At a later stage, after Quickstep has bet en runnin new user registers with email address sem99r@ecs soton ac uk. OntoCoPI identifies this email account as that of stuart Table 5. New-user profile for Middleton Middleton, a PhD candidate within the department, and returns Topic the ranked and normalised communities of practise list displayed Al\AgentslRecommender Systems 176 in table 3. This communities of practise list is identified using relations on conference attendance, supervision, authorship, An\Agents\ Mobile Agents 0.77 search interest, and project membership, using the weights 0.4, ANDistributed Systems 7, 0.3, 0.8, and 0.5 respectively. De Roure was found to be the closest person as he is Middletons supervisor, and has one joint Knowledge Technology oNtology publication co-authored with Middleton and Shadbolt. The people nowledge Technology Knowledge Devices with 0.82 values are other supervisees of De Roure. Alani Knowledge Technology Knowledge Management attended the same conference that middleton went to in 2001 Knowledge Technology \Knowledge Management 0.16 Table 3. OntoCoPI results for Middleton Person Relevance Person Relevance DeTour Alani 0.47 Every day Quicksteps profiles are updated and automatically fed back to the ontology, where they are used to populate the research Revill Shadbolt interest relationships of the relevant people 8. EMPIRICAL EVALUATION to evaluate the effect both the ney The communities of practise list is then sent to Quickstep, which initial profiling algorithms have on our integrated system, we searches for matching user profiles. These profiles will be more onducted an experiment based around the browsing behaviour accurate and up to date than those initially created profiles based logs obtained from the Quickstep [18] user trials. The algorithm on publications. Quickstep manages to find the profiles in table 4 previously described are used, as per the example in the previous in its logs. section, and the average performance for all users calculated 8.1 Experimental approach Table 4. Profiles of similar people to Middleton Users were selected from the Quickstep trials whom had entries Person within the departmental publication database. We selected nine Topic Interest users in total, with each user typically having one or two AlDistributed Systems DeRoure Al\Agents\Recommender Systems 0.73 The URL browsing logs of these users, extracted from 3 months Al\AgentslMobile Agent of browsing behaviour recorded during the Quickstep trials, were Revill then broken up into weekly log entries. Seven weeks of browsing Al\Agents\Recommender Systems 0.4 behaviour were taken from the start of the Quickstep trials, and an Knowledge Technology Knowledge empty log created to simulate the very start of the trial Beals Devices Eight iterations of the integrated system were thus run, the first Al\Agents Mobile agents 0.87 simulating the start of the trial and others simulating the following weeks I to 7. Interest profiles were recorded after each iteration 1.8 Two complete runs were made, one with the ' new-system initial Alani Knowledge Technology Knowledge profiling algorithm and one without. The control run without the Management\ CoP new-system initial profiling'algorithm started with blank profiles for each of its users Knowledge TechnologyKnowledge An additional iteration was run to evaluate the effectiveness of the Shadbolt new-user initial profile' algorithm. We took the communities of Al\Agents\Recommender Systems practice for each user, based on data from week 7, and used the new-user initial profile algorithm to compute initial profiles for These profiles are merged to create a profile for the new user, each user as if they were being entered onto the system at t of the trial. These profiles were recorded. Because we are Middleton, using the 'new-user initial profile algorithm with ay early prototype version of Onto CoPl, communities of value of 2.5. For example, Middleton has a publication on confidence values were not available: we thus used co Recommender Systemsthat is I year old and DeRoure, Revill values of 1 throughout this experiment. and Shadbolt have interest in 'Recommender Systems'-this In order to evaluate our algorithms effect on the cold-start (1.0*0.73+0.82*0.4+0.46*1.0)=1.76. Table5 shows the resulting problem, we compared all recorded profiles to the benchmark week 7 profile. This allows us to measure how quickly profiles converge to the stable state existing after a reasonable amount ofAt a later stage, after Quickstep has been running for a while, a new user registers with email address sem99r@ecs.soton.ac.uk. OntoCoPI identifies this email