Employing chunk size adaptation to overcome concept drift

2013 IEEE Second International Conference on Image Information Processing (ICIIP-2013), 2013

Data Streams are data instances which arrive at a very rapid rate with varying concepts. Many online ensembles of classifiers were developed which handled the drifting concepts and were proved to be better than a single classifier system. In our work, we will discuss our new approach, Early Dynamic Weighted Majority and will empirically prove it to be better than the existing online ensemble approaches. Empirical results would prove that all these online approaches can be quite competitive, and show good accuracy and speed in handling and identifying drifts in data.

Computers, materials & continua, 2023

Every application in a smart city environment like the smart grid, health monitoring, security, and surveillance generates non-stationary data streams. Due to such nature, the statistical properties of data changes over time, leading to class imbalance and concept drift issues. Both these issues cause model performance degradation. Most of the current work has been focused on developing an ensemble strategy by training a new classifier on the latest data to resolve the issue. These techniques suffer while training the new classifier if the data is imbalanced. Also, the class imbalance ratio may change greatly from one input stream to another, making the problem more complex. The existing solutions proposed for addressing the combined issue of class imbalance and concept drift are lacking in understating of correlation of one problem with the other. This work studies the association between concept drift and class imbalance ratio and then demonstrates how changes in class imbalance ratio along with concept drift affect the classifier's performance. We analyzed the effect of both the issues on minority and majority classes individually. To do this, we conducted experiments on benchmark datasets using state-of-the-art classifiers especially designed for data stream classification. Precision, recall, F1 score, and geometric mean were used to measure the performance. Our findings show that when both class imbalance and concept drift problems occur together the performance can decrease up to 15%. Our results also show that the increase in the imbalance ratio can cause a 10% to 15% decrease in the precision scores of both minority and majority classes. The study findings may help in designing intelligent and adaptive solutions that can cope with the challenges of non-stationary data streams like concept drift and class imbalance.

Knowledge and Information Systems, 2011

Mining data streams has become an important and challenging task for a wide range of applications. In these scenarios, data tend to arrive in multiple, rapid and time-varying streams, thus constraining data mining algorithms to look at data only once. Maintaining an accurate model, e.g. a classifier, while the stream goes by requires a smart way of keeping track of the data already passed away. Such a synthetic structure has to serve two purposes: distilling the most of information out of past data and allowing a fast reaction to concept drifting, i.e. to the change of the data trend that necessarily affects the model. The paper outlines novel data structures and algorithms to tackle the above problem, when the model mined out of the data is a classifier. The introduced model and the overall ensemble architecture are presented in details, even considering how the approach can be extended for treating numerical attributes. A large part of the paper discusses the experiments and the comparisons with several existing systems. The comparisons show that the performance of our system in general, and in particular with respect to the reaction to concept drifting, is at the top level. Keywords Mining data streams • Classification • Concept drifting • Ensemble classifier • Data aggregation Definition 2.3 (Training set distribution) Given a feature vector x and a class label y, the distribution of a data stream (x i , y i) can be defined as P(x, y) = P(x) P(y|x) (1) where P(x) is the probability of generating feature vector x, while P(y|x) is the probability to assign class value y to x. As stated by Gao et al. in [30], and with reference to Eq. 1, a data stream distribution evolves from time to time in 4 different ways: No change. There is no change in the data. The model continues to be updated. Feature change. Probability P(x) changes. Some previously infrequent feature vectors become more frequent, or vice versa. The expected error depends on the specific combination between P(x) and P(y|x). The error can be higher, lower or the same. A partial model reconstruction or a retraining using new P(x) can improve accuracy. Conditional change. The conditional probability P(y|x) changes. Some feature vectors previously classified as y are now classified differently. It is likely a completely new model extraction. Dual change. Both P(x) and P(y|x) change. Since P(y|x) has evolved, it is necessary to train a completely new model. Generally, without any assumption about data distribution, in none of the aforementioned four cases, the expected error can be used as an indication of concept drifting. Therefore, it is very difficult to find a general correlation between the expected error of the previously extracted model and any type of concept drifting. Even though the expected error of a model does not change, a variation in the distribution can occur and the classifier is not updated to the current data. 2.2 Related work Data streams processing has recently become an important research domain. This aspect is testified by the number of papers currently published and by the real applications readily

VOLUME-8 ISSUE-10, AUGUST 2019, REGULAR ISSUE, 2019

Data Streams are having huge volume and it can-not be stored permanently in the memory for processing. In this paper we would be mainly focusing on issues in data stream, the major factors which are affecting the accuracy of classifier like imbalance class and Concept Drift. The drift in Data Stream mining refers to the change in data. Such as Class imbalance problem notifies that the samples are in the classes are not equal. In our research work we are trying to identify the change (Drift) in data, we are trying to detect Imbalance class and noise from changed data. And According to the type of drift we are applying the algorithms and trying to make the stream more balance and noise free to improve classifier’s accuracy.

