Direkt zum InhaltDirekt zur SucheDirekt zur Navigation
▼ Zielgruppen ▼

Humboldt-Universität zu Berlin - High Dimensional Nonstationary Time Series

Humboldt-Universität zu Berlin | Wirtschaftswissen­schaftliche Fakultät | High Dimensional Nonstationary Time Series | About us | Research projects | C3- Dynamic E-commerce analytics: analysis, support and automation of real-time decision processes

C3 - Dynamic E-commerce analytics: analysis, support and automation of real-time decision processes

The advent of electronic commerce (e-commerce) has disrupted corporate practices in many ways. New concepts and technologies in the e-commerce environment continue to push the boundaries between markets, industries, and sectors. Corporations internalize clients who participate actively in R&D processes to envision new products. At the same time, many consumers become micro-entrepreneurs and offer highly specialized services or products to a global audience. The economic and societal ramifications of e-commerce are both ubiquitous and pervasive.

A major characteristic of e-commerce applications is tractability. Every process, decision, or transaction generates a digital footprint, which is stored in large databases and available for subsequent analysis. On the one hand, the availability of such data creates substantial interest in empirical forms of decision support among managers, which technology trends such as big data exemplify. On the other hand, easy availability of rich, observational data across long time intervals makes e-commerce a popular environment to develop novel theories that explain real-world phenomena in decision analysis and novel methodologies that distil relevant insights from large amounts of heterogeneous data.

The research projects proposed in C3 embody this two-sided spirit. On the one hand, they share an analytic core, which aims at developing novel machine learning methods. On the other hand, actual decision problems in e-commerce such as real-time decisions related to soliciting e-coupons or catering banner advertisements to individual surfers set algorithmic requirements and thus inspire revision and extension of existing learning algorithms. In both cases, novel methodology will be evaluated against the background of business applications and real-time decision problems in the broad field of e-commerce.

Two overarching research goals encompass the research projects in C3. The first goal is to improve the treatment and coverage of dynamic effects in supervised learning frameworks. The second objective is to devise decision models that combine predictive accuracy with explanatory insight in the form of causal links between predictors and predictions. Time-dependence and the desire to obtain causal insights are commonly encountered in the focal application context.

The research projects in C3 are expected to substantially improve the quality of model-based decisions in e-commerce. The applicability of methodologic contributions may, however, generalize far beyond this application scope. To verify this and exploit potential for cross-pollination, a close collaboration with B3 is envisioned. Dynamic panel data modelling, the focus of B3, is a field that shares several similarities with the tasks studied here. Purchase data from online shops, to which C3 has access from corporate collaborators, exemplifies this. Such data can be considered a panel. Furthermore, e-shop data exhibits large volume, high dimensionality, and heterogeneity across observational units (i.e., shoppers). Last, the objective to infer causal links from data is another connection between B3 and C3. Exploiting ample potential for collaboration will benefit both vertebrae. It allows researches in B3 and C3 to stay abreast of methodological advancements in the realms of causal machine learning and dynamic panel data modelling, respectively; it gives both parties access to highly challenging benchmark methods in the form of the other party’s algorithms; and it offers several opportunities to share data for assessing novel methodology.

 

Coordination

Stefan Lessmann: His research focuses on managerial decision analysis and support using empirical methods for structured and unstructured data processing (supervised machine learning, text mining, sentiment analysis). Prevailing areas of application include marketing analytics and credit risk modelling.

 

Exemplary PhD-Theses

  1. Tensor learning for ensemble classifiers

Supervised classification models are widely used to support decisions in e-commerce applications. Application examples are manifold and include a classification of shoppers into buyers and non-buyers, see Van den Poel & Buckinx (2005) and more generally conversion modelling, see King et al. (2015) to inform advertising decisions or a re-identification of shoppers using contextual information, see Panniello et al. (2016). Tree-based ensemble models such as random forest or gradient boosting are especially popular approaches and have demonstrated their ability to generate highly accurate forecasts of consumer behaviour in several forecasting competitions, see e.g. Lessmann & Voß (2010), Verbeke et al. (2012). However, these methods and supervised machine learning methods in general are ill prepared to model dynamic patterns. Their internal operation require data in the form of a matrix (e.g., with rows describing cases and columns describing variables). Sequential data such as a customer’s purchases are difficult to accommodate in such data structure. The prevailing approach is to first aggregate sequential data and to then incorporate the resulting aggregates as additional variables (no. of purchases in total, no. of purchases in the last month, etc.). Recently, advancements related to learning from tensors have been made. Using a tensor framework, sequential data can easily enter a model as an additional dimension. However, corresponding advancement focus on deep learning methods (e.g., Google’s TensorFlow) or support vector machines, see e.g. Chen et al. (2016). The aim of this thesis is to devise an ensemble framework for tensor learning. The growing of an individual classification tree provides a natural starting point for the thesis. Subsequently, extensions to develop parallel and sequential tensor ensembles, which mimic the popular random forest and gradient boosting approach, respectively, will be developed. In the last stage, the thesis will explore the merit of forming heterogeneous ensembles that integrate and combine multiple (tensor) learning methods.

