Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

So what is data science and why is it so topical? Is it just another fad that will fade away after the hype? We will start with a simple introduction to data science, defining what it is, why it matters, and why it matters now. This chapter will highlight the data science process with guidelines and best practices. It will introduce some of the most commonly used techniques and algorithms in data science. And it will explore ensemble models, a key technology on the cutting edge of data science.

What is Data Science?

Data science is the practice of obtaining useful insights from data. Although it also applies to small data, data science is particularly important for big data, as we now collect petabytes of structured and unstructured data from many sources inside and outside an organization. As a result, we are now data rich but information poor. Data science provides powerful processes and techniques for gleaning actionable information from this sea of data. Data science draws from several disciplines including statistics, mathematics, operations research, signal processing, linguistics, database and storage, programming, machine learning, and scientific computing. Figure 1-1 illustrates the most common disciplines of data science. Although the term data science is new in business, it has been around since 1960 when it was first used by Peter Naur to refer to data processing methods in computer science. Since the late 1990s notable statisticians such as C.F. Jeff Wu and William S. Cleveland have also used the term data science, a discipline they view as the same as or an extension of statistics.

Figure 1-1.
figure 1figure 1

Highlighting the main academic disciplines that constitute data science

Full size image

Practitioners of data science are data scientists, whose skills span statistics, mathematics, operations research, signal processing, linguistics, database and storage, programming, machine learning, and scientific computing. In addition, to be effective, data scientists also need good communication and data visualization skills. Domain knowledge is also important to deliver meaningful results fast. This breadth of skills is very hard to find in one person, which is why data science is a team sport, not an individual effort. To be effective, one needs to hire a team with complementary data science skills.

Analytics Spectrum

According to Gartner, all the analytics we do can be classified into one of four categories: descriptive, diagnostic, predictive, and prescriptive analysis. Descriptive analysis typically helps to describe a situation and can help to answer questions like What happened?, Who are my customers?, etc. Diagnostic analysis help you understand why things happened and can answer questions like Why did it happen? Predictive analysis is forward-looking and can answer questions such as What will happen in the future? As the name suggests, prescriptive analysis is much more prescriptive and helps answer questions like What should we do?, What is the best route to my destination?, or How should I allocate my investments?

Figure 1-2 illustrates the full analytics spectrum. It also shows the degree of sophistication in this diagram.

Figure 1-2.
figure 2figure 2

Spectrum of all data analysis

Full size image

Descriptive Analysis

Descriptive analysis is used to explain what is happening in a given situation. This class of analysis typically involves human intervention and can be used to answer questions like What happened?, Who are my customers?, How many types of users do we have?, etc. Common techniques used for this include descriptive statistics with charts, histograms, box and whisker plots, or data clustering. You’ll explore these techniques later in this chapter.

Diagnostic Analysis

Diagnostic analysis helps you understand why certain things happened and what are the key drivers. For example, a wireless provider would use this to answer questions such as Why are dropped calls increasing? or Why are we losing more customers every month? A customer diagnostic analysis can be done with techniques such as clustering, classification, decision trees, or content analysis. These techniques are available in statistics, data mining, and machine learning. It should be noted that business intelligence is also used for diagnostic analysis.

Predictive Analysis

Predictive analysis helps you predict what will happen in the future. It is used to predict the probability of an uncertain outcome. For example, it can be used to predict if a credit card transaction is fraudulent, or if a given customer is likely to upgrade to a premium phone plan. Statistics and machine learning offer great techniques for prediction. This includes techniques such as neural networks, decision trees, random forests, boosted decision trees, Monte Carlo simulation, and regression.

Prescriptive Analysis

Prescriptive analysis will suggest the best course of action to take to optimize your business outcomes. Typically, prescriptive analysis combines a predictive model with business rules (such as declining a transaction if the probability of fraud is above a given threshold). For example, it can suggest the best phone plan to offer a given customer, or based on optimization, can propose the best route for your delivery trucks. Prescriptive analysis is very useful in scenarios such as channel optimization, portfolio optimization, or traffic optimization to find the best route given current traffic conditions. Techniques such as decision trees, linear and non-linear programming, Monte Carlo simulation, or game theory from statistics and data mining can be used to do prescriptive analysis. See Figure 1-3.

