BinRec: addressing data sparsity and cold-start challenges in recommender systems with biclustering

Applied In elligence (2025) 55:830 h ps://doi.o g/10.1007/s10489-025-06725-6 1 In oduc ion Recommende sys ems (RSs) we e in oduced in he 90 s o suppo use s in a gi en decision making si ua ion [1]. These so wa e ools, om he use ’s poin o iew, educe he sea ch space when accessing a p oduc ca alog and p o- cess he opinions o o he use s. Fo many companies dedi- ca ed o selling p oduc s o any kind, p o iding mul imedia con en , o simply being a da abase o mo ie ankings, he RSs acili a e he use ’s selec ion and/o pu chase p ocess while also ocusing a en ion on a di e en se o less well- known and success ul p oduc s, inc easing hei chances o sale. The use o RSs has become essen ial in many applica- ions, including social ne wo ks [2], online educa ion [3], and Big Da a ecosys em [4, 5]. Al hough he e a e a ious ypes o RSs, he mos widely used app oach is Use Based Collabo a i e Fil e ing (UBCF) [6]. UBCF sys ems p ocess a da abase o use a - ings (’likes,"s a a ings,’ o nume ical da a) on any kind o i em (selling p oduc s, books, mo ies, e c.) o gene a e i em F ancisco Gómez-Vela, Au elio Lopez-Fe nandez, Miguel Ga cía- To es and Fede ico Di ina con ibu ed equally o his wo k. Domingo Rod íguez-Baena [email p o ec ed] F ancisco Gómez-Vela [email p o ec ed] Au elio Lopez-Fe nandez [email p o ec ed] Miguel Ga cía-To es [email p o ec ed] Fede ico Di ina [email p o ec ed] 1 Compu e Science, Uni e sidad Pablo de Ola ide, C a. U e a km 1, Se ille ES-41013, Se illa, Spain 2 Depa men o Compu e Languages and Sys ems, Uni e sidad de Se illa, Se ille ES-41004, Se illa, Spain 3 Da a Science and Big Da a Lab, Uni e sidad Pablo de Ola ide, C a. U e a km 1, Se illa ES-41013, Se illa, Spain Abs ac Recommende Sys ems help use s in making decision in di e en ields such as pu chases o wha mo ies o wa ch. Use - Based Collabo a i e Fil e ing (UBCF) app oach is one o he mos commonly used echniques o de eloping hese so - wa e ools. I is based on he idea ha use s who ha e p e iously sha ed simila as es will almos ce ainly sha e simila as es in he u u e. As a esul , de e mining he nea es use s o he one o whom ecommenda ions a e sough (ac i e use ) is c i ical. Howe e , he massi e g ow h o online comme cial da a has made his ask especially di icul . As a esul , Biclus e ing echniques ha e been used in ecen yea s o pe o m a local sea ch o he nea es use s in subg oups o use s wi h simila a ing beha iou unde a subg oup o i ems (biclus e s), a he han sea ching he en i e a ing da abase. Ne e heless, due o he la ge size o hese da abases, he numbe o biclus e s gene a ed can be ex emely high, making hei p ocessing e y complex. In his pape we p opose BinRec, a no el UBCF app oach based on Biclus e ing. BinRec simpli ies he sea ch o neighbou ing use s by de e mining which ones a e nea es o he ac i e use based on he numbe o biclus e s sha ed by he use s. Expe imen al esul s show ha BinRec ou pe o ms o he s a e-o - he-a ecommende sys ems, wi h a ema kable imp o emen in en i onmen s wi h high da a spa si y. The lexibili y and scalabili y o he me hod posi ion i as an e icien al e na i e o common collabo a i e il e ing p oblems such as spa si y o cold-s a . Keywo ds Recomende sys ems · Biclus e ing · Collabo a i e il e ing · Da a mining Accep ed: 13 June 2025 © The Au ho (s) 2025 BinRec: add essing da a spa si y and cold-s a challenges in ecommende sys ems wi h biclus e ing DomingoRod íguez-Baena1· F anciscoGómez-Vela1· Au elioLopez-Fe nandez2· MiguelGa cía-To es3· Fede icoDi ina3 1 3 D. Rod íguez-Baena e al. ecommenda ions o a speci ic use , answe ing he ques- ion: Wha is he mos popula among people who sha e my as es? As a esul , hey a e based on he idea ha use s who ha e p e iously sha ed simila as es will almos ce ainly sha e simila as es in he u u e. To do so, a subse o use s (nea es use s) who a e simila o he use o whom ecom- menda ions a e sough (ac i e use ) is chosen, and hei a - ings a e combined o p edic unseen i ems. Howe e , UBCF echniques include impo an issues ha a e s ill a challenge o esea che s, such as spa si y, pe o - mance, and he cold-s a p oblem [7, 8]. When a ing da a- bases a e la ge and include a signi ican numbe o i ems, he e a e ew a ings a ailable o each use . This makes i ha de when looking o nea es use s because he e is less in o ma ion a ailable o look o simila beha iou in e ms o a ing (spa si y p oblem). Fu he mo e, in hese cases, UBCF app oaches pe o m poo ly in e ms o scalabili y [9, 10]. Mo eo e , adding a new use o he a ings da a- base may be p oblema ic because his use has an emp y o e y small a ing p o ile, and as a esul , he sys em canno accu a ely calcula e he simila i ies be ween such cold-s a use s and o he s [11]. Thus, since he i s UBCF app oach was p oposed by Resnick e al. [12], many pape s ha e been published in an a emp o imp o e p edic ion esul s and sol e he majo issues associa ed wi h hese echniques [13]. One o he mos impo an asks o UBCF p ocess- ing is de e mining he nea es use (s) o he ac i e one by calcula ing simila i y measu es be ween hem. Gi en he huge amoun s o da a a ailable, his sea ch could be com- plex and ime-consuming [14]. To deal wi h he di e si y o use p o iles, educe dimensionali y and making easie o ind he nea es use s, Clus e ing echniques ha e been used ei he di ec ly o as a p ep ocessing s age in UBCF sys ems [15]. These echniques gene a e disjoin g oups o use s by aking in o accoun he a ings o all he i ems. Howe e , in eal-wo ld scena ios, he e is usually a s ong co ela ion be ween he p e e ences o subse s o use s on subse s o i ems [16]. Thus, du ing he las yea s, UBCFs based on Biclus e ing ha e been p oposed o educe he consequences o spa si y issues in he inpu da ase [17]. Biclus e ing echniques g oup subse s o elemen s ha ha e simila i ies unde a subse o a ibu es and ha e been used success ully in many a eas o s udy [18]. So, ins ead o pe - o ming a global sea ch o ind he nea es use s o he