THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
23
https://www.theamericanjournals.com/index.php/tajet
PUBLISHED DATE: - 02-10-2024
https://doi.org/10.37547/tajet/Volume06Issue10-04
PAGE NO.: - 23-32
UNSUPERVISED MACHINE LEARNING AND
VECTOR MODELS IN DESIGNING AND
OPTIMIZATION OF TELECOM RETAIL
CHANNELS
Andrei Zhuk
Manager, McKinsey & Company, New York
Wharton School of Business (University of Pennsylvania), Philadelphia, USA
MGIMO, Moscow, Russia
INTRODUCTION
In the context of rapid technological development,
both retailers and telecom operators face
challenges in managing physical sales networks, as
maintaining a network of physical stores is one of
the primary cost drivers. The efficiency of
managing these stores is influenced by modern
technological solutions. Traditional methods of
data analysis and demand forecasting are
becoming less effective due to the increasing
volume and complexity of data. In this context, the
use of unsupervised machine learning and vector
models appears particularly promising.
Unsupervised machine learning enables the
analysis of large volumes of unstructured data,
uncovering hidden patterns and optimizing
business processes without the need for pre-
labeled data. This approach is especially relevant
in the telecommunications sector, where
consumer data and behavior play a key role in
decision-making, making this topic highly relevant.
The use of unsupervised machine learning and
vector models significantly enhances operational
efficiency, leading to cost reduction and improved
customer experience. These technologies open
new possibilities for analysis and forecasting,
contributing to more accurate and faster decision-
making.
RESEARCH ARTICLE
Open Access
Abstract
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
24
https://www.theamericanjournals.com/index.php/tajet
The aim of this work is to explore the potential
application of unsupervised machine learning and
vector models for designing and optimizing retail
channels for telecommunications services.
1. Application of Unsupervised Machine
Learning in Consumer Data Analysis
In the process of studying and processing data
within the framework of machine learning,
especially when dealing with models that operate
on information, it is crucial to use accurate and
reliable data. Training models implies that
algorithms must receive the necessary data to
effectively perform tasks using artificial
intelligence. While this approach is a popular
method in machine learning, certain challenges
arise with its use. One of the most significant
challenges is the issue of data labeling, as finding
accurately labeled data for feeding into the model
is often difficult. Moreover, the cost of data can be
high, and its use may not always yield the expected
results. Currently, alternative methods are being
developed that have not yet gained widespread use
but may become important in the future.
One method worth noting is label-free learning. In
this technique, the model receives raw data
without predefined labels or patterns. The
algorithm independently analyzes this data,
generating new patterns and labels. One of the
advantages of this approach is that there is no need
to provide labeled data. The system autonomously
identifies the rules necessary for analysis and
generates its own patterns. The process of
unsupervised learning involves several key stages,
which are illustrated in Figure 1.
Fig.1. The stages of unsupervised learning [1].
Unsupervised learning algorithms can be classified into two main types based on the methods used for
data processing, as shown in Figure 2.
The study of data structure
algorithms and the formation
of templates
Extracting the necessary
knowledge that can be used in
the future for analysis
The decision-making process
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
25
https://www.theamericanjournals.com/index.php/tajet
Fig.2. Classification of unsupervised learning algorithms [1].
Overall, unsupervised learning methods provide
extensive opportunities for data analysis and
processing, making them essential tools in the
development of artificial intelligence [1]. The main
distinction between supervised and unsupervised
learning lies in the use of labeled data. Supervised
learning relies on pre-known labeled input and
output data, allowing the algorithm to learn by
adjusting its predictions until the correct result is
achieved. This requires human involvement and
accurate data labeling. An example of this would be
predicting commute times based on various
factors, where the model is first trained on
historical data [2].
There are other differences between these
approaches. The goal of supervised learning is to
predict outcomes for new data based on known
patterns. In unsupervised learning, the focus is on
uncovering hidden information from a large
volume of new data.
Supervised learning is suitable for classification
tasks, such as determining spam, sentiment
analysis of texts, or weather forecasting.
Unsupervised learning, in turn, is useful for tasks
related to anomaly detection, improving
recommendation systems, predicting customer
behavior, and analyzing medical images [3].
2. Vector Models in Demand Forecasting and
Distribution Channel Optimization
To obtain accurate data on future customer churn,
a deep understanding of classification tasks, which
are part of data mining, is required. The essence of
this task is to create a mathematical model based
on historical customer data to predict their
behavior using the available information. In this
forecast, customers are divided into two classes,
and the term "label" is used to designate them,
which can take two values:
- High probability of customer churn (churn).
- High probability of continuing service usage [4].
An important stage before training is data
Uncontrolled neural networks
These networks are trained on
unlabeled data, which allows
them to solve regression and
classification problems in the
absence of supervision.
Methods of unsupervised learning
Clustering.This method is aimed at
finding patterns in the data and
grouping them.