Figure 1-3.
figure 3figure 3

A smart telco using prescriptive analytics

Full size image

The analytical sophistication increases from descriptive to prescriptive analytics. In many ways, prescriptive analytics is the nirvana of analytics and is often used by the most analytically sophisticated organizations. Imagine a smart telecommunications company that has embedded analytical models in its business workflow systems. It has the following analytical models embedded in its customer call center system:

  • A Customer Churn Model: This is a predictive model that predicts the probability of customer attrition. In other words, it predicts the likelihood that the customer calling the call center will ultimately defect to the competition.

  • A Customer Segmentation Model: This puts customers into distinct behavioral segments for marketing purposes.

  • A Customer Propensity Model: This model predicts the customer’s propensity to respond to each of the marketing offers, such as upgrades to premium plans.

When a customer calls, the call center system identifies him or her in real time from their cell phone number. Then the call center system scores the customer using these three models. If the customer scores high on the customer churn model, it means they are very likely to defect to the competitor. In that case, the telecommunications company will immediately route the customer to a group of call center agents who are empowered to make attractive offers to prevent attrition. Otherwise, if the segmentation model scores the customer as a profitable customer, he/she is routed to a special concierge service with shorter wait lines and the best customer service. If the propensity model scores the customer high for upgrades, the call agent is alerted and will try to upsell the customer with attractive upgrades. The beauty of this solution is that all of the models are baked into the telecommunication company’s business workflow, which enables their agents to make smart decisions that improve profitability and customer satisfaction. This is illustrated in Figure 1-3.

Why Does It Matter and Why Now?

Data science offers organizations a real opportunity to make smarter and timely decisions based on all the data they collect. With the right tools, data science offers you new and actionable insights not only from your own data, but also from the growing sources of data outside your organization, such as weather data, customer demographic data, consumer credit data from the credit bureaus, and data from social media sites such as Facebook, Twitter, Instagram, etc. Here are a few reasons why data science is now critical for business success.

Data as a Competitive Asset

Data is now a critical asset that offers a competitive advantage to smart organizations that use it correctly for decision making. McKinsey and Gartner agree on this: in a recent paper McKinsey suggests that companies that use data and business analytics to make decisions are more productive and deliver a higher return on equity than those who don’t. In a similar vein, Gartner posits that organizations that invest in a modern data infrastructure will outperform their peers by up to 20%. Big data offers organizations the opportunity to combine valuable data across silos to glean new insights that drive smarter decisions.

“Companies that use data and business analytics to guide decision making are more productive and experience higher returns on equity than competitors that don’t.”

—Brad Brown et al., McKinsey Global Institute, 2011

“By 2015, organizations integrating high-value, diverse, new information types and sources into a coherent information management infrastructure will outperform their industry peers financially by more than 20%.”

—Regina Casonato et al., Gartner1

Increased Customer Demand

Business intelligence has been the key type of analytics used by most organizations in the last few decades. However, with the emergence of big data, more customers are now eager to use predictive analytics to improve marketing and business planning. Traditional BI gives a good rear view analysis of their business, but does not help with any forward-looking questions that include forecasting or prediction.

The past two years have seen a surge of demand from customers for predictive analytics as they seek more powerful analytical techniques to uncover value from the troves of data they store on their businesses. In our combined experience, we have not seen as much demand for data science from customers as we did in the last two years alone!

Increased Awareness of Data Mining Technologies

Today a subset of data mining and machine learning algorithms are now more widely understood since they have been tried and tested by early adopters such as Netflix and Amazon, who actively use them in their recommendation engines. While most customers do not fully understand details of the machine learning algorithms used, their application in Netflix movie recommendations or recommendation engines at online stores are very salient. Similarly, many customers are now aware of the targeted ads that are now heavily used by most sophisticated online vendors. So while many customers may not know details of the algorithms used, they now increasingly understand their business value.

Access to More Data

Digital data has exploded in the last few years and shows no signs of abating. Most industry pundits now agree that we are collecting more data than ever before. According to IDC, the digital universe will grow to 35 zetabyes (i.e. 35 trillion terabytes) globally by 2020. Others posit that the world’s data is now growing by up to 10 times every 5 years, which is astounding. In a recent study, McKinsey Consulting also found that in 15 of the 17 US economic sectors, companies with over 1,000 employees store, on average, over 235 terabytes of data–which is more than the data stored by the US Library of Congress! This data explosion is driven by the rise of new data sources such as social media, cell phones, smart sensors, and dramatic gains in the computer industry. The rise of Internet of Things (IoT) only exacerbates this trend as more data is being generated than ever before by sensors. According to Cisco, there will be up to 50 billion connected devices by 2020!

