Nature Inspired Recommender Algorithms for Collaborative Web based Learning Environments

Transcription

1 Iteratioal Joural of Computer Applicatios ( ) Nature Ispired Recommeder Algorithms for Collaborative Web based Learig Eviromets Diesh Kumar Saii Faculty of Computig ad Iformatio Techology Sohar Uiversity Sultaate of Oma Research Fellow ad Adjuct Faculty, Uiversity of Queeslad Australia ABSTRACT The desig of recommeder systems for various domais has bee proposed based o the ature ispired algorithms. I this paper attempt is made to propose a Nature Ispired Algorithms based architecture for recommeder system for web based learig eviromets. The paper also compares betwee the traditioal recommeder systems ad the ature ispired algorithm recommeder systems. Collaborative filterig is proposed for persoalized recommedatios; user ad item attributes are used as filtratio parameter. Attributes ad ratig of the user s similarity is used for collaborative filterig process. Hybrid collaborative filterig is proposed for user ad item attribute that ca alleviate the sparsity issue i the recommeder systems. Traditioal systems are studied i detail ad all the possible limitatios of the traditioal systems are bought uder attetio. Geeral Terms Computig, Nature, Algorithms, Web Sciece. Keywords Recommeder Systems, web based educatioal eviromets, architecture, ature ispired algorithms, optimizatio, ad software testig. 1. INTRODUCTION The role of recommeder systems for decisio-makig is gaiig paramout importace as several domais are ow havig such systems as a itegral compoet of their architectures [1]. The study of recommeder systems was iitiated i the mid-90s. Users are by ad large familiar with websites like Amazo.com, Netflix, YouTube, It was observed that the magitude ad variety of iformatio available o the iteret was overwhelmig for a great majority of the users ad they were ofte perplexed whe it came to selectig or makig a choice or a choice set from a recommeded group of items. The reaso for icorporatig recommeder systems i a service or website is maifold. Of primary importace is the eed to Improve the efficiecy of service offered. Attract more users to use the website or service. Uderstad the requiremets of the user so that the cotets of the system or service ca be improved accordig to this parameter. Icrease the volume of trasactios ad be a aggressive competitor i the olie trasactioal systems eviromet. Assess the cotets available i the website based o ratigs ad rakigs which traslates or coverts ito iformatio that will help recogize or discover the most preferred item i Lakshmi Suil Prakash Faculty of Computig ad Iformatio Techology Sohar Uiversity Sultaate of Oma the item collectio. Develop trust i the service that will i tur lead to users recommedig the items i the service to others surfers, who share similar prefereces or trust the recommedatios made by this particular user. Predictig the demad or ext possible additio to the cotet repository by studyig user patters based o feedback from several user sessios i the website. A learer s activity is guided by Protus which is a itelliget web-based Programmig Tutorig System.It is used for guidig the learer's activities ad recommeds relevat liks ad actios to him/her durig the learig process. I [2] the authors discuss how Nutch s automated crawlig ad idexig techiques as well as stadardized educatioal cotet idexig are used to build cotet profiles, ad Web usage miig techiques (clusterig ad associatio rule miig) are used to build user profiles. Hybrid recommedatios (cotet based filterig ad collaborative based filterig) were used i the recommedatio phase. The approach i this paper is towards filterig the learers accessig the system ito clusters based o their learig styles ad subjects of study. We also take ito accout the ratigs eared by learers based o the umber 2. TRADITIONAL RECOMMENDER SYSTEMS Collaborative filterig systems face the problem of shillig. It is the term used to refer to the ijectio of fake user profiles ito the ratig database of a recommeder system, with the itet of ifluecig the recommedatio behavior of the system. I this the shillig problem will ot arise as the learers will be havig uique id geerated at the time of course registratio, the system will autheticate the user o the basis of their registratio details at the istitutio. Users expect collaboratio based learig eviromets are required to be able to hadle icreasig umber of users ad learig items. However the real challege lies i gettig recommedatios ad ratigs from users. This is called the data sparsity problem [3,4]. Table 1: Traditioal Algorithms Compariso Data Sparsity Algorithms Sigular Value Decompositio (SVD) [23], Latet Sematic Idexig (LSI) SVD [5] Descriptios a closely-related factor aalysis techique remove urepresetative or isigificat users or items to reduce the dimesioalities similarity betwee users is determied by the represetatio of the 16

