Table of Contents

Table of Contents

• Learn about Machine Learning and AI
• The Future of Artificial Intelligence
• AI in Root Cause Analysis
• AI in Human Resources
• AI in Making Recommendations
• AI in Predictive Maintenance
• AI in Bio-metrics Recognition
• AI in Speech Recognition
• AI in Detecting Diseases
• AI in Engineering
• AI in Finance process improvement
• AI in Healthcare process improvement
• AI IN BUSINESS PROCESS MODELING/IMPROVEMENT

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence. @book{Somogyi_2021, doi = {10.1007/978-3-030-60032-7}, url = {https://doi.org/10.1007%2F978-3-030-60032-7}, year = 2021, publisher = {Springer International Publishing}, author = {Zolt{\'{a}}n Somogyi}, title = {The Application of Artificial Intelligence} }

Learn About Machine Learning and AI

Learn About Machine Learning and AI

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

It is not always clear to people, especially if they are new to the subject, what we mean by machine learning and when and why we need it. A lot of people are aware of artificial intelligence (AI) from science fiction but they may not really understand the reality and the connection to machine learning. This article will explain in clear lay terms what machine learning and AI are, and it will also introduce the three major forms of machine learning: supervised, unsupervised and reinforcement learning. The aim is that after reading this article you will understand what, exactly, machine learning is and why we need it.

Machine learning is a process in which computers learn and improve in a specific task by using input data and some kind of rules provided to them. Special algorithms, based on mathematical optimization and computational statistics, are combined together in a complex system to make this possible. Artificial intelligence is the combination of several machine learning algorithms which learn and improve in several connected or independent tasks at the same time. At present, we are able to develop parts of a real artificial intelligence but we cannot yet combine these parts to form a general artificial intelligence which could replace humans entirely.

We could also say that learning in this context is the process of converting past experience, represented by the input data, into knowledge.

There are several important questions that arise: To which kind of tasks should we apply machine learning? What is the necessary input data? How can the learning be automated? How can we evaluate the success of the learning? Why don’t we just directly program the computer with this knowledge instead of providing the input data?

Let us start with answering the last question first. There are three main reasons why we need machine learning instead of just using computer programming:

1. After a computer program is made it is difficult to change it every time the task changes. Machine learning adapts automatically to changes in the input data/task. As an example after software has been programmed to filter out spam e-mails, it cannot handle new types of spam without re-programming. A machine learning system will adapt automatically to the new spam e-mails.
2. If the input is too complex, e.g. with unknown patterns and/or too many data points it is not possible to write a computer program to handle the task.
3. Learning without programming may often be very useful.

Besides the above it is of course also a human desire to try to make an artificial intelligence, towards what we are evolving.

In order to be able to answer the other questions let us first look at a typical machine learning process.

First we need to decide which task to teach to a machine learning model considering the three reasons mentioned above. Next we need to decide which data and rules we need to feed to our machine learning model. Then we need to choose a machine learning model, train the model (this is when the learning takes place) and test the model to see if the learning is correct. Collecting the data, choosing the model, training and testing are all recursive tasks (note the arrows going back to former steps) because if the model cannot be adequately trained then we often need to change the input data, add more data or choose another machine learning model.

Machine Learning tasks can be classified into three main categories:

1. Supervised Learning
2. Unsupervised Learning
3. Reinforcement Learning

We speak about supervised learning when the input to the machine learning model contains extra knowledge (supervision) about the task modeled in the form of a kind of label (identification). For example in the case of an e-mail spam filter the extra knowledge could be labeling whether each e-mail is spam or not. The machine learning algorithm then receives a collection of e-mails labeled spam or not spam and through this we supervise the learning algorithm. Or in the case of a machine learning based speech recognition system the label is a sequence of words (transcribed sentences). Or another example could be the labeling of a collection of images about animals for an animal identification task. With the extra knowledge of which picture contains which animal the learning algorithm is supervised.

There are already many real-world supervised learning applications and many more will be added in the future. Some of the existing applications are as follows:

• E-mail spam detection based on a collection of messages labeled spam and not-spam.
• Voice recognition based on a collection of labeled voice recordings. The labels identify the person who speaks.
• Speech recognition (part of comprehension) based on a collection of labeled voice recordings where the labels are the transcription of sentences.
• Automatic image classification based on a collection of labeled images.
• Face recognition based on a collection of labeled photos. The labels identify which photo belongs to which person.
• Determining whether a patient has a disease or not based on a collection of personal data (temperature, blood pressure, blood composition, x-ray photo, etc.).
• Predicting whether a machine (auto, airplane, manufacturing, etc.) will break down (and when it will break down – for predictive maintenance) based on a collection of labeled data from past experience.