The large volumes of data being collected also enable you to build more accurate predictive models. We know from statistics that the confidence interval (also known as the margin of error) has an inverse relationship with the sample size. So the larger your sample size, the smaller the margin of error. This in turn increases the accuracy of predictions from your model.

Faster and Cheaper Processing Power

We now have far more computing power at our disposal than ever before. Moore’s Law proposed that computer chip performance would grow exponentially, doubling every 18 months. This trend has been true for most of the history of modern computing. In 2010, the International Technology Roadmap for Semiconductors updated this forecast, predicting that growth would slow down in 2013 when computer densities and counts would double every 3 years instead of 18 months. Despite this, the exponential growth in processor performance has delivered dramatic gains in technology and economic productivity. Today, a smartphone’s processor is up to five times more powerful than that of a desktop computer 20 years ago. For instance, the Nokia Lumia 928 has a dual-core 1.5 GHz Qualcomm Snapdragon™ S4 that is at least five times faster than the Intel Pentium P5 CPU released in 1993, which was very popular for personal computers. In the nineties, expensive workstations like the DEC VAX mainframes or the DEC Alpha workstations were required to run advanced, compute-intensive algorithms. It is remarkable that today’s smartphone is also five times faster than the powerful DEC Alpha processor from 1994, whose speed was 200-300 MHz! Today you can run the same algorithms on affordable personal workstations with multi-core processors. In addition, you can leverage Hadoop’s MapReduce architecture to deploy powerful data mining algorithms on a farm of commodity servers at a much lower cost than ever before. With data science we now have the tools to discover hidden patterns in our data through smart deployment of data mining and machine learning algorithms.

We have also seen dramatic gains in capacity, and an exponential reduction in the price of computer memory. This is illustrated in Figures 1-4 and 1-5, which show the exponential price drop and growth in capacity of computer memory since 1960. Since 1990, the average price per MB of memory has dropped from $59 to a meager 0.49 cents–a 99.2% price reduction! At the same time, the capacity of a memory module has increased from 8MB to a whopping 8GB! As a result, a modest laptop is now more powerful than a high-end workstation from the early nineties.

Figure 1-4.
figure 4figure 4

Average computer memory price since 1960

Full size image
Figure 1-5.
figure 5figure 5

Average computer memory size since 1960

Full size image

Note

More information on memory price history is available at John C. McCallum at www.jcmit.com/mem2012.htm .

The Data Science Process

A typical data science project follows the five-step process outlined in Figure 1-6. Let’s review each of these steps in detail.

Figure 1-6.
figure 6figure 6

Overview of the data science process

Full size image
  1. 1.

    Define the business problem: This is critical as it guides the rest of the project. Before building any models, it is important to work with the project sponsor to identify the specific business problem he or she is trying to solve. Without this, one could spend weeks or months building sophisticated models that solve the wrong problem, leading to wasted effort. A good data science project gleans good insights that drive smarter business decisions. Hence the analysis should serve a business goal. It should not be a hammer in search of a nail! There are formal consulting techniques and frameworks (such as guided discovery workshops and six sigma methodology) used by practitioners to help business stakeholders prioritize and scope their business goals.

  2. 2.

    Acquire and prepare data: This step entails two activities. The first is the acquisition of raw data from several source systems including databases, CRM systems, web log files, etc. This may involve ETL (extract, transform, and load) processes, database administrators, and BI personnel. However, the data scientist is intimately involved to ensure the right data is extracted in the right format. Working with the raw data also provides vital context, which is required downstream.

Second, once the right data is pulled, it is analyzed and prepared for modelling. This involves addressing missing data, outliers in the data, and data transformations. Typically, if a variable has over 40% of missing values, it can be rejected, unless the fact that it is missing (or not) conveys critical information. For example, there might be a strong bias in the demographics of who fills in the optional field of “age” in a survey. For the rest, we need to decide how to deal with missing values; should we impute with the average value, median, or something else? There are several statistical techniques for detecting outliers. With a box and whisker plot, an outlier is a sample (value) greater or smaller than 1.5 times the interquartile range (IQR). The interquartile range is the 75th percentile-25th percentile. We need to decide whether to drop an outlier or not. If it makes sense to keep it, we need to find a useful transformation for the variable. For instance, log transformation is generally useful for transforming incomes.

