Machine learning methods, which are used in the framework of predicting the solution of methods of adaptive management of the enterprise's goods sales, have been analyzed. Conduct an analysis of input data obtained during the operation of one enterprise. With the help of input data, train and conduct machine learning simulations with K-Nearest Neighbors, Support Vector Machines classifiers. The most significant factors influencing the client's decision to purchase goods have been identified. The study proposed a business process scenario for solving the problem of increasing the company's profit based on machine learning technology. The performance of the proposed methods was verified on a test sample of data.
In an era where organizations increasingly
rely on data to run their business, the ability of
different companies to manage and analyze
data will have a significant impact on their
performance. In this article, we develop a
model to measure the current level of maturity
of an enterprise's data and analytics systems to
optimize them to maximize the potential of
their data assets. Based on the three critical
components of people, technology and process,
we take the requirements of the enterprise as a
starting point. After a comprehensive review of
the current situation in the industry and the
needs of the enterprise, a methodology for
determining profitability was built [
Profitability [
Let’s consider the main indicators of the company’s profitability:
1. Profitability of invested funds [
2. Another indicator is the profitability of
production assets [
Return on total assets [
2. And another indicator is the profitability
of production [
One of the fundamental prerequisites in the
field of artificial intelligence is the assumption
of the possibility of creating machines that are
capable of performing tasks that usually require
the addition of human intelligence [
The artificial neural network is designed to
simulate learning processes in the human brain.
Artificial neural networks are designed in such
a way that they can recognize basic patterns
(regularities, stable relationships hidden in
data) and learn from them [
The practical implementation of material objects involves the solution of the corresponding problems of synthesis. At the same time, the
mathematical model of the object should not only adequately describe the physical processes that ensure obtaining the necessary initial characteristics of this object, but also allow the optimization process itself to be implemented. The software currently used to solve similar problems is characterized by either narrow specialization, or has a universal calculation device and requires an unacceptably long calculation time for complex objects. One of the
options for implementing fast and flexible methods for solving synthesis problems is the use of approximation properties of artificial neural networks. Most of the research related to artificial neural networks is focused on the solution of combinatorial optimization and forecasting problems, therefore, research on the application of neural networks is relevant.
The process of developing a certain model of
machine learning [
data.
A training sample [
A neural network [
Perceptron is mathematical model proposed by F. Rosenblatt, which is described by the transformation → using the formula a = ∑
=1 where = 1, … , ; is the weight of the perceptron; – value of input signals. After receiving the result, the activation function f is applied to the received value in v. The resulting value is compared with the trigger threshold θ and a decision is made.
Perceptron learning consists of finding weight coefficients. Let there be a set of pairs of vectors ( , ), = 1, … . , which is called a training sample. We will call a neural network trained on a given training dataset if, when applying each vector to the inputs of the network, we get the corresponding vector at the outputs each time. Recognition, processing of text, human language, music, images, 3D objects, tabular data, object detection in photos and videos, even creation of texts, images and videos—all this and much more is included in the practical application of neural networks.
The
learning
method
proposed
by
Rosenblatt [
Step 1. Initial values of all neuron weights ( = 0) are placed randomly.
Step 2. If the input image is loaded, the result will be the output image ̃ ≠ .
Step 3. The error vector = ( − ̃ ) is calculated, carried out by the network at the output. Then it happens that the change in the vector of weighting coefficients in the area of small errors must be proportional to the error at the output and equal to zero if the error is zero.
Step three. The vector of weights is modified by the formula:
( + ∆ ) = ( ) + ∙ ( ) , where 0 < < 1 is learning pace.
Step 4. Steps 1–3 are repeated for all vectors to be trained. One cycle of sequential representation of the entire sample is called an epoch. Training ends after several epochs: (a) when the iterations converge, that is, the vector stops changing, or when (b) or when the complete, rubberized absolute error over all vectors becomes smaller than some small value.
Deep learning is based on complex systems consisting of a huge number of neurons. One neural network can have billions of such structural units as neurons or perceptrons. Accordingly, there are many ways to structure them. And depending on how they will be connected, different architectures of neural networks can be distinguished.
One of the simplest gradient descent optimization algorithms looks like this: +1 = − ( )
We also justify that the Nesterov algorithm
will be the optimal algorithm for finding the
minimum [
− + 1.
A kind of refining operation is used here to find the gradient. Therefore, the method is sometimes called the accelerated Nesterov gradient. However, despite the fact that the proven upper estimate of the convergence time for this algorithm is minimal, it has not found its practical application. The thing is that the top score does not always coincide with the average. And in the case of this algorithm, the average convergence time estimate is close to the upper estimate. Therefore, the algorithm mostly works slowly. 3.3.
Human learning continues hierarchically. In the neural network of our brain, this process is carried out in several stages, each of which is characterized by its own degree of learning. At some stages, simple things are taught, at others—more complex.
In order to simulate the human learning process, layers of neurons are used in the construction of artificial neural networks. The idea of these neurons is suggested by biological processes. Each layer of an artificial neural network is a set of independent neurons, each neuron of a certain layer is connected to the neurons of an adjacent layer. 3.4.
