Skip to content
AvraAvra
Esc
↑↓navigate↵open⌘Jpreview
On this page

Applications

How teams use Avra's embeddings in credit, growth, fraud, and analytics models.

Avra’s embeddings are general-purpose features from Avra GFM. These are the most common ways teams put them to work.

Each of these uses treats the vectors in one of four ways: as features in a model, by similarity to known companies, by distance from a segment, or projected to two dimensions for a picture.

Use case How it uses the embeddings How to measure it
Credit risk Features in your credit models ROC-AUC, Gini, or KS on an out-of-time holdout, against a model without the embeddings
Lookalikes and segmentation Cosine similarity to known good customers The lookalikes’ conversion, compared with your current targeting
Lifetime value Features in your lifetime value model The error on realized value for a recent cohort
Fraud and anomalies Distance from a segment’s typical position Precision at the review capacity your team has
Market analysis A 2D projection with UMAP or t-SNE No metric: use it to frame questions

Credit risk

Use embeddings as features in your own credit models. They carry signals about a company’s network, behavior, and trajectory that are hard to derive from your own records. Compare ROC-AUC, Gini, or KS on an out-of-time holdout, with and without the embeddings, before relying on them.

Lookalikes and segmentation

Find prospects that resemble your best customers. Compute the cosine similarity between the embeddings of known good customers and your prospects. The closest prospects are your lookalikes. Then compare their conversion with your current targeting.

Lifetime value

Forecast what a customer will be worth. Add embeddings to your lifetime value model as features. Measure the error on realized value for a recent cohort.

Fraud and anomalies

Spot companies that don’t look like their peers. Distance from a segment’s typical position in the embedding space can serve as an anomaly feature. Measure precision at the review capacity your team has.

Market analysis

See how companies, customers, and competitors cluster. Project the embeddings to two dimensions with UMAP or t-SNE for visualization, then read the clusters: where your customers concentrate, and where there’s space no one serves. There’s no metric for this one. Use it to frame questions for the models on this page.

Was this page helpful?