Correlation analysis, principal component analysis, or factor analysis are useful techniques that show the relationships between the variables. Finally, feature selection is done at this stage to identify the right variables to use in the model in the next step.

This step can be laborious and time-consuming. In fact, in a typical data science project, we spend up to 75 to 80% of time in data acquisition and preparation. That said, this is the vital step that coverts raw data into high quality gems for modelling. The old adage is still true: garbage in, garbage out. Investing wisely in data preparation improves the success of your project. Chapter 3 provides more details on the data preparation phase.

  1. 3.

    Develop the model: This is the most fun part of the project where we develop the predictive models. In this step, we determine the right algorithm to use for modeling given the business problem and data. For instance, if it is a binary classification problem, we can use logistic regression, decision trees, boosted decision trees, or neural networks. If the final model has to be explainable, this rules out algorithms like boosted decision trees. Model building is an iterative process: we experiment with different models to find the most predictive one. We also validate it with the customer a few times to ensure it meets their needs before exiting this stage.

  2. 4.

    Deploy the model: Once built, the final model has to be deployed in production where it will be used to score transactions or by customers to drive real business decisions. Models are deployed in many different ways depending on the customer’s environment. In most cases, deploying a model involves implementing the data transformations and predictive algorithm developed by the data scientist in order to integrate with an existing decision management platform. Needless to say, it is a cumbersome process today. Azure Machine Learning dramatically simplifies model deployment by enabling data scientists to deploy their finished models as web services that can be invoked from any application on any platform, including mobile devices.

  3. 5.

    Monitor model’s performance: Data science does not end with deployment. It is worth noting that every statistical or machine learning model is only an approximation of the real world, and hence is imperfect from the very beginning. When a validated model is tested and deployed in production, it has to be monitored to ensure it is performing as planned. This is critical because any data-driven model has a fixed shelf life. The accuracy of the model degrades with time because fundamentally the data in production will vary over time for a number of reasons (such as the business may launch new products to target a different demographic). For instance, the wireless carrier we discussed earlier may choose to launch a new phone plan for teenage kids. If they continue to use the same churn and propensity models, they may see a degradation in their models’ performance after the launch of this new product. This is because the original dataset used to build the churn and propensity models did not contain significant numbers of teenage customers. With close monitoring of the model in production we can detect when its performance starts to degrade. When its accuracy degrades significantly, it is time to rebuild the model by either re-training it with the latest dataset including production data, or completely rebuilding it with additional datasets. In that case, we return to Step 1 where we revisit the business goals and start all over.

How often should we rebuild a model? The frequency varies by business domain. In a stable business environment where the data does not vary too quickly, models can be rebuilt once every year or two. A good example is retail banking products such as mortgages and car loans. However, in a very dynamic environment where the ambient data changes rapidly, models can be rebuilt daily or weekly. A good case in point is the wireless phone industry, which is fiercely competitive. Churn models need to be retrained every few days since customers are being lured by ever more attractive offers from the competition.

Common Data Science Techniques

Data science offers a large body of algorithms from its constituent disciplines, namely statistics, mathematics, operations research, signal processing, linguistics, database and storage, programming, machine learning, and scientific computing. We organize these algorithms into the following groups for simplicity:

  • Classification

  • Clustering

  • Regression

  • Simulation

  • Content Analysis

  • Recommenders

Chapter 6 provides more details on some of these algorithms.

Classification Algorithms

Classification algorithms are commonly used to classify people or things into one of many groups. They are also widely used for predictions. For example, to prevent fraud, a card issuer will classify a credit card transaction as either fraudulent or not. The card issuer typically has a large volume of historical credit card transactions and knows the status of each of these transactions. Many of these cases are reported by the legitimate cardholder who does not want to pay for unauthorized charges. So the issuer knows whether each transaction was fraudulent or not. Using this historical data the issuer can now build a model that predicts whether a new credit card transaction is likely to be fraudulent or not. This is a binary classification problem in which all cases fall into one of two classes.

Another classification problem is the customers’ propensity to upgrade to a premium phone plan. In this case, the wireless carrier needs to know if a customer will upgrade to a premium plan or not. Using sales and usage data, the carrier can determine which customers upgraded in the past. Hence they can classify all customers into one of two groups: whether they upgraded or not. Since the carrier also has information on demographic and behavioral data on new and existing customers, they can build a model to predict a new customer’s probability to upgrade; in other words, the model will group each customer into one of two classes.