ac i e one, Biclus e ing can be used o na ow he sea ch space. The a ing da abase biclus e s, which a e subse s o use s wi h simila a ing beha iou unde a subse o i ems, ep- esen dense a eas whe e he nea es use s can be ound in a local sea ch. The gene a ion o biclus e s akes place in a phase p io o he ecommenda ion, called he o line phase. Following ha , and on demand, he nea es use s o he ac i e use a e sea ched in he biclus e s o gene a e a speci ic ecommen- da ion (online phase) [19]. Howe e , due o he la ge num- be o biclus e s ha can be ex ac ed om he inpu a ing da ase s, p ocessing hem o de e mine he simila i y mea- su es when looking o he nea es use s can be challenging. Fu he mo e, he e a e se e al impo an aspec s associa ed wi h he use o Biclus e ing. Fo example, which Biclus- e ing echnique applies, i he same se o biclus e s can be used o e e y ac i e use , how o upda e he biclus e s when new da a is added o he a ing da ase o how o a o d he cold-s a p oblem. This wo k p esen s BinRec, a no el UBCF app oach ha uses biclus e s o make he sea ch o he nea es use s mo e e icien in spa se da abases, while also op imizing he compu a ional complexi y o such a sea ch h ough a simple da a s uc u e. This s uc u e can also be used o add ess he cold-s a p oblem, which is common in a ing da abases. BinRec is lexible, adap ing o any Biclus e ing echnique. Howe e , he use o he BiBi Biclus e ing algo i hm [20] is p oposed, as i allows he inc emen al gene a ion o new biclus e s as new use s a e added o he a ing da abase. This ea u e, combined wi h he a o emen ioned e iciency, makes BinRec well sui ed o managing la ge-scale da a- bases wi h a high numbe o biclus e s. The main con ibu ions o his wo k can be summa ized as ollows: ●In oduc ion o BinRec, a no el collabo a i e il e - ing app oach ha le e ages Biclus e ing echniques o g oup use s wi h simila a ing p e e ences, he eby educing he sea ch space o nea es neighbou s in spa se a ing da abases. ●BinRec educes he compu a ion ime equi ed o iden i y he nea es use s o a gi en one based on biclus- e s, using a simple da a s uc u e ha s o es he numbe o biclus e s sha ed be ween use s, while also achie ing highly accu a e p edic ions o use ecommenda ions. ●The cold s a p oblem can also be add essed using he same in o ma ion s o ed in he p e iously men ioned da a s uc u e. ●BinRec adap s o any Biclus e ing echnique. How- e e , he use o BiBi Biclus e ing algo i hm is p oposed due o i s pe o mance and he possibili y o de eloping an inc emen al biclus e s p ocessing when new use s a e added o he a ing da abase. All hese con ibu ions make BinRec an e icien and accu- a e app oach ha can be applied o la ge a ing da abases and a high numbe o biclus e s gene a ed om hem, while also add essing he issues o spa si y and cold-s a . To demons a e his, he s udy includes expe imen a ion and a compa ison wi h o he benchma k collabo a i e 1 3 830 Page 2 o 17 BinRec: add essing da a spa si y and cold-s a challenges in ecommende sys ems wi h biclus e ing il e ing app oaches and a biclus e -based ecommenda ion echnique. The es o he pape is o ganized as ollows: Sec ion 2 e iews he s a e o he a in FC-based ecommende sys- ems and Biclus e ing. In Sec ion 3, he BinRec app oach is desc ibed in de ail, ollowed by a se ies o expe imen s and analysis o esul s in Sec ion 4. Finally, Sec ion 5 p esen s conclusions and possible u u e lines o esea ch. 2 Backg ound As i has been commen ed in he p e ious Sec ion, se e al UBCF echniques based on Biclus e ing ha e been p o- posed in he las yea s [21]. Following, hey a e analyzed om di e en poin s o iew. In i s place, he gene a ion o he biclus e s: whe he he Biclus e ing echnique is applied o he en i e da ase o only a po ion o i , which Biclus e - ing echnique is used, and so on. Second, how he gene a ed biclus e s a e p ocessed o ind he nea es use s. Finally, how he cold-s a p oblem is handled is an impo an ac o o conside . In e ms o biclus e gene a ion, mos app oaches apply Biclus e ing echniques o he en i e a ing da ase , ega d- less o who is he ac i e use [17, 22–33]. In hese cases, all o he gene a ed biclus e s a e used o p edic any ac i e use ’s ecommenda ions. Howe e , i a new use and hei a ings a e added o he inpu da ase , he Biclus e ing echnique mus be applied again o include his new use . O he esea ch wo ks only gene a e biclus e s ha include he ac i e use . The gene a ion o he biclus e s is hus less compu a ionally expensi e, bu i mus be epea ed o each ac i e use . Fo example, in he wo ks [34, 35], a hie a - chical Biclus e ing me hod is applied o gene a e biclus e s in di e en laye s consis ing o co- a ed i ems among he ac i e one and o he use s. In he i s laye , a biclus e is c ea ed o e e y i em a ed by he ac i e use , including also he es o use s ha ha e a ed he same i em. In he second laye , pai s o biclus e s om he i s laye a e com- bined, so he subse o i ems inc eases while he subse o use s dec eases in he new biclus e s. The p ocess ends in a laye in which he biclus e s con ains he maximum numbe o i ems wi h he minimum numbe o use s. Some esea ch wo ks a e e y lexible when i comes o selec ing which Biclus e ing echnique o use [23, 30–33, 36], allowing you o use wha e e echnique you wan . Ob iously, bo h he inal esul s and he pe o mance may a y depending on he op ion chosen. In he six p e iously ci ed wo ks, he Biclus e ing algo i hms selec ed o conduc he expe imen a ion a e Bimax [37], QUBIC [38] o XMo i [39]. In many o he cases, speci ic Biclus e ing algo i hms ha e been designed o g oup oge he subse s o use s and subse s o i ems. Fo example, BIC-aiNET, p oposed in [22] and used again by Desai e al. [26], is an adap a ion o he A i icial Immune Ne wo k model applied o Clus e - ing echniques [40]. In he wo k o Elnaba awy e al. [27], he Biclus e ing me hod BARTMAP is used [41], al hough he au ho s don‘ cla i y i i is possible o use any o he echnique. A e ob aining he biclus e s, he ollowing s ep is o look o he a ge biclus e s whe e he nea es use s