The Recommendation Mechanism in an Internet Information System with Time Impact Coefficient

Transcription

1 The ecommendation Mechanism in an Internet Information System with Time Impact Coefficient arisz Król, Michał Szymański, Bogdan Trawiński Institte for Applied Informatics Technical University of Wrocław, Poland Abstract: In this paper we propose two generic mechanisms implemented in a cadastre internet information system. The first one is the list of last qeries sbmitted by a given ser and the second one is the list of page profiles recommended to a ser. The idea of page recommendation is based on the concept of a page profile which represents a system option, type of retrieval mechanisms and search criteria. The calclation of rank vales for page profiles is based on the sage freqency and the time impact coefficient. A recommended page is selected by a ser from a list facilitates and accelerates his searches by moving him directly to the chosen option page with search form filled with the most expected criteria vales. As an additional complementary mechanism the list of last sbmitted qeries is available to each ser. Keywords: information filtering, recommendation systems, ser assessment. 1 Introdction e to the growth of the Web systems both in the research and commercial area, the sers impose new methods for predicting their needs. Systems that adapt its behavior to their sers are called ser adaptive systems [5, 1]. One of the ways of helping the ser is to adapt the interface elements. The other way is looking for similarities between a new ser and past sers [1]. Adaptation to the individal ser is flfilled by personalization [8]. Starting in the late 199s personalization became very poplar. First personalized services are based on static qestionnaires that sers fill ot in order to make se of new capabilities. Nowadays we se machine learning techniqes for adaptation [4]. The personalization process sally consists of (a the collection and pre-processing of Web data, inclding content data, strctre data, sage data and ser profile data, (b the analysis and discovery of correlations between sch data, (c the determination of the recommendation methods for hyperlinks [11], qeries [3], prodcts [9] and ser interface [17]. The means to analyse the Web data listed in [12] inclde demographic filtering (F, collaborative filtering (CF, content-based filtering (CBF, case-based reasoning (CB, rle-based filtering (BF, Web mining (WM and hybrid approaches (HA. F systems se the information stored in the profile that contains many different demographic attribtes. CF systems collect visitor opinions on a set of obects, sing ratings provided 1

2 by the sers to predict a particlar ser s interest in an item. The ratings may be implicit or explicit. The main idea is to compare the ser model of an active ser with the previos ser in order to clster similar sers. Many stdies have shown that CF methods prodce recommendations with seflly high accracy. Improving the performance by optimizing featre weight sing genetic algorithm is presented in [14]. CB paradigm based on past experiences allows soltion adaptation that leads to finegrained tning of historical soltions toward the new one. CBF systems track ser s behavior and take descriptions of the known content to learn the relationship between a ser and new items. In BF sers are asked to answer a set of qestion and a reslt is cstomized for their needs. WM specifies three domains: Web content mining, Web strctre mining and Web sage mining. The last domain is the most poplar in the area of recommendation [19]. Web sage mining, also known as Web log mining, aims to discover ser patterns from the data stored in server logs or browser logs while srfing the Web system. The mined knowledge can improve the design of Web pages, and develop adaptive sage scenarios more efficiently and effectively. A hybrid approach to recommendations combines aspects of F, CF, CB, CBF and WM. Users evalate the docments and provide feedback for the system and the system knows more. The trend towards increasing profitability makes voice-browsing very promising approach to deliver proper content for devices sch as mobile phones and PA. Voice web service presented in [13] combines lingistic knowledge, Voice XML and Web ontology into a personalized recommendation system. Traditional recommendation techniqes identified in [12] sch as non-personalized, attribte based, item-to-item correlation and people-to-people correlation have been applied for Web recommendations. The method proposed in the paper [11] based on direct and indirect association rles ses information abot ser behavior to local Web pages. The rles discovered in this process estimate helpflness of one page to make ranking lists of all visited pages. In papers [9, 2] the athors present a model which ses the visiting time and freqency of pages withot considering the access order of page reqests in se sessions. To captre the relationships between pages they extract information from log data. Knowledge discovery sally is exected by periodically mining new contents of the log files and can be smmarized in the following steps: pre-process logs to extract ser sessions, smmarize the session in terms of ser profiles, and create context associations from ser profiles. In sch environment, fzzy reasoning is also a good framework for the recommendation process. The approach proposed in [16] is fit for real-time recommendation. Similar approach is presented in [15] with a two-step recommendation system, which ses specific UL-predictor neral networks. ecommendation techniqes still address several problems and do not represent easy road to sccess [18]. They reqire more sophisticated methods of acqiring ser needs. It is becoming harder to design a system sitable for all sers and contexts. Open systems sffer from a drawback in which sers srf via a proxy, and their identities are anonymos. Finally, it is often tricky to prove empirically ability of adaptation [8]. It is a very reasonable qestion to ask whether or not the recommendation will actally improve the system. In this work we focs on Web sage mining by ser profiling, content and log analysis to recommend qeries profiles. It is organized as follows. In Section 2 we 2

