Five machine learning types to know

Machine learning (ML) technologies can drive decision-making in virtually all industries, from healthcare to human resources to finance and in myriad use cases, like computer vision, large language models (LLMs), speech recognition, self-driving cars and more.

However, the growing influence of ML isn’t without complications. The validation and training datasets that undergird ML technology are often aggregated by human beings, and humans are susceptible to bias and prone to error. Even in cases where an ML model isn’t itself biased or faulty, deploying it in the wrong context can produce errors with unintended harmful consequences.

That’s why diversifying enterprise AI and ML usage can prove invaluable to maintaining a competitive edge. Each type and sub-type of ML algorithm has unique benefits and capabilities that teams can leverage for different tasks. Here, we’ll discuss the five major types and their applications.

What is machine learning?

ML is a computer science, data science and artificial intelligence (AI) subset that enables systems to learn and improve from data without additional programming interventions.

Instead of using explicit instructions for performance optimization, ML models rely on algorithms and statistical models that deploy tasks based on data patterns and inferences. In other words, ML leverages input data to predict outputs, continuously updating outputs as new data becomes available.

On retail websites, for instance, machine learning algorithms influence consumer buying decisions by making recommendations based on purchase history. Many retailers’ e-commerce platforms—including those of IBM, Amazon, Google, Meta and Netflix—rely on artificial neural networks (ANNs) to deliver personalized recommendations. And retailers frequently leverage data from chatbots and virtual assistants, in concert with ML and natural language processing (NLP) technology, to automate users’ shopping experiences.

Machine learning types

Machine learning algorithms fall into five broad categories: supervised learning, unsupervised learning, semi-supervised learning, self-supervised and reinforcement learning.

1. Supervised machine learning

Supervised machine learning is a type of machine learning where the model is trained on a labeled dataset (i.e., the target or outcome variable is known). For instance, if data scientists were building a model for tornado forecasting, the input variables might include date, location, temperature, wind flow patterns and more, and the output would be the actual tornado activity recorded for those days.

Supervised learning is commonly used for risk assessment, image recognition, predictive analytics and fraud detection, and comprises several types of algorithms.

• Regression algorithms—predict output values by identifying linear relationships between real or continuous values (e.g., temperature, salary). Regression algorithms include linear regression, random forest and gradient boosting, as well as other subtypes.
• Classification algorithms—predict categorical output variables (e.g., “junk” or “not junk”) by labeling pieces of input data. Classification algorithms include logistic regression, k-nearest neighbors and support vector machines (SVMs), among others.
• Naïve Bayes classifiers—enable classification tasks for large datasets. They’re also part of a family of generative learning algorithms that model the input distribution of a given class or/category. Naïve Bayes algorithms include decision trees, which can actually accommodate both regression and classification algorithms.
• Neural networks—simulate the way the human brain works, with a huge number of linked processing nodes that can facilitate processes like natural language translation, image recognition, speech recognition and image creation.
• Random forest algorithms—predict a value or category by combining the results from a number of decision trees.

2. Unsupervised machine learning

Unsupervised learning algorithms—like Apriori, Gaussian Mixture Models (GMMs) and principal component analysis (PCA)—draw inferences from unlabeled datasets, facilitating exploratory data analysis and enabling pattern recognition and predictive modeling.

The most common unsupervised learning method is cluster analysis, which uses clustering algorithms to categorize data points according to value similarity (as in customer segmentation or anomaly detection). Association algorithms allow data scientists to identify associations between data objects inside large databases, facilitating data visualization and dimensionality reduction.

• K-means clustering—assigns data points into K groups, where the data points closest to a given centroid are clustered under the same category and K represents clusters based on their size and level of granularity. K-means clustering is commonly used for market segmentation, document clustering, image segmentation and image compression.
• Hierarchical clustering—describes a set of clustering techniques, including agglomerative clustering—where data points are initially isolated into groups and then merged iteratively based on similarity until one cluster remains—and divisive clustering—where a single data cluster is divided based on the differences between data points.
• Probabilistic clustering—helps solve density estimation or “soft” clustering problems by grouping data points based on the likelihood that they belong to a particular distribution.

Unsupervised ML models are often behind the “customers who bought this also bought…” types of recommendation systems.

3. Self-supervised machine learning

Self-supervised learning (SSL) enables models to train themselves on unlabeled data, instead of requiring massive annotated and/or labeled datasets. SSL algorithms, also called predictive or pretext learning algorithms, learn one part of the input from another part, automatically generating labels and transforming unsupervised problems into supervised ones. These algorithms are especially useful for jobs like computer vision and NLP, where the volume of labeled training data needed to train models can be exceptionally large (sometimes prohibitively so).

4. Reinforcement learning

Reinforcement learning, also called reinforcement learning from human feedback (RLHF), is a type of dynamic programming that trains algorithms using a system of reward and punishment. To deploy reinforcement learning, an agent takes actions in a specific environment to reach a predetermined goal. The agent is rewarded or penalized for its actions based on an established metric (typically points), encouraging the agent to continue good practices and discard bad ones. With repetition, the agent learns the best strategies.

Reinforcement learning algorithms are common in video game development and are frequently used to teach robots how to replicate human tasks.

5. Semi-supervised learning

The fifth type of machine learning technique offers a combination between supervised and unsupervised learning.

Semi-supervised learning algorithms are trained on a small labeled dataset and a large unlabeled dataset, with the labeled data guiding the learning process for the larger body of unlabeled data. A semi-supervised learning model might use unsupervised learning to identify data clusters and then use supervised learning to label the clusters.

Generative adversarial networks (GANs)—deep learning tool that generates unlabeled data by training two neural networks—are an example of semi-supervised machine learning.

Regardless of type, ML models can glean data insights from enterprise data, but their vulnerability to human/data bias make responsible AI practices an organizational imperative.

