# Prep with 19 machine learning interview questions and answers

## Aspiring machine learning job candidates should be fluent in varied aspects of machine learning, from statistical theory and programming concepts to general industry knowledge.

Employees with machine learning skills are in high demand. Companies in a variety of industries are investing in big data and artificial intelligence (AI). Machine learning is an effective analytics technique that allows AI to effectively interact with and process big data, making it essential. Whether applying for a job a data engineer, machine learning engineer or data scientist (whose roles are changing with the proliferation of automated machine learning platforms), applicants should have a fairly balanced understanding of the technical aspects of machine learning and their application in the enterprise. Interviewer questions for machine learning jobs -- whatever the title may be -- generally test:

- Your understanding of the mathematical theories and algorithms that machine learning is based on.
- Your ability to apply those theories through programming.
- Your ability to apply those programs to real-world business problems.
- Your general, broader interest in the industry of machine learning.

Interviewers also may evaluate your knowledge of data analysis and data science, your approaches to novel, open-ended problems, and your soft skills and cultural fit with the company.

To be a strong machine learning job candidate, you should have strong answers prepared for questions that touch on each of these categories and prove you are a competent, curious systems thinker who can continuously improve as the field evolves. Interview formats vary, so it's important to have concise answers that indicate your extensive knowledge of the topic without including too many details -- but be able to expand if necessary. Ideally, you already have significant knowledge of machine learning topics. These questions can help you synthesize that knowledge concisely in an interview format and help you review topics that you're less proficient in.

Below is a list of commonly asked machine learning questions and accompanying answers that you can study and review to prepare for your interview.

This article is part of

### In-depth guide to machine learning in the enterprise

### Theory

#### What is the difference between supervised and unsupervised machine learning?

The main difference between supervised learning and unsupervised learning is the type of training data used. Supervised machine learning algorithms use labeled data, whereas unsupervised learning algorithms train on unlabeled data.

Labeled data contains predefined variables that guide the algorithm in learning -- it knows to some degree what it is looking for. Unlabeled data does not have predefined variables or features, so the algorithm scans the data set for any meaningful connection. Basically, the supervised algorithm learns with the help of a teacher, or supervisor, and the unsupervised algorithm learns alone. Supervised algorithms can generally produce results with a high degree of confidence but take a lot of manual effort in the process of labeling the training data. Unsupervised algorithms may be less accurate but can also be implemented with less initial work.

Semi-supervised learning is a combination of both methods in which unsupervised learning algorithms generate the labels that are fed to supervised learning algorithms. In this method, there is some human involvement in augmenting the labels that the unsupervised algorithm initially generates.

Which method to use depends on the results you are trying to produce and the data you have available. Certain types of data are easier to label than other types.

Together, supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning make up the four main types of machine learning algorithms.

#### What is the relationship between "variance and "bias"?

Machine learning bias is when an algorithm produces prejudiced results because of erroneous assumptions made during the ML process. It is error due to oversimplification.

Variance is error due to overcomplexity in a learning algorithm, which causes too much noise in the test data.

The relationship between these two terms is sometimes framed as the "bias-variance dilemma" or the "bias-variance tradeoff" and is considered a main problem in supervised learning models, partially due to the fact that unconscious human biases enter the ML process when humans label the training data. Basically, the lower the bias in an algorithm, the higher the variance, and vice versa. An algorithm with high variance but low bias may represent the training data set well but may also overfit to unrepresentative data. An algorithm with high levels of bias may underfit its training data and fail to capture important characteristics of the training set because it is oversimplifying.

The way to reduce bias is to make a model more complex and add variables (features), and to do the opposite to reduce variance. A method known as the bias-variance decomposition is used to evaluate an algorithm's expected generalization error. Another method -- known as the kernel trick -- can be used to make complex data sets more manageable.

