辅导COMP3425、讲解Data Mining、R辅导、讲解R程序设计
辅导留学生 Statistics统计、回归、迭代|讲解留学生Pro
COMP3425 Data Mining 2019
Assignment 2
Maximum marks 100
Weight 20% 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, 20 May
Submission mode Electronic, via Wattle
Estimated time 15 hours
Penalty for lateness 100% after the deadline has passed
First posted: 1
st April, 9am
Last modified: 1
st April, 9am
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 unumber
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 useful 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 follow a professional style and 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. 2
This is a single-person assignment and should be completed on your own. 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, CECS 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 which always requires a formal reference to the source.
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 assignment submission will be treated confidentially. It will be available to ANU staff involved
in the course for the purposes of marking.
Task
You are to complete the following exercises. 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 are expected, in your own
words, to interpret selected tool output in the context of the learning task. Write just what is needed
to explain the results you see. Similarly, 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.
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
Look at the pairwise correlation amongst all the numeric variables using Pearson product-moment
correlation.
(a) Explain why you would expect DayofYear and WeekofYear to be highly correlated.
(b) Qualitatively describe the correlations amongst the variables High, Low, Open, Close and
Volume. Explain what you see in terms of the source of the data.3
3. Association mining: What factors affect volume of sales?
(a) Compute and give the 5-number summary for Volume. Qualitatively describe what it tells you
about Volume.
(b) For this exercise, bin Volume into 5 categories using Quantiles. Why is quantile binning
appropriate for association mining with Volume?
When you have completed this question 3, remove the extra variable you created so they do not
interfere with other exercises.
(c) 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
Volume, find 3 interesting rules, and explain both objectively and subjectively why they are
interesting.
(d) Comment on whether, in general, association mining could be a useful technique on this data.
4. Study a very simple classification task
Aim to build a model to classify Change. Use Change as the target class and set every other variable
as Input (independent). Using sensible defaults for model parameters is fine for this exercise where
we aim to compare methods rather than optimise them.
(a) This should be a very easy task for a learner. Why? Hint: Think how Change 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 for 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
One investment strategy could rely on the previous day’s price to predict the opening price for a
stock, enabling you to place a buy or sell offer overnight ready for the next day. To predict the
opening price for a day, you cannot use any of the other prices or Volume for that same day as that
information is not available until the close of the day, when it is too late. So, using the variables you
have in the dataset, but ignoring all of High, Low, Close, Volume, Close-Open, Change, High-Low, and
HMLOL, aim to predict the opening price Open using the previous day’s closing price PriorClose. and
the date and stock-related variables. Use a regression tree or a neural net.
(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. 4
6. More Complex Classification
An alternative investment strategy might be to predict where there will be a big proportional change
in the price over a day, once the day has opened, and so a good opportunity for a short term gain
(or loss) that day.
(a) Transform HMLOL to a categoric class variable by binning into 2 classes, using a k-means
clustering of the HMLOL. The skewed distribution of HMLOL is helpful here as the higher values we
need for investing are relatively rare. When you have completed this question 6, remove the extra
variable you created so it does not interfere with other exercises. Now, be sure to ignore most of
the current day price and volume variables, that is ignore all of High, Low, Close, Volume, CloseOpen,
Change, High-Low, and HMLOL. Explain why HMLOL should be ignored. This time, use the
Open price, PriorClose and all the date and company description variables for learning.
Initially, use a small training set, 20%, and where run-time speeds are acceptable, experiment with a
larger training set. Explain how you will partition the available dataset to train and validate
classification models 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) Train an SVM Classifier. Then proceed as for (b) Decision Tree above.
(d) Train a Neural Net classifier. Then proceed as for (b) Decision Tree above.
7. Clustering
(a) Restore the dataset to its original distributed form, removing any new variables you have
constructed above. For clustering, use only the five raw variables, Date, Open, High, Low and Volume
and remove all of the others.
