Key Takeaways
Treat data science as a team discipline, not a superhero job
Combine complementary skills. Data science is not simply statistics with a new label, nor a checklist one person can master. It brings together computation, statistical thinking, domain knowledge, visualization, and communication to solve practical problems. At Google, the team described in the book included statisticians, engineers, a social scientist, a physicist, and a computer scientist. Their different strengths made it possible to build models and products that none could have delivered alone.
Organize around the problem. The book's recurring “data science profile” exercise asks people to compare their strengths across skills, then consider whether a team's combined profile fits its problem. A scientist also needs to ask what data to collect, explain assumptions, and translate findings for people who make decisions.
What's striking is that the team is not merely a staffing convenience. It is a way to distribute cognitive work: domain experts notice meaningful questions, statisticians interrogate uncertainty, and engineers make the analysis reliable at scale. This resembles interdisciplinary research in medicine, where clinical insight and technical measurement must meet. Yet collaboration alone does not guarantee good judgment. Teams can share blind spots, especially when hierarchy silences dissent. Clear ownership of assumptions and a norm of asking basic questions help turn diversity of expertise into genuine scrutiny.
More data never repairs a biased window onto the world
Data is a partial trace. A dataset records selected aspects of real events, through a collection process shaped by people and technology. Even a huge dataset can misrepresent the population or question that matters. After Hurricane Sandy, a study of tweets could suggest New Yorkers were shopping before the storm and partying afterward. But Twitter users were not representative of everyone affected, and people in the hardest-hit areas may have been too busy to post.
Ask who is missing. “N=ALL” is rarely literal: a complete record of platform users is not a complete record of human behavior. Sampling can also solve engineering problems when a smaller dataset is sufficient. Before generalizing, identify who could appear in the data, who could not, and how those limits affect the conclusion.
The useful shift is from asking only how much data exists to asking how the data came to exist. This is a measurement problem, familiar in surveys and the social sciences: the people easiest to observe are not necessarily the people most relevant to a decision. Scale can reduce random error while leaving systematic exclusion untouched. In some settings, missingness is itself informative, but that does not make it harmless. A model's confidence should reflect the boundaries of its evidence, not the size of its storage system.
Inspect the logs before trusting the model
Explore before confirming. Exploratory data analysis, or EDA, means examining data with plots, summaries, and comparisons before settling on a model or hypothesis. The aim is to understand the data's shape, find missing values and outliers, check whether measurements make sense, and notice patterns that may come from a broken logging process rather than real behavior. For online products, this can reveal that an apparent user trend is actually a software bug.
Make it part of the workflow. The authors recommend starting with a small sample, such as plotting individual users' actions over time, then aggregating only after understanding what those records represent. The process loops: cleaning can uncover new problems, and exploration can show that more data or better logging is needed. EDA is private investigation, not yet a polished chart meant to persuade an audience.
EDA is a practical defense against premature certainty. A plot can expose an impossible value or a sudden synchronized spike faster than a sophisticated model can explain it. This echoes software debugging: outputs are evidence about both the system and the instrument producing them. Still, visual inspection is not a substitute for reproducible checks. People tend to see patterns they expect, so exploration should generate questions that can be tested, not license a story simply because a chart looks compelling.
Choose algorithms by problem, assumptions, and deployment constraints
Match method to task. Linear regression estimates a continuous outcome, such as time spent on a site. Logistic regression estimates probabilities for a binary outcome, such as whether a user clicks. Naive Bayes can classify text by combining evidence from words, while k-means groups unlabeled data into clusters. Each method answers a different kind of question, and each relies on choices about variables, assumptions, and measurement.
Account for constraints. A nearest-neighbor model can be intuitive, but it struggles when there are many features, sparse data, or poorly scaled variables. A model that performs well may also be too slow to run for each user, too costly to update, or too hard to explain. Begin with a simple approach, evaluate it, and add complexity only when the problem and evidence warrant it.
The book's emphasis on choosing a method before reaching for a familiar tool is a form of model discipline. Algorithms encode assumptions about what counts as similar, important, or predictable; those assumptions can matter as much as the equations. Deployment constraints also shape scientific validity: a model that cannot be updated or applied consistently is not a useful answer. Simplicity is not automatically superior, though. The right trade-off depends on the cost of errors, the need for explanation, and the consequences of acting on a prediction.
