Attribution Modeling: Markov Chains & Shapley Value in Media Science

In multi-channel programmatic marketing, determining which touchpoint genuinely generated a conversion is one of the most mathematically challenging problems in data science. For decades, the advertising industry relied on simplistic heuristic models: Last-Touch Attribution (LTA), which awards 100% of conversion credit to the final ad clicked, and First-Touch Attribution (FTA), which credits only the introductory discovery channel. These crude heuristics distort marketing investments, overvaluing bottom-funnel branded search while starving top-funnel programmatic display, video, and social channels of credit. Advanced media science replaces heuristics with rigorous game-theoretic and probabilistic algorithms: First-Order Markov Chain removal effects and Cooperative Game Theory Shapley Value equations.

1. The Limitations of Static Rule-Based Attribution Models

Rule-based models allocate credit according to static, subjective assumptions. Linear attribution splits credit evenly across all touchpoints, failing to differentiate between an impactful 30-second non-skippable video and an incidental passive banner impression. Time-decay models assume recent touches matter more, unfairly penalizing high-impact brand awareness campaigns that initiate multi-month enterprise consideration cycles.

Data-driven algorithmic attribution relies entirely on empirical customer journey data. By analyzing hundreds of thousands of conversion paths—both successful conversions and paths that terminated in drop-off—data scientists calculate the exact mathematical contribution each marketing channel provides to conversion probability.

2. Markov Chain Attribution & The Removal Effect

A customer journey can be formally modeled as a discrete-time Markov Chain, where marketing channels represent states in a directed graph, and transitions between channels are governed by state transition probabilities. The model includes two absorbing states: Conversion and Null (Drop-off).

To quantify the attribution weight of Channel $i$, the algorithm computes the Removal Effect: it simulates what happens to total conversion probability if Channel $i$ is completely removed from the graph, routing all transitions entering Channel $i$ directly to the Null state:

# Markov Chain Multi-Touch Attribution Engine in Python
import numpy as np
import pandas as pd

def calculate_markov_attribution(journeys_df):
    """
    Computes first-order Markov Chain transition matrix and removal effects
    across marketing touchpoint sequences.
    """
    # 1. Construct State Transition Graph
    states = list(set([state for path in journeys_df['path'] for state in path]))
    transition_counts = {s1: {s2: 0 for s2 in states} for s1 in states}

    for _, row in journeys_df.iterrows():
        path = row['path']
        weight = row['conversions']
        for i in range(len(path) - 1):
            transition_counts[path[i]][path[i+1]] += weight

    # 2. Normalize to Transition Probabilities P(S_j | S_i)
    transition_matrix = pd.DataFrame(transition_counts).T
    transition_matrix = transition_matrix.div(transition_matrix.sum(axis=1), axis=0).fillna(0)

    # 3. Compute Baseline Conversion Probability via Fundamental Matrix
    # Q: transient-to-transient, R: transient-to-absorbing
    # Removal Effect: RE(i) = 1 - (P_conversion_without_i / P_baseline)
    # Attribution weight for Channel i is normalized across all removal effects.
    return transition_matrix

The Removal Effect measures the proportion of conversions that would vanish if a marketing channel ceased to exist. Channels with high removal effects are critical linchpins in the user journey, even if they never generate a direct final click.

3. Shapley Value Attribution: Cooperative Game Theory

Originating in game theory (Nobel laureate Lloyd Shapley), the Shapley Value models marketing channels as cooperating players in an $N$-player coalitional game, where the total payout is the overall volume of conversions generated. The algorithm calculates the marginal contribution of Channel $i$ across all possible subset coalitions $S subseteq N setminus {i}$:

Where $|N|$ is the total number of channels, $S$ is a coalition of channels excluding Channel $i$, and $v(S)$ represents the conversion value produced by coalition $S$. Shapley value provides four vital mathematical axioms: Efficiency, Symmetry, Dummy Player, and Additivity, guaranteeing that credit distribution is strictly fair, non-arbitrary, and provably balanced.

4. Comparative Matrix: Attribution Methodologies

5. Measuring Cross-Channel Synergy & Carryover Decay

Neither Markov nor Shapley models exist in a vacuum; touchpoints interact synergistically. For example, programmatic CTV impressions followed by targeted display remarketing routinely yield a 35% higher conversion probability than either channel deployed in isolation.

Data science teams model this interaction using cross-channel synergy matrices and adstock decay functions. Adstock transformations model the cumulative psychological retention of advertising exposures over time, applying half-life decay equations that account for delayed purchase decisions across high-consideration B2B and consumer finance purchases. Estimating decay parameters via nonlinear least squares allows the attribution model to sustain credit for foundational awareness exposures across long buying cycles.

6. Algorithmic Budget Optimization & Marginal ROAS

The ultimate goal of multi-touch attribution is capital reallocation. Merely admiring historical attribution weights provides zero commercial value; algorithms must translate Shapley scores into optimal future media spend:

Using convex optimization (e.g., Karush-Kuhn-Tucker conditions), the media mix model plots diminishing returns curves for every channel. Capital is automatically reallocated from saturated channels operating on steep marginal decay curves toward under-invested channels where marginal Return on Ad Spend (mROAS) remains high, maximizing global revenue efficiency.

7. Privacy-Preserving Attribution in Clean Rooms

With third-party cookie deprecation and mobile ID restrictions, stitching complete multi-channel user journeys across disparate platforms requires modern privacy architectures:

Data science teams deploy Data Clean Rooms (Snowflake, BigQuery, AWS Clean Rooms). Advertisers and ad networks join pseudonymized first-party event logs using homomorphic encryption and differential privacy noise. Markov and Shapley attribution algorithms execute securely within the clean room boundary, outputting aggregate channel weights without exposing individual consumer identity records. Differential privacy epsilon budgets prevent malicious reconstruction attacks while delivering mathematically provable measurement integrity.