How works Artificial Intelligence for risk management

Summary

In this webinar, Nicole Koza and data scientist Hi Faguna explore the integration of artificial intelligence in risk management. They define AI, delve into machine learning and deep learning, and discuss their applications in client segmentation and risk recommendation systems using chat GPT. The session emphasizes the importance of aligning AI techniques with business goals, the significance of data cleaning, and the practical steps for achieving accurate client segmentation. It also highlights the potential of natural language processing in understanding and generating human language for risk assessment.

Takeaways

• 😀 The webinar focuses on the theoretical and practical aspects of artificial intelligence in risk management.
• 🔍 The speaker, Faguna, introduces the concept of artificial intelligence (AI) as the ability to reason, solve problems, and learn from experience without human intervention.
• 🤖 AI is composed of two main areas: machine learning and deep learning, which are responsible for developing and training algorithms and models.
• 📚 The importance of data science is highlighted, emphasizing the need for a clear business goal before applying AI techniques.
• 📈 Two types of machine learning are discussed: supervised learning, which requires a target variable, and unsupervised learning, which does not.
• 🏦 An example of supervised learning is identifying fraudulent transactions in a bank's dataset, while unsupervised learning might be used for customer segmentation in a retail company.
• 🤝 The benefits of machine learning for segmentation include accuracy, automation, scalability, adaptability, and personalization.
• 🔍 Two clustering algorithms are mentioned: k-means, which is simple and efficient but sensitive to initialization, and Birch, which is robust to outliers and suitable for large datasets.
• 🧹 The critical step of data cleaning is discussed, including outlier detection, categorical variable encoding, data normalization, and principal component analysis.
• 📊 Two methods for determining the optimal number of segments are presented: the elbow method, which uses inertia, and the silhouette method, which considers cohesion and separation.
• 💡 The platform demonstration shows how to perform client segmentation using the discussed techniques and tools, emphasizing the importance of analyzing results from a business perspective.

Q & A

What is the main focus of the webinar presented by Nicole Koza and Data Scientist Faguna?

The main focus of the webinar is to explore the theoretical and practical aspects of artificial intervention in risk management, including applications of artificial intelligence, machine learning for segmentation, and natural language processing techniques.

What is the definition of artificial intelligence as discussed in the webinar?

Artificial intelligence is defined as the ability of a computer program or application to reason, solve problems, understand complex ideas, learn quickly, and learn from experience, without the need for human intervention.

What are the two main sub-areas of artificial intelligence mentioned in the script?

The two main sub-areas of artificial intelligence mentioned are machine learning and deep learning, which are responsible for developing and training the algorithms and models used in AI applications.

Why is data science important in the context of artificial intelligence applications?

Data science is important because it helps to identify clear business goals or problems that AI applications aim to solve or opportunities they aim to leverage. It ensures that AI techniques are used as tools to achieve specific business objectives rather than being an end in themselves.

What are the two types of machine learning mentioned in the script?

The two types of machine learning mentioned are supervised learning, which requires a target variable, and unsupervised learning, which does not require a target variable and is often used for pattern recognition or anomaly detection.

Can you explain the purpose of client segmentation using machine learning techniques?

The purpose of client segmentation using machine learning is to identify groups of clients with similar characteristics, which can help in understanding the customer base better, discovering unknown patterns, complying with regulations, and improving transaction monitoring.

What are the advantages of using machine learning for client segmentation compared to manual methods?

The advantages of using machine learning for client segmentation include higher accuracy with large datasets, automation, scalability, adaptability, personalization of segmentation characteristics, and improved time efficiency.

What are the two clustering algorithms mentioned in the script, and what are their main differences?

The two clustering algorithms mentioned are k-means and Birch. K-means is simpler and more efficient but can be sensitive to initialization and outliers. Birch is more robust to outliers and better suited for large datasets and high-dimensional data.

Why is data cleaning an essential step in the machine learning process for segmentation?

Data cleaning is essential because it ensures the quality of the data used for training models. Steps like outlier detection, categorical variable encoding, data normalization, and principal component analysis help to prepare the data for accurate and meaningful analysis.

What are the two methods mentioned for determining the optimal number of segments in a dataset?

The two methods mentioned for determining the optimal number of segments are the elbow method, which uses the inertia metric, and the silhouette method, which considers cohesion and separation parameters.

What is the role of natural language processing (NLP) in risk management applications?

In risk management, NLP can be used for sentiment analysis to understand customer feedback, text classification to categorize risks, and chatbots to interact with clients and provide risk recommendations or information.

What is the significance of GPT in the context of the presented risk management application?

