COMP3425 Data Mining S1 2024 

Assignment 2

Maximum marks              100

Weight                            25% of the total marks for the course

Length                            Maximum of 10 pages, excluding cover sheet, bibliography and appendices.

Layout                             A4 margin, at least 11-point type size, use of typeface, margins and headings consistent with a professional style.

Submission deadline          9:00am, Monday, 6 May

Submission mode              Electronic, via Wattle

Estimated time                 15 hours

Penalty for lateness           100% after the deadline has passed

First posted:                      25 th March, 9:00 AM

Last modified:                     9 th April 2:00 PM

Questions to:                       Wattle Discussion Forum

This assignment specification may be updated to reflect clarifications and modifications after it is first issued. -It is strongly suggested that you start working on the assignment right away. You can submit   as many times as you like. Only the most recent submission at the due date will be assessed.

In this assignment, you are required to submit a single report in the form of a PDF file.  You may also   attach supporting information (appendices) as one or more identified sections at the end of the same PDF file. Appendices will not be marked but may be treated as supporting information to your report. Please use a cover sheet at the front that identifies you as author of the work using your u-number and name and identifies this as your submission for COMP3425 Assignment 2. The cover sheet and appendices do not contribute to the page limit.

You are expected to write in a style appropriate to a professional report. You may refer to

http://www.anu.edu.au/students/learningdevelopment/writing-assessment/report-writing for some stylistic advice. You are expected to use the question and sub-question numbering in this assignment to identify the relevant answers in your report.

No particular layout is specified, but you should use no smaller than 11-point typeface and stay within the maximum specified page count. Page margins, heading sizes, paragraph breaks and so forth are not specified but a professional style must be maintained. Text beyond the page limit will be treated as non-existent.

This is a single-person assignment and should be completed on your own.  The use of any Generative AI tools is not permitted. Make certain you carefully reference all the material that you use, although the nature of this assignment suggests few references will be needed.  It is unacceptable to cut and paste another author's work and pass it off as your own. Anyone found doing this, from whatever source, will get a mark of zero for the assignment and, in addition, CECC procedures for plagiarism will apply.

No particular referencing style. is required. However, you are expected to reference conventionally, conveniently, and consistently.  References are not included in the page limit. Due to the context in  which this assignment is placed, you may refer to the course notes or course software where appropriate (e.g. “For this experiment Rattle was used”), without formal reference to original sources, unless you copy text or images which always requires a formal reference to the source. You do not need to reference this specification.

An assessment rubric is provided. The rubric will be used to mark your assignment. You are advised to use it to supplement your understanding of what is expected for the assignment and to direct your effort towards the most rewarding parts of the work.

Your submission will be treated confidentially.  It will be available to ANU staff involved in the course for marking.  It may be shared, de-identified, as an exemplar for other students.

Task

You are to complete the following exercises, using the supplied data set. For simplicity, the exercises are expressed using the assumption that you are using Rattle, however you are free to use R directly  or any other data mining platform. you choose that can deliver the required functions. You should describe the methods used in terms of the language of data mining, not in the terms of commands   you typed or buttons you selected. You are expected, in your own words, to interpretselected tool output in the context of the learning task. Write just what is needed to explain the results you see.

1. Platform

Briefly describe the platform. for your experiments in terms of memory, CPU, operating system, and software that you use for the exercises. If your platform is not consistent throughout, you must describe it for each exercise. This is to ensure your results are reproducible.

2. Data

(a)  In your own words, briefly describe the purpose and means of data collection.

(b) Look at the pairwise correlation amongst ordinal variables using Pearson product-moment

correlation.  Qualitatively describe the pairwise correlations amongst A1 and each of the variables    RA2_a, RA2_b RA2_c, RA2_dand RA2_e. Explain what you see in terms of the meaning of the data.

3. Association mining: What factors affect satisfaction with the country’sfuture?

A1 of the survey asks respondents how they feel about the way the country is heading. Your task is to use association mining to find out which factors might be associated with a person’s response to A1.

(a)  Generate association rules, adjusting min_support and min_confidence parameters as you need.   What parameters do you use? Bearing in mind we are looking for insight into what factors affect RA1 A1, find 3 interesting rules, and explain both objectively and subjectively why they are interesting.

(b) Comment on whether, in general, association mining could be a useful technique on this data.

4.  Study a simple classification task

Restore the dataset to its original distributed form, if you made any changes above.

Aim to build a model to classify how people voted in the referendum, relying on the variable RB1.

