Data Analysis using R and Weka
Overview
The coursework is organized into three parts, each one focusing on a different and important aspect of either Data Pre-processing, Data Analysis or Data Mining. All parts involve the use of the same dataset. The first part focuses on describing and visualizing the data and preparing the data for subsequent treatment ('pre-processing'). The second part focuses on clustering and the third part focuses on classification. The main goal is to give you first-hand experience on working with a relatively large and real data set, from the earliest states of data description to the later stages of knowledge extraction and prediction.
Data Set
The data set is a slightly modified version of a reference data set available from the UCI Machine Learning Repository (https://archive.ics.uci.edu/ml). The data concerns the modeling of wine quality based on physicochemical tests (https://archive.ics.uci.edu/ml/datasets/Wine+Quality). Each record consists of 11 attribute (input) columns, and one class (output) column corresponding to the quality of wine, rated on a ten-point scale. The attributes are various physical / chemical properties of the wine. The entire data set consists of 1600 instances. Figure 1 provides a small snapshot of the data: the header row and 15 instances of the data set. Some of the variables contain missing values, which are indicated by empty entries.
Part 1 - Description, Visualisation and Pre-processing [R Only]
a) Explore the data
i. Use as many functions/techniques in R as necessary to adequately describe and visualize the data. Provide a table for all the attributes of the dataset including the measures of centrality (mean, median etc.), dispersion and how many missing values each attribute has. Use the table to make comments about the data.
ii. Produce histograms for each attribute. Provide details how you created the histograms and comment on the distribution of data. Use also the descriptive statistics you produced above to help you characterise the shape of the distribution.
b) Explore the relationships between the attributes, and between the class and the attributes
i. Calculate the correlations between er and pgr, b1 and b2, and p1 and p2 (three correlations). What do these tell you about the relationships between these variables?
ii. Produce scatterplots between the class variable and er, pgr and h1 variables (note: you may have to recode the class variable as numeric to produce scatterplots). What do these tell you about the relationships between these three variables and the class?
c) General Conclusions
Take into considerations all the descriptive statistics, the visualisations, the correlations you produced together with the missing values and comment on the importance of the attributes. Which of the attributes seem to hold significant information and which you can regard as insignificant? Provide an explanation for your choice.
d) Dealing with missing values in R
i. Write an script in R to find missing values and replace them using three strategies. Replace missing values with 0, mean and median.
ii. Compare and contrast these approaches
f) Attribute transformation
Explore the use of three transformation techniques (mean centering, normalisation and standardisation) to scale the attributes, and compare their various effects.
g) Attribute / instance selection
i. Starting again from the raw data, consider attribute and instance deletion strategies to deal with missing values. Choose a number of missing values per instance or per attribute and delete instances/attributes accordingly. Explain your choice.
ii. Consider using correlations between attributes to reduce the number of attributes. Try to reduce the dataset to contain only uncorrelated attributes.
iii. Use principal component analysis in R to create a data set with ten attributes.
As a result, you will end up with several different sets of data to be used in Part 2 & 3. Give each set of data a clear and distinct name, so that you can easily refer to again in the later stages.
Part 2 - Clustering [R Only]
Using R (only), explore the use of clustering to find natural groupings in the data, without using the class variable - i.e. use only the 20 numeric (input) attributes to perform the clustering. Once the data is clustered, you may use the class variable to evaluate or interpret the results (how do the new clusters compare to the original classes?).
a) Use hierarchical, k-means, PAM as clustering algorithms to create classifications of seven clusters and write the results. Which algorithm produces better results when compared to the class attribute?
b) As each of these algorithms has adjustable parameters, you may explore the 'optimisation' or 'tuning' of these parameters, either manually or (preferably) automatically. Which parameters produce the best results for each clustering algorithm? Provide the reasoning of the techniques you used to find the optimal parameters.
c) Choose one clustering algorithm of the above and perform this clustering on alternative data sets that you have produced as a result of Part 2.
i. The reduced data set featuring only the first 10 Principal Components.
ii. The dataset after deletion of instances and attributes.
iii. The three datasets after you replaced missing values with the three techniques.
iv. Which of these datasets had a positive impact on the quality of the clustering? Provide explanations using the results for each clustering of the alternative data set.
Part 3 - Classification [Weka and R]
You must use Weka to perform the classification, but you may choose to use R to present results. Use Weka to explore the use of various classification techniques to create models that predict the given class from the input attributes. Split the data (randomly) into a training set (2/3 of the data) and a test set (containing 1/3 of the data);
a) Try using the following classification algorithms: ZeroR, OneR, NaïveBayes, IBk (k-NN) and J48 (C4.5) algorithms. Which algorithm produces the best results?
b) Choose one classification algorithm of the above and explore various parameter settings for each of the different splits of data. Which parameters improve the predictive ability of the algorithm?
c) Choose one classification algorithm of the above and use the data sets you created in part 2 [5]:
i. The reduced data set featuring only the first 10 Principal Components.
ii. The dataset after deletion of instances and attributes.
iii. The three datasets after you replaced missing values with the three techniques.
iv. Which of the datasets had a good impact on the predictive ability of the algorithm? Provide explanations using the results for each clustering of the alternative data set.