In statistical theory, if you were to take a poll of people (for instance, leading up to an election) you are sampling a portion of the overall population. One of the first issues that you face in performing such a poll is whether the sample represents the overall population. For instance, if a poll is conducted using land-line phones, this may exclude a significant percentage of the population who only use cell phones, and if land-line phones are found to correlate with one party over another, this could skew the results systematically in one direction or another. This is known as *bias*.

At the time of that poll, a person could be uncertain about which candidate they are going to choose and they could change their mind later. Put another way, the signal from that particular person is weak, and if that's true of the overall electorate, then the uncertainty about the poll's accuracy is likely higher. This uncertainty is known as *variance* or *variability*.

Machine learning data faces the same issues. When the sample set (the learning set) comes from a biased source, meaning that sample set was not representative, then you will have a strong signal that's likely to deviate from reality. You can compensate for this by tweaking the model to weaken the bias, but you are still very much beholden to the quality of the inbound data.

When you have a high degree of variance, and hence low signal strength, this usually means that the margin of error is likely to be higher. Increasing the overall sample size can help, as can attempting to boost the signal by adding correlative factors (e.g., socioeconomic status, gender, geographical location or ethnicity) that may better give an indication of how they are likely to vote.

Note that bias and variance influence one another. Bias can be thought of as the deviance of the sample mean (the average of the sample in all dimensions) as compared to the population mean. In an ideal world, there should be no difference between these two means, but reality is usually not that forgiving, so typically data scientists should assume that there is likely a normal or bell curve that determines the probability that the sample mean is likely to be off by a certain amount. The variance determines the how wide the bell curve is. Both of these factors ultimately determine the margin of error, though bias generally cannot be measured directly.

#### What is the difference between "probability" and "likelihood"?

Roll a six-sided dice. The chance of one number coming up (say a 3) is one in six (assuming a fair dice). If you graphed this distribution based upon frequency of occurrence, the graph will be pretty monotonous -- a horizontal line of unit height. This is called a *uniform* distribution. If you paint one dice blue and the other red, then roll them together, the distribution for each color will be the same. The dice are considered *independent* in that they do not affect one another.

However, when you roll those two die together and add their sum, the sum depends equally on each, so that you are now measuring the number of combinations that will add up this distribution. This distribution will look like a triangle, (2:1,3:2,4:3,5:4,6:5,7:6,8:5,9:4,10:3,11:2,12:1), where the first number in each pair is the sum, and the second is the number of combinations. With three dice, the distribution begins to approach a bell curve (or more properly a binomial distribution). If you divide the total number of states, in this case 6x6 (36) and 6x6x6 (216) respectively, you get what's called a *normalized* distribution: The total of each fraction of the whole will be equal to 1.

As the number of die (each independent but of the same number of faces) becomes larger and larger, this eventually creates a continuous curve or envelope called a Gaussian distribution. Gaussian distributions are significant because they pop up when you are measuring a certain property that may have many independent states.

It also turns out that if you take such a distribution and measure the area underneath it (through calculus), the highest point in the curve (on the y axis) represents even probability -- 50% of the population will be on one side of the symmetric relationship, the other 50% will be on the other, with the probability of a given sample being proportional to its standard deviation from the mean, where a standard deviation is essentially the square root of the sum of the squares of the differences of the variances from the mean.

This means that if you know the mean height of an adult male (about 5'10") and you can show that 67% of the population (1 standard deviation past the mean) are 6'3" or under, then you can use it to predict the number of men who will be 6'9" or 4'10". This number is, for a given distribution, the *probability* that a given sample will be of that value.

Note also that if the distribution is not quite symmetric (perhaps in the die example that someone weights all die to preferentially show 6), then the skew of the distribution curve can be one way to measure the bias in a given sample.

Likelihood is a related concept but is not quite the same thing. A good example of understanding likelihood (at least to American readers) is the question of "if a given candidate wins a certain state, such as Texas or California, in the election, what is the likelihood that they will become president?" This is a surprisingly complex problem, because typically, states are very seldom independent races. Instead, the chances are good that if you win California, then you will likely also win Washington, Oregon, Hawaii, New York, Massachusetts and several others. Similarly, winning Texas will bring with it a different constellation of states.

