Implementation of Data Mining for Retail Chain Sales

Prediction Using Artificial Neural Network

R Y Mahendra

Jurusan Teknik Informatika, Fakultas Teknik dan Ilmu Komputer, Universitas

Komputer Indonesia, Indonesia

*[email protected]

Abstract. The purpose of this research is to help retail chain business predict sales of
each item in each outlet based on the location, type, and size of outlets. The method
used in this research is the experimental method by applying data mining using the
Artificial Neural Network (ANN) algorithm to predict sales of each item in each outlet
in the retail chain business. The data used in this research was sales data from 10 outlets
in different cities in 2013. The results of the prediction using this algorithm have a Root
Mean Squared Error (RMSE) value of 1127.239. This shows great prediction results.
The retail chain business can use the result of this research for managing the right
business strategy in the future.

1. Introduction
Predicting and the ability to plan a business strategy to achieve success in the future is crucial for
companies, especially retail chain business [1]. Nowadays, company management is easier to make
decisions to determine the business strategy that will be used because of the increased accuracy of sales
predictions [1]. Accurate sales prediction offer numerous benefits for the company such as speeding up
the decision-making process, reducing risk, managing the budget effectively, increasing profits, etc [2].
Nowadays, the reliability of predicting is greatly increased due to the development of mathematical
algorithms combined with computer utilization. One of the improved technologies is Data Mining,
which is an advanced technology for extracting implicit information from large data sets using some
methods such as statistics, artificial intelligence, or machine learning so as to produce more reliable
predicting than before[3]. In this Retail Chain sales prediction, the model used is ANN. ANN is a model
built to imitate the workings of neurons in the human brain and is a computational model that has shown
great behavior in problem-solving in numerous fields such as artificial intelligence, engineering, etc[4].
Research related to sales predictions, among others, carried out by [1]. In these studies, [1] predicting
the monthly sales volume of the textile warehouse obtained high accuracy results with the value of Root
Mean Square Error (RMSE) of 3.34e-11. Another research conducted by [3], predicting the German
Automobile Market using Multiple Linear Regression (MLR) and Support Vector Machine (SVM)
models. The data used in this study are the number of annual, monthly, and quarterly registrations of
new cars from 1992 to 2007. The result is that the Non-Linear Trend Estimation (SVM) model is
superior to its predictions (less test error) compared to the Linear Trend Estimation (MLR).
Meanwhile, [5] also conducted research on predictions of possible turnover for potential outlet sites
from a large European food retail company using the Support Vector Regression (SVR) algorithm. The
data used in this study were 245 attributes from 870 outlets. The results obtained show that SVR is
better at predicting outlet locations compared to Huff-prediction with a Root Mean Square Error value
of 98072,306. Another research conducted by [6], predicting sales using Artificial Neural Network
(ANN), Gradient Boosted Tree, Support Vector Machine (SVM), k-Nearest Neighbor (k-NN), Decision
Tree, and Random Forest algorithms. The data used in this study is the sales prediction data of Walmart
companies available on the Kaggle platform. The results obtained prove that ANN is the best algorithm
for predicting sales compared to the other 5 algorithms. But, the performance of the Gradient Boosted
Tree algorithm is better than the SVM, k-NN, Decision Tree, and Random Forest algorithms.
In the research conducted by [7], the results obtained from forecasting the amount of fish production
in September 2016 using the Backpropagation Neural Network (BPNN) is 86573 kg with an average
value of MAPE errors of 22.49%. Another research conducted by [8], predicting car sales in Kia and
Hyundai in the USA using the ANN algorithm. The data used in this study are sales data obtained from
Kia and Hyundai companies in the US and Canada from 2010 to 2015. In this study also compared three
algorithms such as ANN, Linear Regression, and Exponential Regression. The results obtained indicate
that ANN is one of the accurate methods for predicting car sales, because it has a lower Minimum
Square Error (MSE) value compared to other methods. In the research conducted by [9], predicting car
sales using Artificial Neural Network (ANN) and Certainty Factor (CF). The data used in this study is
dealer sales data for the area of Depok and surrounding areas for the period 2005 to 2010. Forecasting
using ANN results in 2015 will sell 29579 Honda cars with an error target value of 4.205%.
Most previous research shows that sales prediction has an important role in planning business
operations and business strategies in the future [1][2][3]. Therefore, the purpose of this research is to
predict the sales of each item in each outlet and retail business chain can use the results of this study to
set the right business strategy going forward. In addition, this research provides visualization of
prediction results in graphical form so that business chain retailers are easy to use the prediction models
produced. The method used in this study is an experimental method by applying data mining using the
Artificial Neural Network (ANN) algorithm to predict the sales of each item in each outlet using retail
chain sales data from 10 Big Mart outlets in different cities in 2013.