GPT, or Generative Pre-trained Transformer, is significant because it can generate new text data based on the input provided. In the risk management application, it uses various client-related variables to suggest potential risks, aiding in the identification and mitigation of such risks.

How does the platform use chat GPT to assist clients in identifying potential risks?

The platform uses chat GPT by inputting variables such as industry, process, and risk system into the model's prompt. GPT then generates responses and recommendations for potential risks based on the pre-trained data it has from the internet.

What are the precautions taken to ensure client data privacy when using chat GPT in the platform?

To ensure client data privacy, the platform does not provide specific client information to chat GPT. Instead, it uses general variables like industry and process, relying on the model's pre-trained data to generate risk recommendations.

Outlines

00:00

👋 Introduction to Risk Management and AI

The script begins with Nicole Koza welcoming participants to a webinar on artificial intervention for risk management. She introduces the agenda, which includes defining artificial intelligence (AI), discussing machine learning for segmentation, and exploring natural language processing with chat GPT. Nicole also invites attendees to visit the Pirani Academy for learning materials and encourages questions throughout the session. The session's first speaker, a data scientist, outlines the day's topics, emphasizing the importance of understanding AI's theoretical and practical aspects and their applications in risk management.

05:02

🤖 Understanding Artificial Intelligence Composition

The speaker delves into the definition of artificial intelligence, distinguishing between human intelligence and AI's ability to learn from experience without human intervention. They explain the composition of AI, highlighting the importance of machine learning and deep learning in developing algorithms and models. The speaker also discusses the role of data science in aligning AI applications with clear business goals, emphasizing that AI is a tool to aid in solving complex problems rather than a standalone solution.

10:04

📚 Machine Learning Techniques: Supervised and Unsupervised Learning

The script explains the two primary methods of machine learning: supervised and unsupervised learning. Supervised learning is illustrated with examples from banking and retail sectors, where historical data is used to predict outcomes such as fraudulent transactions or sales amounts. Unsupervised learning is characterized by the absence of a target variable, with clustering and segmentation as common applications. The speaker also discusses the importance of identifying business problems before applying machine learning techniques.

15:06

🔍 Applications of Machine Learning in Anti-Money Laundering

The speaker introduces a specific application of machine learning in anti-money laundering, focusing on client segmentation to identify similar groups for better understanding and compliance with regulations. They discuss the advantages of using machine learning for scalability, adaptability, and time efficiency, especially when dealing with large datasets. The script also mentions different clustering algorithms, such as k-means and Birch, and their suitability for various data sizes and dimensions.

20:08

🧼 Data Cleaning and Clustering Algorithms

This paragraph emphasizes the importance of data cleaning in the machine learning process, outlining steps such as outlier detection, categorical variable encoding, data normalization, and principal component analysis. The speaker explains that these steps are crucial for accurate results, regardless of the sophistication of the algorithms used. They also discuss the k-means and Birch algorithms, providing insights into their functionality and ideal use cases.

25:10

📊 Optimal Segment Selection and Platform Demonstration

The script describes methods for determining the optimal number of segments in a dataset, including the elbow method and the silhouette metric. These methods help in finding a balance between too many and too few segments for effective client understanding. The speaker then transitions to demonstrating the platform where these steps can be performed, including data selection, cleaning, and the application of machine learning models.

30:12

📝 Natural Language Processing and Its Applications

The speaker introduces natural language processing (NLP) as a subfield of AI that combines computer science and linguistics to process and understand human language. They discuss various applications of NLP, such as sentiment analysis, text classification, and chatbots, which are used to analyze sentiments in social media, classify risks, and interact with customers, respectively.

35:14

🗣️ Chat GPT and Its Risk Management Applications

The script explores the use of chat GPT, a generative pre-trained transformer model, in risk management. The speaker explains how chat GPT can be used to recommend potential risks based on various client-related variables. They also mention other applications of chat GPT, such as risk classification and suggesting mitigation actions and possible causes of risks.

40:16

🏢 Platform Features and Risk Recommendation System

The speaker demonstrates the platform's features for risk management, including the use of chat GPT to suggest potential risks based on industry, process, and other client variables. They show how users can select variables and receive risk suggestions, which can then be used to create new risk entries in the system. The script highlights the platform's ability to provide suggestions across various industries and company processes.

45:19

📉 Conclusion and Future Improvements

In conclusion, the speaker summarizes the importance of aligning AI applications with business goals and the necessity of thorough data cleaning in machine learning. They also emphasize the value of comparing different methodologies to find the best solution for a company's needs. The script ends with an invitation for participants to complete a survey to help improve future webinars and thanks them for their participation.