2 Iteratioal Joural of Computer Applicatios ( ) Priciple Compoet Aalysis (PCA),[6], Eigetaste,[6] hybrid collaborative filterig approach [7] users i the reduced space a closely-related factor aalysis techique remove urepresetative or isigificat users or items to reduce the dimesioalities Goldberg et al. developed which applies to reduce user-item dimesioality How to exploit bulk taxoomic iformatio desiged for exact product classificatio to address the data sparsity problem of CF recommedatios, based o the geeratio of profiles via iferece of supertopic score ad topic diversificatio 3. TRADITIONAL RECOMMENDATION ALGORITHMS The followig are some of the traditioal recommedatio algorithms that have bee developed, these iclude collaborative filterig [3,4], cotet-based aalysis [5], spectral aalysis [6,7] ad Iterative self-cosistet refiemet [8, 9]. What most traditioal collaboratio filterig algorithms have i commo is that they are based o similarity, either of users or items or both[8]. Such approach is uder high risk of providig poor coverage of the space of relevat items. As a result, with recommedatios based o similarity rather tha differece, more ad more users will be exposed to a arrow bad of popular items. Although it seems more accurate to recommed popular items tha iche oes, beig accurate is ot eough [10]. Diversity ad ovelty are also importat criteria of algorithmic performace. The diversity-accuracy dilemma becomes oe of the mai challeges i recommeder system. These algorithms face similar problems like The tasks for which collaborative filterig ca be performed are [3,7] 1. Suggest items i the data set which the user may fid iterestig 2. Create a group of users who share the same iterest 3. Suggest a recurrig set of similar set of items that a user may fid iterestig 4. Suggest details about a selected item.. 5. To group results of previous searches ad predict recommedatios for future 4. REASONS FOR NEW ALGORITHMS IN RECOMMENDER SYSTEMS The large scale of data i recommeder systems is a major reaso for the eed to move away from the traditioal algorithms which iclude the collaborative algorithms (Pearso s coefficiet. Nature Ispired Algorithms have bee very popular i recet years as they have bee able to provide simple ad effective meta-heuristic solutios to complicated problems i the realworld Several Bee Coloy algorithms have bee proposed based o the foragig behaviour which icludes the food searchig ad searchig for ew est behaviours of bees. Table 2: New Algorithms i Recommeder Systems Bee Ispired Algorithm Bees System (BS) Algorithm [9] Bee Coloy Optimizatio (BCO) [10] Hoey Bee Algorithm[11 ] Beehive Algorithm[12] At Coloy Algorithm [15] Bat ispired Algorithm[14] Essece Collects maximum ectar from the hives i the bee trajectory. It determies the route to be take takig ito cosideratio the distace ad demad at various odes i the route. Hoey bee coloies are selforgaised i that they have reach the food source with the help of other bees ivolved i the same activity Based o the local iformatio that a short distace bee aget collects i a food searchig zoe Based o the pheromoe secretio of ats which helps to create a trail for the ats comig after. Echolocatio property of bats Applicatio Tested o travellig salesma problem. Produced good results Vehicle Problem routig Dyamic allocatio of iteret sources Applied to routig i wired computer etworks. I VRS to help vehicles fid the least cogested path Idetifyig the correct object ad discrimiatig betwee objects i a search routie. 5. PROPOSED WORK BASED ON BEE COLONY ALGORITHM I a bee coloy, the quee bee ca be compared to a highly rated user. All the other bees i the bee coloy are proe to the ifluece of this quee bee. I the same way, learers who 17

