Data-driven attribution from a first-order Markov model of the journey table (Anderl, Becker, von Wangenheim and Schumann, 2016). Journeys are treated as walks from a start state through channel states to one of two absorbing states, conversion or null. A channel's removal effect is the proportional drop in the probability of reaching conversion when every transition into that channel is redirected to the null state; conversions are then divided among channels in proportion to their removal effects.
Arguments
- paths
An
mm_pathsobject frombuild_paths(). Keep the non-converting journeys (keep_null_paths = TRUE, the default): they are what the null state is estimated from.
Value
A data frame with one row per channel, ordered by descending
conversions: channel, removal_effect, conversions (credited),
share, value when the journey table carries conversion values, and
p_without, the probability of conversion once the channel is removed.
Three attributes let you check the removal effects by hand:
p_conversionThe baseline conversion probability.
removal_effectis1 - p_without / p_conversion, or 0 when no journey converts.transitionsThe fitted first-order transition matrix, with one row for each state a walk can leave (
"(start)"and every channel) and one column for each state it can enter (every channel,"(conversion)"and"(null)"). Each row sums to 1.transition_countsThe transitions observed, in the same layout, from which
transitionsis estimated.
A zero-row journey table gives a zero-row result without these attributes.
Details
The probabilities come from the absorbing-chain fundamental matrix, so they
are exact for the fitted transition matrix rather than simulated: with
\(Q\) the transitions among transient states and \(R\) those into the
absorbing states, absorption probabilities are \((I - Q)^{-1} R\). One
inverse serves every channel. Once channel \(s\) is removed, a walk
converts only if it converts without reaching \(s\), so with
\(N = (I - Q)^{-1}\) and \(x = N R\) the conversion probability is
\(x_1 - (N_{1s} / N_{ss}) x_s\), where \(N_{1s} / N_{ss}\) is the
probability of ever reaching \(s\) from the start and \(x_s\) that of
converting from \(s\). That probability is the p_without column.
A channel named "(start)", "(conversion)" or "(null)" shares its
label with a special state in transitions; index the matrix by position
in that case.
This is the order-1 model, with no dependency beyond base R. For
higher-order chains, or journey tables with millions of distinct paths,
hand the output of as_channel_paths() to ChannelAttribution, whose
markov_model() implements the same removal-effect definition in C++. For
order = 1 the two agree up to the error of ChannelAttribution's
simulation.
Removal effects are sensitive to journey construction: repeats collapsed or
not, direct traffic kept or dropped, lookback length. That is the argument
for build_paths() making each an explicit choice.
What this is not
A removal effect is a statement about the fitted transition matrix, not
about what customers would do if a channel were switched off. Removing a
state from a chain assumes that everyone who would have passed through it
is lost, and that nobody reaches the same conversion another way. It is a
more data-driven convention than first- or last-touch, and it is still a
convention. Use it alongside the heuristic rules in attribute(), where the
spread between them is the honest summary, and use experiments for
incrementality.
References
Anderl, E., Becker, I., von Wangenheim, F. and Schumann, J. H. (2016). Mapping the customer journey: Lessons learned from graph-based online attribution modeling. International Journal of Research in Marketing, 33(3), 457–474. doi:10.1016/j.ijresmar.2016.03.001
See also
attribute(), which includes this as the "markov" rule,
as_channel_paths() for ChannelAttribution.
Examples
# A three-journey example small enough to check by hand: P(conversion) is
# 2/3; removing "a" halves it, removing "b" makes conversion impossible.
ev <- data.frame(
id = c("j1", "j1", "j2", "j3"),
ch = c("a", "b", "a", "b"),
ts = c(1, 2, 1, 1),
conv = c(0, 1, 0, 1)
)
p <- build_paths(ev, id = "id", channel = "ch", timestamp = "ts",
conversion = "conv")
mk <- markov_removal(p)
mk
#> channel removal_effect conversions share p_without
#> 1 b 1.0 1.3333333 0.6666667 0.0000000
#> 2 a 0.5 0.6666667 0.3333333 0.3333333
# The fitted chain behind those numbers
attr(mk, "transitions")
#> to
#> from a b (conversion) (null)
#> (start) 0.6666667 0.3333333 0 0.0
#> a 0.0000000 0.5000000 0 0.5
#> b 0.0000000 0.0000000 1 0.0
attr(mk, "p_conversion")
#> [1] 0.6666667
1 - mk$p_without / attr(mk, "p_conversion")
#> [1] 1.0 0.5
data(mm_events)
paths <- build_paths(mm_events, id = "customer_id", channel = "channel",
timestamp = "timestamp", conversion = "conversion",
value = "value")
markov_removal(paths)
#> channel removal_effect conversions share value p_without
#> 1 paid_search 0.43712431 487.56069 0.17787694 61314.695 0.2527178
#> 2 display 0.41872296 467.03614 0.17038896 58733.567 0.2609796
#> 3 social 0.40268569 449.14846 0.16386299 56484.047 0.2681799
#> 4 organic_search 0.37610320 419.49882 0.15304590 52755.365 0.2801148
#> 5 email 0.31515234 351.51532 0.12824346 44205.890 0.3074803
#> 6 affiliate 0.25688705 286.52725 0.10453384 36033.116 0.3336401
#> 7 (blank) 0.07606490 84.84143 0.03095273 10669.496 0.4148249
#> 8 (none) 0.06855467 76.46466 0.02789663 9616.049 0.4181968
#> 9 direct 0.06592975 73.53687 0.02682848 9247.856 0.4193754
#> 10 (missing) 0.04022868 44.87035 0.01637007 5642.809 0.4309145