523809523809524, geoMeanTrueClassProb=0.841736920530503, N=15]
INFO 2024-08-13 22:11:28,221 sensai.evaluation.eval_util:perform_cross_validation:349 - Cross-validation results:
mean[accuracy] std[accuracy] mean[balancedAccuracy] std[balancedAccuracy] mean[geoMeanTrueClassProb] std[geoMeanTrueClassProb]
class 0.96 0.053333 0.963571 0.050077 0.89878 0.10223
INFO 2024-08-13 22:11:28,312 sensai.evaluation.eval_util:compare_models:398 - Evaluating model 2/2 named 'DecisionTree' ...
INFO 2024-08-13 22:11:28,313 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 1/10 ...
INFO 2024-08-13 22:11:28,313 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,314 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,318 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,326 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 1/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=1.0, N=15]
INFO 2024-08-13 22:11:28,326 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 2/10 ...
INFO 2024-08-13 22:11:28,328 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,330 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,336 sensai.vector_model:fit:400 - Fitting completed in 0.01 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,345 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 2/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=1.0, N=15]
INFO 2024-08-13 22:11:28,345 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 3/10 ...
INFO 2024-08-13 22:11:28,348 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,349 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,353 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,362 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 3/10: ClassificationEvalStats[accuracy=1.0, balancedAccuracy=1.0, geoMeanTrueClassProb=1.0, N=15]
INFO 2024-08-13 22:11:28,363 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 4/10 ...
INFO 2024-08-13 22:11:28,364 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,365 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,369 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,377 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 4/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9333333333333332, geoMeanTrueClassProb=0.6141303814089187, N=15]
INFO 2024-08-13 22:11:28,377 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 5/10 ...
INFO 2024-08-13 22:11:28,378 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,379 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,382 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,390 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 5/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9333333333333332, geoMeanTrueClassProb=0.9548416039104165, N=15]
INFO 2024-08-13 22:11:28,391 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 6/10 ...
INFO 2024-08-13 22:11:28,393 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,394 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,398 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,405 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 6/10: ClassificationEvalStats[accuracy=0.8666666666666667, balancedAccuracy=0.8888888888888888, geoMeanTrueClassProb=0.39810717055349726, N=15]
INFO 2024-08-13 22:11:28,406 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 7/10 ...
INFO 2024-08-13 22:11:28,408 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,408 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,412 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,418 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 7/10: ClassificationEvalStats[accuracy=0.8666666666666667, balancedAccuracy=0.8611111111111112, geoMeanTrueClassProb=0.39810717055349726, N=15]
INFO 2024-08-13 22:11:28,419 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 8/10 ...
INFO 2024-08-13 22:11:28,419 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,420 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,423 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,429 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 8/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9444444444444445, geoMeanTrueClassProb=0.9548416039104165, N=15]
INFO 2024-08-13 22:11:28,429 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 9/10 ...
INFO 2024-08-13 22:11:28,430 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,431 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,434 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,441 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 9/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9333333333333332, geoMeanTrueClassProb=0.6141303814089187, N=15]
INFO 2024-08-13 22:11:28,441 sensai.evaluation.crossval:eval_model:192 - Training and evaluating model with fold 10/10 ...
INFO 2024-08-13 22:11:28,441 sensai.vector_model:fit:371 - Fitting SkLearnDecisionTreeVectorClassificationModel instance
INFO 2024-08-13 22:11:28,442 sensai.sklearn.sklearn_base:_fit_classifier:281 - Fitting sklearn classifier of type DecisionTreeClassifier
INFO 2024-08-13 22:11:28,446 sensai.vector_model:fit:400 - Fitting completed in 0.00 seconds: SkLearnDecisionTreeVectorClassificationModel[featureGenerator=None, rawInputTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], featureTransformerChain=DataFrameTransformerChain[dataFrameTransformers=[]], fitArgs={}, useBalancedClassWeights=False, useLabelEncoding=False, name=DecisionTree, model=DecisionTreeClassifier(min_samples_leaf=2, random_state=42)]
INFO 2024-08-13 22:11:28,452 sensai.evaluation.crossval:eval_model:201 - Evaluation result for class, fold 10/10: ClassificationEvalStats[accuracy=0.9333333333333333, balancedAccuracy=0.9523809523809524, geoMeanTrueClassProb=0.6309573444801932, N=15]
INFO 2024-08-13 22:11:28,466 sensai.evaluation.eval_util:perform_cross_validation:349 - Cross-validation results:
mean[accuracy] std[accuracy] mean[balancedAccuracy] std[balancedAccuracy] mean[geoMeanTrueClassProb] std[geoMeanTrueClassProb]
class 0.94 0.046667 0.944683 0.04441 0.756512 0.238693
INFO 2024-08-13 22:11:28,561 sensai.evaluation.eval_util:compare_models:467 - Model comparison results, aggregated across folds:
mean[accuracy] std[accuracy] mean[balancedAccuracy] std[balancedAccuracy] mean[geoMeanTrueClassProb] std[geoMeanTrueClassProb]
model_name
RandomForest 0.96 0.053333 0.963571 0.050077 0.898780 0.102230
DecisionTree 0.94 0.046667 0.944683 0.044410 0.756512 0.238693
In addition to the data frame with the aggregated metrics, which was already printed to the log, the results object contains all the data that was generated during the evaluation. We can, for example, use it to plot the distribution of one of the metrics across all the folds for one of our models.