Evaluate the decision you need, not a flattering accuracy score
Choose a consequential metric. Accuracy can be misleading when positive cases are rare. A fraud detector that labels every transaction legitimate may be correct most of the time and still miss nearly all fraud. Precision asks how many flagged cases are truly positive; recall asks how many true positive cases the model catches. In payments, false alarms can anger honest sellers, while missed fraud costs money, so the right balance depends on real costs.
Separate ranking from probability. AUC can assess whether a model ranks cases well, but a good ranking does not guarantee accurate probability estimates. The book's charity example makes this concrete: a model that separately estimated the chance of donating and the amount donated raised $15,000, compared with $9,000 from mailing everyone. The metric should track the actual decision, not merely a convenient model score.
Evaluation is a statement about what the organization values. A fraud system that maximizes recall without accounting for false accusations transfers costs to innocent customers; one that prizes precision may leave victims exposed. Separating ranking from calibration is equally important because decisions often depend on thresholds or expected costs, not order alone. A further safeguard is to report performance across affected groups and operational conditions. A single aggregate score can conceal who bears the model's errors.
Keep future information out of training data
Watch the timeline. Data leakage occurs when a model uses information that would not legitimately be available when making a prediction. In an online retail contest, “free shipping” looked like a strong signal of high spending, but customers received it after crossing a spending threshold. The feature revealed the outcome instead of forecasting it. In a medical prediction example, the way diagnoses were removed and records rearranged left clues that exposed whether a patient had pneumonia.
Recreate real use. For time-dependent questions, train on the past and test on a later period. Record when information became available, not just when an event occurred. Cleaning, joining, or filtering data can accidentally create clues, so inspect the pipeline and understand how every field was produced. A high score that seems implausibly good is a reason to investigate, not celebrate.
Leakage is more than a technical bug. It is a failure to preserve the conditions under which a decision will actually be made. Random splits can quietly mix information from the future into both training and testing, creating an illusion of readiness. Temporal validation is therefore essential when the world changes over time. Even careful separation cannot guarantee validity if the deployment setting differs from the test data, so performance should be monitored after release and treated as conditional evidence.
Recommendation engines shape the tastes they claim to measure
Sparse ratings need structure. A recommendation system uses past preferences to estimate what a user may like. Nearest-neighbor methods can fail when users have rated few items or when thousands of overlapping features make meaningful distance hard to define. The book describes matrix methods such as singular value decomposition and alternating least squares, which represent users and items through a smaller set of hidden preference dimensions. That compressed representation can help predict missing ratings from a sparse user-item table.
Remember the feedback. Once recommendations are shown, they influence what people see and choose. An item promoted near the top may receive more clicks because of its position, not simply because it was already preferred. Repeated exposure can create a filter bubble or a popularity spiral. Logging position and testing alternatives help distinguish user preference from the system's own influence.
A recommender is not a passive window onto taste; it is part of the environment in which taste is expressed. This connects prediction to mechanism design: changing the options people see can change the outcomes used to train the next model. Latent factors offer an efficient way to summarize patterns, but they may also hide assumptions about what dimensions matter. Systems should reserve some opportunity to surface less-promoted items, otherwise their evidence can become self-confirming.
Use experiments for causes; treat observational answers as provisional
Association is not intervention. OK Cupid compared response rates for introductory messages containing different words. Messages with “beautiful” drew fewer replies than the overall average, but that alone could not show that the word caused the difference. The recipients of those messages might differ in important ways, such as receiving more messages overall. Such a confounder affects both the treatment and the outcome.
Randomize when possible. A randomized experiment assigns comparable groups to different conditions, making it easier to attribute outcome differences to the treatment. Experiments are not always ethical, affordable, or practical, so researchers use observational methods such as propensity score matching, which pairs treated and untreated cases with similar estimated chances of treatment. These methods only adjust for factors that were measured and modeled. Simpson's paradox shows why group composition matters: a trend in combined data can reverse within subgroups.
