Recurring concepts

In [1]:

from river.datasets.synth import FriedmanDrift 
from river.preprocessing import MinMaxScaler 
from river.metrics import MAE
from river.utils import Rolling
from river.tree import HoeffdingTreeRegressor
from deep_river.regression import Regressor
from torch import nn 
from tqdm import tqdm 
import matplotlib.pyplot as plt
import torch

from river.datasets.synth import FriedmanDrift from river.preprocessing import MinMaxScaler from river.metrics import MAE from river.utils import Rolling from river.tree import HoeffdingTreeRegressor from deep_river.regression import Regressor from torch import nn from tqdm import tqdm import matplotlib.pyplot as plt import torch

In [2]:

n_samples = 12500
change_points = (5000, 7500)
friedman = FriedmanDrift(drift_type="gra", position=change_points)

def test_train_eval(model, stream, update_interval=100):
    results = []
    steps = []
    step = 0
    metric = Rolling(MAE(), window_size=400)
    scaler = MinMaxScaler()
    for x, y in tqdm(list(stream)):
        x = scaler.learn_one(x).transform_one(x)
        y_pred = model.predict_one(x)
        model.learn_one(x, y)
        metric.update(y, y_pred)
        step += 1
        if step % update_interval == 0:
            results.append(metric.get())
            steps.append(step)
    return steps, results

n_samples = 12500 change_points = (5000, 7500) friedman = FriedmanDrift(drift_type="gra", position=change_points) def test_train_eval(model, stream, update_interval=100): results = [] steps = [] step = 0 metric = Rolling(MAE(), window_size=400) scaler = MinMaxScaler() for x, y in tqdm(list(stream)): x = scaler.learn_one(x).transform_one(x) y_pred = model.predict_one(x) model.learn_one(x, y) metric.update(y, y_pred) step += 1 if step % update_interval == 0: results.append(metric.get()) steps.append(step) return steps, results

In [3]:

class SimpleMLP(nn.Module):
    def __init__(self, n_features):
        super().__init__()
        self.hidden = nn.Linear(n_features, 20)
        self.logit = nn.Linear(20, 1)

    def forward(self, x):
        h = self.hidden(x)
        h = torch.relu(h)
        return self.logit(h)


mlp = Regressor(
    SimpleMLP,
    loss_fn="l1",
    optimizer_fn="adam",
    lr=0.005,
    seed=42,
)
steps, results_mlp = test_train_eval(mlp, friedman.take(n_samples))

class SimpleMLP(nn.Module): def __init__(self, n_features): super().__init__() self.hidden = nn.Linear(n_features, 20) self.logit = nn.Linear(20, 1) def forward(self, x): h = self.hidden(x) h = torch.relu(h) return self.logit(h) mlp = Regressor( SimpleMLP, loss_fn="l1", optimizer_fn="adam", lr=0.005, seed=42, ) steps, results_mlp = test_train_eval(mlp, friedman.take(n_samples))

100%|██████████| 12500/12500 [00:09<00:00, 1363.65it/s]

In [4]:

tree = HoeffdingTreeRegressor()
steps, results_tree = test_train_eval(tree, friedman.take(n_samples))

tree = HoeffdingTreeRegressor() steps, results_tree = test_train_eval(tree, friedman.take(n_samples))

100%|██████████| 12500/12500 [00:03<00:00, 3890.17it/s]

In [5]:

fig, ax = plt.subplots(figsize=(8, 4))
ax.plot(steps, results_mlp, label="MLP")
ax.plot(steps, results_tree, label="Hoeffding Tree")
for change_point in change_points:
    ax.axvline(change_point, color="red", alpha=0.5)
ax.set_xlim(0, n_samples)
ax.set_ylim(1, 5)
plt.text(int(change_points[0] / 2), 4, "Concept 1", horizontalalignment="center")
plt.text(
    int(change_points[0] + (change_points[1] - change_points[0]) / 2),
    4,
    "Concept 2",
    horizontalalignment="center",
)
plt.text(
    int(change_points[1] + (n_samples - change_points[1]) / 2),
    4,
    "Concept 1",
    horizontalalignment="center",
)

ax.set_xlabel("Steps")
ax.set_ylabel("Moving MAE")
ax.legend()

fig, ax = plt.subplots(figsize=(8, 4)) ax.plot(steps, results_mlp, label="MLP") ax.plot(steps, results_tree, label="Hoeffding Tree") for change_point in change_points: ax.axvline(change_point, color="red", alpha=0.5) ax.set_xlim(0, n_samples) ax.set_ylim(1, 5) plt.text(int(change_points[0] / 2), 4, "Concept 1", horizontalalignment="center") plt.text( int(change_points[0] + (change_points[1] - change_points[0]) / 2), 4, "Concept 2", horizontalalignment="center", ) plt.text( int(change_points[1] + (n_samples - change_points[1]) / 2), 4, "Concept 1", horizontalalignment="center", ) ax.set_xlabel("Steps") ax.set_ylabel("Moving MAE") ax.legend()

Out[5]:

<matplotlib.legend.Legend at 0x13da5dfd0>