3 otline the proposed method of recommendation. Section 3 discsses the fnctionalities of the real estate cadastre system and the reqirements of its design. Then we focs on the pages strctre in the ISEG2-INT system in Section 4. Section 5 and 6 introdce some theoretical and implementation details while Section 7 states the conclsions. 2 Otline of the recommendation method 2.1 Internet system and ser model The method is designed for an internet information system which provides for its sers different retrieval mechanisms based on definite nmber of search criteria. The system is bilt in form of an option tree with search forms located at different levels of that tree (see Figre 1. The sers of the system are the members of an organization e.g. a local government or a corporation and flfil their everyday dties freqently se information obtained from the system. For some days they focs on specific topics and after completing one task they move to another one and change their topics of interest. The crcial assmption for the recommendation method presented in the article is that the sers for some time perform searches arond similar or the same topics. The main goal of the method is to prompt the sers the page profiles calclated on the basis of sage freqency of qeries and time when the searches were condcted. When the list of recommended page profiles is available in the main men page, it can facilitate and accelerate ser searches by moving him directly to the chosen option page with search form filled with the most expected criteria vales. Main men Option 1 Option 2 Option i Option N Option 11 Option 12 Option 1K... Search forms Search forms Search form Type 1 Type 1 Type 2 Type 1 Type 2 Type 3 Type 1 Figre 1 General model of an internet information system 2.2 escription of the method The idea of page recommendation is based on the concept of a page profile, which concerns the pages with search forms. The page profile is characterized by the option of the system containing a retrieval mechanism, search criteria and finally by the type of that mechanism (e.g. simplified or extended. A formal model of page profiles can be presented as follows. For representing page profiles we se a finite set O of system options and a finite set C of search criteria, where C ={C 1, C 2,, C N } and a finite set T of retrieval mechanism types. The page profile is defined as a tple P= <O, C 1, C 2,.., C N, T>, where 3

4 O = {o 1, o 2,, o No } a set of system options, where N o is the nmber of options, C i = {c i1, c i2,, c inci } a set of vales of the i-th search criterion, where N ci is the nmber of vales of that criterion, T = {t 1, t 2,, t Nt } a set of types of retrieval mechanisms, where N t is the nmber of types. So the nmber of page profiles is eqal to N p =N o *N c1 *N c2 * *N cn *N t. All criteria occrring in all search forms are taken into accont, becase some of them may be fond in several or even in all forms. The mechanism analyses previos qeries inpt by a given ser into the system and takes into accont the freqency and time of each qery element sage, it calclates the rank vale of each page profile. The rank vale of page profiles is atomatically calclated for each ser separately. In order to lower the contribtion of qery elements sed earlier, the factor 1-d/ called time impact coefficient has been introdced (see Figre 2, where is an identifier of the ser, d is the nmber of expired days since the moment of calclation, is the nmber of days taken into accont for the ser. The vale of shold be determined by each ser as the component of his set of preferences. d d 1 = 1 = Figre 2 Time impact coefficient definition The day of rank recalclation ank vale r (p of a page profile p sed by ser can be calclated in the following way 1 d r (p = f d (p (1 (1 d= where f d (p is the sage freqency of a page profile p by ser and d days back before the day of rank calclation. ank vales of page profiles can be modified by rank vales of options, selected set of search criteria and types of retrieval mechanisms compted in a very similar way bt for all ses, irrespectively of any page profile. Modifying factors seem to be sefl when rank vales of page profiles are eqal or small, then the most freqently sed options or search criteria can play greater role. ank vale r (o of an option o sed by ser within the period of days can be determined sing following formla 1 d r (o = f d (o (1 (2 d= where f d (o is the sage freqency of option o by ser and d days back before the day of rank calclation. ank vale r (c i of the -th vale of i-th criterion sed by ser can be considered as follows 4

