Adding a Trust Layer to Semantic Web Metadata

Adding a Trust Layer to Semantic Web Metadata Paolo Ceravolo, Ernesto Damiani, Marco Viviani Scenario „ „ „ „ In a cooperative scenario of metadata production, members of a community of peers freely generate metadata, potentially containing low-quality or even inconsistent assertions. Trustworthiness values computed on the basis of users' feedback can improve metadata quality by filtering out untrustworthy assertions. In our approach, community members express their opinions on the trustworthiness of each assertion, and the outcome is aggergate to obtain a community-wide assessment Global or local trustworthiness thresholds are applied to eliminate some assertions to improve the metadatabase's overall consistency. Use of metadata in open environments: drawbacks „ „ „ Limited traceability: Metadata can be generated by a number of sources (the data owner, other users) and may or may not be digitally signed by their author. Diversity: Metadata produced by members of a diverse community of peers tend to contain sets of mutually inconsistent assertions. Inconsistency: Metadata can also be generated by automatic metadata generators, whose error rates are not negligible. So: „ Metadata have non-uniform trustworthiness. The proposed solution „ „ „ We leave it to community members to express their views on the trustworthiness of each assertion in the metadata base. We aggregate individual trustworthiness values to obtain a community-wide assessment of each assertion. Finally, we apply a trustworthiness threshold to eliminate some assertions, hopefully reducing the overall inconsistency of the metadatabase. Aims: „ An architecturally-neutral and secure solution for collecting user votes, capable of working at various levels of anonymity (see E. Damiani et al.: Managing and Sharing Servents' Reputations in P2P Systems. IEEE Trans. Knowl. Data Eng. 15(4): 840-854 (2003)) „ A suitable aggregation function to compute community-wide trust values on metadata (this work). Example: A Centralized Architecture „ „ Our approach computes the level of trust of an assertion as the aggregation of multiple fuzzy values representing trustworthiness resulting from human interactions with metadata assertions. ‰ A Metadata Publication Center (MPC) collects and displays metadata assertions, possibly in different formats and coming from different sources. ‰ „ A measure of metadata quality rather than risk Can be adapted to a “Pure” Peer-to-Peer scenario: each peer has a local MPC The MPC will assign different trustworthiness values to assertions depending on their origin: assertions manually provided by a domain expert are much more reliable than assertions automatically generated and submitted by a crawler. Centralized Architecture - 2 „ „ The metadata in the MPC are indexed and a group of Clients interacts with the metadata by navigating them and providing implicitly (with their behavior) or explicitly (by means of an explicit vote) an evaluation about metadata trustworthiness. This trust-related information is passed by the MPC to the Trust Manager. The TM is composed of two modules: ‰ ‰ „ Trust Evaluator: examines metadata and evaluates their reliability; Trust Aggregator: aggregates all inputs coming from the Trust Evaluator clients according to a suitable aggregation function (the WOWA operator). The Trust Manager is the computing engine behind the MPC module thatprovides interested parties with a visual overview on the metadata reliability distribution. Architecture - Summary The voting interface „ „ „ Semantic navigators access data via metadata Explicit vote: user clicks on a button to say that metadata assertion was relevant to the data Implicit vote: the user clicks on the metadata to access data Choice of the aggregation operator „ „ „ „ A simple arithmetic average would perform a rough compensation between high and low values. Weighted mean aggregates trust values from different sources, taking into account the reliability of each source. The OWA operator weights trust values in relation to their size, without taking into account which sources have expressed them. The Weighted OWA operator (WOWA) combines the advantages of both the OWA operator and the weighted mean. The WOWA operator „ WOWA uses two sets of weights of dimension n: ‰ p ‰ w ‰ ‰ corresponds to the relevance of the sources, corresponds to the relevance of the values. p = [p1, p2, …, pn] w = [w1, w2, …, wn] Σpi = 1 Σwi = 1 The WOWA operator – Definition „ A mapping fWOWA: Rn Æ R is a WOWA operator of dimension n if: where is a permutation of {1, 2, …, n} such that weights wi are defined as with w* a monotonic function interpolating the points with the point (0, 0). Selecting the WOWA operator „ „ „ „ Given two weigth vectors p and w Given a data vector a Let S = According to definition we must define the function w* interpolating S. Two ways: ‰ ‰ we first define vector w and then function w* is established; we first define function w*. An example „ „ For the sake of simplicity,we assume that array w is composed of values of the form wi = k/n The following two cases are considered: ‰ ‰ Diffident: variable k ranges from 1 to n (fig. (a)) Confident: variable k ranges from n to 1. (fig. (b)) The diffident approach – 1 „ „ „ „ „ a = [.9 .7 .5 .3 .1] p = [1/15 4/15 2/15 5/15 3/15] w = [1/5 2/5 3/5 4/5 1] wn = [1/15 2/15 3/15 4/15 5/15] (already ordered) (normalized) We have to find the function w* interpolating the points The diffident approach – 2 The diffident approach – 3 „ The equation of the interpolation function is: Trustworthiness Metadata „ „ „ Standard Assertions Refer to original assertions as objects (reification) Do not impose any format for original assertions, other than having a URI Conclusion „ „ „ Algorithm+Votes aggregation = Trustworthiness Assertions A system for computing “Trustworthiness views” over metadata ‰ A measure of metadata quality Compared with some probabilistic approaches with encouraging results (First results see E. Damiani et al., Assessing Efficiency of Trust Management in Peer-to-Peer Systems, Proc. Of 1st International Workshop on Collaborative Peer-toPeer Information Systems)