In other words, unlike in the first case where your variables are independent (there is no correlation or anti-correlation), here the variables are typically interconnected. Mathematicians refer to the first as saying that the variables are *orthogonal* -- no one variable contributes to the value of another variable -- while in the second case they are not. Orthogonal variables are usually easy to decompose into independent problems (or dimensions) while nonorthogonal variables almost always have hidden variables and dependencies.

So likelihoods can usually be stated as, "Given event A, what is the likelihood that event B will occur if event B depends upon event A?" This is the foundation for Bayesian analysis, based upon Bayes Theorem. Additionally, Bayes Theorem in term has dependency upon Markov Chains (or Markov Graphs), which basically perform complex analysis by determining the values of associated nodes in a graph as weighted probabilities. Such solutions can be solved using statistical algorithms, weighting coefficients in neural nets, or through the use of graph theory. If the terms probability and given are contained in the same problem, it's usually a pretty good likelihood that Bayes Theorem is involved.

#### What is the curse of dimensionality and how would you respond to it?

Most modern data science and machine learning makes use of the term *features* instead of *variables*, but the idea is much the same. Most statistical problems come down to determining the number of variables (features) that are relevant to understanding the problem, then converting the values into numeric ranges between zero and one (or occasionally -1 and 1). This is the process of normalization. The process of optimizing variables, normalizing them, and determining which are relevant is known as feature engineering.

This makes a bit more sense when talking about ranges from -1 to 1 (although most algorithms use 0 to 1), but for purposes of illustration, let us use the larger range. What this means is that if you define the height of a person, then a height of 1 indicates the tallest people in that range, while a height of -1 correlates to the shortest (or anti-correlates to the longest), and 0 (the average) then becomes the midpoint. These are typically known as normalized vectors.

In an ideal world (mathematically), each vector is treated as being orthogonal to other vectors. If one vector goes from left to right, a second vector from bottom to top, and a third goes from it to out. These are called dimensions, though they're not necessarily the same thing as the three spatial dimensions that we are aware of. They may represent things like height, age and political orientation. This means that you can represent this cluster of traits -- or features -- as a point in this three-dimensional space for a given individual, or a zero-based vector emanating as a deviation from the mean.

Mathematicians are used to thinking in more than three dimensions. A four-dimensional vector is simply another orthogonal value to the other three. We can't visualize it directly in three dimensions, but we can create shadows (or projections) of all four of these vectors upon a plane or three-dimensional space, typically using other attributes such as color, size or texture of the represented point. Five dimensions are harder to visualize this way, but abstractly the idea is the same. In a machine learning context, it is possible to have dozens, hundreds, or even millions of such features. A computer simply treats these as arrays (or sequences) of the corresponding number of dimensions.

Typically, one of the goals of this kind of analysis is to determine clusters where similar individuals occupy the same general proximity in this "phase space." This is used heavily for classification purposes, defining neighborhoods that identify that the clustered data points identify a class. In a small number of dimensions, such clusters usually are self-evident, but as you increase the number of dimensions, there may be features that are totally opposite one another that increase the apparent distance between those points. For instance, let's say that you have a feature set that identifies characteristics that distinguish between cats and dogs.

However, one of those features is gender, which in fact may be central to the definition of the animal in question, but which places at least half of all elements clear across phase space from the other half. The clustering becomes obvious if you take that feature out, but if you don't know a priori what a given feature actually means (and you may not) then you may be hard pressed to say that male and female cats are both cats.

This is another way of saying that as you add more features, the points that represent these individuals can move so far away from one another that they can't form useful neighborhoods. The vector space is said to become *sparse* at that point. You have too many dimensions, which is where the term *curse of dimensionality* comes from.

The best thing to do in this case is to cut down the number of dimensions in your model. It may sound impressive to say that you are working on a model with a million dimensions, but most of those dimensions likely do not contribute significantly to the overall model accuracy but do have a major impact upon how long (how much data) it takes for a model to be trained. Analyses can be done on manifolds that can determine whether a given manifold (a space made from a smaller number of dimensions) provides meaningful impact to the model overall. Common methods for dimensionality reduction include principal component analysis, backward feature elimination and forward feature selection.