The distinction between prediction and causal explanation is central to responsible analysis. A variable can predict an outcome without being a useful lever for changing it; confusing the two can turn a descriptive pattern into a harmful policy. Randomization is powerful because it balances measured and unmeasured factors on average, but it cannot answer every question or guarantee that an experiment applies to every population. Observational estimates remain valuable when their assumptions are explicit and their uncertainty is not hidden.
Map relationships, not just individuals, when influence is the question
Look at the connections. Case-attribute data records traits of individuals, while social network analysis studies actors and the relationships between them. A person's influence may depend less on their profile than on who listens to them, whom they connect, or how information travels through a group. Network measures such as degree count connections, while betweenness highlights actors who sit on many shortest paths between others.
Find the group pattern. Morningside Analytics mapped international blog communities and used clusters and text analysis to interpret their discussions. In its comparison of Iranian blog networks from 2008 and 2011, only about 15 percent of the individual blogs remained, yet broad community patterns persisted. Visualizations can reveal these “schools of fish,” but the analyst must choose what counts as a connection and which local network matters.
Network analysis changes the unit of explanation from isolated people to relationships and positions. That can reveal brokers or communities invisible in a table of demographics. The Iranian blog example also distinguishes continuity of a social pattern from continuity of its members. Yet a network map is not the network itself. Different definitions of an edge, boundaries, or time windows can produce different pictures, so visual clarity should never be mistaken for neutral measurement.
Make large-scale analysis reliable before making it clever
Distribute the work. MapReduce is a framework for computations that can be divided across many machines. In a word-count task, a mapper emits each word with a count, the system groups matching words, and reducers add the counts. The framework also handles failures through mechanisms such as replicated data and restarting work, which matters because machines fail more often as a cluster grows.
Use the right architecture. MapReduce fits tasks that can be split into local computations and combined through aggregation. It is less natural for algorithms that repeatedly depend on the result of the previous iteration. The practical work begins before modeling: collecting, cleaning, joining, and storing data. Reliable pipelines, readable code, testing, and reproducibility are not support chores. They determine whether the analysis can be trusted and repeated.
The engineering lesson is that scale changes the shape of a problem, not just its size. Fault tolerance, data layout, and runtime become part of the analytical method because they determine which computations are feasible and repeatable. MapReduce is a useful example, not a universal answer: forcing iterative work into a batch framework can add complexity without insight. A sound team first asks what operations the task requires, then selects infrastructure that preserves correctness.
Ask what the model will do to people before shipping it
Models carry values. People decide what to collect, which groups appear in a dataset, what counts as success, and how a score changes a product or policy. Those choices are not neutral. A hiring model trained on past outcomes could learn to favor men if women had historically been promoted less or left more often, mistaking unequal treatment for a useful signal.
Treat privacy as part of design. The book notes a mobility study in which four location points identified 95 percent of people in a dataset of 1.5 million. Removing names therefore does not necessarily make records anonymous. Responsible data science requires clear questions, explicit assumptions, careful communication, and attention to feedback loops, privacy, and who bears mistakes. Technical skill matters, but so does the judgment to question whether a problem should be automated.
The ethical argument is strongest when treated as a design requirement rather than a closing disclaimer. A model can reproduce past inequity even when sensitive attributes are omitted, because other variables may act as proxies. Privacy risk also arises from combining seemingly ordinary records. Auditing, explanation, and meaningful routes to challenge decisions can reduce harm, but they cannot decide which goals are legitimate. That judgment belongs to institutions and affected communities, not to an optimization routine.
Analysis
Doing Data Science is best understood as a field guide to practice, not a comprehensive technical textbook. Its structure moves from statistical foundations and algorithms into products, networks, causal inference, engineering, and professional ethics. The throughline is that the work is a chain of decisions: what to measure, what the data can represent, how to model it, what success means, and what happens when an output enters the world.
Its strongest contribution is the refusal to separate technical competence from epistemic humility. Sampling, exploratory analysis, data leakage, evaluation, and causality are presented not as isolated topics but as ways a plausible answer can go wrong. The examples make this tangible: Sandy's tweets cannot stand in for everyone affected, free-shipping status can reveal a purchase that has already happened, and recommendation rankings can manufacture the behavior they later interpret. The book also convincingly recasts data science as team practice, joining computation with domain expertise and communication.
