The document provides an overview of machine learning concepts including the distinction between expert-driven and data-driven decisions, the process of creating machine learning models, and various algorithms like decision trees and logistic regression. It also covers methods for model evaluation, the use of ensembles, clustering techniques, anomaly detection, and association discovery, explaining their practical applications and underlying principles. Additionally, the document discusses the differences between generative and discriminative models and introduces latent Dirichlet allocation for document analysis.

1. Class summary

2. BigML, Inc.
2
Day 1 – Morning sessions
Class su

3. BigML, Inc.
3
A Gentle Introduction to
Machine Learning
Charles Parker
● Experts who extract some
rules to predict new results
● Programmers who tailor a
computer program that
predicts following the
expert's rules.
● Non easily scalable to the
entire organization
● Data (often easily to be
found and more accurate
than the expert)
● ML algorithms
(faster, more modular,
measurable performance)
● Scalable to the entire
organization
What is your company's strategy based on?
Expert­driven decisions Data­driven decisions

4. BigML, Inc.
4
A Gentle Introduction to
Machine Learning
When data­driven decisions are a good idea
● Experts are hard to find or expensive
● Expert knowledge is difficult to be programmed into
production environments accurately/quickly enough
● Experts cannot explain how they do it: character or speech
recognition
● There's a performance­critical hand­made system
● Experts are easily found and cheap
● Expert knowledge is easily programmed into production
environments
● The data is difficult or expensive to acquire
When data­driven decisions are a bad idea

5. BigML, Inc.
5
A Gentle Introduction to
Machine Learning
Steps to create a ML program from data
● Acquiring data
In tabular format: each row stores the information about the
thing that has a property that you want to predict. Each column
is a different attribute (field or feature).
● Defining the objective
The property that you are trying to predict
● Using an ML algorithm
The algorithm builds a program (the model or classifier) whose
inputs are the attributes of the new instance to be predicted and
whose output is the predicted value for the target field (the
objective).

6. BigML, Inc.
6
A Gentle Introduction to
Machine Learning
Modeling: creating a program with an ML algorithm
● The algorithm searches in a Hypothesis Space the set of
variables that best fits your data
Examples of Hypothesis Spaces:
● Logistic regression: Features coefficients + bias
● Neural network: weights for the nodes in the network
● Support vector machines: coefficients on each training point
● Decision trees: combination of features ranges

7. BigML, Inc.
7
A Gentle Introduction to
Machine Learning
Decision tree construction
● What question splits better you data? try all possible splits
and choose the one that achieves more purity
● When should we stop?
When the size of the subset is totally pure
When the size reaches a predetermined minimum
When the number of nodes or tree depth is too large
When you can’t get any statistically significant
improvement
● Nodes that don’t meet the latter criteria can be removed
after tree construction via pruning
The recursive algorithm analyzes the data to find

8. BigML, Inc.
8
A Gentle Introduction to
Machine Learning
Visualizing a decision tree
Root node
(split at petal length=2.45)
Branches
Leaf
(splitting stops)

9. BigML, Inc.
9
A Gentle Introduction to
Machine Learning
Decision tree outputs
● Prediction: Start from the root node. Use the inputs to
answer the question associated to each node you reach.
The answer will decide which branch will be used to
descend the tree. If you reach a leaf node, the majority class
in the leaf will be the prediction.
● Confidence: Degree of reliability of the prediction. Depends
on the purity of the final node and the number of instances
that it classifies.
● Field importance: Which field is more decisive in the
model's classification. Depends on the number of times it is
used as the best split and the error reduction it achieves.
Inputs: values of the features for a new instance

10. BigML, Inc.
10
A Gentle Introduction to
Machine Learning
Evaluating your models
● Testing your model with new data is the key to measure its
performance. Never evaluate with training data!
● Simplest approach: split your data into a training dataset
and a test dataset (80­20% is costumary)
● Advanced approach: to avoid biased splits, do it repeatedly
and average evaluations or k­fold cross­validate.
● Accuracy is not a good metric when classes are
unbalanced. Use the confusion matrix instead or phi, F1­
score or balanced accuracy.
Which evaluation metric to choose?

