The brighter you are, the more you have to learn.

Targeting Audience Using Factorisation Machines

Author – Author – Author – Ankit Gadi, Data Scientist at Decision Tree Analytics

Identifying a target market helps your company develop effective marketing communication strategies.The use of demographic targeting can help the advertiser to target the right audience reducing unnecessary allocation of budget to those sections of the audience that do not benefit the company.

Taking into account the age and gender of the consumers can greatly help advertisers determine their target audience for particular products in order to get the maximum clicks.

Factorisation Machines: FMs are a combination of linear regression and matrix factorisation that models sparse feature interactions but in linear time. They solve the problem of considering pairwise feature interactions. For instance if we have two variables that have multiple levels, FM considers all possible combinations of these variables. It allows us to train, based on reliable information (latent features) from every pairwise combination of features in the model. FMs replace higher order interaction effects by their factorized analogues.

From a modelling perspective, this is powerful because each feature ends up transformed to a space where similar features are embedded near one another. In simple words, the dot product basically represents similarity of the hidden features and it is higher when the features are in the neighbourhood.

Reduce Churn & Improve Loyalty: The cost of acquiring a new customer is far higher than retaining an already existing one and hence customer retention is a key business driver. You can use analytics to predict the factors that are responsible for customer churn, grow existing relationships, and use insights to reduce attrition and improve loyalty.

The advantages of using FMs over other models are:

  • FMs model equation can be computed in linear time leading to fast computation of the model.
  • FM models can work with any real valued feature vector as input unlike other models that work with restricted input data.
  • FMs allow for parameter estimation even under very sparse data
  • Factorization of parameters allows for estimation of higher order interaction effects even if no observations for the interactions are available

FMs for Target Audience Optimisation:

Understanding your customers and how to engage them is the back-bone of any marketing strategy. It is important to identify the target audience for the business, so that the company can channel its resources effectively while maximising its return.

For our business case we consider demographic targeting based on the variables age and gender. The advantage of using FMs for this purpose is that it considers various combinations of the concerned variables in order to predict the defining metric, like, CTR, clicks per spend, and so on. FMs take into consideration how these combinations respond to the campaigns, in order to make accurate predictions. The added advantage of using FMs is that it can also make predictions for thoseage-gender combinations that have not been explored in the dataset.

The results obtained after training a FM, give us a clear picture of those age-gender combinations which give the maximum values of the metric under consideration. We can then assign weights to each of these combinations based on the results obtained from FM.

The weights obtained can now be used as a constraint in addition to the various other constraints, like budget constraint, constraint for a minimum number of impressions, etc., to optimally allocate the company’s resources among the age-gender combinations.

This gives a holistic view of the working of FM and its use in targeting the best audience in the digital marketing space.