Attribution modelling for measuring brand effectiveness has been on the rise. However, the method, which involves assigning credit for sales and conversions to touchpoints (i.e., clicks) in conversion paths, is said to have shortcomings. For instance, it is difficult to make a causal inference that among all factors, what is to be attributed for conversion or sale. Experts believe that the short-term, direct effects say little about effectiveness, calling for a different evaluation method. As a result, big-tech companies like Meta and Google are moving to econometrics for evaluating brand effectiveness. 

Econometric modelling involves statistically inferring the relationships between large quantities of data (called variables) and analysing their correlation over time. Hence, it will consider parameters beyond clicks, namely, non-media drivers of sales, such as seasonality, price changes, and so on. 

Varun Kumar Angurulia, Head of Marketing Analytics at L&T Financial Services, told AIM that marketing analytics is a diverse field and generally requires you to be closer to data, business, and market trends. Establishing a relationship between them often gets complicated with the levels of attributes included. This gap is filled by econometric modelling, allowing organisations to strategise the best actions. Angurulia added that prior to getting into complex solutions, exploration with data—which includes utilising basic statistical concepts like central tendencies and dispersion hypothesis—gives visibility to the direction organisations need to take. 

Econometrics x Machine Learning 

There has been sufficient growth towards clubbing econometrics with machine learning in recent years. This expands the range of hypotheses that can be explored with the available data. For example, ML-powered Marketing Mix Model (MMM) can aid advertisers in figuring out cross-media Return on Investment (RoI) and budget allocation decision-making. However, achieving this is a complex multidimensional challenge that requires considering aspects like non-linearity and saturation, time-delayed impact (Adstock effect), and interaction effects. Historically, linear models have also poorly represented these aspects while also taking a long time if numerous iterations have to be run to calibrate the model.

• Non-linearity: MMM typically follows a logarithmic relationship. This means that to avoid spending too much or too little, one has to find the optimum middle where the media output is used efficiently for a particular budget. 

• Adstock effect: It is not always the case that a campaign will immediately impact business outcomes. Some strategies may take weeks to have an influence over the consumers. These strategies generally take steps to build brand awareness, only following which sales come into the picture. 

• Interaction: As mentioned earlier, consumers generally interact with various touchpoints before making a buying decision, and it is difficult to identify the effect of certain channels correctly. 

ML algorithms can develop complex models studying the patterns or correlations based on sample data and then generalise and apply them to new unknown data. It takes into account the non-linear correlations, the time delay effect, and the interaction effect of media. Here, in contrast to linear models, complex ML-based algorithms can include many variables without leading to over-specification. This implies that it can realistically represent the interaction effects between the different characteristics. 

Interpretability in the model is crucial for marketing analytics. The major features that affect a model’s outcome should be known, along with their direction and quantity of influence. Hence, linear models are still used in marketing analytics for their ease of interpretability, even if the output is less accurate. However, in recent times, two prominent approaches have emerged to interpret complex non-linear models and show which input is decisive for the forecast results: Shapley Additive Explanation (SHAP) and Local interpretable model-agnostic explanations (LIME). Anurag Pandey, a Data Science Engineer, writes in his blog, “The interesting thing to note is that when complex models are combined with the explanation models, they outperform [Generalised Linear Models] GLMs not only in accuracy (obviously) but also in interpretability/explainability.”

Approaches for Explainable AI

Shapley Additive Explanation (SHAP) is a game theoretic approach which connects optimal credit allocation with local explanations using Shapley values. Shapley values guarantee uniqueness in the solution by telling how to distribute the contribution by the features accurately. So, for example, if we take the case of MMM itself, the dataset can include different marketing channels, such as TV, Radio, and Print. There could be a predicted return value for each instance of data (indicating the marketing spend). The predicted result for the particular dataset can be compared with the overall average prediction value to understand the difference.  

Hence, as Anurag Pandey explains, two things are working here: a machine learning algorithm for prediction and Shapley values to understand the feature behaviour for each instance of data. Thus, SHAP is a tool to make explainable AI by visualising the output. Moreover, it can explain any model’s prediction by computing each feature’s contribution to the prediction. 

Read: A Complete Guide to SHAP – SHAPley Additive exPlanations for Practitioners

Similarly, another approach is Local interpretable model-agnostic explanations (LIME). This method is typically used for texts and images and slightly less for tabular data. A simple example of this would be that it isolates individual predictions and trains an interpretable model for this prediction by changing input values (for example, turning certain pixels on and off) and documenting how these values impact a complex model’s decision. 

The post How Marketing Analytics Leverage Explainable AI for Forecast appeared first on Analytics India Magazine.