Mirko Böttcher

On exploiting the power of time in data mining

We introduce the new paradigm of Change Mining as data mining over a volatile, evolving world with the objective of understanding change. While there is much work on incremental mining and stream mining, both focussing on the adaptation of patterns to a changing data distribution, Change Mining concentrates on understanding the changes themselves. This includes detecting when change occurs in the population under observation, describing the change, predicting change and pro-acting towards it. We identify the main tasks of Change Mining and discuss to what extent they are already present in related research areas. We elaborate on research results that can contribute to these tasks, giving a brief overview of the current state of the art and identifying open areas and challenges for the new research area.

Relevance Feedback for Association Rules by Leveraging Concepts from Information Retrieval

The task of detecting those association rules which are interesting within the vast set of discovered ones still is a major research hallenge in data mining. Although several possible solutions have been proposed, they usually require a user to be aware what he knows, to have a rough idea what he is looking for, and to be able to specify this knowledge in advance. In this paper we compare the task of finding the most relevant rules with the task of finding the most relevant documents known from Information Retrieval. We propose a novel and flexible method of relevance feedback for association rules which leverages technologies from Information Retrieval, like document vectors, term frequencies and similarity calculations. By acquiring a user’s preferences our approach builds a repository of what he considers to be (non-)relevant. By calculating and aggregating the similarities of each unexamined rule with the rules in the repository we obtain a relevance score which better reflects the user’s notion of relevance with each feedback provided.

Predicting Future Decision Trees from Evolving Data

Recognizing and analyzing change is an important human virtue because it enables us to anticipate future scenarios and thus allows us to act pro-actively. One approach to understand change within a domain is to analyze how models and patterns evolve. Knowing how a model changes over time is suggesting to ask: Can we use this knowledge to learn a model in anticipation, such that it better reflects the near-future characteristics of an evolving domain? In this paper we provide an answer to this question by presenting an algorithm which predicts future decision trees based on a model of change. In particular, this algorithm encompasses a novel approach to change mining which is based on analyzing the changes of the decisions made during model learning. The proposed approach can also be applied to other types of classifiers and thus provides a basis for future research. We present our first experimental results which show that anticipated decision trees have the potential to outperform trees learned on the most recent data.

An Algorithm for Anticipating Future Decision Trees from Concept-Drifting Data

Concept-Drift is an important topic in practical data mining, since it is reality in most business applications. Whenever a mining model is used in an application it is already outdated since the world has changed since the model induction. The solution is to predict the drift of a model and derive a future model based on such a prediction. One way would be to simulate future data and derive a model from it, but this is typically not feasible. Instead we suggest to predict the values of the measures that drive model induction. In particular, we propose to predict the future values of attribute selection measures and class label distribution for the induction of decision trees. We give an example of how concept drift is reflected in the trend of these measures and that the resulting decision trees perform considerably better than the ones produced by existing approaches.

Relevance Feedback for Association Rules using Fuzzy Score Aggregation

We propose a novel and more flexible relevance feedback for association rules which is based on a fuzzy notion of relevance. Our approach transforms association rules into a vector-based representation using some inspiration from document vectors in information retrieval. These vectors are used as the basis for a relevance feedback approach which builds a knowledge base of rules previously rated as (un)interesting by a user. Given an association rule the vector representation is used to obtain a fuzzy score of how much this rule contradicts a rule in the knowledge base. This yields a set of relevance scores for each assessed rule which still need to be aggregated. Rather than relying on a certain aggregation measure we utilize OWA operators for score aggregation to gain a high degree of flexibility and understandability.

A framework for discovering and analyzing changing customer segments

Identifying customer segments and tracking their change over time is an important application for enterprises who need to understand what their customers expect from them. Customer segmentation is typically done by applying some form of cluster analysis. In this paper we present an alternative approach based on associaton rule mining and a notion of interestingness. Our approach allows us to detect arbitrary segments and analyse their temporal development. Our approach is assumption-free and pro-active and can be run continuously. Newly discovered segments or relevant changes will be reported automatically based on the application of an interestingness measure.

Matching Partitions over Time to Reliably Capture Local Clusters in Noisy Domains

When seeking for small clusters it is very intricate to distinguish between incidental agglomeration of noisy points and true local patterns. We present the PAMALOC algorithm that addresses this problem by exploiting temporal information which is contained in most business data sets. The algorithm enables the detection of local patterns in noisy data sets more reliable compared to the case when the temporal information is ignored. This is achieved by making use of the fact that noise does not reproduce its incidental structure but even small patterns do. In particular, we developed a method to track clusters over time based on an optimal match of data partitions between time periods.

Exploring Time Series of Patterns: Guided Drill-Down in Hierarchies Using Change Mining on Frequent Item Sets

In the past years pattern detection has gained in importance for many companies. As the volume of collected data increases so does typically the number of found patterns. To cope with this problem different interestingness measures for patterns have been proposed. Unfortunately, their usefulness turns out to be limited in practical applications. To address this problem, we propose a technique for a guided, visual exploration of patterns rather than presenting analysts with static ordered lists of patterns. Specifically, we focus on a method to guide drill-downs into hierarchical attributes, where we make use of change mining on frequent item sets for pattern discovery.