will be chosen. In gene al, o selec he a ge biclus e o biclus- e s i is necessa y o calcula e simila i y measu es be ween all he biclus e s and he ac i e use . This can be a cos ly ask conside ing he la ge numbe o biclus e s ha can be gene a ed. Some wo ks simpli y his ask, jus p ocessing only he biclus e s ha con ain he ac i e use . Fo example, in [22] a esidue is calcula ed only o hose biclus e s ha include his use . The biclus e wi h he smalles esidue and he i em o be a ed is chosen. Then, he ecommenda ion is gene a ed di ec ly as an a e age o he a ings o his i em in he biclus e . The biclus e s o which he ac i e use belongs a e de e mined in i s place in he wo k o Yolda e .al. [33]. Then, he op-N ecommenda ions a e hen chosen based on he join decision o hese biclus e s. Ano he in e es ing example could be he wo k p esen ed by Huan-huan e al. [42] whe e a scalable ecommende sys em based on biclus- e ing and mo h lame op imiza ion algo i hm is p oposed. Ano he ecen s udy, [17] p oposes a no el Biclus e ing me hod based on modi ied uzzy adap i e esonance heo y. This pape p oposes a new measu e o e lec he simila - i y be ween use s ha conside s he e ec o he numbe o common elemen s o he use s. In mos cases, howe e , a simila i y calcula ion p ocess is applied o all biclus e s. In some occasions, his p ocess is applied o sea ch o he nea es use s bu also o he nea es i ems o he one o be a ed. In [24], se e al nea es i ems o he i em o be a ed a e selec ed om he biclus e s, calcula ing a simila i y measu e in wo phases. The same p ocess is ca ied ou o he nea es use s o he ac i e one. Then, he p edic ion is ob ained by means o a o mula ha combines he p e iously calcula ed simila i ies. The nea - es K biclus e s o he ac i e use and he a ge i em a e ob ained in [30], compu ing he CjacMD (Cosine- Jacca d- Mean Measu e o Di e gence) simila i y measu e. In he wo k o Symeonidis e al. [23] and Sing e al. [32], a simi- la i y based on he numbe o common i ems is calcula ed o ind he K nea es biclus e s o he ac i e use . Then, a anking o he N op i ems is ex ac ed om hese biclus e s, based on he appea ance equency o each i em. In o he wo ks, he smalles biclus e o he ac i e use , ha is, he biclus e ha con ains his use and he g ea es numbe o i ems, is selec ed [25, 28]. To accomplish his, all biclus e s mus be p ocessed ia in e sec ion. In a second phase, he 1 3 Page 3 o 17 830 D. Rod íguez-Baena e al. The nea es use s a e calcula ed om wo di e en sou ces o in o ma ion in he wo k o Fa yad e al. [52]. In a i s s ep, a Cosine unc ion is used o calcula e he simila i y be ween use s based on he use -i em a ings da ase . Sec- ondly, demog aphic simila i ies be ween use s a e ob ained h ough a weigh ed a e age o hei demog aphic da a. The inal simila i ies be ween use s a e de e mined using a linea combina ion o bo h a ings and demog aphic simila i ies. Howe e , in many cases, he p oblem o spa si y a ec s no only he da a desc ibing use beha iou ( a ings), bu also he da a desc ibing he use s (e.g. pe sonal in o ma ion). To add ess his issue, a no el Induc i e He e ogeneous G aph Neu al Ne wo k (IHGNN) model is p esen ed [53]. This model con e s new use s, i ems, and associa ed mul- imodal in o ma ion in o a Modali y-awa e He e ogeneous G aph (M-HG), which p ese es he ich and he e ogeneous ela ionship in o ma ion among hem. The app oach p esen ed in his pape , BinRec, educes he complexi y o p ocessing biclus e s o ind he nea es use s o he ac i e one, imp o ing pe o mance on la ge da abases wi h high spa si y while main aining a high le el o accu acy in i s ecommenda ions. To manage he cold- s a issue, i is no necessa y o p ocess addi ional in o ma- ion abou use s, bu a he he in o ma ion on a ings and ela ionships be ween use s ob ained h ough biclus e s is used. Also, BinRec is lexible o using any Biclus e ing ech- nique, hough a speci ic one is p oposed o pa ially sol e he p oblem o gene a ing he en i e se o biclus e s again when new da a a e added o he inpu da ase . Finally, p io o applying he Biclus e ing echnique, a p ep ocessing ask is used o ine- une he p ecision o he ecommenda ions. 3 Ma e ials and me hods In his sec ion, he new app oach p oposed by au ho s, BinRec, is ully explained. The s a ing hypo heses a e he ollowing: ●In a biclus e , a sub-g oup o use s sha e a simila opin- ion abou a sub-g oup o i ems, so i is a e y in e es ing sou ce o knowledge o gene a e new ecommenda ions. ●The use s wi h whom he same as es a e sha ed a e hose wi h whom a g ea e numbe o biclus e s a e sha ed. The p oposal p esen ed in his wo k is di ided in o wo phases: o line and online. The o line phase comes be o e he ecommenda ion and consis s mainly o p ocessing he biclus e s ob ained om he a ing da abase. Du ing he online phase, he ecommenda ion is ca ied ou . Fu he - mo e, he o line and online phases a e sepa a ed in o wo candida e se o i ems o a ecommenda ion is ex ac ed om he biclus e neighbou hood o he smalles biclus- e . The Mean Absolu e Di e ence measu e is calcula ed o all he biclus e s in [29, 43] o ind he nea es biclus e ha has a s ong pa ial simila i y wi h he p e e ences o an ac i e use . Recen wo ks also apply speci ic algo i hms o compu e he neighbou hood o he ac i e use , such as Meme ic algo i hms in [34] o he Mo h Flame Op imiza- ion algo i hm in [35]. In [44], he biclus e s a e p ocessed and mapped in a squa e g id o ep esen di e en s a es in a Ma ko decision p oblem. To do so, biclus e s a e so ed by hei Scaling Mean Squa ed Residue (SMSR) alues and hen me ged o i and place hem in he squa e ma ix. In conclusion, as obse ed in he e iewed wo ks, educing he sea ch space h ough he gene a ion o biclus e s in ol es a massi e p ocessing e o o ind he nea es neighbou s. This ask can be compu a ionally demanding, conside ing he la ge numbe o biclus e s ha can be ob ained om inc easingly ex ensi e a ing da abases. Recen ad ancemen s in ecommende sys ems ex end beyond biclus e ing, pa icula ly u ilising g aph-based me hodologies o enhance pe o mance