Aprendizaje por transferencia de redes bayesianas
In several domains, it is common to have data from different, but closely related problems (this means that the distributions of the data are similars but no equals). For instance, in manufacturing many products follow the same industrial process but with different conditions; or in industrial diagn...
| Autor: | |
|---|---|
| Tipo de documento: | dissertação |
| Estado: | Versión aceptada para publicación |
| Data de publicação: | 2009 |
| País: | México |
| Recursos: | Instituto Nacional de Astrofísica, Óptica y Electrónica |
| Repositório: | Repositorio Institucional del INAOE |
| Idioma: | espanhol |
| OAI Identifier: | oai:inaoe.repositorioinstitucional.mx:1009/401 |
| Acesso em linha: | http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/401 |
| Access Level: | Acceso aberto |
| Palavra-chave: | info:eu-repo/classification/Belie networks/Belie networks info:eu-repo/classification/Aprendizaje automático/Machine learning info:eu-repo/classification/Transferencia de aprendizaje/Transfer learning info:eu-repo/classification/cti/1 info:eu-repo/classification/cti/12 info:eu-repo/classification/cti/1203 |
| Resumo: | In several domains, it is common to have data from different, but closely related problems (this means that the distributions of the data are similars but no equals). For instance, in manufacturing many products follow the same industrial process but with different conditions; or in industrial diagnosis, where there is equipment with similar specifications. In these cases, it is common to have plenty of data for some scenarios but very little for other, for example, for rare products of little production. When there are a lot of data they can induce models from which can be used in diagnosis and classification tasks. However, as the exactitude of the induced model is based on the data available, having relatively little data are obtained very poor models. In order to improve the accuracy of models for learning domains with little data, one possibility is to use data and knowledge of similar domains. Using knowledge of similar domains has already been addressed in previous works introducing techniques known as MULTITASK LEARNING, whose objective is to improve multiple models simultaneously, or otherwise improve a single model using techniques known as TRANSFER LEARNING. The bayesians networks have not been used with the mentioned techniques previously. In this thesis, we propose a transfer learning method to learn Bayesian networks that considers both, structure and parameter learning. For structure learning, we use conditional independence tests, by combining measures from the target domain with those obtained from one or more auxiliary domains, transferring information from the most related domains with the aim of improving the accuracy of the less reliable parts of the network. For parameter learning, it’s compared three techniques for probability aggregation that combine probabilities estimated from the target domain with the auxiliary data. Through these techniques is to address two related problems: the lack of information in the domains with little data using domains related and a way of transferring knowledge from those domains related retaining characteristics of the target model. To validate the approach, are used three standard Bayesian networks commonly used in literature, and generated variants of each model by changing the structure as well as the parameters. Then learned on one of the variants with a small data set and combined it with information from the other variants. |
|---|