display(results.results_df)
esc_random_forest = results.result_by_model_name["RandomForest"].cross_validation_data.get_eval_stats_collection()
esc_random_forest.plot_distribution("accuracy", bins=np.linspace(0,1,21), stat="count", kde=False);
| mean[accuracy] | std[accuracy] | mean[balancedAccuracy] | std[balancedAccuracy] | mean[geoMeanTrueClassProb] | std[geoMeanTrueClassProb] | |
|---|---|---|---|---|---|---|
| model_name | ||||||
| RandomForest | 0.96 | 0.053333 | 0.963571 | 0.050077 | 0.898780 | 0.102230 |
| DecisionTree | 0.94 | 0.046667 | 0.944683 | 0.044410 | 0.756512 | 0.238693 |
We can also compute additional aggregations or inspect the full list of metrics.
esc_random_forest.agg_metrics_dict(agg_fns=[np.max, np.min])
{'amax[accuracy]': 1.0,
'amin[accuracy]': 0.8666666666666667,
'amax[balancedAccuracy]': 1.0,
'amin[balancedAccuracy]': 0.8611111111111112,
'amax[geoMeanTrueClassProb]': 0.976458969438705,
'amin[geoMeanTrueClassProb]': 0.6379522029160006}
esc_random_forest.get_values("accuracy")
[1.0,
1.0,
1.0,
0.9333333333333333,
1.0,
0.8666666666666667,
0.8666666666666667,
1.0,
1.0,
0.9333333333333333]
Feature Generators and Data Frame Transformers#
When dealing with the preparation of input data for a model, we often need to cater to technical requirements of various types of models. sensAI seeks to make the process of supporting multiple input pipelines for different types of models as simple as possible - by focusing on concise, declarative semantics and integrating the model-specific data extraction and transformation mechanisms into the models themselves. In essence, this means:
Starting with the raw or most general representation of the data
This could mean simply starting with the data that is straightforward for us to obtain - or using directly using particular domain specific objects.
For example, if the problem is to classify situations, we might already have a
Situationclass in our code which represents all the data that is is relevant to a situation (e.g. the point in time, the affected user, the location, etc.). Pandas DataFrames can represent arbitrary data, so there is no reason to not simply use as the raw input data frame that is fed to our models a single column containing instances of classSituation. Or we might instead directly observe a set of sensor readings, all of which are real numbers; this scenario would certainly be closer to what we typically see in machine learning data sets, but it isn’t always the case in the real world.Whatever the case may be, we can represent it in a data frame. We call the original input data frame, which we pass to a sensAI
VectorModel, the raw data frame.Extracing features from the raw data, using their “natural” representation (using
FeatureGenerators)We extract from the raw data frame pieces of information that we regard as relevant features for the task at hand. A sensAI
FeatureGeneratorcan generate one or more data frame columns (containing arbitrary data), and a model can be associated with any number of feature generators. Several key aspects:FeatureGenerators crititcally decouple the original raw data from the features used by the models, enabling different models to use different sets of features or entirely different representations of the same features.
FeatureGenerators become part of the model and are (where necessary) jointly trained with model. This facilitates model deployment, as every sensAI model becomes a single unit that can directly process raw input data, which is (usually) straightforward to supply at inference time.
FeatureGenerators store meta-data on the features they generate, enabling downstream components to handle them appropriately.
The feature representation that we choose to generate can be arbitrary, but oftentimes, we will want to extract “natural” feature representations, which could, in priciple, be used by many types of models, albeit in different concrete forms. Sequential data can be naturally represented as an array/list, categorical data can be represented using descriptive category names, and numeric data can be represented using unmodified integers and floating point numbers.
Transforming feature representations into a form that is suitable for the model at hand (using
DataFrameTransformers)In the transformation stage, we address the model-specific idiosynchrasies, which may require, for example, that all features be represented as numbers (or even numbers within a limited range) or that all features be discrete, that no values be missing, etc. A
DataFrameTransformercan, in principle perform an arbitary transformation from one data frame to another, but the typical use case is to apply transformations of feature representations that are necessary for specific types of models to work (their best).