in spa se o in ica e use -i em con ex s. These app oaches ep esen in e ac ions as bipa i e o he e ogeneous g aphs. Fo ins ance, Li eGSR [45] e ines social g aphs using PageRank-based cen al- i y, educing agg ega ion o e head in GNN-based ecom- menda ions. KGCFRec [46] in eg a es knowledge g aphs wi h collabo a i e il e ing h ough a dual-channel GNN and adap i e a en ion, ou pe o ming baselines on mul iple da ase s. Simila ly, he G aph Con olu ional Recommen- da ion Sys em wi h Bila e al A en ion [47] enhances pe - o mance by combining use -i em and knowledge g aphs ia a en ion mechanisms. Beyond g aph-based me hods, Tu bo-CF [48] o e s as ecommenda ion h ough ma ix decomposi ion- ee il e ing, while PolyCF [49] uses spec- al g aph il e s o imp o e collabo a i e il e ing accu acy and scalabili y. Finally, he use s cold-s a p oblem appea s when he exis ing a ing da a abou a use a e insu icien , so he sys- em canno gene a e e icien ecommenda ions. I happens whene e he e a e use s who ha e no a ed any i em (new use s) o who ha e a ed a e y ew i ems (cold use s) [50]. To add ess his issue, mos o he p oposed solu ions ind di e en ways o associa e use s no using a ing da a bu o he ypes o in o ma ion. Fo example, in [26], he inpu da ase is clus e ed as a p ep ocessing s ep based on use s’ age and loca ion, no on hei a ings. O he wo ks, such as [51], use bo h explici ( a ings) and implici da a (b ows- ing his o y, pu chase and click ac i i y, e c.) abou use s o gene a e ecommenda ions. To do so, au ho s combine a P obabilis ic Ma ix Fac o iza ion based on a ing da a and a Bayesian Pe sonalized Ranking based on explici da a. 1 3 830 Page 4 o 17 BinRec: add essing da a spa si y and cold-s a challenges in ecommende sys ems wi h biclus e ing encoding, i use u ecommends an i em i, hen (u, i) is equal o 1; o he wise, i is equal o 0. A de ailed illus a ion o he o line phase is p o ided in Fig. 2 by means o an example. As i can be obse ed, he o line phase is di ided in o h ee s eps: he bina iza ion o he a ing ma ix, he gene a ion o he biclus e s and he c ea ion o he Biclus e s ma ix. The inpu da a o he example, A, is a 4x10 ma ix in which 4 use s a e 10 i ems using sco es om 1 o 5. The A(u, i)=−1 alue ep esen s ha i em i has no been a ed ye by use u. As i has been said be o e, he bina iza ion o he a ing ma ix is he i s s ep. This s ep is necessa y in o de o apply he BiBi biclus e ing algo i hm, which ope a es on bina y da abases. The bina iza ion p ocess—i.e., ans- o ming he da a in o ones and ze os— equi es a h eshold ha de e mines whe he an elemen will be con e ed o 1 o 0. Ra ings da abases a e usually based on a lis o dis- c e e alues, o example, a ings om 1 o 5 s a s. Thus, he h eshold will be a speci ic alue wi hin ha lis . Fo ins ance, i he h eshold is se o 3, i means ha all a ings equal o o g ea e han 3 will be con e ed o he alue 1, and he emaining alues o 0. The esul ing da abase high- ligh s hose a ings equal o o g ea e han 3, which will be conside ed posi i e ecommenda ions. So, i he possible sco e alues o he a ings ange om x o y, hen o each alue S ∈[x, y] , a new bina y ma ix B_S can be gen- e a ed, in which B_S (u, i)=1 i in he o iginal da ase A(u, i)≥S and i means ha i em i is ecommended by use u; B_S (u, i)=0 o he wise. In he example o Fig. 2, S =2 , so a new bina y ma ix is c ea ed ( B_2 ). As i will be explained in Sec ion 4, his p ep ocessing o e s he possibili y o in luencing he sensi i i y o he ecommende sys em. s ages (see Fig. 1). Following, e e y phase, along wi h i s s ages, is desc ibed. 3.1 O line phase The goal o he o line phase is o use he local beha iou pa e ns, ega ding use ecommenda ions, ex ac ed by he biclus e s o gene a e a ma ix, M, ha e lec s he ela ion- ships be ween di e en use s based on he simila i y o hei a ings. This ma ix will con ain, o each use , he numbe o biclus e s sha ed wi h o he s and will be used o quickly and easily iden i y he nea es use s o he ac i e one. When wo use s appea in he same biclus e , i means ha hey ha e ecommended he same subse o i ems. The e o e, i hey sha e many biclus e s, i indica es a highe le el o a ini y. As we ha e seen in he s a e-o - he-a e iew, he sea ch o nea es neighbou s in ol es applying complex measu es o all he biclus e s gene a ed o each ecom- menda ion, esul ing in a high compu a ional load. Wi h his new p oposal, a single da a s uc u e will be gene a ed, he ma ix M, which, wi hou complex calcula ions, will con- ain he necessa y in o ma ion o de e mine he simila i y be ween use s o all he new ecommenda ions. As i can be obse ed in Fig. 1, Biclus e ing is applied o he a ings da ase in he o line phase o gene a e biclus- e s, ha is, subse s o use s wi h simila a ing beha iou unde subse s o i ems. Al hough any Biclus e ing echnique can be used in ou p oposal, in his wo k, bina y Biclus e - ing app oaches ha e been p oposed. This kind o echnique adap s pe ec ly o he classi ica ion o ecommenda ions based on labels: no ecommended and ecommended. In addi ion, as will be seen below, he inpu da a can be p e- p ocessed in a ious ways due o hei bina iza ion. So, when wo king wi h use s and i ems o be a ed in a bina y Fig. 1 This is he gene al schema o he p oposal. I is di ided in o wo di e en phases: o line and online. Biclus e ing is applied o he a ing da ase in he o line phase o gene a e a se o biclus e s. A biclus e is de ined as a subg oup o use s wi h simila a ing beha iou unde a subg oup o i ems. Nex , hese biclus e s a e used o de ine he biclus- e s ma ix, M, ha will be used in he online phase du ing he selec ion o he nea es use s. In he online phase, he nea es use selec ion and ecommenda ion p ocess ake place 1 3 Page 5 o 17 830 D. Rod íguez-Baena e al. o bi s gene a ed as a esul o he AND ope a ion be ween hese ows. The po en ial biclus e is hen de i ed om his pai o ows and he columns in he pa e n ha a e equal o 1. Finally, new ows a e added o he biclus e i hey a e compa ible wi h ha pa e n. So, i a new use is added o he a ing da ase , o example U5 (see