Experiences with complex user profiles for approximate P2P community matching

Transcription

1 Experiences with coplex user profiles for approxiate PP counity atching Patrizio Dazzi ISTI-CNR Pisa, Italy Matteo Mordacchini IIT-CNR Pisa, Fabio Baglini University of Pisa Pisa, Italy Abstract The proble of defining PP overlays where peers characterized by siilar interests are directly connected is currently an iportant research issue. We have recently proposed a two layer PP architecture where the first layer exploits a gossip algorith for the detection of counities of peers characterized by siilar interests, while the second layer defines a DHT storing the profiles of the counities detected in the first layer. The DHT is exploited by peers joining the syste to find out a counity atching its interests. This paper investigates a DHT based approach supporting a siilarity based search of user profiles. Our approach exploits Locality Sensitive Hashing to support the siilarity indexing. The paper investigates several types of profiles to odel user interests and evaluates the indexing echaniss of the DHT with respect to the different types. Experiental evaluation has been conducted by considering a real data set. Keywords-distributed clustering; peer-to-peer; user profiles; I. INTRODUCTION The first recoender syste, dealing with artists and albus, based on the collaborative filtering technique has been proposed in [11]. After this contribution several tools have been proposed and realized, both fro acadeic counities and copanies. These solutions achieve a rearkable success for their ability to odel the behavior of huan exchange inforation by word-of-outh, which is a typical behavior of people sharing one or ore interests. Most of these systes are based on centralized architectures, an approach that on one hand siplifies the overall design, but on the other one presents a set of liitations. These liitations include the reduced scalability, the existence of single-point-of-failure and the concentration of data published by the users at a single authority which is potentially able to access to the whole set of data. The ost proising alternatives to current centralized systes are based on peer-to-peer (PP) architectures. PP architectures allow the ipleentation of systes able to dynaically scale and to be resilient to failures of groups of nodes, even significant ones. Moreover, their exploitation avoid the concentration of inforation in a single entity. Most of recoender systes, both centralized and PP, group hoogeneous users in counities. This siplify the exchange of pertinent recoendations aong users sharing coon interests. The creation of a PP syste able to support the aggregation of users in counities, identified on the basis of coon interests, has been the subject of a nuber of recent research works carried out by different research groups. These groups belong to different scientific counities and the proposed works start fro different points of view. However, the coon idea of all the proposed works is to associate with each user a profile, odeling user s interests, to exploit in order to aggregate the in counities. In a precedent position paper of ours [6], we proposed a syste structured according to two layers. The first one supports the creation and the aintenance of self-eerging user counities, by perforing a continuous exchange of inforation aong users. The other layer is devoted to the aintenance of a distributed index of existing counities. It should allow users to cheaply (in ters of transferred data) discover counities pertinent to their interests, even when does not exist any counity that exactly atches its profile. In this paper we ai at realizing such a distributed index. To allow it to support the entioned features, the index we propose is PP, supports siilarity searches and supports different ways to represent the user profile. We started fro a Distributed Hash Table (DHT). This kind of network provide the functionalities of a distributed data index, with an effective load balancing strategy, a careful anageent of the disconnection of the nodes and a logarithic coplexity routing both in the nuber of hops required to resolve a query and in the size of the routing tables. However, DHTs are not designed to support siilarity searches. Indeed, as with non-distributed hash tables, each value is calculated using an identifier associated with the application of a function designed to distribute the results, avoiding to preserve the locality of the original data. Therefore, in order to achieve the functionalities required by our envisioned distributed index, DHTs have to be extended with echaniss that allow to perfor not-exact searches. To do this, we exploit a probabilistic approach based on the use of Locality Sensitive Hash Functions (LSH), proposed in [13] to index the user profiles on a DHT. This leads to a drastically reduced resources usage. LSH approxiates the exact solution of the proble, issuing a very reduced aount

