Collection Selection with Highly Discriminative Keys Sander Bockting Djoerd Hiemstra Avanade Netherlands B.V. University of Twente Versterkerstraat 6 P.O. Box 217 1322 AP, Almere, Netherlands 7500 AE, Enschede, Netherlands sander.bockting@avanade.com d.hiemstra@utwente.nl ABSTRACT The centralized web search paradigm introduces several prob- lems, such as large data traffic requirements for crawling, index freshness problems and problems to index everything. 1. Query 4. Merged result set Client In this study, we look at collection selection using highly discriminative keys and query-driven indexing as part of a distributed web search system. The approach is evaluated 3. Result 3. Result on different splits of the TREC WT10g corpus. Experimen- tal results show that the approach outperforms a Dirichlet 2. Query 2. Query smoothing language modeling approach for collection selec- tion, if we assume that web servers index their local con- Broker tent. 1. INTRODUCTION The web search approach of major search engines, like Google, Yahoo! and Bing, amounts to crawling, indexing Peer 1 Peer 2 Peer i Peer n-1 Peer n and searching. We call this approach centralized search, because all operations are controlled by the search engines Figure 1: Peers are accessible via a broker to answer themselves, be it from a relatively limited number of loca- queries from clients tions on large clusters of thousands of machines. The cen- tralized web search paradigm poses several problems. The amount of web data is estimated to grow exponen- tially [34]. The changing and growing data requires frequent Figure 1 shows three types of entities: peers, brokers and visits by crawlers, just to keep the index fresh. Crawling clients. We assume that peers are collaborative. Every peer should be done often, but generates a huge amount of traf- runs a search engine enabled web server. The search engine fic, making it impossible to do frequent crawls of all pages. indexes the local website(s), but it may also index other With an estimated four weeks update interval, updates are websites. In this scenario, there can be many millions of performed relatively slow [25, 35]. Also, it is impossible to peers. When a user has an information need, he can pose index everything, as the search engine accessible visible web a query at the client. The client sends the query to the is only a fraction of the total number of web pages [16]. broker. In a response, the broker tries to identify the most Callan [5] identified the distributed information retrieval promising peers to answer the query. This has to be a small problem set, consisting of resource description, resource se- amount of peers, e.g. five to ten peers, so not a lot of traffic lection and result merging. We believe that distributing the is generated. The query is routed to those peers and the search efforts may be an approach to solve the problems de- results are returned to the broker. The broker merges the scribed above. This research focuses on resource description results and sends the merged list to the client. and resource selection [10, 22]. Resource description, i.e. Peers and brokers cooperate to enable brokers to identify indexing of the peers, is distributed; resource selection, or most promising peers. Therefore, every peer sends a small collection selection, is centralized in our research. part of its index to the broker. This part cannot be too small, to still allow for proper judging about the peers’ abil- ity to satisfactory answer queries. The index part cannot be too large due to index data traffic scalability. Techniques have been proposed to manage the index size. Podnar et al [20] used the concept of highly discriminative keys (HDKs) for document retrieval in distributed informa- Copyright c 2009 for the individual papers by the papers’ authors. Copy- tion retrieval. An HDK is a set of terms that is highly ing permitted for private and academic purposes. Re-publication of material discriminative, i.e., that only match a few documents in the from this volume requires permission by the copyright owners. This volume is published by its editors. collection. Because the terms are pre-coordinated (they are LSDS-IR Workshop. July 2009. Boston, USA. combined at index time, not at search time) and because only a few document match all terms in a pre-coordinated tics. Opposed to tf.idf, the probabilities are calculated us- set, the HDK approach is able to very efficiently retrieve the ing document frequencies df and inverse collection frequen- top documents for a query. Although searching can be effi- cies icf (df.icf). The inverse collection frequency is the cient, the HDK indexing process, described in more detail in number of collections that contain the term. An inference Section 3.1, has the negative side-effect that a huge amount network with these properties is called a collection retrieval of highly discriminative keys are generated. To reduce the inference network (cori net). number of keys, Skobeltsyn [32] proposed a query-driven in- Given query q with terms rk , the network is used to obtain dexing strategy