Keywords

💡Artificial Intelligence (AI)

Artificial Intelligence refers to the simulation of human intelligence in computers that are programmed to think like humans and mimic their actions. In the context of the video, AI is the overarching theme, focusing on its applications in risk management. The script discusses how AI can be utilized as a tool to assist with complex problem-solving and efficiency in tasks, exemplified by the use of AI in machine learning models and natural language processing for risk assessment.

💡Risk Management

Risk management is the process of identifying, assessing, and prioritizing potential risks to an organization's capital and earnings. The video's theme revolves around using AI for risk management, specifically in the context of anti-money laundering and operational risks. The script mentions the practical applications of AI in identifying fraudulent transactions and classifying risks, which are key components of managing an organization's exposure to potential losses.

💡Machine Learning

Machine learning is a subset of AI that provides systems the ability to learn and improve from experience without being explicitly programmed. The script explains machine learning techniques such as supervised and unsupervised learning, which are used for tasks like client segmentation and transaction monitoring in the field of risk management.

💡Supervised Learning

Supervised learning is an approach in machine learning where the algorithm is trained on a labeled dataset, learning to predict outcomes for new data. In the script, it is exemplified by training a model to predict fraudulent transactions in a bank's dataset, where the target variable is the classification of transactions as fraudulent or not.

💡Unsupervised Learning

Unsupervised learning is a type of machine learning where the data used to train the algorithm is not labeled. The script discusses its use in scenarios like client segmentation, where the algorithm identifies patterns or groups without a predefined target variable, helping to uncover unknown patterns in the data.

💡Client Segmentation

Client segmentation is the process of dividing a customer base into groups with similar characteristics. In the video, it is highlighted as a goal for understanding clients better and for compliance with regulations. The script describes using machine learning algorithms to identify similar groups of clients for more effective risk monitoring and personalized services.

💡Natural Language Processing (NLP)

Natural Language Processing is a field of AI that enables computers to understand, interpret, and generate human language. The script discusses NLP in the context of applications like sentiment analysis and chatbots, particularly focusing on using GPT (Generative Pre-trained Transformer) models to create a risk recommendation system that can understand and generate text related to potential risks.

💡Generative Pre-trained Transformer (GPT)

GPT is a type of deep learning model that has been pre-trained on a large corpus of text data, enabling it to generate human-like text. The script explains the use of GPT in risk management, specifically for creating a system that suggests potential risks based on various client and company characteristics, without needing to input sensitive client information directly.

💡Data Science

Data science is a field that uses scientific methods, processes, and algorithms to extract knowledge and insights from data. In the script, data science is mentioned as an essential part of AI applications, emphasizing the need for a clear business goal before applying machine learning or deep learning techniques, ensuring that these tools are used to solve specific business problems effectively.

💡Principal Component Analysis (PCA)

PCA is a statistical technique used to emphasize the variation and bring out strong patterns in a dataset. The script refers to PCA as a step in the data cleaning process, where it helps identify the most relevant characteristics for analysis by reducing the dimensionality of the data set and excluding variables that do not contribute to the segmentation process.

💡K-means

K-means is a clustering algorithm used in machine learning to partition a dataset into a specified number of clusters. The script explains the process of using K-means for client segmentation, where the algorithm identifies the optimal number of segments and assigns clients to clusters based on their similarity, helping in understanding and managing client relationships more effectively.

Highlights

Introduction to the webinar on artificial intervention for risk management by Nicole Koza and data scientist Faguna.

Invitation to learn more about risk management through the Pirani Academy and a provided link.

Discussion on the theoretical and practical aspects of artificial intelligence (AI) in risk management.

Definition of AI as the ability to reason, solve problems, and learn from experience without human intervention.

Explanation of the composition of AI, including machine learning, deep learning, and data science.

Emphasis on having a clear business goal before applying AI techniques like machine learning.

Introduction to supervised and unsupervised learning in machine learning.

Example of using supervised learning for fraud detection in banking transactions.

Explanation of unsupervised learning for segmentation and anomaly detection without a target variable.

Application of machine learning for anti-money laundering through client segmentation.

Importance of data cleaning in the machine learning process for accurate results.

Description of the k-means clustering algorithm for segmenting clients into groups.

Introduction of the Birch algorithm for large datasets and its robustness to outliers.

Discussion on the advantages of machine learning for scalability, adaptability, and personalization in client segmentation.

Process of choosing the optimal number of segments using the elbow method and silhouette metric.

Demonstration of the platform's capability to perform segmentation steps and data cleaning.