Manage a range of machine learning models with watstonx.ai

Nearly everyone, from developers to users to regulators, engages with applications of machine learning at some point, whether they interact directly with AI technology or not. And the adoption of ML technology is only accelerating. The global machine learning market was valued at USD 19 billion in 2022 and is expected to reach USD 188 billion by 2030 (a CAGR of more than 37 percent).

The scale of ML adoption and its growing business impact make understanding AI and ML technologies an ongoing—and vitally important—commitment, requiring vigilant monitoring and timely adjustments as technologies evolve. With IBM® watsonx.ai AI studio, developers can manage ML algorithms and processes with ease.

IBM watsonx.ai—part of the IBM watsonx AI and data platform—combines new generative AI capabilities and a next-generation enterprise studio to help AI builders train, validate, tune and deploy AI models with a fraction of the data, in a fraction of the time. Watsonx.ai offers teams advanced data generation and classification features that help businesses leverage data insights for optimal real-world AI performance.

In the age of data proliferation, AI and machine learning are as integral to day-to-day business operations as they are to tech innovation and business competition. But as new pillars of a modern society, they also represent an opportunity to diversify enterprise IT infrastructures and create technologies that work for the benefit of businesses and the people who depend on them.

Source: ibm.com

Continue reading

How to build a decision tree model in IBM Db2

After developing a machine learning model, you need a place to run your model and serve predictions. If your company is in the early stage of its AI journey or has budget constraints, you may struggle to find a deployment system for your model. Building ML infrastructure and integrating ML models with the larger business are major bottlenecks to AI adoption. IBM Db2 can help solve these problems with its built-in ML infrastructure. Someone with the knowledge of SQL and access to a Db2 instance, where the in-database ML feature is enabled, can easily learn to build and use a machine learning model in the database.

In this post, I will show how to develop, deploy, and use a decision tree model in a Db2 database.

These are my major steps in this tutorial:

1. Set up Db2 tables

2. Explore ML dataset

3. Preprocess the dataset

4. Train a decision tree model

5. Generate predictions using the model

6. Evaluate the model

I implemented these steps in a Db2 Warehouse on-prem database. Db2 Warehouse on cloud also supports these ML features.

The machine learning use case

I will use a dataset of historical flights in the US. For each flight, the dataset has information such as the flight’s origin airport, departure time, flying time, and arrival time. Also, a column in the dataset indicates if each flight had arrived on time or late. Using examples from the dataset, we’ll build a classification model with decision tree algorithm. Once trained, the model can receive as input unseen flight data and predict if the flight will arrive on time or late at its destination.

1. Set up Db2 tables

The dataset I use in this tutorial is available here as a csv file.

Creating a Db2 table

I use the following SQL for creating a table for storing the dataset.

db2start

connect to <database_name>

db2 "CREATE TABLE FLIGHTS.FLIGHTS_DATA_V3  (

ID INTEGER NOT NULL GENERATED BY DEFAULT AS IDENTITY,

YEAR INTEGER ,

QUARTER INTEGER ,

MONTH INTEGER ,

DAYOFMONTH INTEGER ,                       

DAYOFWEEK INTEGER ,                       

UNIQUECARRIER VARCHAR(50 OCTETS) ,               

ORIGIN VARCHAR(50 OCTETS) ,                      

DEST VARCHAR(50 OCTETS) ,                       

CRSDEPTIME INTEGER ,                       

DEPTIME INTEGER ,                       

DEPDELAY REAL ,                       

DEPDEL15 REAL ,                       

TAXIOUT INTEGER ,                       

WHEELSOFF INTEGER ,                       

CRSARRTIME INTEGER ,                       

CRSELAPSEDTIME INTEGER ,                       

AIRTIME INTEGER ,                       

DISTANCEGROUP INTEGER ,                       

FLIGHTSTATUS VARCHAR(1) )

ORGANIZE BY ROW";

After creating the table, I use the following SQL to load the data, from the csv file, into the table:

db2 "IMPORT FROM 'FLIGHTS_DATA_V3.csv' OF DEL COMMITCOUNT 50000 INSERT INTO FLIGHTS.FLIGHTS_DATA_V3"

I now have the ML dataset loaded into the FLIGHTS.FLIGHTS_DATA_V3 table in Db2. I’ll copy a subset of the records from this table to a separate table for the ML model development and evaluation, leaving the original copy of the data intact. 

SELECT count(*) FROM FLIGHTS.FLIGHTS_DATA_V3

 — — — 

1000000

Creating a separate table with sample records

Create a table with 10% sample rows from the above table. Use the RAND function of Db2 for random sampling.

CREATE TABLE FLIGHT.FLIGHTS_DATA AS (SELECT * FROM FLIGHTS.FLIGHTS_DATA_V3 WHERE RAND() < 0.1) WITH DATA

Count the number of rows in the sample table.

SELECT count(*) FROM FLIGHT.FLIGHTS_DATA

— — — 

99879

Look into the scheme definition of the table.

SELECT NAME, COLTYPE, LENGTH

FROM SYSIBM.SYSCOLUMNS

WHERE TBCREATOR = 'FLIGHT' AND TBNAME = 'FLIGHTS_DATA'

ORDER BY COLNO

FLIGHTSTATUS is the response or the target column. Others are feature columns.

Find the DISTINCT values in the target column.

From these values, I can see that it’s a binary classification task where each flight arrived either on time or late. 

Find the frequencies of distinct values in the FLIGHTSTATUS column.

SELECT FLIGHTSTATUS, count(*) AS FREQUENCY, count(*) / (SELECT count(*) FROM FLIGHT.FLIGHTS_DATA) AS FRACTION

FROM FLIGHT.FLIGHTS_DATA fdf

GROUP BY FLIGHTSTATUS

From the above, I see the classes are imbalanced. Now I’ll not gain any further insights from the entire dataset, as this can leak information to the modeling phase. 