This is especially true in unsupervised learning systems, Sometimes, clusters form that are random noise that don't really have any significance in the model at all. In general, by reducing the number of dimensions by analysis of manifold impact, you can get down to a point where a model not only provides more meaningful data, but it does so requiring less data be processed in a shorter period of time.

#### Can you explain the difference between a discriminative and generative ML model?

Let's revisit the concept of distributions again. A distribution can the thought of as the continuous version of a frequency diagram. In taking the sum of three dice, for instance, there is a minimum value (3) and a maximum value (18) that constitute buckets. The 3 bucket has only one potential configuration (all three dice have a value of 1). The 18 bucket is the same -- in both cases the probability of rolling a 3 or an eighteen is 1/216, a shade under half of one percent. In the middle, there are also two buckets, buckets 10 and 11, that each have 36 possible configurations that will add up to 10 or 11 respectively, so the probability there 36/216 = 1/6=16.67%.

If you view each collection of buckets as a feature, then the probability associated with the bucket can be determined from the frequency distribution divided by the total number of states (or the normalized equivalent for a continuous distribution). In the case of a uniform distribution (such as rolling one red dice and comparing it with the role of another blue dice), this calculation becomes trivial -- it is (1/r)*(1/b) where r and b are the number of pips on each dice.

Suppose, however, that you and I each had a set of three dice each, yours red and mine blue. The probability that your red dice would roll a 10 and my blue dice would roll an 18 would be P(10)*P(18) = 1/6 * 1/216 = 1/1296. On the other hand, the probability of my rolling a ten and you rolling an eleven is much higher. P(10)*P(11)= 1/6 *1/6 = 1/36 (or thirty six times more likely). This is called a *bivariate* distribution, and is usually written as P(X,Y) where X and Y vary over their respective ranges.

When generalized to N different features, this is referred to as *multivariate distributions*, or more frequently, as *joint distributions*. At any given point in this N-dimensional feature space P(X_{1},X_{2},…,X_{N}), the associated probability will be the product of all of the independent probabilities at those respective points, what's often referred to as the *probability density. *

A *generative* machine learning model is one where the functions that determine the various coefficients of that model are determined by using the overall joint distribution of all of the features. This can be comprehensive, because each feature fully informs the overall probability, but it is also computationally expensive, as you also have to calculate the probability density across a mesh of N dimensions, which increased by the number of points in each message (say you sample across 10 points in any given dimension) to the Nth power. A model with 10 dimensions would take 10 to the 10^{th} power calculations (about 10 gigaflops), which is quite doable (about 1 microsecond) on something like the recently announced Xbox X Series GPUs, which is capable of 13 teraflops. A model with 15 features or dimensions, though, would take 10 minutes to do the same thing, and a model with 100 dimensions would take longer than the universe is likely to be in existence.

A *discriminative* model, however, uses Bayesians to partition the problem domain in different ways, reducing the number of dimensions by filtering out those that don't contribute materially to the overall model. Discriminative models basically look at the differences between sets, meaning that they don't necessarily have as much information as generative models do, and so are likely to be less accurate, but they can calculate results in a much shorter period of time.

#### Can you explain the Bayes Theorem and how it's useful in ML?

So what, exactly, is a Bayesian? Joint distributions work upon the assumption that two or more distributions are independent of one another. Many times, however, that's not the case. For instance, in a human population, height and weight are definitely correlated. In general, your weight varies according to two factors -- your height and your girth, in effect, your average radius across the sagittal (top to bottom) line of your body. You may be tall and skinny or short and stout and have the same weight, but if you constrain one of those two, the other will have a linear relationship. I'm assuming, for ease of discussion, that average density remains constant, as mass = volume x density.

Now, height, girth and mass all are distributions. Bayes Theorem can be stated as follows:

P(A|B) = P(B|A) x P(A)/P(B).

