Crafte Sessions analysis 2026 09 29 2026 10 29
sessions_analysis-2026-09-29-2026-10-29.csv · 2,246 rows · 15 columns · definition v3 saved 29 Oct 2026 14:08
Segments v24 built 29 Oct 2026 16:27 on feature set v6
A newer feature set (v8) exists. These segments were built on v6.
Segments found
15
fixed
In a segment
99%
2,223 of 2,246 snapshots
Unknown
1.0%
0 history · 0 between · 23 novel
Rules match the model
98%
for named segments
device was segmented separately: each device group has its own segments.
What to do
From these segments' numbers. Setting buttons fill in the Build segments tab (then Build segments); merge buttons change the table below (then Save labels, no refit).- Worth doingS8, S9, S10, S11, S12, S13, S14, S15 are very small (1 to 26 snapshots)
Under 30 snapshots, a segment's rates swing with a couple of people, and here the fit is mostly splitting off a few odd snapshots. Merge each into the segment it's most like (S8 → S4, S9 → S2, S10 → S4, S11 → S2, S12 → S2, S13 → S7, S14 → S3, S15 → S2). Keep one only if it's a group you specifically care about.
- Worth doingordered: use S6, S5, S7 as the targets on Outcomes
S6, S5, S7 hold 31% of ordered in 10% of the snapshots (3.1× the overall rate).
- Fine as isUnknowns: 1.0% of snapshots
A small unknown share means nearly everyone fits a segment, while the odd ones aren't forced in.
Outcome concentration
Share of each segment's snapshots whose entity eventually had the outcome. Lift = rate ÷ overall rate.S6 has the highest rate: 19% for ordered, 4.8× the overall 3.8%.
- S6 has the highest rate: 19% for ordered, 4.8× the overall 3.8%.
- Rate: share of a segment's snapshots whose entity eventually had the outcome. Everyone together: 3.8% for ordered.
- Lift = rate ÷ overall rate. S6 has 19% for ordered, so 4.8× the average; above 1 is better than average.
- Captured: the segment's share of all the outcomes (each Captured column adds to 100%). Compare it with Share: S1 is 55% of snapshots and holds 30% of ordered, so it dilutes it.
- Segments that capture much more than their share are the ones to take to Outcomes as targets.
| Segment | Snapshots | Share | ordered | ||
|---|---|---|---|---|---|
| Rate | Lift | Captured | |||
| S1 | 1,246 | 55.5% | 2.1% | 0.5× | 30% |
| S2 | 267 | 11.9% | 3.0% | 0.8× | 9% |
| S3 | 225 | 10.0% | 2.7% | 0.7× | 7% |
| S4 | 192 | 8.5% | 4.2% | 1.1× | 9% |
| S5 | 142 | 6.3% | 9.9% | 2.6× | 16% |
| S6 | 54 | 2.4% | 18.5% | 4.8× | 12% |
| S7 | 33 | 1.5% | 9.1% | 2.4× | 3% |
| S8 | 26 | 1.2% | 3.8% | 1.0× | 1% |
| S9 | 15 | 0.7% | 13.3% | 3.5× | 2% |
| S10 | 7 | 0.3% | 0.0% | 0.0× | 0% |
| S11 | 6 | 0.3% | 0.0% | 0.0× | 0% |
| S12 | 4 | 0.2% | 0.0% | 0.0× | 0% |
| S13 | 4 | 0.2% | 25.0% | 6.5× | 1% |
| S14 | 1 | 0.0% | 100.0% | 26.1× | 1% |
| S15 | 1 | 0.0% | 0.0% | 0.0× | 0% |
| Unknown: fits no segment | 23 | 1.0% | 26.1% | 6.8× | 7% |
| Everyone | 2,246 | 100% | 3.8% | ||
How each segment differs
Average of each input, in standard deviations from everyone. Red higher, blue lower.source=Facebook separates the segments most: from -4.8 sd in S6 to +0.2 sd in S1.- Each column is a segment and each row an input. A cell is how far the segment's average sits from everyone's, in standard deviations: +1 is clearly higher than typical (red), −1 clearly lower (blue), and near 0 is typical. Below, "slightly" means under 0.6 and "much" means over 1.5.
source=Facebookseparates the segments most: from -4.8 sd in S6 to +0.2 sd in S1.- S1 (1,246 snapshots): slightly more often device is mobile (+0.6); slightly more often landingPage is / (+0.5).
- S2 (267 snapshots): much more often device is tablet (+2.6).