5 d r (3 1 (ci = f d (ci (1 d= where f d (c i is the sage freqency of of the -th vale of i-th criterion by ser and d days back before the day of rank calclation. ank vale r (t of retrieval mechanism type t sed by ser can be obtained in the following way d r (4 1 (t = f d (t (1 d= where f d (t is the sage freqency of the type t by ser and d days back before the day of rank calclation. Total rank vale of a page profile assmes that primarily calclated rank vale is then modified by rank vales of other elements according to ser preferences. So it can be calclated as follows + w (p = r (p + w c1 r (c 1 o w r (o + w ck where w o significance weight for options determined by -th ser, w t significance weight for retrieval mechanism types determined by -th ser, w c1, w c2,..,w ck significance weights for C 1, C 2, C k criteria respectively determined by -th ser and k is the nmber of criteria chosen by the ser to be sed as rank vale modifiers. Each ser can set his preferences by determining the nmber of days for calclation of rank vales and the weights of modifying rank vales of other elements. So if the weight assigned to a given element eqals zero, it means that the ser does not want to se that element to modify his page profiles. 2.3 Idea of the implementation When a ser rns a qery at any option page with a qery form, the featres of that qery i.e. option, type of retrieval mechanisms and the vales of criteria sed are immediately saved to the log of sbmitted qeries. Total rank vales are calclated for each page profile, for each ser separately, every day sing data saved in the log of sbmitted qeries and preferences set by individal sers. This can be performed in delayed mode, e.g. at night, when almost nobody ses the system. The reslt is saved in the log of recommended pages (see Figre 3. r (c k t r (t (5 5

6 ecommendation mechanism Log of sbmitted qeries Log of recom. page profiles ecommended page profiles Users Last qeries User preferences Figre 3 Schema of the implementation of the method eslting lists of page profiles with the highest scores are presented to each ser on the main men page and on all option pages with search forms. Having chosen an item from the list the ser control is moved to the option page with search form determined by the profile. The criteria fields are filled with vales according to the profile selected. e to this mechanism the ser is relieved of an ardos navigation throgh the option tree and from laborios filling of criteria fields in the case when he intends to carry ot search which is the same or similar to searches previosly performed. Then the ser can immediately rn the qery prompted by the system or modify that qery. Additional mechanism is list of last qeries sbmitted by individal sers available in the same pages as the list of recommended profiles. This cold be very sefl especially when a ser performed many similar searches dring one day and they were not taken into accont yet by the recommendation mechanism. 3 The real estate cadastre system The real estate cadastre system is designed to maintain the register of all parcels, bildings and apartments as well as their owners and sers at a given territory. The maintenance of real estate cadastre registers in Poland is dispersed. There are above 4 information centres located by administrative district governments as well as by the mnicipalities of bigger towns which exploit different cadastre systems. The ISEG2- INT system presented in the paper is an internet information system designed for the retrieval of real estate cadastre data and is complementary to the main system in which cadastre database is pdated. The system has been deployed in abot 5 intranets and extranets in local governments throghot Poland while the main EGB2 cadastre system is sed by above 1 centres. The ISEG2-INT system has been implemented sing PHP script langage and accommodated for cooperation with Apache or IIS Web servers. It assres commnication with MS SQL Server and MySQL database management systems. Using the system mainly rests on formlating qeries, browsing the list of retrieved obects, choosing the obects to reports and generating reports in PF format. As shown in 6

