Modeling Evolving User Behavior via Sequential Clustering

Abstract

In this paper we address the problem of modeling the evolution of clusters over time by applying sequential clustering. We propose a sequential partitioning algorithm that can be applied for grouping distinct snapshots of streaming data so that a clustering model is built on each data snapshot. The algorithm is initialized by a clustering solution built on available historical data. Then a new clustering solution is generated on each data snapshot by applying a partitioning algorithm seeded with the centroids of the clustering model obtained at the previous time interval. At each step the algorithm also conducts model adapting operations in order to reflect the evolution in the clustering structure. In that way, it enables to deal with both incremental and dynamic aspects of modeling evolving behavior problems. In addition, the proposed approach is able to trace back evolution through the detection of clusters' transitions, such as splits and merges. We have illustrated and initially evaluated our ideas on household electricity consumption data. The results have shown that the proposed sequential clustering algorithm is robust to modeling evolving behavior by being enable to mine changes and update the model, respectively.

SECLEDS: Sequence Clustering in Evolving Data Streams via Multiple Medoids and Medoid Voting

“…In recent years, sequential data has increasingly become popular because of the powerful insights that it provides regarding behavior analytics [5], e.g., for attacker strategy profiling [21], fraud detection [13], human activity recognition [7]. Clustering sequences in an offline setting is challenging in itself because sequences are often out-of-sync, requiring expensive alignment-based distance measures, which are often not supported by many clustering algorithms.…”

EvolveCluster: an evolutionary clustering algorithm for streaming data

“…In this study, we propose a novel evolutionary clustering algorithm capable of modeling data streams containing evolving data, entitled EvolveCluster, a continuation of our previous work (Boeva and Nordahl 2019). Instead of processing elements individually, we collect data over a defined period (creating segments) to trace how the data evolves.…”