Remember that we speak about supervised learning when the input to the machine learning model contains extra knowledge (supervision) about the task modeled in the form of a kind of label. When we do not have this extra knowledge or label then we speak about unsupervised learning. The aim of unsupervised learning is the identification of this extra knowledge or label. In other words, the goal of unsupervised learning is to find hidden patterns in the data and classify or label unlabeled data and use this to group similar items (similar properties and/or features) together, and thus put dissimilar items into different groups. Another name for unsupervised learning is clustering (grouping). 

There are already many real-world unsupervised learning applications and many more may be added in the future. Some of the existing applications are as follows:

• Grouping shoppers together based on past purchases and other personal properties; for example, as part of a recommendation system.
• Market segmentation based on chosen properties, e.g., for marketing applications.
• Segmentation of a social network or a group of people, e.g., for connecting people together (as on a dating site).
• Detecting fraud or abuse (by applying unsupervised learning to better understand complex patterns in the data).
• Grouping songs together based on different properties of the music, e.g., on streaming platforms.
• Grouping news articles together depending on the contents or keywords, e.g., as part of a news recommendation application.

We could define reinforcement learning as a general purpose decision making machine learning framework used for learning to control a system. There are several important keywords in this definition which need some explanation. General purpose means that reinforcement learning can be applied to an unlimited number of different fields and problems; from very complex problems such as driving an autonomous vehicle to less complex problems such as business process automation, logistics, etc. Decision making means carrying out any kind of decision/action depending on the specific problem, for example, accelerating a car, taking a step forward, initiating an action, buying stocks, etc. Controlling a system means taking actions in order to reach a specific goal, where the specific goal depends on the problem (e.g., reaching a destination, having profit, being in balance, etc.).

There are currently many real-world reinforcement learning applications and no doubt more will be developed in the future. Some of the existing applications are as follows:

• Self-driving cars. A control system based on reinforcement learning is used to adjust acceleration, braking and steering.
• Automated financial trading. The reward is based on the profit or loss for each trade. The reinforcement learning Environment is built using historical stock prices.
• Recommendation systems. The reward is given when, for example, the users click on an item. Real-time learning improves the machine learning model or recommendation systems are trained on historical data.
• Traffic light control. 
• Logistics and supply chain optimization.
• Control and industrial applications, e.g., for optimizing energy consumption, efficient equipment tuning, etc.
• Optimizing treatment policies or medication dosage in healthcare.
• Advertising optimization.
• Various types of automation.
• Robotics.
• Automated game play.

This article is a slightly modified excerpt from the book “The Application of Artificial Intelligence”. If you are interested in the subject then it is strongly recommended to read the book which contains much more details and real world case studies for several sectors and disciplines! The book explains several examples by using the AI-TOOLKIT. The book is going through the publishing process at the time of writing this article. You may use the contact form for info about pre-ordering the book.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence. @book{Somogyi_2021, doi = {10.1007/978-3-030-60032-7}, url = {https://doi.org/10.1007%2F978-3-030-60032-7}, year = 2021, publisher = {Springer International Publishing}, author = {Zolt{\'{a}}n Somogyi}, title = {The Application of Artificial Intelligence} }

AI in Root Cause Analysis

AI in Root Cause Analysis

Are you still using techniques like 5 Whys, fish-bone diagram or even guessing for root cause analysis? Would you like to have a root cause instantly with high accuracy? If yes, then read further!

Detecting anomalies, and finding the root cause of the anomaly, is an important application in the field of machine learning in nearly all sectors and disciplines. An anomaly may mean different things in different applications, for example, fraudulent use of credit cards or suspicious transactions in the financial sector, a specific disease or the outbreak of a disease in healthcare, the signs of intrusion in a computer network, a fault in a production system or product in the manufacturing industry, an error in a business process, etc.

Machine learning discovers patterns in the data and therefore it is well suited for discovering unusual patterns which are the signatures of anomalies.

Anomaly detection can be combined with root cause analysis (RCA) in a machine learning (ML) model, or ML anomaly detection may assist traditional RCA techniques in finding the root cause. Machine learning automated RCA may also be useful for the following reasons:

• To perform complex RCA several domain experts are often needed, who may not always be available or may be expensive or difficult to deploy. In this case ML automated RCA can also be considered as a knowledge management tool.
• ML automated RCA may save a lot of time because traditional RCA projects may often take even several days.

There are three main types of machine learning models (methods) which can be used for anomaly detection and/or RCA:

• Supervised,
• Unsupervised and
• Semi-supervised learning based method.

