reduces the subjectivity in judging similarity and gives us more realistic values for icipant was asked to judge the similarity between their profile other profiles. Additionally, each of the participants judged the similarity between every pair of profiles(third person view ). The mean of the subjective judgments provided by the participants were used as the base/reality values to evaluate our similarity measures The comparison of the computed similarity value with the reality values were actually made across all user pairs. However, for evaluating the algorithms in the context of expert finding, we consider a user q to represent the query profile and evaluate similarity results of user pairs(q, a)where a is every other user(experts) Comparison of Precision and Recall PRecision F-measure Co:. SAN L+ D D- LHU L C0s-sr (a) Precision and Recall observations b) F-measure Observations Fig 3: Effectiveness of Similarity Measures for Expert Search(Threshold-based Match) We first evaluate the effectiveness of our similarity measures in the context of short-listing a group of experts(eg: for recruitment interview). ere the selected expert profiles are those that exceed a pre-determined similarity threshold. We repeat the search for 10 query profiles over the derived expert profiles using all the approaches listed in Table 1. Figure 3 shows the results from our candidate search process where we measured precision, recall, and F-measure* of all the approaches. In short, precision represents the fraction of the correctly determined experts from those selected by our algorithms(based on how many of the matched results are the experts we wanted to get) Recall represents the effectiveness of the algorithm to find all experts(based on how many experts did we miss). F-measure is a compound measure, basically a harmonic mean of precision and recall. Higher values are better. From Figure 3, we are able to make the following observations. All the metrics based on bipartite graph mapping(Bi-*)work very well over the standard cosine similarity measurement techniques(Cos-Word and Cos-Con) The accuracy of set-based measures increases with the increase in the number of spreading iterations(Cos-1On performs much better than Cos-5n) The precision of all our approaches are almost equal while the recall varies. Our algorithms show significant improvements in recall when compared with the stan- dard approaches Our approaches Bi-*and Cos-10n exhibit upto 20% improvement. The recall of our Bi-PATH approach is 100% while Bi-EUby2 and Cos-1On approaches exhibit around 90% recall Spreading with 5 iterations(Cos-5n)is almost equal performance to other path based/reachability conditions for termination in a general semantic search approach (SAN). This may be suggestive of the maximum diameter of the relevant subgraph consisting of the user's concept We use the standard definitions of Precision, Recall and F-measure as defined in [8reduces the subjectivity in judging similarity and gives us more realistic values for comparison. Every participant was asked to judge the similarity between their profile and other profiles. Additionally, each of the participants judged the similarity between every pair of profiles (third person view). The mean of the subjective judgments provided by the participants were used as the base/reality values to evaluate our similarity measures. The comparison of the computed similarity value with the reality values were actually made across all user pairs. However, for evaluating the algorithms in the context of expert finding, we consider a user q to represent the query profile and evaluate similarity results of user pairs (q, x) where x is every other user (experts). (a) Precision and Recall Observations (b) F-measure Observations Fig. 3: Effectiveness of Similarity Measures for Expert Search (Threshold-based Match) We first evaluate the effectiveness of our similarity measures in the context of short-listing a group of experts (eg: for recruitment interview). Here the selected expert profiles are those that exceed a pre-determined similarity threshold. We repeat the search for 10 query profiles over the derived expert profiles using all the approaches listed in Table 1. Figure 3 shows the results from our candidate search process where we measured precision, recall, and F-measure4 of all the approaches. In short, precision represents the fraction of the correctly determined experts from those selected by our algorithms (based on how many of the matched results are the experts we wanted to get). Recall represents the effectiveness of the algorithm to find all experts (based on how many experts did we miss). F-measure is a compound measure, basically a harmonic mean of precision and recall. Higher values are better. From Figure 3, we are able to make the following observations. – All the metrics based on bipartite graph mapping (Bi-*) work very well over the standard cosine similarity measurement techniques (Cos-Word and Cos-Con). – The accuracy of set-based measures increases with the increase in the number of spreading iterations (Cos-10n performs much better than Cos-5n). – The precision of all our approaches are almost equal while the recall varies. – Our algorithms show significant improvements in recall when compared with the stan￾dard approaches. Our approaches Bi-* and Cos-10n exhibit upto 20% improvement. – The recall of our Bi-PATH approach is 100% while Bi-EUby2 and Cos-10n approaches exhibit around 90% recall. – Spreading with 5 iterations (Cos-5n) is almost equal performance to other path￾based/reachability conditions for termination in a general semantic search approach (SAN). This may be suggestive of the maximum diameter of the relevant subgraph consisting of the user’s concepts. 4 We use the standard definitions of Precision, Recall and F-measure as defined in [8]