that uses caching techniques to adapt to a ranked list of collections by calculating the belief p(rk |ci ) the changing querying behavior of users. The combination in collection ci due to the observation of rk . The collec- of HDK with query-driven indexing allows for completely tion ranking score of ci for query q is the sum of all beliefs distributed document retrieval that in theory grows to web p(rk |ci ), where r ∈ q. The belief is calculated using For- scale proportions [31]. mula 1. In this formula, b and l are constants, cw i is the This paper contributes to the field of distributed informa- number of words in ci , cw is the mean number of words tion retrieval by the applying HDKs and query-driven in- in the collections and |C| is the number of collections. df dexing to select collections, instead of documents. Such an and cf respectively are the number of documents and col- approach would in theory scale the distributed search sce- lections that contain rk . Finally, dt and db respectively are nario described above to millions of peers: The broker lists the minimum term frequency component and minimum be- for every HDK a small number of peers to send the query to, lief component when rk occurs in ci . and the peers retrieve the documents; possibly many. Un- To improve retrieval, the component L is used to scale like a traditional inverted file index that typically consists the document frequency in the calculation of T [6, 23]. L of huge posting lists and a, in comparison, tiny dictionary is made sensitive to the number of documents that contain [36], our HDK index consists of a huge dictionary and, in term rk (using b) and is made large using l. L should be comparison, tiny posting lists. The system is fitted into the large, because df is generally large. Proper tuning of these previously sketched scenario, which allows for control at the two parameters is required for every data set, but deemed broker. This control can for example be used to prevent impossible as the correct settings are highly sensitive to data misuse or to allow for domain-specific search. set variations [12]; the value of l should be varied largely even This paper is organized as follows: the next section dis- when keeping the data set constant while varying the query cusses earlier collection selection methods, Section 3 intro- type. duces our collection selection system and Section 4 describes Further research showed that it is not justified to use stan- the evaluation. The paper concludes with results and con- dard cori as a collection selection benchmark. D’Souza et clusions. al. showed that HighSim outperformed cori in 15 of 21 cases [11]. Si and Callan [28] found limitations with differ- ent collection sizes. Large collections are not often ranked 2. EARLIER WORK ON COLLECTION SE- high by cori, although they often are the most promising LECTION collections. Collection selection systems have been developed to se- lect collections containing documents that are relevant to a user’s query. The generalized Glossary-Of-Servers Server L = l · ((1 − b) + b · cw i /cw ) (gGlOSS) is such a system [13]. It uses a vector space model log (df ) representing index items (document collections) and user T = dt + (1 − dt ) · log (df + L) queries as weight vectors in a high dimensional Euclidean   space to calculate the distance (or similarity) between doc- log |C|+0.5 cf ument collections and queries [24]. I= log |C| + 1.0 Another well-known approach is CVV, which exploits the p(rk |ci ) = db + (1 − db ) · T · I (1) variation in cue validity to select collections [38]. The cue validity CV i,j of query term tj for collection ci measures the extent that tj discriminates ci from the other collec- tions, by comparing the ratio of documents in ci containing 2.2 Language modeling tj to the ratios of documents in other collections containing A language model is a statistical model to assign a proba- tj . The larger the variation in cue validities for collections bility to a sequence of words (e.g. a query) being generated with respect to a term, the better the term is for selecting by a particular document or document collection. Language collections. models can be used for collection selection in distributed This section will describe two collection selection methods search, by creating a language model for each collection [29, in more detail: inference networks and language modeling. 37]. They have also been used for collection selection for other tasks, for instance for blog search [2, 26]. 