A Machine Learning Approach to Marketing Attribution

LIMITLESS INFORMATION!

People today are increasingly bombarded with data across different media, more so in the digital space. A recent finding says over 90% of the data have been created in the last 2 years, which is nearly 2.5 quintillion (2.5 followed by 18 zeros!) in volume. This has opened up a whole new world of information for customers, who now have better visibility and awareness about various brands and products at their fingertips. They come across information in different formats, across diverse platforms like print, TV, radio, search engines, social media, content sites, and blogs, before finally making a purchase.

A BRAND’S DATA HURDLE

Brands need to engage with customers across these platforms, be it through phone calls, marketing emails, app/website notifications, social media engagement, etc. These different channels generate impression-level data that can be linked to unique customers, individuals or households. By recognizing every step in this journey, brands can gain a significant advantage over competitors.

However, in reality, they’re far from the ideal world. Brands today are always playing catch-up, trying to figure out and improve processes that effectively map the entire customer journey in this dynamic and omnichannel environment. On many occasions, the data trail left behind by the customers is left untouched. A Forrester report states that nearly 75% of all data retained by an organization is never analyzed or used. This goes on to show how crucial it is for brands and ad platforms to realize and leverage the immense potential of data analytics. We shall look at a few important attribution models in the coming sections.

OVERCOMING THE MARKETING CHALLENGES

Continuing from the brand’s perspective, the primary challenge here is being able to correctly attribute ROI to different marketing campaigns and efforts. They also face hurdles in effectively mapping the customer journey, which is usually based on accurate touch-points or different levels of engagement to lead them closer to the end goal of conversion. Such marketing challenges often arise due to ineffective tracking or measurement, rather than campaign performance issues.

Here’s where marketing attribution modelling comes into the picture.

Marketing Attribution is the process of identifying and measuring the impact of a series of user actions (touch-points) that contribute to the desired outcome of customer conversion. Each touchpoint is assigned a value relative to the total value of customer engagement during that journey. The main goal is to help brands understand the value of each interaction through those engagements in the entire journey. Marketing personnel can better optimize their spending and messaging by understanding which channels or touchpoints can lead to a higher conversion rate to their desired outcomes. This provides a level of understanding and clarity to marketers as to which set and sequence of user actions would typically influence users to engage in a desired behaviour, resulting in conversion.

WHY IS MARKETING ATTRIBUTION IMPORTANT?

Effective attribution is a holy grail of digital advertising. A customer can encounter your brand in many ways: organic results on a search engine, display media campaigns, social media links, or even re-targeting on external sites, etc. Access to different platforms gives customers multiple options to research and buy things using different devices in different locations. Hence capturing every touch-point is highly critical to the success of attribution modeling.

Marketing attribution helps you understand what your customers want, leading to smarter spending decisions. By helping you understand how different interactions affect movement along the customer journey, attribution makes it easy to prioritize the right content and channels.

A recent HubSpot report shows nearly 74% of companies are prioritizing converting their contacts/leads to customers in the immediate future. This again underlines the importance of efficiently mapping the entire customer journey and streamlining marketing efforts towards customer conversion.

Source: HubSpot

To do so, there are 3 main parts in any Marketing Attribution exercise:

1. Measuring the relative effectiveness of different strategies on a single device
2. Tracking campaigns across different devices
3. Attributing offline outcomes to a particular campaign

Here are a few key things to keep in mind while implementing your marketing attribution strategy:

Key steps to implement Marketing Attribution

CHOOSING THE RIGHT ATTRIBUTION MODEL

There are different attribution allocation models to achieve the common goal of customer conversion. A large number of marketers use rule-based attribution methods. Sometimes more than a single weighting system is needed to attribute correct credit for any given campaign sequence. For each interaction, credit varies depending on the time, order and interaction nature.

