USER: User-Sensitive Expert Recommendations 85 Expert resolution PageRank content network Management Syste Recommender Engine Wab Usage USER Learner Cleaned Fig 3 System architecture of the Class Website Recommendation Systen 5 System Architecture The recommender system architecture was designed to closely reflect a philosophical model of learning. In its initial implementation, the recommender system incorpo rates both the expert and learner portions of the learning model as shown in Figure 3 However, analysis of queries by an expert to a learner is left for future work. The recommender system works by integrating the output of three offline analysis-the expert knowledge and learner queries, and the PageRank of our Web graph-into a set of online recommendations determined by the user's current context. Expert: As an introduction, the website data used to build and test the prototype re- commender system was from the University of Minnesotas College of Liberal Arts StudentServices(class)website(httP://www.class.umn.edu).Thesiteisdeployed and maintained through the use of an in-house content management system(CMS) called Crimson. Crimson distributes the workload of creating and maintaining web- sites to CLASS domain experts; the academic advisors. In this paper, conceptual information is derived from the Web graph and anchor tag text. For every Web page, we construct a term frequency vector using the anchor text from incoming links Each term in this vector is then used as a seed for phrase growth, which builds an array of all possible left-to-right phrase candidates. These candidates are then pruned according to the following criteria: If a phrase occurs only once in the corpus, delete it. If a phrase is a stop word (using the mysQL 4.1 stop word list [26 ), has a string length of l,or delete it If a candidate phrase is a left-side perfect substring of a phrase one word longer than the candidate and their frequencies are equal, delete the candidate. If a candidate phrase is a right-side perfect substring of a phrase one word longer than the candidate and their frequencies are equal, delete the candidateUSER: User-Sensitive Expert Recommendations 85 Fig. 3. System architecture of the CLASS Website Recommendation System 5 System Architecture The recommender system architecture was designed to closely reflect a philosophical model of learning. In its initial implementation, the recommender system incorpo￾rates both the expert and learner portions of the learning model as shown in Figure 3. However, analysis of queries by an expert to a learner is left for future work. The recommender system works by integrating the output of three offline analysis - the expert knowledge and learner queries, and the PageRank of our Web graph – into a set of online recommendations determined by the user’s current context. Expert: As an introduction, the website data used to build and test the prototype re￾commender system was from the University of Minnesota’s College of Liberal Arts Student Services (CLASS) website (http://www.class.umn.edu). The site is deployed and maintained through the use of an in-house content management system (CMS) called Crimson. Crimson distributes the workload of creating and maintaining web￾sites to CLASS domain experts; the academic advisors. In this paper, conceptual information is derived from the Web graph and anchor tag text. For every Web page, we construct a term frequency vector using the anchor text from incoming links. Each term in this vector is then used as a seed for phrase growth, which builds an array of all possible left-to-right phrase candidates. These candidates are then pruned according to the following criteria: • If a phrase occurs only once in the corpus, delete it. • If a phrase is a stop word (using the MySQL 4.1 stop word list [26]), has a string length of 1, or is numeric, delete it. • If a candidate phrase is a left-side perfect substring of a phrase one word longer than the candidate and their frequencies are equal, delete the candidate. • If a candidate phrase is a right-side perfect substring of a phrase one word longer than the candidate and their frequencies are equal, delete the candidate