- S3 (225 snapshots): much less often landingPage is / (-2.2); slightly more often device is mobile (+0.6).
- S4 (192 snapshots): much more often device is desktop (+2.7); slightly more often landingPage is / (+0.3).
- S5 (142 snapshots): much longer-standing (returning) (+2.5); much more visits so far (+2.4); slightly more often device is mobile (+0.6).
- S6 (54 snapshots): much more often source is Direct (+4.6); lower newVisitor (-1.2); slightly more often landingPage is other (+0.6).
- Read a segment down its column to describe it; read across a row to see which segments differ on that input.
| S1 | S2 | S3 | S4 | S5 | S6 | S7 | S8 | S9 | S10 | S11 | S12 | S13 | S14 | S15 | Novel | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
actions[add_search] | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | +47.4 | -0.0 | -0.0 |
actions[sub_start] | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | -0.0 | +47.4 | -0.0 |
newVisitor | +0.1 | +0.1 | +0.1 | +0.1 | +0.1 | -1.2 | +0.1 | -0.8 | +0.1 | +0.1 | +0.1 | -11.8 | +0.1 | +0.1 | +0.1 | -2.0 |
landingPage=/product/75-christmas-cake-boxed | -0.1 | -0.1 | +0.4 | -0.1 | -0.0 | -0.1 | +0.5 | -0.1 | -0.1 | -0.1 | +9.4 | -0.1 | -0.1 | -0.1 | -0.1 | +1.6 |
landingPage=/product/474-port-and-prune-pudding-800g-log-in-cloth | -0.1 | -0.1 | +0.7 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | +8.3 | -0.1 | -0.1 | +1.0 |
pages | -0.1 | -0.2 | +0.2 | -0.1 | +0.3 | +1.0 | -0.1 | +0.2 | +0.3 | +0.6 | -0.4 | +1.8 | -0.4 | +6.3 | +4.3 | +2.1 |
landingPage=/product/79-chocolate-fruit-cake-500g-boxed | -0.2 | -0.2 | +1.2 | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 | +5.0 | +5.0 | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 | +0.9 |
source=Facebook | +0.2 | +0.2 | +0.2 | +0.2 | +0.2 | -4.8 | +0.2 | -4.8 | +0.2 | +0.2 | +0.2 | -4.8 | +0.2 | +0.2 | +0.2 | -1.3 |
source=Direct | -0.2 | -0.2 | -0.2 | -0.2 | -0.2 | +4.6 | -0.2 | +3.7 | -0.2 | -0.2 | -0.2 | +1.3 | -0.2 | -0.2 | -0.2 | +0.8 |
rows | -0.3 | -0.1 | -0.3 | -0.3 | +2.4 | +0.2 | +3.0 | -0.3 | +0.3 | +0.0 | +0.1 | +2.5 | +0.8 | -0.3 | -0.3 | +2.3 |
device=desktop | -0.4 | -0.4 | -0.4 | +2.7 | -0.4 | -0.4 | +2.7 | +2.7 | -0.4 | +2.7 | -0.4 | -0.4 | +2.7 | -0.4 | -0.4 | +0.0 |
tenure_days | -0.3 | -0.0 | -0.3 | -0.3 | +2.5 | +0.1 | +2.6 | -0.3 | +0.4 | -0.1 | -0.1 | +2.4 | +1.1 | -0.3 | -0.3 | +2.2 |
device=tablet | -0.4 | +2.6 | -0.4 | -0.4 | -0.4 | -0.4 | -0.4 | -0.4 | +2.6 | -0.4 | +2.6 | +2.6 | -0.4 | -0.4 | +2.6 | -0.0 |
minutes | -0.1 | -0.1 | +0.0 | -0.1 | +0.3 | +1.0 | -0.0 | +0.3 | -0.2 | -0.2 | -0.4 | -0.2 | -0.4 | +2.4 | +2.3 | +1.7 |
landingPage=/ | +0.5 | +0.3 | -2.2 | +0.3 | -0.2 | -0.9 | -0.5 | -1.0 | -2.2 | -2.2 | -2.2 | -0.2 | -2.2 | +0.5 | +0.5 | -1.6 |
actions[add_recommended] | -0.1 | -0.1 | -0.1 | -0.1 | +0.5 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | -0.1 | +2.1 |
landingPage=other | -0.2 | -0.1 | +1.0 | -0.2 | +0.3 | +0.6 | +0.3 | +2.0 | -0.2 | -0.2 | -0.2 | +0.9 | -0.2 | -0.2 | -0.2 | +0.5 |
device=mobile | +0.6 | -1.7 | +0.6 | -1.7 | +0.6 | +0.6 | -1.7 | -1.7 | -1.7 | -1.7 | -1.7 | -1.7 | -1.7 | +0.6 | -1.7 | -0.0 |
Movement between segments
Where each entity's next snapshot lands, as a share of the rowThere are 232 steps from one snapshot to the next. 39% stayed in the same segment, so people often change segment between visits.