For supervised learning we need labeled (classified) data. In the case of anomaly detection, the data must contain data records labeled as ‘normal’ and data records labeled as ‘anomaly’. We can combine anomaly detection with root cause analysis by defining several types of anomaly labels which all identify a specific root cause. For example, in the case of the root cause analysis of a defective product in a production process, we could define root causes as “wrong material”, “handling error”, “machine error”, etc. instead of just using “wrong product”.

Supervised machine learning based anomaly detection and root cause analysis is a very powerful technique, but collecting and labeling the data is a lot of work and must be done by domain experts. You may also consider using a semi-supervised method (see later) in order to decrease the amount of work needed for labeling the data.

One of the most important parts of anomaly detection and root cause analysis is the data collection and feature selection phase. It is very important to select the right features which can be used to distinguish between normal and anomalous phenomena. In most cases (especially if root cause analysis is integrated) domain experts must design the data collection and feature selection process, and they also must take care of the labeling of the data records if supervised learning models are used.

Usually a traditional structured and systematic approach is used to investigate anomalies and their root causes (when appropriate) and in order to determine the best set of features used in the input data for the machine learning model.

It is often possible to simulate a real situation or environment and introduce all forms of potential anomalies and record the attributes (responses) of the system. It is much easier and faster to study the signatures of anomalies in this way than to monitor real systems. Try to think about a solution for modeling your real situation in some way and try to simulate potential anomalies.

This article is a slightly modified excerpt from the book “The Application of Artificial Intelligence”. If you are interested in the subject then it is strongly recommended to read the book which contains many more details and real world case studies for several sectors and disciplines! The book explains several examples step-by-step by using the AI-TOOLKIT. The book is going through the publishing process at the time of writing this article. You may use the contact form for information about pre-ordering the book.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

AI in Human Resources

AI in Human Resources

In this article, for beginners in machine learning, you will learn how to use the machine learning (ML) tools in the AI-TOOLKIT to make difficult HR decisions automatically. In this simple example we will train an ML model which can be used to predict if an employee will leave the company. We could use the same principles to predict the reason of leave or if it is worthwhile to offer a promotion to an employee. The article will also explain and compare some of the ML models available in the AI-TOOLKIT.

You can apply the same principles to any other sector or business case, for example, you could predict if a client will leave, why it will leave, or if it is worthwhile to offer a discount, etc.

The Dataset

The dataset contains 15,000 rows (records) and 10 columns (variables or features). You can download the data at the end of this article. If you are doing this in your company you should first study which variables influence the most the specific business case, in this example an HR problem, and select the variables accordingly. We call this step Feature Engineering. Selecting not enough or too many variables or features (not enough knowledge or unneeded noise) will result in a less useful or less accurate ML model. The accuracy of the trained ML model depends mainly on the input data (quantity and quality) and also on the parameters of the models.

The 10 columns (features) are as follows:

• Satisfaction Level (0-1)
• Last evaluation (0-1)
• Number of projects (integer)
• Average monthly hours (integer)
• Time spent at the company (integer)
• Whether they have had a work accident (0-no, 1-yes)
• Whether they have had a promotion in the last 5 years (0-no, 1-yes)
• Department name (text)
• Salary (text: low, medium, high)
• Whether the employee has left (0-no, 1-yes)

Depending on which variable (column) you choose as decision variable you can train a model for different purposes, for example, to predict whether the employee will leave in the future, whether it is worthwhile to offer a promotion, etc.

In this example we will choose the ‘Left’ (whether the employee has left) column as decision variable in order to predict if an employee will leave or not.

satisfaction  level	last evaluation	number  of projects	average  monthly hours	time  spend company	work  accident	left	promotion  last 5 years	sales	salary
0.38	0.53	2	157	3	0	1	0	7	1
0.8	0.86	5	262	6	0	1	0	7	2
0.11	0.88	7	272	4	0	1	0	7	2
0.72	0.87	5	223	5	0	1	0	7	1
...	...	...	...	...	...	...	...	...	...

Training the AI Model

There are different types of machine learning models available in the AI-TOOLKIT. Each model/algorithm has its advantages and disadvantages. Some algorithms are well suited for one type of data but not for another type of data. Neural network based models can be tuned so that they can be applied to all kinds of problems, but with the cost of complexity (often with many layers of different type and with many nodes) and processing speed (more layers and nodes mean more processing time and more computer resources). Furthermore, neural networks also need much more data than other types of machine learning models. Therefore, it is worthwhile to choose the machine learning model you want to use in a clever way!

Let us choose the SVM model for this example.

Support Vector Machine (SVM) model

You can easily import your numerical delimited data into the AI-TOOLKIT. The SVM model has several parameters, which can be automatically optimized by the built-in parameter optimization module.