Creating train/test partitions of the dataset

Before collecting deeper insights into the data, I’ll divide this dataset into train and test partitions using Db2’s RANDOM_SAMPLING SP. I apply stratified sampling to preserve the ratio between two classes in the generated training data set.

Create a TRAIN partition.

call IDAX.RANDOM_SAMPLE('intable=FLIGHT.FLIGHTS_DATA, fraction=0.8, outtable=FLIGHT.FLIGHTS_TRAIN, by=FLIGHTSTATUS')

Copy the remaining records to a test PARTITION.

CREATE TABLE FLIGHT.FLIGHTS_TEST AS (SELECT * FROM FLIGHT.FLIGHTS_DATA FDF WHERE FDF.ID NOT IN(SELECT FT.ID FROM FLIGHT.FLIGHTS_TRAIN FT)) WITH DATA

2. Explore data

In this step, I’ll look at both sample records and the summary statistics of the training dataset to gain insights into the dataset.

Look into some sample records.

SELECT * FROM FLIGHT.FLIGHTS_TRAIN FETCH FIRST 10 ROWS ONLY

Some columns have encoded the time as numbers:

 — CRSDEPTIME: Computer Reservation System (scheduled) Departure Time (hhmm)

 — DepTime: Departure Time (hhmm)

 — CRSArrTime: Computer Reservation System (scheduled) Arrival Time

Now, I collect summary statistics from the FLIGHTS_TRAIN using SUMMARY1000 SP to get a global view of the characteristics of the dataset.

CALL IDAX.SUMMARY1000('intable=FLIGHT.FLIGHTS_TRAIN, outtable=FLIGHT.FLIGHTS_TRAIN_SUM1000')

Here the intable has the name of the input table from which I want SUMMARY1000 SP to collect statistics. outtable is the name of the table where SUMMARY1000 will store gathered statistics for the entire dataset. Besides the outtable, SUMMARY1000 SP creates a few additional output tables — one table with statistics for each column type. Our dataset has two types of columns — numeric and nominal. So, SUMMARY1000 will generate two additional tables. These additional tables follow this naming convention: the name of the outtable + column type. In our case, the column types are NUM, representing numeric, and CHAR, representing nominal. So, the names of these two additional tables will be as follows:

FLIGHTS_TRAIN_SUM1000_NUM

FLIGHTS_TRAIN_SUM1000_CHAR

Having the statistics available in separate tables for specific datatypes makes it easier to view the statistics that apply to specific datatype and reduce the number of columns whose statistics are viewed together. This simplifies the analysis process. 

Check the summary statistics of the numeric column.

SELECT * FROM FLIGHT.FLIGHTS_TRAIN_SUM1000_NUM

For the numeric columns, SUMMARY1000 gather the following statistics:

◉ Missing value count

◉ Non-missing value count

◉ Average

◉ Variance

◉ Standard deviation

◉ Skewness

◉ Excess kurtosis

◉ Minimum

◉ Maximum

Each of these statistics can help uncover insights into the dataset. For instance, I can see that DEPDEL15 and DEPDELAY columns have 49 missing values. There are large values in these columns: AIRTIME, CRSARRTIME, CRSDEPTIME, CRSELAPSEDTIME, DEPDELAY, DEPTIME, TAXIOUT, WHEELSOFF, and YEAR. Since I will create a decision tree model, I don’t need to deal with the large value and the missing values. Db2 will deal with both issues natively. 

Next, I investigate the summary statistics of the nominal columns.

select * from FLIGHT.FLIGHTS_TRAIN_SUM1000_CHAR

For nominal columns, SUMMARY1000 gathered the following statistics:

◉ Number of missing values

◉ Number of non-missing values

◉ Number of distinct values

◉ Frequency of the most frequent value

3. Preprocess data

From the above data exploration, I can see that the dataset has no missing values. These four TIME columns have large values: AIRTIME, CRSARRTIME, DEPTIME, WHEELSOFF. I’ll leave the nominal values in all columns as-is, as the decision tree implementation in Db2 can deal with them natively. 

Extract the hour part from the TIME columns — CRSARRTIME, DEPTIME, WHEELSOFF.

From looking up the description of the dataset, I see the values in the CRSARRTIME, DEPTIME, and WHEELSOFF columns are encoding of hhmm of the time values. I extract the hour part of these values to create, hopefully, better features for the learning algorithm. 

Scale CRSARRTIME COLUMN: divide the value with 100 gives the hour of the flight arrival time:

UPDATE FLIGHT.FLIGHTS_TRAIN SET CRSARRTIME = CRSARRTIME / 100

Scale DEPTIME COLUMN: divide the value by 100 gives the hour of the flight arrival time:

UPDATE FLIGHT.FLIGHTS_TRAIN SET DEPTIME = DEPTIME / 100

Scale WHEELSOFF COLUMN: divide the value by 100 will give the hour of the flight arrival time:

UPDATE FLIGHT.FLIGHTS_TRAIN SET WHEELSOFF = WHEELSOFF / 100

4. Train a decision tree model

Now the training dataset is ready for the decision tree algorithm. 

I train a decision tree model using GROW_DECTREE SP. 

CALL IDAX.GROW_DECTREE('model=FLIGHT.flight_dectree, intable=FLIGHT.FLIGHTS_TRAIN, id=ID, target=FLIGHTSTATUS')

I called this SP using the following parameters:

◉ model: the name I want to give to the decision tree model — FLIGHT_DECTREE

◉ intable: the name of the table where the training dataset is stored

◉ id: the name of the ID column

◉ target: the name of the target column

After completing the model training, the GROW_DECTREE SP generated several tables with metadata from the model and the training dataset. Here are some of the key tables:

◉ FLIGHT_DECTREE_MODEL: this table contains metadata about the model. Examples of metadata include depth of the tree, strategy for handling missing values, and the number of leaf nodes in the tree. 

