What is Data Science?

PSTAT 234 (Fall 2025)

Sang-Yun Oh

University of California, Santa Barbara

Data Science Example: Google Ads

• Backbone of Google’s revenue model

• 80.20% market share in pay-per-click (PPC) ads market ($237.8 billion revenue in 2023)

• Precision targeting: search intent, demographics, location, time

• Scalable for all businesses: $50 to $5M budgets

• Global reach with local control

• “Leveling the Playing Field: Great for Small Businesses”

• 82% of small businesses attribute revenue growth to digital ads and 79% say these tools help them compete with larger companies.

• Faced accusations of anti-competitive practices, market dominance, and potential favoring its own services in search results.

Google Ads

Types of Google Ads (source: Zapier).

Ad type	Location
Search ads	Throughout ranking links in the Google Search Engine Results Page
Display ads	On webpages, apps, and Google properties in the Display network
Shopping ads	Shopping, SERP, Images, Maps, and search partner search results
Video ads	Before, during, and after YouTube videos; YouTube search feeds; YouTube home feed; and across Google video partners network
App ads	Play Store search suggestions/results, Google SERP, YouTube feeds, Google Discover, and across the Google search partner network
Discovery ads	Google Discover feed, YouTube homepage and Watch Next feeds, and Gmail Promotion and Social tabs
Local Services Ads	Google SERP, Google Maps
Performance Max ads	All Google advertising channels
Smart ads	Google SERP, Maps, YouTube, Gmail, and search partner sites

How Google Ads Work

For this example, assume we are looking at search ads.

• Advertisers bid for ads to appear during searches.
• An auction determines which ads are shown based on bid and estimated user interest.
• High-placing ads are shown to users
• Advertisers pay Google if users click on the ads.
• Enhanced Campaigns allow customized bidding based on user context (e.g., time, device, location).

Enhanced Campaigns offered advertisers the ability to customize their bidding, e.g. to show one ad to someone searching for “pizza” at 1 p.m. on their PC at work (perhaps a link to an online order form or menu) and a different ad to someone searching for pizza at 8 p.m. on a smartphone a half-mile from the restaurant (perhaps a click-to-call phone number and restaurant locator).

Internet Advertising Data

Advertising data often have the following variables:

Variable	Description
impr	Number of ad impressions (ads shown)
click	Number of clicks
cost	What advertisers paid
conv	Number of conversions (purchases)
value	Value of conversions as reported by advertisers
cpm	Cost per impression (cost/impr)
cpc	Cost per click (cost/click)
cpa	Cost per conversion (cost/conv) (or 0, if conv is 0)
cpr	Cost per return (cost/value) (or 0 if value is 0)
roi	Return on investment (value/cost)

Data variations

	Prefix indicates platform
m.*	Mobile, e.g. m.impr
d.*	Desktop/tablet, e.g. d.impr

	Suffix indicates when
*_pre	Before experiment, e.g. m.impr_pre
*_post	In experiment, e.g. m.impr_post

From the data, derive new features for analysis:

	Derived features
error.cpr*	m.cpr - d.cpr (pre, post)
mult.change	Change in mobile multiplier

Common Questions from Advertisers

• How much am I paying for ads? On desktop? On mobile?
• How much am I paying for clicks?
• How much am I paying for conversions (purchases)?
• How much was spent for each dollar of value generated (cost per return)?
• How much am I getting back in comparison to what I spent? (return on investment)

Google Ads Metrics

• If the roi = value/cost was higher on mobile than desktop,
  \[\texttt{m.roi} = \frac{\texttt{m.value}}{\texttt{m.cost}} > \frac{\texttt{d.value}}{\texttt{d.cost}}= \texttt{d.roi} \]

• Increasing “mobile multiplier” means adjusting mobile bids by a factor: \[\texttt{mult} = \frac{\texttt{m.cost}}{\texttt{d.cost}}\]

• Due to auction structure, more mobile and fewer desktop ads will show.

• Better “mobile multiplier” should make two rois more equal:
  \[\texttt{m.roi} \approx \texttt{d.roi}\]

• Why would Google and advertisers care about this?

Google Ads Experiment: Goal

• Setting appropriate multiplier equalizes return on ad investment.

• Better experience for advertisers benefits Google in the long run.

• Experiment to equalize return on ad investment (desktop vs. mobile).

• Determine better “mobile multiplier” recommendation.

• Simpler advertising cost estimates.

• Suppose a new multiplier recommendation algorithm is being tested.

• After experiment, how do you measure success?

Google Ads Experiment: Naive approach

Chihara and Hesterberg (2018)

Question from a Googler

I have a pre vs. post comparison I’m trying to make where alternative hypothesis is pre.mean.error > post.mean.error (mean.error is mean of cpr.error = m.cpr - d.cpr). My distribution for these samples are both right skewed as shown below. Anyone know what test method would be best suited for this type of situation?

Density plot for cpr.error. (Recall cpr is 0 if value is 0)

Google Ads Experiment: Feature Engineering

Is CPR error (m.cpr - d.cpr) a good measure of performance difference?

Figure 8.7 (c) offers interpretability of zeros whereas (a) doesn’t and is less non-linear than (b)

Google Ads Experiment: Results

Error means error = log(1+m.roi) - log(1+d.roi) for pre/post experiment.

Hypothesis test can be performed on error_pre and error_post.

What does this example tell us about data science?

Data Science is …

• A multi-disciplinary field that uses scientific methods, processes, algorithms and systems to extract knowledge and insights from structured and unstructured data (reference)

• Merger of statistics, data analysis, machine learning and their related methods in order to understand and analyze actual phenomena with data (reference)

• Composed of techniques and theories drawn from many fields within the context of mathematics, statistics, computer science, and information science. (reference)

[International Statistical Review]

• Domain expertise: data analysis collaborations in subject matter areas.

• Mathematics/Statistics: models, estimation, and distribution based on probabilistic inference.

• Computing: hardware and software; computational algorithms

• Theory: foundations of data science; mathematical investigations of models and methods

Data Science Approach

[UC Berkeley, School of Information]

Statistics and Data Science

• Many statistical methods make (optimistic) assumptions.

• Data science often focus on practical benefits (predictions).

• Data science often iterate to make improvements.

• Data science process emphasize entire data science lifecycle

• Data science develops processes (often custom built)

Data Science Diagram

Drew Conway: The Data Science Venn Diagram

References

Chihara, Laura M., and Tim C. Hesterberg. 2018. Mathematical Statistics with Resampling and R. 1st ed. Wiley. https://doi.org/10.1002/9781119505969.