| Resumen |
The paper describes a novel associative model for time series data mining. The model is based on the Gamma classifier, which is inspired on the Alpha-Beta associative memories, which are both supervised pattern recognition models. The objective is to mine known patterns in the time series in order to forecast unknown values, with the distinctive characteristic that said unknown values may be towards the future or the past of known samples. The proposed model performance is tested both on time series forecasting benchmarks and a data set of oil monthly production. Some features of interest in the experimental data sets are spikes, abrupt changes and frequent discontinuities, which considerably decrease the precision of traditional forecasting methods. As experimental results show, this classifier-based predictor exhibits competitive performance. The advantages and limitations of the model, as well as lines of improvement, are discussed. |
