ASSIGNMENT TWO

Semester 2 - 2022


PAPER NAME: Data Mining and Machine Learning
PAPER CODE: COMP809
DUE DATE: Sunday 30 Oct 2022 at midnight
TOTAL MARKS: 100
Students’ Names: ………………………………………………………….……………………………………………….

Students’ IDs: ………………………………………………………….………………………………….………………….

Due date: 30 Oct 2022 midnight NZ time.
Late penalty: maximum late submission time is 24 hours after the due date. In this case,
a 5% late penalty will be applied.
Include your actual code (no screenshot) in an appendix with appropriate comments for
each task.
Note: This assignment should be complemented by a group of two students.
Submission: a soft copy needs to be submitted through the canvas assessment link.


INSTRUCTIONS:
1. The following actions may be deemed to constitute a breach of the General Academic
Regulations Part 7: Academic Discipline,
Communicating with or collaborating with another person regarding the Assignment
Copying from any other student work for your Assignment
Copying from any third-party websites unless it is an open book Assignment
Uses any other unfair means
2. Please email DCT.EXAM@AUT.AC.NZ if you have any technical issues with your submission
on Canvas immediately
3. Attach your code for all the datasets in the appendix section.
Student IDs Numbers: COMP809 S2 2022
2 | P a g e

Part A: Clustering Methods (40 marks)
For this question, you will explore the clustering methods you have learnt in this course. You have
been given datasets from three very different application environments and you are required to
explore three widely used clustering algorithms and deploy each of them on the different datasets.
The three algorithms that you have decided to explore are 1) K-Means 2) DBSCAN and 3)
Agglomerative.
The three datasets that you have been given are:
Dow Jones Index
Facebook Live Sellers in Thailand
Sales Transactions

You need to complete the following three tasks as detailed below.
Task 1
For each activity in this task, you must explain each dataset and perform data exploration, data pre-
processing and apply a suitable feature selection algorithm before deploying each clustering
algorithm. Your clustering results should include the following measures:
The time taken, Sum of Squares Errors (SSE), Cluster Silhouette Measure (CSM). You may use the
Davis-Bouldin score as an alternative to SSE. Submit Python code used for parts a) to c) below. You
only need to submit the code for one of the 3 datasets.
a) Run the K means algorithm on each of the three datasets. Obtain the best value of K using
either SSE and/or CSM. Tabulate your results in a 3 by 3 table, with each row corresponding
to a dataset and each column corresponding to one of the three measures mentioned above.
Display the CSM plot for the best value of the K parameter for each dataset. [5 marks]

b) Repeat the same activity for DBSCAN algorithm and tabulate your results once again, just as
you did for part a). Display the CSM plot and the 3 by 3 table for each dataset. [5 marks]

c) Finally, use the Agglomerative algorithm and document your results as you did for parts a)
and b). Display the CSM plot and the 3 by 3 table for each dataset. [5 marks]

Task 2
a) For each dataset identify which clustering algorithm performed best. Justify your answer.
In the event that no single algorithm performs best on all three performance measures, you
will need to carefully consider how you will rate each of the measures and then decide how
you will produce an overall measure that will enable you to rank the algorithms. [5 marks]

b) For each winner algorithm and for each dataset explain why it produced the best value for the
CSM measure. This explanation must refer directly to the conceptual design details of the
algorithm. There is no need to produce any further experimental evidence for this part of the
question. [5 marks]

c) Based on what you produced in a) above, which clustering algorithm would you consider to
be the overall winner (i.e., after taking into consideration performance across all three
datasets). Justify your answer. [5 marks]
Student IDs Numbers: COMP809 S2 2022

Part B: Predictions of Particulate Matter (PM2.5 or PM10) [60 marks]
Air pollution causes serious damage to public health and based on existing research, particulate matter
(PM) smaller than PM2.5 is currently considered to have the strongest correlation with the effects of
cardiovascular disease. Therefore, making accurate predictions of PM2.5 is a crucial task. In this part,
you are required to build prediction models based on multi-layer perceptron (MLP) and long short-term
memory (LSTM).