2 of queries while aintaining very high quality results. Beside supporting siilarity searches on DHTs, our distributed index is able to represent the interests of users in different ways. In particular, in this paper we present a study on three different characterizations of the user interests. The study is aied at finding a good trade-off between the accuracy and the network overhead. The three different types of profiles are based on sets of attributes defined in spaces ade of an arbitrary nuber of diensions and coparable by using siple siilarity etrics. This allows to build profiles fro objects of any kind (e.g. text stored by users, log browsing web, etadata about the edia, sets of coordinates about places visited, etc.). The three types of profiles are, respectively, a vector of ites, an adjacency atrix of ites occurrences and a set of connected coponents extracted profile by profile fro the adjacency atrixes. The index we propose has been studied fro the point of view of coplexity and evaluated experientally using a real dataset, consisting of a list of articles of edical literature consulted by users of PubMed [1], where each publication is identified by title, keywords and abstracts. Fro this dataset were extracted profiles fro over 700 users. The results obtained by searching for siilarities on our proposed DHT index has been copared against the ones obtained for the sae queries, indexing the profiles of individual attributes. This evaluation was carried out using various configurations of LSH indexing functions, for all types of profiles. The reaining of this paper is organized as follows: Section II reviews the state of the art in the field, Section III present the overall architecture of the syste describing the two layers it is ade of, with a particular focus on the indexing layer. Section IV describes both the three kind of profiles we propose in this paper. Section Vpresent the coplexity, both in ters of network usage and storage space required. Section VI shows the experiental results we obtained. Finally, in Section VII we derive our conclusions. A. P-RING II. RELATED WORK P-Ring [5] is a PP index structure that supports range queries. P-Ring is designed to support fault-tolerance and provides logarithic search perforance even for highly skewed data distributions. It supports large sets of data ites for each peer. The authors experientally tested P-Ring by using siulations as well as by using a real case of study they deployed on PlanetLab. According to theauthors achieveents, P-ring provide perforance coparable with Skip Graphs, Online Balancing and Chord. B. MAAN The Multi-Attribute Addressable Network, MAAN [4] is a structured PP syste targeted to the definition of a support for resource discovery in a large distributed systes that provides support for ulti-attribute range queries. It odifies the Chord hashing function by a locality preserving hashing function. In MAAN resources are identified by a set of attribute-value pairs, and each attribute is apped on a Chord ring through the locality preserving function. The node target of the hashing function stores the full resource description so that a resource is stored as any ties as its nuber of attributes. The resolution of a ulti-attribute range query consists in executing a single 1-diensional query on the doinant (i.e. ost selective) attribute, while the other attributes are checked using the replicated data. One of the ain drawbacks of MAAN is related to the large aount of eory required to store resource indices, and the cost to update the. C. Web Retrieval PP with high selective keys In this contribution [9] the authors foralize an indexing and retrieval odel designed to achieve high perforance, cost-efficient retrieval by exploiting highly discriinative keys (HDKs) stored in a distributed global index aintained in a structured PP network. HDKs are carefully selected ters that appear only in a sall set of collected docuents. The authors present both a theoretical analysis of the scalability of the odel and experiental results achieved by using the HDK-based PP retrieval engine. These results deonstrated that the total traffic generated exploiting the HDK approach is saller than the one obtained with distributed single-ter indexing strategies. The authors also perfored a deep scalability analysis that deonstrated that the HDK approach is able to scale to large networks of peers indexing very large docuent collections. D. Efficient seantic search on DHT overlays This paper [13] starts stating that distributed hash tables (DHTs) excel at exact-atch lookups, but they do not directly support coplex queries such as seantic search that is based on content. In this paper, the authors propose an approach to support seantic search on DHT overlays. Their idea is to place, with high probability, indexes of seantically close files into sae peer nodes by exploiting inforation retrieval algoriths and locality sensitive hashing. This approach adds only index inforation to peer nodes, causing an acceptable storage overhead. Via network siulations, the authors show that their approach is viable because the nuber of nodes visited for a query is about 100, independently fro the overlay size.