◉ FLIGHT_DECTREE_NODES: this table provides information about each node in the decision tree. 

◉ FLIGHT_DECTREE_COLUMNS: this table provides information on each input column and their role in the trained model. The information includes the importance of a column in generating a prediction from the model. 

5. Generate predictions from the model

Since the FLIGHT_DECTREE model is trained and deployed in the database, I can use it for generating predictions on the test records from the FLIGHTS_TEST table.

First, I preprocess the test dataset using the same preprocessing logic that I applied to the TRAINING dataset. 

Scale CRSARRTIME COLUMN: divide the value by 100 will give the hour of the flight arrival time:

UPDATE FLIGHT.FLIGHTS_TEST SET CRSARRTIME = CRSARRTIME / 100

Scale DEPTIME COLUMN: divide the value by 100 will give the hour of the flight arrival time:

UPDATE FLIGHT.FLIGHTS_TEST SET DEPTIME = DEPTIME / 100

Scale WHEELSOFF COLUMN: divide the value by 100 will give the hour of the flight arrival time:

UPDATE FLIGHT.FLIGHTS_TEST SET WHEELSOFF = WHEELSOFF / 100

Generating predictions

I use PREDICT_DECTREE SP to generate predictions from the FLIGHT_DECTREE model:

CALL IDAX.PREDICT_DECTREE('model=FLIGHT.flight_dectree, intable=FLIGHT.FLIGHTS_TEST, outtable=FLIGHT.FLIGHTS_TEST_PRED, prob=true, outtableprob=FLIGHT.FLIGHTS_TEST_PRED_DIST')

Here is the list of parameters I passed when calling this SP:

◉ model: the name of the decision tree model, FLIGHT_DECTREE

◉ intable: name of the input table to generate predictions from

◉ outtable: the name of the table that the SP will create and store predictions to

◉ prob: a boolean flag indicating if we want to include in the output the probability of each prediction

◉ outputtableprob: the name of the output table where the probability of each prediction will be stored 

6. Evaluate the model

Using generated predictions for the test dataset, I compute a few metrics to evaluate the quality of the model’s predictions.

Creating a confusion matrix

I use CONFUSION_MATRIX SP to create a confusion matrix based on the model’s prediction on the TEST dataset. 

CALL IDAX.CONFUSION_MATRIX('intable=FLIGHT.FLIGHTS_TEST, resulttable=FLIGHT.FLIGHTS_TEST_PRED, id=ID, target=FLIGHTSTATUS, matrixTable=FLIGHT.FLIGHTS_TEST_CMATRIX')

In calling this SP, here are some of the key parameters that I passed:

◉ intable: the name of the table that contains the dataset and the actual value of the target column

◉ resulttable: the name of the table that contains the column with predicted values from the model

◉ target: the name of the target column

◉ matrixTable: The output table where the SP will store the confusion matrix

After the SP completes its run, we have the following output table with statistics for the confusion matrix. 

FLIGHTS_TEST_CMATRIX:

This table has three columns. The REAL column has the actual flight status. PREDICTION column has the predicted flight status. Since flight status takes two values – 0 (on time) or 1 (delayed), we have four possible combinations between values in the REAL and the PREDICTION columns: 

1. TRUE NEGATIVE: REAL: 0, PREDICTION: 0 — The model has accurately predicted the status of those flights that arrived on schedule. From that CNT column, we see that 11795 rows from the TEST table belong to this combination.

2. FALSE POSITIVE: REAL: 0, PREDICTION: 1 — these are the flights that actually arrived on time but the model predicted them to be delayed. 671 is the count of such flights. 

3. FALSE NEGATIVE: REAL: 1, PREDICTION: 0 — these flights have arrived late, but the model predicted them to be on time. From the CNT table, we find their count to be 2528.

4. TRUE POSITIVE: REAL: 1, PREDICTION: 1 — the model has accurately identified these flights that were late. The count is 4981. 

I use these counts to compute a few evaluation metrics for the model. For doing so, I use CMATRIX_STATS SP as follows:

CALL IDAX.CMATRIX_STATS('matrixTable=FLIGHT.FLIGHTS_TEST_CMATRIX')

The only parameter this SP needs is the name of the table that contains the statistics generated by the CONFUSION_MATRIX SP in the previous step. CMATRIX_STATS SP generates two sets of output. The first one shows overall quality metrics of the model. The second one includes the model’s predictive performance for each class. 

First output — overall model metrics include correction predictions, incorrect prediction, overall accuracy, weighted accuracy. From this output, I see that the model has an overall accuracy of 83.98% and a weighted accuracy of 80.46%. 

With classification tasks, it’s usually useful to view the model’s quality factors for each individual class. The second output from the CMATRIX_STATS SP includes these class level quality metrics. 

For each class, this output includes the True Positive Rate (TPR), False Positive Rate (FPR), Positive Predictive Value (PPV) or Precision, and F-measure (F1 score). 

Source: ibm.com

Continue reading

The three-layered approach to network data monetization

Automated operations in the modern state-of-the-art telco networks, such as 5G Core, O-RAN and SDN, all follow the same pattern. The main enabler is the entity that streamlines collection and consumption of the network state. This is Network Data and Analytics Function (NWDAF) in 5G Core, Radio Intelligent Controller (RIC) in O-RAN and SDN Controller in Software Defined Networks. Each of these entities applies a three-layered approach to network data monetization. While the functionality of these network state consumers is very similar, their implementations are different. This results in vendor specific solutions, silos per network domain, and more importantly, less opportunity to monetize the potential of network data.

