COGS 9 Exam 1

data science

the scientific and ethical process of extracting value from data

steps of the data science process

1. Ask an interesting question

2. Get the data

3. Explore the data

4. Model the data

5. Communicate and visualize the results

why do we start with questions?

• to know what to look for

• to know what kind of analysis to do

• to think about ethics

goal

start with a ____ in mind

questions

____ drive the choice of data

criteria of a data science question

• specific

• measureable

• answerable with data that’s possible to collect

• have relevance or value to something or someone

• big enough to be interesting

• small enough to be accomplishable

ingredients for a data science question

• topic

• outcome/target

• features

• timeframe

topic

who or what are you studying?

outcome/target

what exactly are you measuring?

features

what are the factors or comparisons that matter?

timeframe

over what timescale?

some goals of data science questions…

• describe or understand patterns

• predict

• test cause-and-effect

• decide what to do or optimize something

ethics

principles and standards that guide choices about how data is collected and how models are built and used

privacy

a person’s ability to control access to themselves and their own information

The Cobra Effect

the unintended consequences that result from an attempt to solve a problem

class imbalance problem

the data itself are biased (e.g. a trained hiring-recommendation model discriminates against women)

what your devices do know

• demographic data

• location data

• search history

• browsing history

• purchase history

  • physical interactions

cookies

small pieces of data a website stores in your browser to remember you, but can also track your activity across pages

how to practice more responsible data science

• learn from the past

• use checklists: how data will be dealt with ethically and responsibly before the project is started

• incorporate ethics into hiring decisions

• organize governing bodies to determine how data will be regulated

  • ensure safeguards (e.g. when to “pull the plug”)

Dean’s Ethics Checklist

meant to provoke discussion around:

• data collection

• data storage

• analysis

• modeling

• deployment

data

anything that can be measured or observed about the world

quantitative

a value that can be objectively measured or counted

qualitative

describes categories or qualities

categories of qualitative data

• categorical

  • nominal (unordered)

  • ordinal (ordered)

• text, audio, images (in raw form)

categories of quantative data

• discrete: countable values

• continuous: any value in range

quantitative data visualization

• histogram

• box plot/violin plot

• scatter plot

  • line chart (over time)

qualitative data visualization

• bar chart

• grouped/stacked bar chart

• pie chart

structured data

can be organized into an n x p table of values

n

the number of observations/sample size

p

the number of features

structured data: storage

• spreadsheets (.csv, Excel (.xls, .xlsx))

• databases: SQL (structured query language)

SQL as an example of a structured database

• enables you to relate tables of data together, perform actions on data

  • can ask questions (query), add, update, delete information

• manage + analyze large datasets

structured data: JOIN operation

combine rows from 2+ tables by matching columns

unstructured data

no simple way to represent in table format

unstructured data: data storage

• text documents

• emails + messaging

• images

• audio

• video

• presentations

• web content

sentiment analysis as unstructured data

using Natural Language Processing (NLP) to identify and categorize opinions expressed in text

semi-structured data

doesn’t fit neatly into rows and columns, but has labels + keys that give it organization

how to get data

1. collect it yourself

2. use public/open-source data

3. buy it

4. request it from an API

5. scrape it from a website

6. source it from your company

steps to request data from an API

1. send a request

2. API sends back semi-structured data, often JSON

web scraping

a way to automatically collect information from websites

• used when there’s no API or an insufficient API

API: definition, pros, cons

a site’s official way to access data

• structured + stable

• may require a key + have rate limits

web scraping: pros + cons

• useful when there’s no API

• fragile; websites change

  • can violate terms & conditions

questions to ask before using data

• allowed? restrictions? ethical, privacy, legal concerns?

• why and how was the data collected?

• how was the data cleaned or altered? missing data?

• who/what is included or excluded? biased in data collection/measurement?

• documentation?

  • trustworthy?

robots.txt

small text file a website can publish that tells automatic crawlers which parts of the site they’re allowed to access

• voluntary

• websites enforce real restrictions with logins, paywalls, server-side blocking

“Identity crisis” and the stats vs. data science debate

Donoho is pushing back on the idea that data science is simply a subset of stats or a rebrand. The argument isn’t “stats is irrelevant”—it’s that data science includes major concerns that weren’t historically central in stats curricula (especially computing and end-to-end workflow).

The “two cultures” (generative vs. predictive)

This frames tension between modeling for explanation/interpretation and understanding mechanisms vs. modeling for prediction and performance on held-out data

The Common Task Framework (CTF)

CTF is a way to compare methods using a shared dataset, a defined task, and a scoring metric—often with a holdout test set. It encourages a prediction-centric culture and can drive rapid progress, but it can also reward narrow optimization that may not generalize.

“Science about data science”: cross-study and cross-workflow thinking

Donoho’s examples highlight something uncomfortable but important: different reasonable workflows can produce different results, even on the same data. This raises questions about reliability, reproducibility, and the credibility of conclusions.

key ideas: “Data’s Day of Reckoning”

1. Data-driven products can be weaponized (misuse, abuse) when incentives don’t align with the public good. 

2. Ethics must be integrated into everyday practice and data science education, and not treated as an optional afterthought.

3. Use concrete mechanisms like checklists and “stop-the-line” culture to make responsibility actionable. Build routines for consent, bias checks, abuse cases, monitoring, and the ability to pause/shut down systems.

key ideas: “Myths & Fallacies of Personal Identifiable Information (PII)”

1. The “PII fallacy”: removing a fixed list of identifiers doesn’t guarantee privacy. PII is hard to define and context-dependent.

2. Re-identification can happen without names or obvious identifiers. Any sufficiently rich set of behavioral/transactional attributes can uniquely “fingerprint” people, especially when linked with other data.

3. Privacy protection should move beyond “de-identification” toward stronger approaches.