Fig. 3), i will no be necessa y o apply BiBi o he whole new bina y ma ix. Ins ead, Bibi will be pa ially applied, jus c ea - ing new pa e ns wi h he ow U5 and he es o ows (new biclus e Bic6) and checking i ha ow is compa ible wi h The p ocedu e o gene a ing biclus e s is he second s ep. In he example, i e biclus e s ha e been ex ac ed om he bina y ma ix. The Bibi Biclus e ing algo i hm was used o accomplish his [20]. Se e al easons a e gi en: Fi s and o emos , Bibi pe o ms g ea wi h bina y ma i- ces o all sizes and shapes [54]. Besides, when a new use is in oduced o he a ings da ase , Bibi o e s he op ion o de eloping an inc emen al biclus e s p ocessing a he han ha ing o p ocess he en i e inpu da ase again. BiBi gene a es a pa e n o each pai o ows, which is a g oup Fig. 3 In he example, a new use U5 has been added o he inpu a - ing da ase A (enhanced in ed). The bina y Biclus e ing algo i hm BiBi is applied only his new ow. Thus, Bibi gene a es a new pa - e n wi h ha ow and e e y one o he p e ious ows and om hese pa e ns a new biclus e is c ea ed (Biclus e Bic6). A he same ime, Bibi checks i he new ow is compa ible wi h he pa e ns ob ained in p e ious execu ions, modi ying he exis ing biclus e s i necessa y (changes in ed in biclus e s Bic1, Bic2 and Bic5) Fig. 2 In he i s s ep o he o line phase, he inpu a ings da ase , A, is bina ized using a h eshold alue ( S =2 in he example). Then, a new bina y e sion o A is c ea ed, B2 . In he second s ep, a bina y Biclus e ing algo i hm (BiBi ) is applied o B2 , ex ac ing a g oup o biclus e s. Fo example, he Biclus e Bic1 is composed by wo use s, U1 and U2, bo h ecommending he i ems I4 and I10. Finally, in he hi d s ep, a use s squa e ma ix, M, is c ea ed om he biclus e s gene a ed 1 3 830 Page 6 o 17 BinRec: add essing da a spa si y and cold-s a challenges in ecommende sys ems wi h biclus e ing sha es a conc e e numbe o biclus e s wi h UJ, ep esen ed by he weigh o ha edge. In ou example, U2 is connec ed wi h he es o use s and he sum o i s edges ep esen s he highes alue, 5. The e o e, U2 will be chosen o ep esen he a ing da ase ’s o e all opinion and his/he ecommen- da ions will be assign o he new use s. 3.2 Online phase The online phase is di ided in o wo s eps. The ac i e use ’s nea es use s a e de e mined in he i s s ep. The ecom- menda ion p ocess is hen ca ied ou based on he nea es use s. Following, he examples o Figs. 5 and 6 a e used o clea ly explain he online phase. BinRec, unlike mos Biclus e ing-based ecommenda- ion p oposals (see Sec ion 2), employs a simple measu e based on he M ma ix a he han calcula ing complex simi- la i y measu es ha en ail p ocessing all o he biclus e ele- men s. Ins ead, i is assumed ha a use who sha es a la ge numbe o biclus e s wi h ano he use can de e mine ha his o he opinion is mo e impo an han o he use s’ opin- ions. Le ’s suppose ha he ac i e use is U2. To ind i s nea es use s, he column # o M, which con ains he a e - age o he numbe o biclus e s ha a speci ic use sha es wi h he es , is used (see Fig. 5). To calcula e ha a e - age, only hose use s wi h whom e e y use sha es biclu - e s a e aken in o accoun . So, he nea es use s o U2 a e hose wi h whom U2 sha es a numbe o biclus e s g ea e o equal han ha a e age (1.6), ha is, U3 (he/she sha es 2 biclus e s wi h U2) and U4 (he/she sha es 2 biclus e s wi h U2). As i can be obse ed, he use U1 is disca ded because he/she only sha es 1 biclus e wi h U2. Nex , he po en ially ecommendable i ems a e selec ed, ha is, hose ha he ac i e use had no a ed up o ha momen : I1, I2, I6 and I8. To know which o hese i ems will be ecommended, he pa e ns c ea ed in p e ious execu ions (modi ica ions in Bic1, Bic2 and Bic5). Nex , in he hi d s ep, a Biclus e s ma ix M is c ea ed. M is a squa e ma ix and i s dimension is he numbe o use s in he da abase. The alue o M(I, J) e lec s he numbe o biclus e s sha ed by use s UI and UJ, whe eas he main diagonal ep esen s he numbe o biclus e s in which e e y use UX occu s (see Fig. 2). This ma ix ul ils wo impo - an unc ionali ies (see Sec ion 3.2): M will be used in he online phase du ing he nea es use selec ion p ocess and M also be used o de e mine he use who appea s in he la g- es numbe o biclus e s. This use will be used in he online phase o add ess he cold-s a issue [55]. The hypo hesis is ha he use who appea s in he highes numbe o biclus- e s is he one ha b ings oge he he g ea es numbe o simila as es wi h he es o he use s. In Fig. 4, a g aph gene a ed om M is shown. In his g aph, e e y node is a use and e e y edge, om UI o UJ, de e mines ha UI Fig. 5 The i s s age o online phase is ep esen ed in his image, in which he nea es use s o he ac i e use , U2, a e de e mined using he Biclus e s Ma ix M. M s o es, o e e y use , he numbe o biclus e s sha ed wi h he es o use . The las column, # , s o es he a e age o he numbe o biclus e s sha ed o e e y use . In he case o U2, he use sha es biclus e s wi h U1 (1 biclus e ), U3 (2 biclus e s) and U4 (2 biclus e s). Since he nea es use s a e hose wi h whom a use sha es a numbe o biclus e s equal o o g ea e han he a e age (1.6 in he case o U2), he U3 and U4 use s a e selec ed. In he inpu da ase A, he po en ially ecommendable i ems om U2 a e highligh ed in ed Fig. 4 This igu e shows a g aph gene a ed om ma ix M. E e y node is a use and e e y edge, om UI o UJ, de e mines ha UI sha es a conc e e numbe o biclus e s wi h UJ, ep esen ed by he weigh o ha edge. U2, as i can be obse ed, is he use which appea s in he la ges numbe o biclus e s 1 3 Page 7 o 17 830 D. Rod íguez-Baena e al. in common wi h a ce ain use , he mo e aluable hei opin- ion will be in gene a ing ecommenda ions. Fu he mo e, i he Biclus e ing echnique is in eg a ed in o he BinRec amewo k, his da a s uc u e can be gene a ed in pa allel wi h he biclus e s. Besides ha , he ecommenda ion p o- cess o a speci ic use can be ca ied ou independen ly, esul ing in signi ican scalabili y. As a esul , because he sea ch o nea es use s is execu ed du ing he online phase (on demand), he pu pose o his p oposal is o educe p o- cessing complexi y. Finally, he cold-s a p oblem is sol ed by aking ad an age o he in o ma ion s o ed in he biclus- e s ma ix M. In he ollowing sec ion, he way BinRec beha es unde di e en condi ions is analysed. Besides, a compa ison be ween he new app oach and exis ing UBCF me hods is ca ied ou . 