Google announced that it is officially killing last-click attribution at Marketing Next 2017, the companies annual event to promote ad products, analytics, and Double Click. Google launched Google Attribution, the companies new product which uses machine learning to assign a weighted value to every different touch point along the consumer’s path to purchase.
Last-click model used by digital marketers gives the last-click credit for sale or conversion. With the convergence of offline and online channels, the model was always debated to be not suitable when it comes to evaluating media effectiveness. With this new attribution model, the goal is to bring in more efficiency when it comes to media spends across channels and devices. For instance, if the last click before purchase from Google search, an earlier e-mail campaign may also get its share of credit. Machine learning will enable the brand to build a better prediction model as it learns by weighting a set of points and learns how likely a user is to purchase something.
Machine learning has already made inroads into marketing with use-cases that range from pattern recognition to predicting and forecasting trends. Machine learning is not the same as Artificial Intelligence. ML, unlike AI, doesn’t try to surpass human intellect or develop cognitive abilities. ML is focused on optimizing certain problems – solving processes. As the late American pioneer, Arthur Samuel said: “Machine learning gives computers the ability to learn without being explicitly programmed.”
Target a leading discount chain retailer in the United States was among the first brands to use machine learning to identify buying patterns and behaviors using existing data. The brand was able to determine a teenager’s pregnancy before her parents as machine learning examined the shopping history and started giving context to past purchases from the customer. Based on the shopping behavior of the customer the team at Target was able to arrive at the conclusion that the girl was pregnant.
Zomato a leading restaurant search and discovery service based out of India was seeing an increasing number of biased reviews posted on its platform for restaurants. To curb the biased reviews on the platform, the brand used machine learning to create a new anti-bias algorithm that cleaned up biased reviews retroactively and also puts in bias check for future reviews. The algorithm mainly helps Zomato by hiding biased reviews, assigns credibility scores to users based on the reviews posted and reduces the effect of older bad ratings received by the restaurants.
Machine learning has now been widely adopted by leading brands like Flipkart with the Microsoft Azure partnership), FreshDesk (Amazon Lex for Machine Learning & NLP) and Ola’s Data Science Group is using machine learning to solve a broad range of issues like pricing, navigation/ETA prediction, and demand forecasting.
Machine Learning & Marketing
Kahuna, a leading marketing automation platform, is using machine learning to help brands to deliver better messaging to customers. Kahuna collects data from apps & services customers use and analyzes to understand which messages work the best across multiple variables. The platform then enables brands to craft the perfect message for each user delivered to the right channel at the right time.
So how do you get started with machine learning? To solve the problem it’s important that you define the problem first.
Identifying the problem
Machine learning should be considered when you cannot code the rules. For example, if your e-mail system has to identify spam it’s practically impossible to code the rules considering the multiple or more variables involved. Also, ML has to be considered when you are working on scale.
It’s important you define the problem in terms of what is observed and what answers do you want the model to predict.
Collecting & Analyzing Labelled Data
Gather and sort the data to be made available to the ML model training algorithm. Validate the quality of data running sanity checks. Often the raw data (input variables) and answer (target) are not suitable to train a highly predictive model.
ML practitioners, therefore, typically should attempt to construct more predictive input representations or features from the raw variables. Data has to be ideally be split between training the ML model and evaluating the model’s performance.
Model Accuracy
Once you have a model it’s important you evaluate the model’s accuracy on unseen examples that you have not used for the training model. To do this you can add the data previously held back from the evaluation dataset and then compare the results from the previous dataset.
Machine learning is not as easy as it sounds but as a marketer, it’s important to know how the solution works and how you can use it to your advantage. Brands that strive to be relevant to their consumer need to experiment with technology and machine learning should certainly be on your checklist.