Contribution Explorer and Top Insights
What Is Contribution Explorer?
SNOWFLAKE.ML.CONTRIBUTION_EXPLORER identifies which dimensions (categorical columns) contribute most to a change in a metric. When a KPI moves unexpectedly — revenue drops, error rates spike, conversion decreases — Contribution Explorer pinpoints the segments driving the shift.
What Is Top Insights?
SNOWFLAKE.ML.TOP_INSIGHTS is a streamlined function that wraps Contribution Explorer into a single call. It compares two groups of data (e.g., this week vs last week) and returns the top contributing segments ranked by impact.
When to Use These Functions
| Scenario | Function |
|---|---|
| Revenue dropped 15% week-over-week — why? | Top Insights or Contribution Explorer |
| Conversion rate changed after a deployment | Contribution Explorer |
| Support ticket volume spiked — which categories? | Top Insights |
| Ad spend ROI declined — which campaigns? | Contribution Explorer |
Data Requirements
The input data must contain:
| Column Type | Description | Example |
|---|---|---|
| Metric column | The numeric value that changed | revenue, error_count, conversion_rate |
| Dimension columns | Categorical attributes to analyze | region, product, channel, device_type |
| Label column | Identifies the two groups to compare | period = 'current' vs period = 'previous' |
Using Top Insights
-- Prepare comparison data with a label column
CREATE OR REPLACE VIEW revenue_comparison AS
SELECT
region,
product_category,
customer_segment,
channel,
revenue,
CASE
WHEN sale_date BETWEEN '2024-07-01' AND '2024-07-31' THEN 'CURRENT'
WHEN sale_date BETWEEN '2024-06-01' AND '2024-06-30' THEN 'PREVIOUS'
END AS period
FROM daily_sales
WHERE period IS NOT NULL;
-- Run Top Insights
CALL SNOWFLAKE.ML.TOP_INSIGHTS(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'REVENUE_COMPARISON'),
METRIC_COLNAME => 'REVENUE',
LABEL_COLNAME => 'PERIOD',
LABEL_VALUE_OF_INTEREST => 'CURRENT'
);
Top Insights Output
The result ranks contributing segments:
| Column | Description |
|---|---|
CONTRIBUTOR | The dimension value(s) driving the change |
METRIC_CURRENT | Metric value in the current period |
METRIC_PREVIOUS | Metric value in the previous period |
RELATIVE_CHANGE | Percentage change for this segment |
OVERALL_RELATIVE_CHANGE | Percentage change across all data |
GROWTH_RATE | How much this segment's growth differs from overall |
SURPRISE | Statistical measure of how unexpected this change is |
-- Review the top contributors
SELECT * FROM TABLE(RESULT_SCAN(LAST_QUERY_ID()))
ORDER BY surprise DESC
LIMIT 10;
Using Contribution Explorer
Contribution Explorer provides more control over the analysis:
-- Create the explorer object
CREATE OR REPLACE SNOWFLAKE.ML.CONTRIBUTION_EXPLORER revenue_explorer(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'REVENUE_COMPARISON'),
METRIC_COLNAME => 'REVENUE',
LABEL_COLNAME => 'PERIOD',
LABEL_VALUE_OF_INTEREST => 'CURRENT'
);
-- Get top contributors
CALL revenue_explorer!GET_CONTRIBUTIONS();
Practical Example: Revenue Drop Analysis
-- Revenue dropped 20% this month. Find out why.
-- Step 1: Prepare the comparison dataset
CREATE OR REPLACE VIEW revenue_investigation AS
SELECT
r.region,
r.country,
p.category AS product_category,
c.segment AS customer_segment,
o.channel,
o.revenue,
CASE
WHEN o.order_date BETWEEN '2024-08-01' AND '2024-08-31' THEN 'CURRENT'
WHEN o.order_date BETWEEN '2024-07-01' AND '2024-07-31' THEN 'PREVIOUS'
END AS period
FROM orders o
JOIN regions r ON o.region_id = r.id
JOIN products p ON o.product_id = p.id
JOIN customers c ON o.customer_id = c.id
WHERE period IS NOT NULL;
-- Step 2: Run the analysis
CALL SNOWFLAKE.ML.TOP_INSIGHTS(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'REVENUE_INVESTIGATION'),
METRIC_COLNAME => 'REVENUE',
LABEL_COLNAME => 'PERIOD',
LABEL_VALUE_OF_INTEREST => 'CURRENT'
);
-- Step 3: Interpret results
SELECT
contributor,
ROUND(relative_change * 100, 1) AS segment_change_pct,
ROUND(overall_relative_change * 100, 1) AS overall_change_pct,
ROUND(surprise, 3) AS surprise_score
FROM TABLE(RESULT_SCAN(LAST_QUERY_ID()))
ORDER BY surprise DESC
LIMIT 15;
Example output interpretation:
| Contributor | Segment Change | Overall Change | Surprise |
|---|---|---|---|
region = 'EMEA' | -38.2% | -20.1% | 0.872 |
product_category = 'Enterprise' | -45.1% | -20.1% | 0.814 |
region = 'EMEA' AND product_category = 'Enterprise' | -62.3% | -20.1% | 0.791 |
This tells us: EMEA Enterprise revenue dropped 62%, far exceeding the overall 20% decline. This combination is the primary driver.
Practical Example: KPI Investigation
-- Error rate spiked after a release — which services are affected?
CREATE OR REPLACE VIEW error_rate_investigation AS
SELECT
service_name,
endpoint,
http_status_code,
deployment_region,
error_count,
CASE
WHEN log_date = CURRENT_DATE() THEN 'CURRENT'
WHEN log_date = DATEADD('day', -1, CURRENT_DATE()) THEN 'PREVIOUS'
END AS period
FROM daily_service_metrics
WHERE period IS NOT NULL;
CALL SNOWFLAKE.ML.TOP_INSIGHTS(
INPUT_DATA => SYSTEM$REFERENCE('VIEW', 'ERROR_RATE_INVESTIGATION'),
METRIC_COLNAME => 'ERROR_COUNT',
LABEL_COLNAME => 'PERIOD',
LABEL_VALUE_OF_INTEREST => 'CURRENT'
);
Best Practices
- Include diverse dimensions — The more categorical columns you provide, the more specific the insights. Include region, product, channel, segment, and other relevant breakdowns.
- Compare like for like — Ensure the two periods are comparable (same number of business days, no holidays skewing one period).
- Focus on the SURPRISE column — High surprise indicates segments that changed disproportionately relative to the overall trend.
- Drill down iteratively — If the top contributor is a broad category (e.g.,
region = 'APAC'), create a new analysis scoped to that region with more granular dimensions. - Automate investigations — Schedule Top Insights as a Snowflake Task to run after each reporting cycle, storing results for review.