Dataset
The dataset for this experiment can be downloaded from the Environmental Auckland Data Portal. Your
dataset includes PM2.5 or PM10 (output) and different predictors such as air pollution, Air Quality Index
(AQI), and meteorological data collected on an hourly basis from only one air quality monitoring
stations station listed below:
Penrose Station (ID:7)
Takapuna Station (ID:23)
Two PMlag measurements, lag1 and lag2, should be included in your dataset. For example, lag1 for PM2.5
is the measurement for the previous hour (h-1) and lag2 is PM2.5 concentration for h-2.
Download relevant PM concentration, air pollution data (SO2, NO, NO2), and meteorological data
Solar Radiation (W/m2), Air Temperature (°C), Relative Humidity (%), Wind Direction (°), and
Wind Speed (m/s)). The dataset should be hourly measurement starting from January 2017 to
December 2021 (5 years).
Note 1: Not all mentioned independent variables are collected at these monitoring stations.
Note 2: The unit of measurement for PM and air pollution data should be (μg/m3).
Data Pre-processing [5 marks]
Make sure your dataset all has the same temporal resolution (i.e. hourly measurement). Perform data
exploration and identify missing data and outliers (data that are out of the expected range). For example,
unusual measurements of air temperature of 40(°C) for Auckland, Relative Humidity measurements
above 100, and negative or unexplained high concentrations are outliers.
Provide attribute-specific information about outliers and missing data. How can these affect
dataset quality?
Based on this analysis, decide, and justify your approach for data cleaning. Once your dataset
is cleaned move to the next step for feature selection.
Feature Selection [5 marks]
Choose five attributes of your dataset that has the highest correlation with PM2.5 concentration using
Pearson Correlation or any other feature selection method of your choice with justification.
Provide the correlation plot (or results of any other feature selection method of your choice)
and elaborate on the rationale for your selection.
Describe your chosen attributes and their influence on PMconcentration.
Provide graphical visualisation of variation of PMvariation.
Provide summary statistics of the PM concentration.
Provide summary statistic of predictors of your choice that has the highest correlation in tabular
format.
Student IDs Numbers: COMP809 S2 2022

Experimental Methods
Use 70% of the data for training and the rest for testing the MLP and LSTM models. Use a Workflow
diagram to illustrate the process of predicting PM concentrations using the MLP and LSTM models.
[5 marks]

For both models, provide root mean square error (RMSE), Mean Absolute Error (MAE), and correlation
coefficient (R2) to quantify the prediction performance of each model.

Multilayer Perceptron (MLP)
1) In your own words, describe multilayer perceptron (MLP). You may use one diagram in your
explanation (one page). [5 marks]

2) Use the sklearn.neural_network.MLPRegressor with default values for parameters and a single
hidden layer with k= 25 neurons. Use default values for all parameters and experimentally
determine the best learning rate that gives the highest performance on the testing dataset. Use this
as a baseline for comparison in later parts of this question. [5 marks]

3) Experiment with two hidden layers and experimentally determine the split of the number of
neurons across each of the two layers that gives the highest accuracy. In part 2, we had all k
neurons in a single layer, in this part we will transfer neurons from the first hidden layer to the
second iteratively in step size of 1. Thus, for example in the first iteration, the first hidden layer
will have k-1 neurons whilst the second layer will have 1, in the second iteration k-2 neurons
will be in the first layer with 2 in the second, and so on. [5 marks]

4) From the results in part 3 of this question, you will observe a variation in the obtained
performance metrics with the split of neurons across the two layers. Give explanations for some
possible reasons for this variation and which architecture gives the best performance.
[5 marks]

Long Short-Term Memory (LSTM)
1) Describe LSTM architecture including the gates and state functions. How does LSTM differ
from MLP? Discuss how does the number of neurons and batch size affect the performance of
the network? [5 marks]

2) To create the LSTM Model, apply Adaptive Moment Estimation (ADAM) to train the networks.
Identify an appropriate cost function to measure model performance based on training samples
and the related prediction outputs. To find the best epoch, based on your cost function
results, complete 30 runs keeping the learning rate and the number of batch sizes
constant at 0.01 and 4 respectively. Provide a line plot of the test and train cost function
scores for each epoch. Report the summary statistics (Mean, Standard Deviation,
Minimum and Maximum) of the cost function as well as the run time for each epoch.
Choose the best epoch with justification. [5 marks]

3) Investigate the impact of differing the number of the batch size, complete 30 runs keeping the
learning rate constant at 0.01, and use the best number of epochs obtained in previous step 2.
Report the summary statistics (Mean, Standard Deviation, Minimum and Maximum) of the cost
Student IDs Numbers: COMP809 S2 2022

function as well as the run time for each batch size. Choose the best batch size with
justification... [5 marks]

4) Investigate the impact of differing the number of neurons in the hidden layer while keeping the
epoch (step 2) and Batch size (step 3) constant for 30 runs. Report the summary statistics (Mean,
Standard Deviation, Minimum and Maximum) of the cost function as well as the run time.
Discuss how does the number of neurons affect performance and what is the optimal number of
neurons in your experiment? [5 marks]
Model Comparison
1) Plot model-specific actual and predicted PM to visually compare the model performance. What
is your observation? [2.5 marks]

2) Compare the performance of both MLP and LSTM using RMSE. Which model performed
better? Justify your finding. [2.5 marks]

Report Presentation [5 marks]