2.1 Inference networks indri is an improved version of the inquery retrieval sys- cori [7] is a collection ranking algorithm for the inquery tem [33], as it is capable of dealing with larger collections, retrieval system [6], and uses an inference network to rank allows for more complex queries due to new query constructs collections. A simple document inference network has leafs d and uses formal probabilistic document representations that representing the document collections. The terms that oc- use language models. The combined model of inference net- cur in those collections are represented by representation works and language modeling is capable of achieving more nodes r. Flowing along the arcs between the leaves and favorable retrieval results than inquery [18]. Due to these nodes are probabilities based on document collection statis- differences, term representation beliefs are calculated in an- other way than with cori (as described in Section 2.1): Document collections tf r,D + αr P (r|D) = |D| + αr + βr The belief is calculated for representation concept r of docu- ment node D (in document collection C). Examples of rep- Generated resentation concepts are a term in the body or title of a doc- index keys ument. D and r are nodes in the belief network. The term frequency of representation node r in D is denoted by tf r,D . Prune keys with query log αr and βr are smoothing parameters. Smoothing is used (Query-driven Indexing) to make the maximum likelihood estimations of a language Stored model more accurate, which could have been less accurate index keys due to data sparseness, because not every term occurs in a Broker Ranked list of document [39]. Smoothing ensures that terms that are un- ScoreFunction Index Query collections seen in the document, are not assigned zero probability. The results default smoothing model for Indri is Dirichlet smoothing, which affects the smoothing parameter choices [17]. In set- Figure 2: General overview of the indexing and col- ting the smoothing parameters, it was assumed that the like- lection selection system Sophos liness of observing representation concept r is equal to the probability of observing it in collection C, given by P (r|C). This means that αr /(αr + βr ) = P (r|C). The following term frequencies of every word in the collection, without parameter values were chosen: looking at document boundaries in the collection. A term αr = µP (r|C) set is called frequent when it occurs more times than a term set frequency maximum tf max . Every infrequent single term βr = µ(1 − P (r|C)) is added to the peer index together with its frequency count. This results in the following term representation belief, where The frequent keys are stored for further analysis. the free parameter µ has a default value of 2500: The next step consists of obtaining double term set statis- tics. For every frequent term in the collection, frequency tf r,D + µP (r|C) statistics are created for term combinations that consist of P (r|D) = |D| + µ the frequent term and a term that appears after that term within a window size ws. The result is a list of double terms 2.3 Discussion with corresponding frequency counts. Once again, the term The language modeling approach of indri has a better set frequency maximum defines which term sets are frequent formal probabilistic document representation than cori and and will be used for further analysis, and which term sets indri is an improved version of inquery (which is the foun- are infrequent and will be stored in the peer index. This dation of cori). We will use the language model on doc- procedure can be repeated as long as the generated term ument collections as implemented by indri as our baseline sets do not contain more than ws terms, or when a prede- collection selection system. Si et al. [29] showed that a fined maximum number of terms in a term set, hmax , has language modeling approach for collection selection outper- been reached. forms cori. Furthermore, cori outperforms algorithms like Summarizing, the peer index consists of tuples with term cvv and gGlOSS in several studies [9, 21]. sets and corresponding frequency counts. 3.1.2 Broker indexing 3. SOPHOS The infrequent term sets from the peer indices are sent Sophos is a collection selection prototype that uses HDKs to the broker. The broker index contains term sets with to index collections. The keys are used to assign scores to associated sets of collection identifier counters. A collection the collections. Using a scoring function, collection scores identifier counter is a tuple of a collection identifier and a can be calculated to rank collections for a particular query. term set frequency counter. A collection identifier is a short A general overview is depicted in Figure 2. This section de- representation of a collection where the term set occurs. scribes how the broker index is created, explains index size When a term set is inserted into the broker index, it is reduction using a query-driven indexing approach, identi- called a highly discriminative key (HDK). The broker index fies query result parameters, and concludes with a collection will contain a maximum number of collections per HDK, de- ranking formula (ScoreFunction in Figure 2). noted by the collection maximum cm. As soon as the maxi- mum number of collections is reached for a particular HDK, 3.1 Highly discriminative keys the cm collections with the largest term set frequencies will Building the index is done in two phases. First, every be stored in the broker index. peer builds an index of its document collection and sends that index to the broker. Second, the broker constructs a 3.2 Query-driven indexing broker index from all peer