The usual approach to improve targeted customer interaction is to analyze the converting part (e.g. a sale) of the past online traffic. Smart attribution is hugely important to evaluate the cost effectiveness of marketing campaigns and to optimize the media spending and budget mix. But attribution is per definition only focusing in the converting traffic and purely retrospective. Information of the major traffic share from none or better 'not yet' converting visitors is very much neglected. There are multiple communication/ interaction levels in the sales process and different marketing actions have different strength to capture certain level customers and transfer them from level to level. From transforming site visitor data into customer journeys of users to predicting the future behavior of these users and their next steps with time scales - this case study based on Dutch Online travel company BookIt.com will demonstrate how you can learn much more from the non-converts and use these insights in the predictions.

Lloyds Banking Group has over 20 million Personal Current Account (PCA) customers and vast amounts of data available for use in credit decisions. In the past there has been a reliance on ‘Expert Judgement’ to build characteristics which are used in modelling. As both the quantity and quality of data grows it is imperative to research new modelling techniques that effectively use all data to ensure the Bank makes the best decisions for our customers. Currently Lloyds uses a traditional logistic regression scorecard for PCA overdraft application decisions. These scorecards use characteristics derived from a number of data sources, however they are largely based on expert judgement and do not use the full extent of the data to build models, as a result the predictive power of the data is not fully exploited by these models.As both the quantity and quality of data has grown there are a number of sources of data that are underused, but easily accessible for making PCA credit decisions. One example is the Transactional data, this contains information on monetary transaction on accounts such as transfers, over the counter and point of sales. This data contains thousands of variables over a hierarchy that is completely unsuitable to feed raw into the traditional statistical methods.Here we present two approaches using evolutionary algorithms to mine these previously unwieldy quantities of data to find which characteristics add most incremental power to the model. One based on Genetic Algorithms creates a library of characteristics based on the raw variables before selecting a subset for use in logistic regression models. The second method uses Genetic Programming to evolve the characteristics from the raw data in a method that can be used both to create new characteristics and as a stand-alone predictive model.The predictive power of both is compared to the traditional scorecard to show the improvement both in terms of discriminatory power and business benefit.

Successful Internet commerce sites have tens of thousands of unique daily visitors, creating large quantities of noisy traffic data. In the midst of this data are insights about customer behaviour and journeys that can help form commercial strategies, marketing decisions and data driven decision making, however deriving these insights can be a complex process. Sophisticated statistical predictive analytics techniques are required to indicate when there are subtle changes in the traffic data which can be acted upon. Using time series, textual analysis and statistical techniques, changes in customer profile traffic data can be isolated to give a daily or weekly detailed snapshot of the underlying customer behaviour prompting these changes.

The Belgian Social Security Institution is a federal agency that registers and monitors every active company in Belgium, and is responsible for the collection of employer and employee tax contributions. The taxes are levied on employer level, making this process highly sensitive to fraud. One of the main analytical techniques used to spread fraud through a network, are the so-called collective inference (CI) techniques. Here, one assumes that the label of a node in the network depends on the label of the neighboring nodes. Although CI is proven very useful, it can be easily mislead. A wrongful estimate of one node’s label (e.g., a true legitimate node is classified as fraudulent) might further impact the network and misclassify all other nodes in the neighborhood. In this presentation, we will discuss how (1) active inference – a subdomain of active learning for data analytics – is able to improve CI algorithms by selecting a set of uncertain nodes in the network to be labeled by inspectors, and (2) given that fraud networks are highly adaptive and evolve over time, we temporally integrate an inspector’s decision, knowing that this decision does not necessarily hold in the near future.