If we are dealing with n-dimensional input data, then the input layer will consist of n neurons. If m different classes are distinguished among our training (training) data, then the output layer will consist of m neurons. The layers nested between the input and output layers are called hidden. A simple neural network consists of a pair of layers, and a deep neural network consists of many layers.
Consider the case when we want to use a neural network for data classification. The first step is to collect relevant training data and label it. Each neuron acts as a simple function, and the neural network is trained until the error is less than a certain set value. The difference between predicted and actual outputs is mostly used as an error. Based on how big the error is, the neural network corrects itself and retrains until it gets close to the solution.
A perceptron is a single neuron that receives input data, performs calculations on it and outputs an output signal. The perceptron uses a simple function to make decisions. Suppose we are dealing with an N-dimensional input data point. The perceptron calculates a weighted sum of N numbers, and then adds a constant to them to get the original result. The ego constant is called the bias of the neuron. It is interesting to note that such simple perceptrons are used to design very complex deep neural networks.
The capabilities of the perceptron are limited. You need to get a set of neurons to act as one and evaluate it to achieve a goal. Let's create a single-layer neural network consisting of independent neurons that influence the input data to obtain the output result.
To obtain higher accuracy, we must give more freedom to the neural network. This means that the neural network must have more than one layer to capture the underlying patterns that exist among the test data. Let’s create a multilayer neural network that will ensure this.
A neural network can be used as a classifier.
You can use a neural network like regressor
[
Let’s define a multilayer neural network with two hidden layers. You can design a neural network in any other way. In this case, we will have 10 neurons in the first layer and six neurons in the second layer. Our task is to predict a single value, so the output layer will contain only one neuron.
We will use the creation of the model for the forecasting task.
To do this, we will help predict the probability that the enterprise will be profitable and will have profits in the future.
The dataset we used in this chapter is taken from this database: ML_Manufacture_Prifit_ Companies.csv. This data set contains several independent predictors and one goal, to make the enterprise profitable. Its features are as follows: rating, change in rating, income, management assets, market value, percentage change in income, percentage change in profit, employees.
The data set contains 1001 records and all businesses are successful and profitable.
For this example, the dataset was
downloaded locally and named
ML_Manufacture_Prifit_Companies.csv
(Fig. 1). We observe the columns with the
following data (Fig. 2). The first task for
estimating a company’s profit is to clean the
data so that there are no missing or erroneous
values (Fig. 3 and 4) [
The next step is to study how different independent characteristics affect the result (which characteristics affect the profitability of the enterprise).
The corr() function calculates the pairwise correlation of columns. For example, the following result shows that the level of management assets has little relationship with the market value of the enterprise 0.127, but it has a significant relationship with the profitability of the firm 0.49.
We find out which functions significantly affect the result.
The following code snippet uses the matshow() function to plot the results returned by the corr() function as a matrix. At the same time, various correlation coefficients are also displayed in the matrix:
Fig. 5 shows the matrix. Cubes with colors closest to black represent the highest correlation coefficients, and those closest to blue represent the lowest correlation coefficients.
Another way to construct a correlation matrix is to use Seaborn’s heatmap() function as follows: Fig. 5 shows a heat map created using the Seaborn library (Fig. 6).
Let’s single out four signs at enterprises that have the highest correlation with the result (Fig. 7–9).
We will evaluate several algorithms to find the most optimal one and the one that will provide the best performance. Therefore, we will use the following methods: Logistic regression. K-Nearest Neighbors (KNN). Support vector machines (SVM) using linear kernels and RBF kernels.
For the first method, we will use logistic regression. Instead of splitting the dataset into training and testing sets, we will use 10-fold cross-validation to get the average score of the algorithm used. 8.2.
The next method we will use is K-Nearest Neighbors (KNN). In addition to using 10-fold cross-validation to average the algorithm, different values of k should be tried to find the optimal one so that the best accuracy can be obtained. 8.3.
Another method we will use is Support Vector Machine (SVM). We will use two types of kernels for SVM: linear and RBF.
Since the most efficient algorithm for our data set is KNN with k = 3, we can now continue to train the model using the entire data set: After training the model, we need to save the files so that we can get the model for prediction later. The trained model is now saved to a file.
After loading the model, let’s make some predictions:
The output prints the word “Profitable Enterprise” if the return value of the prediction is 1; otherwise, the word “Unprofitable enterprise” is printed. It is also necessary to obtain the probability of the forecast in order to obtain the probability in percent (Fig. 10–12).
The printed probabilities show that the probability of the outcome is 0 and the probability of the outcome is 1. The prediction is based on the prediction with the highest probability, and we take that probability and convert it to a confidence percentage. 10.Conclusions
In this study, a dataset of top 1,000 profitable companies was selected and machine learning was implemented without a teacher. Companies were divided into three clusters based on revenue, earnings, current assets, market value and number of employees. Clustering was implemented based on the KMeans algorithm. The Inertia metric was also used to determine the optimal number of clusters. Inertia shows us the sum of the distances to each center of mass. If the total distance is large, it means that the points are far apart and may be less similar to each other. In this case, one can continue to evaluate larger values of K to see if the overall distance can be reduced. However, it is not always the smartest idea to reduce the distance. Using the elbow (bend) method, we can choose the value of the number of clusters 2 or 3 (the first or second obvious bend). 11.References