Fast adaptive real-time classification for data streams with concept drift

An important application of Big Data Analytics is the real-time analysis of streaming data. Streaming data imposes unique challenges to data mining algorithms, such as concept drifts, the need to analyse the data on the fly due to unbounded data streams and scalable algorithms due to potentially high throughput of data. Real-time classification algorithms that are adaptive to concept drifts and fast exist, however, most approaches are not naturally parallel and are thus limited in their scalability. This paper presents work on the Micro-Cluster Nearest Neighbour (MC-NN) classifier. MC-NN is based on an adaptive statistical data summary based on Micro-Clusters. MC-NN is very fast and adaptive to concept drift whilst maintaining the parallel properties of the base KNN classifier. Also MC-NN is competitive compared with existing data stream classifiers in terms of accuracy and speed.

Discovering frequent sets from data streams with CPU constraint

Data streams are usually generated in an online fashion characterized by huge volume, rapid unpredictable rates, and fast changing data characteristics. It has been hence recognized that mining over streaming data requires the problem of limited computational resources to be adequately addressed. Since the arrival rate of data streams can significantly increase and exceed the CPU capacity, the machinery must adapt to this change to guarantee the timeliness of the results. We present an online algorithm to approximate a set of frequent patterns from a sliding window over the underlying data stream - given apriori CPU capacity. The algorithm automatically detects overload situations and can adaptively shed unprocessed data to guarantee the timely results. We theoretically prove, using probabilistic and deterministic techniques, that the error on the output results is bounded within a pre-specified threshold. The empirical results on various datasets also confirmed the feasiblity of our proposal.

EStream: online mining of frequent sets with precise error guarantee

In data stream applications, a good approximation obtained in a timely  manner is often better than the exact answer that’s delayed beyond the window of opportunity. Of course, the quality of the approximate is as important as its timely delivery. Unfortunately, algorithms capable of online processing do not conform strictly to a precise error guarantee. Since online processing is essential and so is the precision of the error, it is necessary that stream algorithms meet both criteria. Yet, this is not the case for mining frequent sets in data streams. We present EStream, a novel algorithm that allows online processing while producing results strictly within the error bound. Our theoretical and experimental results show that EStream is a better candidate for finding frequent sets in data streams, when both constraints need to be satisfied.

Incremental and adaptive clustering stream data over sliding window

Cluster analysis has played a key role in data stream understanding. The problem is difficult when the clustering task is considered in a sliding window model in which the requirement of outdated data elimination must be dealt with properly. We propose SWEM algorithm that is designed based on the Expectation Maximization technique to address these challenges. Equipped in SWEM is the capability to compute clusters incrementally using a small number of statistics summarized over the stream and the capability to adapt to the stream distribution’s changes. The feasibility of SWEM has been verified via a number of experiments and we show that it is superior than Clustream algorithm, for both synthetic and real datasets.

Stream management within the CloudMiner

Nowadays cloud computing has become a major trend that enterprises and research organizations are pursuing with increasing zest. A potentially important application area for clouds is data analytics. In our previous publication, we introduced a novel cloud infrastructure, the CloudMiner, which facilitates data mining on massive scientific data. By providing a cloud platform which hosts data mining cloud services following the Software as a Service (SaaS) paradigm, CloudMiner offers the capability for realizing cloud-based data mining tasks upon traditional distributed databases and other dataset types. However, little attention has been paid to the issue of data stream management on the cloud so far. We have noticed the fact that some features of the cloud meet very well the requirements of data stream management. Consequently, we developed an innovative software framework, called the StreamMiner, which is introduced in this paper. It serves as an extension to the CloudMiner for facilitating, in particular, real-world data stream management and analysis using cloud services. In addition, we also introduce our tentative implementation of the framework. Finally, we present and discuss the first experimental performance results achieved with the first StreamMiner prototype.

An efficient one-pass method for discovering bases of recently frequent episodes over online data streams

The knowledge embedded in an online data stream is likely to change over time due to the dynamic evolution of the stream. Consequently, infrequent episode mining over an online stream, frequent episodes should be adaptively extracted from recently generated stream segments instead of the whole stream. However, almost all existing frequent episode mining approaches find episodes frequently occurring over the whole sequence. This paper proposes and investigates a new problem: online mining of recently frequent episodes over data streams. In order to meet strict requirements of stream mining such as one-scan, adaptive result update and instant result return, we choose a novel frequency metric and define a highly condensed set called the base of recently frequent episodes. We then introduce a one-pass method for mining bases of recently frequent episodes. Experimental results show that the proposed method is capable of finding bases of recently frequent episodes quickly and adaptively. The proposed method outperforms the previous approaches with the advantages of one-pass, instant result update and return, more condensed resulting sets and less space usage.