Information Fusion, 2017

In many applications of information systems learning algorithms have to act in dynamic environments where data are collected in the form of transient data streams. Compared to static data mining, processing streams imposes new computational requirements for algorithms to incrementally process incoming examples while using limited memory and time. Furthermore, due to the non-stationary characteristics of streaming data, prediction models are often also required to adapt to concept drifts. Out of several new proposed stream algorithms, ensembles play an important role, in particular for nonstationary environments. This paper surveys research on ensembles for data stream classification as well as regression tasks. Besides presenting a comprehensive spectrum of ensemble approaches for data streams, we also discuss advanced learning concepts such as imbalanced data streams, novelty detection, active and semi-supervised learning, complex data representations and structured outputs. The paper concludes with a discussion of open research problems and lines of future research.

International Journal of Multimedia and Ubiquitous Engineering, 2015

We propose ensemble-based modeling for classifying streaming data with concept drift. The concept drift is a phenomenon in which the distribution of streaming data changes. In this paper, the types of the concept drift are categorized into the change of data distribution and the change of class distribution. The proposed ensemble modeling generates a meta-ensemble which consists of ensembles of classifiers. Whenever a change of class distribution occurs in streaming data, our modeling builds a new classifier of an existing ensemble and whenever a change of data distribution occurs, it builds a new ensemble which consists of an only one classifier. In our approach, new classifiers of a meta-ensemble on streaming data will be generated dynamically according to the estimated distribution of streaming data. We compared the results of our approach and of the chunk-based ensemble approach, which builds new classifiers of an ensemble periodically. In experiments with 13 benchmark data sets, our approach produced an average of 21.95% higher classification accuracy generating an average of 61.7% fewer new classifiers of an ensemble than the chunk-based ensemble method using partially labeled samples. We also examine that the time points when our approach builds new classifiers are appropriate for maintaining performance of an ensemble.

International Journal of Data Mining & Knowledge Management Process, 2012

Data mining is a user-centric process that is used to extract useful patterns from large volumes of data. With the growth of the Internet, a data stream is today a key area of advanced analysis and data mining. Handling data streams is a difficult task due to the variations in the data and the frequent occurrences of concept drifts. No single classifier can be relied upon to correctly classify data stream data since they are developed through a specific learning approach. Hence we use a multi-chunk ensemble of classifiers to classify evolving data streams and improve the prediction accuracy over single classifiers. We evaluate our ensemble on synthetic as well as real time data, compute the precision and represent it graphically using both majority voting as well as new proposed weighted averaging and compare its performance against individual classifiers. Current techniques include a single chunk approach, where the entire set of data is considered as a whole, and the used method shows better efficiency.

These data streams are potentially huge in size and thus it is impossible to process many data mining techniques (e.g., sensor

Wiley Interdisciplinary Reviews-Data Mining and Knowledge Discovery, 2019

Many real-world data mining applications have to deal with unlabeled streaming data. They are unlabeled because the sheer volume of the stream makes it impractical to label a significant portion of the data. The data streams can evolve over time and these changes are called concept drifts. Concept drifts have different characteristics, which can be used to categorize them into different types. A trade-off between performance and cost exists among many concept drift detection approaches. On the one hand, high accuracy detection approach usually requires labeled data, possibly involving high cost for labeling. On the other hand, a variety of methods have been devoted to the topic of concept drift detection with unlabeled data, but these approaches often are most suited for only a subset of the concept drift types. The objective of this survey is to present these methods, categorize them and give recommendations of usage based on their behaviors under different types of concept drift.

TURKISH JOURNAL OF ELECTRICAL ENGINEERING & COMPUTER SCIENCES

Concept drift is the phenomenon where underlying data distribution changes over time unexpectedly. Examining such drifts and getting insight into the executing processes at that instance of time is a big challenge. Prediction models should be capable of handling drifts in scenarios where statistical properties show abrupt changes. Various strategies exist in the literature to deal with such challenging scenarios but the majority of them are limited to the identification of a particular kind of drift pattern. The proposed approach uses online drift detection in a diversified adaptive setting with pruning techniques to formulate a concept drift handling approach, named ensemblebased online diversified drift detection (En-ODDD), with an aim to identify the majority of drifts including abrupt, gradual, recurring, mixed, etc. in a single model. En-ODDD is equipped with a dynamically updated ensemble to speed up the adaptability to changing distributions. Unlike prevalent approaches, which do not consider correlations between experts, En-ODDD entails experts using varying randomized subsets of input data. Different levels of sampling having been applied for diversity generation to promote generalization. Prediction accuracy has been used to evaluate the effectiveness of the proposed approach using Massive Online Analysis software and compared with ten state-of-theart algorithms. Experimental results on fifteen benchmark datasets (artificial and real-world) having up to one million instances depict that En-ODDD outperforms the existing approaches irrespective of nature of drift.