Experiment with clustering using the k-means algorithm. Rescale the variables to fall in the range 0-
1 prior to clustering. Use the full dataset for clustering (do not partition) by building cluster models
for each of k= 2, 5 and the recommended default (i.e. squareroot (n/2) for dataset of size n)
clusters. Choose your preferred k and it’s cluster model for k-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 withincluster-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 a scatterplot to plot (a sample 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 5
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 (approx 1/2 page)
Would you use these results to advise your investment decisions?
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 to assist your investment decisions?6
Assessment Rubric
This 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 assignment will be marked out of 100, and marks will be scaled back to contribute to
the defined weighting for assessment of the course.
Review
Criteria
Max
Mark
Exemplary Excellent Good Acceptable Unsatisfactory
1. Platform & 2.
Data
10 9-10
1.Platform description
complete (memory, CPU,
operating system, software).
2. All required correlations
(Year/DayofYear & all pairs
of High, Low, Open, Close
and Volume) clearly
explained in terms of the
data domain, in the correct
directions and for correct
reasons demonstrating an
understanding of the data.
7-8
1. Platform description
complete.
2a partial or unclear
2b partial or unclear
2b partial explanation
5-6
1. Platform description
complete.
2a attempt but with
correlation direction wrong
2b Partial description of 4
variables or unclear
2b Partial explanation
0-4
1. Platform description
incomplete.
2a. correlation reason
missing or unrelated to
DayofYear and WeekofYear
2b. Description unrelated to
High, Low, Open, Close and
Volume
2b. Explanation unrelated to
data source
3. Association
mining
10 9-10
Answers demonstrate deep
understanding of association
mining, by the careful
selection of interesting and
differentiated rules and clear
rationale for interestingness.
Comment shows original and
insightful analysis of
association mining on the
problem.
7-8
a 5-number summary
complete
b 5-number explanation for
Volume clear
b Binning understood
c Support and confidence
clear
c 3 interesting rules given
c objective interestingness is
given for all 3
c subjective interestingness
attempted
d Comment makes sense
5-6
a 5number summary ok
b 5-number explanation for
Volume poor
b Binning misunderstood
c Support or confidence not
clear
c < 3 interesting rules given
c objective interestingness is
incomplete
c subjective interestingness
is incomplete
d Comment is cursory or offtrack
0-4
Required information not
provided and/or incorrect or
misleading, demonstrating
lack of engagement with the
problem7
Review
Criteria
Max
Mark
Exemplary Excellent Good Acceptable Unsatisfactory
4. Simple
classification
10 9-10
Deep understanding of the 4
models demonstrated
through analysis of
performance on the change
task.
7-8
a correctly explains why
definition of change makes it
seem easy
b 4 confusion matrixes given
b confusion matrixes
explained in terms of the
data and the method and
the model learnt.
c some evidence of
understanding what the
models are doing
c reasoning for comparative
performance demonstrating
understanding of the
methods behind them
5-6
a partially explains why
definition of change makes it
seem easy
b 4 confusion matrixes given
b confusion matrixes
explained at face value only
c weak understanding of
learnt models
c comparative performance
only cursorily presented
c reason for comparative
performance is shallow
0-4
a inadequate explanation
b confusion matrix missing
b confusion matrix
misunderstood
c Interpretation of confusion
matrix missing
c no apparent understanding
of what the models are doing
c missing or unexplained
comparative analysis
5. Prediction 20 17-20
Approach to problem
demonstrates effort to
produce good results and a
deep understanding of the
relative benefits of the 2
models in the context of the
problem domain.
Results are interpreted in
the context of the problem
domain.