Follow the next steps in order to train the ML model:

1. Create a new AI-TOOLKIT project (Open AI-TOOLKIT Editor + New Project).
2. Insert the SVM model template (Insert ML Template + choose Supervised Learning + Support Vector Machine).
3. Save the project.
4. Download the data (at the end of the article) and change the extension to ‘.tsv’. Import the data into a new AI-TOOLKIT database (On the DATABASE tab: Import Data Into Database + follow the instructions on the screen. It is important that you indicate the correct number of header rows (non-numerical) and the zero based index of the decision column (6 in this example)). Use as table name: ‘hr_data’.
5. Save the database into the same folder as the project is saved. Use the name ‘hr.sl3’.
6. Run the SVM parameter optimization module to find the optimal parameters (SVM Parameter Optimizer on the AI-TOOLKIT tab). You may stop the optimization earlier if you see a high enough accuracy or just skip the optimization and use the values shown below.
7. Adjust the SVM model template as shown below (some of the unneeded parameters and comments are not shown). The optimal parameters are filled in.

model:
    id: 'ID-EFnMmvBNWr'
    type: SVM
    path: 'hr.sl3'
    params:
        - svm_type: C_SVC 
        - kernel_type: RBF 
        - gamma: 15.0 
        - C: 281.8 
    training: 
        - data_id: 'hr_data' 
        - dec_id: 'decision' 
    test: 
        - data_id: 'hr_data' 
        - dec_id: 'decision'
    input: 
        - data_id: 'input_data' 
        - dec_id: 'decision'
    output:
        - data_id: 'output_data'
        - col_id: 'decision'

8. Save the project.
9. Train AI model (AI-TOOLKIT tab).

After the training is ready you will see the performance evaluation results:

Performance Evaluation Results
Confusion Matrix [predicted x original] (number of classes: 2):
	(0)	(1)
(0)	11427	0
(1)	1	3571
Accuracy	99.99%
Error	0.01%
C.Kappa	99.98%
	(0)	(1)
Precision	100.00%	99.97%
Recall	99.99%	100.00%
FNR	0.01%	0.00%
F1	100.00%	99.99%
TNR	100.00%	99.99%
FPR	0.00%	0.01%

The accuracy of the trained model is very good (nearly 100%). The trained model only makes one mistake in 15,000 cases. In this example we will not go more in detail about all performance measures and discuss the so called generalization error (testing with unknown data) because this is not the aim of this simple example.

DeepAI Educational Neural Network Model

The deep neural network model in DeepAI Educational is based on a semi-automatic multi-layer and multi-node neural network implementation. The software designs the neural network semi-automatically, you only need to define the number of layers and nodes per layer (you can of course adjust some more parameters but this is most of the time not necessary). DeepAI Educational does not use complex state of the art neural network architectures and extensive model performance evaluation, but it often provides a good result. For real world problems use the machine learning models and tools in AI-TOOLKIT Professional.

DeepAI uses the SSV data file format (delimited text file). Adjust the settings in the ‘Settings/AI’ tab according to the following if needed:

• Number of iterations: 10
• Learning rate: 0.01
• Regularization rate: 0.001
• Batch size: 10
• Activation Function: TANH
• Regularization Function: NONE
• Test data %: 10
• Treat data as X-Y Classification / Regression

Download the data (at the end of the article). You can load an external training data file with the 'Load Data File (SSV)' command. The data must have the AI-TOOLKIT SSV data file format (.ssv), which is tab delimited, without a header row, contains only numbers and with the decision variable (classes in case of classification, continuous numbers in case of regression) in the first column. 

Since the decision variable must be in the first column for DeepAI we need to open the data file in MS Excel and move the decision column (‘Left’) to the first column. We also must remove the first header row! After you are ready save the file in tab delimited format and with ‘.ssv’ extension.

Use the ‘Load Data File’ command and load the above prepared data file. DeepAI will automatically design a neural network for the data file. This neural network will provide good results but let us add an extra layer (4 layers in total), adjust the number of nodes to 24 on the second layer and 10 on the third layer. The first and the last layers have a fixed number of nodes depending on the input functions and the output (1).

Change the number of iterations to 480 and start the training process with the Run command. After a while (2-3 min) the results will appear which indicate 98.3 % accuracy for the training data. You can still fine tune the model and obtain a higher accuracy but this is not a fast and simple process. Fine tuning a neural network is a tedious and often long lasting process (adjusting the number of layers, adjusting the number of nodes per layer, adjusting the learning rate, the activation function, etc.). It is also not sure that more layers and nodes will provide better results but you will need to find the optimal solution also depending on the other parameters.

You can use the trained ML model for making automatic and precise decisions about this HR problem.

