Profiling knowledge workers using open online profiles

In this paper we investigate the accuracy of user terminology models extracted from open online profiles. In our project, user models are used for the profiling of knowledge workers in order to assist them with information tasks such as email filtering and professional search. We created terminology models from profiles on LinkedIn, Twitter and ArnetMiner (scientific publications) using a term scoring function based on Kullback-Leibler Divergence. The resulting term profiles were evaluated by their owners. Overall, all the models were of reasonable quality, scoring between 0.55 and 0.80 Average Precision. We analyzed the overlap between the models, the subjects’ rating for specificity of the models and the distinction between personal and professional interests. We experimented with the potential of the network context by adding information from connected users to the model. However, this did not improve the quality of the model. In future work, we plan to compare the user models created from online profiles with user models created from local documents in their ability to improve personalized information filtering.