indices. A list of popular keys can be created by extracting all unique queries from a query log. Every key that is infrequent 3.1.1 Peer indexing at a peer, and which is present in the unique query list, will Peer indexing starts with the generation of term set statis- be sent to the broker; the other keys are filtered out to reduce tics. First, single term statistics are generated by counting the broker index size and to reduce traffic. c sum of query term set occurrence within one col- calculates a score s for a collection for a given query lections (grouped by sets with h terms)  h #terms in an HDK n−1 s = log10 h−1+ + hmax maximum #terms that HDK can consist of q q #terms in query q c n #distinct query terms found in a collection (hmax +1−h)·(n · α(q−n)  ! h) ·tf max tf max maximum frequency of a term set in a collection /hmax (2) q Table 1: Parameter definitions for query result han- It consists of three components; one component per de- dling sired ranking property. The properties, and corresponding components, are listed below in order of importance. 1. Collections that contain longer HDKs should be ranked higher. Component 1: h − 1. This index pruning strategy was used before by Shokouhi et al. [27]. It is not the most desirable strategy for query- 2. A collection should be ranked higher if it contains more driven indexing, because it is unable to deal with unseen distinct query terms. Component 2: (n − 1)/q. query terms. However, it will give a good idea about the possible index size reduction and the loss of retrieval perfor- 3. More occurrences of query term sets within a collection mance. should result r in a higher collection ranking. Compo- c ·α(q−n) n (hmax +1−h)·( ) h ·tf max nent 3: q . 3.3 Identifying query result parameters The general intuition behind this component is that a collection is more important when it contains more Once the broker index has been built, the system is ready query terms. This is controlled by a damping factor α. to be queried. The broker index contains HDKs with h The term set counts have less impact when they get terms, where h varies from 1 to hmax . In Sophos, hmax is set closer to tf max , because longer keys would be generated to 3. Every key has an associated posting list, which con- for a term set in a collection when its frequency exceeds tains tuples of collection identifiers and counters. A counter tf max . We refer to earlier work for a more detailed indicates the number of occurrences of a certain key within explanation about ScoreFunction [4]. the corresponding collection. The counter value cannot ex- ceed the term set frequency maximum, tf max , as a key would An important detail to mention about querying and rank- otherwise have been locally frequent and new HDKs would ing is that collection X can be found after looking for HDKs have been generated when the maximum number of terms, with length h. When the same collection X is found af- hmax , was not yet reached. ter looking for HDKs with length h − 1, those results are When a user poses a query with q terms, e.g. abcde with discarded as the collection was already found using larger q = 5, the query is first decomposed in query term combi- HDKs. Counts c are only compared with other counts that nations withlength hmax (i.e. abc, abd, . . . , bde, cde). This are retrieved after looking for HDKs of the same length. The results in hq combinations. Each combination is looked up motivation for this behavior is that smaller HDKs tend to in the broker index. Note that this introduces additional have higher counts. load on the broker, but these lookups do not require net- Each of the three components has a share in the collec- work access to the peers. The results of each of those smaller tion score. The component share of a less desired property queries are merged by summing the number of occurrences never exceeds that smallest possible share of a more desired per collection. The sum of all counters, c, has a potential property’s component value. maximum of hq · tf max . It may happen that this procedure results in little or no collections where an HDK occurs. The procedure is then repeated for smaller term sets; first term 4. EXPERIMENT SETUP sets of two terms will be looked up in the index. When even This section describes the corpus, query set and query logs this gives too few results, single terms will be looked up in that were used in the evaluation process, and continues to the index. In the case that one collection contains two differ- describe how the collection selection effectiveness of Sophos ent combinations, e.g. both abc and bce occur in collection was measured. X, the number of occurrences are summed (this is c that was just introduced). However, it also interesting to note that 4.1 Data collections 4 out of 5 query terms can be found in collection X. The number of query terms that can be found using HDKs of a 4.1.1 WT10G corpus splits particular length is indicated by n. The different parameters The Web Track 10GB corpus (WT10g) was developed for are displayed in Table 1. the Text REtrieval Conference1 and consists of 10GB