References

• The Application of Artificial Intelligence, Zoltan Somogyi.
• HR Analytics Dataset: Attribution-Share Alike 4.0 International (CC BY-SA 4.0) license, Source: https://www.kaggle.com/ludobenistant/hr-analytics.
  You can download the dataset in MS Excel format here: HR_COMMA_SEP_U.XLS

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

AI in Predictive Maintenance

AI in Predictive Maintenance

In many industries the reliability of machines is very important. In aerospace, transportation, manufacturing, utilities, etc. complex machines containing many components undergo periodic inspection and repair (preventive maintenance). The main challenge is to schedule preventive maintenance and component replacement in an optimal way such that the machines can work reliably and the components are not replaced too early. Reliability, high asset utilization and operational cost reduction are, in short, the aims of each company in these industries.

By using machine learning and historical data we can train a model which can predict when the next failure will occur and thus when preventive maintenance should be scheduled. We call these kinds of machine learning models predictive maintenance machine learning (PMML) models. There are two main types of PMML models:

• Regression models predict the remaining useful lifetime (RUL) of the machine or components.
• Classification models predict the failure within a pre-defined time period (time window).

In order to build a useful PMML model we need to go through some important steps which are summarized hereunder:

• Data collection
• Feature Engineering
• Data labeling
• Defining the training and test datasets
• Handling imbalance in the data

The input data may come from different sources and usually contains failure history, maintenance history, machine operating conditions and usage, machine properties and operator properties.

After we have collected all necessary data we must combine them into one synchronized dataset which can be fed into the machine learning model. We call this step feature engineering because we are building a dataset from features fabricated from the collected data. This is often a complex process in the case of a predictive maintenance model and the performance of the model will entirely depend on it.

The method for combining the collected data into the final dataset is usually very similar but of course business case and data dependent. Remember that the aim is to predict when the next failure of the machine will occur by using historical data.

There are two types of data, time series and static data. Static data can usually be simply combined with the other data by grouping them per machine ID. For example, if the maintenance history is defined with “time | machine ID | component”, and the machine properties are defined with “machine ID | property 1 | property 2…”, then we can simply add the static machine properties per machine ID as follows: “time | machine ID | component | property 1 | property 2…”

In the case of times series data we need to aggregate the data according to some pre-defined rules based on the business case.

We usually want to predict machine failures in a future time period (time window) based on a historical time period. The data may be collected with a frequency of seconds, minutes, hours, etc. and we need to aggregate it into a pre-defined time period based on the business case. The evolution of the features in the time window is captured by the aggregated values. The machine learning model will learn which aggregated values result in a failure in the next time window. For example, if we want to predict whether a machine will fail in the next 24 hour period, then we can use a time window of 24 hours and label the aggregated records which fall into a 24 hour window just before a failure occurs as FAILURE and all other records as NORMAL. It is of course business case dependent as to how long the time window should be. Sometimes 24 hours is appropriate but sometimes we need to use a longer period, for example, to allow for a longer period of supply of repair parts. If it takes one week to get repair parts then we need to predict a failure much earlier in time.

This article is a slightly modified excerpt from the book “The Application of Artificial Intelligence”. If you are interested in the subject then it is strongly recommended to read the book which contains many more details and real world case studies for several sectors and disciplines! The book explains several examples step-by-step by using the AI-TOOLKIT. The book is going through the publishing process at the time of writing this article. You may use the contact form for info about pre-ordering the book.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

AI in Making Recommendations

AI in Making Recommendations

Making a recommendation means that we recommend one or more items to potential users. The recommended items may be many different things, for example, physical products which are being sold (e.g., cars, smart phones, etc.) or articles, web pages, documents, etc. There are many recommender systems in use today, for example, for recommending books, movies, clothing, holiday destinations, etc., and usually they help to increase revenue and/or help users to find the most relevant, interesting and/or important information or product.

Recommendation machine learning models work with explicit and/or implicit feedback data collected from users while they are interacting with items (products, documents, etc.). 

Explicit feedback is when the user provides some kind of rating or like/dislike of the items he/she is interacting with. There are many types of rating scales, for example, the five-star rating in which one star means low appreciation and five stars mean a very highly appreciated product. All of these ratings can be expressed on a numerical scale (e.g., 1,2,3,4 and 5; or 0 and 1 for dislike and like).

Implicit feedback is when the user does not directly provide some kind of rating but we collect information about user actions, for example, buying a product, viewing a document or web page, etc.

The basic principle behind how recommender machine learning models work is that correlation exists between how different users appreciate similar items, how different items are appreciated by similar users and the combination of the two (joint correlation). The user’s appreciation is expressed with explicit or implicit feedback. These correlations or behaviors can be learned by a machine learning model based on the collected explicit and/or implicit feedback data from the users (user + item + feedback).