7 Figre 4 two main search criteria are mnicipality and the section which reflects the main spatial division of data in the cadastre system. Mnicipality Section Search criteria Obects retrieved egistration nits eports in pdf format Login Main men (List of options Parcels Hipo-parcels Bildings Apartments Figre 4 Generalized schema of the ISEG2-INT cadastre information system At present the data in cadastre systems have not been completed yet. However, descriptive data of land premises are flly completed. The information centres have been still gathering the data of bildings and apartments. Nmeric plans of real estate are being created or complemented too. Nmbers of obects contained in databases of for selected information centres are given in Table 1. Table 1 Nmber of obects in cadastre database in selected centres Centre no egistration nits Parcels Hypo-parcels Bildings Apartments egistered sers Active sers The access to the system is limited. Each ser shold be registered in the system and the rights shold be assigned to the data from a given territory. The sers of the system are the workers of local governments who tilize data to prepare administrative decisions, to inform real estate owners and to prepare reports for management boards of local governments. Usage statistics of the most freqently reqested pages (1, 11, 13, 15, 41 in for information centres dring months from November 24 to Febrary 25 7

8 are presented in Figre 5. It is expected that the sage of the system will increase significantly according to the greater nmber of bildings and apartments registered. nmber of ses center 1 center 2 center 3 center pages Figre 5 Usage statistics of selected pages in selected information centres 4 The strctre of pages in the system The strctre of pages in the ISEG2-INT system is presented in Figre 6, where denotes pages with retrieved obects. In trn Table 2 contains the list of the options of the main men with option codes, which are sed to represent recommended pages. Table 2 Options of main men Option no Name of the option Option code 1 Land registration nit search LU 2 Parcel search PA 3 Parcel search from a list PLI 4 Bilding registration nit search BU 5 Bilding search BUI 6 Apartment registration nit search AU 7 Apartment search APA 8 Hypo-parcel search HYP 9 Price and vale of premises search PI 1 System sage monitoring MON 11 Statistics STA Bilding a recommendation mechanism in this case is a challenging problem for several reasons. We need to bild a flexible qery sbsystem with the width range of topics which cold interest the ser. Since the ser s past interests is of limited practical se the system shold place less weight on history observations. 8

9 Figre 6 Strctre of pages in the ISEG2-INT system Table 3 Pages with search criteria Page No Name of the option Type of criteria 1 Login 3 Main men 11 Land registration nit search simplified 12 Land registration nit search extended 13 Parcel search simplified 14 Parcel search extended 15 Parcel search based on a list 21 Bilding registration nit search simplified 22 Bilding registration nit search extended 23 Bilding search simplified 24 Bilding search extended 31 Apartment registration nit search simplified 32 Apartment registration nit search extended 33 Apartment search simplified 34 Apartment search extended 41 Hypo-parcel search 9

10 51 Price and vale of premises search 61 System sage monitoring 7 Statistics men 71 Statistics list of parcels on a map 72 Statistics area of nits in sections 73 Statistics area of sections 5 The recommendation method sed in the cadastre system The idea of recommendation sed in the cadastre system is based on the concept of a page profile introdced in Section 2. Only two search criteria i.e. a mnicipality and a section were selected. Both are presented on almost every page enmerated in Table 3. The page profile of the cadastre system comprises therefore the option of the main men, the type of search criteria (simplified or extended and finally a mnicipality and a section chosen dring retrieval process. A formal model of page profiles can be presented as follows. The page profile is defined as a qadrple <O, M, S, T>, where O = {o 1, o 2,, o No } is a set of system options, where N o is the nmber of options, M = {m 1, m 2,, m Nm } is a set of mnicipalities registered in the system, where N m is the nmber of mnicipalities, S = {s 1, s 2,..., s Ns } is a set of sections registered in the system, where N s is the nmber of sections, T = {t 1, t 2,..., t Nt } a set of types of retrieval mechanisms, where N t is the nmber of types. The time impact coefficient i.e. the factor 1 - d/ has been also taken into accont to reflect the following principle: the earlier sed the element the lower its rank vale. way: 5.1 ank vale of a page profile ank vale r (p of a page profile p sed by ser is calclated in the following d r (6 1 (p = f d (p (1 d= where f d (p is the sage freqency of page profile p by ser and d days back before the day of rank calclation. In order to reveal the natre of this element the qery history log has been analysed for one of the most active ser in a chosen information centre for the period of two months from November 25 to ecember 25. This ser formlated 981 qeries with simplified criteria, sed 6 options and 254 sections dring this period. Figre 7 shows how the rank vales of three pages change in fnction of time. 1

