Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit

SPCE0038 - Machine Learning with Big Data

27-Apr-2022

Question 1

Consider a multi-class classification problem.

(a) Describe the one-versus-rest (OvR) (also called one-versus-all; OvA) strategy to perform multiclass classification given a binary classifier. [2 marks]

(b) Describe the one-versus-one (OvO) strategy to perform multiclass classification given a binary classifier. [2 marks]

(c) Given N classes, how many binary classifiers are required for the one-versus-rest (OvR) and one-versus-one (OvO) strategies. [2 marks]

(d) Specify the advantages and disadvantages of the one-versus-rest (OvR) and one- versus-one (OvO) strategies. [4 marks]

(e) Describe how a neural network can be used to perform multi-class classification. Include a diagram of the neural network architecture to support your discussion (for the purpose of your diagram, assume a fully-connected network with 5 inputs, 2 hidden layers and, 4 classes to be classified). [6 marks]

(f) Define the softmax function mapping inputs aj  to outputs pj , for j = 1, . . . , n. [1 mark]

(g) Show the outputs of the softmax function pj  satisfy the following properties:

(i)     j(n)11 pj  = 1;

(ii) 0 = pj  = 1 for all j . [3 marks]

Question 3

In a decision tree define the Gini threshold measure.  Describe what value of 0

and value of 0.5 mean. [5 marks]

(b) What is an alternative to using the Gini measure?  Define this, and comment on what effect this may have on a decision tree compared to the Gini measure. [5 marks]

(c) Describe and define the CART algorithm,  including what the acronym CART stands for. Define the cost function that CART uses. [5 marks]

(d) List five ways in which decision tree classification may be regularised. [5 marks]

Question 4

You are working on a project involving data from a number of sources.  Before you build your model, you must prepare the data to make it easier to work with.

(a) Give 2 reasons why you may need to perform some preprocessing rather than work directly with the raw data. [2 marks]

(b) List 2 data cleaning practices that may result in having fewer data points after preprocessing. [2 marks]

Some of the data you need is located in a database and must first be extracted.

(c) Briefly describe how the data in a relational database is structured. [2 marks]

(c) What is the standard language for interacting with relational databases, and which keyword (or kind of query) is used to extract a subset of entries? [1 marks]

You decide to use DVC to automate the different steps of your analysis. With this in mind, you have written the following dvc .yaml file:

s t a g e s :

load _ a nd _ split :

cmd :   python   p r e p a r e _ a l l _ d a t a . py

deps :

_  p r e p a r e _ a l l _ d a t a . py

o u t s :

_  d a t a _ t r a i n . j s o n

_  data _ test . j s o n

t r a i n :

cmd :   python   model . py

deps :

_  d a t a _ t r a i n . j s o n

_  model . py

o u t s :