- Rows are where an entity's snapshot was; columns are where its next snapshot landed. Each row adds to 100%. The diagonal is staying put; anything else is a move.
- There are 232 steps from one snapshot to the next. 39% stayed in the same segment, so people often change segment between visits.
- Most common moves: S1 → S5 (92); S4 → S7 (12); S3 → S5 (9).
- S6 has the highest ordered rate (19%). Entities arriving there mostly come from S8 (1): that's the path worth watching.
| From ↓ to → | S1 | S2 | S3 | S4 | S5 | S6 | S7 | S8 | S9 | S10 | S11 | S12 | S13 | S14 | S15 | Novel | moves |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| S1 | 1% | 92% | 7% | 100 | |||||||||||||
| S2 | 88% | 8% | 4% | 26 | |||||||||||||
| S3 | 6% | 24% | 53% | 18% | 17 | ||||||||||||
| S4 | 86% | 14% | 14 | ||||||||||||||
| S5 | 94% | 6% | 36 | ||||||||||||||
| S6 | 11% | 78% | 11% | 9 | |||||||||||||
| S7 | 100% | 15 | |||||||||||||||
| S8 | 50% | 50% | 2 | ||||||||||||||
| S9 | 67% | 33% | 3 | ||||||||||||||
| S10 | 100% | 1 | |||||||||||||||
| S12 | 100% | 3 | |||||||||||||||
| S13 | 100% | 1 | |||||||||||||||
| Unknown: fits no segment | 40% | 20% | 40% | 5 |
Segments over time
Share of snapshots per weekFrom the first week (05 Oct 2026) to the last (26 Oct 2026), S3 grew the most (+19% points) and S1 shrank the most (-23% points).
- Each bar is one week: how that week's snapshots split across the segments (100% per bar). Hover a block for the count.
- From the first week (05 Oct 2026) to the last (26 Oct 2026), S3 grew the most (+19% points) and S1 shrank the most (-23% points).
- The table underneath gives each segment's ordered rate per week. A rate that holds steady is dependable; one that swings may be seasonal, or the segment may be too small. The first and last weeks are often only part weeks.
- Watch here for change: a segment growing, or its rate drifting, is the signal to rebuild the segments or rethink an action.
S1S2S3S4S5S6S7S8Other (7 segments)Unknown
ordered rate by week (grey: under 20 snapshots, unreliable)
| 05 Oct | 12 Oct | 19 Oct | 26 Oct | |
|---|---|---|---|---|
| S1 | 1% | 3% | 2% | 2% |
| S2 | 0% | 4% | 1% | 6% |
| S3 | 2% | 4% | ||
| S4 | 5% | 5% | 2% | 6% |
| S5 | 50% | 11% | 5% | 12% |
| S6 | 22% | 22% | 10% | 0% |
| S7 | 0% | 0% | 17% | 11% |
| S8 | 12% | 0% | 0% | 0% |
| Other (7 segments) | 0% | 0% | 14% | 10% |
| Unknown | 100% | 33% | 25% | 12% |
How good are these segments?
Separation 0.66 (clear) and stability 0.54 (fairly stable). Least stable: S3, S5, S10, S15.
- Separation (silhouette): for each snapshot, how much closer it sits to its own segment than to the next nearest. 1 is perfectly apart, 0 is on the border, below 0 is closer to another segment. Above 0.25 is reasonable, above 0.5 clear. Here: 0.66, clear.
- Stability: the segmentation was refitted 3 times on different 80% samples. 1 means the same segments came out every time, 0 no better than chance. Here: 0.54, fairly stable.
- Least stable: S3, S5, S10, S15. A segment that doesn't come back when refitted is a quirk of this fit: don't build an action on it alone.
- Confidence: the average probability with which snapshots were placed in their segment.
- The best segmentation for a decision is one that is stable and concentrates the outcome. Compare runs in the table at the bottom of the page.
