| With the "information explosion" and the "data surplus" phenomenon of currentsociety, in the face of increasingly complex information system, how to eliminateredundant noise and get the useful knowledge have become the urgent needs of thepeople. Rough set is a relatively new soft computing method. In recent years, it hasgot increasing attention of the scholars and experts. It is often used to deal withimprecise, uncertain data mining problems and plays a more and more important rolerelying on its advantages in the field of data mining.In this paper, with respect to the problem that the traditional data mining methodcannot handle noisy data effectively, rough set attribute reduction is in-depth studiedin the theory and application. Main topics of this dissertation are as follows:Firstly, the basic principle and theory of rough set theory in data mining areanalyzed. And the attribute reduction of decision tables in rough set data mining isstudied in-depth in this thesis. Through reviewing the existing reduction methods, thesignificant research direction which is combing rough set theory with intelligentalgorithm is pointed out.Secondly, on the theory basis of the standard adaptive genetic algorithm, thispaper proposes an improved adaptive genetic algorithm. And in this algorithm, a newmutation operator and crossover operators are put forward to enhance theconvergence rate of the genetic evolution process, effectively solving the problemssuch as the standard adaptive genetic algorithm is easy to fall into a local optimum.Experiment shows that this genetic algorithm compared to other traditional geneticalgorithms has stronger capability of the global optimal solution.Thirdly, an attribute reduction method is put forward which based on theimproved self-adaptive genetic algorithm. At the same time, a new kind of fitnessfunction of the genetic algorithm is also put forward in this method. Experimentshows that this algorithm compared to other traditional algorithms has strongerattribute reduction ability and can accelerate the speed of the attribute reduction.Finally, on the basis of studying and analyzing the uncertain data miningmethods in stock market, a new stock prediction model is put forward to forecastclosing price of stock. As the first step, in the process of the data cleaning,transformation and discretion, utilizing a method of cluster analysis effectively solved the problem of continuous attributes discretization. As the second step,10attributions are obtained from26ones using the improved attribution reductionalgorithm. At last, the paper employs the improved attribute reduction method andbuilds up the time series prediction model which is using neural network to domachine learning by245stock indicators. In addition, the experiment shows thatimproved algorithms of this paper are feasible and effective in dealing with noisydata and the new stock prediction model has a good guiding role for short-terminvestment in the stock market of China. |
