Michael Hahsler

Implications of Probabilistic Data Modeling for Mining Association Rules

Mining association rules is an important technique for discovering meaningful patterns in transaction databases. In the current literature, the properties of algorithms to mine association rules are discussed in great detail. We present a simple probabilistic framework for transaction data which can be used to simulate transaction data when no associations are present. We use such data and a real-world grocery database to explore the behavior of confidence and lift, two popular interest measures used for rule mining. The results show that confidence is systematically influenced by the frequency of the items in the left-hand-side of rules and that lift performs poorly to filter random noise in transaction data. The probabilistic data modeling approach presented in this paper not only is a valuable framework to analyze interest measures but also provides a starting point for further research to develop new interest measures which are based on statistical tests and geared towards the specific properties of transaction data.

SOStream: self organizing density-based clustering over data stream

In this paper we propose a data stream clustering algorithm, called Self Organizing density based clustering over data Stream (SOStream). This algorithm has several novel features. Instead of using a fixed, user defined similarity threshold or a static grid, SOStream detects structure within fast evolving data streams by automatically adapting the threshold for density-based clustering. It also employs a novel cluster updating strategy which is inspired by competitive learning techniques developed for Self Organizing Maps (SOMs). In addition, SOStream has built-in online functionality to support advanced stream clustering operations including merging and fading. This makes SOStream completely online with no separate offline components. Experiments performed on KDD Cup99 and artificial datasets indicate that SOStream is an effective and superior algorithm in creating clusters of higher purity while having lower space and time requirements compared to previous stream clustering algorithms.

A Model-Based Frequency Constraint for Mining Associations from Transaction Data

Mining frequent itemsets is a popular method for finding associated items in databases. For this method, support, the co-occurrence frequency of the items which form an association, is used as the primary indicator of the associationss significance. A single, user-specified support threshold is used to decided if associations should be further investigated. Support has some known problems with rare items, favors shorter itemsets and sometimes produces misleading associations.In this paper we develop a novel model-based frequency constraint as an alternative to a single, user-specified minimum support. The constraint utilizes knowledge of the process generating transaction data by applying a simple stochastic mixture model (the NB model) which allows for transaction datas typically highly skewed item frequency distribution. A user-specified precision threshold is used together with the model to find local frequency thresholds for groups of itemsets. Based on the constraint we develop the notion of NB-frequent itemsets and adapt a mining algorithm to find all NB-frequent itemsets in a database. In experiments with publicly available transaction databases we show that the new constraint provides improvements over a single minimum support threshold and that the precision threshold is more robust and easier to set and interpret by the user.

Comparing two recommender algorithms with the help of recommendations by peers

Since more and more Web sites, especially sites of retailers, offer automatic recommendation services using Web usage mining, evaluation of recommender algorithms has become increasingly important. In this paper we present a framework for the evaluation of different aspects of recommender systems based on the process of discovering knowledge in databases introduced by Fayyad et al. and we summarize research already done in this area. One aspect identified in the presented evaluation framework is widely neglected when dealing with recommender algorithms. This aspect is to evaluate how useful patterns extracted by recommender algorithms are to support the social process of recommending products to others, a process normally driven by recommendations by peers or experts. To fill this gap for recommender algorithms based on frequent itemsets extracted from usage data we evaluate the usefulness of two algorithms. The first recommender algorithm uses association rules, and the other algorithm is based on the repeat-buying theory known from marketing research. of usage data from an educational Internet information broker and compare useful recommendations identified by users from the target group of the broker (peers) with the recommendations produced by the algorithms. The results of the evaluation presented in this paper suggest that frequent itemsets from usage histories match the concept of useful recommendations expressed by peers with satisfactory accuracy (higher than 70%) and precision (between 60% and 90%). Also the evaluation suggests that both algorithms studied in the paper perform similar on real-world data if they are tuned properly.