Separation
0.66
clear
Stability
0.54
fairly stable
Confidence
100%
average placement
| Segment | Separation | Stability | Confidence |
|---|---|---|---|
| S1 | 0.85 clear | 0.82 very stable | 100% |
| S2 | 0.72 clear | 0.92 very stable | 100% |
| S3 | -0.07 overlapping | 0.17 unstable | 100% |
| S4 | 0.85 clear | 0.92 very stable | 100% |
| S5 | 0.07 overlapping | 0.37 shaky | 100% |
| S6 | 0.22 weak | 0.73 fairly stable | 100% |
| S7 | 0.07 overlapping | 0.75 fairly stable | 100% |
| S8 | 0.20 weak | 0.84 very stable | 100% |
| S9 | 0.63 clear | 0.70 fairly stable | 100% |
| S10 | 0.69 clear | 0.20 unstable | 100% |
| S11 | 0.91 clear | 1.00 very stable | 100% |
| S12 | 0.53 clear | 1.00 very stable | 100% |
| S13 | 0.70 clear | 0.57 fairly stable | 100% |
| S14 | 0.00 overlapping | 1.00 very stable | 100% |
| S15 | 0.00 overlapping | 0.34 shaky | 100% |
Choosing the number of segments
Groups with fewer than 50 snapshots can't be segmented on their own and are counted as too little history.
- Each device group was segmented on its own, so each chose its own number of segments by the same score (BIC, lower is better).
- Segments are numbered across all groups by size; the grey tag in the segment table shows each one's group.
- Groups with fewer than 50 snapshots can't be segmented on their own and are counted as too little history.
| device | Snapshots | Segments | |
|---|---|---|---|
| mobile | 1,685 | 5 | fixed |
| tablet | 296 | 5 | fixed |
| desktop | 265 | 5 | fixed |
The unknowns
23 snapshots (1.0%) are unknown: 0 too little history, 0 between segments, 23 fit no segment.
- 23 snapshots (1.0%) are unknown: 0 too little history, 0 between segments, 23 fit no segment.
- Too little history (0, 0.0%): fewer rows so far than the minimum you set.
- Between segments (0): no segment was a clear best fit; often someone changing pattern.
- Fits no segment (23): the oddest snapshots. Worth a look: new behaviour (or fraud) tends to show up here first.
- Nearest segment says which segment each would have joined if forced.
| Kind | Snapshots | Nearest segment |
|---|---|---|
| Unknown: fits no segment Doesn't resemble any segment. Worth a look: new patterns and fraud often start here. | 23 | S5 (8), S3 (5), S6 (4) |
Boundary rules
How to recognise each segmentTogether the rules agree with the model 98% of the time: good enough to use as a description.
- Read each line as IF these conditions THEN this segment, using the inputs in their own units (minutes, page counts).
- Covers is the share of the segment the rule finds; correct is how many of the snapshots it finds really are in that segment.
- Together the rules agree with the model 98% of the time: good enough to use as a description.
Each segment gets its own short rules for picking it out from everyone else, in each feature's own units. Taken together, simple rules reproduce 98% of the assignments. Rules show where the lines fall, not why: use the profile above for what defines a segment. Unknowns have no rules; they come from how well a snapshot fits, not from where it sits.
S1 1 rule covering 100% of the segment
- device = mobile and landingPage = / and tenure_days ≤ 0.00638 · covers 100% of the segment, 98% correct
S2 1 rule covering 100% of the segment
- device = tablet and landingPage is not /product/79-chocolate-fruit-cake-500g-boxed and landingPage is not /product/75-christmas-cake-boxed · covers 100% of the segment, 98% correct
S3 1 rule covering 100% of the segment
- landingPage is not / and device = mobile and source = Facebook · covers 100% of the segment, 87% correct
S4 1 rule covering 100% of the segment
- device = desktop and tenure_days ≤ 0.389 and source = Facebook · covers 100% of the segment, 94% correct
S5 1 rule covering 97% of the segment
- rows > 1.5 and device = mobile and source = Facebook · covers 97% of the segment, 96% correct
S6 1 rule covering 100% of the segment
- source is not Facebook and device = mobile · covers 100% of the segment, 100% correct
S7 2 rules covering 85% of the segment
- device = desktop and tenure_days > 0.389 and landingPage = / · covers 64% of the segment, 100% correct
- device = desktop and tenure_days > 0.389 and landingPage is not / · covers 21% of the segment, 70% correct
S8 1 rule covering 100% of the segment
- source is not Facebook and device = desktop · covers 100% of the segment, 100% correct
S9 2 rules covering 100% of the segment
- landingPage = /product/79-chocolate-fruit-cake-500g-boxed and device = tablet and pages ≤ 1.5 · covers 53% of the segment, 100% correct
- landingPage = /product/79-chocolate-fruit-cake-500g-boxed and device = tablet and pages > 1.5 · covers 47% of the segment, 100% correct
S10 1 rule covering 100% of the segment
- landingPage = /product/79-chocolate-fruit-cake-500g-boxed and device = desktop · covers 100% of the segment, 100% correct
S11 1 rule covering 100% of the segment
- landingPage = /product/75-christmas-cake-boxed and device = tablet · covers 100% of the segment, 100% correct
S12 1 rule covering 100% of the segment
- newVisitor ≤ 0.5 and device is not mobile · covers 100% of the segment, 67% correct
S13 1 rule covering 100% of the segment
- landingPage = /product/474-port-and-prune-pudding-800g-log-in-cloth and device is not mobile · covers 100% of the segment, 67% correct
S14 no clean rule: this segment overlaps others on these features
S15 no clean rule: this segment overlaps others on these features
Compare runs
Find the best segmentationThe pick is v20: its targets hold 83% of ordered in 25% of snapshots (3.3× the rate), with target stability 0.52.
