subspace a separate linear model is fit. These models all decide where to split the input space according to the observed input variables. There are several such segmentation-based approaches in market segmentation and it has been established that this tactic can help build better customer models. Clusterwise regression is a method for simultaneous clustering (not using the mixture model) and building predictive models. In a regression context the method clusters subjects non-hierarchically in such a way that the fit of the regression within clusters is optimized.
Under the mixture model framework, mixture regression models simultaneously group subjects into unobserved segments and estimate a regression model within each segment by relating a dependent variable to a set of independent variables. The mixture regression methods represent the mixture analog to the clusterwise regression methods. The identification of segments and simultaneous estimation of the response functions within each segment have been accomplished using a variety of mixture regression models.
At the crossing of the connectionist community and the time series community, Gated Expert models introduce chosen external variables to detect the switching of regimes in time series data. It consists of a gating neural network and several competing neural networks. The gating network learns to predict the probability of the prediction of each expert. The input of the gating network includes chosen external variables which are picked manually. When the driving force behind the splitting of the input space is unknown, it is not guaranteed that the hand-picked external variables will cover the hidden driving factors. In addition, effort needs to be taken to gather information about those external variables possibly for every data point in the training data set. Again, the gated expert method focuses on time series data, while our research emphasizes transaction data. The significant difference between our approach and the above methods is the use of pattern-based clustering approach to learn the individual segments.