Warning
The JupyterLab development team is excited to have a robust
third-party extension community. However, we do not review
third-party extensions, and some extensions may introduce security
risks or contain malicious code that runs on your machine. Moreover in order
to work, this panel needs to fetch data from web services. Do you agree to
activate this feature?
Please read the privacy policy.
Installed
Discover
Open Tabs
Kernels
Python 3 (ipykernel)- GiftWrapping_ConvexHull_Solution.ipynb (9fd49195)
- Roof_example.ipynb (47d0d05c)
- GiftWrapping_ConvexHull.ipynb (79bdac9b)
- 04E - Raster Data Exploration.ipynb (30fd7f20)
- 04E - Raster Data Exploration (Solution).ipynb (78656c61)
- 05 - Logistic Regression Decision Boundary.ipynb (b3c2fe0d)
- 05 - Land Cover Prediction from Hyperspectral Satellite Image.ipynb (3f349317)
- E06 - SVMs on Hurricanes.ipynb (423bce23)
- SVM_parameter_exploration.ipynb (64095fd1)
- Classification-Metrics.ipynb (c00c001b)
Language servers
Recently Closed
Workspaces
- auto-B
- auto-L
- default
Terminals
Classification-Metrics.ipynb
- Classification Metrics: Accuracy, Precision, and Recall
- Accuracy
- Precision
- Recall
- F1-Score
- Simple Raster Classification Example
- Classification-Metrics.ipynblast mo.5.5 KB
- E06 - SVMs on Hurricanes.ipynb43m ago150.3 KB
- hurricanes.csvlast mo.21.4 KB
- SVM_parameter_exploration.ipynblast mo.6.7 KB
Classification Metrics: Accuracy, Precision, and Recall¶
分类指标：准确性、精确度和召回率 ¶
- TP: True Positives TP：真阳性
- FP: False Positives FP：误报
- TN: True Negatives TN：真阴性
- FN: False Negatives FN：假阴性
Accuracy¶ 精度 ¶
Definition: 定义：
The ratio of correctly predicted samples to the total number of samples.
正确预测的样本与样本总数的比率。
Interpretation: 解释：
It measures overall correctness of the model, but can be misleading if classes are imbalanced.
它衡量模型的整体正确性，但如果类不平衡，可能会产生误导。
Precision¶ 精度 ¶
Definition: 定义：
The ratio of true positives to all predicted positives.
真阳性与所有预测阳性的比率。
Interpretation: 解释：
It answers: "When the model predicts positive, how often is it correct?"
它回答：“当模型预测为正时，它多久是正确的？
Useful when false positives are costly (e.g., spam detection).
当误报成本高昂时很有用（例如，垃圾邮件检测）。
Recall¶ 召回 ¶
Definition: 定义：
The ratio of true positives to all actual positives.
真阳性与所有实际阳性的比率。
Interpretation: 解释：
It answers: "Of all actual positives, how many did the model identify?"
它回答说：“在所有实际的积极因素中，模型识别了多少？
Important when missing positives is costly (e.g., disease detection).
当漏失阳性代价高昂时（例如，疾病检测），这一点很重要。
F1-Score¶ F1 分数 ¶
Definition: 定义：
The harmonic mean of precision and recall.
精度和召回率的调和平均值。
Or, in terms of TP, FP, and FN:
或者，就 TP、FP 和 FN 而言：
Interpretation: 解释：
F1-Score balances precision and recall in a single metric.
F1-Score 在单个指标中平衡了精度和召回率。
It is low when either precision or recall is low, and is most useful when:
当精度或召回率较低时，它为低，在以下情况下最有用：
- You need to balance false positives and false negatives.
您需要平衡误报和漏报。 - There is a class imbalance and you want a single metric to compare models.
存在类不平衡，您需要一个指标来比较模型。
The harmonic mean gives a more conservative estimate than the arithmetic mean, favoring models that perform well on both precision and recall.
调和均值给出了比算术均值更保守的估计值，有利于在精度和召回率方面都表现良好的模型。
Simple Raster Classification Example¶
简单栅格分类示例 ¶
We simulate a simple land-sea classification on a 10x10 raster grid:
我们在 10x10 栅格网格上模拟简单的陆海分类：
- Each cell in the raster represents either landmine (
1) or safe (0).
栅格中的每个像元表示地雷 （1） 或安全 （0）。 - The ground truth raster has:
地面实况栅格具有：- One cell labeled as landmine (Positive).
一个细胞被标记为地雷（阳性）。 - All other cells labeled as safe (Negative).
所有其他细胞都标记为安全（阴性）。
- One cell labeled as landmine (Positive).
- The predicted raster contains:
- All cells labeled as safe (Negative), so the model fails to detect the single land cell.
What is the accuracy of our prediction?
Note that we don't actually train a model, we just look at a potential prediction and compute the metrics for it.
Use the model to make predictions from the test data.
Use the predicted classes and the true classes of the test data to calculate the accuracy score.
0.8970588235294118
Tasks¶
- Train the SVM with different parameters (kernels, C value etc) and explore how the training and test score change.
- How does the scores relate to overfitting?
- You can also play around with other settings, like for instance the stratification of the train-test-split (see the documentation), which makes sure that the hurricane types are distributed equally between test and train set.
Further questions:
Which coordinate (latitude, longitude) is more significant when predicting the formation of tropical hurricanes?
Try and see if the maximum intensity makes a difference in the model!
Would it improve the model to add all other features as well? Which features would make sense? Can we create other interesting features from the data we have?
The continuous grey line represents the decision boundary, the dashed lines denote the margins. Support vectors are enclosed in circles.
- E06 - SVMs on Hurricanes.ipynb
E06 - Hurricanes.ipynb 上的 SVM - 04E - Raster Data Exploration.ipynb
04E - 栅格数据探索.ipynb - SVM_parameter_exploration.ipynb
- Classification-Metrics.ipynb
分类指标.ipynb