Intro to classifiers

Training a classifier

To train a classifier in Amplified you need three things:

  1. A tag
  2. Positive training examples: patents that are relevant to the tag
  3. Negative training examples: patents that are not relevant to the tag

If you don't have access to classifiers yet and would like to explore adding them to your plan, email us at support@amplified.ai.

The secret to training top notch classifiers? Classifiers learn what you teach them, so give them good data.

The results you get from the classifier are closely tied to the quality of the positive and negative examples that you provide. For example, imagine teaching a classifier to identify apples. You provide it apples as positive examples and bananas as negative examples. Then you show it an orange. The classifier has never seen an orange, so it has to guess. Does this look more like the positives or the negatives that I've seen?

The lesson to learn here is that the classifier needs to see relevant negative examples.

Let's get training!

Once you have a tag with both confirmed and rejected patents, all you need to do is click "Train" under Classifier. You will see a spinning icon while it trains. After training completes the Tag will now say "View" instead of "Train".

Using a classifier to predict tags

OK, now that we have a trained classifier let's go use it! Open a project with results that we want to use the classifier on.

You can select individual patents to predict tags or use the Predict All shortcut to run the classifier on your entire result set.

Predicting tags on selected patents only
Predicting tags on your entire result set

Reviewing predicted tags

After running your classifier, you will see that in addition to your regular tags you now have predicted tags on your results. Unlike regular tags your predicted tags will have a score on them like 0.11 or 0.96. This tells you how confident the classifier is in it's prediction.

What do the green and red colors mean?

Green predicted tags are likely positive and red are likely negative. This is based on a cut-off score which can be customized.


Sort by prediction

You can also use the predicted scores to re-sort your result list! So if you want to see the patents that are most likely related to Qubit generation, we could run that classifier on the full result set and then sort by Qubit generation Predicted tag scores.

Re-training with new data

As you review predictions, you can convert the predicted tags to regular tags by clicking on the predicted tag's check or X button. This will automatically add those patents to your tag which you can then use to re-train and improve the classifier.

To re-train a classifier, navigate back to the Manage tag page and find the tag in the table. Click on View under Classifier. Here you'll see detailed information about your classifier including training history. You can click on the Train button to re-train using the latest data.

Improving classifiers quickly

The best way to improve a classifier is to identify where it is confidently wrong. For example, a high confidence score on a patent that is actually not relevant or a low confidence score on a patent that is, in fact, relevant.

You can use the sort by Predicted tag score to bring high and low confidence patents to the top. Reviewing these and confirming or rejecting the predictions then re-training can give you much bigger classifier performance improvements from much less work.

Did this answer your question? Thanks for the feedback There was a problem submitting your feedback. Please try again later.

Still need help? Contact Us Contact Us