- Each row is one segmentation run. Click the version to open it; the highlighted row is the one shown above.
- Separation and Stability describe the whole segmentation: are the segments distinct, and do they come back when refitted?
- The last four columns are about ordered: Targets are the segments at 1.5× the overall rate or more; Captured is the share of all ordered they hold (and the share of snapshots they are); Lift is how concentrated they are; Target stability is whether those particular segments come back when refitted.
- Read Captured and Lift together: a high lift on a sliver of the outcome is less useful than a slightly lower lift that holds most of it. pick marks the best mix: captured × lift, with full credit only when the targets' stability is 0.5 or more. ★ marks the best value in each column.
- Change one thing per run (the group handling, the number of segments, the shape or the feature set) so you can see what made the difference. Runs built before this table existed show '-'.
| Run | Features | Groups | Shape | Segments | Unknown | Separation | Stability | Targets | Captured | Lift | Target stability |
|---|---|---|---|---|---|---|---|---|---|---|---|
| v29 30 Oct 2026 12:50 |
v8 | device evened out | Full | 8 | 1.1% | 0.64 | 0.88 | S2, S5, S6 | 33% in 12% | 2.8× | 0.96 ★ |
| v28 30 Oct 2026 09:20 |
v7 | device evened out | Diagonal | 8 | 1.3% | 0.54 | 0.92 ★ | S2, S3, S5, S6 | 81% in 30% | 2.7× | 0.66 |
| v27 29 Oct 2026 17:17 |
v6 | device evened out | Diagonal | 8 | 1.3% | 0.54 | 0.92 ★ | S2, S3, S5, S6 | 81% in 30% | 2.7× | 0.66 |
| v26 29 Oct 2026 16:29 |
v6 | - | Diagonal | 5 | 1.0% | 0.45 | 0.59 | S3, S4 | 19% in 5% | 3.4× ★ | 0.21 |
| v25 29 Oct 2026 16:28 |
v6 | device evened out | Diagonal | 5 | 1.0% | 0.49 | 0.41 | S2, S3 | 83% in 31% ★ | 2.7× | 0.39 |
| v24 29 Oct 2026 16:27 |
v6 | device separately | Diagonal | 15 | 1.0% | 0.66 ★ | 0.54 | S5, S6, S7 | 31% in 10% | 3.1× | 0.62 |
| v23 29 Oct 2026 16:26 |
v6 | device ignored | Full | 5 | 0.7% ★ | 0.63 | 0.37 | S2, S3 | 41% in 18% | 2.2× | 0.42 |
| v22 29 Oct 2026 16:26 |
v6 | - | Full | 5 | 1.0% | 0.44 | 0.60 | S3, S4 | 19% in 6% | 3.4× | 0.21 |
| v21 29 Oct 2026 16:26 |
v6 | device separately | Full | 15 | 1.0% | 0.65 | 0.51 | S5, S6 | 27% in 9% | 3.0× | 0.57 |
| v20 pick 29 Oct 2026 16:24 |
v6 | device evened out | Full | 5 | 1.0% | 0.50 | 0.27 | S2, S4 | 83% in 25% ★ | 3.3× | 0.52 |
| v19 29 Oct 2026 16:22 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v18 29 Oct 2026 16:20 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v17 29 Oct 2026 16:20 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v16 29 Oct 2026 16:19 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v15 29 Oct 2026 14:10 |
v6 | - | Full | 5 | 1.0% | - | - | - | - | - | - |
| v14 29 Oct 2026 14:09 |
v6 | - | Diagonal | 4 | 1.0% | - | - | - | - | - | - |
| v13 29 Oct 2026 14:08 |
v6 | - | Diagonal | 9 | 1.1% | - | - | - | - | - | - |
| v12 29 Oct 2026 14:08 |
v6 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v11 29 Oct 2026 14:01 |
v5 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v10 29 Oct 2026 13:10 |
