Evaluating interest profiles based on users' judgment, interest change, and class specificity in the context of filtering medical documents
✍ Scribed by Luz M. Quiroga; Javed Mostafa
- Publisher
- Wiley (John Wiley & Sons)
- Year
- 2005
- Tongue
- English
- Weight
- 732 KB
- Volume
- 39
- Category
- Article
- ISSN
- 0044-7870
No coin nor oath required. For personal study only.
✦ Synopsis
Abstract
Quality of profiles is critical for the performance of information filtering systems. A series of experiments were conducted to investigate factors that can impact the profile acquisition process. Profiles in MedSIFTER, the system used in the study, are represented as degrees of interests over a set of fixed classes and for this research were based on established MeSH categories. We collected data from 12 subjects, who used the system in a mode where profile acquisition is exclusively based on feedback provision. User‐provided profiles, machine‐built profiles, transaction log records and written statements of subjects were the main empirical data. Profile similarities and consistency of feedback with explicit preferences for classes were measured. Results suggested that the system captured user preferences better when these preferences involved specific classes and when the user preferences didn't change significantly. Feedback assessments were more conducive to the representation of user preferences when these preferences involved specific classes. Discrepancies between what the subjects stated, when explicitly expressing their preferences, and what they actually did, when providing feedback, pointed to key factors related to the representation and modeling of the documents and profiles.