Constraint-based recommenders support customers in identifying relevant items from complex item assortments. In this paper we present WeeVis, a constraint-based environment that can be applied in di erent scenarios in the e-government domain. WeeVis supports collaborative knowledge acquisition for recommender applications in a MediaWiki-based context. This paper shows how Wiki pages can be extended with recommender applications and how the environment uses intelligent mechanisms to support users in identifying the optimal solutions to their needs. An evaluation shows a performance overview with di erent knowledge bases.
Constraint-based recommender applications help users navigating in complex
product and service assortments like digital cameras, computers, nancial
services and municipality services. The calculation of the recommendations is based
on a knowledge base of explicitly de ned rules. The engineering of the rules for
recommender knowledge bases (for constraint-based recommenders) is typically
done by knowledge engineers, mostly computer scientists [
Other recommendation approaches like collaborative ltering use
information about the rating behavior of other users to identify recommendations [
In the line of Wikipedia4 where users build and maintain Wiki pages collaboratively we introduce WeeVis5. WeeVis is a MediaWiki6 based environment that exploits the properties of MediaWiki and enables community based development and maintenance of knowledge bases for constraint-based recommenders. WeeVis is freely available as a platform and successfully applied by four Austrian universities (in lectures about recommender systems), in the nancial services domain and in e-government .
In the e-government domain o cials as well as the community residents can take numerous advantages of knowledge-based recommenders: { WeeVis can be used as an online advisory service for citizens for example for documents that are necessary to apply for a private construction project. The online recommendation of necessary documents in advance to on-site appointments can lead to a time reduction for community residents and community o cials. { WeeVis can be used for modeling internal processes like the signing of travel applications for example a community o cial wants to visit a conference, based on di erent parameters like the conference type, or if it's abroad or in the domestic area, di erent o cials have to sign the travel request. In WeeVis the appropriate rules for such internal processes can be mapped and especially for new employees WeeVis recommenders can provide substantial assistance. { WeeVis can be used as an information platform for example with integrated knowledge-based recommenders for community residents e.g. to identify the optimal waste disposal strategy for a household (this example is used as a running example in this paper, see Section 2). Instead of providing plain text information, like common municipality web pages, the knowledge representation as a recommender provides an easier way for community members to identify the optimal solution for their situation.
A recommender development environment for single users is introduced in [
The WeeVis interface also provides intelligent mechanisms for an instant
presentation of alternative solutions in situations where it is not possible to nd
a solution for a given set of user (customer) requirements, i.e., the requirements
are inconsistent with the recommendation knowledge base and the user is in
the need for repair proposals to nd a way out from the no solution could be
found dilemma. Model-based diagnosis [
A Semantic Wiki-based approach to knowledge acquisition for collaborative
ontology development is introduced, for example, in [
The remainder of this paper is organized as follows. In Section 2 we present an overview of the recommendation environment WeeVis and it's application in the e-government domain. In Section 3 we present results of a performance evaluation that illustrates the performance of the integrated diagnosis technologies. With Section 4 we conclude the paper. 2
Since WeeVis is based on the MediaWiki platform, it can be installed on freely available web servers. On the website www.weevis.org a selection of di erent WeeVis recommenders is publicly available. For internal processes WeeVis can be deployed in the local intranet. Standard wiki pages can be complemented easily by recommender knowledge bases. Currently, WeeVis calculates recommendations based on previously entered requirements. If the requirements would result in a no solution could be found message Weevis calculates alternative solutions based on diagnoses (see Section 2.4). In line with the Wiki idea, WeeVis provides the ability to build knowledge bases collaboratively, a valuable feature in e-government domain, because depending on the community department multiple people are responsible for data management and administration. Furthermore, WeeVis exploits the basic functionalities provided by MediaWiki and allows rapid prototyping processes where the result of a change can immediately be seen by simply switching from the edit mode to the corresponding read mode. This approach allows an easy understanding of the WeeVis tags and also of the semantics of the provided WeeVis language. 2.1
Since WeeVis is a MediaWiki-based environment the user interface relies on the common Wiki principle of the read mode (see Figure 1) for executing a recommender and the write mode (see Figure 2) for de ning a recommender knowledge base. The development and maintenance of a knowledge base is supported a textual fashion with a syntax that is similar to the standard Wiki syntax (see Figure 2). In the following we will present the concepts integrated in the WeeVis environment on the basis of a working example from the e-government domain. More speci cally we present a recommender that supports households in identifying their optimal waste disposal strategy. In this recommendation scenario, a user has to specify his/her requirements regarding, for example, the number of persons living in the household or how frequently the containers should be emptied. A corresponding WeeVis user interface is depicted in Figure 1. Requirements are speci ed on the left hand side and the corresponding recommendations for the optimal waste disposal plan are displayed in the right hand side.
For each solution, a so-called support score is determined. If a solution ful lls all requirements, this score is 100%, otherwise it is lower and, when clicking on the score value, a corresponding repair action is displayed on the left-hand side (see Figure 1). Due to the automated alternative determination, WeeVis is always able to present a solution and users are never ending up in the no solution could be found dilemma (see Figure 1).
An example of the de nition of a (simpli ed) e-government recommender knowledge base is depicted in Figure 2. The de nition of a recommender knowledge base is supported in a textual fashion on the basis of a syntax similar to MediaWiki. Basic syntactical elements provided in WeeVis will be introduced in the next subsection. 2.2
A WeeVis recommender consists of three necessary aspects, the de nition of questions and possible answers, items and their properties, and constraints (see Figure 2).
The de nition of an item assortment in WeeVis starts with the &PRODUCTS tag (see Figure 2). The rst line represents the attributes separated by the exclamation mark. In our example, the item assortment is speci ed by the name, sizep, the container size, emptyingp, the emptying frequency, and pricep, the price of the waste disposal plan. Each of the next lines represents an item with the values related to the attributes, in our example there are three items speci ed: Small Plan, Medium Plan, and Large Plan.