account as that of Stuart Middleton, a PhD candidate within the department, and returns the ranked and normalised communities of practise list displayed in table 3. This communities of practise list is identified using relations on conference attendance, supervision, authorship, research interest, and project membership, using the weights 0.4, 0.7, 0.3, 0.8, and 0.5 respectively. De Roure was found to be the closest person as he is Middleton’s supervisor, and has one joint publication co-authored with Middleton and Shadbolt. The people with 0.82 values are other supervisees of De Roure. Alani attended the same conference that Middleton went to in 2001. Table 3. OntoCoPI results for Middleton Person Relevance value Person Relevance value DeRoure 1.0 Alani 0.47 Revill 0.82 Shadbolt 0.46 Beales 0.82 The communities of practise list is then sent to Quickstep, which searches for matching user profiles. These profiles will be more accurate and up to date than those initially created profiles based on publications. Quickstep manages to find the profiles in table 4 in its logs. Table 4. Profiles of similar people to Middleton Person Topic Interest AI\Distributed Systems 1.2 DeRoure AI\Agents\Recommender Systems … 0.73 AI\Agents\Mobile Agents 1.0 Revill AI\Agents\Recommender Systems … 0.4 Knowledge Technology\Knowledge Devices 0.9 Beals AI\Agents\Mobile Agents … 0.87 Knowledge Technology\Ontology 1.8 Alani Knowledge Technology\Knowledge Management\ CoP … 0.7 Knowledge Technology\Knowledge Management 1.5 Shadbolt AI\Agents\Recommender Systems … 1.0 These profiles are merged to create a profile for the new user, Middleton, using the ‘new-user initial profile’ algorithm with a γ value of 2.5. For example, Middleton has a publication on ‘Recommender Systems’ that is 1 year old and DeRoure, Revill and Shadbolt have interest in ‘Recommender Systems’ – this topics value is therefore 1/1 + 2.5/5 * (1.0*0.73+0.82*0.4+0.46*1.0) = 1.76. Table 5 shows the resulting profile. Table 5. New-user profile for Middleton Topic Interest AI\Agents\Recommender Systems 1.76 AI\Agents\Mobile Agents 0.77 AI\Distributed Systems 0.6 Knowledge Technology\Ontology 0.42 Knowledge Technology\Knowledge Devices 0.37 Knowledge Technology\Knowledge Management 0.35 Knowledge Technology\Knowledge Management\ CoP 0.16 … Every day Quickstep’s profiles are updated and automatically fed back to the ontology, where they are used to populate the research interest relationships of the relevant people. 8. EMPIRICAL EVALUATION In order to evaluate the effect both the new-system and new-user initial profiling algorithms have on our integrated system, we conducted an experiment based around the browsing behaviour logs obtained from the Quickstep [18] user trials. The algorithms previously described are used, as per the example in the previous section, and the average performance for all users calculated. 8.1 Experimental approach Users were selected from the Quickstep trials whom had entries within the departmental publication database. We selected nine users in total, with each user typically having one or two publications. The URL browsing logs of these users, extracted from 3 months of browsing behaviour recorded during the Quickstep trials, were then broken up into weekly log entries. Seven weeks of browsing behaviour were taken from the start of the Quickstep trials, and an empty log created to simulate the very start of the trial. Eight iterations of the integrated system were thus run, the first simulating the start of the trial and others simulating the following weeks 1 to 7. Interest profiles were recorded after each iteration. Two complete runs were made, one with the ‘new-system initial profiling’ algorithm and one without. The control run without the ‘new-system initial profiling’ algorithm started with blank profiles for each of its users. An additional iteration was run to evaluate the effectiveness of the ‘new-user initial profile’ algorithm. We took the communities of practice for each user, based on data from week 7, and used the ‘new-user initial profile’ algorithm to compute initial profiles for each user as if they were being entered onto the system at the end of the trial. These profiles were recorded. Because we are using an early prototype version of OntoCoPI, communities of practice confidence values were not available; we thus used confidence values of 1 throughout this experiment. In order to evaluate our algorithms effect on the cold-start problem, we compared all recorded profiles to the benchmark week 7 profile. This allows us to measure how quickly profiles converge to the stable state existing after a reasonable amount of