Prediction is an exciting, relatively new addition to the data types that marketers can use to enhance customer engagement. However, as with any new technology, idea or trend, it is not without its challenges.
In this blog, we briefly discuss four major challenges involved in developing predictive analytics applications.
Ethical challenge – privacy and ownership
A prime challenge is privacy and ownership of profiles aggregated on users. One major question is who owns user profiles? Is it the user herself, or the company collecting the profiles? Who has access to this data? What can be inferred form user behaviour and attributes and who can we use this inferred knowledge?
The next question is security. What is the security measure around storages keeping user profiles? Is the cloud based storage or data is stored in on premise infrastructure ?… Despite significant advancements in privacy preserving predictive analytics solutions and recommendation methods, and a multitude of highly secure cloud based storage solutions, privacy is still (and probably will continue to be) an important issue that needs to be addressed properly. When developing predictive applications, this presents an ethical challenge.
Interpretability challenge – Accuracy vs interpretability
The second challenge is the trade-off between prediction model accuracy and interpretability of predicted results. The more accurate the prediction model becomes, (more often) the more complex the model and, as a result, the more difficult it becomes to interpret how such prediction is achieved based on given predictive variables.
While, for a machine learning expert, understanding the result can be relatively straightforward, it would be challenging for people with less technical expertise in the field to grasp how a complex prediction model behaves under different circumstances of predictive variables (i.e. Inputs). The data gained from predictive analytics needs to be easily understandable for marketers, many of whom will not have as much technical expertise as a machine learning expert.
Scalability challenge – coping with more
The third challenge is scalability of the system in implementing the model.
Scalability means that, for example, if you are providing a personalised news recommendation system, how the system copes with growth in volume, variety and velocity of user and news items. Not to mention peaks in the number of users in case of disasters, world cups, elections, etc…
Scalability involves both scaling underlying infrastructure (e.g. memory, bandwidth, storage, etc.), utilised platforms (e.g. Message queues, stream processing system, database, etc.) as well as scalability of a core prediction or recommendation algorithm. For example, collaborative filtering recommendation systems suffer from expensive computations that grow significantly with the number of users and items in the database. On top of this, many of the recommendation systems need to operate in real time and respond in a fraction of a second!
Feasibility challenge – goal setting
The fourth and last major challenge comes from a business perspective: defining a clear goal and identifying the benefits for targeted predictive analytics application. A predictive analytics or recommendation system is not the solution for every e-commerce problem – sometimes even the benefit gained from utilising such a predictive system is marginal. Without a clear and sound feasibility study together with prediction objectives, businesses are unlikely to gain real benefit.
And it’s not just the business – close collaboration between data science and marketing teams is required, to ensure that the proposed predictive solution is an appropriate solution for customers too.
For predictive analytics to add real value, businesses need to consider the privacy and interpretability of the data, together with the business factors of scalability and feasibility. With these four boxes ticked, prediction can integrate with your data and marketing strategies and help to deliver an unrivalled customer experience.
The Profile Cloud predictive functionality is being developed with support from Vinnova – a Swedish Government Agency for Innovation Systems.
This post was written by Shahab Mokarizadeh