Advances in Databases and Information Systems by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9783030548322
Publisher: Springer International Publishing

4 Experimental Results

4.1 Analysis of Meta-feature Importance

Here we present an analysis of the importance rate of each meta-feature to predict the meta-target. These rates were measured by the meta-models performing a 5-fold cross validation over the meta-dataset. Results are presented in Fig. 4. Overall, for both meta-targets evaluated, the most relevant meta-features were the construction parameter NN, the search parameters R and k and, the graph type. Nonetheless, other features also contribute to the prediction. In the meta-models, for each tree split from a set of descriptors, all the following splits depend on the graph type and its parameters, as these are the meta-features that refine and determine the final behavior of the proximity graph over a given set of dataset descriptors. Regarding recall, we can observe that the most relevant meta-feature is the construction parameter NN. This is implied by the fact that the more edges a graph has, the better its query recall rate is. Similarly, for query time, we have the query parameter R as the most important meta-feature as the higher the number of restarts R is, the longer the query execution time is, and vice-versa. Moreover, the high importance rate of the ID measure for recall analysis is an interesting result. Excluding the graph configuration meta-features, the ID had the highest rate, thus, the embedding dimensionality showed no relevance in this case.

Fig. 4.The importance rate of each meta-feature per meta-target and category: (a) graph configurations, (b) general and info-theoretical, and (c) statistical.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.