Runs a set of credit rules over the same journey table and returns their results stacked, so the spread between rules is visible rather than hidden by a choice of default. That spread is the most useful output of rule-based attribution: it bounds how much a channel's apparent value depends on the convention rather than the data.
Usage
attribute(
paths,
rules = c("linear", "first", "last", "position", "time_decay", "markov"),
first_weight = 0.4,
last_weight = 0.4,
decay = decay_from_half_life(7),
period = 1,
by = NULL,
times = 0L,
level = 0.9,
seed = NULL
)Arguments
- paths
An
mm_pathsobject frombuild_paths().- rules
Character vector of rules to apply: any of
"linear","first","last","position","time_decay"and"markov". The first five are heuristic conventions;"markov"is the data-driven removal-effect model ofmarkov_removal(). All six are run by default. A rule named twice is run once.- first_weight, last_weight
Passed to
credit_position().- decay, period
Passed to
credit_time_decay().- by
Optional names of columns of
pathsthat divide the journeys into segments, such as a device column carried by thekeepargument ofbuild_paths(). Every rule then runs on each segment's journeys alone, so shares sum to 1 within each segment and rule (or are all 0 in a segment with no conversions), and the Markov chain is fitted per segment. Each column must take one value per journey and have no missing values. See Segments.- times
Number of bootstrap replicates behind intervals on the shares: 0, the default, for no intervals, or from 2 to 1,000,000. See Intervals.
- level
Confidence level of the percentile intervals, in
(0, 1).- seed
Optional random seed, one whole number, for reproducible intervals. The caller's random number stream is restored afterwards. Giving
levelorseedwithtimes = 0has no effect, and a warning says so.
Value
A data frame with one row per rule and channel, or per segment,
rule and channel when by is given:
- The
bycolumns The segment, in the columns' own types. Present only with
by.ruleWhich credit rule produced the row.
channelChannel label.
conversionsFractional conversions credited to the channel.
shareconversionsas a share of all conversions credited under that rule. Shares sum to 1 within each rule, and within each segment; where there is no conversion to share, every share is 0.lower,upperPercentile interval for
shareatlevel, present only whentimesis above 0.NAwhere fewer than two replicates define the share, as in a segment with no converting journey.valueCredited conversion value, present only when the journey table carries values.
touchesNumber of touchpoints on the channel across all journeys, converting or not.
Rows are ordered by segment, then by rule in the order of rules, then
by descending conversions, with ties in channel order (C locale, so the
same on every platform). Segments are sorted by their values, factors in
level order.
The result has class mm_attribution. With by it also has class
mm_attribution_segments and the attribute "segments", naming the
segment columns. With intervals it also has class
mm_attribution_intervals and the attributes "level" and "times".
For these two classes plot() draws each rule's share per channel; see
plot.mm_attribution_intervals().
Segments
A segment holds whole journeys, so a column given to by must take one
value in each journey. A device column kept per touch can change within a
cross-device journey; give each journey one value first, for example the
device of its last touch:
last <- paths$touch_rank == paths$touch_n
at_last <- match(paths$path_id, paths$path_id[last])
paths$journey_device <- paths$device[last][at_last]Each segment is attributed as if it were the whole table: a channel's
share is its share of that segment's credited conversions, touches
counts that segment's touches, and a channel appears only in segments
where it has a touch. With the "markov" rule, one warning names every
segment with no non-converting journeys, as markov_removal() warns about
a whole table without them.
Intervals
With times above 0, the journeys of each segment are resampled with
replacement times times, every rule is recomputed on each resample, and
lower and upper are percentile intervals of the resampled shares. They
show how far the shares move with the sample of journeys observed, with
the journey table's construction and each rule held fixed. They say
nothing about the choice of rule, which attribution_spread() measures.
A resample gives the same shares as attribute() on a journey table
holding the drawn journeys, each draw a separate journey: a heuristic rule
credits each journey on its own, so its resampled totals reuse the credit
every journey received on the full data, and the Markov chain is refitted
from the transitions of the drawn journeys. Journeys are drawn
independently, so a customer's several journeys are treated as unrelated.
A resample with no converting journey defines no share; it is left out of
the intervals and counted in a message.