 

2.    Causal inference, machine learning, and uplift models

Algorithms to generate operationally accurate forecasts from empirical data are well advanced and enjoy much popularity in industry. Applications in e-commerce include the distribution of e-promotions, see Nassiri-Mofakham et al. (2009) and, more generally, online advertisement, see Graepel et al. (2010). The thesis addresses two limitations of corresponding methods and proposed suitable remedies. First, powerful prediction methods are typically opaque and do not disclose the predictor-response-relationship, which they infer from data, see Shmueli & Koppius (2011). Second, the relationships inferred by such methods are correlational but not causal. Most (online) marketing applications require both, interpretability and causality. The former enables a marketer to understand which factors govern (predicted) customer behaviour and in which way. The latter is important since marketers want to establish that it is the marketing message (i.e., treatment) that triggers a behaviour (a purchase). Due to these requirements, econometric methods dominate in marketing literature, while data-driven prediction methods have received little attention. Uplift models are a noteworthy exception and comprise of different approaches to introduce a causal element into ordinary prediction models, see e.g. Sołtys et al. (2015). In particular, an uplift model measures the marginal causal effect of a marketing action. Previous work on uplift models provide a starting point for the thesis. Consolidating and benchmarking existing approaches is a first research goal. Afterwards, the thesis will survey the field of causal machine learning to identify additional opportunities to develop uplift model and develop corresponding prototypes. An important design objective to be respected in this step concerns model interpretability. To that end, the thesis will draw upon the rule extraction literature and propose an algorithm to summarize complex uplift models in a concise and understandable set of rules.

 

3.    Dynamic targeting automation

The thesis focuses on decision problems in online marketing and conversion modelling. The subject of analysis is the session of an individual visitor at some website (e.g., an e-shop). The decision problem of the site owner is to select out of a set of multiple marketing stimuli (displaying an info message, offering a percentage rebate, offering a time-bound absolute rebate, etc.) and time points (e.g., visits of individual web pages on the site) an optimal policy so as to reach some objective, broadly referred to as conversion (e.g., newsletter sign-up, site registration, purchase, etc.). The complexity of the problem arises from the multiple dimensions of the decision; namely whether to perform an action (e.g., show coupon or not), when to take the action (e.g., after how many clicks), and how to configure the action, given it will be taken (e.g., discount amount of the coupon). Previous work in machine learning commonly abstracts the problem. For example, a binary classification of visitors into a treatment group and a no-treatment group is a common simplification; see e.g. Van den Poel & Buckinx (2005). More advanced econometric approaches have also been proposed, but suffer limitations in terms of scalability, see e.g. Ding et al. (2015). The goal of the thesis is to develop a scalable learning framework which is able to cover the full range of options and constraints. One idea to address the different dimensions of the decision problem is to exploit the large-scale character of e-commerce applications. In particular, identifying statistical twins (i.e., website visitors with similar characteristics) facilitates exploring alternative marketing policies in parallel. Based on the feedback obtained from such experiments, policies can be adapted and progressed in real-time. Previous work in reinforcement learning and Bayesian learning, see e.g. Graepel et al. (2010) will provide a useful starting point for the thesis.

 

References

  • Chen Z-Y, Fan Z-P, Sun M (2016) A multi-kernel support tensor machine for classification with multitype multiway data and an application to cross-selling recommendations. European Journal of Operational Research, 255, 110-120.
  • Ding A W, Li S, Chatterjee P (2015) Learning user real-time intent for optimal dynamic web page transformation. Information Systems Research, 26, 339-359.
  • Graepel T, Candela J Q, Borchert T, Herbrich R (2010) Web-Scale Bayesian Click-Through Rate Prediction for Sponsored Search Advertising in Microsoft's Bing Search Engine. In Fürnkranz J & Joachims T (Eds.), Proc. of the 27th Intern. Conf. on Machine Learning (pp. 13-20). Haifa, Israel: Omnipress.
  • King M A, Abrahams A S, Ragsdale C T (2015) Ensemble learning methods for pay-per-click campaign management. Expert Systems with Applications, 42, 4818-4829.
  • Lessmann S, Voß S (2010) Customer-centric decision support: A benchmarking study of novel versus established classification models. Business & Information Systems Engineering, 2, 79-93.
  • Nassiri-Mofakham F, Nematbakhsh M A, Baraani-Dastjerdi A, Ghasem-Aghaee N (2009) Electronic promotion to new customers using mkNN learning. Information Sciences, 179, 248-266.
  • Panniello U, Hill S, Gorgoglione M (2016) Using context for online customer re-identification. Expert Systems with Applications, 64, 500-511.
  • Shmueli G, Koppius O R (2011) Predictive analytics in information systems research. MIS Quarterly, 35, 553-572.
  • Sołtys M, Jaroszewicz S, Rzepakowski P (2015) Ensemble methods for uplift modeling. Data Mining and Knowledge Discovery, 29, 1531-1559.
  • Van den Poel D, Buckinx W (2005) Predicting online-purchasing behaviour. European Journal of Operational Research, 166, 557-575.
  • Verbeke W, Dejaeger K, Martens D, Hur J, Baesens B (2012) New insights into churn prediction in the telecommunication sector: A profit driven data mining approach. European Journal of Operational Research, 218, 211-229.