3.3 Expe imen a ion Following, he me hodology used in he expe imen a ion, along wi h he desc ip ion o he da ase s and measu es used, a e in oduced. 3.3.1 Expe imen s wo k low E e y expe imen ollows he gene al schema shown in Fig. 7 and is di ided in o 4 di e en phases. The inpu consis s o a a ing da ase A and a h eshold alue o bina iza ion, S , chosen om he da ase ’s ange o possible sco es, wi h ex eme alues disca ded. In ou case, because all o he da ase s used in he expe imen a ion ha e sco es anging om 1 o 5, S can be 2, 3, o 4 (1 and 5 a e disca ded). Fol- lowing he selec ion o S , he aining and es da ase s a e gene a ed. To do so, a subse o he inpu da ase is chosen he second s ep is ca ied ou (see Fig. 6). I is an i e a i e p ocess in which a new a ing alue, Final Rec, is gene a ed o each i em selec ed in he p e ious s ep by calcula ing he a e age o he a ing alues om he nea es use s. The a - ing alues equal o −1 a e igno ed. Finally, o decide i an i em is ecommended o no , he new a ing alue mus be equal o g ea e o he sco e alue used in he o line phase o ans o m he inpu ma ix A in o a bina y ma ix, B_S ( S =2 in ou example). The new a ing alues o ac i e use U2 a e gene a ed in he example o Fig. 6. In he case o i em I1, he a e - age o he opinions o he nea es use s, A(U3,I1) = 5 and A(U4,I1) = 1 , is used o gene a e a new a ing alue: Final Rec = (5 + 1)/2=3>=S . As his new alue is g ea e han S =2 , he i em I1 is conside ed as ecom- mended. On he con a y, in he case o I2, he use U3 has no a ed ye his i em, A(U3,I2) = −1 , so i s opin- ion is no aken in o accoun . Then, he new a ing alue is Final Rec = A(U4,I2)/1=1/1=1< S , so I2 is no ecommended. Finally, he cold-s a issue is add essed. When a new use is added o a ecommenda ion sys em, commonly i has no i ems a ed ye . This is called a cold-s a p oblem [51] and i implies ha his new use will no be pa o any biclus e . In hese cases, he solu ion p oposed by his wo k is o use he ecommenda ions om he mos connec ed use , ha is, he use ha appea s in he la ges numbe o biclus e s (see Sec ion 2). In ou example, U2 is he mos connec ed use , so i is selec ed as ep esen a i e o he opinions o he es o use s. As a conclusion, he use o he biclus e ma ix, M, implies a oiding he calcula ion o a complex simila i y measu e be ween he ac i e use and all he biclus e s gene - a ed. The idea behind i is ha he mo e biclus e s you ha e Fig. 6 The second s ep o he online phase is ep esen ed in his image. Fo e e y i em no a ed ye by he ac i e use U2, a new a ing alue, Final Rec, is gene a ed by he a e - age o he a ing alues o he nea es use s (U3 and U4). The i ems no a ed a e igno ed. I he new a ing alue is g ea e o equal o S , he i em is ecommended 1 3 830 Page 8 o 17 BinRec: add essing da a spa si y and cold-s a challenges in ecommende sys ems wi h biclus e ing 6040 use s abou 39523 mo ies (1 million a ings). The CiaoDVD da ase has 278,483 a ings on 99,746 i ems p o- ided by 7,375 use s. Also, hese da ase s ha e been used in he expe imen a- ion because o hei high spa si y le el. Spa si y e e s o he phenomenon ha occu s when he numbe o use s and i ems in a a ings da abase is e y la ge, bu he numbe o a ings gi en by use s o i ems is e y low. The e o e, a a - ings da abase wi h a spa si y le el o X% only includes ha pe cen age o all possible a ings. Speci ically, Fig. 8 shows he le el o spa si y o he h ee da ase s. I can be seen ha CiaoDVD has he highes alue (99,9%). These le els o spa si y make hem sui able o measu ing he pe o mance o he p oposals p esen ed in he con ex o ecommenda- ion sys ems. 3.3.3 Pe o mance me ics A ecommende sys em’s pe o mance is ypically assessed using wo ypes o me ics: label-based o alue-based [58]. In classi ica ion, label-based measu es a e used o de e mine he p ecision wi h which a class is assigned. In ou case, he e a e wo classes o i ems: ecommended by a use and no ecommended by a use . Thus, like in simila wo ks [35], P ecision and Recall measu es a e used in his pape . To compu e hese me ics, a con usion ma ix mus be c e- a ed [59], in which he ou di e en cases a e conside ed: ●TP ( ue posi i e): The algo i hm co ec ly ecom- mends an i em I ha is ac ually ele an o he use U. and s o ed in he o ma Use - I em - Value (Tes da ase , T). The c i e ia o ha selec ion a e as ollows: he elemen s a e chosen a andom, wi h he idea ha he inal selec ion mus be balanced, wi h hal based on i ems ecommended by use s ( a ing alue equals o exceeds S ) and he o he hal based on i ems no ecommended by use s ( a ing alue less han o equal o S ). Then, he a ing alues o T a e eplaced in he inpu da ase wi h he alue −1 and nex A is bina ized using S as a h eshold, esul ing in he aining da ase A′ . To de e mine he numbe o elemen s in he es da ase , he 20% o he o al ecommenda ions p esen in he o iginal da abase ha been used. Following, he BiBi bina y Biclus e ing algo i hm is applied o he aining da ase A′ ( see jus i ica ion in Sec- ion 2). In phase 3, new a ing alues o he pai s Use - I em s o ed in T a e p edic ed. Finally, in phase 4, hese new a ing alues a e compa ed o hose s o ed in T, and se e al pe o mance measu es a e ob ained o assess he quali y o he new ecommenda ions. 3.3.2 Da ase s desc ip ion The da ase s used in he expe imen a ion a e he Mo ieLens 100 K and 1M da ase s [56] and he CiaoDVD da ase [57]. These h ee da ase s a e conside ed he s anda d da ase s in e alua ing he ecommenda ion echniques. They include he use ’s a ings o mo ies using he 5-poin a ing scale; ha is, he sco e alue 5 is highly liked, and he sco e alue 1 is mos disliked. The da ase Mo ieLens_100K is composed o a ings om 943 use s abou 1682 mo ies (100,000 a - ings). The da