Machine Learning View

What the Models Suggest (Plain Language)

This page translates model output into practical answers: expected traffic, unusual surges, who returns, where people go next, recurring error themes, and countries with unusual traffic patterns.

How To Read This Page

Forecasts: directional estimates, not exact counts.
Behavior charts: percentages are easier to trust than raw scores.
Anomalies: these are review candidates, not confirmed issues.
Recommendations: treat as hints for UI prompts and cross-links.
Sessions Analyzed16,407216,036 tracked events
Return Model Score63/10050 = random guess, 100 = perfect
Unusual Traffic Surges13FTU Explorer has the biggest jump
Error Themes450,130 error rows clustered
Countries To Review135 look likely automated

Expected Visits Over The Next 6 Months

Each line is a tool. Hover to see likely low/high range.

Forecast
Biggest projected month in this run: KG Explorer at 8,075 visits in Nov 2026 (likely range 4,579 - 11,793).

Big Traffic Jumps We Should Explain

Bars show extra visits compared with the previous month.

Surges

Largest jump in this run: FTU Explorer in Mar 2024.

Who Comes Back After A Session?

Blue bars = number of sessions. Orange line = real return rate.

Return Behavior
Strongest return signal: early page views · strongest drop-off signal: long session duration.

Cohort Retention — Do Users Come Back?

Each line is a monthly cohort. Drops show how retention decays over time.

Cohort Analysis

Hollow circles = May '26 data (partial month, undercounts real retention).

Reading the chart: Row = cohort month, Column = months after first visit. Darker blue = more users returned. Tracked via persistent cookie (anon_id).
Month 0 = 100% by definition — these are the users that define the cohort. Watch for how quickly each row fades.
Cookie tracking started Oct 2025, so only recent cohorts appear. Older visits lack persistent IDs for user-level retention measurement.

Session Types (Simple Grouping)

The model groups sessions by depth, duration, and interaction style.

Audience Types

Largest group: Regular Researchers (92.8%).

Where People Go Next Between Tools

Rows are current tool, columns are next tool in the same session.

User Journeys

Most common next-step path: KG Explorer → CDE (42.7%, 41 transitions).

Behaviors That Happen Together

How often the right-side action happens after the left-side action.

Behavior Pairings

Strongest pairing right now: click + EUI -> organ selection (confidence 51.3%).

Main Error Themes

~55% are uncontrollable network failures · ~25% are fixable KG Explorer icon bugs · ~3% is dev noise

Errors

Largest cluster contributes 48.1% of error rows.

Countries With Unusual Traffic Patterns

Bar = bot-like traffic share. Use as a review list, not a final verdict.

Geo Review

Highest-priority review country in this run: Finland.

Confidence Note

Return model score is 63/100, which is useful for ranking risk but not perfect for exact individual prediction.

Bot Signal Driver

Most influential bot feature is user-agent string length in this run.

Journey Coverage

150 sessions contained multi-step tool journeys we could model.

Data Coverage

Last ML run processed 126 monthly tool points and 16,407 session-level transactions.