Dimensionality reduction simplifies
data analysis by reducing the
number of variables and identifying
key features
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
26
https://www.theamericanjournals.com/index.php/tajet
preparation, including cleaning outliers and
eliminating duplicate features. Correlation
analysis helped identify and remove redundant
features, improvi
ng the model’s quality. To test the
model’s accuracy, the data was split into training
and test sets. As a result, it was found that the
model's accuracy reached 92%, confirming the
high effectiveness of using the support vector
machine (SVM) method in predicting customer
churn.
Vector data models are a powerful tool for
analyzing and optimizing business processes,
including logistics and inventory management in a
retail network. These models represent objects
and their characteristics as multidimensional
vectors, enabling a more accurate analysis of the
relationships between different parameters. In this
context, vector model construction methods
mainly include factor analysis, clustering, and
neural networks.
Factor analysis is used to reduce data
dimensionality and highlight the key factors that
most influence stock dynamics and logistical
processes. This allows for models that focus on the
most significant characteristics, reducing excess
information and simplifying the optimization
process.
Neural networks, in turn, are used for demand
forecasting based on historical data, a key element
in inventory management. Recurrent neural
networks (RNNs) and their modifications, such as
LSTM, can model complex temporal dependencies,
making them effective tools for predicting the
demand for specific products in the retail network.
The practical application of vector models for
optimizing logistics and inventory management
manifests in several areas. First, they allow for
more accurate forecasts of inventory needs,
reducing the risk of overstocking or understocking.
Second, these models help optimize delivery
routes, minimizing time and logistics costs. Third,
vector models can be used to automate inventory
management processes, reducing labor costs and
increasing decision-making accuracy.
Additionally, a program implementing an SVM
classifier (support vector machine used for
classification and regression analysis tasks) was
tested for accuracy and performance speed. The
conclusions drawn from the analysis indicate that
the model can be successfully used to classify new
customers, with a high degree of confidence in its
predictions [5].
Thus, vector models represent an effective tool for
improving logistics and inventory management in
the retail network, providing more accurate
forecasting, process optimization, and automation
of routine operations.
3. Interaction of Unsupervised Machine
Learning and Vector Models in Designing Sales
Channels
One of the key methods of unsupervised machine
learning applied in this field is clustering.
Clustering allows grouping data based on similar
features, identifying segments of consumers,
product types, or geographic regions with shared
characteristics. In this context, vector data models,
which represent objects as multidimensional
vectors, serve as the foundation for conducting
cluster analysis. For example, customer profiles,
represented as vectors, can be divided into
clusters, enabling more precise adjustments to
marketing strategies and sales offers for different
audience segments [6].
Another important area of applying unsupervised
machine learning in combination with vector
models is cost reduction and logistics optimization.
Dimensionality reduction algorithms, such as
principal component analysis (PCA), can be used to
simplify multidimensional vector models that
represent various logistics process parameters.
This simplification helps identify key factors
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
27
https://www.theamericanjournals.com/index.php/tajet
influencing the efficiency of sales channels, such as
delivery speed, inventory levels, or transportation
routes. The patterns identified are then used to
design more efficient logistics systems.
Neural networks that utilize unsupervised
learning, such as autoencoders, also find
application in the design of sales channels. These
models are capable of identifying latent data
representations that are not obvious using
traditional analysis methods. Vector models
trained with autoencoders can assist in
automatically detecting anomalies in sales
processes or predicting changes in consumer
behavior, allowing sales channels to adapt to new
market conditions and customer needs [7].
Thus, the interaction of unsupervised machine
learning and vector models in the design of sales
channels provides deeper data analysis and opens
up opportunities for adaptive management of sales
processes. This interaction not only reveals hidden
patterns and structures in the data but also utilizes
them to create more efficient, flexible, and resilient
sales channels.
4. Practical Experience
As part of professional activities, an advanced
analytical vector model was created to assess the
potential of micro-markets based on both internal
and external data, for both existing and new stores.
An information dashboard was also developed to
display the potential of individual stores and
compare them with similar stores in the network.
As a result, actual gross profit from sales in 245
new telecom retail locations increased by 5-10%
over the course of 6 months.
In terms of automation of store format and
assortment management, four key tools were
developed to manage and optimize store space:
1. Location Monitoring Dashboard. This tool allows
for the assessment of a store’s potential in terms of
revenue, margin, or other key performance
indicators of choice. It can then determine the key
success factors and risks for the location, as well as
compare performance metrics with those of other
similar locations.
2. Heat Map of Potential Stores. This tool enables
the organization to create visual results on a map,
highlighting areas with the highest potential. It
also provides the ability to view detailed
information (e.g., demographics, competition,
store density) through a "double-click" function.
3. Optimization of New Store Locations. This
feature determines the optimal number of stores
and their locations, evaluates the return on
investment (ROI) for each location, and allows
scenario analysis based on business assumptions.
4. Setting Sales Targets, Managing Performance,
and Choosing Store Closure Options. This includes
a comparative analysis of the store’s key
performance indicators with other similar stores
in similar locations (see Fig.4).
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
28
https://www.theamericanjournals.com/index.php/tajet
Fig.3. An example of a map.