3 III. THE OVERALL ARCHITECTURE As in [6] the syste we propose (depicted in Figure 1) consists in a two-layered peer-to-peer (PP), fully decentralized architecture. This choice is due to a series of considerations about the syste goals. First of all, such systes allow to propose a service without any centralized authority (e.g., a single server that would store all user profiles), with the service ipleented through the collaboration of the peers. Figure 1. Overall Syste Architecture Moreover, PP systes allow to address the iportant issue of scalability, as they do not require the over- and proportional-provisioning of resources that would be required with a centralized approach. The ore peers participate, the ore power is added to the syste. If carefully designed, a PP approach scales up to hundred of thousands peers. Finally, PP systes are known to deal gracefully with syste dynais at no or very reduced additional cost, whereas centralized systes need expensive and coplex techniques to ensure continuous operation under node and link failures. Starting fro the peer-to-peer approach, we envision a two-layered solution. This solution ai is two-fold. Indeed, it is designed to support both a highly dynaic participation, aggregations of users and a stable and scalable aintenance of the description of the counities the users belong to. The lower layer is in charge for the foration processes of the user counities. It has to support the construction of a interest-based network, so that peers are effectively grouped with other peers that share siilar interest in their various interest doains. This requires to construct and aintain a representation of these interests (user profiling), to copare these profiles to deterine their siilarity (siilarity etrics), to propose distributed algoriths that cluster peers in interest-based groups based on this etric (clustering algorith) and to give a representation to the counities. Possible solutions for the constructions of this layer are beyond the scope of this paper, actually we are investigating possible solutions like [10] or []. The upper layer is a distributed index structure (we also refer as backbone layer). It has the responsibility to store the signature of each counity as well as the contact points needed to establish a connection with that counity. This allow new peers to rapidly find, aong the registered counities, the best ones that atch their interests. However, an index able to find only perfect atches between users and their interests could not be as effective as needed in a very dynaic and potentially heterogeneous environent. As an exaple consider a peer that has in coon with a certain counity the 90% of its characterizing ites. Even if they do not share all the interests, they are probably enough siilar to benefit fro inforation exchange flows. Hence, in order to provide a effective solution, the distributed structure ust be able to support the detection of the counities sharing a certain nuber of ites with the new joining peer. To this end the index structure have to provide a proper support for siilarity based searches. IV. THE PROPOSED SOLUTION As we introduced before (see Section I), a viable solution for obtaining a support to siilarity searches on DHTs has been proposed in [13]. In that paper the authors applied the in-wise independent perutations approach to user profiles built according the Vector Space Model (VSM). The VSM odel (also known as ter vector odel) is an algebraic odel for representing text docuents as vectors of identifiers. Docuents and queries are represented as vectors of ters. Each diension corresponds to a separate ter. If a ter occurs in the docuent, its value in the vector is nonzero. This ethod is widely used due to its effectiveness in situations where a query should be copared against a given docuent to easure its relevance. The ain drawback of this approach is its need to have a coplete knowledge on the existing ters in the considered set of docuents. This is not a big proble in centralized systes that can access to the whole set of docuents and, consequently, can built the whole dictionary of existing ters but is unfeasible in peer-to-peer systes where there is not a global knowledge on existing docuents and ters. In order to address this proble, soe coplex solutions have been presented in literature [1]. However, even sipler solutions, both fro the ipleentation and coputational coplexity point of view, deonstrated to achieve pretty good results [7]. One of the is the application of the

