Abstract not found

. Landmarking is a novel technique for data characterization in metalearning. While conventional approaches typically describe a database with its statistical measurements and properties, landmarking proposes to enrich such a description with quick and easy-to-obtain performance measures of simple learning algorithms. In this paper, we will discuss two novel aspects of landmarking. First, we investigate relative landmarking, which tries to exploit the relative order of the landmark measures instead of their absolute value. Second, we propose to use subsampling estimates as a different way for efficiently obtaining landmarks. In general, our results are mostly negative. The most interesting result is a surprisingly simple rule that predicts quite accurately when it is worth to boost decision trees. 1

Current data mining tools are characterized by a plethora of algorithms but a lack of guidelines to select the right method according to the nature of the problem under analysis. Producing such guidelines is a primary goal by the field of meta-learning; the research objective is to understand the interaction between the mechanism of learning and the concrete contexts in which that mechanism is applicable. The field of meta-learning has seen continuous growth in the past years with interesting new developments in the construction of practical model-selection assistants, task-adaptive learners, and a solid conceptual framework. In this paper, we give an overview of different techniques necessary to build meta-learning systems. We begin by describing an idealized meta-learning architecture comprising a variety of relevant component techniques. We then look at how each technique has been studied and implemented by previous research. In addition, we show how metalearning has already been identified as an important component in real-world applications. 1

The use of subsampling for scaling up the performance of learning algorithms has become fairly popular in the recent literature. In this paper, we investigate the use of performance estimates obtained on a subsample of the data for the task of recommending the best learning algorithm(s) for the problem. In particular, we examine the use of subsampling estimates as features for meta-learning, thereby generalizing previous work on landmarking and on direct algorithm recommendation via subsampling.

exploration

There are two main paradigms in combining different classification algorithms: Classifier Selection and Classifier Fusion. The first one selects a single model for classifying a new instance, while the lat-ter combines the decisions of all models. The work presented in this paper stands in between these two paradigms aiming tackle the dis-advantages and benefit from the advantages of both. In particular, this paper proposes the use of statistical procedures for the selection of the best subgroup among different classification algorithms and the subsequent fusion of the decision of the models in this subgroup with simple methods like Weighted Voting. Extensive experimental results show that the proposed approach, Selective Fusion, improves over sim-ple selection and fusion methods, leading to performance comparable with the state-of-the-art heterogeneous classifier combination method of Stacking, without the additional computational cost and learning problems of meta-training. 1

Abstract — Meta-learning helps us find solutions to computational intelligence (CI) challenges in automated way. Metalearning algorithm presented in this paper is universal and may be applied to any type of CI problems. The novelty of our proposal lies in complexity controlled testing combined with very useful learning machines generators. The simplest and the best solutions are strongly preferred and are explored earlier. The learning algorithm is augmented by meta-knowledge repository which accumulates information about progress of the search through the space of candidate solutions. The approach facilitates using human experts knowledge to restrict the search space and provide goal definition, gaining meta-knowledge in an automated manner. I.

There are many knowledge-based data mining frameworks and it is common to think that new ones can not come up with anything new. This article refutes such claims. We propose a sophisticated unification mechanism and two-tier machine cache system aimed at saving time and memory. No machine is run twice. Instead, machines are reused wherever they are repeatedly requested (regardless of request context). These mechanisms are integrated with hierarchical task scheduler. Its unique design facilitates efficient automatic management of large numbers of tasks with natural adjustment to available computational resources. The system is also equipped with extremely powerful and general module for analysis of learning processes and evaluation of machine configurations with huge bunch of testing procedures of different kinds. Uniform results collection and query system is independent from particular learning machines. All the solutions are possible thanks to very general and universal definition of machine, configuration, machine context, machine information exchange and other necessary concepts. The ideas presented here, have been implemented and tested.

There are many data mining systems derived from machine learning, neural network, statistics and other fields. Most of them are dedicated to some particular algorithms or applications. Unfortunately, their architectures are still too naive to provide satisfactory background for advanced meta-learning problems. In order to efficiently perform sophisticated meta-level analysis, we have designed and implemented a very versatile, easily expandable system (in many independent aspects), which uniformly deals with different kinds of models and models with very complex structures (not only committees but also deeper hierarchic models). We present our requirements and their motivations for an advanced data mining system, and describe some of our solutions facilitating advanced meta-level model management at the scope of each system component, optimization of computation time and memory consumption and much more.

There are hundreds of algorithms within data mining. Some of them are used to transform data, some to build classifiers, others for prediction, etc. Nobody knows well all these algorithms and nobody can know all the arcana of their behavior in all possible applications. How to find the best combination of transformation and final machine which solves given problem? The solution is to use configurable and efficient meta-learning to solve data mining problems. Below, a general and flexible meta-learning system is presented. It can be used to solve different problems with computational intelligence, basing on learning from data. The main ideas of our meta-learning algorithms lie in complexity controlled loop, searching for most adequate models and in using special functional specification of search spaces (the meta-learning spaces) combined with flexible way of defining the goal of metasearching.