v4 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v9 29 Oct 2026 13:10 |
v4 | - | Diagonal | 7 | 89.8% | - | - | - | - | - | - |
| v8 29 Oct 2026 13:10 |
v4 | - | Diagonal | 10 | 1.0% | - | - | - | - | - | - |
| v7 29 Oct 2026 13:09 |
v4 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v6 29 Oct 2026 12:51 |
v3 | - | Diagonal | 7 | 1.0% | - | - | - | - | - | - |
| v5 29 Oct 2026 12:51 |
v3 | - | Diagonal | 10 | 1.0% | - | - | - | - | - | - |
Search results
49 setups · 30 Oct 2026 12:43 · judged on orderedBest: K-means · device evened out (8 segments). Its target segments (S2, S3, S4) hold 77% of ordered in 20% of the training snapshots, 3.8× the overall rate, and come back at 0.96 stability when refitted.
Every setup was fitted on the same 1,828 training snapshots (1,611 people); the 20% of people held out for testing weren't used, so Outcomes can still test the winner honestly.
- Every setup was fitted on the same 1,828 training snapshots (1,611 people); the 20% of people held out for testing weren't used, so Outcomes can still test the winner honestly.
- Score = share of ordered captured × lift, with full credit only when the target segments' stability is 0.5 or more. Setups with over 25% unknown are halved.
- Stability is for the whole segmentation, Target stability just for the target segments. A setup with a slightly lower score but much higher stability can be the safer choice.
- Build on full data makes the real segmentation with that setup on everyone. It then appears below as a normal run: name the segments, compare, take it to Outcomes, and export it.
| # | Setup | Segments | Unknown | Separation | Stability | Captured | Lift | Target stability | Score | |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | K-means · device evened out | 8 | 1.3% | 0.64 | 0.96 | 77% in 20% | 3.8× | 0.96 | 2.93 | |
| 2 | Bayesian Gaussian mixture · Diagonal · device evened out | 8 | 1.0% | 0.54 | 0.63 | 84% in 31% | 2.7× | 0.75 | 2.25 | |
| 3 | Bayesian Gaussian mixture · Full · source ignored | 8 | 1.0% | 0.60 | 0.69 | 78% in 27% | 2.9× | 0.54 | 2.23 | |
| 4 | K-means · device separately | 18 | 1.3% | 0.73 | 0.90 | 57% in 14% | 3.9× | 0.94 | 2.23 | |
| 5 | K-means · source evened out | 8 | 1.1% | 0.63 | 0.76 | 59% in 16% | 3.7× | 0.72 | 2.19 | |
| 6 | Gaussian mixture · Diagonal · device evened out | 8 | 1.0% | 0.55 | 0.63 | 81% in 31% | 2.6× | 0.72 | 2.14 | |
| 7 | Gaussian mixture · Full · device evened out | 8 | 1.0% | 0.55 | 0.61 | 81% in 31% | 2.6× | 0.78 | 2.13 | |
| 8 | Bayesian Gaussian mixture · Diagonal · device ignored | 8 | 1.0% | 0.67 | 0.81 | 81% in 31% | 2.6× | 0.77 | 2.11 | |
| 9 | Gaussian mixture · Full · device ignored | 8 | 1.0% | 0.66 | 0.77 | 81% in 33% | 2.5× | 0.72 | 2.00 | |
| 10 | Bayesian Gaussian mixture · Full · device evened out | 8 | 1.0% | 0.55 | 0.59 | 78% in 31% | 2.5× | 0.74 | 2.00 | |
| 11 | Gaussian mixture · Diagonal · device ignored | 8 | 1.0% | 0.67 | 0.84 | 78% in 31% | 2.5× | 0.75 | 1.99 | |
| 12 | Gaussian mixture · Full · device separately | 19 | 1.1% | 0.67 | 0.82 | 62% in 20% | 3.1× | 0.56 | 1.96 |