I am plotting the precision-recall curves for my models which I have built using an imbalanced dataset.

I initially plotted the precision-recall curve for my models using the plot_precision_recall_curve function directly, like so:

# split into train/test sets
trainX, testX, trainy, testy = train_test_split(X, y, test_size=0.5, random_state=2, stratify=y)

dt = DecisionTreeClassifier()
from sklearn.metrics import plot_precision_recall_curve
plot_precision_recall_curve(dt, testX, testy, ax = plt.gca(), name = "Decision Tree")

Which resulted in this plot:

enter image description here

However, I then wanted to apply threshold tuning to achieve the optimal F0.5 score for my models. To do this, I plotted the precision-recall curve like so:

# predict probabilities
y_pred = dt.predict_proba(X_test)

# keep probabilities for the positive outcome only
y_pred = y_pred[:, 1]

precision, recall, thresholds = precision_recall_curve(testy, y_pred)

# convert to F0.5 score
beta = 0.5 
f05score = ( (1 + pow(0.5, 2)) * precision * recall ) / (pow(0.5, 2)* precision + recall )

# locate the index of the largest f 0.5 score
ix = argmax(f05score)

no_skill = len(testy[testy['0']==1]) / len(testy)

pyplot.plot([0,1], [no_skill,no_skill], linestyle='--', label='No Skill')
pyplot.plot(recall, precision, marker='.', label='DT', zorder=1)

# set zorder so dots appear over line
pyplot.scatter(recall[ix], precision[ix], marker='o', color='black', label='Best F0.5 Score', zorder=2)

# axis labels

# show the plot

Which resulted in this plot:

enter image description here

Therefore, as you can see from both precision-recall curves, they look different and I want to know why this is as I assume they should look the same? Are there any mistakes in my code?

  • $\begingroup$ It's the same plot. In second plot you interpolate in the PR domain (and you shouldn't). $\endgroup$
    – usεr11852
    Commented Oct 18, 2022 at 2:44
  • $\begingroup$ I see, so how do I stop the interpolation from happening in the second plot? @usεr11852 $\endgroup$
    – sums22
    Commented Oct 18, 2022 at 8:37
  • $\begingroup$ You just change what we plot, we don't have enough points there. $\endgroup$
    – usεr11852
    Commented Oct 18, 2022 at 10:58
  • $\begingroup$ @usεr11852 could you include an answer that clearly shows how to do this? Currently I am plotting precision-recall pairs for different thresholds which I calculated through: precision, recall, thresholds = precision_recall_curve(testy, y_pred). How do I modify this code to return more precision-recall pairs? $\endgroup$
    – sums22
    Commented Oct 20, 2022 at 8:52
  • $\begingroup$ I think you don't plot everything, for example, there should be a point at approximately 0 recall, 0.22 precision. Where is that? $\endgroup$
    – usεr11852
    Commented Oct 20, 2022 at 11:10


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.