Overview of the platform's features for compliance, operational risk, information security, and anti-money laundering.

Introduction to natural language processing (NLP) as a subfield of AI for text data processing and generation.

Examples of NLP applications in sentiment analysis, text classification, and chatbots.

Explanation of GPT (Generative Pre-trained Transformer) model and its use in risk management.

Application of chat GPT for recommending potential risks in a risk management system.

Description of the process for using chat GPT to suggest risks based on industry, process, and other variables.

Summary emphasizing the importance of focusing on business goals and the significance of data cleaning in AI applications.

Transcripts

hello everyone Welcome to our rates

management School my name is Nicole Koza

and I'm a member of the purani team I'm

excited to be here with you today to

explore the theoretical and practical

aspect of the different topics that we

will see

to help ourselves into this session on

artificial intervention for risk

management we are happy to have with us

data scientist at kirani Hi faguna how

are you

hi everyone Thanks for for being part of

this webinar how are you

cool just exciting to to get started and

before we begin I would love to invite

everyone to continue learning about the

world of risk management school by

visiting the pirani academy where you

will find valuable learning materials I

will leave the link in that chat now so

you can see the website

so there you are

done

and if you have any questions during

that webinar please feel free to make

them through the question and answer

section we'll be happy to answer them at

the end of this webinar now let's begin

with this exciting topic over to you

okay hey thank you again everyone for

being part of this webinar

today we're going to be talking about

artificial intelligence and some of its

applications into risk management

is to First Define this concept of

artificial intelligence and understand

how is it composed

and then we are going to start

explaining our first application of

machine learning for segmentation

and finally we are going to be talking

about natural language processing

techniques

and an application that we also have

using chat GPD to create a risk

recommendation system

so well that's the the agenda for today

obviously if you have questions Nicole

feel free to interrupt me and we are

going to try to to solve them in the

moment

um I'm going to try to to explain like

the theory first and then we're going to

to go to the platform to

to show you all the steps

um okay so what we we're gonna start

um I think that the best way to

understand this new concept of

artificial intelligence is to First

Define both words separately and so what

do we mean by intelligence well that's

the ability of reasoning of solving

problems

understanding complex ideas and being

able to learn quickly and learn from

experience and well why why do we say

that this is artificial well because we

don't want to employ a human resource to

develop those tasks or or those problems

that we can get done using a computer

program or an artificial intelligence

application

I mean that the idea here is that we can

use the human intelligence for more

complex problems and deliver like the

more simpler task or to a computer

program or an application that can help

us

um obviously we don't pretend that this

artificial intelligence to do all the

job but but we are going to use it as a

tool to help us and

and maybe achieve better results or or

to get the task done more more quickly

so well another thing that it's

important is to to understand how is it

composed uh the this artificial

intelligence and you need to know

that there are two main soup areas

um these two areas are called machine

learning and deep learning

and these two areas are responsible for

developing and training the algorithms

and models that we are going to use in

our final artificial intelligence

application

so

these areas use like a huge amount of

data to train the models and this model

is what we are going to use in the final

application for example

a a really common nowadays application

of this is chat CPT so we can consider

the GPT as the final result as the final

artificial intelligence application that

we use to process and generate a text

and also called but

we need to understand that this

application was trained before using a

huge amount text Data I'm using some

deep learning techniques so well this is

a how is it composed this this whole box

um and another important thing is the

other Circle that you can see in the

right

it's called data science and what's

what's the importance of this well

because we need to understand that we

should always need to have like a

business goal for example a business

problem that we want to solve or a

business opportunity that we want to

take advantage of

but that's that's need to be like really

clear we never need to use like a

machine learning or deep learning uh a

goal itself these are just tools and

techniques that we can use to help us

achieve a business goal in this case for

example to solve a business problem but

that's all so so the process should

should ever be to First identify these

business problems or opportunities

and to try to fix them uh in a simpler

way and then if we think that the

problem is really complex well in that

case yes let's use machine learning for

example but only if if this techniques

is going to to help us to achieve a

better result

so that's I consider it really important

and well before explaining our first

application another thing that we need

to know is

that in machine learning there are two

ways of training a model these two ways

are called supervised learning and

unsupervised learning

and well the supervised learning uh the

the most important part in this

technique is that we need to have like a

Target variable

and I will explain this with an example

so suppose that we are a bank for

example

um we have a data set with a lot of

transactions from our clients and in

this data set we have a column that

identified is if this transaction is

problem or not problem

so in that case our Target variable will

be that column the column that

classifies the transactions into

fraudulent and not from

so we can use this whole data set to

train a machine learning model using

supervised learning techniques and we

want that this model is able to predict