To do this, define a new variable Ass2_voted to map those who voted Yes to TRUE, those who voted No to FALSE and all others to NA.  For example, in Excel you can create a new column, give it the header Ass2_voted, enter the following formula to row 2, then copy into all other rows:

=IF(AS2=1,TRUE, IF(AS2=2, FALSE,"NA"))

Now remove or otherwise ignore all rows in which people responded to RB1 with something other   than yes or no. Use RB1 Ass2_voted as the target class and set every other variable (except srcid) as Input (independent).  Remove (or Ignore) variable RB1 itself.

Using sensible defaults for model parameters is fine for this exercise where we aim to compare methods rather than optimise them.  You may choose to remove (or Ignore) some other variables from the data. If you do so, explain what they were and why you removed them.

Keep a copy of your modified form. of the data for a later purpose.

(a) Why were you asked to Ignore or otherwise remove the original RB1 variable for this purpose? Why? Hint: Think how Ass2_voted  is defined.

(b) Train each of a Linear, Decision tree, SVM and Neural Net classifier, so you have 4 classifiers.

Hint: Because the dataset is large, begin with a small training set, 20%, and where run-time speeds are acceptable, move up to a 70% training set. Evaluate each of these 4 classifiers, using a confusion matrix and interpreting the results in the context of the learning task.

(c) Inspect the models themselves where that is possible to assist in your evaluation and to explain the performance results.  Which learner(s) performed best and why?

5. Predict a Numeric Variable

Return to working with the original form. of the data as distributed.

Weight_final_ref is a factor that is calculated to weight each respondent’s answers according to the extent to which they represent a demographic portion of the population, based on the values for the variables p_geography,p_region, .., p_state_sdc (excluding  weight_final_ref   and weight_ref). You   are to train a regression tree or a neural net to predict weight_final_ref, using those demographic variables as input (excluding weight_ref).

(a) Explain which you chose of a regression tree or neural net and justify your choice.

(b) Train your chosen model and tune by setting controllable parameters to achieve a reasonable performance. Explain what parameters you varied and how, and the values you chose finally.

(c) Assess the performance of your best result using the subjective and objective evaluation appropriate for the method you chose and justify why you settled with that result.

6. More Complex Classification

Return to using the modified data you created for Question 4.

Here you are going to work harder to attempt to build a great model to classify Ass2_voted. Hint: Ensure variable RB1 is not used.

(a) Explain how you will partition the available dataset to train and validate classification models in (b) to (d) below.

(b) Train a Decision Tree Classifier. You will need to adjust default parameters to obtain optimal

performance. State what parameters you varied and (briefly) their effect on your results.

Evaluate your optimal classifier using the error matrix, ROC, and any quality information specific to the classifier method.

(c)  Proceed as for (b) above, training  and discussing an SVM Classifier instead of a Decision Tree.

(d) Proceed as for (b) above, training  and discussing a Neural Net instead of a Decision Tree.

7. Clustering

Restore the dataset to its original distributed form.

For clustering, use 5 variables of your choice from RD7_a  to RD7_j . Ignore all the other variables.

Rescale the variables to fallin the range [0-1] prior to clustering.  Use the full dataset for clustering (i.e. do not partition into train-test sets).

Experiment with clustering using the k-means algorithm by building cluster models for each of k= 2, 5, √ (the latter is a recommended default for dataset of size n) clusters. Choose your preferred k    and its cluster model fork-means to answer the following.

(a) Justify your choice of k as your preferred (Hint: have look at parts b-d below for each cluster model).

(b) Calculate the sum of the within-cluster-sum-of-squares for your chosen model.  The within-cluster-sum-of-squares is the sum of the squares of the Euclidean distance of each object from its cluster mean.  Discuss why this is interesting.

(c) Look at the cluster centres for each variable.  Using this information, discuss qualitatively how each cluster differs from the others.

(d) Use ascatterplot to plot (asample of) the objects projected on to each combination of 2 variables with objects mapped to each cluster by colour (Hint: The Data button on Rattle’s Cluster tab can do this).  Describe what you can see as the major influences on clustering.  Include the image in your answer.

8. Qualitative Summary of Findings (Hint: approx 1/2 page)

Comparatively evaluate the techniques you have used and their suitability or not for mining this data. This should be a qualitative opinion that draws on what you have found already doing the exercises above.  For example, what can you say about training and classification speeds, the size or other aspects of the training data, or the predictive power of the models built?  Finally, what else would you propose to investigate as a follow-up to your work presented here?