of web pages (1,692,096 documents on 11,680 servers). Compared to regular TREC corpora, WT10g should be more suited 3.4 ScoreFunction: ranking query results for distributed information retrieval experiments, due to the existence of hyperlinks, differences in topics, variation in ScoreFunction, given in Formula 2, is a ranking formula quality and presence of duplicates [3, 8]. that uses the query result parameters to enforce a collection 1 ranking conforming to our desired ranking properties. It http://trec.nist.gov/ Four splits of the WT10G document corpus were made to version of the query log that only contains the actual queries look at the effect of document corpora on collection selec- issued in random order (and none of the other metadata). tion. Every split is a set of collections; every collection is a We also used two query logs that were published by Ex- set of documents. The numbers 100 and 11512 indicate the cite2 that were stripped in the same way. One query log amount of collections in the corpus split. from 1997 contains 1,025,907 queries and another query log from 1999 contains 1,777,251 queries. IP Split: Documents are put in collections based on the IP Finally, a fourth query log with 3,512,320 unique queries addresses of the site where a document was residing. was created by removing all queries from the AOL query log This results in 11,512 collections. that were issued only once. This query log will be referred to as AOL2. The other logs are called AOL, Excite97 and IP Merge 100: A cluster is created by grouping up to 116 Excite99. collections, which results in 100 collections. Grouping is done in order of IP address. This split simulates 4.2 Method the scenario of indexing the search engines that index To evaluate the performance of our collection selection many servers. system, we adopted the precision and recall metrics for col- Random 100 and Random 11512: Two random splits lection selection as defined by Gravano et al. [13]. We start with 100 collections and with 11,512 collections were by obtaining the retrieval system ranking (S ) for a query, created. Documents were randomly assigned to a col- which contains up to 1,000 collections. We also create the lection. The number of 11,512 collections was chosen best ranking (B ) which is the best possible ranking for a to be able to compare a random split with the IP Split. particular query; collections are ranked by their amount of merit with respect to a query. The number of documents in random splits is relatively Given query q and collection c i , the merit within a col- constant, but varies in IP based collections from less than 10 lection can be expressed using merit(q, c i ). The merit of up to more than 1000 documents. This is typical for the size the ith ranked collection in rankings S and B is given by of sites on the Internet; the number of documents per server S i = merit(q, c si ) and B i = merit(q, c bi ). follows a Zipf distribution on the Internet [1]. The IP based The obtained recall after selecting n collections can be splits show signs of conformance to a Zipf distribution [4]. calculated by dividing the merit selected by the best possible This means that the IP based splits can be compared to the retrieval system: Internet in terms of distribution of the number of documents Pn Si and the sizes of the collections. Rn = Pni=1 (3) B i The merit of a collection is the number of relevant docu- i=1 ments in a collection for a particular query. The IP based Precision Pn is the fraction of top n collections that have corpus splits have a larger deviation in merit among the non-zero merit: collections. This contrasts with random splits, which by ap- |{sc ∈ Topn (S)|merit(q, sc) > 0}| proximation have equal merit for each collection [4]. Pn = (4) |Topn (S)| 4.1.2 WT10g retrieval tasks The precision and recall obtained by Sophos is compared Fifty WT10g informational ‘ad-hoc’ queries were used for to the collection selection results from a baseline of Language evaluation (query numbers 501–550). The queries have a Modeling with Dirichlet Smoothing (lmds) as implemented query number, title, description and a narrative description by indri. The baseline system has one parameter µ that of the result that is considered relevant. The title is a small is set to 2500. Krovetz word stemming is applied to the set of words which was used as the query text. The narra- collections. tive descriptions were used by humans to assign relevance Section 3 introduced Sophos and its parameters. There judgments to documents. The relevance judgments can be are three parameters for peers: the maximum key length used to count the number of relevant documents in the col- h max , the maximum key frequency before calling a key fre- lections, which in turn can be used to measure collection quent (tf max ) and the window size ws. Based on num- selection effectiveness. bers used by Luu et al. [15], we use the following settings: There are three types of relevance judgments: not rele- tf max = {250, 500, 750} and ws = {6, 12, 18}. The average vant (0), relevant (1) and authoritative (2). There