Part A: Clustering Methods (40 Marks)

For this question, you will explore the clustering methods you have learned 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 to explore are:

1.   K-Means

2.   DBSCAN and

3.   Agglomerative.

The three datasets are:

1.  Seoul Bike Sharing Demand

2.  Facebook Live Sellers in Thailand Data Set

3.  Water Treatment Plant

You need to complete 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 is taken, Sum of Squares Errors (SSE), Cluster Silhouette Measure (CSM). You may use the Davis-Bouldin score as an alternative to SSE. Submit the 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.                                           [7 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.                 [7 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.                                     [7 marks]

Task 2

a)   For each dataset identify which clustering algorithm performed best. Justify your answer. If no single algorithm performs best on all three performance measures, you will need to carefully consider how you will rate each ofthe measures and then decide how

you will produce an overall measure that will enable you to rank the algorithms.        [7 marks]

b)   For each winner algorithm and 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.

[7 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]

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 theEnvironmental 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 PM lag measurements, lag1 and lag2, should be included in your dataset. For example, lag1 for PM2.5 is measurements 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.

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 training iteration 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 ofneurons 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 size constant at 0.01 and 4 respectively. Provide a line plot ofthe 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

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

2)   Compare the performance of both MLP and LSTM using RMSE. Which model performed