Volume Two - Academic Conferences

4.1 The Mining Plans Selector system
Cristina Wanzeller and Orlando Belo
Figure 1 shows the MPS system fundamental functional components. We adapted the acknowledge
CBR cycle (Aamod and Plaza 1994) to the tasks at hands, devising (six) MPS constituent steps and
forming a problem solving and learning from experience strategy. The R’s steps, (Retrieve, Reuse,
Revise and Retain) are the original ones. Other steps are specific of the MPS system.
Figure 1: MPS functional components
The problem solving part of the system takes as inputs the clickstream data and the requirements
and, based on the cases kept on the knowledge base, delivers appropriate WUM plans. One MPS’s
specific task is to “characterize” the target usage data, producing metadata. We need a data
characterization consistent and independent from KD tool, in order to compare different datasets
systematically. Further, the extraction of some metadata must be automatized, as much as possible,
since clickstream data is typically huge. Another particular task is to “construct” a new WUM problem,
guiding, gathering and organizing the user’s explicit analysis constraints specification, through the
provided abstractions (e.g. analysis goals and application areas).
The “retrieve” step is a typical one and plays a vital role on problem solving. This step selects the
most plausible cases to found the construction of mining plans to recommend, according to the target
problem. The retrieve is based on the similarity threshold value, to choose the cases potentially more
effective. Next, the “reuse” task generates WUM plans, mostly based on the mining methods and the
levels of similarity of the retrieved cases, and considering also the evaluation criteria most important
to the analyst. This step does not performs extensive adaptation of the solution to the current
problem, namely in the wide sense intended by the original step. Yet, it focus the main parts of the
candidate cases that may be transferred to the target problems, by recommending mining methods
instead of mere cases. So, it prepares the reuse of the methods that constructed the solution. The
“revise” step is accomplished by the analyst, outside of the system, using a KD tool.
From the learning point of view, the system operates accepting heterogeneous descriptions of new
WUM processes and acquiring knowledge, through a semi-automated approach. The goal is to
simplify arduous activity, due to the great number of details. The accepted incomings are: documents
describing mining activities, generated by the KD tool in PMML; the process complementary
description, which would be exhaustive, if the used tool does not supports the PMML standard. The
learning starts with a “conciliate” task, to transform and combine the heterogeneous descriptions
items, supplied by user interaction and documents in PMML. Then the traditional “retain” task
essentially adds a new case to the knowledge base. This is realised by integrating and structuring the
incoming elements, considering the internal schema of the cases’ representation. After that, the case
becomes available for edition, to obtain additional information, such as items omitted in the
documents, decisions’ and settings’ explanations/justifications, discoveries and even new stages.
Indeed, the conciliate step may be accomplished in several iterations to enrich progressively the
description and benefit from the availability of the items that are extracted automatically.
4.2 Exemplifying problem solving
To exemplify problem solving we use a simple dataset and specify requisites comprising all the D, T
and P dimensions of problem description. We consider the improving site structure WUM typical
problem. We want to discover meaning relationships among learning materials. The idea is to find out
contents relevance and relations, taking into account the student’ point of view.
857