11 rank vale Profile 1 Profile 2 Profile date of rank calclation Figre 7 ank vales of page profiles for a selected ser 5.2 ank vale of an option ank vale r (o of an option o sed by ser is shown as follows 1 d r (o = f d (o (1 (7 d= where f d (o is the sage freqency of option o by ser and d days back before the day of rank calclation. Similarly, in order to reveal the natre of this element analogos analysis for the same ser and the same period has been carried ot. Figre 8 shows how rank vales of three options change in fnction of time. It shold be noted that the graph in this figre was plotted sing data taken from different information centre and for different ser than graphs presented in other figres. 11

12 rank vale Option 7 Option 5 Option date of rank calclation Figre 8 ank vales of options for a selected ser way 5.3 ank vale of a mnicipality ank vale r (m of a mnicipality m sed by ser is presented in the following d r (8 1 (m = f d (m (1 d= where f d (m is the sage freqency of mnicipality m by ser and d days back before the day of rank calclation. In the cadastre system tested the rank vale of a mnicipality was not calclated, becase the system covers only one registration nit and a mnicipality is always the same. 5.4 ank vale of a section ank vale r (s of a section s sed by ser is obtained in the following way 1 d r (s = f d (s (1 (9 d= where f d (s is the sage freqency of section s by ser and d days back before the day of rank calclation. In order to reveal the natre of this element analogos analysis for the same ser and the same period has been carried ot. Figre 9 shows how the rank vales of three sections change in fnction of time. 12

13 rank vale Section Section Section date of rank calclation Figre 9 ank vales of sections for a selected ser way 5.5 ank vale of a criterion type ank vale r (t of a criterion t sed by ser can be calclated in the following d r (1 1 (t = f d (t (1 d= where f d (t is the sage freqency of criterion type t by ser and d days back before the day of rank calclation. 5.6 Total rank vale of a page profile Total rank vale of a page profile assmes that the rank vale calclated primarily is modified by rank vales of other elements according to ser preferences. So it can be calclated as follows (p = r (p + w s + w o r (o + w r (s + w r (t where W o significance weight of an option determined by -th ser, (.5 sed, W m significance weight of a mnicipality determined by -th ser, ( sed, W s significance weight of a section determined by -th ser, (.25 sed, W t significance weight of a criterion type determined by -th ser (.4 sed. Similarly, in order to reveal the natre of this element Figre 1 shows how the total rank vales of three profiles change in fnction of time. t m r (m (11 13

14 rank vale ank of page 1 ank of page 2 ank of page date of rank calclation Figre 1 Total rank vales of page profiles for a selected ser 6 The implementation of the method Two mechanisms have been designed and implemented. The first one is the list of last qeries sbmitted by a given ser and the second one is the list of page profiles recommended to a ser. In order to implement the recommendation mechanisms logs of sbmitted qeries have been created. Users have the possibility to determine their preferences by assigning significance weights to the elements of page profiles, by pointing the nmber of days when the sage freqency is taken into accont and by stating the nmber of items in the lists of recommended pages and last qeries. The strctre of main obects implemented is shown in Figre 11. The qery history comprises all qeries sbmitted by sers to the system inclding information of the option and the type of search mechanism sed. On the basis of the qery history the list of last sbmitted qeries can be created as well as the rank vales of page profiles can be calclated. 14

