1/86
Looks like no tags are added yet.
What did redman say quality actually is
fitness for intended use. Quality is not a property of data, it depends on the use —> marketing data isn’t the same as finance data
Pivot table
summarizes a table you CANT read by eye
What’s the alternative for using a pivot table
the alternative is writing a formula for every question you have
Why are they called pivot tables
you can pivot the data around to a different axis/ format
how can the same set of data, create 4 different data
different functions —> count, sum, aa verage, %of grand total
Relational database
one way to store data related to eachother in a predefined way
Predefined way
at the time of creation of database, yo ucan identify relationship that exist between different entities or groups of data
anatomy of relational database (4)
tables
properties you want to store
relationships - between table
index- useful for connecting tables for quick look-ups
Schema
structure that defines tables, relationship between tables, fields, index
When designing schema and using a database, what are 4 things to keep in mind
understand business needs (Who are the customers? spending to date, age, employee sales, products)
identify entities (customers, products, employees) —> title of table
identify properties/fields on those entities (values inside table)fl
define relationship between tables
flat file + give 3 examples
stores data in a single, two-dimensional table without links or relationships to other file
ex: pivot, BASIC data base, spreadsheet
Spreadsheet v database
spreadsheet- flat file
Database- Automatically fed, can have many tables, linked, queries across all of it, written in SQL, hold more data than spreadsheet can, collection of files
A database is a collection of ____q____ organized in ___x_____ . ____x____ are made of records/ ____y____ , which are made of ____z____
q= files
x= tables
y= rows
z= columns
Flat file v relational
Flat file = single table
relational - multiple tables
metadata
data about data. Data that provides information about other data. describes, contextualizes, and organizes other data. A column means, where it came from, who owns it, when it last changed
For meadata, the shared key is a ________ that the same names mean the same thing. There are 3 systems, 3 definitions of a customer , need a key that records them
promise
For the cost of bad data, who pays
the analyst, customers, person who has to call bacck
For the cost of bad data, when does it catch up to you
months later when someone finally asks question the record can’t answer how
when it comes to the cost of bad data, how do you find out
usually you don’t find out. it shows up as time, rework, and a decision made on a wrong number
Why is bad data called a tax and not a cost
cost is paid once, where the mistake happened. Tax is paid by every piece of work downstream/accumulated of it. A trained model on the data pays the tax on every answer it gives
6 dimensions of data quality
completeness
validity
consistency
uniqueness timeliness
accuracy
completeness
is it there
validity
could it be true
consistency
does it agree with itself
uniqueness
is the same thing recorded twice
timeliness
is it current enough to act on
accuracy
is it right
Cleaning is deciding what the value should have been —> to clean data, it requires personal ___ on whether something is needed or not
judgement
3 questions that decide whether cleaning can happen
will the organization accept a standard
can the souce system be fixed or only the copy
when 2 systems disagree, who decides which one wins
A defect class in data quality is a ___, not a ___ ___
pattern, broken cell
what’s an example of a symptom
row 7 looks wrong (observation)
interprete, a class is a pattern, not a broken cell
data quality classification focuses on identifying the structural type or behavior orf data issues, rather than pointing out individual isolated errors
what’s an example of a class
the state is written 2 ways
3 ways to find a defect class
profile the column- list the distinct values
compare against a rule - a date field has to hold a valid date
compare against otehr systems - 2 sources that should agre, and a list of the rows where they don’t
after finding defect classes, what can you do (2)
standardize it, flag it and leave it —> both will get written down!
Generative AI can be __________, looking for future trends and stuff
predictive
What did the reading say snake oil was
AI is often sold for far more than money, but decisions about ppls lives. It can be seen as a “cure” but it harms ppl as well
Machine learning identifies patterns using _______ _______ and compunters by unearthing _________ in data sets. You can use it to make predictions
statistical learning
boundaries
One method for making predictions is called _________ __________,
decision tree wh
decision tree function
series of if- then statements to identify boundaries and define patterns in the data
Overfitting
model does well on training examples but poorly on new examples it has never seen
occurs when a machine learning model fits its training data too closely, memorizing random noise and minor fluctuations instead of learning generalizable patterns
models can be adjusted to change the way they fit the data. These settings are called _____________
parameters
Simplest version of decision tree = ________. Why is it prone to bias
stump
ignores complexities. Black and white interpretation of the data
overly complex trees suffer from errors due to _________. Data tends to be _____________
variance
inconsistent
A CEO reads that an AI chatbot passed a bar exam and tells the board the company's AI resume-screening vendor "clearly works — AI handles law exams now." Per Narayanan & Kapoor, the reasoning fails because:
The bar-exam claim must be exaggerated — no AI can really do that The screener just needs more training data to catch up to the chatbot
The two tools share a label, not a technology — one proves nothing about the other
Neither tool can be judged until vendors publish accuracy numbers
c
A procurement lead is comparing four AI pitches. Per the authors' categories, which claim deserves the hardest skepticism before buying?