in the future if a new transaction is

frozen or not from so that's that's the

concept of targeted balance

another example let's suppose that we we

are a retail company

um we have a data set with a lot of

products and we have a column A that

shows them the amount of sales for each

product so in this case our alternative

variable what we want to predict is that

column the amount of size so using this

historical data we can train a model and

this model this model is going to

understand and be able to predict in the

future the amounts of sales for each

product

uh yes Nicole you have a question I know

okay sorry sorry

um okay so that's that's the case of a

supervised learner

what what happened in the other case on

unsupervised learning well here we don't

have like our

our Target variable

what do we have here it's like we have a

data set for example the most common

example here is an application for

clustering or segmentation

so let's suppose that we have a data set

with characteristics from from all of

our clients and we are trying to find

something find a pattern find some

groups that have a like similar

characteristics uh but we don't have in

this case our Target value that's why

it's called unsupervised learning

so uh as I told you the most common news

is for for clustering for example uh

where we want to find like patterns and

another really common use case is for

cyber security or for anomaly detection

for example we have a data set with all

of the transactions that our clients

um made in in our in our bank

but we don't have the column that

classifies these transactions in problem

or not problem

so in that case we can apply an

algorithm an unsupervised learning

algorithm that identifies like uncommon

transactions or weird patterns and that

can help us to finally identify maybe a

fraudible in transaction but that's the

main difference between them

so well now that we know this we can

start with our first application in this

case is for

is for the section of anti-money

laundering that we have in our platform

and the application is the use of

machine learning techniques for

segmentation of our clients and so well

remember what I told you at the

beginning we always need to have like

this business call in mind and that's

our pre our main goal

so which which our objective to find

groups

clients that are similar between them

and different from others so that's the

main goal

and what's the purpose of doing this

well this will help us to understand the

customers that we are dealing with this

also can be applied for example to

suppliers or products and this will help

us to discover patterns that we didn't

know and also it's a really good first

step towards the transactional

monitoring

um and also in some companies it's an

obligation to perform a segmentation so

we we can comply with regulations

um so well that that's the goal the the

objective is to identify groups of

clients that are similar between them

and

when I say similar I mean that they are

close to to each other because for

example if we plot our data set in a

graph uh like you can see here in the

right

um

we can see that there are some

observations that are closer to each

other and are like more far away from

from others so that's the way that these

algorithms Works they calculate like

distance between observations and in

that way they identify like patterns or

groups so so we need to to know that

um and well another important thing is

to to know like the advantages that we

can obtain

by applying machine learning techniques

um well obviously if if we are working

for example a with a data set of 100

clients what I would told you is to okay

just to a manual segmentation because

never forget the goal is to segment our

clients so if you have a data set with

100 clients I consider that you it will

be more easy to perform a manual

segmentation

but obviously if you are dealing with

for example a huge amount of data for

example with 2000 clients but in that

case yes you can

um like have more advantages by applying

a machine learning technique

in that case I consider that you can

achieve a higher level of accuracy in

those cases another advantages are for

example automation scalability

adaptability for example in our platform

you will be able to perform all the

steps like really easy and for example

another case if you experience and

and incrementation for example you go

from 2 000 clients and now you have like

10 000 well it's the same you can apply

the same algorithms and it will be good

you are not going to suffer like

of time efficiency and like if you have

to do it manually another good thing is

that you can apply these algorithms to

segment clients or suppliers or products

and another good thing is the

personalization

um

for example you can decide if you're you

are going to use

um for example different characteristics

in their segmentation you can choose if

you can consider for example the country

of your clients or you can you want to

consider the income of your client or

not and you can perform and apply these

algorithms and compare results and and

obviously well the time efficiency of

applying these techniques it's a lot

so these are some of the advantages

um and well now I'm going to talk about

um the the classroom algorithm the

clustering algorithms that we have in in

the platform one of them it's called

k-means

and this algorithm

tries to partition

our data set into K different groups

so later we are going to talk about how

to identify the optimal number segments

I mean the the K number of segments

but the idea of of this algorithm is to

to do that to partition our data set

into K different groups and the way it

performs is to First

identify like

different points randomly in the in our

in our data set and then it starts like

every performing calculations to well

identify which observations are closer

to that that initial point that we are

going to call

a the cluster Center or the centroid of

the cluster and this process it's it's

going to be repeating like a like a

hundred of times or like many times

until there are no more movements and we

can just say okay these are the the

final segment

and what thing about what's good about

this is that it's really simple

uh and it's like really efficient uh

also maybe it's like kind of sensitive

to initialization and maybe it's

sensitive also to Appliance but well we

are we're going to explain a way to

to remove these outliers before