3 Iteratioal Joural of Computer Applicatios ( ) have high success rates ifluece the learig decisios of other learers i the group. Each cluster ca be compared to a bee coloy with its ow quee bee Cotets i the Learig Maagemet System The compoets of the LMS are divided ito Learers, Istructors ad Learig items. Learer Clusters Learig Items Learig Profiles Bat Coloy Alg to discrimiate betwee useful ad oiterestig learig items. Recommeder System The most suitable learig object for each cluster At Coloy Alg to cluster learers with similar iterests. Highly Rated Learig i each cluster Bee Coloy Alg selects the best learer i each cluster. Learig Maagemet System I Figure 1, we discuss the three algorithms which determie the recommeder s ability to provide the most optimised search results to its users. The At Coloy Algorithm [18] is required by the recommeder system to cluster similar learers. These clusters have dimesios such as learig style, ad subject iterest. Oce the learig style ad subject iterest are gleaed from the learer profiles, the a trail is created for other users with similar iterests to be clusters together o the basis of these two traits. Similarly the Bee Coloy algorithm helps to idetify the learer with best ratigs o the basis of the recommeder systems calculatios of access time ad assessmet scores of the learers. This helps to filter the best learer i each cluster. Figure 1 While the Bat Algorithm helps to discrimiate betwee the useful learig objects ad others which are ot useful, so that the highly rated learer i the cluster is ow able to receive the best recommedatios for his /her learig module. The Learig Maagemet System cosists of the followig etities: Course Name Subject Course Coordiator Course Descriptio Course Learer Profile Advaced, Itermediate, Begier No of learers is deoted by N 18

4 Each course will have learig items.its attributes will be as follows Learig item_id uique idetifier Learig item_type assessmet item, learig material, group assigmet etc. Learig_outcome expected learig outcome achieved after completig the learig item. Learig item_filetype - audio, video, presetatio, word documet. Iteratioal Joural of Computer Applicatios ( ) Cotet advaced, itermediate, begier The learer group is categorised by the learig style prefereces collected from the learer profile. Suggested for Learig_style Usig Vark Learig Styles[7] - Verbal, aural, visual, logical, kiaesthetic,solitary or social Frequecy of use (F q ) - total score of accesses eared by the item durig the duratio of the module. Professioal Advaced Small Cluster Size Small Learig Space Itermediate Large Cluster Size Begier Large Learig Space Largest Cluster Size Total Learig Space Larger Cluster Size Larger Learig Space Recommedatios_eared (LR )- calculated by the recommeder system o the basis of learer access ad duratio of use. Item_Ratig (IR ) - ratigs eared by the item, calculated by frequecy of access by top-rated learers ad recommedatios eared. Each learer will have the followig key attributes Learer_id studet registratio umber. Learig_style (L s )- Verbal, aural, visual, logical, kiaesthetic,solitary or social Assessmet_result (R) achieved by the learer o completio of a module. Learer_ratig (Lr ) ratigs eared by the learer o the basis of assessmet results. Learer Cluster (LC) category or categories to which the learer may belog Figure 2: Learer Groups ad Space While clusterig learers by the learig style, we also eed to deliver the most suitable learig cotet to the learer. Normally suitability of cotet is measured by the earest eighbor algorithm or Pearso s coefficiet, however usig The suitability of the cotet ca be assessed by the recommedatios of the learers who score higher assessmet results; this learer becomes the learer with the highest learer ratig. Accordig to the QBE algorithm, the quee bee is the learer with the most authority to lead the group, i this maer the recommeder system ca suggest to each learer the most suitable items for his study based o the recommedatio ratigs eared by each item The recommeder based learig systems will ot suffer from sparsity problems if the system ca rate ay item by the Lr L ( Lr ) s umber of items that is available i the cotet database by the umber of users accessig the item multiplied by the access times. LR R recommede d 19