can be query length on the Internet is 2.3 [14, 30]. We use this ob- multiple authoritative documents in the document corpus servation to set hmax to 3. Setting it smaller would require for a query, but for some queries there are no authoritative many intersections of term sets (with associated collections) documents. All authoritative judgments are converted to 1, to answer queries. Setting it larger would result in many so documents are either relevant or not relevant. This allows term sets that are rarely queried for. For the broker, the for evaluation of the collected merit. collection maximum cm is tested for values 5, 10, 20 and 50. To evaluate Sophos, the collections are processed by re- 4.1.3 Query logs moving stop words – drastically reducing the number of in- AOL published a query log with 21,011,340 queries [19]. valuable keys and speeding up term set generation – and The log has been anonymized and consists of several data applying Porter word stemming. fields: the actual query issued and the query submission date Finally, we will look at the precision and recall with query- and time, and an anonymous user ID number. The release driven indexing using four different query logs. By pruning of the anonymized data set was controversial at the time the keys from the peer indices that do not occur in a query because it was proven possible to link an ID to a real person. 2 To respect the anonymity of the users, we used a stripped http://www.excite.com/ log. At the same time, we will look at the number of term 6e+007 Sophos with QDI AOL sets (keys) in the broker index to get an idea about its size. Sophos with QDI AOL2 #ColID Counters in global index 5e+007 Sophos with QDI Excite 97 Sophos with QDI Excite 99 4e+007 5. RESULTS 3e+007 5.1 Index size 2e+007 Figure 3 shows the number of collection identifier coun- 1e+007 ters within the broker index for different indexing settings of indexing the IP Split. Spread over single, double and 0 Single terms Double terms Triple terms triple term set collection identifier counters, the number of counters are a good indication for the actual broker index size. The figure shows that the number of counters decreases Figure 5: Number of collection identifier counters when the term set frequency maximum is increased. stored per QDI strategy for IP Split Single terms 5.2 Collection selection performance Double terms #ColID Counters in global index Triple terms Table 2 shows the average precision and recall over 50 108 queries that were calculated with Formula 3 and Formula 4. The numbers are calculated for four different corpus splits with which the baseline (lmds) and Sophos were tested. Sophos was used with the following settings: query-driven indexing with the AOL query log, tf max = 250, cm = 20 and ws = 6. Due to memory constraints, we were unable to run Sophos with a window size larger than 6. The table 107 shows that the baseline outperforms Sophos on the Random tf2 tf2 tf2 tf2 tf5 tf5 tf5 tf5 tf7 tf7 tf7 tf7 11,512 corpus split, but Sophos outperforms the baseline on 50 50 50 50 00 00 00 00 50 50 50 50 cm cm cm cm cm cm cm cm cm cm cm cm the IP split. 5 1 2 5 5 1 2 5 5 1 2 5 0 0 0 0 0 0 0 0 0 This is displayed in more detail in Figure 6, which shows the average recall of Sophos and the baseline after selecting Figure 3: Number of collection ID counters with IP n collections. Sophos was tested using different query logs, Split. tf max = 250 and cm = 50. Interestingly, pruning with the smallest Excite query logs results in the best recall figures, A more substantial reduction of collection identifier coun- possibly because the queries were logged at the same time ters – of roughly 70% – can be achieved by using query- as when the corpus was crawled. driven indexing, as shown in Figure 4. The figure shows the number of collection identifier counters after indexing the 1 LMDS Random 11,512 corpus split with or without query-driven in- Sophos with QDI AOL tf250 cm50 ws6 Sophos with QDI AOL2 tf250 cm50 ws6 dexing. Figure 5 depicts the obtainable index size reduction Sophos with QDI Excite 97 tf250 cm50 ws6 Sophos with QDI Excite 99 tf250 cm50 ws6 by using different query logs. The Excite query logs contain 0.8 significantly less query term sets than the AOL query log. Recall after selecting n collections The figure shows that more keys are pruned from the peer indices, resulting in a smaller broker index. The figure shows 0.6 that using Excite query logs instead of the standard AOL query log can result in roughly 75% less collection identifier counters. 