A tool that transcribes meeting audio into searchable text
A tool that predicts which job applicants will still be employed in five years
A tool that drafts marketing images from text prompts
A tool that autocompletes code as developers type
b
After a text generator fabricates a citation in a research memo, a director declares "it's all snake oil" and moves to cancel every AI project, including a working transcription rollout. On the authors' framework, the director is:
Right — one fabrication is enough to reveal the category as fraud
Wrong only tactically — wait for a second failure before canceling
Wrong on both counts — vendor accuracy claims already settle the question
Collapsing categories — generative AI is imperfect but real; snake oil is AI that can't work as advertised
d
An analyst grows a decision tree to flag which support tickets will escalate. It classifies 100% of the historical tickets used to build it correctly — and 68% of the next month's tickets. Per R2D3's walkthrough, the gap signals:
Overfitting — the tree memorized the training data's quirks, not the general pattern
Bad inputs — next month's tickets were probably logged incorrectly
Too simple a model — add splits until the new tickets classify correctly
Nothing — accuracy always drops on new data, so the test is uninformative
a
What does power query applied steps do
records every data transformation step in order so the procedure can be rerun on multiple or updated files
what is deterministic computing
computing where a person defines the rules and the same input produces the same output every time
what is deterministic compuputing examples
sql queries and power query
What is probabilistic computing
computing where a model learns patterns from past data and produces estimates (can be wrong
examples of probabilistic computing
fraud scores, chatbot answers
What is AI
broad field of technology that enables systems to perform tasks associated with human intelligence
what is machine learning (ML)
field within AI in which models learn patterns from data
Deep learning
Type of machine learning that uses neutral network to learn complex patterns
Large language model LLM
A type of deep learning model trained on large amounts of text to predict likely next words
how are AI, ML. Deep learing, and LLMs related
AI is the broadest field
ML is within AI
Deep learning is within ML
LLM are a type of deep learning
what are the 3 main ways of ML
supervised learning, unsupervised learning, reinforcement learning
Supervised learning
learns from examples that have labels identifying the correct answers
What’s labeled: known correct outcomes like who definitely left or stayed
unsupervised learning
learning patterns or structure from data without labled answers
Reinforcement learning
learning which actions work best through rewards and penalties
ex super mario bros
classification
predicting which category smt belongs to
regression
predicting a numerical value such as next months sales c
churn
when customer stops using company products or services p
purpose of a cusotmer churn model
to predict which customers are likely to leave company so compnay can try to retain them
Decision tree
machine learning model that makes predictions using a sequence of questions about features. Learns through using historical labeled examples to identify questions that best separate different outcomes
why might a decision tree ignore a feature
if that feature doesn’t help determine outcomes
training vs scoring
training teaches a model using past data
scoring appies trained data to new data
unsupervised association mining
method that finds items frequently appeared together, such asprodcuts purchased in the same basket
lift in association mining
measure of how much more often 2 items occur together than expected by chance. Lift above 1 is a positive association
is deep learning check fraud detection supervised or unsupervised
supervised learning cause the model learns from past checks labeled as forged or genuine
neutral network and how it learns
model made of interconnected units who connection strenghts can be adjusted learn patterns.
learns through predicting labels and compares it’s predictions with already known labels. It adjust connection strengths and repeats process across examples
How does decision tree differ from neutral network in explainability
decision tree can be read directly
neutral networks learned connections are harder to interpret
Latent capacity
ability such as solving problems or producing creative looking writing, that emerges as model becomes capable even thhough it wasn’t explicitly trained for that exact teask
confabulation in AI
model fills GAPS with plausible sounding info that may be made up
hallucination
perceiving something that’s not there
why do language model confabulate
predict likely words, but don’t guarantee truth
3 ways to reduce confabulation
provide a source and require source based answer
verify specific details like names and dates
ask the model to admit when it deesn’t know
RAG
Retrieval augmented generation- method that provides model with SOURCE material so it can answer using retrieval information
anthropomorphization
describing a model as if it were a person with thoughts, beliefs, intentions, feelings . should avoid cause it can make a model text prediction seem like human understanding, belief or intention
Chatbot v AI agent
chatbot answers questions
agent uses steps, tools, to carry out task
model context protocol
common standard through which agents can conneect to tools and other software
explainability
whether a person can say why a model produced a particular output
Why can some models explain their outputs and others cannot?
Explainability
6. Why is “AI” a label for many technologies, and which claims deserve skepticism?
it’s an umbrella term of different systems, ML, deep learning, LLM, etc
Format for a spec
Part a = executives (problem)
part b = systems (scope