Identifying the three layers of complexity

Modern telco networks produce a lot of data. Each of the above-mentioned network entities first collect the network data and exposes it to the interested network state consumers. The network state consumers are various analytical modules that may vary from a simple threshold-based algorithms that analyze the current network state to very complex AI/ML-based algorithms that predict the future network state. Finally, based on the insights from these algorithms different actions are pushed towards the network, depending on the implemented automation use cases, and thus the loop is closed.

Let us now take the example of a European CSP which has a lot of ideas on how to improve their network operations. They have performed many different Proof-of-Concepts, all focusing on the Analytics part of the model. The data- and automation aspects were neglected in these PoC’s, resulting in very sub-optimal results. Many use cases yielded disappointing results, due to data limitations, and when the results were good, the required automation capabilities were not (yet) available, which significantly reduced the business benefits of the use case. In one other case, an energy saving use case was not deployed, because the required policy management function was not available.

IBM advocates to disaggregate data and analytics functions and combine them with automation capabilities, as actionable data-driven analytical insights without automation do not bring the full business value.

Here we identify three layers of complexity, where each of the three proposed layers brings specific value: Data Democratization in the Data Layer, an open “plug & play” AI/ML module in the Analytics Layer, and finally the ability to act directly on the actionable insights in the Automation Layer.

1. Data Layer: collecting the data and making it accessible and understandable to all interested consumers.

2. Analytics Layer: analyzing the data for the various Use Cases to provide actionable insights. 

3. Automation Layer: acting upon the actionable insights in an automated way.

A lot of the industry hype today is around AI/ML and Automation. However, when it comes to Telco Networks the complexity of the Data layer is often overlooked.

Figure 1 Three-Layered Model for Network Data Monetization

Data is the new natural resource; however, this does not imply that it is easy to use and to have it monetized. To get the maximum benefit from it, the right data must be efficiently mined from its hiding places and only afterwards valuable insights can be produced out of it. In addition to this, network data is generated all the time and everybody has it – indeed, each CSP has an abundant unlimited data source that never stops. Therefore, data mining is the business of every CSP nowadays. The question is – how to ensure that your data mining is efficient and that it allows you to unlock the power of your data?

Use cases on how to get started

Automated operations opened a world of possibilities for the telco operators. The concept came at the same time as the necessity in the era of technological evolution towards disaggregation which enormously increased the complexity of modern telco network operations, but also as the main enabler for the continued evolution of telco services towards SLAs never seen before, such as ultra-low latency or enhanced mobile broadband. From today’s perspective more and more operations in telco will be automated as manual operations cannot cope with the challenges. However, it is yet not clear what are the winning use cases that will bring the best monetary values to the operators. We call this the Business Value Hypothesis. Today we can distinguish the two categories:

◉ The obvious or foundational use cases where the business value hypothesis is intuitively clear. Examples of such use cases are energy saving, anomaly detection, automated root cause analysis, failure prediction. The value of these use cases has already been proven by others, the task at hand is to develop and deploy these use cases in the own environment, which may not always be easy to do.

◉ The less obvious or aspirational use cases where the business value hypothesis is yet to be proven. Data can be combined from network, customer domains and call data records as well as transactions that go over the network to arrive at novel insights that are of value, e.g., unsuspected correlations. We refer here to the ideas, internal gut feelings, etc. of combining network data with other sources of data that may yield results but their exact value and monetary of achieved quality improvements has yet to be proved.

The above elaboration on the still existing uncertainty of the monetary values of the data-driven analytics use cases in telco underscores the need for the CSPs to host their very own playgrounds for analytics. In fact, CSPs need an open platform, in which they can easily develop and deploy their own analytical modules but also deploy 3rd party best of breed analytical modules, as CSPs are unlikely to create all analytical models by themselves.

It requires the easy access to network data in an understandable way, to avoid massive time loss in trying to find the right network data sets. Having readily available data models can reduce time and effort to implement use cases by 30 to 50%.

The first step towards the monetization of your Network Data is to create a Value Tree. What are the various categories of value that can be identified and specifying these categories in more and more detail. The Network Data Value Tree is not static, it continues to evolve as new ideas emerge, and successfully implemented use case spark new ideas.

Figure 2 Network Data Value Tree Example

In the analysis of which use case to develop, it is important to assess the availability of the right data, and what you want to do with results of the data analytics: what is the required automation to act upon based on these newly obtained insights. Insights that cannot be acted upon have far less value than insights that can be acted upon in an automated way.

Source: ibm.com

Continue reading

Six innovation strategies for Life Science organizations in 2023

Last year, as life sciences organizations were consumed by the recovery from COVID-19, their focus had to shift rapidly to mitigating supply chain constraints, labor and skill shortages, and by the end of the year, inflationary pressures—all of which were exacerbated by the Russia-Ukraine war.

Along with these challenges, the ongoing drive to reduce costs, improve efficiency and productivity, drive better decision-making and reduce risk will continue to drive pharma investment in cloud, AI/ML, analytics and automation in 2023 despite higher interest rates.

Organizations must rethink their business models to serve a variety of strategic goals, and AI will play an increasingly important role in all of them: supporting drug discovery, trial diversity, forecasting and supply chain functions, and supporting engagement and adherence to decentralized trials and on-market regimens. These shifts will set the stage for further industry transformation as gene therapy and precision health become more widely available.

The macro-environment favors steady long-term focus and growth

The Inflation Reduction Act of 2022 puts pressure on companies to reduce drug and device prices in the USA, causing pharma to leverage technology to drive cost efficiencies and maintain margins. Globally, the shifting policy debate around access and affordability of patented pharmaceuticals exerts additional pressures. An inflationary environment will slow down traditional R&D and leave pharma no choice but to investigate AI/ML and related techniques to accelerate drug discovery and repurposing while reducing costs. This will also stimulate new partnerships (for example, pharma companies working with research labs or providers working with payers) through federated learning and cloud-based digital ecosystems.