Statistics and data mining offer many great tools for classification: this includes logistic regression, which is widely used by statisticians for building credit scorecards, or propensity-to-buy models, or neural networks algorithms such as backpropagation, radial basis functions, or ridge polynomial networks. Others include decision trees or ensemble models such as boosted decision trees or random forests. For more complex classification problems with more than two classes, you can use multimodal techniques that predict multiple classes. Classification problems generally use supervised learning algorithms that use label data for training. Azure Machine Learning offers several algorithms for classification including logistic regression, decision trees, boosted decision trees, multimodal neural networks, etc. See Chapter 6 for more details.

Clustering Algorithms

Clustering uses unsupervised learning to group data into distinct classes. A major difference between clustering and classification problems is that the outcome of clustering is unknown beforehand. Before clustering we do not know the cluster to which each data point belongs. In contrast, with classification problems we have historical data that shows the class to which each data point belongs. For example, the lender knows from historical data whether a customer defaulted on their car loans or not.

A good application of clustering is customer segmentation where we group customers into distinct segments for marketing purposes. In a good segmentation model, the data within each segment is very similar. However, data across different segments is very different. For example, a marketer in the gaming segment needs to understand his or her customers better in order to create the right offers for them. Let’s assume that he or she only has two variables on the customers: age and gaming intensity. Using clustering, the marketer finds that there are three distinct segments of gaming customers, as shown in Figure 1-7. Segment 1 is the intense gamers who play computer games passionately every day and are typically young. Segment 2 is the casual gamers who only play occasionally and are typically in their thirties or forties. The non-gamers rarely ever play computer games and are typically older; they make up Segment 3.

Figure 1-7.
figure 7figure 7

Simple hypothetical customer segments from a clustering algorithm

Full size image

Statistics offers several tools for clustering, but the most widely used is the k-means algorithm that uses a distance metric to cluster similar data together. With this algorithm you decide a priori how many clusters you want; this is the constant K. If you set K = 3, the algorithm produces three clusters. Refer to Haralambos Marmanis and Dmitry Babenko’s book for more details on the k-means algorithm. Machine learning also offers more sophisticated algorithms such as self-organizing maps (also known as Kohonen networks) developed by Teuvo Kohonen, or adaptive resonance theory (ART) networks developed by Stephen Grossberg and Gail Carpenter. Clustering algorithms typically use unsupervised learning since the outcome is not known during training.

Note

You can read more about clustering algorithms in the following books and paper:

Haralambos Marmanis and Dmitry Babenko, Algorithms of the Intelligent Web (Stamford, CT: Manning Publications Co., January 2011).

T. Kohonen, Self-Organizing Maps. Third, extended edition (Springer, 2001).

“Art2-A: an adaptive resonance algorithm for rapid category learning and recognition”, Carpenter, G., Grossberg, S., and Rosen, D. Neural Networks, 4:493-504. 1991a.

Regression Algorithms

Regression techniques are used to predict response variables with numerical outcomes. For example, a wireless carrier can use regression techniques to predict call volumes at their customer service centers. With this information they can allocate the right number of call center staff to meet demand. The input variables for regression models may be numeric or categorical. However, what is common with these algorithms is that the output (or response variable) is typically numeric. Some of the most commonly used regression techniques include linear regression, decision trees, neural networks, and boosted decision tree regression.

Linear regression is one of the oldest prediction techniques in statistics and its goal is to predict a given outcome from a set of observed variables. A simple linear regression model is a linear function. If there is only one input variable, the linear regression model is the best line that fits the data. For two or more input variables, the regression model is the best hyperplane that fits the underlying data.

Artificial neural networks are a set of algorithms that mimic the functioning of the brain. They learn by example and can be trained to make predictions from a dataset even when the function that maps the response to independent variables is unknown. There are many different neural network algorithms, including backpropagation networks, Hopfield networks, Kohonen networks (also known as self-organizing maps), and adaptive resonance theory (or ART) networks. However, the most common is backpropagation, also known as multilayered perceptron. Neural networks are used for regression or classification.

