Research On Periodicity Detecting And Appling For The Sequence Of Events With Cross Entropy

If a particular event periodically recurs in the sequence, we believe that theparticular event has periodically in this continuous sequence. Observing from thehistorical sequence data and finding the periodic of the event can reveal the future trendof the event, and thus provide more valuable information for management activities,help decision-makers to implement more effective decisions. The previous studies aboutthe related fields generally need to know the mode of the incident to find the cycle, ordetect the period with a user-specified value. In order to find the periodicity of thespecial event in time-series effectively in the case of the absence of any information topredict, this paper proposes a method based on cross-entropy to detect periodic of agiven event in the sequence data.The entire thesis consists of six chapters. First, it introduced the background to thestudy, specifically analyzed the existing methods of data mining, its branch——timeseries mining and commercial applications. The second chapter is a review of theresearch, it synthetically analyzed various types of periodic detection algorithm and theadvantages and disadvantages of their respective applications, so there need a simpledetection algorithm without pre-set cycle which can be used in commercial sequence.The third chapter is a general overview of the entire periodicity detection algorithm; itexplained the research-related concepts and the function of each method in thealgorithms. The fourth chapter detailed decomposed and interpreted the algorithm, andused two sets of real data to validate the feasibility of the algorithm. The fifth chapter isthe application of the algorithm. On the EC platform, the algorithm can detect theconsumers’ purchase period of various products, and the algorithm can combinepurchase period and the basic functions of the recommendation system to make therecommendations are more accurate. Finally it’s the full summary and outlook for futureresearch.The main contribution of this thesis is that it investigated the problem of discovering periodicity of a certain event in a binary data series and a new methodbasing on cross entropy is proposed. First, a series of rational partition methods forbinary data series are introduced, which can divide the data series into differentsegments. Then, we use cross entropy to calculate both the partition feasibility andstructure feasibility, which could be the good measurements for the feasible of eventperiodicity. Finally, a periodicity evaluation method is proposed to obtain the feasibleperiodicity of the given event. The results of calculation example show that the methodcan be used to explore feasible event periodicity in binary data series. Through purchasedata analysis, it proved that this algorithm can detect the consumers’ purchase period ofvarious products, Whereby, in the future recommendation system, we will be able tointroduce the time factor as a dimension of recommendation, then the Recommendedresults will be more accurate and reliable.