5 Iteratioal Joural of Computer Applicatios ( ) Similarly each learer profile will be havig a ratig oce he completes the module depedig o his/her performace i the assessmet for that module. 6. THE ABB ALGORITHM I this algorithm a user cluster is created based o the similarity i learig styles ad similarity of subject iterest. Here the best performig learers for a module receive the highest ratigs from the module or course coordiator. These top-rated learers are the filtered by their learig styles; these learig styles ca be termed L s The Mea average recommedatios eared R by the item are the calculated. The Mea average ratigs for the learer are also calculated across each assessmet, MLR The Learig Style factor L s iflueces the categorisatio of learers ito clusters. Fq + LR i1 LR + IR LR Fq i1 LR i1 IR With time ad duratio of access, the recommedatio eared ad frequecy of access Lr d dx Learig Ratig, Fq d IR Lr = dx i1 LR Cetroid distace F 2= N i 1 j 1, K d(zi,mj ) Variace Ratio criterio = F 4 = VRC = trace B /(K-1) / trace W/(N-K) Itra ad iter cluster distace = F 5 = K it ( ), i D c 1 iter i w D ra c j w is a parameter. Du s, idex F 6 ( ci, c j ) DI / K i K, j 1{ k { K( ( c where c i, c ) = mi {d(z i, z j ) : z i, c i z } ( j, j 7. TESTING THE RECOMMENDER SYSTEMS Recommeder systems are testig based o the accuracy ad closeess of the recommedatio suggested by the algorithm used. [19] The scope of the system will be tested the used of the best algorithm, assumptios made for learers, baselie documets, methodology adapted to desigig the proposed systems, etry criteria. As show i figure 3 cocept ad formulas will be the basis of the recommedatio with structure ad relatios. k )}, Figure3: Epistemological Triagle ad recommeder systems 7.1 Testig Process For recommeder systems we eed to test how the systems adapt recommedatio process, which algorithm comes closer to the expectatios ad preadaptatio i the process. The systems eed to be tested o sufficiet explosio ad for performace ad accuracy [20,21]. The system must be tested for fault tolerace, prevetio ad forecastig of faults i the system is difficult to predict but it is still eeded i the recommedatio systems. Implemetatio of supervised learig mechaism i the recommedatio systems is very much desired to that false recommedatios ca be miimized [22]. Cotext perspective i recommedatio systems usig qualitative research is very subjective ad situatios arisig from the qualitative research are ot easy to hadle. Moreover, qualitative research methodologies are cocered with the opiios, experieces ad feeligs of idividuals [16]. Testig such recommedatios is ot easy task but various testig techiques will be employed i the give situatio. [23] As show i figure o 4 various testig strategies will be adopted for checkig the accuracy ad perfectio of the system. Recommedatio fuctios, GUI compoets, systems acceptace ad accuracy will be tested ad validated before adaptig the particular algorithm for the recommedatio system [17]. 20

6 Iteratioal Joural of Computer Applicatios ( ) S.No Test Case ID 1 Geeral Fuctio Figure 4: Software Testig Process for Recommeder Systems Table4: Software Test Cases for the recommeder system Objective Id Category Coditio Expected Result Actual Result Performace ad Fuctioality Sposor /developmet /Testig recommeder Which algorithm better is Best Recommeder Accuracy Req.ID Which is better 8. CONCLUSIONS AND FUTURE WORK I this paper we proposed recommeder systems for various Kowledge domais based o ature ispired algorithms. Recommeder systems architecture based o ature ispired algorithm is for web based learig eviromets. The paper also compares betwee the traditioal recommeder systems ad the ature ispired algorithm recommeder systems. Collaborative filterig is proposed for persoalized recommedatios; user ad item attributes are used as filtratio parameter. Attributes ad ratig of the user s similarity is used for collaborative filterig process. Hybrid collaborative filterig is proposed for user ad item attribute that ca alleviate the sparsity issue i the recommeder systems. This system eed to be tested ad validated that ature ispired algorithm perform better tha traditioal algorithms. First Bee coloy optimizatio algorithm was used to desig ad propose the recommedatio systems, ad it is suggested that ca be itegrated i the Learig cotet maagemet systems. 