Now that you're familiar with the key phases of your ML project, it's time to walk you through some of my favorite advanced features of BigQuery ML. Recall that you can create a model with just CREATE MODEL. If you do want to override an existing model, you can do CREATE OR REPLACE MODEL.

Models, again, take options which you can specify if you want. The most important and the only required one is that model type. Linear regression for forecasting, logistic regression for classification, with many more types coming soon. I'll add a link in the resources, where you can view all of the available model options. You can set things like the model learning rate, for how fast it should learn and even hyperparameters for things like regularization, to prevent overfitting.

You can inspect the importance the model placed on each feature by looking at the weights it learned. You do this by using the ML.WEIGHTS in filtering on a given input column. In the output, each feature column will have a weight from -1 to 1. The closer the number is to -1 or 1 means the more useful that field is in the model size to predicting the value for that label. It's by far one of my favorite things to do. To evaluate the models performance, next you could just run simply ML.EVALUATE against a trained model. You'll get different performance metrics as we covered depending upon the type of model that you choose. Also you can look at the model's performance in the UI just by clicking on the model object and your dataset and taking a look at all that available metadata. I do that pretty often because I'm pretty lazy in BigQuery. Making predictions is as simple as calling ML.PREDICT on a trained model and passing through the dataset that you want to predict on.

Now, let's do one final big review and what I'll call the "cheatsheet." If you're going to print something out, it's going to be this page.
Воспроизведите видео, начиная с :1:58, и следуйте текстовым инструкциям1:58
First, in BigQuery ML. You need to have a field in your training dataset titled label. Or, you need to specify which field or fields are your labels by using the input_ label_cols in your model options. Second, your model features are simply the data columns that are part of your SELECT statement after your CREATE MODEL statement. After a model is trained, you can use ML.FEATURE_INFO to get statistics and metrics about that column for additional analysis.

Next is the model object itself. You train many different models, which will all be based on objects stored inside your BigQuery dataset, much like your data tables and views. Try clicking on a model object and you can view information about when it was last updated or how many training runs it completed.

Creating a new model is as easy writing CREATE MODEL, choosing that model type, and passing in a training dataset. And again, if you're predicting in a numeric field, like sales for next year, consider looking at linear regression for forecasting. If it's a discrete class like high, medium, or low, spam or not spam, consider using logistic regression for classification.

While the model's running, and even after it's complete, you could view training progress with ML.TRAINING_INFO.

As we mentioned previously, you can see what the model learned about the importance of each feature as it relates to the label that you're predicting. That's my favorite part. Those are what's called you model's weight.

You can see how well the model did against this evaluation dataset by using ML.EVALUATE.

Lastly, it's as simple as writing ML.PREDICT in referencing your trained model, and then your prediction dataset to return back predictions. An important note here is that when using ML.PREDICT in passing in a new dataset with an unknown label that you can now add other columns that you didn't want to train on initially. The model is not being retrained naturally during prediction. Note that if you do happen to remove or rename columns from your prediction dataset, that the model is expecting and it saw in training, you'll be given an error.

So the four major steps look like this. Write a SQL query to extract training data from BigQuery. Create a model specifying that type, evaluate the model and verify that it meets the requirements. And then predict on that model using data that's extracted from BigQuery. If you are explaining BigQuery ML as a whole to others, I often just list these main points. So as a broad recap, BigQuery ML allows you to, write machine learning models with SQL, experiment and iterate right where your data lives, inside of BigQuery.

Currently, build classification, both binary and multi-class and forecasting models, with more model types coming soon.

And if you know machine learning already, you can really get into the details of your model options and weights very easily. Common adjustments that you can make above and beyond what BigQuery ML defaults to everyone include things like the learning rate, regularization, the training evaluation dataset split, predefined weights for classes, and much, much more. Check out the BigQuery ML documentation link and bookmark it as a reference, especially as new features continue to be being added.