Where A|B should be read as A *given *B. In English, this translates to:

The probability of A given B is the same as the probability of B given A x the probability of A divided by the probability of B.

This is not necessarily that much clearer, so again, let's think of this in terms of the example given above. The probability that you will have a certain weight given you have a specific height (how much do I weigh if I'm six feet tall) is going to be the probability that I'm going to have a certain height if I weigh 200 pounds, times the probability that I weigh 200 pounds divided by the probability that I am six feet tall.

We can determine P(A) and P(B) by surveying a population and coming up with a distribution. If you assume that each of the distributions is a Gaussian function, we can determine that the probability of you being six foot tall puts you at the 84^{th} percentile (see this handy dandy height calculator) and the probability that you are 200 lbs is the 58^{th} percentile (similar handy dandy weight calculator). This simplifies the equation down (for this one data point) to:

P(W=200|H=6.0) = P(H|W) * P(H)/P(W) = P(H|W) * .84 / .58 = 1.45 * P(H=6.0|W=200)

Now, this doesn't necessarily tell you a lot, but it tells you something important -- if you go out and sample 10,000 men for their height and weight and create a bivariate distribution, you can use this to create a table that will let you determine roughly what P(H|W) looks like. Once you have that distribution function, you can use it to determine the distribution function for P(W|H).

To carry this through, let's say that you surveyed 1,000 men, 30 of them were 200 lbs and of those 15 were 6' even. This means that P(H=6.0|W=200) = 15/30 = 0.5. Completed this equation, you could then say that:

P(W=200|H=6.0) = P(H|W) * P(W)/P(H) = 0.5 * 0.84 / 0.58 = 0.72, or 72%.

The expression P(A|B) is typically known as the *posterior distribution*, with P(B|A) referred to as the likelihood, P(A) being called the *prior distribution*, and P(B) being called the *evidence distribution*. In our sample above, P(W/H) is the model about weight given height we don't know yet but are trying to determine, P(H/W) is the model of height given weight that we do know (or at least have a pretty good idea about), P(H) is the prior knowledge and P(W) is the data.

#### What is deep learning, and how does it differ from machine learning?

Deep learning is a subset of machine learning that imitates the way humans learn using neural networks, backpropagation, and various neuroscience principles to accurately model unstructured or semi-structured data. Because machine learning is a broader category than deep learning, deep learning can be thought of as a machine learning algorithm that specifically uses neural nets to learn data representations.

Neural nets were first hypothesized in the late 1940s by Donald Olding Hebb as a theory of neuroplasticity. In the 1960s, Frank Rosenblatt proposed a specialized computer circuit, based upon Hebb's theories, which is now considered the first computer model of an artificial neuron. Rosenblatt's untimely death in 1971, as well as scathing criticism by Marvin Minsky and Seymour Papert, who had published their own models about neurons called *perceptrons*, put a significant damper on additional research in the field until the 1980s. This additional research ultimately proved that Rosenblatt's models were more correct. Even there, however, most neural nets, which involved non-linear mathematics, were simply too slow given the technology to be fully developed.

It would take another 30 years before the technology caught up with the theory, and for neural nets to gain ascendancy. Neural nets work by taking a particular programming problem (such as analyzing the shape of handwritten characters), decomposing them into specific sub-patterns, then using arrays of non-linear equations to determine the coefficients involved in those equations. This process, called backpropagation, takes an initial guess and associates weights with each coefficient to determine the strength of a particular input.

Supervised learning, in which each data instance is labeled to identify the relevant characteristics, can usually produce higher quality results, but requires the curation of source materials. Conversely, unsupervised learning looks for patterns without initial prejudice. It is not as accurate and requires a large starting set but doesn't require that curation. Recently, more complex algorithms have emerged which use systems of ongoing "rewards" and "punishments" to change the coefficients over time, which means that a given actor will appear to learn to optimize changing conditions. This is called reinforcement learning, and it is increasingly used to drive agented systems and computer games, for example.