2. Method
2.1. Experimental Dataset
In order to predict sales on retail chain business, this study uses the "Big Mart Sales Prediction" public
dataset available on the Kaggle platform [6]. This dataset describes sales made by 10 Big Mart outlets
in different cities in 2013. This dataset contains 8523 records and 12 variables as shown in table 1.

Table 1. Variables of the Big Mart Sales Prediction dataset

Variable Name Type Description Segment
Item_Identifier Numeric Unique product ID Product
Item_Weight Numeric Product weight Product
Item_Fat_Content Categorical Low fat or not Product
Item_Visibility Numeric Percentage of the total display area of all Product
products in the store allocated for a
particular product.
Item_Type Categorical Item category Product
Item_MRP Numeric Maximum Retail Price (price list) of the Product
product
Outlet_Identifier Numeric Unique outlet ID Outlet
Outlet_Establishment_Year Numeric The year the outlet was established Outlet
Outlet_Size Categorical Outlet size Outlet
Outlet_Location_Type Categorical Types of cities where outlets are located Outlet
(Tier 1, Tier 2, Tier 3)
Outlet_Type Categorical Grocery store or supermarket Outlet
Item_Outlet_Sales Numeric Sales of products at certain outlets and Product
outcome variables to be predicted

2.2 Data Mining

Data mining is generally defined as the extraction of implicit information from a big database using
some methods such as statistics, artificial intelligence, or machine learning so that the information is
understandable and useful. [5]
Data mining commonly involves four tasks: [10]
 1. Classification
The purpose of classification is to classify or divide data into some categories, for example
classifying employee income into the low, medium, or high categories.
2. Clustering
The purpose of clustering is to group data into some groups based on similarities.
3. Regression
The purpose of regression is to find a function that becomes the data model with the least error.
4. Association Rule Learning
The purpose of the Association Rule Learning is to find relationships between available
variables.

2.3. Artificial Neural Network (ANN)

Artificial Neural Network (ANN) is a model built to imitate the workings of neurons in the human brain
because humans are considered the most perfect system [4]. The architecture of this model is shown in
figure 1.

Figure 1. ANN Architecture

Based on figure 2, the following is an explanation of some common types of layers in ANN [4]:

1. Input Layer: Layer that receives input data features from the outside world and distributes it to
the hidden layer.
2. Hidden Layer: Layer that accepts input from input layer and performs calculation process and
converts input value into output value by using activation functions such as sigmoid, tanh
(Hyperbolic Tangent), or ReLu (Rectified Linear Units). After the input value is converted to
an output value, the hidden layer will distribute it to the output layer.
3. Output Layer: Layer that receives input from the hidden layer and distributes information from
the network to the outside world.
3. Results and Discussion
Prediction of item sales at each Big Mart outlet is done using Orange data mining tool. Orange is a
Python-based tool for a general-purpose machine learning and data mining tool developed at the
Bioinformatics Laboratory of the Faculty of Computer and Information Science at the University of
Ljubljana. programming. It offers a structured view of supported functionalities grouped into some
categories such as data operations, visualization, classification, regression, evaluation, unsupervised
learning, association, etc[11]. After the prediction results are obtained, the predicted data will be
visualized into the scatter plot. Based on the results of predictions, it appears that low-fat items have
the highest sales as shown in Figure 2 because low-fat items are generally used more every day than
others.

Figure 2. Scatter Plot Item_Fat_Content

Meanwhile, the types of items that have the highest sales are other types compared to the types of
baking goods, canned, dairy, etc. (see figure 3).
 Figure 3. Scatter Plot Item_Type

However, outlets located in cities or Tier 1 will produce the highest sales as shown in figure 4. Outlets
that produce the highest sales are OUT049 (see figure 5).

