How Does a Result of the Model Developed through Machine Learning Turn Out?

22 Mar 2022

Summoning up previous results, DAVinCI LABS would have made its identity very clear- it automates the development of machine learning prediction models. Other than that, we also contain modules that optimize decision-making processes according to business strategies, and we will further introduce this module later on.

This post will cover the results of model development through DAVinCI LABS. Just like we see from sci-fi movies, once DAVinCI LABS receives data, can we gain intuitive answers to our questions straight away- who the criminal is, which companies would end up in ashes?

Target variable we would like to find out has generated a prediction value.

To start with the answer, what DAVinCI delivers to the user is a predicted value (score) of new data. Once DAVinCI LABS undergoes the process of developing a machine learning predictive model, the model is loaded into the service, where then the user uploads new data that has no records yet. As a result, the user runs the model and ends up with prediction results for the samples.

According to the subject column used for training, the prediction result comes out in a form out of these three types: a possibility in numerical figure between 0 and 1, a class out of many clusters, or a continuous value.

In DAVinCI LABS, the system automatically reads the target type and then automates the development process of the appropriate model. However, the data preparation process requires knowledge of different data forms to achieve the initial analysis goal. That is precisely why we are going to learn today “model types decided by prediction target” and “the final output of each type”.

1. Different types of prediction subject result in different output

Supervised learning in machine learning is classified into two types of models: “regression model that makes predictions in continuous values” and “classification model that predicts the category the input belongs to”.

And depending on the form of the target value, these models are divided into subclassifications. Each classification is listed below.

Different results according to the form of the prediction target

1.    Does the target value (variable) consist of only 0 and 1?

→  Logistic regression model (Binary-class classification linear model)

2.   Does the prediction target (variable) consist of continuous real numbers?

→ Regression model

3.   Is the prediction target (variable) classified into two categories?

→  Binary-class classification model

4.   Does the prediction target (variable) consist of three or more categories?

→  Multiple-class classification model

It depends on the model type when it comes to selecting the performance index, applicable algorithms, and additional functions as well as the final output, which is why we must be aware of each model type.

Index and algorithms will be covered in the future- let’s dig into what exact model types there are.


2. Logistic regression model (Binary-class classification linear model)

First, let’s examine the logistic regression model, in which the prediction target consists of numbers 0 and 1. Binary classification is a matter of sorting between ‘Yes’ and ‘No’, ‘A’ and ‘B’, or 0 and 1, so the output also appears in either of these forms.

But then, what would it mean to solve this problem with regression? It means finding out the possibility of the target value turning out to be ‘1’. For example, if the challenge is to predict whether a client would buy a product or not, generating the possibility of purchase is what the logistic regression model (binary-class classification linear model) covers.


Problem to solve in this logistic regression model: What are the purchase possibilities of each client cluster?

Along with the whole process naturally arises a question:

“Why necessarily convert the form of the output to real numbers to get the possibility value?”

The answer is ‘concreteness’. Compared to models that are limited to acquiring either ‘retention’ or ‘churn’ as prediction results, regression analysis performs a more concrete action. Classification models only offer you a uni-dimensional distinction between the customer being likely to churn or stay loyal (retention). But then, would all samples classified as ‘churn’ imply the same probability?

Certain samples could hold a 0.99 churn rate while another could contain a rather ambiguous output of 0.51. Such variance in percentage is addressed through regression conversion. In other words, if a more careful observation is needed in predicting an indicator with significant impact on the business, it is better to find an accurate threshold by receiving a prediction value in the form of a probability, as shown below.

Let’s expand from merely addressing customer churn in two cases of 0 and observe the topic with probability values.

3. Regression model

A regression model is applied when real numbers come into the prediction range. For example, let’s assume developing a prediction model that measures customer values based on their amount of payment. What we could fix as a learning material (prediction target) at this point of model development would be “the total amount of expenditure inside the service range by the client”.

So, once new customer data is added to the developed model, “payment prediction” can be retrieved now. This means we could then speculate in advance the amount the customer is likely to spend. Like so, the amount of expenditure in form of real numbers is demonstrated as the target value, and the model is developed under this initiative. Once all the conditions are set and the model runs a prediction, the user will get the prediction value in the continuous real number form- similar to that of the training process.

The generated prediction value becomes a good reference along with sales goals, sales policy to draw an estimated margin (threshold) of salesforce clients.


If the amount of purchase is included in past data, let’s figure out how much the new client is likely to spend.

4. Binary-class classification model

This model shares the same fundamental mechanism with the previously mentioned logistic regression. When we face a traditional binary-class classification problem, aiming to get a categorical target as the result, the system (not the user) sets the threshold based on the target average and then provides the classified class according to the created margin. This is the core difference the two classification models. (For example, if the value is bigger than 0.3, the decision leans towards selecting B).

As the model’s final output, the user is offered two classes identical to the sets that have been used for training. To be more precise, you are provided a class per sample as the prediction value- a choice between the two.

Therefore, in order to explicitly decipher the prediction values and actively utilize them, we recommend developing a classification model rather than labeling a target in 0, 1 form.

Let’s try to predict whether the new customer would churn, referring to past records of customers’ churn/ retention decisions.

5. Multiple-class classification model

If a subject column to be used for prediction consists of three or more classes (categories), we choose a multiple-class classification model. As a common example, there is a product recommendation model.

While running the service, the user can retrieve records on the type of product each client has chosen. Setting this as a subject, we develop a prediction model to get the predictable equation: once a certain group of customers similar to the studied group is added, we can identify in advance the type of products they would be likely to choose.

As such, if the ultimate goal is to sort different groups into separate segments, in other words, classifying specific groups or categories, we can always utilize a multiple-class classification model.

Based on past records of customer purchases, let’s make a new prediction on what a new client would buy.

This session, we examined how machine learning prediction models generate results. #Regression analysis might already be familiar to you. However, machine learning is not just run with regression analysis, it can approach classification problems too. We observed that in case of binary-class classification problems we could approach them with regression models as well. DAVInCI LABS of Ailys is applicable in all of these forms.

Next time we’ll dive into how these model outputs could be used in real business fields.