XEvolutionaryNetwork
Overview
XEvolutionaryNetwork is a novel optimisation framework for XRegressor models.
It works by taking a pre-trained XRegressor model and fitting it with training data to a network of optimisation layers. Each network layer is responsible for optimising the model weights given a set of constraints.
The inspiration for the network concept came from deep learning frameworks but is applied over additive models for weight optimisation that is understandable and explainable.
What are Layers?
Layers are the building blocks of XEvolutionaryNetwork. Each layer runs sequentially and optimises the model’s weights given a set of constraints. There are currently two types of layers:
Tighten
This layer is a leaf-boosting method that optimises the weights of each leaf node in the model. It does this by using a gradient descent method to minimise the model’s loss function. The default loss function is the mean absolute error of predictions calculated using the initial model training data.
The name Tighten comes from the visual effect that the model has when plotting the predictions of the model before and after the layer is run as the predictions are “tightened” around the training data.
The Tighten layer brings determinism to the network and is used to improve the model’s accuracy at a granular level. The deterministic nature of the layer means that it will always find a better set of weights for the training data on each iteration – this can make it prone to overfitting.
Evolve
This layer is a genetic algorithm that optimises the model weights by starting with a population of model weights and mutating them continuously until they produce a more optimal set of weights. The initial chromosomes are mutations of the current model weights, and the default fitness function is the mean absolute error of predictions. The genetic algorithm runs for a specified number of generations, and the best chromosome updates the final model weights.
The Evolve layer brings stochasticity to the network and is used to escape local minima in the loss function. Its stochastic nature means that it will unlikely find weights that perfectly fit any minima, making it a stronger layer for avoiding overfitting earlier in the network, and a weaker layer later in the network.
While each layer is effective in isolation, they are more powerful when used together.
A Typical Network
A typical network will start and end with a Tighten layer, with one or more Evolve layers in between. The Tighten layers find the nearest minima to the current model weights, and the Evolve layers help to escape local minima and find a set of weights that exist closer to a better minimum.
Example
from xplainable.core.optimisation.genetic import XEvolutionaryNetwork
from xplainable.core.optimisation.layers import Tighten, Evolve
from xplainable.core.models import XRegressor
from sklearn.model_selection import train_test_split
Load the data
data = pd.read_csv("data.csv")
x, y = data.drop("target", axis=1), data["target"]
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2)
Create the initial model
model = XRegressor()
model.fit(x_train, y_train)
Create the network
network = XEvolutionaryNetwork(model)
Add the layers
Start with an initial Tighten layer
network.add_layer(
Tighten(
iterations=100,
learning_rate=0.1,
early_stopping=20
)
)
Add an Evolve layer with a high severity
network.add_layer(
Evolve(
mutations=100,
generations=50,
max_severity=0.5,
max_leaves=20,
early_stopping=20
)
)
Add another Evolve layer with a lower severity and reach
network.add_layer(
Evolve(
mutations=100,
generations=50,
max_severity=0.3,
max_leaves=15,
early_stopping=20
)
)
Add a final Tighten layer with a low learning rate
network.add_layer(
Tighten(
iterations=100,
learning_rate=0.025,
early_stopping=20
)
)
Fit the network (before or after adding layers)
network.fit(x_train, y_train)
Run the network
network.optimise()
Predict the Test data
y_pred = model.predict(x_test)
The above example has a lot to unpack, so let’s go through it step by step. First, we load the data and split it into training and test sets. Then we create the initial model and fit it to the training data. This process is vanilla data science and is the starting point for the network.
Next, we create the network:
network = XEvolutionaryNetwork(model)
This line creates the network and allows it to update the model weights in place. This characteristic is essential as each layer will permanently affect the model weights from the point that the layer finishes.
Next, we add the layers to the network. We generally start with a Tighten layer as this will find the nearest minima to the current model weights:
network.add_layer(
Tighten(
iterations=100,
learning_rate=0.1,
early_stopping=20
)
)
The Tighten layer has three parameters:
iterationslearning_rateearly_stopping
The iterations parameter is the number of iterations that the leaf boosting method will run for, and the learning_rate specifies how much a given weight will update on each iteration. The early_stopping parameter is the number of iterations that the layer will run without improving the loss function before it stops.
Next, we add two Evolve layers:
network.add_layer(
Evolve(
mutations=100,
generations=50,
max_severity=0.5,
max_leaves=20,
early_stopping=20
)
)
Other layer...
We generally add one or more Evolve layers after the initial Tighten layer as this will allow the network to escape local minima and find its way to a better minima.
The Evolve layer has five parameters:
mutationsgenerationsmax_severitymax_leavesearly_stopping
The mutations parameter is the number of mutations created for each generation, and the generations parameter is the number of generations that the genetic algorithm will run for.
The max_severity and max_leaves parameters dictate the significance of each mutation. The max_severity parameter is the maximum severity of the mutation relative to the current weights, and the max_leaves parameter is the maximum number of leaf nodes the mutation can affect. Generally, at the start of the network, we want to allow for significant mutations and then reduce the severity and reach of the mutations as the network progresses.
The early_stopping parameter is the number of generations that the layer will run without improving the loss function before it stops.
Finally, we add a final Tighten layer:
network.add_layer(
Tighten(
iterations=100,
learning_rate=0.025,
early_stopping=20
)
)
You will notice that this layer has a lower learning rate than the initial Tighten layer. A low learning rate makes smaller adjustments to the model weights as the network progresses to maximise our chances of finding a strong minima.
Now that we have added the layers, we can fit the network to the training data:
network.fit(x_train, y_train)
This line fits the network to the training data. At its core, it creates a matrix of the model weights based on the model leaf nodes. The layers then use this matrix to calculate better model weights. The fit method is independent of the layers and is run before or after adding the layers.
Finally, we can run the network:
network.optimise()
This line sequentially runs each layer in the network, updating the model weights in place at the end of each layer. Once the network has finished running, the model weights are updated to the final weights of the network.
It is possible to add more layers to an existing XEvolutionaryNetwork object and continue to run the network. More, shorter layers can be useful if you want closer control of performance monitoring.
Once the network finishes, we can simply use the model to make predictions and explanations as we would typically:
y_pred = model.predict(x_test)
Considerations
It’s important to note the limitations of XEvolutionaryNetwork. While networks consistently improve the accuracy of a model, they have drawbacks.
Overfitting
Long-running networks can be prone to overfitting if the network isn’t well-designed. Networks should be 3-6 layers with parameters suitable to the training data and model structure. We recommend monitoring the performance of the network with a validation set.
Time Complexity
The Evolve layer can be computationally expensive, especially if the mutations parameter is high. We recommend experimenting with lower numbers when training on a large dataset and increasing the number of mutations when you understand how these parameters affect the network.
Reproducibility
Due to the stochastic nature of the Evolve layer, it is not always possible to reproduce the same results for long-running networks. We recommend using a random seed to ensure that the network is reproducible.