There are two main types of state-of-the-art machine learning recommendation models:

• Collaborative filtering (CF),
• Content-based (CB).

Collaborative filtering models may use explicit and/or implicit feedback data in the form of a triplet consisting of a user ID, an item ID and the feedback value. These triplets form a three dimensional space which can be represented in a matrix or table.

Collaborative filtering models use the collected feedback data from all users (collaborative), but it usually does not use content information (description of the items). When the predicted explicit (ratings) or implicit (user actions) feedback is known, a top-k number of recommendations can be made to any user.

This article is a slightly modified excerpt from the book “The Application of Artificial Intelligence”. If you are interested in the subject then it is strongly recommended to read the book which contains many more details and real world case studies for several sectors and disciplines! The book explains several examples step-by-step by using the AI-TOOLKIT. The book is going through the publishing process at the time of writing this article. You may use the contact form for info about pre-ordering the book.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

AI in Biometrics Recognition

AI in Biometrics Recognition

The Merriam-Webster dictionary describes biometrics as follows: “the measurement and analysis of unique physical or behavioral characteristics (such as fingerprint or voice patterns) especially as a means of verifying personal identity.”

There are many types of biometrics used today, for example, DNA matching, the shape of the ear, eye matching (iris, retina), facial features, fingerprinting, hand geometry, voice, signature, etc. Verifying personal identity may be very important in many applications for law enforcement, security and access control, and even in smart offices and homes where person dependent services may improve processes and everyday life for people.

Most biometrics identification systems work in a very similar manner and involve two main steps, feature extraction and feature (or pattern) matching. Feature extraction means that we analyze the chosen biometrics (a human face in this case) and extract a collection of features which are necessary to distinguish between different people. The aim is, of course, to limit the extracted information to the minimum amount necessary in order to optimize the machine learning training and prediction phases. Too much information would not only make everything much slower but it would also confuse the machine learning model, which should focus on the features that are really important for distinguishing different people. Feature matching is the process in which we use the extracted features in order to determine the identity of a person. We usually compare extracted features in a reference database to the input features for recognition.

The main steps of building and using a face recognition machine learning system can be divided into two major tasks:

• Training a machine learning model for feature extraction, and
• Performing face recognition with the help of the trained machine learning model.

The two major tasks explained above are further divided into several sub-tasks. First we need to train a machine learning model based on a huge number of input images (an image database) for feature extraction. The training of such a model may take several days or even weeks and may involve millions of images. The aim is that the ML model (a large scale convolutional neural network (CNN)) learns how to distinguish between the faces of different people. Deep inside of the system the CNN learns which face patterns are important in order to distinguish between different people.

As usual ML model training and testing are both important in order to arrive to a good final ML model.

The face recognition branch of the whole process involves the detection of face(s) in the input image, normalization of the extracted face image (we will see later how and why), feature extraction using the previously trained ML model and, finally, effective face recognition based on the extracted features.

After we have trained our CNN model we are ready to assemble a professional face recognition system.

As a first step we need to find automatically all of the faces in the input image and their exact location in order to extract the face images. Face detection is a complex problem because of the many possible face poses, rotations, scales, facial expressions, occlusions, etc.

Before we can perform face recognition we need to build a reference face recognition database with high quality frontal face images of people we want to recognize. The size of the face images should be similar to the size of the images we used during the training of the ML model for feature extraction. Face recognition systems usually extract and scale the face images automatically (AI-TOOLKIT) from selected input images.

The trained residual convolutional neural network (RCNN) can now be used to extract the feature vector from each detected and normalized face in an input image for recognition. Next we need to extract the feature vector from each image in the reference database. When we have all of the above feature vectors we can simply use a clustering algorithm in order to group (cluster) all feature vectors. If the detected image corresponds to one of the reference images, then both images will be grouped into the same cluster because the feature vectors are close to each other in the Euclidean space learned by the ML model if they are both face images of the same person. If the detected face (represented by its feature vector) is assigned to a cluster without any other face, then the face is an unknown face (it does not exist in the reference database).

Speaker recognition is similar to face recognition (feature extraction and identification) and based on some acoustic patterns (features) in human speech which are unique between individuals. The uniqueness of these acoustic patterns is due to the unique anatomy of humans (the shape and size of organs in the mouth called the vocal tract) and due to learned speech patterns and style.

The AI-TOOLKIT has built-in Apps which can be used for professional automatic face, speaker and fingerprint recognition.

