imilarity between the remaining profile terms are computed using our bipartite graph pproaches(Equations 3, 4, and 5). More details follow Given two user profiles u1 and u2, the intersecting profile parts are denoted as uy and u such that terms(ui)=terms(u2)= terms(un)terms(u2). Th overlapping profile parts are denoted as ui and i2 such that terms(ui)=terms(un)\ terms(u2)and terms(ui2)=terms(un)\ terms(u1). The combined size of the two profiles is denoted as N= terms(u1)l+terms(u2). The size of the intersecting profile parts is N= terms(u1)l terms u2). The size of the non-overlapping profile parts is N= terms(ui)+ terms(ui2) The compound similarity measure based on simpath(Equation 3)is (x1,u2)×N (ui1,2)×N The compound similarity measure based on simeupath(Equation 4)is simc.= simcos(u1,12)XN+simeupath(ui, ti2)xN The compound similarity measure based on simeuhalf(Equation 5)is simcos(u1, u2)xN+simeuhalf(u1, ti2) 6 Evaluation and results We evaluate the different algorithms described in the previous section in the context of expert finding. We use an inhouse-built software called Profile Builder to generate expert profiles using techniques described in [7] to create profiles by analysing the documents such as web pages visited by the expert). Both the Bow(word profiles)and BOC(terms are Wikipedia concepts: Wiki profiles)representations of the experts are generated by the profile builder software. An expert finding query is correspondingly in the form of either a Bow or a BOC. For a given query profile, matching expert profiles are determined by computing similarity between the expert profile and the query profile. Measure Description COS-Word Cosine similarity measure between expert and query bow profiles(Equation r iterations of Bi-path Compound similarity measure after graph spreading as defined in equation 6 Bi-EU Compound similarity measure. raph spreading as defined in Equation 7 Similarity measure after grap ing as defined in E Table 1: Glossary of the Similarity Measures A pilot study conducted as a part of the evaluation process interviewed 10 partic ipants with expertise in different fields of computer science research. From each of the participants, 5 to 10 documents that in the participants opinion best describe their research were collected. Along with the documents, the participants were asked to give 5 keywords for each of their document that in their opinion best described the ince these keywords somewhat described the expertise of the participants. used by the participants to provide two similarity judgments. We believe thissimilarity between the remaining profile terms are computed using our bipartite graph approaches (Equations 3, 4, and 5). More details follow. Given two user profiles u1 and u2, the intersecting profile parts are denoted as u 0 1 and u 0 2 such that terms(u 0 1 ) = terms(u 0 2 ) = terms(u1) ∩ terms(u2). The remaining non￾overlapping profile parts are denoted as uˆ1 and uˆ2 such that terms( ˆu1) = terms(u1) \ terms(u2) and terms( ˆu2) = terms(u2) \ terms(u1). The combined size of the two profiles is denoted as N = |terms(u1)| + |terms(u2)|. The size of the intersecting profile parts is N0 = |terms(u 0 1 )| + |terms(u 0 2 )|. The size of the non-overlapping profile parts is Nˆ = |terms( ˆu1)| + |terms( ˆu2)|. The compound similarity measure based on simpath (Equation 3) is (6) simC path = simcos(u 0 1, u0 2) × N 0 + simpath( ˆu1, uˆ2) × Nˆ N The compound similarity measure based on simeupath (Equation 4) is (7) simC eupath = simcos(u 0 1, u0 2) × N 0 + simeupath( ˆu1, uˆ2) × Nˆ N The compound similarity measure based on simeuhalf (Equation 5) is (8) simC euhalf = simcos(u 0 1, u0 2) × N 0 + simeuhalf ( ˆu1, uˆ2) × Nˆ N 6 Evaluation and Results We evaluate the different algorithms described in the previous section in the context of expert finding. We use an inhouse-built software called Profile Builder to generate expert profiles using techniques described in [7] to create profiles by analysing the documents (such as web pages visited by the expert). Both the BOW (word profiles) and BOC (terms are Wikipedia concepts; Wiki profiles) representations of the experts are generated by the profile builder software. An expert finding query is correspondingly in the form of either a BOW or a BOC. For a given query profile, matching expert profiles are determined by computing similarity between the expert profile and the query profile. Measure Description COS-Word Cosine similarity measure between expert and query BOW profiles (Equation 1) COS-Con Cosine similarity measure between expert and query BOC profiles (Equation 1) COS-5n Mean cosine similarity between BOC profiles after 5 iterations of set spreading COS-10n Mean cosine similarity between BOC profiles after 10 iterations of set spreading Bi-PATH Compound similarity measure after graph spreading as defined in Equation 6 Bi-EU Compound similarity measure after graph spreading as defined in Equation 7 Bi-EUby2 Compound similarity measure after graph spreading as defined in Equation 8 SAN Similarity measure after graph spreading as defined in Equation 2 Table 1: Glossary of the Similarity Measures A pilot study conducted as a part of the evaluation process interviewed 10 partic￾ipants with expertise in different fields of computer science research. From each of the participants, 5 to 10 documents that in the participant’s opinion best describe their research were collected. Along with the documents, the participants were asked to give 5 keywords for each of their document that in their opinion best described the document. Since these keywords somewhat described the expertise of the participants, they were used by the participants to provide two similarity judgments. We believe this approach