performing the algorithm but well that's

the idea behind okay means

and the other one it's called Birch and

this algorithm is better for more larger

data sets

but this algorithm does is to First

identify like observations that are like

really similar to to each other

um and confirm like a summary of the

data set a summary using like this like

uh aggregating like observations that

are like really similar and then using

like this summary of the data set and

he decides like which are going to be

the the final segments of or clusters uh

well this algorithm is like it's more

robust to outliers for example and it's

better for

for high dimensional data

um for larger data sets so in those

cases we should use this but well what I

recommend is to let's use both and

compare results that's really important

and decide which result is better for

you and for your your use case of

business

um okay uh well another important step

in this process is to perform a good

data cleaning

and because this is like the most

important step in the whole process you

can use for example the most advanced

algorithms and the most advanced and

machine learning techniques but if you

didn't perform a good data cleaning you

are not going to to obtain a good result

anyway

so well in the platform you are going to

see that we have like different

uh four steps and you can perform this

directly in the platform and the first

one is to is an outlier detection

and why this is important

because for example we have our data set

of clients

um we identify that there is a client

which income is like super different

from the others so this is going to

uh this is going to generate a

distortion like in in the distance

between the observations like you can

see in this graph in the right for

example the presence of an outlier

it's generating a distortion that it

seems that the the whole other group of

observations belongs to the same cluster

but this isn't true because it's just a

distortion generated by by the presence

of the supplier so

it's important to detect them and to

remove them from the process but never

to eliminate them we should confirm a

different segment of the suppliers and

we we should analyze them like in a

separate way

and another important step is the

categorical variable encoding and in

this case what we are trying to do is

for example if we have a a column that

with the country of each client

so we have for example that one client

belongs to for example Spain and the

other one to to France uh well but we

can't measure like distance between

these words so what we are trying to do

is to express this categorical variable

into numerical ones

um then another step is data

normalization

and why this is important because for

example we have one column with the H of

the clients that assumes values from 20

to 80 for example and we have another

column that

with the income of the of the client so

well in that case the income for example

can assume assume values from one

thousand to five thousand for example so

it would be like really difficult for

the algorithm to calculate distance

between numbers like 20 and numbers like

and five thousands so what we are doing

is to express like all numerical data in

in a in a scale that makes them

comparable between them

and finally another important step is

the principal component analysis

um this this is also an algorithm itself

that identifies which of all the

characteristics from our clients are the

most important and the more the most

relevant

for the analysis

so

for example a really simple example

maybe we have this country column and

maybe all of our clients belongs to

Spain for example so well the algorithm

is going to tell us look all your

clients are from the same country you

are not going to be able to segment

clients using this variable because they

are all from the same country so you

should not include this in the process

and what in that case we are going to

exclude this column

and that will be also good for the

process because we are going to reduce

the dimensionality of the whole data set

and so what these are the like the four

steps that we are using

um

well as I told you another important

thing is to choose the optimal number of

segments

in this case we are going to use like

two methods two different methods we

always try to offer like different

Alternatives as we were talking about

the algorithm kamins and built to apply

both of them and then compare results

well in this case to choose the optimal

number of segments we are going to use

two methods

one of them it's called the elbow method

and this method use a metric that is

called inertia

and this metric

it's the average of the distance from

each point to its nearest Center it it

nearest the center of the cluster so

what it's trying to do is to calculate

the distance of the distance of a client

to the center of the client of the

cluster that this client belongs to

um well so what what happened here that

if we have for example 100 clients the

the lowest inertia that we can achieve

is if we have like 100 clusters because

each client is going to be a cluster

itself

but well we need to find like an optimal

point so the optimal point

will be when the inertia is starting to

to slow down

um and it well we need to choose like

like a number because we need to achieve

like the final goal of the segmentation

that is to to segment and create like

group of clients and we can just use

like 100 clusters and have like every

client is separate in a different

cluster so we are going to find this

this optimal point

and um another

um method that we have is the silhouette

metric and this is like really common to

use and this metric

um use like two different parameters one

of them is called cohesion

and they are one it's called separation

um decoration parameter

tries to

calculate like the distance or how like

like how

close the observations from a cluster

are

so for example in this in this picture

it tries to to calculate the distance

from all the observations from the same

cluster from the cluster that is in red

color and the

the separation parameter

is trying to calculate the distance from

the observations of the cluster bread

to the cluster blue so we are trying to

achieve like the optimal situation is

when for example like this like the

picture number one where the

observations are like really similar the

observations from from one cluster are

like really similar between them and are

