# Low precision on classification model

I am working since some months on a prediction from lead to a sale. Someone makes a lead on my website and I want to predict if this user will make a sale. I have these metrics on the test data.

Now when I check for example for the month May_2020 all the leads I had predicted as 1 (probability >=0.5) and compare with the real sales I receive a conversion rate of 22%. That means, only 22% of all the predicted with 1 lead becomes sale. What do you think?

• Is this caused because of low precision? The precision is similar to the conversion rate or is this only luck or the precision implicates the conversion rate.
• Any idea of how to reduce this big error of 78%?

Precision gives the percentage of correctly predicted positives out of total positives predicted by the model. Therefore, low conversion rate of your model is due to low precision. Looking at the support, this looks like a class imbalance problem. 92% of your training examples belong to class '0'.Therefore, while training, your model has more examples of class '0' to learn from.

There are a few ways to deal with this problem:

1. Over-sampling - This when you randomly draw samples from the minority class with replacement and use it in the training set
2. Under-sampling - This is when you draw random samples from the majority class with or without replacement and decrease the number of examples of the majority class given to the model while learning.
3. Synthetic sampling - There are also techniques like SMOTE that generate samples of minority class synthetically.

These are only few techniques I have highlighted. You can check this link for further reading https://machinelearningmastery.com/tactics-to-combat-imbalanced-classes-in-your-machine-learning-dataset/

• Thanks, Ankita, I will check everything you suggest! – Mutatos Oct 25 '20 at 8:01

1.The conversion rate is actually the precision of the model.
Your low conversion rate is because of the low precision as mostly your dataset consist of 0.
Thats why precision of 0 is quite high as compared to 1

2.As for your second question , In order to increase precision can you upload the dataset from which you are predicting the results

• Thanks, Shiv! I can't upload it, because it is a private data set. – Mutatos Oct 25 '20 at 8:02