what is ai ? (normally required human intelligence)
visual perception
speech recognition decision-making
translating between languages etc.
History
In 1950 alan turing : if a machine could carry on a conversion that was that indistinguishable from a conversation with a human being , then it was a responsible to that the machine was "thinking".
Dartmouth conference 1956.
1970 - AI winter.
1990 - MI is separate field.
1997 - IBM's Deep Blue Beats the world champion at chess.
2010 - ML becomes integral to many widely used services.
Data and Alg give you model .
business > data accquisition & Understading > modeling > deployment .
training set (usually 70-80% of data)
validate set (usually 10-20% of data)
test set (usually 10-20% of data)
training > Alg> model (result in)
Type of alg.
supervised alg :- data already label . used to predict target attribute/label.
unsupervised alg :- doesn't have label . not have target attribute . find the pattern among the input datasets. problems can be further grouped into clustering.
semi supervised :- combination of supervised and unsupervised alg. eg. photo archive.
Algorithms
1. regression (supervised)
Linear regression :- y=mx+c where m and c are constant.
adv
simple
based on math
lower training time
continuous value prediction
dis-adv
not solve complex problem
uses
predicting
stock price prediction
credit assessment
2. Classification
used to categorical response values, where the data can separated into specific "classes".
Two- class classification : - predicts between two category .
Multi -class classification :- predicts between several categories.
3. Decision Tree
flowchart like structure .
created upside down with root at top.
adv.
white box , easy to interpret and explain.
useful to find the most imp attribute.
not affected by outliers , less data cleaning required.
once crated , can provide fast classification .
dis adv.
work with discrete value better than continuous value.
requires a lot of prior data.
limited to one output without probability.
Not great in regression .
Uses
Astronomy
financial analysis
power system
4. clustering (unsupervised)
4.1. K-means
adv.
no nned for classified input data.
no need to have information about attributes.
dis adv.
results for different for 2 successive runs.
hard to find good means.
specify no of clusters.
uses
E-commerce
credit
super-markets
doc.classification
5. Neural Network
Artificial Neural Networks (ANN)
adv.
great with even continuous value attributes.
can have 1000 of attributes.
no need to understand to domain/problem
higher accuracy
dis. adv.
Black box , not possible to check results .
training takes a long time.
sometimes lack of domain knowledge exposed later.
uses
predictive analysis
click stream analysis
fraud detection
image recognize
Comments
Post a Comment