Every draw comes after seed is set, and the caller's random number stream
is restored afterwards, as in block_bootstrap(). Without seed the
draws advance the session's stream.
Examples
data(mm_events)
paths <- build_paths(mm_events, id = "customer_id", channel = "channel",
timestamp = "timestamp", conversion = "conversion",
value = "value")
res <- attribute(paths)
head(res, 10)
#> rule channel conversions share value touches
#> 1 linear paid_search 494.40675 0.18037459 63269.896 3203
#> 2 linear display 482.35595 0.17597809 60241.653 3011
#> 3 linear social 460.37460 0.16795863 59182.995 2818
#> 4 linear organic_search 425.06786 0.15507766 51503.251 2741
#> 5 linear email 329.57579 0.12023925 42200.946 2178
#> 6 linear affiliate 265.93214 0.09702012 33558.401 1734
#> 7 linear (blank) 91.17024 0.03326167 11382.201 465
#> 8 linear (none) 81.73690 0.02982010 9640.961 427
#> 9 linear direct 66.63452 0.02431030 7771.148 444
#> 10 linear (missing) 43.74524 0.01595959 5951.438 261
# Total credited conversions equals observed converting journeys, per rule
tapply(res$conversions, res$rule, sum)
#> first last linear markov position time_decay
#> 2741 2741 2741 2741 2741 2741
# How much does the answer depend on the rule?
attribution_spread(res)
#> channel min_share max_share mean_share spread
#> 1 display 0.07187158 0.29113462 0.17382396 0.2192630427
#> 2 paid_search 0.09157242 0.27252827 0.18342704 0.1809558555
#> 3 email 0.05654870 0.18314484 0.12349952 0.1265961328
#> 4 social 0.11638088 0.22072236 0.16553747 0.1043414812
#> 5 organic_search 0.12623130 0.17876687 0.15470603 0.0525355710
#> 6 affiliate 0.08062751 0.11711054 0.09861140 0.0364830354
#> 7 direct 0.02225465 0.02682848 0.02410676 0.0045738324
#> 8 (blank) 0.02955126 0.03378661 0.03175589 0.0042353471
#> 9 (none) 0.02736228 0.02982010 0.02849282 0.0024578273
#> 10 (missing) 0.01568771 0.01637007 0.01603912 0.0006823617
#> first_last_ratio
#> 1 4.0507614
#> 2 0.3360107
#> 3 0.3087649
#> 4 1.8965517
#> 5 0.7061224
#> 6 1.4524887
#> 7 1.0491803
#> 8 0.9418605
#> 9 1.0133333
#> 10 1.0232558
# How much does it depend on the sample? 90% intervals from 50 resamples
boot <- attribute(paths, rules = c("first", "last", "markov"), times = 50,
seed = 1)
head(boot)
#> rule channel conversions share lower upper value
#> 1 first display 798 0.29113462 0.27584561 0.30603943 100719.02
#> 2 first social 605 0.22072236 0.20889362 0.23312958 78372.54
#> 3 first organic_search 346 0.12623130 0.11601449 0.13602582 41455.44
#> 4 first affiliate 321 0.11711054 0.10586710 0.12843015 40943.15
#> 5 first paid_search 251 0.09157242 0.08385965 0.09976528 30259.63
#> 6 first email 155 0.05654870 0.05136549 0.06496043 20063.21
#> touches
#> 1 3011
#> 2 2818
#> 3 2741
#> 4 1734
#> 5 3203
#> 6 2178
# Credit per segment, here a device recorded once per customer
ev <- mm_events
ev$device <- ifelse(grepl("[02468]$", ev$customer_id), "mobile", "desktop")
kept <- build_paths(ev, id = "customer_id", channel = "channel",
timestamp = "timestamp", conversion = "conversion",
keep = "device")
by_device <- attribute(kept, rules = c("first", "last"), by = "device")
by_device[by_device$channel == "display", ]
#> device rule channel conversions share touches
#> 1 desktop first display 376 0.27769572 1506
#> 16 desktop last display 100 0.07385524 1506
#> 21 mobile first display 422 0.30425379 1505
#> 36 mobile last display 97 0.06993511 1505