The breadth is also a limitation. The survey format offers uneven depth, and its 2013 technology landscape predates many subsequent shifts in machine learning and data infrastructure. Readers seeking a complete account of modern algorithms will need other sources. Some advice also depends on organizational context: sampling may be a practical solution for analysis, while serving a product may require comprehensive user-level records. The book's own emphasis on asking the right question helps resolve that tension.
The durable lesson is methodological rather than technological. A model is not validated because software runs, a metric improves, or a dataset is large. It earns trust only when its evidence matches the claim, its evaluation resembles its use, and its effects on people are considered. That standard applies whether the tool is a simple regression or a system distributed across thousands of machines.
Review Summary
Doing Data Science receives mixed reviews. Some praise it as a good introduction to data science, highlighting its broad coverage and real-world insights. Others criticize it for being disjointed and superficial. Readers appreciate the practical advice and industry perspectives but note it's not a comprehensive textbook. The book's structure, based on lecture notes, is seen as both a strength and weakness. While some find it engaging and informative, others feel it lacks depth and coherence. Overall, it's viewed as a decent overview for beginners but not sufficient for in-depth learning.
People Also Read
Glossary
Datafication
Turning life into recorded dataDatafication is the process of converting aspects of life and behavior into recorded data that can be stored and analyzed. Online actions such as liking a post are explicit examples; browsing or moving through a monitored space can also generate records without a person's active intention. Once captured, data may be used for purposes beyond the original action.
Exploratory data analysis
Inspecting data before modelingExploratory data analysis, or EDA, is the practice of examining data with plots, summary statistics, and comparisons before committing to a formal model. It helps analysts understand distributions, find missing values and outliers, check that records were collected as expected, and develop questions. It is both a set of tools and an open-ended investigative stance.
Data science process
From raw data to productThe data science process describes the path from a real-world activity to useful analysis or a data product. It typically includes collecting and cleaning raw data, exploring it, choosing and fitting models, interpreting and communicating results, and sometimes building the output into a product. When users interact with that product, their behavior generates new data and can alter the phenomenon being studied.
Naive Bayes
Probabilistic classification from featuresNaive Bayes is a classification method that applies Bayes' rule to estimate the probability of a class given observed features. In the book's spam-filter example, words provide evidence for whether a message is spam or ordinary email. The method treats features such as words as conditionally independent within each class, a simplifying assumption that makes estimation efficient.
Laplace Smoothing
Stabilizing sparse probability estimatesLaplace Smoothing adjusts probability estimates based on counts by adding pseudocounts. In a Naive Bayes text classifier, this prevents a word that never appeared in a training example from receiving a probability of exactly zero, and prevents certainty based on a tiny sample. The pseudocount values act like a prior and influence how strongly sparse evidence affects the estimate.
Data leakage
Invalid information entering predictionData leakage occurs when a predictive model uses information that would not legitimately be available at the moment the prediction is made, or when data preparation unintentionally exposes the target. A model may score exceptionally well in testing but fail in real use. Preventing leakage requires understanding how fields were generated and preserving the real timeline of information availability.
Rubin Causal Model
Potential outcomes frameworkThe Rubin Causal Model describes causal effects by defining, for each unit, an outcome under treatment and a separate outcome under control. The individual causal effect is the difference between those potential outcomes. Because a unit can experience only one condition at a time, the other outcome is unobserved, a limitation known as the fundamental problem of causal inference.
MapReduce
Distributed map-and-aggregate computingMapReduce is a framework for processing large datasets across multiple machines. A mapper transforms records into key-value pairs; the system groups values with matching keys; and reducers aggregate each group. For example, a word-count task maps each word to a count of one, then adds counts for each word. The framework also coordinates distributed work and handles machine failures.
Exponential downweighting
Discounting older observations graduallyExponential downweighting gives recent observations more influence than older ones, with each step into the past reducing a value's weight by a decay factor. In the book's financial example, it supports a running volatility estimate that can be updated with little stored data. The decay factor determines how quickly older information loses influence.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.