notes.txt

1. No offer in training data.

2. Preprocessing -
    - Remove split, PII and original fields
    - Select some topics in which help will be offered
    - Selected columns are - food, shelter and water

3. Column Description

    - id: Unique ID for each individual row
    - split: Test, tune, validation split
    - message: English text of actual messages related to disaster
    - original: Text of column 3 in native language as originally written
    - genre: Type of message, including direct messages, social posting, and news stories or bulletins
    - related: Is the message disaster related? 1= yes, 2=no
    - PII: Does the message contain PII? 1= yes, 2=no
    - request: Does the message contain a request? 1= yes, 2=no
    - offer: Does the message contain an offer? 1= yes, 2=no
    - aid_related: Is the message aid related? 1= yes, 2=no
    - medical_help: Does the message concern medical help? 1= yes, 2=no
    - medical_products: Does the message concern medical products? 1= yes, 2=no
    - search_and_rescue: Does the message concern search and rescue? 1= yes, 2=no
    - security: Does the message concern security? 1= yes, 2=no
    - military: Does the message concern military? 1= yes, 2=no
    - child_alone: Does the message mention a child alone? 1= yes, 2=no
    - water: Does the message concern water? 1= yes, 2=no
    - food: Does the message concern food? 1= yes, 2=no
    - shelter: Does the message concern shelter? 1= yes, 2=no
    - clothing: Does the message concern clothing? 1= yes, 2=no
    - Build various feature matrices - count matrix and tfidf matrix on message, message_stem, message_lemma. Play around with values of max feature size.

4. Train LogisticRegression models on all feature matrices for all chosen columns

5. Initial Training Conclusion: TFIDF feature matrices are performing better than Count matrices

6. Use SVM Classifier instead of LogisticRegression since the accuracy was poor