Enhance Predictive Models with Text Clustering in Einstein Discovery

Supervised machine learning models often rely on tabular datasets consisting of numerical, categorical, and temporal variables. However, additional value can be unlocked by incorporating insights from unstructured text data. Examples of unstructured text include:

• Salesforce records: Notes from opportunities, service cases, activities, etc.
• Customer feedback: Surveys, product reviews, or complaints.
• Meeting notes and product descriptions.
• Digital communication: Emails, social media posts, live chat logs, and chatbot data.

It is estimated that 80% of business data is unstructured. While this figure is just an approximation, it highlights the wealth of untapped information available in text data. Harnessing this data with machine learning opens doors to enhanced business insights and improved outcomes.

Introducing Text Clustering in Einstein Discovery

Previously, unstructured text data required external preprocessing before it could be used in Einstein Discovery predictive models. With the Summer ’22 Release, Salesforce introduced a Text Clustering feature, enabling Einstein Discovery to seamlessly analyze and incorporate unstructured text.

This functionality not only improves model accuracy but also provides new insights to drive actionable business decisions and streamline workflows.

How Text Clustering Works

Algorithms Behind Text Clustering

Two core algorithms power text clustering in Einstein Discovery:

1. TF-IDF (Term Frequency-Inverse Document Frequency):

   • Term Frequency (TF): Measures how often a word appears in a document relative to the total number of words in that document.
   • Inverse Document Frequency (IDF): Identifies how unique a word is across the entire dataset by penalizing terms that appear too frequently in all documents.
   • TF-IDF assigns scores between 0 and 1 to each word, prioritizing terms that are frequent within specific documents but rare across the dataset.

2. K-Means Clustering:

   • Once TF-IDF scores are calculated, K-Means groups the top 75 most significant terms into 10 clusters (K=10).
   • From these clusters, the top 3 terms per cluster are selected for use in the predictive model, helping identify the text most relevant to the predicted outcomes.

Salesforce handles the technical complexities of these algorithms, allowing users to focus on insights without delving into the math.

Using Text Clustering in Einstein Discovery: A Step-by-Step Guide

Step 1: Create a Dataset

Begin with a dataset that includes a column of unstructured text. This text can come from Salesforce data, external sources, or combined datasets.

For example, imagine you are a wine distributor analyzing a dataset of wine reviews to predict the average price of wine sold.

Step 2: Create an Einstein Discovery Story

1. Navigate to Einstein Discovery and create a new story.
2. Choose Insight & Predictions and select Manual Configuration to include specific columns in the model.
3. The unstructured text column (e.g., “Customer Review”) will initially appear unavailable.

Step 3: Apply the Text Clustering Transformation

1. In the Story Settings, click the text column (e.g., “Customer Review”) to reveal the Transform dropdown menu.
2. Select Text Clustering.
3. Proceed to create the story. Einstein Discovery will analyze the text and add the derived clusters to your model.

Step 4: Explore Story Insights

After the story is complete:

• Review the Insights Page, where text clusters and their associated keywords are displayed as variables.
• Visualize the top 10 clusters, each with 3 key terms, to understand how unstructured text impacts your predicted outcome.

Step 5: Leverage Predictions in Your Workflow

Text clusters become part of the predictive model:

• Model Predictions: Text clusters influencing predictions are highlighted in the Prediction Examination section.
• Einstein Discovery Component: Deploy predictions directly into Salesforce using the Einstein Discovery Lightning Web Component. This ensures actionable insights for each record.

Example Walkthrough: Predicting Wine Prices

Scenario: A wine distributor uses Einstein Discovery to analyze customer wine reviews.

• Dataset: Contains structured data (e.g., wine variety, price) and unstructured text (e.g., reviews).
• Goal: Predict the average price of wine sold and use customer reviews to uncover factors driving pricing decisions.

Through Text Clustering:

1. Customer Review text is transformed into clusters, identifying key terms like “oak,” “fruity,” or “acidity.”
2. These insights help the distributor target marketing efforts or adjust pricing strategies based on customer sentiment.

Benefits of Text Clustering in Einstein Discovery

• Improved Model Accuracy: Adds context and nuance to predictions using unstructured data.
• Enhanced Insights: Identifies patterns in text that influence outcomes, offering deeper understanding.
• Actionable Predictions: Enables tailored interventions directly within Salesforce workflows.

By leveraging Text Clustering, organizations can transform unstructured data into meaningful insights, driving smarter decisions and greater business impact.