Multiple data sets (MDS) emerge from the progress of measurement tools, technologies of data storage, and accumulation of data in VLDB and data warehouses. Methods of solving problems of ordinary data sets analysis, e.g. global modeling, discovery of dependencies, exploratory data analysis and manipulation, get less applicable when dealing with massive data owing to huge overheads in time and computer memory to produce models. Even if workable, the models reveal poor performance on the whole MDS. The main approach of adequate solving the problems that this thesis is based on is using artificial intelligence methods for flexible modelling data and providing the tradeoff between generality of models being created and the size of a MDS, and computation resources. Besides special algorithms implemented, a necessary component of MDS analysis software is means for user's control of the algorithms' run, representation of results, decision making support, control of data input - that is, the user interface. A ccordingly, a part of research in the thesis was aimed at integration of modelling techniques developed in it as well as known ones within the so called intelligent user interface. The principal idea behind this integration is an approach proposed in the thesis - modelling automation which requires that composing training subsets of possible MDS's models, training them, selection of the most adequate one, its timely adjustment, retraining and discarding as the training sets grow or new examples get available be performed without human intervention. Some domains like geophysical interpretation require more complex analysis, e.g. multi-phase building different models of an MDS, its consequent modification by a human expert by the knowledge about models' performance, support of analysis alternatives (models developed and their training data including intermediate ones). Features of this type of analysis are inadequacy of automatic modelling and moreover the need of means for interactive analysis. In the thesis, a specialized programming language with constructs for data analysis operations is proposed. Its features are opeations for manipulations of MDS's and their constituents (parameter vectors) as instances of special data types - collections, sub-collections, and single elements. In the language, various MDS analysis objects (vectors of parameters and their sets - classes, clusters created via different types of analysis) can be represented and manipulated with various domain-specific operations, e.g. (a) forming classes out of parameter vectors or their projections stored in the database and treating them via set theory operations, (b) processing single parameter vectors, e.g. classifying and projecting vectors, (c) performing control of a previousely built-up classification configuration, e.g. automatic classification by universal methods (decision trees, nearest neighbours, fuzzy techniques), retraining, discarding and merging classes. In the thesis, specifics of MDS primarily those resulting in inadequacy of their models were considered and a principle of limited history impact was proposed. It claimes that the joint probability distribution of the unknown function modelled and its parameters drift in time as the effect of the underlying process of an MDS is stable only locally but not globally; so local models only can be trained on the data set. Based on the order of locally stable subsets, an MDS can be further analysed as an equivalent time series. (A substantial fraction of massive datasets are in fact temporal.) An algorithm for local training models of temporal MDS's were presented along with a criterion of model performance. Another algorithm based on the same criterion seeks a boundary between adjasent local subsets as a point of unimprovable model performance. Statistical aspects of training on MDS's were considered. By assuming parameter drift of an underlying process as a series of discrete, binomially distributed "revolutions", estimates of the VC confidence of the expected risk in the classification and regression regain problems for the case of a training set being an MDS. With the probability of "revolution" vanish, the estimates approach to the Vapnik-Chervonenkis ones.