Manufacturing costs and costs of clinical trials will continue to rise. The cost of active pharmaceutical ingredients has increased by up to 70% since 2018. Recruiting on-site patients and maintaining on-site trials remains prohibitively expensive. Pharma will focus on digital patient recruitment through social advertising and digital engagement (including wearables) throughout trials to manage adherence and persistence. Decentralized trial structures will push pharma to focus more closely on cybersecurity and protected health information (PHI) while managing the associated costs.

The costs of operating manufacturing facilities will increase as energy prices continue to rise. We predict companies will make significant strides towards digitization to reduce cost, improve quality, reduce recalls and improve safety. Pre-digital facilities with manual processes supported by expensive labor will no longer be the norm.

Strategy 1: Prioritize around novel drug development, generics or consumer engagement

Strategic reprioritization should be top of mind for the C-suite. Last year, Pfizer and GSK left the consumer health sector to prioritize novel drug and vaccine development and core innovation. We saw targeted mergers and acquisitions to replenish pipelines, which will continue in 2023. Novartis spun off Sandoz, their generics business, and is streamlining their research efforts around innovative pharmaceuticals. Sanofi and the generics giant Sun Pharmaceuticals are repositioning to enter specialty pharma with the latter releasing a new highly competitive biologic for psoriasis.

Companies are recognizing that divergent business models are required: innovative pharma requires significant capital investments to support state-of-the-art research enabled by new technology, while generic and consumer health business models demand unparalleled scale, access to distribution and close partnerships with pharmacies. A rising risk-return ratio for innovators will lead companies to form research and technology partnerships that combine top talent with the most innovative computation techniques, rather than outright mergers or acquisitions.

Strategy 2: Use AI-driven forecasting and supply chains to improve operational efficiency and sustainability

The upheaval and disruption caused by COVID-19 has given pharma leaders a heightened awareness of resiliency in delivering innovative drugs and therapeutics to communities, underscoring the importance of investing in forecasting and supply chains.

Forecasting

Forecasting continues to be a pain point for many organizations. According to a recent IBM IBV study, 57% of CEOs view the CFO as playing the most crucial role in their organizations over the next two to three years. Legacy processes, demand volatility and increasing data scale and complexity demand a new approach. Traditional quarterly forecasting cycles (which are manual and burdensome) yield inaccurate predictions.

Leading companies will invest in AI, ML and intelligent workflows to deliver end-to-end forecasting capabilities that utilize real-time feeds from multiple data sources, leveraging hundreds of AI and ML models, to deliver more granular and accurate forecasts and customer insights. These capabilities will fundamentally change the role of finance organizations by emphasizing speed of insight, adoption of data-driven decision making and scaling of analytics within the enterprise.

Supply chain

Business leaders will focus on supply chain solutions that drive transparency across sourcing, manufacturing, delivery and logistics while minimizing cost, waste and time. CSCOs are modernizing supply chain operations by using AI to leverage unstructured data in real time and integrating automation, blockchain and edge computing to manage operations and collect and connect information across multiple sources.

Priorities driving supply chain innovation, according to CSCOs

In the wake of COVID-19, we observe leaders viewing the supply chain as a core organizational function rather than a supportive one. David Volk, executive director of clinical supply chain planning at Roche states, “We are a networked organization… collaborating much more broadly across all our partners and the industry. We view ourselves as a supply chain organization, and a significant part of the value we bring to patients lies in optimizing our global supply chain and inventory. That’s a very different mindset, and it’s changed how we run the organization.”

Supply chain sustainability also ranks among the highest priorities for CEOs. 48% of CEOs surveyed say increasing sustainability is a top priority—up 37% since 2021. 44% cite a lack of data-driven insights as barriers to achieving sustainability objectives. End-to-end visibility into sustainability impact, such as metrics on emissions and waste from raw material to delivery, will unlock a new level of information that position CSCOs as key enablers for companies to achieve their sustainability and ESG vision.

Strategy 3: Prepare for an influx of cell and gene therapies

Gene therapy is the new frontier of medicine. It focuses on targeting a person’s genes for modification to treat or cure disease, including cancer, genetic diseases and infectious diseases. The US Food & Drug Administration (FDA) approved the first gene therapy in the United States in 2017. Since then, more than 20 cell and gene therapy products have been approved.

According to the Alliance for Regenerative Medicine, we could see five more gene therapies for rare diseases introduced to the U.S. market in 2023, including new treatments for sickle cell disease, Duchenne muscular dystrophy and hemophilia.

These therapies will challenge life sciences organizations to rethink their business models. How will they efficiently determine which patients are eligible for these therapies? How will they obtain the patient’s blood as part of the therapy? How will they contract with payers for reimbursement, given these therapies can cost upwards of $3M per treatment? How will they track outcomes from treatment for outcome-based agreements? These questions and many more spanning payment models, consumer experience, supply chain and manufacturing will need to be addressed.

A key driver in the growth of gene therapies and adoption of precision health is the growth and accessibility of next-generation DNA sequencing (NGS). NGS will become more mainstream, moving the science out of the lab to deliver improved patient care and outcomes at scale. NGS delivers ultra-high throughput, scalability and speed and has revolutionized the industry by enabling a wide variety of analyses across multiple applications at a level never before possible. This includes delivering whole-genome sequencing at accessible and practical price points for researchers, scientists, doctors and patients. An example is the new Illumina NovaSeq X sequencer released in September 2022, which is twice as fast as prior models and capable of sequencing 20,000 genomes per year at a cost of $200 per genome. As the price of sequencing genomes declines, the ability to support personalized healthcare and gene therapy at scale will continue to grow.

