As the initial hype and uncertainty around machine learning settles down there are still many unanswered questions around its application to business challenges. In the media planning space, one of these burning questions is “do we really need to invest in machine learning or are our traditional statistical models offering the same level of performance”. In this article we delve into the answer to this question.
As far back as history can recollect, businesses have explored various techniques to expose people to their products in different ways. Ranging anywhere from shouting the effects of your product audibly around the local market to enriching people’s attention via televised advertisements. This stretching of the vocal cords, while being a good way to get some attention, isn’t as cost effective and efficient for your business on a large scale as many hope.
As techniques have come and gone, we have arrived at a point in time where direct response media planning is a significantly more effective (and measurable) way to sell your product. With televised, direct response media planning one simply has to book advert spots to be aired; wait for your advert to show the audience the value of your product and efficiently respond to customer requests as they come in.
Sounds simple enough but how do you pick the best spots to achieve your desired target within the allotted budget? Marketers are solving this challenge through the use of modelling to determine which spot to book (considering which time, channel and at most efficient cost) and how many responses to expect from each spot.
But which model build choice should savvy marketers make? Will the newer machine learning models perform better or will the traditional, statistical modelling approach give you the information you need? Let’s consider the main differences between the two approaches:
- Purely statistical models have many prior assumptions regarding the variable distributions and often require the relationship to be linear between the dependent and independent variables. Machine learning models have no rigid prior assumptions and calculate the relationship according to their algorithm.
- The redundancy in variables is more permissible in machine learning models as they can handle greater numbers of variables and usually focus on more useful variables, identifying predictive pockets of data.
- Machine learning models usually train by iterating through different parameter combinations to optimize specific selection criteria, e.g. errors or RMSE, and therefore many ideal variations could exist. Statistical models assume there is only one best fit.
Currently, purely statistical models are still widely used as they have established their usability and stability. However machine learning models are slowly but surely gaining in popularity as they inspire more confidence in their abilities due to ease of application and higher accuracy of predictions. It also helps that it has become significantly easier to gain access to the computing power required for the machine learning models.
At this point it sounds like machine learning is edging into the lead however I must issue a word of warning. The one size fits all rule is not always advised for model building and hence each environment requires its own choice between the two model approaches in accordance to the situation. Choosing one approach does not necessarily exclude the other from being useful. Consider the following situations where the two models can work hand in hand:
- Machine learning models can be used alongside statistical models to provide insights on variable creation or to evaluate the importance of the variables to be used for prediction. For example, let’s say we include a variable in our model that consists of whether the advert spot falls on a school holiday or not. Using the variable in the statistical model, it might not be evident that a specific few days are more predictive than others. If however you use machine learning models to identify the predictive pocket of data you could adjust the variable of the statistical model to cater for the pocket or let the variable be the pocket itself.
- Machine learning models can be used alongside statistical models or alongside different variations of themselves to create ensemble models. These models work by using different models to predict the amount of the responses for the advert and then either assigning their predictions different weightings or building a model using their predictions as variables to predict more accurately. Adding more prediction layers sounds overly complex, but if the end predictions are still even more accurate and stable enough, it is a viable option.
The logic behind ensemble models is fundamentally relatable to occurrences in our daily society. Imagine writing an essay at school and handing it to the teacher to mark. The teacher marking the essay might inherently have specific preferences, insights or bias towards a subject, in the end reflecting in your mark achieved. How much the mark is affected depends on the teacher. Now imagine handing your essay to 3 teachers who mark it together. Even though each one might have preferences, insights or bias they could help one another to evaluate your essay from different viewpoints (even if only different by a small amount) with the effect that your mark will be more accurate.
In conclusion, which ever model approach is chosen, there is more than enough reason to support the use of machine learning models in direct response media planning. While there is no clear winner yet, in the long run the most accurate models will become the standard, especially as so much depends on them.
Businesses that advertise but don’t have access to machine learning models are likely to struggle over time to stay ahead of their competitors, in terms of keep their campaigns efficient and effective.
If your strategic objectives in terms of core competencies don’t align with creating your own data scientist team at this point in time; contact Incline at email@example.com. Incline’s application of machine learning models has proven benefit for our local and international clients.