The advantage that neural networks have over more algorithmic approaches is that they can facilitate transactions when the underlying model driving the system is unknown. The disadvantage of such systems is that it is often difficult to understand why a given model makes the decisions that it does, a process known as *lack of explainability*. Most recent deep learning work is now focusing on that particular problem.

#### What is ensemble learning and what are its benefits?

In weather forecasting, especially when trying to determine the likely tracks of hurricanes, it is not uncommon for different organizations to create different models working with different core assumptions but usually the same base data. Each model in turn is updated with new information in real time. These agencies then frequently publish the models they used in comparison to how the actual hurricane proceeded, so that all agencies can see what proved to be consistently predictive over several storms.

An ensemble in data science can mean several things, but usually involves the creation of models that may have the same dimensional features but have different starting conditions or underlying assumptions. These can then be combined via straightforward linear algebra principles.

### Programming

#### Do you have experience using big data tools like Hadoop, Spark or Avro?

Apache Hadoop started out as a project by Doug Cutting and Mike Cafarella, then at Yahoo, to take advantage of a processing paradigm called *MapReduce* to better manage distributed search. In effect, MapReduce breaks up (maps) a large processing job (such as indexing search pages) into smaller chunks, runs each of these chunks in parallel, then reduces (synthesizes) each subsequent result into a single consolidated result.

For a while, Hadoop took the computer world by storm, eventually being spun off into two companies, Hortonworks and Cloudera. While Hadoop became less significant later in the 2010s, (and Hortonworks and Cloudera eventually merged in 2019), the movement nonetheless drove the rise of cloud-based data computing platforms, and the shift towards containerized data management through the rise of Apache Spark, which generalized many of the components of Hadoop that had emerged as part of its database phase.

More recently Apache Avro created a remote-procedure call (web services) framework for converting JSON content produced by Hadoop into an implementation neutral binary format for faster access and processing.

While machine learning does not specifically rely upon any of these tools, they tend to play an outsized role in high volume data analysis.

#### Can you write pseudocode for a parallel implementation of a given algorithm?

To make use of MapReduce-like techniques, you need to parallelize your content. Typically, this involves looking at the algorithm and locating where you are iterating over items in an array or sequence. In a serial iteration, each item gets processed one at a time, typically with a buffer acting as an accumulator of the results.

Parallel implementations typically involve exporting the processing of each activity on a separate *thread*. Because such threads may take different lengths of time, the system needs to have some way of determining whether a given process has completed or not, and then only doing the consolidation once all of the separate threads have returned.

For instance, suppose that you were trying to extract metadata keywords from a spidered set of web pages. A typical parallelization might look something like this (code given here is Javascript)

async function processPage(page){

// do the separate processing at the page level.

return results

}

function reducer(acc,cur,idx,src){

//Take an accumulator (an object or array) and process

// it with the current object, with an optional index and src

// for comparison operations, then return a result, such as

// updating a database index

}

function processBatch(pages){

let pages = [page1,page2,…,pageN]

let partialResults = await (Promise.all(

pages.map((page)=>processPage(page))

))

let finalResult = partialResults

.reduce({acc,cur,index,src}=>processBatch({acc,cur,index,src})

Most languages now have similar asynchronous function calls and linear map and reduce routines.

#### How would you address missing, invalid or corrupted data?

It has been said that 85%-90% of the work that most data scientists do is not, in fact building models but rather is involved with cleaning and preparing data for use. This means understanding where data comes from, and understanding how to take those large data sets and put them into a form that are easier to work (a discipline that's coming to be called *data engineering*). Of course, it has also been said that 80% of all statistics are made up, but data scientists do spend a significant amount of time trying to turn bad data into at least workable data.

Data quality, and its counterpart, data cleansing, are an integral part of any data scientist's work. The unfortunate truth is that most data that data scientists work with is *forensic* -- it exists primarily as an artifact of some other, usually unrelated process that was typically tied into specific applications. Data scientists likely won't have access to the operational code used to generate that data in the first place, so they have to play detective.