ase Mo ieLens_1M includes he opinion o Fig. 7 This is he gene al schema ollowed by all he expe imen s. I is di ided in o ou di e en phases. The inpu consis s o he inpu a ings da ase , A, and he h eshold alue o bina iza ion, S . The aining ( A′ ) and es (T) da ase s a e gene a ed in he i s phase. Then, biclus e s a e gene a ed and ecommenda ions a e p edic ed based on he es da ase using he no el BinRec me hodology. Finally, se e al pe o mance measu es a e calcula ed in phase 4 1 3 Page 9 o 17 830 D. Rod íguez-Baena e al. 4. Khadija A, Almohsen H (2015) Recommende sys ems in ligh o big da a. In J Elec ic Compu Eng 5 5. Zhang Q, Lu J, Jin Y (2021) A i icial in elligence in ecom- mende sys ems. Complex In ell Sys 7:439–457 6. Fkih F (2022) Simila i y measu es o collabo a i e il e ing- based ecommende sys ems: Re iew and expe imen al compa i- son. J King Saud Uni e -Compu In Sci 34:7645–7669 7. Sai udin I (2024) & Widiyaning yas, T. App oach, p oblem, e al- ua ion echniques, da ase s. IEEE Access, Sys ema ic li e a u e e iew on ecommende sys em 8. Nan hini M, P adeep Mohan Kuma K (2022) Cold s a and da a spa si y p oblems in ecommende sys em: A concise e iew, 107–118 (Sp inge ) 9. Ismai S, Özlem N, Özgü U (2013) Clus e sea ching s a egies o collabo a i e ecommenda ion sys ems. In P ocess Manage 49:688–697 10. Heida i N, Mo adi P, Koocha i A (2022) An a en ion-based deep lea ning me hod o sol ing he cold-s a and spa si y issues o ecommende sys ems. Knowl-Based Sys 256:109835 11. Tahmasebi F, Meghdadi M, Ahmadian S (2021) A hyb id ecom- menda ion sys em based on p o ile expansion echnique o alle i- a e cold s a p oblem. Mul imed Tools Appl 80:2339–2354 12. Resnick P, Iaco ou N, Suchak M, Be gs om P, Riedl J (1994) G ouplens: An open a chi ec u e o collabo a i e il e ing on ne - news, 175–186 13. Ko H, Lee S, Pa k Y, Choi A (2022) A su ey o ecommenda ion sys ems: ecommenda ion models, echniques, and applica ion ields. Elec onics 11:141 14. Shen J, Zhou T, Chen L (2020) Collabo a i e il e ing-based ecommenda ion sys em o big da a. In J Compu Sci Eng 21:219–225 15. Papadakis H, Papag igo iou A, Panagio akis C, Kosmas E, F ago- poulou P (2022) Collabo a i e il e ing ecommende sys ems axonomy. Knowl In Sys 64:35–74 16. Sim K, Gopalk ishnan V, Zimek A e a (2013) A su ey on enhanced subspace clus e ing. Da a Min Knowl Disc 26:332–397 17. Sun J, Zhang Y (2022) Recommenda ion sys em wi h biclus e - ing. Big Da a Mining Anal 5:282–293 18. Madei a S, Oli ei a A (2004) Biclus e ing algo i hms o biologi- cal da a analysis: a su ey. IEEE/ACM T ans Compu Biology Bioin 1:24–45 19. Cas anho E N, Aidos H, Madei a S C (2024) Biclus e ing da a analysis: a comp ehensi e su ey. B ie Bioin 25:bbae342 20. Rod iguez-Baena DS, Pe ez-Pulido A, Aguila -Ruiz JS (2011) A biclus e ing algo i hm o ex ac ing bi -pa e ns om bina y da ase s. Bioin o ma ics 27:2738–2745 21. Sil a G, Madei a C, Rui S (2024) A comp ehensi e su ey on biclus e ing-based collabo a i e il e ing. ACM Compu Su 56. h p s : / / d o i . o g / 1 0 . 1 1 4 5 / 3 6 7 4 7 2 3 22. de Cas o P, de F anca F, Fe ei a H, Von Zuben F (2007) Apply- ing biclus e ing o pe o m collabo a i e il e ing, 421–426 23. Symeonidis P, Nanopoulos A, Papadopoulos A, Manolopoulos Y (2008) Nea es -biclus e s collabo a i e il e ing based on con- s an and cohe en alues. In Re ie al 11 24. Zhang D e al (2014) Cold-s a ecommenda ion using bi-clus e - ing and usion o la ge-scale social ecommende sys ems. IEEE T ans Eme g Topics Compu 2:239–250 25. Alqadah F, Reddy C, Hu J, Alqadah H (2015) Biclus e ing neigh- bo hood-based collabo a i e il e ing me hod o op-n ecom- mende sys ems. Knowl In Sys 44:475–491 26. Desai T, e al (2016) An en e p ise- iendly book ecommenda- ion sys em o e y spa se da a, 211–215 27. Elnaba awy I, Wunsch D, Abdelba A (2016) Biclus e ing a map collabo a i e il e ing ecommende sys em, 2986–2991 deep lea ning and a ional echniques o imp o ing pe - sonaliza ion, al hough hese me hods end o ha e a highe demand on compu a ional esou ces. In his con ex , he use o biclus e s o educe he sea ch space and acili a e ecom- menda ion in spa se en i onmen s is shown o be scalable. In addi ion, he e is a g owing end owa ds he c ea ion o hyb id RSs ha combine he bene i s o Biclus e ing wi h ma ix ac o ing and deep lea ning models, which could o e a balance be ween accu acy and e iciency in u u e ecommende sys ems. Au ho con ibu ions Domingo S. Rod íguez-Baena was esponsible o he concep ion and design o he s udy. Ma e ial p epa a ion, da a collec ion and analysis we e pe o med by Domingo S. Rod íguez- Baena. All au ho s w o e he i s d a o he manusc ip and com- men ed on ea lie e sions o he manusc ip . All au ho s ead and ap- p o ed he inal manusc ip . Funding Funding o open access publishing: Uni e sidad Pablo de Ola ide/CBUA. Da a a ailabili y All da a used is a ailable a he ollowing u l: h p s : / / g o u p l e n s . o g / d a a s e s / m o i e l e n s / and h p s : / / w w w . c s e . m s u . e d u / a n g j i l i / d a a s e c o d e / u s s u d y . h m BinRec sou ce code is a ailable a h p s : / / g i h u b . c o m / d s o d b a e / B i n R e c. Decla a ions E hical and in o med consen o da a used The au ho s ha e no con- lic s o in e es o e hics. The da a used in he documen a e public and accessible o esea ch pu poses. Con lic s o in e es The au ho s decla e ha hey ha e no con lic s o in e es . Open Access This a icle is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional License, which pe mi s use, sha ing, adap a ion, dis ibu ion and ep oduc ion in any medium o o ma , as long as you gi e app op ia e c edi o he o iginal au ho (s) and he sou ce, p o ide a link o he C ea i e Commons licence, and indica e i changes we e made. The images o o he hi d pa y ma e ial in his a icle a e included in he a icle’s C ea i e Commons licence, unless indica ed o he wise in a c edi line o he ma e ial. I ma e ial is no included in he a icle’s C ea i e Commons licence and you in ended use is no pe mi ed by s a u o y egula ion o exceeds he pe mi ed use, you will need o ob ain pe mission di ec ly om he copy igh holde . To iew a copy o his licence, isi h p : / / c e a i e c o m m o n s . o g / l i c e n s e s / b y / 4 . 