4 Jaccard siilarity coefficient to the ter vectors, that also the authors of [13] exploited in their work. Despite of its siplicity, the Jaccard Index is a very flexible easure and can be exploited to copare set of eleents without specific knowledge about spatial structure of data. In order to exploit this indexing echanis in the envisioned scenario, and pursuing the ai of finding an effective and efficient way to handle interest inforation, in this paper we propose three different types of profiles to odel user interests. All of the are build starting fro the collections of attributes that characterize a user. The three profile types are: Weighted attribute vector: this profile consists of collections of attributes properly weighted according to a certain nuerical weight indicating the attribute relevance with respect to the user profile. Attribute Adjacency Matrix: in this case the user profile is represented with a weighted attribute vector enriched with values estiating the correlation between attributes. The adjacency atrixes are squared and syetrical. Set of connected coponents extracted fro Attribute Adjacency Matrix: the profile is represented with a set of connected coponent obtained by applying to the Attribute Adjacency Matrix an algorith that extract a set of connected coponents. Each coponent describes a specific user interest. In order to analyze the advantages deriving fro the usage of all types of profiles, we perfored both a theoretical and experiental coparison between two different syste configurations. In the first one, we use the profiles in a DHT exploiting LSH based on the in-wise independent perutations. The second configuration consists a naive solution. As a naive solution we ean one like [4] or [3]. In those approaches, profiles are stored attribute per attribute and the query resolution process is perfored by finding all the atches for every attribute and by intersecting the. Obviously, this probabilistic ethod can lead to inexact query resolution but the network traffic is bounded and the storage space required is only a fraction with respect to the per attribute indexing. V. THEORETICAL ANALYSIS Tables I, II and III show the coparison between the two indexing approaches when applied to index the user profiles respectively in the for of Weighted attribute vector, Attribute Adjacency Matrix and Set of connected coponents extracted fro Attribute Adjacency Matrix. In this Analysis U and C indicate respectively the nuber of attributes coposing the profile of each user and each counity. A represents the nuber of access points for each counity. X indicates the nuber of nodes of the DHT, Co the nuber of existing counities. k is the axiu nuber of profiles requested to each node when it resolves a query. s indicates the axiu nuber of counity profiles following fro a store request for a new profile. R corresponds to the nuber of reoved attributes following to a profile update. When exploiting LSH indexing, n represents the nuber of identifiers associated with each profile. When the profiles are structured as Set of connected coponents I indicates the nuber of attributes describing the l -th user interest with P I apple U, whereas Int indicates the nuber of interests of a user. VI. EVALUATION The validation of the proposed solution was conducted using a real dataset of users sharing a subset of the PubMed [1] scientific publications. The dataset is part of a dataset released by the Mendeley group [8]. It contains inforation about 5,177 users and 11,878 docuents, with an average of 48.8 docuents consulted by each user. The profile of each user is characterized by the ters occurring in the title, keywords and abstracts of the papers she has accessed. In order to have a ore precise characterization of the users, the whole set of docuents contain 17,49 distinct ters have been filtered to eliinate the stop-words, getting 17,075 real ters. Finally, through a process of steing, 107,57 roots were extracted. The final set of,754 users exploited for the following evaluation was obtained by selecting the users with at least 150 root ters in their profiles. Vector-based user profiles are obtained directly by these ters. Adjacency atrixes are obtained by using the cooccurrences frequencies of ters within each user collection. Finally, the profiles built according the connected coponents odel have been extracted staring fro the adjacency atrices by identifying the connected coponents, obtained through the application of a pruning technique on the relationships occurring with a frequency less than a certain threshold infreq and selecting, aong the obtained coponents, those with a size greater than or equal to a certain, fixed value insize. The following tests ai to stress the ability of the proposed syste in finding a right counity for a user that is joining the network. We used a network of,000 nodes, each denoting a different counity representative. Users coing for the dataset are grouped in 300 different counities, obtained by using K-Means. The centroids obtained so far are used as counity representatives. Moreover, the syste is perturbed by the insertion of 1,700 other profiles, generated randoly in the sae vector space of the user profiles. These counities, identified using K-Means, have been used as references to copare against the ones we identified by exploiting the peer-to-peer approach. This has been perfored for all three kinds of profiles.