0.4 4.5e+007 Sophos without QDI 4e+007 Sophos with QDI AOL 0.2 #ColID Counters in global index 3.5e+007 3e+007 2.5e+007 0 1 10 100 2e+007 Number (n) of selected collection 1.5e+007 1e+007 Figure 6: Recall of different collections selection 5e+006 methods on IP Split 0 Single terms Double terms Triple terms 6. CONCLUSIONS Figure 4: Number of collection identifier counters We introduced the collection selection system Sophos that stored per QDI strategy for Random 11512 corpus uses highly discriminative keys in peer indices to construct split a broker index. The broker index contains keys that are good discriminators to select collections (or peers). To limit Corpus split Collection selection method P@1 P@10 P@20 P@50 R@1 R@10 R@20 R@50 Random 100 lmds 0.290 0.251 0.249 0.217 0.314 0.435 0.526 0.683 Sophos QDI AOL tf250 cm20 0.330 0.254 0.237 0.208 0.379 0.436 0.493 0.644 Random 11512 lmds 0.140 0.096 0.084 0.069 0.233 0.196 0.186 0.198 Sophos QDI AOL tf250 cm20 0.040 0.036 0.037 0.026 0.067 0.073 0.082 0.073 IP Merge 100 lmds 0.280 0.202 0.188 0.154 0.489 0.489 0.567 0.755 Sophos QDI AOL tf250 cm20 0.300 0.254 0.214 0.160 0.211 0.485 0.626 0.825 IP Split lmds 0.170 0.110 0.083 0.056 0.070 0.149 0.183 0.289 Sophos QDI AOL tf250 cm20 0.170 0.147 0.121 0.091 0.280 0.466 0.466 0.548 Table 2: Average precision and recall over 50 queries after selecting n collections, high scores shown in bold the number of keys transferred to the broker, and to reduce [3] P. Bailey, N. Craswell, and D. Hawking. Engineering a the broker index size, we employed query-driven indexing to multi-purpose test collection for web retrieval only store the keys that are queried for by users. Similar experiments. Inf. Process. Manage., 39(6):853–871, studies were performed for document retrieval [15], but to 2003. the best of our knowledge, we are the first to apply this [4] S. Bockting. Collection selection for distributed web approach for collection selection. search. Master’s thesis, University of Twente, Feb. Precision and recall was measured using 50 queries on 2009. the WT10g TREC test corpus and compared to a state- [5] J. Callan. Distributed information retrieval. Advances of-the-art baseline that uses language modeling with Dirich- in Information Retrieval, pages 127–150, 2000. let smoothing (lmds). The results showed that our system [6] J. Callan, W. B. Croft, and S. M. Harding. The outperformed the baseline with the IP split as test corpus. inquery retrieval system. In Proc. of the 3rd This is promising, because the IP based splits most closely International Conference on Database and Expert resemble the information structure on the Internet. lmds Systems Applications, pages 78–83, Valencia, Spain, was better capable of selecting information in random based 1992. Springer-Verlag. splits, because it stores all available information about the [7] J. Callan, Z. Lu, and W. B. Croft. Searching collections. In random based splits, relevant documents (and distributed collections with inference networks. In their corresponding terms) are mixed over many collections, SIGIR ’95: Proc. of the 18th annual international making it hard for our approach to select highly discrimi- ACM SIGIR conference on Research and development native keys that can discriminate collections with relevant in information retrieval, pages 21–28, New York, NY, documents. USA, 1995. ACM. Query-driven indexing is required to keep the broker index [8] N. Craswell, P. Bailey, and D. Hawking. Is it fair to size manageable; a 70% index size reduction can be obtained evaluate web systems using trec ad hoc methods. In by pruning keys using the AOL query log, another 75% re- Workshop on Web Evaluation. SIGIR’99, 1999. duction is possible by using a smaller query log. Our results on the IP split showed that pruning using the smaller Excite [9] N. Craswell, P. Bailey, and D. Hawking. Server query logs resulted in higher precision and recall than with selection on the world wide web. In DL ’00: Proc. of AOL query logs. The use of any query log resulted in higher the 5th ACM conference on Digital libraries, pages precision and recall than the baseline results. This motivates 37–46, New York, NY, USA, 2000. ACM. further research using more advanced query-driven indexing [10] D. D’Souza, J. Thom, and J. Zobel. A comparison of strategies, such as described by Skobeltsyn [32], to further techniques for selecting text collections. In ADC ’00: reduce the index size while improving the performance. It Proceedings of the Australasian Database Conference, would also be interesting to make tf max depending on the page 28, Washington, DC, USA, 2000. IEEE collection size. Computer Society. [11] D. D’Souza, J. A. Thom, and J. Zobel. Collection selection for managed distributed document Acknowledgements databases. Information Processing & Management, We are grateful to the Yahoo Research Faculty Grant pro- 40(3):527–546, May 2004. gramme and to the Netherlands Organisation for Scientific [12] D. D’Souza, J. Zobel, and J. A. Thom. Is cori effective Research (NWO, Grant 639.022.809) for supporting this work. for collection selection? an exploration of parameters, We would like to thank Berkant Barla Cambazoglu and the queries, and data. In Proc. of the 9th Australasian anonymous reviewers for their valuable comments that im- Document Computing Symposium, pages 41–46, Dec. proved this paper. 