The task is to develop a machine learning model that can classify tweets about real disasters or not. The tweets in the dataset have been hand-classified, and the goal is to create a model that can accurately classify new, unseen tweets. The challenge is that some tweets may contain words that are commonly associated with disasters, but are not actually about real disasters. For example, a tweet containing the word "ABLAZE" may be metaphorical and not refer to a real fire.
This is a binary classification task. Given the text, keyword, and location of a tweet, the model needs to predict whether the tweet is about a real disaster or not. ML Method To determine the best machine learning method to use for this project, we will be experimenting with the following. Convolutional Neural Networks (CNNs), Random Forest, Logistic Regression, and Naive Bayes.
We will be using the F1 score as the evaluation metric, which is the harmonic mean of precision and recall. The F1 score balances both precision and recall and is a good metric for imbalanced datasets. The higher the F1 score, the better the model's performance. Data The dataset consists of three CSV files: train.csv, test.csv, and sample_submission.csv. The training set contains 7,613 tweets and the test set contains 3,263 tweets. Each tweet is represented by its text, a keyword (which may be blank), and the location from which it was written (which may also be blank). The training set also includes a target column, which indicates whether the tweet is about a real disaster or not.
- id - a unique identifier for each tweet
- text - the text of the tweet
- location - the location the tweet was sent from (may be blank)
- keyword - a particular keyword from the tweet (may be blank)
- target - in train.csv only, this denotes whether a tweet is about a real disaster (1) or not (0)
The goal of the project is to build a machine learning model that can accurately predict whether a tweet is about a real disaster or not. The outcome of the project will be a model that can be used to classify new, unseen tweets, helping disaster relief organizations and news agencies to quickly identify tweets that are relevant to disasters.
If time persists, some ideas of stretch features are as follow.
- Model tuning: The performance can be improved by tuning the hyperparameters. For example trying different hyperparameters to see which combination results in the best performance.
- Preprocessing techniques: Implementing different preprocessing techniques such as removing stop words, stemming, and handling misspelled words.
- Multilingual classification: The kaggle competition is focused on English language tweets only. However, expanding the task to include multiple languages could make it more universal. This would require building a model that can handle multiple languages, which can require additional preprocessing steps.