High-performance methods for impressive results in real time

Tackle any challenge equipped with these innovative methods in intelligent data analysis.

prudsys AG is the technology leader in the fields of data mining and real-time analytics. Right from the start our in-house research and development department has been working on innovative methods for analyzing high-dimensional data. Current research focal points include machine learning, dynamic pricing and real-time scoring.

We own numerous patents and are continually expanding our leading technological position by cooperating closely with renowned scientific institutions. We belong to the DMG and OMG standards committees, taking an active role in shaping the guiding principles of intelligent data analytics.

Reinforcement learning

Machine learning in real time.

Reinforcement learning is a variant of machine learning and forms the central framework of the prudsys Realtime Decisioning Engine (prudsys RDE for short). Conventional recommendation engines, especially those based on shopping basket analysis and collaborative filtering, are based on the assumption that the content which should be recommended to users is the content that users are most likely to choose. Although this principle is empirically sound, it should still be critically questioned: Why recommend something to a user that he probably would have chosen anyway? The challenge when it comes to the topic of recommendations is to change from a static to a control theoretical point of view, to think “cybernetically” to a certain extent.

Reinforcement learning is a central framework based essentially on methods of dynamic programming and one that primarily investigates the aspect of real-time learning. Along the way, reinforcement learning always considers the chains of all subsequent transactions, working not only in a target-oriented manner but also sustainably over the long term. The challenge when using reinforcement learning for recommendations is to let the process converge easily, even for extremely sparse transaction data. To do this, robust approximation architectures are used and hierarchical methods play a key role. prudsys AG has been dealing with this topic since 2004, achieving outstanding results.

Benefits of reinforcement learning

Feedback incorporated into learning

Maximization of a preset parameter

Optimization over the entire sequence of future user actions

Interplay between application and exploration

Uniform operator-based theory

Seamless combination of offline and online learning

Real-time scoring

Sparse grids for classification and regression.

The focus in the area of real-time scoring is on the development of adaptive sparse grid techniques. The “sparse grids” technique is the first, universal multivariate scoring procedure for classification and regression problems which is linearly scalable to the number of data sets and which can therefore be used on extremely large amounts of data. The basic idea is to solve classification and regression problems using their operator equations (usually differential equations), where the label space is discretized. This formulation has been used for decades, particularly in the form of finite element analysis, to solve physical problems but has failed in the data mining field because of the complexity of the calculations needed to deal with exponentially increasing amounts of data.

Sparse grid methods represent multidimensional wavelets, which, through the elimination of partial grids, can penetrate into the high dimensions for the first time without losing the considerable advantages of wavelets (approximation, spectral representation, orthogonality). prudsys AG developed sparse grid technology for scoring applications in cooperation with their partners from the University of Bonn and the Fraunhofer Institute for Algorithms and Scientific Computing. Current research focuses on the further increase of the dimension number, the transfer to other operator formulations and different real-time aspects.

Benefits of real-time scoring

Solves classification and regression problems regardless of the number of data sets

Inherent incremental approach allows direct use for real-time learning

Interpretable representation of the regression function

Spectral representation allows for direct analysis, compression and smoothing of the regression function

Expansion to a wide variety of operator formulations, integration of a-priori knowledge

Combination of offline and online learning

Dynamic Pricing

Fully automatic calculation of price elasticity.

Methods for automatic price optimization are becoming dramatically more important. They are among the most innovative of analytical methods. prudsys AG is a pioneer in the field of dynamic pricing in retail and has developed a comprehensive modular system of dynamic price optimization methods. The methods are based primarily on regression analysis for estimating price elasticity (real-time scoring) combined with control theory approaches. Methods for dynamic price optimization deliver optimized prices in real time. The reinforcement learning framework plays a role here. Existing algorithms for automatic price optimization have the drawback that they vary price too little and thus do not provide sufficient empirical data on which to base forecasts.

Strategically and continuously changing prices systematically fills the gaps in the data, enabling a precise estimate of price elasticity. Dynamic pricing can be used in a wide variety of scenarios. These scenarios include dealing with competitors’ prices, buying and selling, auctions, bundling, individual discounts and couponing. For this reason, the framework for dynamic price optimization is very wide-ranging and enables the definition of a variety of optimization parameters such as turnover, sales and earnings; strategies such as high price/low price, demand/competition oriented, anonymized/personalized; constraints such as price limits, time frames, variance, degradability. This results in a number of current research topics, including aspects of the game theory.

Benefits of dynamic price optimization

Optimization of different target parameters, strategies and complex constraints

Optimization using continuous transaction chains

Immediate adaptation to changes in customer purchase patterns

Automatic price setting and price tests

Simulation function for testing different price strategies

Combination of offline and online learning

Your contact

prudsys AG, Forschung, Chemnitz

André Müller

Head of Data Science
Phone: +49 371 27093 40
Email: moc.sysdurp@relleuma