Strategy 4: Accelerate development and delivery of lifesaving therapies through decentralized clinical trials

Limitations of traditional clinical trials were amplified during the COVID-19 pandemic and have accelerated the use of decentralized clinical trials (DCTs). There is a clear need to improve study formats so broader, more equitable populations are accessed and included. New technologies will help integrate patient data points and derive holistic insights like never before. Life sciences organizations will increase their use of DCTs to run global studies and bring new therapies to market. We expect a record number of decentralized trials in 2023.

Key benefits of DCTs include:

◉ Faster recruitment. Participants can be identified and engaged without the need to travel and be evaluated in person.

◉ Improved retention. Participants are less likely to drop out of a trial due to the typical in-person requirements.

◉ Greater control, convenience and comfort. Participants are more comfortable engaging at home and at local patient care sites.

◉ Increased diversity. Participants in legacy trials lacked diversity and contributed to gaps in understanding of diseases.

As DCTs are more broadly adopted, designing trials around the patient experience will be critical to ensuring clear, transparent engagement and willing and active participation. Methodologies such as Enterprise Design Thinking can provide a useful framework. Likewise, integrating patient data from multiple sources such as electronic health and medical records, electronic data capture platforms, clinical data management systems, wearables and other digital technologies will require a more open approach to information sharing.

Quantum computing will enable more advanced DCT capabilities for recruitment, trial site selection, and optimization and patient surveillance. Quantum-based algorithms can outperform existing computer algorithms, enabling better analysis of integrated patient data at scale.

In the coming years, decentralized trials will become the norm, improving the ability to recruit, select and deliver clinical trials at scale, ensuring full and diverse populations are represented and lifesaving treatments are more quickly approved and launched.

Strategy 5: Explore AI-driven drug discovery

AI-driven drug discovery continues to gain momentum and achieve critical milestones. The first AI-designed drug candidate to enter clinical trials was reported by Exscientia in early 2020. Since then, companies such as Insilico Medicine, Evotec and Schrödinger have announced phase I trials. Several candidates have had their clinical development accelerated through AI-enabled solutions. Within drug companies focused on AI-based discovery, there is publicly available information on about 160 discovery programs, of which 15 products are reportedly in clinical development.

Some execs may think AI can be delivered through the “tool in the public cloud” or by a single team. From our experience working with life sciences companies, this is not the case. Achieving full value from AI requires transformation of the discovery process spanning new tech, new talent and new behaviors throughout the R&D organization.

The AI-driven discovery process delivers value across four dimensions: finding the right biological target, designing a small molecule as a preclinical candidate, improving success rates and delivering overall speed and efficiency.

Search for new biological targets

We see the research community and industry scientists pursuing integration of multiomics and clinical data with machine learning to achieve drug repositioning. Leveraging experimental data and literature analysis, it is possible to uncover new disease pathways and polypharmacological and protein interactions. Application of AI to imaging (and other diagnostic techniques that rigorously analyze phenotypic outputs) may offer opportunities to identify new biological targets. Some of our clients look to understand protein interactions, function and motion using traditional computation techniques as well as quantum computing.

Use new techniques to search for new molecules

Using a deep search technique, it is possible today to mine the research literature and published experimental data to predict new small molecule structures and how molecules will behave. This and other techniques can be used to predict pharmacokinetic and pharmacodynamic properties and help identify off-target effects.

Explore the promise of quantum computing

Since 2020, there have been numerous quantum-related activities and experiments in the field of life sciences, spanning genomics, clinical research and discovery, diagnostics, treatments and interventions. Quantum-driven machine learning, trained on diverse clinical and real-world data sets, has been applied to molecular entity generation, diagnostics, forecasting effectiveness and tailoring radiotherapy.

Strategy 6: Use digital engagement to increase sales efficiency, patient loyalty and adherence

For healthcare providers

Conventional face-to-face visits to healthcare providers (HCPs) have reached the limit of effectiveness. HCPs now expect personalized approaches and instant access to knowledge. Increased scrutiny by public authorities, along with COVID-19, disrupted a traditional approach where sales reps had HCP offices and hospitals as their second home. A virtual engagement model emerged that is less effective in its current form.

At the same time, industry sees the value of an omnichannel HCP engagement strategy: our analysis shows 5-10% higher satisfaction with a new HCP experience, 15-25% more effective marketing spend, 5-7% boost in active prescribers and up to 15% lift in recurring revenue depending on the indication.

Pharma companies have enough data on certain products to enable a personalized experience for HCPs. An analytics and AI-driven approach to engagement with clinicians provides the highest impact as it improves both their speed-to-decision and their awareness of the latest clinical evidence. Well-defined technology and data strategies, along with change management and talent identification programs, are key to success.

For patients

Adherence and persistence are major challenges in an industry that caters to chronic patients. Additionally, with new reimbursement models, payers incentivize “complete” cases that achieve prolonged remission or, for acute patients, functional recovery. To keep selling meds and getting paid for them, patients need to be taking them continuously. For many indications, patients have many pharmaceutical options. Successful companies will differentiate themselves in the market by offering digital support for their pharmaceuticals, engaging patients in their care on their smartphones through gamification and incentive programs.

How technology can advance new biologics for treatment of plague psoriasis

Bills and regulations will increase the adoption and application of AI

AI underpins the trends mentioned above. While AI technology has been around for decades, its adoption in life sciences has accelerated over the last several years, impacting drug development, clinical trials and supply chains. AI is infused into many of our daily interactions, from calling an airline to rebook tickets, asking Alexa to play music and turn on the lights, receiving an approval for a loan, to providing automated treatment recommendations to patients based on their clinical history and the latest treatment guidelines.