0 /. Re e ences 1. Sha ma R, Singh R (2016) E olu ion o ecommende sys ems om ancien imes o mode n e a: A su ey. Indian J Sci Technol 9 2. Da oodi E, Kianmeh K, A sha chi M (2013) A seman ic social ne wo k-based expe ecommende sys em. Appl In ell 39:1–13 3. Ahmad HK, Qi C, Wu Z, Muhammad BA (2023) Abine-c s: cou se ecommende sys em in online educa ion using a ibu ed bipa i e ne wo k embedding. Appl In ell 53:4665–4684 1 3 830 Page 16 o 17 BinRec: add essing da a spa si y and cold-s a challenges in ecommende sys ems wi h biclus e ing 49. Qin Y, Ju W, Gu Y, Zhang M (2025) Polyc : Towa ds op imal spec al g aph il e s o collabo a i e il e ing. ACM T ans In Sys (TOIS) 43:1–25 50. Roy D, Du a M (2022) A sys ema ic e iew and esea ch pe - spec i e on ecommende sys ems. J Big Da a 9 51. Feng J, Xia Z, Feng X, Peng J (2021) Rbp : A hyb id model o he new use cold s a p oblem in ecommende sys ems. Knowl- Based Sys 214 52. Fa yad T, Majid M, Sajad A (2021) A hyb id ecommenda ion sys em based on p o ile expansion echnique o alle ia e cold s a p oblem. Mul imed Tools Appl 80:2339–2354 53. Cai D, Qian S, Fang Q, Hu J, Xu C (2023) Use cold-s a ec- ommenda ion ia induc i e he e ogeneous g aph neu al ne wo k. ACM T ans In Sys 41 54. Lopez-Fe nandez A, Rod iguez-Baena D, Gomez-Vela F, Di ina F, Ga cia-To es M (2021) A mul i-gpu biclus e ing algo i hm o bina y da ase s. J Pa allel Dis ib Compu 147:209–219 55. Scia one F, Yada U, Duhan N, Bha ia K (2020) Dealing wi h pu e new use cold-s a p oblem in ecommenda ion sys em based on linked open da a and social ne wo k ea u es. Mobile In Sys 2020 56. Ha pe F, Kons an J (2015) The mo ielens da ase s: His o y and con ex . ACM T ans In e ac In ell Sys 5 57. Guo G, Zhang J, Thalmann D, Yo ke-Smi h N (2014) E a : An ex ended us an eceden s amewo k o us p edic ion, 540–547 58. Desai M, Ansa i N (2023) An inno a i e me hod o inc ease ag i- cul u al p oduc i i y using machine lea ning-based c op ecom- menda ion sys ems, IEEE, 645–651 59. Hasnain M e al (2020) E alua ing us p edic ion and con u- sion ma ix measu es o web se ices anking. Ieee Access 8:90847–90861 60. Kuan M, Mohapa a P (2021) Assessmen me hods o e alua- ion o ecommende sys ems: a su ey. Found Compu Dec Sci 46:393–421 61. Hug N (2020) Su p ise: A py hon lib a y o ecommende sys- ems. J Open Sou ce So w 5:2174 62. Ukey N e al (2023) Su ey on exac knn que ies o e high- dimensional da a space. Senso s 23:629 63. Jain K, Jindal R (2023) Sampling and noise il e ing me hods o ecommende sys ems: A li e a u e e iew. Eng Appl A i In ell 122:106129 64. Yan Y, Mo eau C, Wang Z, Fan W, Fu C (2024) T ans o ming mo ie ecommenda ions wi h ad anced machine lea ning: A s udy o nm , s d, and k-means clus e ing, IEEE, 178–181. 65. Ai en S, Ag awal J (2023) Mo ie ecommende sys em using pa ame e uning o use and mo ie neighbou hood ia co-clus- e ing. P ocedia Compu Sci 218:1176–1183 Publishe 's No e Sp inge Na u e emains neu al wi h ega d o ju is- dic ional claims in published maps and ins i u ional a ilia ions. 28. Chand alekha M, Sa anya K, Sudha-Sadasi am G (2016) Biclus- e ing based collabo a i e il e ing algo i hm o pe sonalized web se ice ecommenda ion. In J Compu Appl 142:18–24 29. Vi nala S, Vi ekanandan K (2018) A no el biclus e ing wi h mean absolu e di e ence simila i y measu e o collabo a i e il- e ing ecommende sys em. In J Pu e Appl Ma h 118 30. Kan S, Maha a T (2018) Nea es biclus e s collabo a i e il e ing amewo k wi h usion. J Compu Sci 25:204–212 31. Choi S, e al (2018) Rein o cemen lea ning based ecommende sys em using biclus e ing echnique 32. Singh M, Meh o a M (2018) Impac o biclus e ing on he pe - o mance o biclus e ing based collabo a i e il e ing. Expe Sys Appl 113:443–456 33. Yolda M, Özcan U (2019) Collabo a i e a ge ing: Biclus e ing- based online ad ecommenda ion. Elec on Comme ce Res Appl 35 34. Bansal S, Baliyan N (2020) Bi-ma s: A bi-clus e ing based meme ic algo i hm o ecommende sys ems. Appl So Compu 97 35. Wu H, Ke G, Wang Y, Chang Y (2022) P edic ion on ecom- mende sys em based on bi-clus e ing and mo h lame op imiza- ion. Appl So Compu 120 36. Ka a ia S, Ba a U (2023) Implemen a ion o nea es co-clus e collabo a i e il e ing using a no el simila i y index. Indian J Sci Technol 16:2204–2216 37. P elić A, e al (2006) A sys ema ic compa ison and e alua ion o biclus e ing me hods o gene exp ession da a. Bioin o ma ics 1 38. Li G, Ma Q, Tang H, Pa e son A, Xu Y (2009) QUBIC: a quali a- i e biclus e ing algo i hm o analyses o gene exp ession da a. Nucleic Acids Res 37 39. Mu ali T, Kasi S (2003) Ex ac ing conse ed gene exp ession mo i s om gene exp ession da a 8:77–88 40. Abbass H A, Sa ke R A, New on C S (2001) aine : An a i icial immune ne wo k o da a analysis, 231–260 41. Rui X, Donald C (2011) Ba map: A iable s uc u e o biclus e - ing. Neu al Ne w 24:709–716 42. Wu H-H, Ke G, Wang Y, Chang Y-T (2022) P edic ion on ecom- mende sys em based on bi-clus e ing and mo h lame op imiza- ion. Appl So Compu 120:108626 43. Sa a anan S, B i o A, P abin S (2023) No el k means biclus e - ing usion based collabo a i e ecommende sys em. Cloud Da a Sci 124:607–616. h p s : / / d o i . o g / 1 0 . 4 0 2 8 / p - 5 1 0 d 44. A a I, e al (2024) A ein o cemen lea ning ecommende sys- em using bi-clus e ing and ma ko decision p ocess. Expe Sys Appl 237 45. Ga S, Cho PH, Moon GE, Jung S (2025) E icien gnn-based social ecommende sys ems h ough social g aph e inemen . The J Supe compu 81:1–24 46. Peng J e al (2024) Kgc ec: Imp o ing collabo a i e il e ing ec- ommenda ion wi h knowledge g aph. Elec onics 13:1927 47. Yang H, Yang C (2024) G aph con olu ional ecommenda ion sys em based on bila e al a en ion mechanism. J Eng 48. Pa k J, Shin Y, Shin W (2024) Tu bo-c : Ma ix decomposi ion- ee g aph il e ing o as ecommenda ion, 1234–1243 (ACM) 1 3 Page 17 o 17 830