Initial data used:
- Residential buildings with aggregated data on the
number of apartments and population breakdown
by age;
Based on the data, an information panel was
developed,
providing
details
on
store
characteristics, key performance indicators, and its
driving forces (Fig.4).
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
29
https://www.theamericanjournals.com/index.php/tajet
Fig. 4. Information panel.
Next, let's consider the process of creating a
heatmap based on the available data. The map
displays address points that are further analyzed.
Each location is scored, with the monthly revenue
forecast based on geospatial analysis using:
- Demographic data (e.g., population density,
income, age, education);
- Competitor data (e.g., proximity to competing
stores);
- Pedestrian traffic data;
- POI (proximity to traffic-generating objects like
offices or tourist attractions);
- Weather data (e.g., deviations from typical
weather conditions based on data from 100
measurement points);
- Performance data (e.g., store size, quality audit
results).
The results obtained from the model are
interpolated to identify areas with similar
potential (Fig. 5).
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
30
https://www.theamericanjournals.com/index.php/tajet
Fig. 5. Heatmap of sample residential estate.
Evaluation of return on investment (ROI) for a new
store. A separate EBITDA assessment model was
developed for new stores based on:
- Expected revenue;
- Assortment and store segment affecting the
margin;
- Operating costs associated with the type of
location.
For each location, the model forecasts:
- Gross profit;
- Logistics costs;
- Rent expenses;
- Personnel costs;
- Other expenses.
Based on the above, the payback period was also
calculated (Fig. 6).
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
31
https://www.theamericanjournals.com/index.php/tajet
Fig. 6. Heatmap of sample residential estate.
Next, we look at geocoding customers and
forecasting probabilities using deep analytics
models
with
individual
characteristics
(residence/workplace, digitalization, grocery
penetration, etc.). The optimization process
simultaneously compares the possibility of
relocation
to
multiple
locations,
which
significantly complicates the calculations. This
method can be used to refine customer flow for the
final footprint and assess individual actions.
Advantages:
- Universal function for calculating churn/flow;
- Depends on the number of neighboring branches
and distance;
- Accounts for actual city/district-level distribution
to reflect travel tendencies;
- Has a probabilistic interpretation.
The main disadvantage of this method is that
actual customer distribution and flow per specific
branch may vary significantly.
Following changes in the system, cannibalization
variables from the current network are updated,
and the potential in all areas within the impact
zone is recalculated. Effective data processing tools
are required to use this method. It can be used to
refine customer potential for the final footprint
and evaluate specific actions. By using dynamic
updates of geo-potential maps, it is possible to
account for non-linear relationships and the actual
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE10
32
https://www.theamericanjournals.com/index.php/tajet
distribution of traffic generators on the map.
However, this method requires significant
computational power, and, considering the
specifics of the training set, the results may not
align with economic logic in edge scenarios.
CONCLUSION
Thus, the application of unsupervised machine
learning and vector models in the design and
optimization
of
retail
channels
for
telecommunications services opens new horizons
for improving business efficiency. These methods
allow for the discovery of hidden patterns in large
data sets, enhance the accuracy of demand
forecasting, and optimize logistics processes.
Despite
the
challenges
associated
with
implementing these technologies, their use
provides significant competitive advantages,
improving customer experience and the
company’s operational performance. In the future,
further research in this area may contribute to the
development of more advanced models and
methods capable of even more effectively
addressing the optimization tasks of sales channels
in the telecommunications industry.
REFERENCES
1.
Zaadnoordeik L., Besold T. R., Cusack R. Infant
learning lessons for unattended machine
learning //Nature Machine Intelligence.
–
2022.
–
Vol. 4.
–
No. 6.
–
pp. 510-520.
2.
Allogani M. et al. A systematic review of
supervised and unsupervised machine
learning algorithms for data science
//Supervised and unsupervised learning for
data science. - 2020.
–
pp. 3-21.
3.
Fadokun D. O., Oshilike I. B., Onyekonwu M. O.
An approach to machine learning with and
without control in forecasting facies // Annual
International Conference and Exhibition SPE in
Nigeria.
–
SPE,
2020.
–
Number
D013S014R007.
4.
Ma S., Fildes R. Forecasting third-party mobile
payments,
taking
into
account
the
consequences for forecasting the flow of
customers
//International
Journal
of
Forecasting.
–
2020.
–
vol. 36.
–
No. 3.
–
pp. 739-
760.
5.
Guven I., Shimshir F. Forecasting demand using
color parameters in the retail clothing industry
using artificial neural network (ANN) and
support vector machine (SVM) methods
//Computers and Industrial Engineering. -
2020.
–
Vol. 147.
–
p. 106678.
6.
Jansen S. Machine Learning for Algorithmic
Trading: Predictive models for extracting
signals from market and alternative data for
systematic trading strategies using Python.
–
Packt Publishing Ltd, 2020.
7.
Ho L., Getals P. Application of machine learning
in river research: trends, opportunities and
challenges //Methods in ecology and
evolution.
–
2022.
–
vol. 13.
–
No. 11.
–
pp.
2603-2621.