As AI continues to permeate our lives, oversight will be front and center. Both the United States and EU consider regulation to be essential to the development of AI tools that consumers can trust. Life sciences companies must understand the impact AI regulations have on their business models and that they play a proactive role in influencing these policies in the interest of better patient outcomes.

As an example, IBM’s Policy Lab takes a proactive approach to providing policymakers with a vision and actionable recommendations to harness the benefits of innovation while ensuring trust in a world being reshaped by data. IBM works with organizations and policymakers to share our perspective to support responsible innovations. One such bill was the Biden-Harris administration’s Blueprint for an AI Bill of Rights released in September 2022. As stated in the Bill, “AI systems have the potential to bring incredible societal benefits, but only if we do the hard work of ensuring AI products and services are safe and secure, accurate, transparent, free of harmful bias and otherwise trustworthy.” The Bill lays outs five commonsense protections to which everyone in America should be entitled in the design, development and deployment of AI and other automated technologies:

◉ Right to safe and effective systems. You should be protected from unsafe or ineffective systems.

◉ Algorithmic discrimination protections. You should not face discrimination by algorithms, and systems should be used and designed in an equitable way.

◉ Data privacy. You should be protected from abusive data practices via built-in protections and have agency over how your data is used.

◉ Notice and explanation. You should know that an automated system is being used and understand how and why it contributes to outcomes that impact you.

◉ Human alternatives, consideration and fallback. You should be able to opt out where appropriate and have access to a person who can quickly consider and remedy problems you encounter.

The application of AI is not slowing down, nor is scrutiny of it. Life sciences organizations will differentiate themselves by having a seat at the table. They will seek opportunity to influence AI-health policy and deliver ethical and responsible AI-powered solutions that augment their existing product portfolio and improve patient and provider experiences and healthcare outcomes at reduced costs.

Embrace new technologies to offer major advances

Life sciences companies, particularly in pharma and biotech, can prove resilient despite inflationary pressures. They must focus on business model specialization across innovation and invention, generics business and consumer health. Strong demand can help companies overcome business challenges and position the industry for steady innovation-led growth. It is crucial to embrace new technologies, particularly state-of-the-art computing and AI, to offer major advances that may represent a paradigm shift in drug discovery, clinical trial site optimization, and, ultimately, engagement with a person receiving care. Acting boldly in 2023 with a clearly articulated strategy and prioritization will set both mature life sciences organizations and new players on the right path. Companies that focus on strategy and innovation will be the biggest winners.

Source: ibm.com

Continue reading

AI can help in the fight against racism

In my role as Open Source Community Manager for the Call for Code for Racial Justice, I oversee a community of developers, data scientists, designers and general problem-solvers all looking to use technology to fight for racial justice. Just like any role, there are challenges I must deal with on a daily basis, but the one thing that has pleasantly surprised me since I started almost a year ago has been the interest and enthusiasm from people all around the world and from different backgrounds who are invested in advancing racial equity using data and artificial intelligence (AI).

The Call for Code for Racial Justice is an initiative external to IBM, so the people I deal with come from big and small organizations from around the globe — yet they all share this common belief and that drives them to give up weekends and work nights building tech for social good.

What is this community building to fight racial injustice?

We currently have seven projects in the Call for Code for Racial Justice. These were originally incubated by the black community inside of IBM as a response to the racial injustice highlighted through the #BlackLivesMatter campaign in 2020. When looking across these projects, you can see that there are certain areas where technology has the greatest opportunity to fight racial bias in society:

◉ Accessing information: When information is dense and difficult to consume, it tends to be hard for people to come together and use it in an effective way. This often happens in the government and policy space, where information can have a significant impact on our lives — especially for underserved communities. Policy related to schools, roads, availability of local shops and resources can often be written in legalese that is hard for people to comprehend. AI can help rectify this. The Legit-Info project utilized Watson Natural Language Understanding to identify titles, summaries, locations and impacts. The results can then be further curated to improve readability and make these meaningful to all members in a community.

◉ Identifying racial bias: Racial bias can creep into all kinds of places — from a police write-up of a crime to technical documentation on a software tool. In some cases, this may be explicit and driven by the bias of the individual writing the document, but in just as many cases, this may be implicit and the result of societal norms carried over from the past. TakeTwo is an API-based tool that can take a document as its input and highlight potential racial bias based on a trained machine learning model. Looking for insights in data is another way to identify racial bias — the Open Sentencing project looks specifically at incarceration rates based on racial demographics to help defense lawyers make the case for black defendants who often face tougher sentences for the same crimes as those committed by people of other races.

Why get involved in building AI solutions to fight racism?

In the case of these open-source projects, community involvement is as important as the technology itself. In a recent survey of community members, many were motivated by a desire for social good. Others were interested in networking and connecting with those sharing similar interests. The development of skills is also a big component — working with industry-leading technology and building skills that they can take into other areas of their lives.

For myself, starting as a contributor and progressing to Community Manager, I’ve experienced all these benefits, but there is another factor that is important when it comes to technology helping with social justice. After earning a post-graduate degree in Mechanical Engineering, I started my career as a product manager for AI products. One thing that has become clear as I have progressed through my career is the need to have the right people in the room when making all kinds of decisions. We need to ensure the AI systems we build are trustworthy. Beyond that, whether it’s the policies that impact communities, the products we build and how we market them or, indeed, almost any facet of our lives, we need proper representation and diversity of thought if we are to realize the dream of creating a more just society. AI has a growing role to play in the fight for social justice, but we can’t rely on it alone.

Get involved with the Call for Code for Racial Justice Projects

We are always looking for new participants in the Call for Code for Racial Justice Projects — find out more about how you can get involved.

Source: ibm.com

Continue reading