11. BigML, Inc.
11
● Confusion matrix can tell the number of correctly classified
(TP, TN) or misclassified instances (FP, FN) but this does not
tell you how misclassifications will impact your business.
● As a domain expert, you can assign a cost to each FP or FN
(cost matrix). This cost/gain ratio is the significant
performance measure for your models.
A Gentle Introduction to
Machine Learning
Domain specific evaluation

12. BigML, Inc.
12
● Ensembles are groups of different models built on
samples of data.
● Randomness is introduced in the models. Each
model is a good approximation for a different
random sample of data.
● A single ML Algorithm may not adapt nicely to
some datasets. Combining different models can.
● Combining models can reduce the over-fitting
caused by anomalies, errors or outliers.
● The combination of several accurate models gets us
closer to the real model.
Ensembles
Can a group of weaker models outperform a stronger
single model?
Poul Petersen

13. BigML, Inc.
13
● Bootstrap aggregating (bagging) models are built on
random samples (with replacement) of n instances.
● Random decision forest in addition to the random samples
of bagging, the models are built by choosing randomly the
candidate features at each split (random candidates).
● Plurality majority wins
● Confidence weighted each vote is weigthed by confidence
and majority wins
● Probability weighted each tree votes according to the
distribution at its prediction node
● K-Threshold a class is predicted only if enough models vote
for it
● Confidence Threshold votes for a class are only computed
if their confidence is over the threshold
Ensembles
Types of ensembles
Types of combinations

14. BigML, Inc.
14
● How many trees?
● How many nodes?
● Missing splits?
● Random Candidates?
● SMACdown: automatic optimization of ensembles
by exploring the configuration space.
Ensembles
Configuration parameters
Too many parameters? Automate!

15. BigML, Inc.
15
● Regressions are typically
used to relate two numeric
variables
● But using the proper function
we can relate discrete
variables too
Logistic Regression
How comes we use a regression to classify?
Logistic Regression is a classification ML Algorithm
Poul Petersen

16. BigML, Inc.
16
● We should use feature engineering to transform
raw features in linearly related predictors, if
needed.
● The ML algorithm searches for the coefficients to
solve the problem
by transforming it into a linear regression problem
In general, the algorithm will find a coefficient per
feature plus a bias coefficient and a missing
coefficient
Logistic Regression
Assumption: The output is linearly related to the
predictors.

17. BigML, Inc.
17
• Bias: Allows an intercept term. Important if
P(x=0) != 0
• Regularization
L1: prefers zeroing individual coefficients
L2: prefers pushing all coefficients towards
zero
• EPS: The minimum error between steps to
stop.
• Auto-scaling: Ensures that all features
contribute equally. Recommended unless there is
a specific need to not auto-scale.
Logistic Regression
Configuration parameters

18. BigML, Inc.
18
• Multi-class LR: Each class has its own LR computed
as a binary problem (one-vs-the-rest). A set of
coefficients is computed for each class.
• Non-numeric predictors: As LR works for numeric
predictors, the algorithm needs to do some encoding
of the non-numeric features to be able to use them.
These are the field-encodings.
– Categorical: one-shot, dummy encoding, contrast
encoding
– Text and Items: frequencies of terms
● Curvilinear LR: adding quadratic features as new
features
Logistic Regression
Extending the domain for the algorithm

19. BigML, Inc.
19
Logistic Regression
Logistic Regressions versus Decision Trees
● Expects a "smooth" linear
relationship with predictors
● LR is concerned with
probability of a discrete
outcome.
● Lots of parameters to get
wrong: regularization,
scaling, codings
● Slightly less prone to over-
fitting
● Because fits a shape, might
work better when less data
available.
● Adapts well to ragged
non-linear relationships
● No concern:
classification, regression,
multi-class all fine.
● Virtually parameter free
● Slightly more prone to
over-fitting
● Prefers surfaces parallel
to parameter axes, but
given enough data will
discover any shape.