The second aspect starts with the &QUESTIONS tag. In our example the following user requirements are de ned: persons, speci es the number of persons living in the household (one to two, three to four, more than four) and maxprice speci es the upper limit regarding the price of the waste disposal plan. Furthermore, emptying represents the sequence in which the dustbins will be emptied, weekly or monthly, and container size, the preferred size of the dust container, 120 or 60.
The third aspect represents the de nition of the constraints. Starting with the &CONSTRAINTS tag in WeeVis di erent types of constraints can be dened. For the rst constraint in our example the &INCOMPATIBLE keyword is used to describe incompatible combinations of requirements. The rst incompatibility constraint describes an incompatibility between the number of persons in the household (persons) and the container size. For example, a waste disposal plan with (container size) 60 must not be recommended to users who live in a household with more than four persons. Filter constraints describe relationships between requirements and items, for example, maxprice pricep, i.e., the price of a waste disposal plan must be equal or below the maximum accepted price. 2.3
A recommendation knowledge base can be represented as a CSP (Constraint
Satisfaction Problem) [
The knowledge base speci ed in Figure 2 can be transformed into a constraint satisfaction problem where &QUESTIONS represents U , &PRODUCTS represents P and &CONSTRAINTS represents P ROD, COM P , and F ILT . Based on this knowledge representation WeeVis is able to determine recommendations that take into account a speci ed set of user requirements. The results collected are represented as unary constraints (R = fr1; r2; :::; rkg). Finally the determined set of solutions (recommended items) is presented to the user. 2.4
In situations where requirements ri 2 R (unary constraints de ned on variables of U such as emptying = monthly ) are inconsistent with the constraints in C, we are interested in a subset of these requirements that should be adapted to be able to restore consistency. On a formal level we de ne a requirements diagnosis task and a corresponding diagnosis (see De nition 1).
De nition 1 (Requirements Diagnosis Task). Given a set of requirements R and a set of constraints C (the recommendation knowledge base), the requirements diagnosis task is to identify a minimal set of constraints (the diagnosis) that has to be removed from R such that R [ C is consistent.
As an example R = fr1 : persons = morethanf our, r2 : maxprice = 600,
r3 : emptying = monthly; r4 : containersize = 60g is a set of requirements
inconsistent with the de ned recommendation knowledge. The recommendation
knowledge base induces two minimal con ict sets (CS) [
Determining con ict sets [
De nition 2 (Repair Task). Given a set of requirements R = fr1; r2; :::; rkg inconsistent with the constraints in C and a corresponding diagnosis R ( = frl; :::; rog), the corresponding repair task is to determine an adaption A = frl0; :::; ro0g such that R [ A is consistent with C.
In WeeVis, repair actions are determined conform to De nition 2. For each diagnosis determined by FastDiag, the corresponding solution search for R [ C returns a set of alternative repair actions (represented as adaptation A). In the following, all solutions that satisfy R [ A are shown to the user (see the right hand side of Figure 1).
Diagnosis determination in FastDiag is based on a total lexicographical
ordering of the customer requirements [
support(p) =
#adaptions in A #requirements in R (1) 3 3.1
We have conducted a performance evaluation with the goal to highlight the
ability of WeeVis to calculate repair actions and if no solutions could be found.
Therefore we set up an experiment with three WeeVis recommenders based
on the e-government example presented in Section 2. To illustrate the
performance of WeeVis, the knowledge base was extended and deployed with di
erent complexity regarding the number of solutions (&PRODUCTS tag in
WeeVis), user requirements (&QUESTIONS tag WeeVis), and constraints
(&CONSTRAINTS tag WeeVis) (see Table 1). According to these three attributes the
knowledge bases were classi ed as Small, Medium, and Large. To t the
attributes of knowledge base Small from Table 1, the running example (see Figure
2) was adapted by adding one question, two products and removing the last two
constraints. The Medium and Large knowledge base are extended versions of the
running example.
To provide an optimal user experience a focus of WeeVis is to provide instant
feedback after every interaction. Interacting with a WeeVis recommender starts
with the the entering of new requirements and the subsequent calculation of
solutions for these requirements. If no solution could be found WeeVis calculates
one or more diagnoses and the complementing alternative products. With this
performance evaluation we show that WeeVis can identify at least one
alternative solution even for large knowledge bases within recommended user interface
response times [
For the rst performance evaluation the goal was to measure the time needed for calculating the corresponding solutions to given requirements. After assigning answers to the questions for the three di erent knowledge bases, the resulting values are depicted in Table 2. The performance values in Table 2 show that for each of the knowledge bases WeeVis identi es solutions fast enough to provide instantaneous feedback from the user interface. If no solution could be found due to inconsistencies between the requirements and the knowledge base, Table 3 shows the time needed to identify at least one alternative solution on the basis of one preferred diagnosis, Table 4 shows the time consumption of calculating all possible solutions. WeeVis is able to calculate either one, two, three or all diagnoses and the corresponding alternative solutions. By taking the response time boundaries for user interfaces into account, the experiment shows that for small and medium knowledge bases it's possible to calculate all minimal diagnoses within acceptable response times (see Table 3). When it comes to large knowledge bases the presented alternative solutions can be reduced to increase the performance of the user interface instead (see Table 4). In this paper we presented WeeVis which is an open constraint-based recommendation environment. By exploiting the advantages of Mediawiki, WeeVis provides an intuitive basis for the development and maintenance of constraintbased recommender applications. The results of our experiment show that due to the integrated direct diagnosis algorithms the WeeVis user interface provides good the response times for common interactive settings.