also really far away or different from

the other clusters

and so what this is the the way we can

like choose the optimal number

um well now we are going to show you

the platform so I'm going to show you

all these steps and how you can perform

them there

okay this is the platform here you are

going to find like different sections

for example we have a section for

compliance another one for operational

risk another one for informational

information security risk and another

one for a anti-money laundering this is

what we we were talking about

so here in this section

um sorry my my screenshots

gold goes off

and just just I mean

okay so here

we can create and perform all the steps

from the segmentation we will be able to

create a new one I'm going to show you a

signification uh segmentation I have

already created and so so it will be

more quickly

but here is a news case for for clients

segmentation

um

sorry I left in the link the link of

this pirani tool in the chat so you can

access to our software and you can do

the Practical exercise with a condo with

this link you will be able to access a

15-day free trial so you can test the

fifth the five system that we have in

kirani

okay perfect perfect

well in this section you will be able to

perform like all the steps uh we are

going to start by choosing the variables

the variables that we are going to that

we want to include in our segmentation

process

what is important here is to First

calculate this the that is the

population of the variable and this is a

metric that indicates how many not new

values or not empty values we have in

that variable

and so well a recommendation is to

always use like those characteristics

that are over 80 percent for example

if we have like this this type of

characteristics that are empty or are

like this like lower than 80 percent

well you we should like exclude them and

then the the second step is to perform

the data cleaning that we were talking

about here you will be able to perform

like all the four steps

transform like categorical data to

numerical

then we are going to detect the outliers

in our data and also an important thing

as I told you we are not going to remove

or eliminate the suppliers we are just

removing them from the segmentation

process but we are going to like store

them in a different file in a different

CSV file so you can analyze these

suppliers like separately and decide if

it was just a mistake in the data or

maybe it's a really important client

that you should analyze and separately

another important thing it's the well

the data normalization that we were

talking about and finally the principal

component analysis to identify the the

most important characteristics to

include in our model

and then when we finish this this part

we we can choose like the model the the

we prefer as I told you I think that the

best thing is to try two different

segmentations one using comments and one

and another one using the Birch

algorithm

and then just analyze the results and

compare compare them to well to decide

which one is better for your use case

um okay then after we we choose the the

model we are going to calculate like the

optimal number of segments

here for example I have already

calculated so using the album method uh

well it is identifying that the optimal

number should be six and but well

obviously we should try with this method

and also with the silhouette method and

and then finally when we have done all

the steps we are going to execute the

the segmentation

so when this process is conclude

um we are going to receive like a PDF a

file

here I can show you with all the

a all important characteristics and

descriptive statistics of your segments

and you are going to here you can see

like the amounts of data that you have

the number of segments uh you can see

for example some characteristics from

your numerical variables you are going

to see for example the characteristics

from the cluster one from cluster 2 and

it's important that you start here

analyzing all the results

uh well yeah try to find a business

perspective on these results and try to

well

achieve the goal that is to understand

more your your clients

and well this is how you can perform it

in the platform

now we are going to go back to the

presentation

I don't know if you have any questions

but now we can continue with this the

final part that is the

the natural language processing

techniques and the chat GPT application

um well here well it's important

is to to know what what it means this

this Superior that it's also a soup area

in artificial intelligence and the main

purpose of this is to combine the

computer science

the computer science

applications with Linguistics to obtain

this new file that is called natural

language processing and the main

objective here it's not only that we are

able to process a text Data because

before we were talking about only

numerical data but here we are trying to

process like text but on not only that

on or also we try to that this model can

understand what it means like this text

and also will be able to generate like

new human language

so that's the the main goal and some of

the the examples of these applications

are for sentiment analysis most

companies use like these techniques to

analyze sentiment in in the tweets for

example when someone writes a tweet in

about their company and try we are

trying to understand if the sentiment

behind this tweet is like positive or

negative

um also this is really useful for

comments or service in in our

in our company's platform to well to

understand if we are offering a good

service and if the sentiment of these

comments are like positive or negative

um well another common

application is to text classification

for for example an application into risk

management it would be to classify risk

we can like read all these risks that

people write and classify them into

different categories to well to

understand which category is like more

dangerous or it's more probable that

that it happens for example

and another application that it's really

common is to the chatbot and this is

what this is really common to ask

questions to this chatbot and receive an

answer

um

um well now

in in our platform we decided to to

apply the the model John CPT

but it's important to understand what it

means this maybe a lot of people don't

know what it means like the the letters

the G the B and the T the C refers to