You can do some tasks to improve data quality. The first task is generalizing the data into a format that can be serialized as JSON, XML, or ideally, RDF. Once in these formats, it is usually easier to see where the holes are. If possible, once serialized, you should spend some time putting together a validating schema in your language of choice, which will then help you identify fields that are non-compliant. If you can identify that something is a date, for instance, then it is usually a straightforward process to convert from one date format to another. Similarly, validators can identify errors such as dates located far into the future (the year 9999, for instance) which was a frequent hack in older data systems.

Nulls typically indicate that no data was known for the record in question. Relational data didn't really have the option of specifying that a given data field was optional, which was where the use of nulls arose from. More contemporary data formats usually do have this capability, which is one reason why moving out of a specific relational database format makes sense.

Corrupted data should simply be removed. The problem is that you do not know how the data was corrupted, and without that piece of information the data that you have may introduce errors into your processing.

A number of commercial tools provide a certain amount of data cleansing, but it's worth spending time to understand what these tools do, and what impact they have on the data you work with.

#### What's more important, accuracy or performance?

While you should strive for accuracy in your models, the reality is that accuracy often highly depends upon the data that you are working with, and in the real world, uncontrollable factors can significantly skew the data. However, model performance is something very much under your control. As a general rule of thumb, you should work your way up the ladder of complexity and performance.

Many forms of data analysis come down to simple linear regression testing, which can often at least help you determine what the best tools to work with the data may be (and which tools are a waste of time). As you get a better feel for the "shape" of the data, you can start to put together an *analysis plan* that can be ranked by performance (quick and dirty vs. slow but precise), and often you'll discover that the faster performing tools will be good enough for all concerned. Logistic regression can also be used, which is slightly more complex than linear regression, and is used when the dependent variable is binary, not continuous.

Indeed, one of the dangers that most data scientists are prone to is the pursuit of precision. By nature, most data is filled with biases and potentially weak variances. The higher the degree of imprecision of the source, the less precise the final values will be. Simply because your R or Python program can give you precision to 32 places does not mean that anything beyond the first two or three of those places have any real meaning.

### Business Applications

#### How do you envision your ML skills benefiting the company's bottom line? How will you use them to generate revenue for the company?

This is one of those questions that is worth thinking about in depth, because ultimately it comes down to whether or not the company is a good fit for you. Data science, whether of the data analytics kind or the machine learning variant, is *very* new to most companies. Chances are very good that the reason that you are being hired in the first place was that someone in management went to a conference or saw a TED Talk on all of the benefits of data science, AI and machine learning to your bottom line without have more than a basic understanding about what was actually involved.

In some cases, this can be a good thing -- you will be in a better position to help the company shape an effective data analytics strategy -- but it can also mean that your existence there is to justify that the company take certain actions, regardless of what the data actually says. This means that going in you should understand what the actual process involved is with regards to what you do, and it also means that up front you should be very clear that your intent is to recommend what the data says, not to justify a decision at odds with that data. It may mean you don't get the job, but you will be in a much healthier head space having turned it down.

Your skills as a data scientist will be applied two ways: forensically or proactively. Forensic data science involves the analysis of information to determine the past state of some environment, such as the company's finances and spending or changes in consumer sentiment. The resulting analysis is intended to better predict the future based upon the near-term past, but usually humans in forensic roles make these decisions.

Proactive data science, however, is involved in creating products that use data science and machine learning techniques to do things. Example of this include developing chatbots and natural language processing systems, building autonomous vehicles, drones and AI activated sensors, or even handling the purchasing of goods, stock or other instruments.

Understanding the distinction is important. People skilled with applications such as TensorFlow may not actually have (or need) strong statistical knowledge, but they usually do need to have above-average programming skills (most likely in Python, Scala or similar tools). Conversely, forensic data scientists typically need a deeper statistical background but may never even touch a neural network. In short, proactive data scientists are programmers who use machine learning as their tool of choice, while forensic data scientists are much more likely to be domain subject matter experts and analysts who use machine learning tools as simply one tool in a suite.

