Skip to content

Latest commit

 

History

History
65 lines (46 loc) · 4.6 KB

paperReport.md

File metadata and controls

65 lines (46 loc) · 4.6 KB
Raw

Data Mining Term Project

1.Introduction :

  Wine quality is widely used in machine learning and data mining, therefore, this are our study subject of data mining.
  Wine quality frequently influenced by climate, weather, temperature and the winemaking process that are not the indicators we used in dataset, instead ,there are objectivie attributes, such as, volatile acidity, sulphate, sulfur dioxide.
  After we learnt plenty of methods of data mining to find the rules of classification, we can now practice and compare it by Weka.

2.Preprocess(Related Work) :

  Because we grab the famous data set from UCI website，and we don’t find any missing data in this dataset after checking.

  • The following method is the Data Preprocess method we try in this dataset

    • remove outliers
    • discretization
    • delete attribute

  • remove outliers:
    We use InterquartileRange function in filter/unsupervise/attribute to find the outliers, and remove the outlier by Removewithvalue function in filter/unsupervise/instance.

  • discretization:
    We use the function at filter/unsupervise/discretization to discretize to data.

  • delete attribute:
    We use CfbsubsetEval function with Bestfirst method to find the best related attribute with "quality", and delete it.

  In Random Forest, we try all these method, the result isn’t always turn to positive way. Among these method “remove outliers” is the best method that can improve our accuracy. And other two method are always turn into lower accuracy.
  In Multilayer, remove outliers and delete attribute can improve our accuracy, but discretization can’t.
  In K-nearest, remove outliers and delete attribute can improve our accuracy, but discretization can’t.
  You can see the brief summary result below.

  According the result above ,after discretize will always cause accuracy decrease or even malfunction. In my hypothesis, it is because the discretize we use force the data seperate to N-bins and the attribute will turn to nominal so, it is diffical to seperate different attribute by the same rule.

3.Mine the Data :

  After preprocessing, we tested the following methods to find the rule of data where we could calculate accuracy by classifying the test set.
  There are two major way to find the rule. First, using the classifier functions or Decision trees .Secondly, clustering the data and we evaluation the clusters to their possible classes.
  It is a simple strategy that we used is choosing the highest accuracy in these methods with non-preprocessed dataset. Processing to the next step, adjusting the arguments try to get the higher accuracy, therefore the final goal we want to achieve is to find the best accuracy as possible to classify the quality of wine by these attributes experimentally filtered or chosen by heuristic.

Methods we will try:

  • K-means with Classes to clusters evaluation
  • MultiLayerPerceptron (Neural Network)
  • IBK (K-nearest)
  • RandomForest
  • J48 (Decision Trees)

4.Experiments :

  First, clustering the data by Simple K Means.Using the Euclidean Distance and number of clusters is 6. As we can see, it is not the result we want to get and this method can't applied to classifying that the accuracy calculated by confusion matrix is 26.8293% which is extremely low.

Fig.1 Confusion Matrix of Simple K Means

  Secondly, comparing the different classifier by accuracy, and put the other results of two independent preprocessing method in one table. RandomForest stays the highest accuracy in all results, undoubtedly, it would be test forward by arguments adjusting and multiple preprocessing. Besides, reducing attributes results making accuracy of the methods related to decision trees lower. In the other hand, after InterquartileRange preprocessing, most of them improved especially for Decision Tree group.

Fig.2 Comparison of Different Methods
[1] "4 attr." presents to volatile acidity, total sulfur dioxide, sulfates, alcohol, these 4 attributes selected by "CfsSubsetEval" evaluator with "Best First"

Finally, we can evolve the study of RandomForest now and it would shows the interaction of accuracy with diverse preprocess simultaneously.

5.Conclusion :

Reference :

Dataset contributed by Paulo Cortez (Univ. Minho), Antonio Cerdeira, Fernando Almeida, Telmo Matos and Jose Reis (CVRVV) @ 2009