5 We perfored a series of tests by injecting,500 queries to the syste. Each query is coposed of a user profile. For each query, the syste returns as an answer, the ost siilar counity descriptor registered. In order to evaluate the perforances of our solution, we evaluated both its efficacy and efficiency. Efficacy is easured using both network requireents for query resolution and network and storage consuption for aintaining the distributed index structure. and network bandwidth, respectively, to create and to aintain the distributed index. Figure 4 depicts the global aount of network bandwidth consuption both to issue queries to the distributed index and to retrieve the results. In both the figures the results are related, respectively, to profiles built according to a vector of ters, an adjacency atrix and a set of connected coponents. 1e e Ter Vectors Adj. Matrixes Coponents LSH N=15 M=1 LSH N=15 M= LSH N=15 M=3 LSH N=0 M=1 LSH N=0 M= LSH N=0 M=3 Figure. Average difference between the ost siilar counity and the actual one Figure shows the average difference (Jaccard distance) between the ost siilar counity found by each query and the counity identified by the centralized ethod. As can be observed, the usage of co-occurrences data highly enhance the perforance of the syste. In particular, the connected coponents ethod achieves the best results. In this case, ethods with ore than one hashing function sees to perfor better than the others. On the other hand, such achieveents should be copared with the bandwidth and storage needs required to aintain indexes based on ore coplex profiles, like atrixes and connected coponents. 1e Figure 4. Ter Vector - Naive Ter Vector - LSH 0 ID Ter Vector - LSH 15 ID Query routing Adj. Matrixes - Naive Adj. Matrixes - LSH 0 ID Adj. Matrixes - LSH 15 ID Query resolution Coponents - Naive Coponents - LSH 0 ID Coponents - LSH 15 ID Network Bandwidth consuption for Query anageent It can be noticed that, even in this case, the connected coponents ethod is the ost perforing one, in every kind of etrics we considered. Moreover, we copared those results with the ones obtained by a naive approach. It can be observed that every profile approach obtains a considerable gain using the LSH indexing ethod with respect to the pure naive solution. VII. CONCLUSIONS This paper has presented a DHT based approach for storing the profiles of counities of users characterized by siilar interests. The siilarity search over the DHT exploits the Min Wise Independent Perutation approach. We have proposed several types of user s profiles and evaluated the indexing echaniss with respect to the different kind of profiles. Experiental results conducted by considering a real data set have shown the effectiveness of our approach ACKNOWLEDGEMENT We acknowledge the support of the S-CUBE (EU-FP ), CONTRAIL (EU-FP ) and RECOGNI- TION (EU-FP ) EU projects. 10 Ter Vector - Naive Ter Vector - LSH 0 ID Ter Vector - LSH 15 ID Index Maintenance Adj. Matrixes - Naive Adj. Matrixes - LSH 0 ID Adj. Matrixes - LSH 15 ID Overall Index Size Coponents - Naive Coponents - LSH 0 ID Coponents - LSH 15 ID Figure 3. Network Bandwidth and Storage Space required by the distributed index Figure 3 shows the global size in ters of storage space REFERENCES [1] Pubed central. [] Ranieri Baraglia, Patrizio Dazzi, Matteo Mordacchini, Laura Ricci, and Luca Alessi. Group: A gossip based building counity protocol. In Proc. of the 11th International Conference on Next Generation Wired/Wireless Networking, to appear.

