Classify news titles into categories like business, technology, etc using pretained google's word2vec model and applying logistic regression.
Raw data set can be downloaded from the link below.
https://www.kaggle.com/arjunchandrasekhara/news-classification/data
wget -c "https://s3.amazonaws.com/dl4j-distribution/GoogleNews-vectors-negative300.bin.gz"
This model is trained by google and provides a 300 dimensional vector for each word.
50,000 clean records are used as features. Clean records means the records are without stop words, out of vocab words and punctuation marks. These records have clean labels that belong to either of the category of b(business), t(technology), m(medicine), e(entertainment). The records have a 300 dimensional vector for each title and its respective label has single char.
Shape of Features is 50000*300 and shape of labels is 50000*1. The ratio of train:test set is 4:1. i.e. 40,000 records used for training the model and 10,000 records used for validation of the model.
Outcomes the logistic regression model are as follows.
| Train | Test | Dimensions | Accuracy | Time(sec) | Comments |
|---|---|---|---|---|---|
| 40000 | 10000 | 300 | 86 | 12 | Data randomly shuffled |
| 80000 | 20000 | 300 | 78 | 53 | Not shuffled |
| 40000 | 10000 | 300 | 74 | 12 | Not shuffled |
| 80000 | 20000 | 300 | 78 | 53 | Not shuffled |
| 24000 | 6000 | 300 | 73 | 11 | Not shuffled |
| 160000 | 40000 | 300 | 72 | 53 | Not shuffled |
Precision and Recall are as follows
For labels in sequence as ['b', 't', 'e', 'm']
precision: [82.94729775 83.25883787 89.44781729 87.46594005]
recall: [84.79752917 80.92417062 91.4843288 82.16723549]
ROC curves are as follows