#### What are the KPIs of the company?

This is a question that you should ask your interviewer because it is likely that the answer will be both informative and signal that you're interested in the business side of what you bring to the table. KPIs evolved from management practices in the late 1990s and beyond, attempting to both quantify specific targets of activity and qualify the impact of certain actions in terms of fuzzier qualities such as morale and customer satisfaction.

KPIs for data science are still evolving, given how new this particular domain is, but typically come down to a few significant factors. For forensic analysis, this typically includes the following:

- Are data analyses accurate?
- Is the amount of data processing necessary to get to meaningful measures too onerous?
- Is the analysis available in a timely fashion? (Is the data reasonably fresh enough to provide an accurate snapshot?)
- Are the results of the data analyses actionable? (Can key decision makers make decisions based on the data?)
- Are the analyses sufficiently predictive? (Given an analysis, do the conclusions occur at a statistically relevant frequency?)

Proactive machine learning is even more tenuous:

- Do the models ML specialists create produce expected results given data?
- Is the cost of producing the results sustainable for the customer or the company?
- Do the products that use the machine learning results provide increased customer satisfaction?

This is an area that you can also use to get a gauge of how seriously the company takes machine learning. If they don't have any formal KPIs in place, it means they don't know how to manage data science and machine in their company. The job may not effectively use your talents, unless that role is more administrative and entrepreneurial.

### Personal interest

#### What is your favorite algorithm or ML model use case?

This question is designed to probe your interest in the field in ML and the answer should show that you have enough experience to develop this opinion. You've used multiple algorithm types and can articulate the benefits of each and why you prefer one over another, either based on its use case or your work style.

#### What would you have done as a contestant in the "Netflix Prize" competition?

Netflix makes very heavy use of machine learning algorithms in order to determine movie preferences and recommendations. This particular question is similar to Kaggle Competition test questions, and ultimately is an indicator of how well you stay abreast of developments in the field and have thought about solutions to problems like these.

#### What was the last ML research paper you've read?

Similarly, this recognizes that machine learning is still enough in its infancy that most of the innovations are still coming from academic sources. Sites such as Data Science Central (a subsidiary of TechTarget) or ArXiv.org are useful places to find information about current research in the fields of artificial intelligence, machine learning and graph analytics.

#### How is data trained for autonomous vehicles or self-driving cars?

This is a question that is intended to ferret out your understanding of how data science and machine learning is handled in the real world. Not surprisingly, the answer is complicated. Typically training involves a combination of taking human-driven vehicles mounted with cameras and other sensors through various environments and conditions, then feeding this information as multiple synchronized data streams to a comprehensive integration system. The resulting content is then synthesized into different scenarios that can be mixed and matched, often augmented by superimposed models on the environment.

#### How and where do you source data sets?

This is always a contentious question. For a while, the cost of data acquisition dropped dramatically, and by 2000 data was pretty ubiquitous, albeit often of fairly marginal quality. However, that equation has changed as the need for large-scale training data sets of good quality rises again.

For some time, most governments moved towards a period of enhanced data transparency, but this too has changed as cost and political attitudes have shifted. Additionally, the COVID-19 pandemic exposed the fact that data collection globally is nowhere near as comprehensive nor as consistent as it was once believed to be. Privacy laws such as the European GDPR and the California CCPA also raise issues about data collection and consequently data availability.

Github repositories, linked data sets and similar public data sets are available, but more than a few companies are thriving precisely because they've focused on the mostly manual process of collecting and curating data content in specific domains.

### Summary

Machine learning is an intriguing new field, and like any new field, is still experiencing both opportunities and growing pains. When going into a company, try to ascertain as quickly as possible whether the role is likely to be analytical and forensic or product-centric and proactive. Spend time researching not just the technology but also the company, and spend some time attempting to parse between the lines to determine how far along a company is in its own digital transformation. Most importantly, believe in yourself. The world needs more data scientists.