A Customer Purchase Incidence Model Applied to Recommender Services

In this contribution we transfer a customer purchase incidence model for consumer products which is based on Ehrenbergs repeat-buying theory to Web-based information products. Ehrenbergs repeat-buying theory successfully describes regularities on a large number of consumer product markets. We show that these regularities exist in electronic markets for information goods, too, and that purchase incidence models provide a well founded theoretical base for recommender and alert services.The article consists of two parts. In the first part Ehrenbergs repeat-buying theory and its assumptions are reviewed and adapted for web-based information markets. Second, we present the empirical validation of the model based on data collected from the information market of the Virtual University of the Vienna University of Economics and Business Administration from September 1999 to May 2001.

Clustering Data Streams Based on Shared Density between Micro-Clusters

As more and more applications produce streaming data, clustering data streams has become an important technique for data and knowledge engineering. A typical approach is to summarize the data stream in real-time with an online process into a large number of so called micro-clusters. Micro-clusters represent local density estimates by aggregating the information of many data points in a defined area. On demand, a (modified) conventional clustering algorithm is used in a second offline step to recluster the micro-clusters into larger final clusters. For reclustering, the centers of the micro-clusters are used as pseudo points with the density estimates used as their weights. However, information about density in the area between micro-clusters is not preserved in the online process and reclustering is based on possibly inaccurate assumptions about the distribution of data within and between micro-clusters (e.g., uniform or Gaussian). This paper describes DBSTREAM, the first micro-cluster-based online clustering component that explicitly captures the density between micro-clusters via a shared density graph. The density information in this graph is then exploited for reclustering based on actual density between adjacent micro-clusters. We discuss the space and time complexity of maintaining the shared density graph. Experiments on a wide range of synthetic and real data sets highlight that using shared density improves clustering quality over other popular data stream clustering methods which require the creation of a larger number of smaller micro-clusters to achieve comparable results.

Selective association rule generation

Mining association rules is a popular and well researched method for discovering interesting relations between variables in large databases. A practical problem is that at medium to low support values often a large number of frequent itemsets and an even larger number of association rules are found in a database. A widely used approach is to gradually increase minimum support and minimum confidence or to filter the found rules using increasingly strict constraints on additional measures of interestingness until the set of rules found is reduced to a manageable size. In this paper we describe a different approach which is based on the idea to first define a set of “interesting” itemsets (e.g., by a mixture of mining and expert knowledge) and then, in a second step to selectively generate rules for only these itemsets. The main advantage of this approach over increasing thresholds or filtering rules is that the number of rules found is significantly reduced while at the same time it is not necessary to increase the support and confidence thresholds which might lead to missing important information in the database.

Building on the Arules Infrastructure for Analyzing Transaction Data with R

The free and extensible statistical computing environment R with its enormous number of extension packages already provides many state-of-the-art techniques for data analysis. Support for association rule mining, a popular exploratory method which can be used, among other purposes, for uncovering cross-selling opportunities in market baskets, has become available recently with the R extension package arules. After a brief introduction to transaction data and association rules, we present the formal framework implemented in arules and demonstrate how clustering and association rule mining can be applied together using a market basket data set from a typical retailer. This paper shows that implementing a basic infrastructure with formal classes in R provides an extensible basis which can very efficiently be employed for developing new applications (such as clustering transactions) in addition to association rule mining.

Recommendations for Virtual Universities from Observed User Behavior

Recently recommender systems started to gain ground in commercial Web-applications. For example, the online-bookseller amazon.com recommends his customers books similar to the ones they bought using the analysis of observed purchase behavior of consumers.

myVU: A Next Generation Recommender System Based on Observed Consumer Behavior and Interactive Evolutionary Algorithms

myVU is a next generation recommender system based on observed consumer behavior and interactive evolutionary algorithms implementing customer relationship management and one-to-one marketing in the educational and scientific broker system of a virtual university. myVU provides a personalized, adaptive WWW-based user interface for all members of a virtual university and it delivers routine recommendations for frequently used scientific and educational Web-sites.