Some of the rule-based methods are as follows:

Rule-based MTA models

1. First-Touch Attribution In this model, the first touchpoint that begins a conversion journey receives a full 100% credit. This model is best used with conversions that have a short consideration cycle. It is widely used by Display and Social Marketing teams to analyze the effectiveness of their brand awareness and customer acquisition efforts.
2. Last Touch Attribution In this model, the last touchpoint gets 100% of the credit in the conversion. Similar to the earlier model, the last-touch attribution is used with conversions that have a short consideration cycle. Teams that manage search marketing or analyze search keywords often use this model.
3. Linear Attribution: This model gives equal credit to each touchpoint leading up to the conversion. It is best used for campaigns with long consideration cycles or customer experiences that require frequent or consistent engagement. Mobile app teams use this model to measure notification effectiveness. While linear attribution arbitrarily allocates an equal credit weighting to every interaction along the customer journey, it is only slightly better than the first-touch and last-touch approaches. Linear attribution often tends to under-credit or over-credit specific interactions.
4. Time Decay In this model, credit for each touchpoint depends on the time that passes between the interaction and the conversion. It is best used for promotions that run across a predetermined number of days. Teams that schedule marketing around sporting events use this model to give more credit to all touches closest to conversion. This attribution method arbitrarily biases the channel weighting toward the most recent channel or touchpoint across the customer’s journey.

First-Touch and Last-Touch Attribution have a fundamental weakness in that they ignore all other interactions with your brand across a multi-touch journey.

Web Analytics like Google Analytics have traditionally defaulted to First-Touch and Last-Touch Attribution approach for performing attribution analysis. This is also because these attributions are easy and simplistically identifies ownership of the converting visit.

DATA-DRIVEN ATTRIBUTION STRATEGY

Marketers usually rely on last-touch or simple rule-based attribution approaches. However, businesses can gain significantly by adopting a data-driven attribution approach. With the increasing number of channels to reach out to customers, it is easy to divert attention and lose track of which channels are best suited for your customers. That’s when the data-driven attribution model puts you back on track.

This model is created by capturing every possible variation of cross-channel and cross-section activities at a granular user level. This process accords no predetermined weight to first-touch, last-touch, linear or other channels. Data-driven attribution is an intensive data-modelling exercise in which complex algorithms find and analyze statistically relevant patterns across huge volumes of quality data. It creates a “Single Source Of Truth” (SSOT) about marketing performance that unifies and completes your analytics. Organizational silos are required to work together and communicate closely for capturing usable data and driving data-driven insights into action. Data is the fuel behind all the insight, so maximizing this is crucial for accuracy and granular analysis.

Companies need an effective combination of technical knowledge and communication skills to build a data-driven attribution program. The main benefit is that results are measured commonly and consistently. Companies looking to solve their attribution problem need to lay a solid foundation: a platform that puts data at the centre of the strategy, standard metrics in place for all channels, communication across organizational silos, and a plan for integrating existing data and tactics for accommodating future permutations.

WHAT ARE THE VARIOUS DATA-DRIVEN ATTRIBUTION MODELS?

There are several algorithmic methods for allocating online/offline advertising exposure to multiple channels through attribution models. The most popular and effective options include:

GENERALIZED LINEAR MODELS (GLMs & GAMs)

Sometimes, marketing attribution can be simplified as a classification or a regression use case.

All sorts of campaign-related attributes like the type of campaign, markdowns, discounts, offers, bundles, segments of products covered customers and time attributes can be used to build an attribution model.

Generalized linear models and additive models can be used for data-driven revenue-based attribution modelling to determine the attribution of a given action (any marketing/advertising activity on any given channel) in terms of sales/revenue for a given duration. Also, they provide a more accurate overview of the multi-touch attribution problem by capturing the co-occurrences and other interactions among advertising channels.

The advantage of such methods is the ability to evaluate the predictive power of the models using the common model validation techniques (validation set \ AUC \ pseudo R squared etc).

But there exists a few issues such asexist too including the difficulty of interpreting the spurious correlations among the variables and the coefficients in business terms.

MARKOV CHAIN MODELING

Markov chains are a type of Probabilistic model. They represent a set of events that occur sequentially and connect each event with every other event through conditional probabilities. In the context of marketing attribution, they represent the probability that a customer would move from one step (channel/action) in the customer journey to another and only depend on the past steps.

