Explain personalized product recommendations
Use Kumo explanation artifacts to answer a merchandiser's questions, and check each answer against the patterns planted in the synthetic data.
What does the model respond to across customers?
Unsigned sensitivity of the model output to each input column, aggregated over retained test examples. Tags show what each column is known to carry in the synthetic data.
| Column | Known role | Score |
|---|
Is a sensitive column signal or known noise?
Three checks for each known-noise column, with region as the planted positive control. Group differences compare the product mix between the column's groups with randomly relabeled groups (95th percentile in parentheses). The change test sets new values for every fitted customer at the review cutoff and rescores with the frozen model.
| Column | Known role | Gradient rank | Purchases differ by group | Recommendations differ by group | Top 5 kept after change | Customers with a changed top 5 |
|---|
Highest-scored context
Center: customer. Middle ring: the highest-scored purchases, newest at the top; lower-scored purchases are grouped and not drawn. Outer ring: purchased products. Darker fill means a higher local score in this example. Select a node to see its record.
Select a node.
Highest-scored purchases
| # | Product | Collection | Role | Days before | Row score | Product score |
|---|
Customer columns
Compare Product A with Product B
| Product A | Product B |
|---|
Do the subgraph scores point to the planted starter?
For retained examples with a starter bought in the recent window, compare the top-scored purchase with the starter. Every prior purchase is in the model context; the subgraph shows the highest-scored ones and groups the rest. Taking the most recent purchase is the simple rule to beat.
Remove the evidence and rescore
For every fitted customer with a recent starter at the review cutoff, remove that starter purchase, or instead an everyday purchase of similar age, and rescore with the unchanged model.
| History | Customers | Accessory in top 5 | Median accessory rank | Accessory score lower | Mean accessory score change | Repeat favorite in top 5 | Median favorite rank |
|---|
Is the model only repeating past purchases?
Where do recommendations beat simple rules?
| Method | Pairs with purchases | Recall@5 | NDCG@5 |
|---|
Paired difference in recall@5 (Kumo minus comparison)
| Comparison | Difference | 95% interval | Result |
|---|