"... In PAGE 19: ... In other words any query routing technique could be applied in our setting in order to define the top-k candidate objects. Table3 compares HT-p2p with dominant approaches in top-k query processing for distributed networks. HT-p2p and the approach of [11, 17], are the only ones that are adapted to super-peer p2p networks and, as such, take advantage of the heterogeneity of peers and of a topology that reduces bandwidth by eliminating the transferred messages between peers.... ..."

"... In PAGE 6: ... QFmin was set to 1, 3, 5 and 1, where QFmin = 1 means that no key is activated and only the basic single term index is used to process the queries. Table1 shows the achieved precisions at K (P@K). The highest value in each line of the table is highlighted in bold.... In PAGE 6: ... However, we believe this should not be a problem in the context of Web search where users are usually only in- terested in the top 10-20 documents. In addition, for K gt;20, Table1 also shows that, with a higher value for DFmax, our system is becoming similar to ST-BM25 (in fact, if DFmax = jDj, our system is ex- actly equivalent to ST-BM25). In the worst case, when DFmax = 100 (we only keep top-100 documents in the post- ing lists) and QFmin = 1 (the query driven mechanism is not applied), our system retrieves 75% of the relevant doc- uments retrieved by ST-BM25 at top-50 (0.... ..."

"... In PAGE 10: ... There are very few occurrences of \photographic quot; under keyword, while all occurrences are under dataset. Table2 shows the results of our experiment. For each value of k, we report the speedup obtained through our algorithm, measured as the ratio of the time taken to fully execute the query on the database to the time taken by our algorithm.... ..."

"... In PAGE 8: ...94 As search engines demands for quick response, the computation time is an important factor for consideration. The average number of keyframe pair comparison for top k re-ranking over 24 queries is listed in Table6 . Compared to fast re-ranking with global signatures and time duration, the hierarchical method is more expensive.... ..."

"... In PAGE 7: ...Table 1: Precision results of the top k term-based algorithm for k = 10 and varying epsilon1. Table1 summarizes the results obtained by the top k term-based algorithm, for long and short queries, where k is fixed to be 10 and where epsilon1 is varied. As the results show, the algorithm can achieve high levels of pruning even for modest values of epsilon1.... ..."

"... In PAGE 5: ... special case of Gapped Top K (i.e. when the gap equals to 0). We do experiments varying the gap to test whether a non-zero gap can perform better than Top K. The results on the HARD data set and AP88-89 data set are shown in Table1 , where we show the MAP, the precision at 10 documents, and the number of judged relevant documents per query. From the results, we can see Top K (gap = 0) is clearly not the best strategy.... In PAGE 5: ...3 Comparison of Different Algorithms Since the effectiveness of the underlying feedback mechanism ( the mixture model method in our case) is an important factor that may affect our evaluation, we compare several different feedback algorithms with the non-feedback baseline in Table 3. The perfor- mance for the Gapped Top K and the K Cluster Centroid is the best performance from Table1 and Table 2, respectively. From these results, we can see that the performance of both ac- tive feedback and pseudo feedback are better than that of baseline retrieval.... ..."

"... In PAGE 3: ... It merges docu- ments and clusters based on a first-in first-out ba- sis. Table1 describes the data structures needed to pro- cess the algorithms. Each of these lists can be rep- resented as a simple linear vector.... ..."