Ontology-based access to sensor data streams

Sensor networks are increasingly becoming one of the main sources of Big Data on the Web. However, the observations that they produce are made available with heterogeneous schemas, vocabularies and data formats, making it difficult to share and reuse these data for other purposes than those for which they were originally set up. In this thesis we address these challenges, considering how we can transform streaming raw data to rich ontology-based information that is accessible through continuous queries for streaming data. Our main contribution is an ontology-based approach for providing data access and query capabilities to streaming data sources, allowing users to express their needs at a conceptual level, independent of implementation and language-specific details. We introduce novel query rewriting and data translation techniques that rely on mapping definitions relating streaming data models to ontological concepts. Specific contributions include: • The syntax and semantics of the SPARQLStream query language for ontologybased data access, and a query rewriting approach for transforming SPARQLStream queries into streaming algebra expressions. • The design of an ontology-based streaming data access engine that can internally reuse an existing data stream engine, complex event processor or sensor middleware, using R2RML mappings for defining relationships between streaming data models and ontology concepts. Concerning the sensor metadata of such streaming data sources, we have investigated how we can use raw measurements to characterize streaming data, producing enriched data descriptions in terms of ontological models. Our specific contributions are: • A representation of sensor data time series that captures gradient information that is useful to characterize types of sensor data. • A method for classifying sensor data time series and determining the type of data, using data mining techniques, and a method for extracting semantic sensor metadata features from the time series.

Mining recurring concepts in a dynamic feature space

Most data stream classification techniques assume that the underlying feature space is static. However, in real-world applications the set of features and their relevance to the target concept may change over time. In addition, when the underlying concepts reappear, reusing previously learnt models can enhance the learning process in terms of accuracy and processing time at the expense of manageable memory consumption. In this paper, we propose mining recurring concepts in a dynamic feature space (MReC-DFS), a data stream classification system to address the challenges of learning recurring concepts in a dynamic feature space while simultaneously reducing the memory cost associated with storing past models. MReC-DFS is able to detect and adapt to concept changes using the performance of the learning process and contextual information. To handle recurring concepts, stored models are combined in a dynamically weighted ensemble. Incremental feature selection is performed to reduce the combined feature space. This contribution allows MReC-DFS to store only the features most relevant to the learnt concepts, which in turn increases the memory efficiency of the technique. In addition, an incremental feature selection method is proposed that dynamically determines the threshold between relevant and irrelevant features. Experimental results demonstrating the high accuracy of MReC-DFS compared with state-of-the-art techniques on a variety of real datasets are presented. The results also show the superior memory efficiency of MReC-DFS.

Approximate processing of massive continuous quantile queries over high-speed data streams

Quantile computation has many applications including data mining and financial data analysis. It has been shown that an is an element of-approximate summary can be maintained so that, given a quantile query d (phi, is an element of), the data item at rank [phi N] may be approximately obtained within the rank error precision is an element of N over all N data items in a data stream or in a sliding window. However, scalable online processing of massive continuous quantile queries with different phi and is an element of poses a new challenge because the summary is continuously updated with new arrivals of data items. In this paper, first we aim to dramatically reduce the number of distinct query results by grouping a set of different queries into a cluster so that they can be processed virtually as a single query while the precision requirements from users can be retained. Second, we aim to minimize the total query processing costs. Efficient algorithms are developed to minimize the total number of times for reprocessing clusters and to produce the minimum number of clusters, respectively. The techniques are extended to maintain near-optimal clustering when queries are registered and removed in an arbitrary fashion against whole data streams or sliding windows. In addition to theoretical analysis, our performance study indicates that the proposed techniques are indeed scalable with respect to the number of input queries as well as the number of items and the item arrival rate in a data stream.

Adaptive unsupervised learning of human actions

Automatic detection of suspicious activities in CCTV camera feeds is crucial to the success of video surveillance systems. Such a capability can help transform the dumb CCTV cameras into smart surveillance tools for fighting crime and terror. Learning and classification of basic human actions is a precursor to detecting suspicious activities. Most of the current approaches rely on a non-realistic assumption that a complete dataset of normal human actions is available. This paper presents a different approach to deal with the problem of understanding human actions in video when no prior information is available. This is achieved by working with an incomplete dataset of basic actions which are continuously updated. Initially, all video segments are represented by Bags-Of-Words (BOW) method using only Term Frequency-Inverse Document Frequency (TF-IDF) features. Then, a data-stream clustering algorithm is applied for updating the system's knowledge from the incoming video feeds. Finally, all the actions are classified into different sets. Experiments and comparisons are conducted on the well known Weizmann and KTH datasets to show the efficacy of the proposed approach.