20. BigML, Inc.
20
Day 1 – Evening sessions

21. BigML, Inc.
21
● Clustering is a ML technique designed to find
and group of similar instances in your data
(group by).
● It's unsupervised learning, as opposed to
supervised learning algorithms, like decision
trees, where training data has been labeled
and the model learns to predict that label.
Clusters are built on raw data.
● Goal: finding k clusters in which similar data
can be grouped together. Data in each cluster
is similar self similar and dissimilar to the rest.
Clusters
Clusters: looking for similarity
Poul Petersen

22. BigML, Inc.
22
● Customer segmentation: grouping users to act on
each group differently
● Item discovery: grouping items to find similar
alternatives
● Similarity: Grouping products or cases to act on each
group differently
● Recommender: grouping products to recommend
similar ones
● Active learning: grouping partially labeled data as
alternative to labeling each instance
Clustering can help us to identify new features shared
by the data in the groups
Clusters
Use cases

23. BigML, Inc.
23
● K-means: The number of expected groups is given by the user. The
algorithm starts using random data points as centers.
– K++: the first center is chosen randomly from instances and each
subsequent center is chosen from the remaining instances with
probability proportional to its squared distance from the point's
closest existing cluster center
Clusters
Types of clustering algorithm
The algorithm computes distances based
on each instance features. Each instance
is assigned to the nearest center or
centroid. Centroids are recalculated as the
center of all the data points in each
cluster and process is repeated till the
groups converge.
● G-means: The number of groups is also
determined by the algorithm. Starting
from k=2, each group is split if the data
distribution in it is not Gaussian-like.

24. BigML, Inc.
24
How distance between two instances is defined?
For clustering to work we need a distance function that must
be computable for all the features in your data. Scaled
euclidean distance is used for numeric features. What about
the rest of field types?
Categorical: Features contribute to the distance if categories
for both points are not the same
Text and Items: Words are parsed and its frequencies are
stored in a vector format. Cosine distance (1 – cosine
similarity) is computed.
Missing values: Distance to a missing value cannot be
defined. Either you ignore the instances with missing values
or you previously assign a common value (mean, median,
zero, etc.)
Clusters
Extending clustering to different data types

25. BigML, Inc.
25
● Anomaly detectors use ML algorithms
designed to single out instances in your data
which do not follow the general pattern (rank
by).
● As clustering, they fall into the unsupervised
learning category, so no labeling is required.
Anomaly detectors are built on raw data.
● Goal: Assigning to each data instance an
anomaly score, ranging from 0 to 1, where 0
means very similar to the rest of instances
and 1 means very dissimilar (anomalous).
Anomaly Detection
Anomaly detection: looking for the unusual
Poul Petersen

26. BigML, Inc.
26
● Unusual instance discovery
● Intrusion Detection: users whose behaviour does not
comply to the general pattern may indicate an intrusion
● Fraud: Cluster per profile and look for anomalous
transactions at different levels (card, user, user groups)
● Identify Incorrect Data
● Remove Outliers
● Model Competence / Input Data Drift: Models
performance can be downgraded because new data has
evolved to be statistically different. Check the
prediction's anomaly score.
Anomaly Detection
Use cases

27. BigML, Inc.
27
Anomaly Detection
Statistical anomaly indicators
● Univariate­approach: Given a single
variable, and assuming normal distribution
(Gaussian). Compute the standard
deviation and choose a multiple of it as
threshold to define what's anomalous.
● Benford's law: In real­life numeric sets
the small digits occur disproportionately
often as leading significant digits.

