It is mentioned in this article ''it's plausible to use the ANN method to establish a model, nevertheless, large amount of data is required; therefore the idea was neglected'' but in another published article in CBM 20 (2006) 441-449 the same data are used to train an ANN model!
Would you justify the performance of this ann model with a 45 data set?
Article Prediction of sulfate expansion of PC mortar using adaptive ...
Ok, it is apparent that, larger the data sets are, more confident you are. Nevertheless, this was the data at hand. It was not easy to collect this kind of data since it was time-consuming. On the other hand,, the data sets scattering and distribution were good enough to characterize the behavior (tendency or relationship). These studies are generally conducted in a trial-and-error approach, selection of training and testing data sets are quite effective on the performance of the system. Last but not the least, if we had the opportunity to use a larger data set, be sure we would. As soon as we obtain additional data, we plan to republish the results. However, the performance of the network in the study is sufficiently good.
The number of training data points alone bears little significance towards the validity of a model. I agree with A.Sezer that a representative distribution of the data set is more important.
A reliable assessment of model performance and ability to generalize is only possible with either fresh data that was not part of the training set or (slightly less reliable) using techniques like cross-validation ot bootstrapping.
- Badges
- Science topic
- Similar topics
- Engineering
- Computer Engineering