Stephan Spiegel

Pattern recognition and classification for multivariate time series

Nowadays we are faced with fast growing and permanently evolving data, including social networks and sensor data recorded from smart phones or vehicles. Temporally evolving data brings a lot of new challenges to the data mining and machine learning community. This paper is concerned with the recognition of recurring patterns within multivariate time series, which capture the evolution of multiple parameters over a certain period of time. Our approach first separates a time series into segments that can be considered as situations, and then clusters the recognized segments into groups of similar context. The time series segmentation is established in a bottom-up manner according the correlation of the individual signals. Recognized segments are grouped in terms of statistical features using agglomerative hierarchical clustering. The proposed approach is evaluated on the basis of real-life sensor data from different vehicles recorded during car drives. According to our evaluation it is feasible to recognize recurring patterns in time series by means of bottom-up segmentation and hierarchical clustering.

Hydra: a hybrid recommender system [cross-linked rating and content information]

This paper discusses the combination of collaborative and content-based filtering in the context of web-based recommender systems. In particular, we link the well-known MovieLens rating data with supplementary IMDB content information. The resulting network of user-item relations and associated content features is converted into a unified mathematical model, which is applicable to our underlying neighbor-based prediction algorithm. By means of various experiments, we demonstrate the influence of supplementary user as well as item features on the prediction accuracy of Hydra, our proposed hybrid recommender. In order to decrease system runtime and to reveal latent user and item relations, we factorize our hybrid model via singular value decomposition (SVD).

Link prediction on evolving data using tensor factorization

Within the last few years a lot of research has been done on large social and information networks. One of the principal challenges concerning complex networks is link prediction. Most link prediction algorithms are based on the underlying network structure in terms of traditional graph theory. In order to design efficient algorithms for large scale networks, researchers increasingly adapt methods from advanced matrix and tensor computations. This paper proposes a novel approach of link prediction for complex networks by means of multi-way tensors. In addition to structural data we furthermore consider temporal evolution of a network. Our approach applies the canonical Parafac decomposition to reduce tensor dimensionality and to retrieve latent trends. For the development and evaluation of our proposed link prediction algorithm we employed various popular datasets of online social networks like Facebook and Wikipedia. Our results show significant improvements for evolutionary networks in terms of prediction accuracy measured through mean average precision.

A Recurrence Plot-Based Distance Measure

Given a set of time series, our goal is to identify prototypes that cover the maximum possible amount of occurring subsequences regardless of their order. This scenario appears in the context of the automotive industry, where the goal is to determine operational profiles that comprise frequently recurring driving behavior patterns. This problem can be solved by clustering, however, standard distance measures such as the dynamic time warping distance might not be suitable for this task, because they aim at capturing the cost of aligning two time series rather than rewarding pairwise recurring patterns. In this contribution, we propose a novel time series distance measure, based on the notion of recurrence plots, which enables us to determine the (dis)similarity of multivariate time series that contain segments of similar trajectories at arbitrary positions. We use recurrence quantification analysis to measure the structures observed in recurrence plots and to investigate dynamical properties, such as determinism, which reflect the pairwise (dis)similarity of time series. In experiments on real-life test drives from Volkswagen, we demonstrate that clustering multivariate time series using the proposed recurrence plot-based distance measure results in prototypical test drives that cover significantly more recurring patterns than using the same clustering algorithm with dynamic time warping distance.

Energy disaggregation meets heating control

Heating control is of particular importance, since heating accounts for the biggest amount of total residential energy consumption. Smart heating strategies allow to reduce such energy consumption by automatically turning off the heating when the occupants are sleeping or away from home. The present context or occupancy state of a household can be deduced from the appliances that are currently in use. In this study we investigate energy disaggregation techniques to infer appliance states from an aggregated energy signal measured by a smart meter. Since most household devices have predictable energy consumption, we propose to use the changes in aggregated energy consumption as features for the appliance/occupancy state classification task. We evaluate our approach on real-life energy consumption data from several households, compare the classification accuracy of various machine learning techniques, and explain how to use the inferred appliance states to optimize heating schedules.

