Temporal data mining is concerned with the exploration of large sequential (or temporally ordered) data sets to discover some nontrivial information that was previously unknown to the data owner. Sequential data sets come up naturally in a wide range of application domains, ranging from bioinformatics to manufacturing processes. Pattern discovery refers to a broad class of data mining techniques in which the objective is to unearth hidden patterns or unexpected trends in the data. In general, pattern discovery is about finding all patterns of 'interest' in the data and one popular measure of interestingness for a pattern is its frequency in the data. The problem of frequent pattern discovery is to find all patterns in the data whose frequency exceeds some user-defined threshold. Discovery of temporal patterns that occur frequently in sequential data has received a lot of attention in recent times. Different approaches consider different classes of temporal patterns and propose different algorithms for their efficient discovery from the data. This thesis is concerned with a specific class of temporal patterns called episodes and their discovery in large sequential data sets.
In the framework of frequent episode discovery, data (referred to as an event sequence or an event stream) is available as a single long sequence of events. The ith event in the sequence is an ordered pair, (Et,tt), where Et takes values from a finite alphabet (of event types), and U is the time of occurrence of the event. The events in the sequence are ordered according to these times of occurrence. An episode (which is the temporal pattern considered in this framework) is a (typically) short partially ordered sequence of event types. Formally, an episode is a triple, (V,<,9), where V is a collection of nodes, < is a partial order on V and 9 is a map that assigns an event type to each node of the episode. When < is total, the episode is referred to as a serial episode, and when < is trivial (or empty), the episode is referred to as a parallel episode. An episode is said to occur in an event sequence if there are events in the sequence, with event types same as those constituting the episode, and with times of occurrence respecting the partial order in the episode. The frequency of an episode is some measure of how often it occurs in the event sequence. Given a frequency definition for episodes, the task is to discover all episodes whose frequencies exceed some threshold. This is done using a level-wise procedure. In each level, a candidate generation step is used to combine frequent episodes from the previous level to build candidates of the next larger size, and then a frequency counting step makes one pass over the event stream to determine frequencies of all the candidates and thus identify the frequent episodes.
Frequency counting is the main computationally intensive step in frequent episode discovery. Choice of frequency definition for episodes has a direct bearing on the efficiency of the counting procedure. In the original framework of frequent episode discovery, episode frequency is defined as the number of fixed-width sliding windows over the data in which the episode occurs at least once. Under this frequency definition, frequency counting of a set of |C| candidate serial episodes of size N has space complexity O(N|C|) and time complexity O(ΔTN|C|) (where ΔT is the difference between the times of occurrence of the last and the first event in the data stream). The other main frequency definition available in the literature, defines episode frequency as the number of minimal occurrences of the episode (where, a minimal occurrence is a window on the time axis containing an occurrence of the episode, such that, no proper sub-window of it contains another occurrence of the episode). The algorithm for obtaining frequencies for a set of |C| episodes needs O(n|C|) time (where n denotes the number of events in the data stream). While this is time-wise better than the the windows-based algorithm, the space needed to locate minimal occurrences of an episode can be very high (and is in fact of the order of length, n, of the event stream).
This thesis proposes a new definition for episode frequency, based on the notion of, what is called, non-overlapped occurrences of episodes in the event stream. Two occurrences are said to be non-overlapped if no event corresponding to one occurrence appears in between events corresponding to the other. Frequency of an episode is defined as the maximum possible number of non-overlapped occurrences of the episode in the data. The thesis also presents algorithms for efficient frequent episode discovery under this frequency definition. The space and time complexities for frequency counting of serial episodes are O(|C|) and O(n|C|) respectively (where n denotes the total number of events in the given event sequence and |C| denotes the num-ber of candidate episodes). These are arguably the best possible space and time complexities for the frequency counting step that can be achieved. Also, the fact that the time needed by the non-overlapped occurrences-based algorithm is linear in the number of events, n, in the event sequence (rather than the difference, ΔT, between occurrence times of the first and last events in the data stream, as is the case with the windows-based algorithm), can result in considerable time advantage when the number of time ticks far exceeds the number of events in the event stream. The thesis also presents efficient algorithms for frequent episode discovery under expiry time constraints (according to which, an occurrence of an episode can be counted for its frequency only if the total time span of the occurrence is less than a user-defined threshold). It is shown through simulation experiments that, in terms of actual run-times, frequent episode discovery under the non-overlapped occurrences-based frequency (using the algorithms developed here) is much faster than existing methods.
There is also a second frequency measure that is proposed in this thesis, which is based on, what is termed as, non-interleaved occurrences of episodes in the data. This definition counts certain kinds of overlapping occurrences of the episode. The time needed is linear in the number of events, n, in the data sequence, the size, N, of episodes and the number of candidates, |C|. Simulation experiments show that run-time performance under this frequency definition is slightly inferior compared to the non-overlapped occurrences-based frequency, but is still better than the run-times under the windows-based frequency. This thesis also establishes the following interesting property that connects the non-overlapped, the non-interleaved and the minimal occurrences-based frequencies of an episode in the data: the number of minimal occurrences of an episode is bounded below by the maximum number of non-overlapped occurrences of the episode, and is bounded above by the maximum number of non-interleaved occurrences of the episode in the data. Hence, non-interleaved occurrences-based frequency is an efficient alternative to that based on minimal occurrences.