14-16
a justification for choice
shows understanding of the
comparative benefits of
each and extensive
experiments with
performance.
b parameter variations
shows a combination of
experimentation and
understanding of the
parameters with justification
for stopping at selected
parameters.
c several subjective and
objective evaluation
measures used as
appropriate to method
chosen
c justification for stopping
demonstrates awareness of
appropriateness of best
12-13
a justification for choice
shows understanding of the
comparative benefits of
each and experiments with
performance.
b parameter variations
shows a combination of
experimentation and
understanding of the
parameters
c multiple subjective and
objective evaluation
measures used as
appropriate to method
chosen
c justification for stopping
demonstrates awareness of
appropriateness of best
result
10-11
a justification for choice
shows some understanding
of the comparative benefits
of each or experiments with
performance.
b parameter variation
demonstrates some
experimentation
c cursory evaluation given
c justification for stopping
perfunctory
0-9
a weak justification for
choice
b parameter variation
insufficient
c evaluation fails to
demonstrate effort or
understanding of evaluation
c justification for stopping
effectively absent8
Review
Criteria
Max
Mark
Exemplary Excellent Good Acceptable Unsatisfactory
result and scope of potential
for further improvement
6. Complex
Classification
30 26-30
Exemplary use of
classification models with
comprehensive and fit-forpurpose
performance
analysis on the problem
22-25
a explanation correct
b,c,d parameter variation
clear and extensive
demonstrating
understanding of effect in all
3 methods
b.c.d error matrix and ROC
correctly interpreted
in all 3 methods
b,c,d extensive use of
specific evaluation methods
used and significance clearly
explained in all 3 methods
18-21
a explanation correct
a satisfactory approach to
dataset partitioning
b parameter variation clear
and sufficient for good
results
b error matrix correctly
interpreted
b ROC correctly interpreted
b some specific evaluation
methods used
c parameter variation clear
and sufficient for good
results
c error matrix correctly
interpreted
c ROC correctly interpreted
c some specific evaluation
methods used
d parameter variation clear
and sufficient for good
results
d error matrix correctly
interpreted
d ROC correctly interpreted
d some specific evaluation
methods used
15-17
a explanation correct
a satisfactory approach to
dataset partitioning
b parameter variation
perfunctory
b error matrix given
b ROC given
b few specific evaluation
methods used
c parameter variation
perfunctory
c error matrix given
c ROC given
c few specific evaluation
methods used
d parameter variation
perfunctory
d error matrix given
d ROC given
d few specific evaluation
methods used
0-14
a explanation incorrect
a unsound use of
training/testing/validation
data
b no parameter variation
b no error matrix
b no or faulty ROC
b specific evaluation
methods missing
c no parameter variation
c no error matrix
c no or faulty ROC
c specific evaluation
methods missing
d no parameter variation
d no error matrix
d no or faulty ROC
d specific evaluation
methods missing9
Review
Criteria
Max
Mark
Exemplary Excellent Good Acceptable Unsatisfactory
7. Clustering 10 9-10
The application of k-means
algorithm to the dataset and
its evaluation demonstrates
exemplary understanding of
the algorithm, its evaluation,
and its limitations.
Suitable evaluation methods
or clustering experiments
beyond those required here
may be used.
7-8
a justification convincing
b measure calculated
correctly. Discussion
recognises value and
limitations
c discussion on centres
reflects numeric results and
emphasises the interesting
parts that relate to the
significance in domain terms
d correct image included and
description shows
understanding linked to data
domain
5-6
a justification offered but
not clear or unconvincing
b measure calculated
correctly
c discussion on centres
reflects numeric results
d correct image included
0-4
Clustering experimentation
and discussion inadequate
8. Qualitative
Summary
10 9-10
Many aspects of evaluation
are discussed and a clear
conclusion is drawn, with
direct reference to potential
goals of the domain of the
data.
Proposal for further
investigation demonstrates
creativity and thoughtful
engagement with the
problem, clearly building on
the work reported.
8
A clear conclusion is drawn
from the work reported and
a defended proposal for
further investigation is
proposed, with clear links to
both the work reported and
the domain of application.
7
A rounded, balanced
summary of the work is
presented with a justified
proposal given.
6
A summary of the work is
presented and a proposal
made.
0-4
Answer does not
demonstrate adequate
engagement with the
problem nor a qualitative
understanding of the work
reported.