Pattern recognition in multivariate time series: dissertation proposal

Nowadays computer scientists are faced with fast growing and permanently evolving data, which are represented as observations made sequentially in time. A common problem in the data mining community is the recognition of recurring patterns within temporal databases or streaming data. This dissertation proposal aims at developing and investigating efficient methods for the recognition of contextual patterns in multivariate time series in different application domains based on machine learning techniques. To this end, we propose a generic three-step approach that involves (1) feature extraction to build robust learning models based on significant time series characteristics, (2) segmentation to identify internally homogeneous time intervals and change points, as well as (3) clustering and/or classification to group the time series (segments) into the sub-population to which they belong to. To support our proposed approach, we present and discuss first experiments on real-life vehicular data. Furthermore we describe a number of applications, where pattern recognition in multivariate time series is practical or rather necessary.

Approximate Recurrence Quantification Analysis (aRQA) in Code of Best Practice

Recurrence quantification analysis (RQA) is a well-known tool for studying nonlinear behavior of dynamical systems, e.g. for finding transitions in climate data or classifying reading abilities. But the construction of a recurrence plot and the subsequent quantification of its small and large scale structures is computational demanding, especially for long time series or data streams with high sample rate. One way to reduce the time and space complexity of RQA are approximations, which are sufficient for many data analysis tasks, although they do not guarantee exact solutions. In earlier work, we proposed how to approximate diagonal line based RQA measures and showed how these approximations perform in finding transitions for difference equations. The present work aims at extending these approximations to vertical line based RQA measures and investigating the runtime/accuracy of our approximate RQA measures on real-life climate data. Our empirical evaluation shows that the proposed approximate RQA measures achieve tremendous speedups without losing much of the accuracy.

Discovery of Driving Behavior Patterns

Given a set of time series, our goal is to identify prototypes that cover the maximum possible amount of occurring subsequences regardless of their order. This scenario appears in the context of the automotive industry, where the objective is to determine operational profiles that comprise frequently recurring driving behavior patterns. This problem can be solved by clustering, however, standard distance measures such as the dynamic time warping distance might not be suitable for this task, because they aim at capturing the cost of aligning two time series rather than rewarding pairwise occurring patterns. In this work, we propose a novel time series distance measure, based on the theoretical foundation of recurrence plots, which enables us to determine the (dis)similarity of multivariate time series that contain segments of similar trajectories at arbitrary positions. We use recurrence quantification analysis to measure the structures observed in recurrence plots and to investigate dynamical properties, such as determinism, which reflect the pairwise (dis)similarity of time series. In experiments on real-life test drives from Volkswagen, we demonstrate that clustering multivariate time series using the proposed recurrence plot-based distance measure results in prototypical test drives that cover significantly more recurring patterns than using the same clustering algorithm with dynamic time warping distance.

BestTime: finding representatives in time series datasets

Given a set of time series, we aim at finding representatives which best comprehend the recurring temporal patterns contained in the data. We demonstrate BestTime, a Matlab application that uses recurrence quantification analysis to find time series representatives.

Optimization of In-House Energy Demand

Heating control is of particular importance, since heating accounts for the biggest amount of total residential energy consumption. Smart heating strategies allow to reduce such energy consumption by automatically turning off the heating when the occupants are sleeping or away from home. The present context or occupancy state of a household can be deduced from the appliances that are currently in use. In this chapter, we investigate energy disaggregation techniques to infer appliance states from an aggregated energy signal measured by a smart meter. Since most household devices have predictable energy consumption, we propose to use the changes in aggregated energy consumption as features for the appliance/occupancy state classification task. We evaluate our approach on real-life energy consumption data from several households, compare the classification accuracy of various machine learning techniques, and explain how to use the inferred appliance states to optimize heating schedules.