Leverage Generative AI for Your Initial Training

Welcome to the demo of leverage generative AI for your initial training in Spend Classification within Oracle's Procurement Cloud.

Generate initial training sets to predict categories based on the taxonomy you have selected. Gen AI uses text data from your business transactions and category descriptions to predict spend categories. Review the top three predictions for each category and use these predictions to finalize your training data for future classifications.

Supervised learning models, which Oracle Fusion Spend Classification uses to classify the business transactions, needs to be trained with good quality training data spread across the different categories. For customers already doing some sort of classification are the ones that plan to use same taxonomy to classify using spent classification, providing training data is easier as they can start working on any analysis they have done in the past and upload it.

But if you need to train the system with new taxonomy or you're just beginning your classification and analysis journey, preparing the training data from scratch can be quite time-consuming. You can now use Gen AI to get a prediction for the sample training set and review the top three predictions to finalize the training. Gen AI uses description-based columns on a business transaction like an invoice, and then reads the descriptions assigned to each of your category and works out the initial prediction for each of the invoice distribution. Description-based columns, like invoice description, item description, line descriptions, are used for this prediction.

This demo shows how you can use Gen AI to reduce the time and effort otherwise spent on building a training set from scratch. Click Spend Classification within the Procurement offering. This is the landing page of Spend Classification, which shows you a recently concluded batch and a few visual insights. Click the configuration on navigation bar at the bottom of the page.

This is the Data Set tab in the configuration section. Here, you can see all the data sets seeded or uploaded by users. Data sets can be used to train the machine learning model or can be classified with the knowledge base built using the training. Click on the AI Assist action. This opens a new drawer where you can select the data set, taxonomy, and narrow down the transactions using filters available.

Please note, by default, the sampling percentage is set to 1. This means that the system will look at all the transactions, satisfying the filter criteria and then pick a unique representative. That would be approximately 1% of the available transactions. We use clustering and sampling algorithms to extract this representative sample. You can check how many transactions will be there in the training set by clicking the get transactions count button.

You may use advanced sampling options to improve your data selection. Most of the users should proceed with the default settings. You can choose to generate a training set or a knowledge base from this draw. If you choose a training set, we will use Gen AI to come up with the top three predictions for your transactions. You can review them and pick the best predictions to finalize your training set.

If you choose to generate the knowledge base directly, the system will pick the top prediction for you and use it to create the knowledge base. Once the knowledge base is ready, you can start classifying your business transactions. For this demo, let's create the training set. Click the generate training set button at the bottom.

A background process will be initiated that will take the invoices and categories and start calling Gen AI to generate predictions. We can check the status of this process in the activity log. Click on the Activity Log in the message banner. Once the process completes, we can download the sample training set.

Close the Activity Log. Search for the newly created training set. Go to the row level action and click Download. I have a sample file ready. Let's take a look at it. This is a training set file, and you will notice there are six new columns added: Best Prediction, Confidence for the Best Prediction, Second Best Prediction, its confidence, Third Best Prediction, and its confidence. If you are happy with the best prediction, you can proceed to create the knowledge base as is. Otherwise, pick the best value out of these three and populate into the auto code column for the purchasing category or the category that you have chosen.

Once you have done that, you can upload the data set as the final training set and proceed with knowledge base creation. This concludes the demo.