Depiction of Markov chain-based Interactions between channels

To implement Markov Chains to build an attribution model, the following steps are followed:

1. Model the customer journey based on the conditional probabilities
2. Simulate the conversion of users based on the probability matrix to obtain the conversion rates of the entire system
3. Understand the contribution of every channel by removing each of them from the system and re-calculating the conversions, known as the Removal Effect
4. Deduce which channels are the most important

GAME THEORY & SHAPLEY VALUE

Game theory attribution uses algorithms and the Shapley value to identify the impact of each touchpoint and then fairly distribute credit to each touchpoint in a conversion path. It paints a clearer and more accurate picture of the customer journey and identifies which touchpoints perform the best and which ones don’t.

In the context of marketing attribution, game theory can be used to model the customer interactions with the marketing channels as a cooperative game where each marketing channel can be seen as a player in the game. The set of all players/channels can be thought of as working together to drive the conversions and assign each touchpoint fair credit (using the Shapley Value) for a conversion based on their true contribution.

Depiction of Shapley values of channels

REAL-WORLD IMPACT

In this section, let us look at the implementation of a Generalized Additive Model (GAM) model for a real multi-touch attribution (MTA) problem.

One of our clients — a global CPG giant, wanted

“To improve the performance of digital marketing campaigns and increase Return on Marketing investments by understanding the contribution of different factors to sales”.

To solve this, Sigmoid built a GAM-based attribution model to obtain attribution of different factors contributing to sales and to generate results at multiple granularities like campaign, tactic, creative, audience, product, segment, year, week, zip etc.

Introduction to GAMs

GAMs are simply a class of statistical Models in which the usual Linear relationship between the Response and Predictors is replaced by several Nonlinear smooth functions to model and capture the non-linearities in the data.

They also help us to fit Linear Models which can be either linearly or non-linearly dependent on several Predictors to capture non-linear relationships between Response and Predictors.

Approach

To build the model, multiple Marketing factors such as TV advertising and digital advertising as well as non-marketing factors such as demographic and seasonality were considered.

Sales was modelled as:-

Sales = constant + F(factor 1) + F(factor 2) + F(factor 3) ….

where F(x) represents a nonlinear/linear function over any factor x

It represents the contribution of each factor x to sales.

A few key factors considered for the Model were:-

• Past Sales
• Retailer Price
• TV Spend
• TV GRP (Gross Rating Point) ratio of Non-Peak vs Peak Time
• Temperature (Seasonality)
• Week (Seasonality)
• Households (Demographic)
• Facebook Impressions
• DoubleClick Campaign Manager(DCM) Impressions

A GAM model for sales as a function of Marketing channel spending and other factors

The contribution provided us with the Incremental Volume generated due to various marketing actions which in turn is used to estimate various KPIs — Efficiency, ROI, Effectiveness for a given action.

Predictive KPIs helped the marketers tweak the campaign planning parameters to achieve business goals of ROI and effectiveness.

An overview of the approach and the business outcomes

Business Impact

• Reduced campaign evaluation timelines from 6 months to 1 month
• Improved the performance of planned campaigns through predictive KPIs and recommendations
• Employed Predictive KPIs as feedback on planned actions
• Created recommendations that were based on rules identified using pattern and rule-mining algorithms
• Overall 11% improvement in planned campaigns
• Developed the ability to modify campaigns mid-flight leading to better ROI
• Created Potential savings of 220k USD over 15 weeks in digital campaigns in a single product segment

FINAL WORDS

Today, marketing has evolved and become highly personalized thanks to focused analysis of diverse data points. Considering that 64% of advertisers globally are now leveraging big data from third-party sources, it’s proving to be the workhorse behind modern omnichannel attribution.

Effective data-driven marketing attribution, therefore, plays a vital role in helping marketers navigate the daily challenges of taking the customer another step closer to conversion. It correctly highlights the interplay between channels and helps them identify and focus on those steps that are taking the customer closer towards conversion.