004 Informatik
Refine
Document Type
- Journal article (1)
- Conference proceeding (1)
- Doctoral Thesis (1)
Language
- English (3) (remove)
Is part of the Bibliography
- yes (3)
Institute
- Informatik (3)
Publisher
- Gesellschaft für Informatik e.V (1)
- IARIA (1)
Data collected from internet applications are mainly stored in the form of transactions. All transactions of one user form a sequence, which shows the user´s behaviour on the site. Nowadays, it is important to be able to classify the behaviour in real time for various reasons: e.g. to increase conversion rate of customers while they are in the store or to prevent fraudulent transactions before they are placed. However, this is difficult due to the complex structure of the data sequences (i.e. a mix of categorical and continuous data types, constant data updates) and the large amounts of data that are stored. Therefore, this thesis studies the classification of complex data sequences. It surveys the fields of time series analysis (temporal data mining), sequence data mining or standard classification algorithms. It turns out that these algorithms are either difficult to be applied on data sequences or do not deliver a classification: Time series need a predefined model and are not able to handle complex data types; sequence classification algorithms such as the apriori algorithm family are not able to utilize the time aspect of the data. The strengths and weaknesses of the candidate algorithms are identified and used to build a new approach to solve the problem of classification of complex data sequences. The problem is thereby solved by a two-step process. First, feature construction is used to create and discover suitable features in a training phase. Then, the blueprints of the discovered features are used in a formula during the classification phase to perform the real time classification. The features are constructed by combining and aggregating the original data over the span of the sequence including the elapsed time by using a calculated time axis. Additionally, a combination of features and feature selection are used to simplify complex data types. This allows catching behavioural patterns that occur in the course of time. This new proposed approach combines techniques from several research fields. Part of the algorithm originates from the field of feature construction and is used to reveal behaviour over time and express this behaviour in the form of features. A combination of the features is used to highlight relations between them. The blueprints of these features can then be used to achieve classification in real time on an incoming data stream. An automated framework is presented that allows the features to adapt iteratively to a change in underlying patterns in the data stream. This core feature of the presented work is achieved by separating the feature application step from the computational costly feature construction step and by iteratively restarting the feature construction step on the new incoming data. The algorithm and the corresponding models are described in detail as well as applied to three case studies (customer churn prediction, bot detection in computer games, credit card fraud detection). The case studies show that the proposed algorithm is able to find distinctive information in data sequences and use it effectively for classification tasks. The promising results indicate that the suggested approach can be applied to a wide range of other application areas that incorporate data sequences.
A sequence of transactions represents a complex and multi dimensional type of data. Feature construction can be used to reduce the data´s dimensionality to find behavioural patterns within such sequences. The patterns can be expressed using the blue prints of the constructed relevant features. These blue prints can then be used for real time classification on other sequences.
When forecasting sales figures, not only the sales history but also the future price of a product will influence the sales quantity. At first sight, multivariate time series seem to be the appropriate model for this task. Nonetheless, in real life history is not always repeatable, i.e., in the case of sales history there is only one price for a product at a given time. This complicates the design of a multivariate time series. However, for some seasonal or perishable products the price is rather a function of the expiration date than of the sales history. This additional information can help to design a more accurate and causal time series model. The proposed solution uses an univariate time series model but takes the price of a product as a parameter that influences systematically the prediction based on a calculated periodicity. The price influence is computed based on historical sales data using correlation analysis and adjustable price ranges to identify products with comparable history. The periodicity is calculated based on a novel approach that is based on data folding and Pearson Correlation. Compared to other techniques this approach is easy to compute and allows to preset the price parameter for predictions and simulations. Tests with data from the Data Mining Cup 2012 as well as artificial data demonstrate better results than established sophisticated time series methods.