The Lift Chart (lift chart) and the gain graph (gain chart) are a very useful graphical representation in evaluating the predictive capability of a model. In SPSS, a typical gain graph is as follows:
in today's blog post, bloggers will discuss with you the logic of making the gain graph and how to interpret the gain and lift graphs. In the following blog post, we will use an example of a direct mail company to explain to you. Assuming that, based on previous experience, the company knows that the average response rate for their direct mail campaigns is 10%. Next we continue to assume that: * Cost per Mail AD = $
* Rebate for each response = $
* Number of email advertisements =
Based on the above hypothesis, if a company issues 10,000 e-mail ads, the following table summarizes the results of the campaign:
Now let's assume that this company uses SPSS Modeler to build predictive models using historical marketing data. "There is no response" is the target value, while other demographic variables, socioeconomic variables, and behavioral variables are predictor variables (predictors). And the company can forecast the results to sort out the sales expectations in descending order. So instead of randomly distributing 10,000 ad messages to the audience, the company can send messages to the 10,000 customers who are most likely to respond, and then send a message to 10,000 customers who might respond next, and so on. The end result is as follows:
As shown above, the results in the second table are significantly better than the first chart. If the company is interested in maximizing revenue, using the above forecast method, the company can get $400,000 profit by spending $50,000 cost, and through the first method, if the company to get the same revenue, it needs to spend $80,000 around the cost!
By comparing the response rates of the two marketing methods, we got the following results:
The above data is made into a line chart, we can get the following results:
The gap between the Green Line (using the response rate after predicting the model) and the Red Line (using the response rate before the forecast model) shows how the company uses the predictive model to maximize profits and randomly send messages in a way that yields a gap between the profit and the random sending of mail. So how are the gain and lift values calculated? We can find the answer by defining the two terms:
* gain = (expected response of the application Prediction model)/(expected response sent randomly) * elevation = (expected response of the first 10,000 users applying the forecast model)/(expected response of the first 10,000 users randomly sent)
By using the data in the example, the gain and lift values are as follows:
To summarize, the gain and lift graphs can help solve two of the following problems:
* How to assess the quality of a predictive model?
* How to compare the response rate after using the predictive model and the response rate of the randomly sent? Gain >1 means that the results of the predictive model are better than the results sent randomly.
From Wiznote
How to evaluate the model using gain and lift graphs
