We develop new methods, strategies and algorithms for individualised marketing, customer retention, optimised communication with users, personalised pricing and personalised recommendations or to maximise the probability of purchase of a product or other actions of the users. We exploit users’ behavioural measurements in addition to their more standard characteristics and external data (including competitors’ activity, market indicators, financial information, and geographic information). We exploit network topologies, informative missigness and temporal relations. A key point is to identify the actionable causes of customer behaviour.
Stochastic customer growth dynamics
Understanding how networks of customers grow in time and topology is important. The Vipps transaction data may be viewed as a graph with users corresponding to the nodes and the financial transactions between the users defining the edges. With an advanced statistical model we have analysed the growth of this graph. Our experiments show that the intrinsic quality of the nodes plays an important part in the evolution of the network. This insight may be used to identify influential nodes for viral marketing. The approach will be published.
Predicting customer behavior from time series
Multiple time series, related to individual customers have been used to predict, as early as possible, whether the customer will pay back his loan or not. We have developed a new solution based on deep learning of time series, which gives excellent results on the DNB cases it has been tested on. The method will be transferred to DNB and has been accepted for publication in the journal Expert systems and Applications.
Bayesian methodology for recommender systems
In many important situations, we wish to recommend to each individual user or customer, the items she might be most interested in, or the ones she would benefit most. Starting from data where user either rate or compare items or click on them, we want to predict user’s preferences on other items. In some applications, we are also interested in estimating the shared consensus preference of a homogeneous group of people. We have invented a new Bayesian approach based on extensions of the Mallows model, which allow making individualized recommendations, equipped with a level of uncertainty. We are working on public data and NRK data and plan to analyse other industrial cases, including an AB testing experiment. The methodology has been published on the Journal of Machine Learning Research.
Stochastic models for early prediction of viral customer behavior on networks
Can we predict if a new service or product will be go viral in a market, or if it will be a flop? We look to markets where adopters are organised as nodes in a network, where links represent contacts which allow a user to "convince" a neighbor to adopt. We worked on a case from Telenor. We are able to predict, already after a few weeks, if the adoption process will spread virally (or not) on the network and how the adoptions will happen in the future. Our simulation based approach is generic and looks very promising.
CLUSTERING OF CLICK STREAM DATA
Web stream data are routinely collected to study how users browse the web or use a service. The ability to identify user behaviour patterns from such data may be very valuable for different businesses. It may help to produce better marketing strategies and a better user experience. We use model-based clustering to segment users based on web clickstream data from Skatteetaten and Gjensidige. Model-based clustering assumes that users’ behaviours are generated by a set of probabilistic models and each model corresponds to a cluster.