generative

and this means that this model is able

to create or generate like new data

points

remember that in this case we are

talking about text their data so this

model is going to be able to generate

like new paragraph or new text

another important thing is the p and it

refers to pre-train uh well this is

because it was pre-trained using like a

huge amount of text Data available in

the internet

and another thing is the the tea that

refers to a Transformers and well this

is a like the Deep learning or neural

network architecture that works from

behind the model uh well you you need to

know this in detail but an important

thing about this is the

attention mechanism mechanism that it

has and this allows the model to for

example process like a big paragraph of

text and identify in in that paragraph

like the keywords or the context or the

of the whole paragraph and to retain

that information and well this is like

it was like really revolutionary in that

moment

um well there are some applications of

chat GPT into risk management

um

for example one of them is what I'm

going to show you in the platform and is

a recommendation of potential risk

because well we identified that a lot of

like new clients access to the platform

and they couldn't like identify the risk

that that they were suffering or maybe

they know like the risk that they were

suffering but they couldn't like write

it properly in the platform so they were

like starting like calling our

consultants in race and what we decided

to create like this system to recommend

potential risk

um and this system

is going to be using like different K

variables about the company of the

client for example one of the the K

variables are the industry of the client

uh the process in the company and the

risk system that this refers to for

example if they are talking about

compliance or operational risk or

anti-money laundering risks and and of

course the language of the client

um well the process is like this picture

that we have in the lab all these K

variables are an input

into the prompt of the chat CPT model

and then the well GPT use like all the

information all these text data that it

was used to pre-train this model from

the internet

um so chat GPT used all this information

to generate a response and to make some

recommendations of potential risks to

decline

um but also as I told you this is one

application but we also have like like

other ones for example risk

classification we can use like this

small to classify like the risk into

different categories

and and another one that we are working

on is to make suggestions of mitigations

actions and possible causes of the risk

and so the idea is not only to recommend

potential risk but also to recommend

actions to mitigate that risk and maybe

possible causes so that's what they're

working on and so now I'm going to show

you

in the platform where you can find this

um

okay so here

we can go for example to another section

for example we can go to the operational

risk section

um

okay so here we are going to have the

module of risk

where we can create like new ones

a and for example before this

application like people start like

writing here please

something but now we are trying to offer

like this suggestion risk using chat gbt

so you can get some ideas

and and here you are going to select for

example the industry that you are

working on and maybe for example the

hull industry

and you can also select the process and

for example we are going to choose this

technology process

uh let's just press suggest and subject

is going to start working using all the

information

of the internet and well here are the

suggestions you you can read them uh

well maybe this can be helpful to

identify to to use these suggestions or

to identify potential risk

um if you want to use one of these

recommendations you just click the

um and and that's all and you can create

in this way a new risk

also you will find like suggestions for

a lot of Industries like the financial

the insurance

retail and also a lot of process into

the company from the company for example

a marketing commercial customer service

it's a lot to

that you can use and well that's that's

all of this

of this presentation that's all what I

have to to show you I don't know if

there are some questions

we have one question

keeps information we provide

um well no in this case we are not like

giving the information that we have for

example from our clients to to chat gbd

we are just using like this uh K

variables for example the the industry

and the process and that's all we are

using like the information that is from

behind of the model of charge GPT to

help us but we are not giving life

information actually

um well we are doing this because we

don't know like which which are the

the security behind this and

if data is keeping private or not so

so that's what what we are doing

great

I think we know how

to have questions

so we can do a brief summary about your

presentation

to our participant to

okay okay yeah so well to summary uh we

can say that it's important to don't

lose the the focus or the goal in

business we need to First understand

like the business problem or the

business opportunity that we want to

solve and then start thinking if if we

need to use machine learning or

artificial intelligence for that if if

the problem is like really complex so

let's do it

um but if not just keep it simple and

try to always focus on solving the

problem

and then another important thing is if

you are performing a segmentation to

well to do like all the steps and give

the importance to the data cleaning part

because that's the most important part

as I told you you can use like all the

algorithms that you want but if you

didn't perform a good data cleaning you

are not going to get a good result

and then obviously always compare

results like using both different

methodologies but at the end you need to

compare and try to choose which one is

better for for your company

um well that's it and if you if you need

some recommendations of risk just use

our section of gbd to to get them

thank you so much for your value

participation in swearing us finally I

would love to invite everyone to

participate in the short survey that

will help you when this meeting ends

that will help us a lot to continue

improving our webinars thank you so much

for your participate to for your

participation and see you soon thank you

bye

[Music]