Decision tree algorithms are hierarchical techniques that work by splitting the dataset iteratively based on certain statistical criteria. The goal of decision trees is to maximize the variance across different nodes in the tree, and minimize the variance within each node. Some of the most commonly used decision tree algorithms include Iterative Dichotomizer 3 (ID3), C4.5 and C5.0 (successors of ID3), Automatic Interaction Detection (AID), Chi-Squared Automatic Interaction Detection (CHAID), and Classification and Regression Tree (CART). While very useful, the ID3, C4.5, C5.0, and CHAID algorithms are classification algorithms and are not useful for regression. The CART algorithm, on the other hand, can be used for either classification or regression.

Simulation

Simulation is widely used across many industries to model and optimize processes in the real world. Engineers have long used mathematical techniques like finite elements or finite volumes to simulate the aerodynamics of aircraft wings or cars. Simulation saves engineering firms millions of dollars in R&D costs since they no longer have to do all their testing with real physical models. In addition, simulation offers the opportunity to test many more scenarios by simply adjusting variables in their computer models.

In business, simulation is used to model processes like optimizing wait times in call centers or optimizing routes for trucking companies or airlines. Through simulation, business analysts can model a vast set of hypotheses to optimize for profit or other business goals.

Statistics offers many powerful techniques for simulation and optimization. One method, the Markov chain analysis, can be used to simulate state changes in a dynamic system. For instance, it can be used to model how customers will flow through a call center: how long will a customer wait before dropping off, or what are their chances of staying on after engaging the interactive voice response (IVR) system? Linear programming is used to optimize trucking or airline routes, while Monte Carlo simulation is used to find the best conditions to optimize for given business outcome such as profit.

Content Analysis

Content analysis is used to mine content such as text files, images, and videos for insights. Text mining uses statistical and linguistic analysis to understand the meaning of text. Simple keyword searching is too primitive for most practical applications. For example, to understand the sentiment of Twitter feed data with a simple keyword search is a manual and laborious process because you have to store keywords for positive, neutral, and negative sentiments. Then, as you scan the Twitter data, you score each Twitter feed based on the specific keywords detected. This approach, though useful in narrow cases, is cumbersome and fairly primitive. The process can be automated with text mining and natural language processing (NLP), which mines the text and tries to infer the meaning of words based on context instead of simple keyword search.

Machine learning also offers several tools for analyzing images and videos through pattern recognition. Through pattern recognition, we can identify known targets with face recognition algorithms. Neural network algorithms such as multilayer perceptron and ART networks can be used to detect and track known targets in video streams, or to aid analysis of X-ray images.

Recommendation Engines

Recommendation engines have been used extensively by online retailers like Amazon to recommend products based on users’ preferences. There are three broad approaches to recommendation engines. Collaboration filtering (CF) makes recommendations based on similarities between users or items. With item-based collaborative filtering, we analyze item data to find which items are similar. With collaborative filtering, that data is specifically the interactions of users with the movies, such as ratings or viewing, as opposed to characteristics of the movies such as genre, director, and actors. So whenever a customer buys a movie from this set we recommend others based on similarity.

The second class of recommendation engines makes recommendations by analyzing the content selected by each user. In this case, text mining or natural language processing techniques are used to analyze content such as document files. Similar content types are grouped together, and this forms the basis of recommendations to new users. More information on collaborative filtering and content-based approaches are available in Haralambos Marmanis and Dmitry Babenko’s book.

The third approach to recommendation engines uses machine learning algorithms to determine product affinity. This approach is also known as market basket analysis. Algorithms such as Naïve Bayes, the Microsoft Association Rules, or the Arules package in R are used to mine sales data to determine which products sell together.

Cutting Edge of Data Science

Let’s conclude this chapter with a quick overview of ensemble models that are at the cutting edge of data science.

The Rise of Ensemble Models

Ensemble models are a set of classifiers from machine learning that use a panel of algorithms instead of a single one to solve classification problems. They mimic our human tendency to improve the accuracy of decisions by consulting knowledgeable friends or experts. When faced with important decisions such as a medical diagnosis, we tend to seek a second opinion from other doctors to improve our confidence. In the same way, ensemble models use a set of algorithms as a panel of experts to improve the accuracy and reduce the variance of classification problems.

The machine learning community has worked on ensemble models for decades. In fact, seminal papers were published as early as 1979 by Dasarathy and Sheela. However, since the mid-1990s, this area has seen rapid progress with several important contributions resulting in very successful real-world applications.

Real-World Applications of Ensemble Models

In the last few years, ensemble models have been found in great real-world applications including face recognition in cameras, bioinformatics, Netflix movie recommendations, and Microsoft’s Xbox Kinect. Let’s examine two of these applications.

