Geoscience Reference
In-Depth Information
To our knowledge, no GIRS implements partially balanced aggregation.
3.3.3. Summary and positioning of a partially compensatory GIR
Section 3.3 presents different studies concerning the standardization and
combination of criteria in different fields, such as IR, multimedia IR, GIR and
decision support. It has allowed us to grasp the elements in published literature
dealing with our issue, which is the proposal of models and methods for the
generalization of indexed data representation, on the one hand, and multicriteria IR,
on the other.
From this study, it follows that the IR systems targeting several dimensions,
such as GIRSs, usually federate mono-dimensional IRSs, each supporting their own
indexes and matching operators dedicated to their field. GIRSs have, in general, done
withoutthepreliminaryindexstandardizationstep,byimplementingrelativelysimple
combination approaches such as filtering of linear combination. Thus, we retain the
approachproposedbyPham et al. [PHA 07]inmultimediaIR,dealingwithimagesin
a way similar to texts, in other words using visual lemmas (visterms) and calculating
weights relative to their raw frequency. This approach leads to homogeneous forms
of representing different types of information. We propose representing spatial and
temporal information according to an approach of standardization by spatial and
temporal tiles followed by testing our first two working hypotheses:
- the generalization of data representations is adapted to spatial and temporal
information;
- the generalization associated with a well-tried IR model does not lead to a loss
of quality with respect to the initial dedicated IRSs.
We thus propose measuring the eventual loss of precision induced by
generalization by comparing the ad hoc IR models, implemented in different versions
of the PIV prototype, with the vectorial IR model applied to the new generalized
indexes. We choose the vectorial model [SAL 71, SAL 75], well-tried and giving
good results [BAE 99] and, although dedicated to terms, we apply it to spatial and
temporal tiles. We study the interest of integrating the frequencies of tiles into the
calculation of relevance scores by the vectorial model. Moreover, we observe the
impact of the different range choices for spatial and calendar tiles on the precision of
vectorial IR.
From the previous study, it can also be seen that the approaches of criteria
combination in GIR use dedicated IRSs for each dimension, on the one hand, and do
notallowanyconfigurationofthecombinationontheother.Weretaintheapproaches
of multicriteria decision support whose flexibility allows us to specify, for each
criterion,thelevelofcompensationdesiredbytheuser[MAL 03,BÜY 10].Similarly,
Search WWH ::
Custom Search