Ансамбль нейромереж GRNN на підставі зміщених поверхонь відгуку для задач електронної комерції
| Autor: | Ivan Izonin, Pavlo Vitynskyi, Roman Tkachenko |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Mean squared error
Artificial neural network business.industry Computer science 05 social sciences Sample (statistics) ансамбль алгоритми без навчання зміщення поверхонь відгуку нейронна мережа узагальненої регресії прогнозування електронна комерція Machine learning computer.software_genre Regression Support vector machine ComputingMethodologies_PATTERNRECOGNITION 0502 economics and business Linear regression lcsh:SD1-669.5 General Earth and Planetary Sciences 050211 marketing Artificial intelligence lcsh:Forestry business computer 050203 business & management Transformation geometry General Environmental Science Test data |
| Zdroj: | Науковий вісник НЛТУ України, Vol 29, Iss 9, Pp 142-146 (2019) |
| ISSN: | 2519-2477 1994-7836 |
| DOI: | 10.36930/40290925 |
| Popis: | Solving e-commerce problems being represented in most cases by non-linear response surfaces is an important task. The use of existing computing intelligence methods is not always appropriate due to the significant complexity of training and debugging procedures. Non-iterative tools and neural networks without training also do not provide satisfactory accuracy of the result. The accuracy can be improved using different ensembling techniques. Therefore, the paper describes a new ensemble method based on generalized regression neural networks. The basic idea of the developed ensemble is to linearize the response surface given by the data of the available sample. Therefore, the surface obtained by means of the general regression neural network is given to the input of a linear neural structure. This combination helps improve the accuracy of solving the task by the ensemble. The described ensemble is used to solve the problem of predicting the price of used cars. Application of the ensemble developed enables predicting the price of the used cars based on the most suitable independent attributes. In common practice solving this task requires expert knowledge. The urgency of solving this problem is substantiated. The dataset contains vehicle characteristics and sale prices of 1436 used cars. The main attributes of the considered dataset are provided. The optimal parameters were experimentally selected. Performances of different existing methods were compered. The methods were evaluated by the root mean square error using a test data sample. By comparison with known methods such as General Regression Neural Network, Radial Basis Function Neural Network, Linear Regression, Lasso Regression, and Support Vector Machines Regressor, the highest accuracy of its work is established. The results are compared with Condorcet's jury theorem estimations. The implementation of the proposed method was done on Python programming language. Thus, we can summarize that the developed general regression neural network ensemble based on offsets of response surfaces and with additional use of the linear-type Neural-like Structure of a Successive Geometric Transformations Model can be used to solve various high-precision e‑commerce tasks. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|