A suspicious behaviour detection using a context space model for smart surveillance systems

Video surveillance systems using Closed Circuit Television (CCTV) cameras, is one of the fastest growing areas in the field of security technologies. However, the existing video surveillance systems are still not at a stage where they can be used for crime prevention. The systems rely heavily on human observers and are therefore limited by factors such as fatigue and monitoring capabilities over long periods of time. This work attempts to address these problems by proposing an automatic suspicious behaviour detection which utilises contextual information. The utilisation of contextual information is done via three main components: a context space model, a data stream clustering algorithm, and an inference algorithm. The utilisation of contextual information is still limited in the domain of suspicious behaviour detection. Furthermore, it is nearly impossible to correctly understand human behaviour without considering the context where it is observed. This work presents experiments using video feeds taken from CAVIAR dataset and a camera mounted on one of the buildings Z-Block) at the Queensland University of Technology, Australia. From these experiments, it is shown that by exploiting contextual information, the proposed system is able to make more accurate detections, especially of those behaviours which are only suspicious in some contexts while being normal in the others. Moreover, this information gives critical feedback to the system designers to refine the system.

A comparison of methods to identify alcohol involvement in youth injury-related emergency department presentation data

Aims: To compare different methods for identifying alcohol involvement in injury-related emergency department presentation in Queensland youth, and to explore the alcohol terminology used in triage text. Methods: Emergency Department Information System data were provided for patients aged 12-24 years with an injury-related diagnosis code for a 5 year period 2006-2010 presenting to a Queensland emergency department (N=348895). Three approaches were used to estimate alcohol involvement: 1) analysis of coded data, 2) mining of triage text, and 3) estimation using an adaptation of alcohol attributable fractions (AAF). Cases were identified as ‘alcohol-involved’ by code and text, as well as AAF weighted. Results: Around 6.4% of these injury presentations overall had some documentation of alcohol involvement, with higher proportions of alcohol involvement documented for 18-24 year olds, females, indigenous youth, where presentations occurred on a Saturday or Sunday, and where presentations occurred between midnight and 5am. The most common alcohol terms identified for all subgroups were generic alcohol terms (eg. ETOH or alcohol) with almost half of the cases where alcohol involvement was documented having a generic alcohol term recorded in the triage text. Conclusions: Emergency department data is a useful source of information for identification of high risk sub-groups to target intervention opportunities, though it is not a reliable source of data for incidence or trend estimation in its current unstandardised form. Improving the accuracy and consistency of identification, documenting and coding of alcohol-involvement at the point of data capture in the emergency department is the most desirable long term approach to produce a more solid evidence base to support policy and practice in this field.

An hybrid approach for efficient multicast stream authentication over unsecured channels

We study the multicast stream authentication problem when an opponent can drop, reorder and inject data packets into the communication channel. In this context, bandwidth limitation and fast authentication are the core concerns. Therefore any authentication scheme is to reduce as much as possible the packet overhead and the time spent at the receiver to check the authenticity of collected elements. Recently, Tartary and Wang developed a provably secure protocol with small packet overhead and a reduced number of signature verifications to be performed at the receiver. In this paper, we propose an hybrid scheme based on Tartary and Wang’s approach and Merkle hash trees. Our construction will exhibit a smaller overhead and a much faster processing at the receiver making it even more suitable for multicast than the earlier approach. As Tartary and Wang’s protocol, our construction is provably secure and allows the total recovery of the data stream despite erasures and injections occurred during transmission.

Representing team behaviours from noisy data using player role

Due to their unobtrusive nature, vision-based approaches to tracking sports players have been preferred over wearable sensors as they do not require the players to be instrumented for each match. Unfortunately however, due to the heavy occlusion between players, variation in resolution and pose, in addition to fluctuating illumination conditions, tracking players continuously is still an unsolved vision problem. For tasks like clustering and retrieval, having noisy data (i.e. missing and false player detections) is problematic as it generates discontinuities in the input data stream. One method of circumventing this issue is to use an occupancy map, where the field is discretised into a series of zones and a count of player detections in each zone is obtained. A series of frames can then be concatenated to represent a set-play or example of team behaviour. A problem with this approach though is that the compressibility is low (i.e. the variability in the feature space is incredibly high). In this paper, we propose the use of a bilinear spatiotemporal basis model using a role representation to clean-up the noisy detections which operates in a low-dimensional space. To evaluate our approach, we used a fully instrumented field-hockey pitch with 8 fixed high-definition (HD) cameras and evaluated our approach on approximately 200,000 frames of data from a state-of-the-art real-time player detector and compare it to manually labeled data.