28. BigML, Inc.
28
Anomaly Detection
Isolation forests
● Train several random
decision trees that over­fit
data till each instance is
completely isolated
● Use the medium depth of
these trees as threshold to
compute the anomaly
score, a number from 0 to 1
where 0 is similar and 1 is
dissimilar
● New instances are run
through the trees and
assigned an anomaly score
according to the average
depth they reach

29. BigML, Inc.
29
● Association Discovery is an unsupervised technique, like
clustering and anomaly detection.
● Uses the “Magnum Opus” algorithm by Geoff Webb
Association Discovery
Geoff Webb and Poul Petersen
Looking for “interesting” relations between variables
date customer account auth class zip amount
Mon Bob 3421 pin clothes 46140 135
Tue Bob 3421 sign food 46140 401
Tue Alice 2456 pin food 12222 234
Wed Sally 6788 pin gas 26339 94
Wed Bob 3421 pin tech 21350 2459
Wed Bob 3421 pin gas 46140 83
Tue Sally 6788 sign food 26339 51
{class = gas} amount < 100
{customer = Bob, account = 3421} zip = 46140
Antecedent Consequent

30. BigML, Inc.
30
Association Discovery
Use Cases
Market Basket Analysis
Web usage patterns
Intrusion detection
Fraud detection
Bioinformatics
Medical risk factors

31. BigML, Inc.
31
Association Discovery
Problems with frequent pattern mining
●
Often results in too few or too many patterns
●
Some high value patterns are infrequent
●
Cannot handle dense data
●
Cannot prune search space using constraints on
relationship between antecedent and consequent eg
confidence
●
Minimum support may not be relevant
●
Cannot be low enough to capture all valid rules cannot
be high enough to exclude all spurious rules

32. BigML, Inc.
32
● Very high support patterns can be spurious
● Very infrequent patterns can be significant
So the user selects the measure of interest
System finds the top-k associations on that
measure within constraints
– Must be statistically significant interaction
between antecedent and consequent
– Every item in the antecedent must increase
the strength of association
Association Discovery
It turns out that:

33. BigML, Inc.
33
Association Discovery
Measures:
Coverage
Support
Confidence
Lift Leverage
Support/
Coverage
Ratio Difference

34. BigML, Inc.
34
Association Discovery
Meaningful relations:

35. BigML, Inc.
35
● Generative models try to fit the coefficients of a
generic function to use it as data generating
function. This conveys information about the
structure of the model (looking for causality).
● Discriminative models, do not care about how the
labeling is generated, they only find how to split the
data into categories
● Generative models are more probabilistically sound
and able to do more than just classify
● Discriminative models are faster to fit and quicker to
predict
Latent Dirichlet Allocation
Generative vs discriminative models
Charles Parker
Pros and Cons

36. BigML, Inc.
36
A document can be analyzed from different levels
● According to its terms (one or more words)
● According to its topics (distributions of terms ~
semantics)
● Documents are generated by repeatedly drawing
a topic and a term in that topic at random
● Goal: To infer the topic distribution
How? Dirichlet Process is used to model the
term|topic, and topic|document distributions
Latent Dirichlet Allocation
Thinking of documents in terms of Topics
Generative Models for documents

37. BigML, Inc.
37
● We're more likely to think a word came from a topic if
we've already seen a bunch of other words from that
topic
● We're more likely to think the topic was responsible
for generating the document if we've already seen a
bunch of words in the document from that topic.
● Visualizing topic changes in documents over time
(specially for dated historical collections)
● Search by topics (without keywords)
● Using topics as a new feature instead of the bag of
words approach in modeling
Latent Dirichlet Allocation
Dirichlet Process intuitions
Applications

38. BigML, Inc.
38
● Topics can reduce the feature space
● Are nicely interpretable
● Automatically tailored to the document
● Need to choose the number of topics
● Takes a lot of time to fit or do inference
● Takes a lot of text to make it meaningful
● Tends to focus on “meaningless minutiae”
● While sometimes makes nice classifications, it's usually not a
dramatic improvement over bag-of-words
● Nice for exploration
Latent Dirichlet Allocation
Nice properties about topics
Caveats