6 [3] A.R. Bharabe, M. Agrawal, and S. Seshan. Mercury: supporting scalable ulti-attribute range queries. In ACM SIGCOMM Coputer Counication Review, volue 34, pages ACM, 004. [4] Min Cai, Martin Frank, Jinbo Chen, and Pedro Szekely. Maan: A ulti-attribute addressable network for grid inforation services. In Journal of Grid Coputing, page 184. IEEE Coputer Society, 003. [5] Adina Crainiceanu, Prakash Linga, Ashwin Machanavajjhala, Johannes Gehrke, and Jayavel Shanugasundara. P-ring: an efficient and robust pp range index structure. In Proceedings of the 007 ACM SIGMOD international conference on Manageent of data, SIGMOD 07, pages 3 34, New York, NY, USA, 007. ACM. [6] P. Dazzi, P. Felber, L.B. Anh, L. Leonini, M. Mordacchini, R. Perego, M. Rajan, and E. Riviere. Peer-to-peer clustering of web-browsing users. In C. Lucchese, G. Skobeltsyn, and W. G. Yee, editors, Proceedings of the 7th Workshop on Large-Scale Distributed Systes for Inforation Retrieval, co-located with ACM SIGIR 009, pages 71 78, 009. [7] Wa el Musa Hadi, Fadi A. Thabtah, and Hussein Abdel-jaber. A coparative study using vector space odel with k-nearest neighbor on text categorization data. In World Congress on Engineering, pages , 007. [8] Kris Jack, Jaes Haerton, Dan Harvey, Jason J Hoyt, Jan Reichelt, and Victor Henning. Mendeleys reply to the datatel challenge. Procedia Coputer Science, 1():1 3, 010. [9] Ivana Podnar, Martin Rajan, Toan Luu, Fabius Kle, and Karl Aberer. Scalable peer-to-peer web retrieval with highly discriinative keys. In ICDE, pages , 007. [10] Lakshish Raasway, Bugra Gedik, and Ling Liu. A distributed approach to node clustering in decentralized peer-topeer networks. IEEE Transactions on Parallel and Distributed Systes, 16:814 89, 005. [11] Upendra Shardanand and Pattie Maes. Social inforation filtering: algoriths for autoating word of outh. In Proc. of the SIGCHI conference on Huan factors in coputing systes, pages 10 17, New York, NY, USA, ACM, Addison-Wesley. [1] Chengxiang Zhai and John Lafferty. A study of soothing ethods for language odels applied to inforation retrieval. ACM Trans. Inf. Syst., :179 14, April 004. [13] Yingwu Zhu and Yiing Hu. Efficient seantic search on dht overlays. J. Parallel Distrib. Coput., 67(5): , 007. Type Cost Query Naive O( U log(x)) LSH O(n U log(x)) Query resolution Naive O(k C U ) LSH O(k C n) Counity insertion Naive O( C log(x)) LSH O(n C log(x)) Profile update Naive O( C log(x)+r log(x)) LSH O(n C log(x)+r log(x)) Descriptor reoval Naive O( C log(x)) LSH O(n log(x)) Index size Naive O( C Co) LSH O(n C Co) Table I COMPLEXITY ANALYSIS FOR INDEXING PROFILES STRUCTURED AS TERM VECTORS. Type Cost Query Naive O( U 3 log(x)) LSH O(n U log(x)) Query resolution Naive O(k C U ) LSH O(k C n) Counity insertion Naive O( C 3 log(x)) LSH O(n C log(x)) Profile update Naive O( C 3 log(x)+r log(x)) LSH O(n C log(x)+r log(x)) Descriptor reoval Naive O( C log(x)) LSH O(n log(x)) Index size Naive O( C 3 Co) LSH O(n C Co) Table II COMPLEXITY ANALYSIS FOR INDEXING PROFILES STRUCTURED AS ADJACENCY MATRIXES. Query Type Cost Naive O(( P I ) log(x)) LSH O(n Int P I log(x)) Query resolution Naive O(k C ( P I ) LSH O(k n Int C ) Counity insertion Naive O( C log(x)) LSH O(n C log(x)) Profile update Naive O( C log(x)+r log(x)) LSH O(n C log(x)+r log(x)) Descriptor reoval Naive O( C log(x)) LSH O(n log(x)) Index size Naive O( C Co) LSH O(n C Co) Table III COMPLEXITY ANALYSIS FOR INDEXING PROFILES STRUCTURED AS CONNECTED COMPONENTS.