This article is a slightly modified excerpt from the book “The Application of Artificial Intelligence”. If you are interested in the subject then it is strongly recommended to read the book which contains many more details and real world case studies for several sectors and disciplines! The book explains several examples step-by-step by using the AI-TOOLKIT. The book is going through the publishing process at the time of writing this article. You may use the contact form for info about pre-ordering the book.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

AI in Speech Recognition

AI in Speech Recognition

One of the most complex fields of machine learning is natural language processing (NLP). Automatic speech recognition is a sub-category of NLP. The aim of ASR is to provide a word sequence (often called a transcription) which corresponds to an input speech. The input speech is captured in the form of a speech waveform (acoustic signal), which is the combination of sound waves representing the change in air pressure over time (the creation and spread of sound - theoretical physics of sound waves).

There are many useful applications of ASR, for example, voice dialing, call routing (e.g., in order to automate call centers), data entry by speech, command and control (computers, vehicles, appliances…), etc. A GPS in a car or a smart phone which can be commanded by human speech is a good example of ASR.

The open source software VoiceBridge in the AI-TOOLKIT can be used to develop speech recognition applications with only a few lines of C++ code. The distribution of VoiceBridge contains two voice recognition application examples. The ‘Yes-No’ example trains an ML model for recognizing people saying ‘yes’ or ‘no’. The ‘LibriSpeech’ example trains a more realistic ML model to recognize English speech. More information about VoiceBridge can be found in the description of the software.

By using VoiceBridge you can train an ML model for a specific language which can transcribe the spoken words into written text. The transcription may be used for many useful applications in different sectors. Transcribing spoken text is the first step in comprehension and acting on the spoken text (e.g. answering or initiating an action).

This article is a slightly modified excerpt from the book “The Application of Artificial Intelligence”. If you are interested in the subject then it is strongly recommended to read the book which contains many more details and real world case studies for several sectors and disciplines! The book explains several examples step-by-step by using the AI-TOOLKIT. The book is going through the publishing process at the time of writing this article. You may use the contact form for info about pre-ordering the book.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.

AI in Detecting Diseases

AI in Detecting Diseases

The breast cancer diagnosis process is a complex and unpleasant process for the patients. This example will present a possible improvement of this process by using machine learning (ML). 

The patient goes through several process steps from which one step is where a digitized image of a breast mass is created and analyzed by the computer and the so called cell nucleus characteristics are measured and recorded. By studying and comparing the characteristics of the cell nucleus for many patients, who have or do not have cancer, and feeding the collected data to an ML model, the ML model can learn which characteristics result in cancer of the patient. The necessary ML training data attributes are decided by specialist and computed from a digitized image of a fine needle aspirate (FNA) of a breast mass. 

Building and using an ML model in the decision process does not only decrease process time significantly but it also makes the process more reliable (this depends of course on the accuracy of the ML model) because it eliminates possible human error. Another advantage could be that the input data can automatically be fed to the ML model and by this eliminating a very time consuming manual process step.

The input data

Each record contains a series of attributes and the final diagnosis whether the patient with these attributes has cancer (malignant tumor) or not. The aim is to collect all possible combinations of the attributes in a way that the ML model can be trained well and that it then can decide very accurately whether the patient has breast cancer or not.

Two digitized images with the cell nucleus present are shown below.

The different attributes in the data are as follows:

• Column 1
• Diagnosis: Malignant=1, Benign=2)
• Columns 2-31
• Ten real-valued features are computed for each cell nucleus. The mean, standard error, and "worst" or largest (mean of the three largest values) of these features were computed for each image, resulting in 30 features:
• Radius (mean of distances from center to points on the perimeter)
• Texture (standard deviation of gray-scale values)
• Perimeter
• Area
• Smoothness (local variation in radius lengths)
• Compactness (perimeter^2 / area - 1.0)
• Concavity (severity of concave portions of the contour)
• Concave points (number of concave portions of the contour)
• Symmetry
• Fractal dimension ("coastline approximation" - 1)

digitized images with the cell nucleus present, source [2]

The data file (can be downloaded at the end of the article) has a simple tab separated format. In order to use the data in the AI-TOOLKIT we need to change the extension of the data file to ‘.TSV’ (the AI-TOOLKIT expects this extension for tab delimited data files).

In order to use the fully numerical ML model all attributes need to be converted to numerical values. In our case there is only one non-numerical attribute and that is the Decision variable which is the Diagnosis whether the patient has breast cancer or not. The two possible options can be simply converted to Malignant=1, Benign=2. The AI-TOOLKIT can do this conversion automatically for you while importing the data (select the ‘Automatically Convert Categorical or Text values’ option) or you can just do a text replace in a text editor.

After preparing the input data in the appropriate format (tab separated values) the type of the ML model must be chosen. Let us choose an SVM model for this example.

