Section: New Results
Content-Based Publish/Subscribe System for Web Syndication
Participants: Zeinab Hmedeh CNAM, Harry Kourdounakis (FORTH-ICS, Vassilis Christophides, Cedric du Mouza (CNAM), Michel Scholl (CNAM), and Nicolas Travers (CNAM)
Content syndication has become a popular way for timely delivery of frequently updated information on the Web. Today, web syndication technologies such as RSS or Atom are used in a wide variety of applications spreading from large-scale news broadcasting to medium-scale information sharing in scientific and professional communities. However, they exhibit serious limitations for dealing with information overload in Web 2.0. There is a vital need for efficient real-time filtering methods across feeds, to allow users to effectively follow personally interesting information.
To efficiently check whether all keywords of a subscription also appear in an incoming item (i.e., broad match semantics), we need to index the subscriptions. Count-based (CI) and tree-based (TI) are two main indexing schemes proposed in the literature for counting explicitly and implicitly the number of contained key-words. The majority of related data structures cannot be employed for conjunctions of keywords (rather than attribute-value pairs) due to the space high-dimensionality. In this paper, we are interested in efficient implementations of both indexing schemes using inverted lists (IL) for CI and a variant for distinct terms of ordered tries (OT) for TI and study their behavior for critical parameters of realistic web syndication workloads. Although these data structures have been employed to evaluate broad match queries in the context of selective information dissemination and sponsored search or for mining frequent item sets, their memory and matching time requirements appear to be quite different in our setting. This is due to the peculiarities of web syndication systems which are characterized 1) by information items of average length (25?36 distinct terms) which are greater than advertisement bids (4?5 terms) and smaller than documents of Web collections (12K terms) and 2) by very large vocabularies of terms (up to 1.5M terms). Note also that due to broad match semantics, information retrieval techniques for optimizing ILs (e.g., early pruning) are not suited in our setting.
We present analytical models for memory requirements and matching time and we conduct a thorough experimental evaluation to exhibit the impact of critical parameters of realistic web syndication workloads. We found that for small vocabularies, POT matching time is one order of magnitude faster than the best IL (RIL), while for large vocabularies (like the one used on the Web), RIL outperforms the matching POT, which uses almost four times more memory space. The actual distribution of term occurrences has almost no impact on the size of the three indexing structures while it significantly affects the number of nodes that need to be visited upon matching something that justifies OT performance gains. The smaller the subscription length, the larger the OT factorization gain w.r.t. IL and the larger the rank of the term from which the OT substructure degenerates to an IL.