15 Option dictionary User preferences ecommended options Qery criteria Qery history Users ecommended profiles ecommended criteria Qery type dictionary ecommended qery types Figre 11 Strctre of the obects in the recommendation method 6.1 The list of last qeries sbmitted The list of last qeries sbmitted by a ser is pdated online. The list is provided for the ser in a combo box available on each page with search criteria as well as on the page with main men. Qeries on the list are described by data of sbmission, option code, type of criterion, mnicipality code, section code and the nmber of obect retrieved. An example of the list of last qeries ordered descending by data is shown in Table 4. The item at the list consists of following elements: - nmber of a qery sbmitted by a given ser, e.g data of qery rn, e.g option code as enmerated in Table 2, e.g. PA - type of criteria: S for simplified and E for extended - name of a section, e.g ; in other systems names are not codes bt short names of settlements or villages. The name of a mnicipality has been omitted, becase the system tested covers only one registration nit and the name of mnicipality is always the same. When a given item element was not sed in a qery the mask of xxxxx is inserted instead of a code of a section in order to point ot this fact. Table 4 The list of last qeries sbmitted by a given ser Qeries sbmitted on Qeries sbmitted on : : PA : S : xxxxxxxx 1165 : : PA : S : : : PA : S : : : PA : S : : : PA : S : xxxxxxxx 1163 : : PA : S : : : PA : S : xxxxxxxx 1162 : : PA : S : : : PA : S : : : PA : S :

16 When a ser chooses a qery from the list, the system moves the control to a page with a search form pointed by the option and the type of criteria and fills all the fields with the vales pt previosly by the ser. 6.2 The list of page profiles recommended The recommendation mechanism analyses previos activity of a ser and calclates and assigns a rank vale to each page profile sed by him. Similarly, the list of recommended page profiles is provided for the ser in a combo box available on each page with search criteria as well as on the page with main men. Page profiles on the list are characterized by an option code, type of criterion, mnicipality code and section code. An example of the list of recommended page profiles ordered descending by rank vale is shown in Table 5. The item of the list consists of following elements: - option code as enmerated in Table 2, e.g. PA - type of criteria: S for simplified and E for extended - code of a section, e.g name of a section, e.g ; in other systems names are not codes bt short names of settlements or villages. Analogosly to the list of last sbmitted qeries the name of a mnicipality has been omitted and also when given item does not occr in a profile, the mask of xxxxx is inserted instead of a code or a name of a section. Table 5 The list of page profiles with search criteria recommended to a ser Profiles recommended on Profiles recommended on PA : S : xxxx : xxxxxxx PA : S : xxxx : xxxxxxx PA : S : 733 : PA : S : 1231 : PA : S : 732 : PA : S : 133 : PA : S : 1115 : PA : S : 121 : When a ser chooses a page profile from the list, the system moves the control to a page with a search form pointed by the option and the type of criteria and fills all the fields with the vales contained in the profile. 7 Conclsions The recommendation method has been designed for an internet information system which provides for its sers different retrieval mechanisms based on definite nmber of search criteria. It has been assmed that the sers for some time perform searches concerning similar or the same topics and then accomplishing next task they focs for some days on other topic. Each qery sbmitted to the system is saved in a special log which is sed to calclate rank vales of page profiles separately for each ser. The page profiles, which represent a system option, type of retrieval mechanisms and search criteria, are given vales according to their freqency of sage and the time impact coefficient. A recommended page profile selected by a ser from a list facilitates and accelerates his searches by moving him directly to the chosen option page with search form filled with the most expected criteria vales. As an additional complementary mechanism the list of last sbmitted qeries is available to each ser. The recommendation mechanism has been implemented and provided to the sers of the ISEG2-INT cadastre information system. The list of last qeries is pdated in 16

17 online mode whereas the list of recommended pages is created for individal ser every night. rank vale calclation days: 15 calclation days: 2 calclation days: 25 calclation days: date of rank calclation Figre 12 Total rank vales of page profiles for a selected time period The performance of the proposed recommendation method may be evalated in comparison with the last qeries method. Experimental reslts have shown that the recommendation algorithm with the time impact coefficient improves the efficiency of the sage, especially when the information needed becomes more diverse and the nmber of accessed pages gets larger. In addition, we have measred rank vale with respect to different time periods. The reslts are given in Figre 12. When the period of calclation becomes to short, the rank vale decreases. Therefore, we sed 3 days period for all evalations. Ftre work will focs on observing how sers will se the recommendation mechanism and what vales will gain the ranks depending on different vales of parameters. For example, we want to test whether a system with recommendation works better for experienced ser or new sers. It will be interesting to investigate the behavior of system sers sing association rles and fzzy logic methods. Acknowledgements We thank r. Ngoc Thanh Ngyen who gave s a lot of valable advices. We also thank the Compter Association for Information BOGAT Ltd for financial spport and participation in the experiments. We thank anonymos reviewers for their helpfl sggestions. eferences [1] L. Ardissono, L. Console, I. Torre: An adaptive system for the personalized access to news. AI Commnications, vol. 14(3, 21, pp