First, in case of an SVM model the ML model parameters need to be optimized. This can be done automatically by the AI-TOOLKIT by using the built-in SVM Parameter Optimization module. The AI-TOOLKIT will report the best parameter combination for the input data which then can be filled in as follows:

model:

    id: 'ID-WFcqHlreYm'

    type: SVM

    path: 'wdbc.sl3'

    params:

        - svm_type: C_SVC 

        - kernel_type: RBF 

        - gamma: 15.0 

        - C: 1.779 

    training: 

        - data_id: 'wdbc' 

        - dec_id: 'decision' 

    test: 

        - data_id: 'wdbc_t' 

        - dec_id: 'decision'

    input: 

        - data_id: 'input_data' 

        - dec_id: 'decision'

    output:

        - data_id: 'output_data'

        - col_id: 'decision'

After importing the data, defining the data table names (wdbc and wdbc_t) and entering the optimal model parameters the ML model can be trained. 

When the ML model is ready learning the problem it will let you know the accuracy of the model on the training dataset:

Performance Evaluation Results: TRAINING
Confusion Matrix [predicted x original] (number of classes: 2):
	(0)	(1)
(0)	199	0
(1)	0	342
Accuracy	100.00%
Error	0.00%
C.Kappa	100.00%
	(0)	(1)
Precision	100.00%	100.00%
Recall	100.00%	100.00%
FNR	0.00%	0.00%
F1	100.00%	100.00%
TNR	100.00%	100.00%
FPR	0.00%	0.00%

The ML model is able to predict correctly whether the patient has breast cancer or not in all of the cases but do not forget that the model still needs to be tested with an appropriate number of data records (attribute sets) unseen during the training of the ML model in order to make sure that the ML model is learned enough about the phenomena and that it generalizes well!

In the case we use 5% of the input data for testing (removing it from the training data) and let the AI-TOOLKIT test the trained ML model with this test data then we get the results as follows:

Performance Evaluation Results: TEST
Confusion Matrix [predicted x original] (number of classes: 2):
	(0)	(1)
(0)	12	1
(1)	1	14
Accuracy	92.86%
Error	7.14%
C.Kappa	85.64%
	(0)	(1)
Precision	92.31%	93.33%
Recall	92.31%	93.33%
FNR	7.69%	6.67%
F1	92.31%	93.33%
TNR	93.33%	92.31%
FPR	6.67%	7.69%

The test results are less good than what we have seen during the training but the trained ML model can still predict 26 cases well from 28 which is still a very good result especially if we consider that we use new data! The ML model makes 1 mistake predicting incorrectly cancer when it should not and 1 mistake predicting no cancer when it should. Predicting incorrectly cancer is a less sever mistake because the diagnosis can still be checked by a medical doctor but the mistake of predicting no cancer when there is cancer should be eliminated! This is an important special way of ML model evaluation in the healthcare sector, not all mistakes have the same weight!

The above SVM model can still be improved by adding more data or/and changing the input features. It is of course also possible to choose another ML model e.g., a neural network model.

The extended performance evaluation results of the AI-TOOLKIT allows us to make a thorough analysis of the performance of the ML model but this is left as an exercise for the reader.

The trained ML model can be used to make important decisions and the input data could be fed to the ML model automatically and the results could also be collected automatically. The ML algorithm could even be integrated into different digital devices in order to have an all-in automatic analysis possible.

Conclusion

As we have seen above an ML model can be very useful in the improvement of business processes. The techniques explained in this article can be used not only in the healthcare sector but in many other sectors too! There are two important considerations while using an ML model:

1. The attributes and the data records (attribute sets) used to train the ML model are very important. The capabilities of the ML model will depend on the data it gets for learning a specific phenomenon. You can of course add more data and/or attributes and re-train the model. Not only the amount of input data but the selection of the right attributes (features) is also very important.
2. Extensively testing the ML model is very important in order to make sure that it is trained well in all aspects of the studied phenomena and that the model generalizes well (performs well in case of using during the training unseen input data).

References

1. The Application of Artificial Intelligence, Zoltan Somogyi.
2. Breast Cancer Wisconsin (Diagnostic) Data Set: Dr. William H. Wolberg, General Surgery Dept. University of Wisconsin, Clinical Sciences Center Madison, WI 53792. You can download the dataset here: Breast Cancer Diagnosis data set.

Learn about the application of Artificial Intelligence and Machine Learning from the book "The Application of Artificial Intelligence | Step-by-Step Guide from Beginner to Expert", Springer 2020 (~400 pages) (ISBN 978-3-030-60031-0). Unique, understandable view of machine learning using many practical examples. Introduces AI-TOOLKIT, freely available software that allows the reader to test and study the examples in the book. No programming or scripting skills needed! Suitable for self-study by professionals, also useful as a supplementary resource for advanced undergraduate and graduate courses on AI. More information can be found at the Springer website: Springer book: The Application of Artificial Intelligence.