In addition to being superior in terms of both time and space complexities compared to all other existing algorithms for frequent episode discovery, the non-overlapped occurrences-based frequency has another very important property. It facilitates a formal connection between discovering frequent serial episodes in data streams and learning or estimating a model for the data generation process in terms of certain kinds of Hidden Markov Models (HMMs). In order to establish this connection, a special class of HMMs, called Episode Generating HMMs (EGHs) are defined. The symbol set for the HMM is chosen to be the alphabet of event types, so that, the output of EGHs can be regarded as event streams in the frequent episode discovery framework.
Given a serial episode, α, that occurs in the event stream, a method is proposed to uniquely associate it with an EGH, Λα. Consider two N-node serial episodes, α and β, whose (non-overlapped occurrences-based) frequencies in the given event stream, o, are fα and fβ respectively. Let Λα and Λβ be the EGHs associated with α and β. The main result connecting episodes and EGHs states that, the joint probability of o and the most likely state sequence for Λα is more than the corresponding probability for Λβ, if and only if, fα is greater than fβ. This theoretical connection has some interesting consequences. First of all, since the most frequent serial episode is associated with the EGH having the highest data likelihood, frequent episode discovery can now be interpreted as a generative model learning exercise. More importantly, it is now possible to derive a formal test of significance for serial episodes in the data, that prescribes, for a given size of the test, a minimum frequency for the episode needed in order to declare it as statistically significant. Note that this significance test for serial episodes does not require any separate model estimation (or training). The only quantity required to assess significance of an episode is its non-overlapped occurrences-based frequency (and this is obtained through the usual counting procedure). The significance test also helps to automatically fix the frequency threshold for the frequent episode discovery process, so that it can lead to what may be termed parameterless data mining.
In the framework considered so far, the input to frequent episode discovery process is a sequence of instantaneous events. However, in many applications events tend to persist for different periods of time and the durations may carry important information from a data mining perspective. This thesis extends the framework of frequent episodes to incorporate such duration information directly into the definition of episodes, so that, the patterns discovered will now carry this duration information as well. Each event in this generalized framework looks like a triple, (Ei, ti, τi), where Ei, as earlier, is the event type (from some finite alphabet) corresponding to the ith event, and ti and τi denote the start and end times of this event. The new temporal pattern, called the generalized episode, is a quadruple, (V, <, g, d), where V, < and g, as earlier, respectively denote a collection of nodes, a partial order over this collection and a map assigning event types to nodes. The new feature in the generalized episode is d, which is a map from V to 2I, where, I denotes a collection of time interval possibilities for event durations, which is defined by the user. An occurrence of a generalized episode in the event sequence consists of events with both 'correct' event types and 'correct' time durations, appearing in the event sequence in 'correct' time order. All frequency definitions for episodes over instantaneous event streams are applicable for generalized episodes as well. The algorithms for frequent episode discovery also easily extend to the case of generalized episodes. The extra design choice that the user has in this generalized framework, is the set, I, of time interval possibilities. This can be used to orient and focus the frequent episode discovery process to come up with temporal correlations involving only time durations that are of interest. Through extensive simulations the utility and effectiveness of the generalized framework are demonstrated.
The new algorithms for frequent episode discovery presented in this thesis are used to develop an application for temporal data mining of some data from car engine manufacturing plants. Engine manufacturing is a heavily automated and complex distributed controlled process with large amounts of faults data logged each day. The goal of temporal data mining here is to unearth some strong time-ordered correlations in the data which can facilitate quick diagnosis of root causes for persistent problems and predict major breakdowns well in advance. This thesis presents an application of the algorithms developed here for such analysis of the faults data. The data consists of time-stamped faults logged in car engine manufacturing plants of General Motors. Each fault is logged using an extensive list of codes (which constitutes the alphabet of event types for frequent episode discovery). Frequent episodes in fault logs represent temporal correlations among faults and these can be used for fault diagnosis in the plant. This thesis describes how the outputs from the frequent episode discovery framework, can be used to help plant engineers interpret the large volumes of faults logged, in an efficient and convenient manner. Such a system, based on the algorithms developed in this thesis, is currently being used in one of the engine manufacturing plants of General Motors. Some examples of the results obtained that were regarded as useful by the plant engineers are also presented.
| Identifer | oai:union.ndltd.org:IISc/oai:etd.ncsi.iisc.ernet.in:2005/375 |
| Date | 03 1900 |
| Creators | Laxman, Srivatsan |
| Contributors | Sastry, P S |
| Source Sets | India Institute of Science |
| Language | en_US |
| Detected Language | English |
| Type | Thesis |
| Rights | I grant Indian Institute of Science the right to archive and to make available my thesis or dissertation in whole or in part in all forms of media, now hereafter known. I retain all proprietary rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertation. |
Page generated in 0.0035 seconds