As I understand it, this depends on the model you're using. Some models, like linear classifiers and decision trees can train on a weighted data set. They will simply let the first 3 examples weigh much more heavily in the training than the rest. If your classifier can't do this, you can resample a new dataset with replacement: each instance is sampled according to its weight, and replaced. You keep doing this until your dataset is the same size as the original. Then you retrain on this new dataset. In your example the new dataset would have 10 instances, probably roughly equal amounts example 1, 2 and 3, with a small probability of one of the other examples occurring.

Hi Sarah. There's certainly nothing wrong with that choice. XGBoost is known for working pretty well out-of-the-box. I would recommend comparing with a few simpler algorithms like Naive Bayes and Logistic regression, to see if the complexity of XGBoost is warranted. To be honest, the choice of classifier will probably matter a lot less than _which features_ you choose to extract to convert your text to feature vectors. If you want to try something more involved, you could also go for a model that is specifically designed to consume text (and doesn't require feature extraction). See lecture 11 for some examples.

thank you so much @@riskone1 👍