Discover Dynamic Evolutional Regularity Based On Higher Order Mining

In the real world, data changes with time, accordingly, useful rules show some development changes during several mining periods. Massive and complex time-series data are growing larger, increased the difficulties of human to understand of objective things, making the existing data mining techniques can no longer meet the needs of the dynamic changes information. Meanwhile, the diversification of first-order rules sources, making it difficult to select knowledge. With the value of the knowledge patterns continued to weaken, it is urgent to discover some new and interesting regularity.In this thesis, the basic concepts, theories and technology of the higher order mining are reviewed and analyzed systematically first. And then find a uniform way of representing mining patterns that otherwise exist in different formats and sources based on the ternary relation theory. A predetermined threshold is used to define the sources that will contribute to the knowledge selection; eventually a new first-order rule measure sequence is created. Furthermore, in order to discover the dynamic evolution regularity of the new sequence, we proposed diffusion estimation applicable to small, based on the information diffusion theory. Finally, combined the relational representation and information diffusion estimation, a complete higher order mining process is developed.There are two innovation points:Based on the traditional data mining method, a unified expression is proposed for knowledge from different sources, and appropriate method is developed to classify and preserve the most support knowledge patterns. The other is based on the principle of information diffusion, proposed an estimation method to dig the evolution of the new rules.The experiments show that, after analyzing of the formalized first-order rules, the new first-order rules are more stable and close to the original knowledge patterns of the actual situation. The results of the descriptive and predictive modeling with new first-order rules show that, the information diffusion estimation outperform EM algorithm and least squares estimation, the one without the prior assumed distribution is easier to operate and able to get more accurate estimates.