2004. [13] L. Gravano and H. Garcia-Molina. Generalizing 7. REFERENCES GlOSS to vector-space databases and broker hierarchies. In International Conference on Very Large [1] L. A. Adamic and B. A. Huberman. Zipf’s law and the Databases, VLDB, pages 78–89, 1995. internet. Glottometrics, 3:143–150, 2002. [14] T. Lau and E. Horvitz. Patterns of search: analyzing [2] J. Arguello, J. L. Elsas, J. Callan, and J. G. Carbonell. and modeling web query refinement. In UM ’99: Proc. Document representation and query expansion models of the 7th international conference on User modeling, for blog recommendation. In Proc. of the 2nd Intl. pages 119–128, Secaucus, NJ, USA, 1999. Conf. on Weblogs and Social Media (ICWSM), 2008. Springer-Verlag New York, Inc. M. Moricz. Analysis of a very large web search engine [15] T. Luu, F. Klemm, M. Rajman, and K. Aberer. Using query log. SIGIR Forum, 33(1):6–12, 1999. highly discriminative keys for indexing in a [31] G. Skobeltsyn. Query-driven indexing in large-scale peer-to-peer full-text retrieval system. Technical distributed systems. PhD thesis, EPFL, Lausanne, report, TR: 2005041, EPFL Lausanne, 2005. 2009. [16] P. Lyman and H. R. Varian. How much information, [32] G. Skobeltsyn, T. Luu, I. Podnar Zarko, M. Rajman, 2003. http://www.sims.berkeley.edu/ and K. Aberer. Query-driven indexing for scalable how-much-info-2003, retrieved on April 23, 2008. peer-to-peer text retrieval. Future Generation [17] D. Metzler. Indri retrieval model overview, July 2005. Computer Systems, 25(1):89–99, June 2009. http://ciir.cs.umass.edu/~metzler/ [33] T. Strohman, D. Metzler, H. Turtle, and W. B. Croft. indriretmodel.html, retrieved on January 20, 2008. Indri: A language model-based search engine for [18] D. Metzler and W. B. Croft. Combining the language complex queries. In Proc. of the International model and inference network approaches to retrieval. Conference on Intelligence Analysis, 2004. Information Processing and Management, [34] C. Tang, Z. Xu, and S. Dwarkadas. Peer-to-peer 40(5):735–750, 2004. information retrieval using self-organizing semantic [19] G. Pass, A. Chowdhury, and C. Torgeson. A picture of overlay networks. Proc. of the 2003 conference on search. In InfoScale ’06: Proc. of the 1st international Applications, technologies, architectures, and protocols conference on Scalable information systems, page 1, for computer communications, pages 175–186, 2003. New York, NY, USA, 2006. ACM. [35] Y. Wang and D. J. DeWitt. Computing pagerank in a [20] I. Podnar Zarko, M. Rajman, T. Luu, F. Klemm, and distributed internet search system. In Proc. of the K. Aberer. Scalable peer-to-peer web retrieval with International Conference on Very Large Databases highly discriminative keys. IEEE 23rd International (VLDB), Aug. 2004. Conference on Data Engineering (ICDE), 2007. [36] I. H. Witten, A. Moffat, and T. C. Bell. Managing [21] A. L. Powell and J. C. French. Comparing the Gigabytes: Compressing and Indexing Documents and performance of collection selection algorithms. ACM Images. Morgan Kaufmann, 1999. Trans. Inf. Syst., 21(4):412–456, 2003. [37] J. Xu and W. B. Croft. Cluster-based language models [22] D. Puppin, F. Silvestri, and D. Laforenza. for distributed retrieval. Proc. of the 22nd annual Query-driven document partitioning and collection international ACM SIGIR conference on Research and selection. In InfoScale ’06: Proc. of the 1st development in information retrieval, pages 254–261, international conference on Scalable information 1999. systems, page 34, New York, NY, USA, 2006. ACM. [38] B. Yuwono and D. L. Lee. Server ranking for [23] S. E. Robertson, S. Walker, S. Jones, M. M. distributed text retrieval systems on the internet. In Hancock-beaulieu, and M. Gatford. Okapi at trec-3. In Proc. of the 5th International Conference on Database TREC-3 Proc., pages 109–126, 1995. Systems for Advanced Applications (DASFAA), pages [24] G. Salton and M. J. McGill. Introduction to Modern 41–50. World Scientific Press, 1997. Information Retrieval. McGraw-Hill, Inc., New York, [39] C. Zhai and J. Lafferty. A study of smoothing NY, USA, 1986. methods for language models applied to ad hoc [25] N. Sato, M. Udagawa, M. Uehara, Y. Sakai, and information retrieval. In SIGIR ’01: Proc. of the 24th H. Mori. Query based site selection for distributed annual international ACM SIGIR conference on search engines. Distributed Computing Systems Research and development in information retrieval, Workshops, 2003. Proc.. 23rd International pages 334–342, New York, NY, USA, 2001. ACM. Conference on, pages 556–561, 2003. [26] J. Seo and W. B. Croft. Umass at trec 2007 blog distillation task. In Proc. of the 2008 Text REtrieval Conference. NIST, 2007. [27] M. Shokouhi, J. Zobel, S. Tahaghoghi, and F. Scholer. Using query logs to establish vocabularies in distributed information retrieval. Information Processing and Management, 43(1):169–180, 2007. [28] L. Si and J. Callan. Relevant document distribution estimation method for resource selection. In SIGIR ’03: Proc. of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, pages 298–305, New York, NY, USA, 2003. ACM. [29] L. Si, R. Jin, J. Callan, and P. Ogilvie. A language modeling framework for resource selection and results merging. Proc. of the 11th international conference on Information and knowledge management, pages 391–397, 2002. [30] C. Silverstein, H. Marais, M. Henzinger, and