9. REFERENCES [1] Zhag, Fuzhi, ad Quaqiag Zhou. "A Meta-learigbased Approach for Detectig Profile Ijectio Attacks i Collaborative Recommeder Systems." Joural of Computers 7.1 (2012). [2] Khribi, Mohamed Koutheaïr, Mohamed Jemi, ad Olfa Nasraoui. "Toward a hybrid recommeder system for e- learig persoalizatio based o web usage miig techiques ad iformatio retrieval." World Coferece o E-Learig i Corporate, Govermet, Healthcare, ad Higher Educatio. Vol No [3] Prakash, Lakshmi Suil, Diesh Kumar Saii, ad Narayaa Swamy Kutti. "Itegratig EduLear learig cotet maagemet system (LCMS) with cooperatig learig object repositories (LORs) i a peer to peer (P2P) architectural framework." ACM SIGSOFT Software Egieerig Notes 34.3 (2009): 1-21

7 Iteratioal Joural of Computer Applicatios ( ) [4] Ludford, P.J., Cosley, D., Frakowski, D., Tervee, L. Thik Differet: Icreasig Olie Commuity Participatio Usig Uiqueess Ad Group Dissimilarity. Proceedigs of the SIGCHI coferece o Huma factors i computig systems (2004),ACM Press: Viea, Austria p [5] Suil Prakash, Lakshmi, Narayaa Swamy Kutti, ad A. S. M. Sajeev. "Review of challeges i cotet extractio i web based persoalized learig cotet maagemet systems." Proceedigs of the 12th Iteratioal Coferece o Iformatio Itegratio ad Web-based Applicatios & Services. ACM, [6] Goldberg,T.Roeder,D.Gupta,,Perkis, Eigetaste: a costat time collaborative filterig algorithm, Iformatio Retrieval, vol. 4, o. 2, pp , [7] T. Ladauer, M. Littma, ad Bell Commuicatios Research (Bellcore), Computerized cross-laguage documet retrieval usig latet sematic idexig, US patet o , April [8] K. Pearso, O lies ad plaes of closest fit to systems of poits i space, Philosophical Magazie, vol. 2, pp [9] Prakash, Lakshmi Suil, ad Diesh Kumar Saii. "Eassessmet for e-learig." Egieerig Educatio: Iovative Practices ad Future Treds (AICERA), 2012 IEEE Iteratioal Coferece o. IEEE, [10] Flemig, N. (1995), VARK a guide to learig styles, available at: h/idex.asp [11] Billsus ad M. Pazzai, Learig collaborative iformatio filters, i Proceedigs of the 15th Iteratioal Coferece o Machie Learig (ICML 98), [12] P. Lucic, ad D. Teodorovic, Bee system: Modelig Combiatorial Optimizatio Trasportatio Egieerig Problems by Swarm Itelligece, Preprits of the TRISTAN IV Trieial Symposium o Trasportatio Aalysis, Sao Miguel, Azores Islads, pp , 2001 [13] D. Teodorovic, ad M. Dell'Orco, Bee Coloy Optimizatio - A Cooperative Learig Approach to Complex Trasportatio Problems, Advaced OR ad AI Methods i Trasportatio, pp , 2005 [14] S. Nakrai, ad C. Tovey, O Hoey Bees ad Dyamic Allocatio i a Iteret Server Coloy, Proceedigs of 2d Iteratioal Workshop o the Mathematics ad Algorithms of Social Isects, Atlata, Georgia, USA, 2004 [15] H.F. Wedde, M. Farooq, ad Y. Zhag, BeeHive: A Efficiet Fault-Tolerat Routig Algorithm Ispired by Hoey Bee Behavior, [16] H.F. Wedde, M. Farooq, T. Paebaecker, B. Vogel, C. Mueller, J. Meth, ad K. Jeruschkat, BeeAdHoc: A eergy efficiet routig algorithm for mobile ad hoc etworks ispired by bee behavior, GECCO 2005, Washigto DC, USA, 2005 [17] Yılmaz, S., E. Ugur Kucuksille, ad Y. Cegiz. "Modified Bat Algorithm." Electroics & Electrical Egieerig 20.2 (2014). [18] At Coloy, Optimizatio ad Swarm Itelligece, Eds. M. Dorigo, LNCS 3172, Spriger Berli, pp , 2004 [19] Liamputtog, P. Qualitative data aalysis: Coceptual ad practical cosideratios. Australia Joural of Health Promotio, 20(2), (2009) [20] WM Omar, DK Saii, M Hasa Credibility of Digital Cotet i a Healthcare Collaborative Commuity, Software Tools ad Algorithms for Biological Systems, pp , 201. [21] N Gupta, D Saii, H Saii Class Level Test Case Geeratio i Object Orieted Software Testig, Web Egieerig Advacemets ad Treds: Buildig New Dimesios of Iformatio Techology, pp , [22] Saii, Diesh Kumar, Lakshmi Suil Prakash, ad M. Goyal. "Emergig iformatio techology ad cotemporary challegig R & D problems i the area of learig: A artificial itelligece approach." Egieerig Educatio: Iovative Practices ad Future Treds (AICERA), 2012 IEEE Iteratioal Coferece o. IEEE, [23] Lima, Salvador, ad José Moreira. "A Sematic Framework for Touristic Iformatio Systems." Cases o Ope-Liked Data ad Sematic Web Applicatios (2013): 132. IJCA TM : 22