18 [2] L. Ardissono, A. Goy, G. Petrone, M. Segnan, P. Torasso: Intrige: Personalized ecommendation of Torist Attractions for esktop and Handset evices. Applied Artificial Intelligence, vol. 17(8-9, 23, pp [3]. Baeza-Yates, C. Hrtado, M. Mendoza: Qery ecommendation Using Qery Logs in Search Engines. Lectre Notes in Compter Science vol. 3268, 24, pp [4].. Billss, M. Pazzani: User Modelling for Adaptive News Access. User Modelling and User-Adapted Interaction, vol. 1, 2, pp [5] P. Brsilowsky: Adaptive Hypermedia. User Modelling and User-Adapted Interaction, vol. 11, 21, pp [6] Z. Chedrawy, S. S.. Abidi: Intelligent Knowledge Sharing Strategy Featring Item-Based Collaborative Filtering and Case Based easoning. In Proc. International Conference on Intelligent Systems esign and Applications, 25, pp [7]. N. Chin: Empirical Evalation of User Models and User-Adapted Systems. User Modeling and User-Adapted Interaction, vol. 11, 21, pp [8] M. Eirinaki, M. Vazirgiannis: Web Mining for Web Personalization. ACM Transactions on Internet Technology vol. 3, no. 1, 23, pp [9] S. Gndz, T. Ozs: A User Behavior Model for Web Page Navigation. University of Waterloo, 22 [1] A. Jameson: Adaptive Interfaces and Agents. In J. Jacko, A. Sears (Eds., Hman-Compter Interaction Handbook, Mahwah, NJ: Erlbam, chapter 15, 23, pp [11] P. Kazienko: Mlti-agent Web ecommendation Method Based on Indirect Association les. Lectre Notes in Artificial Intelligence, vol. 3214, 24, pp [12] P. Kazienko, M. Kiewra: Personalized ecommendation of Web pages. International Series on Advanced Intelligence, vol. 1, 24, pp [13] H. T. Macedo, J. obin: Increasing Profitability: Voice-Based Browsing to ecommendation System Web Services. Lectre Notes in Artificial Intelligence vol. 3528, 25, pp [14] S-H. Min, I. Han: Optimizing Collaborative Filtering ecommender Systems. Lectre Notes in Artificial Intelligence vol. 3528, 25, pp [15] O. Nasraoi, M. Pavlri: Complete this Pzzle: A Connectionist Approach to Accrate Web ecommendations based on a Committee of Predictors. In Proc. WebK Workshop on Web Mining and Web Usage Analysis, Seattle, WA, 24 [16] O. Nasraoi, C. Petenes: Combining Web Usage Mining and Fzzy Inference for Website Personalization. In Proc. K Workshop on Web mining as a Premise to Effective and Intelligent Web Applications, Washington C, 23 [17] J. Sobecki: Web-Based Systems User Interfaces ecommendation Using Hybrid Methods. International Series on Advanced Intelligence, vol. 1, 24, pp [18] M. Vozalis, K. G. Margaritis: Applying SV on Item-based Filtering. In Proc. International Conference on Intelligent Systems esign and Applications, 25, pp

19 [19] B. Zho, S. Hi, A. Fong: Efficient seqential access pattern mining for web recommendation. International Jornal of Knowledge-based and intelligent Engineering Systems, vol. 1, no 2, 26, pp [2] T. Zh: Web Usage Mining for Internet ecommendation. University of Alberta, 21 19