Figure 4. Scatter Plot Outlet_Location_Type

 Figure 5. Scatter Plot Outlet_Identifier

Outlet size does not guarantee high sales, because in this study medium size outlets produce the highest
sales compared to high-size outlets (see figure 6).

Figure 6. Scatter Plot Outlet_Size

The types of outlets that produces the highest sales is Supermarket Type1 as shown in figure 7.
 Figure 7. Scatter Plot Outlet_Type

Prediction results using the ANN algorithm with the Root Mean Squared Error (RMSE) value of
1127.239 show great predictive results (see Table 2).

Table 2. Prediction results of the ANN algorithm

Evaluation Criteria Value
Mean Squared Error (MSE) 1270667.116
Root Mean Squared Error (RMSE) 1127.239
Mean Absolute Error (MAE) 847.305
R Squared (R2) 0.370

4. Conclusion
Sales predictions play a crucial role in the continuation of future business operations for all companies,
especially for business chain retailers such as Big Mart. Accurate sales predictions can have a major
impact on the effectiveness of retail chain business operations management. The ANN algorithm that
is applied to predict sales of each item in each outlet works great with the RMSE value of 1127.239.
ANN is often superior to other algorithms in certain studies. Prediction results that have been done
show that the location of an outlet greatly affects the sales results, and the size of the outlet capacity
does not affect the sales results. The retail chain business can use the result of this research for managing
the right business strategy in the future.

5. Acknowledge
This research was supported by Universitas Komputer Indonesia, Indonesia.

References
[1] Scherer, M. (2018). Multi-layer neural networks for sales forecasting. Journal of Applied
Mathematics and Computational Mechanics, 17(1).
[2] Penpece, D., & Elma, O. E. (2014). Predicting sales revenue by using artificial neural network in
grocery retailing industry: a case study in Turkey. International Journal of Trade, Economics
and Finance, 5(5), 435.
[3] Brühl, B., Hülsmann, M., Borscheid, D., Friedrich, C. M., & Reith, D. (2009, July). A sales forecast
model for the german automobile market based on time series analysis and data mining
methods. In Industrial Conference on Data Mining (pp. 146-160). Springer, Berlin, Heidelberg.
[4] Kuo, R. J., Wang, Y. C., & Tien, F. C. (2010). Integration of artificial neural network and MADA
methods for green supplier selection. Journal of cleaner production, 18(12), 1161-1170.
[5] Krause-Traudes, M., Scheider, S., Rüping, S., & Meßner, H. (2008). Spatial data mining for retail
sales forecasting. In 11th AGILE International Conference on Geographic Information
Science (pp. 1-11).
[6] Massaro, A., Maritati, V., & Galiano, A. (2018). Data Mining model performance of sales
predictive algorithms based on RapidMiner workflows. Int. J. Comp. Sci. Inf. Technol, 10, 39-
56.
[7] Razak, A., & Riksakomara, E. (2017). Peramalan Jumlah Produksi Ikan dengan Menggunakan
Backpropagation Neural Network (Studi Kasus: UPTD Pelabuhan Perikanan
Banjarmasin. Jurnal Teknik ITS, 6(1), 138-141.
[8] Farahani, D. S., Momeni, M., & Amiri, N. S. (2016). Car Sales Forecasting Using Artificial Neural
Networks and Analytical Hierarchy Process. DATA ANALYTICS 2016, 69.
[9] Pakaja, F., Naba, A., & Purwanto, P. (2012). Peramalan Penjualan Mobil Menggunakan Jaringan
Syaraf Tiruan dan Certainty Factor. Jurnal EECCIS, 6(1), 23-28.
[10] Chauhan, A., Mishra, G., & Kumar, G. (2011). Survey on data mining techniques in intrusion
detection. International Journal of Scientific & Engineering Research, 2(7), 1-4.
[11] Jovic, A., Brkic, K., & Bogunovic, N. (2014, May). An overview of free software tools for general
data mining. In 2014 37th International Convention on Information and Communication
Technology, Electronics and Microelectronics (MIPRO) (pp. 1112-1117). IEEE.