First, ensemble models were very instrumental to the success of the Netflix Prize competition. In 2006, Netflix ran an open contest with a $1 million prize for the best collaborative filtering algorithm that improved their existing solution by 10%. In September 2009, the $1 million prize was awarded to BellKor’s Pragmatic Chaos, a team of scientists from AT&T Labs joining forces with two lesser known teams. At the start of the contest, most teams used single classifier algorithms: although they outperformed the Netflix model by 6–8%, performance quickly plateaued until teams started applying ensemble models. Leading contestants soon realized that they could improve their models by combining their algorithms with those of the apparently weaker teams. In the end, most of the top teams, including the winners, used ensemble models to significantly outperform Netflix’s recommendation engine. For example, the second-place team, aptly named The Ensemble, used more than 900 individual models in their ensemble.

Microsoft’s Xbox Kinect sensor also uses ensemble modeling. Random Forests, a form of ensemble model, is used effectively to track skeletal movements when users play games with the Xbox Kinect sensor.

Despite success in real-world applications, a key limitation of ensemble models is that they are black boxes in that their decisions are hard to explain. As a result, they are not suitable for applications where decisions have to be explained. Credit scorecards are a good example because lenders need to explain the credit score they assign to each consumer. In some markets, such explanations are a legal requirement and hence ensemble models would be unsuitable despite their predictive power.

Building an Ensemble Model

There are three key steps to building an ensemble model: a) selecting data, b) training classifiers, and c) combining classifiers.

The first step to build an ensemble model is data selection for the classifier models. When sampling the data, a key goal is to maximize diversity of the models, since this improves the accuracy of the solution. In general, the more diverse your models, the better the performance of your final classifier, and the smaller the variance of its predictions.

Step 2 of the process entails training several individual classifiers. But how do you assign the classifiers? Of the many available strategies, the two most popular are bagging and boosting. The bagging algorithm uses different subsets of the data to train each model. The Random Forest algorithm uses this bagging approach. In contrast, the boosting algorithm improves performance by making misclassified examples in the training set more important during training. So during training, each additional model focuses on the misclassified data. The boosted decision tree algorithm uses the boosting strategy.

Finally, once you train all the classifiers, the final step is to combine their results to make a final prediction. There are several approaches to combining the outcomes, ranging from a simple majority to a weighted majority voting.

Ensemble models are a really exciting part of machine learning, and they offer the potential for breakthroughs in classification problems.

Summary

This chapter introduced data science, defining what it is, why it matters, and why it matters now. We outlined the key academic disciplines of data science, including statistics, mathematics, operations research, signal processing, linguistics, database and storage, programming, and machine learning. We covered the key reasons behind the heightened interest in data science: increasing data volumes, data as a competitive asset, growing awareness of data mining, and hardware economics.

A simple five-step data science process was introduced with guidelines on how to apply it correctly. We also introduced some of the most commonly used techniques and algorithms in data science. Finally, we introduced ensemble models, which is one of the key technologies on the cutting edge of data science.

Bibliography

  1. 1.

    Alexander Linden, 2014. “Key trends and emerging technologies in advanced analytics.” Gartner BI Summit 2014, Las Vegas, USA.

  2. 2.

    “Are you ready for the era of Big Data?”, McKinsey Global Institute - Brad Brown, Michael Chui, and James Manyika, October 2011.

  3. 3.

    “Information Management in the 21st Century”, Gartner - Regina Casonato, Anne Lapkin, Mark A. Beyer, Yvonne Genovese, Ted Friedman, September 2011.

  4. 4.

    John C. McCallum, www.jcmit.com/mem2012.htm .

  5. 5.

    Marmanis, Haralambos and Dmitry Babenko, Algorithms of the Intelligent Web (Stamford, CT: Manning Publications Co., January 2011).

  6. 6.

    Kohonen, T., Self-Organizing Maps. Third, extended edition (Springer, 2001).

  7. 7.

    “Art2-A: an adaptive resonance algorithm for rapid category learning and recognition”, Carpenter, G., Grossberg, S., and Rosen, D. Neural Networks, 4:493-504. 1991a.

  8. 8.

    MacLennan, Jamie, ZhaoHui Tang, and Bogdan Crivat, Data Mining with Microsoft SQL Server 2008 (Indianapolis, Indiana: Wiley Publishing Inc., 2009).