LT2212 V20 Assignment 2
Part 1
I split the news samples by spaces and lowercase all the words. Then I use isalpha() function to filter the words containing numbers or punctuations. In order to make the dataset smaller, I filter the words occur less than 5 times in the whole corpus. I get 18846 news samples and 29543 words ultimately.
Part 2
I use TruncatedSVD class doing dimensionality reduction.
Part 3
model_id #1: GaussianNB
model_id #2 : DecisionTreeClassifier
Part 4
100%(29543), 50%(14772), 25%(7386), 10%(2954), 5%(1477)
| Algorithms | D-Reduction | Accuracy | Precision | Recall | F-measure |
|---|---|---|---|---|---|
| GaussianNB | 1.0, 0.5, 0.25, 0.10, 0.05 | 0.6334, 0.1095, 0.1308, 0.1257, 0.1263 | 0.6525, 0.1706, 0.2024, 0.2460, 0.2755 | 0.6334, 0.1095, 0.1308, 0.1257, 0.1263 | 0.6367, 0.0853, 0.1065, 0.1051, 0.1126 |
| DecisionTreeClassifier | 1.0, 0.5, 0.25, 0.10, 0.05 | 0.4966, 0.1271, 0.1456, 0.1568, 0.1642 | 0.5022, 0.1290, 0.1490, 0.1583, 0.1675 | 0.4965, 0.1271, 0.1456, 0.1568, 0.1642 | 0.4967, 0.1274, 0.1468, 0.1572, 0.1653 |
My most time-consuming bug is that in part 1, when I filter the words which occur less than 5 times in my corpus, I forget to resize the index(value) of each unique word(key) in my dictionary, so I get a wrong values dictionary. Then in Part 3, I found both classifier training algorithms GaussianNB and DecisionTreeClassifier have a really bad performance, all the scores are less than ideal.
In general, GaussianNB has a better performance compared to DecisionTreeClassifier. After 50% features reducting, all the scores in both classifiers are dramatically decreased. However, all the reductions in different reducing levels share a similar scores.