Issuu

Mood-Based Music Recommendation System Using Facial Expression Recognition and Hybrid Filtering

Riya

Sharma1 ,

Namrita Kanoujiya

Shreya

Patel3, Devesh Katiyar4, Gaurav Goel5

¹ ² ³ Students, ´ µ Assistant Professors

Dr. Shakuntala Misra National Rehabilitation University, Lucknow, Uttar Pradesh, India.

Abstract - Imagine you come home after a bad day when nothing went right. You feel stressed and your mind is still overthinking. You open your music app hoping to hear something that matches your tired mood. Instead it fires up a playlist full of the upbeat songs you were playing last weekend. Technically the app did its job. But it got the moment completely wrong. This paper is about a system called MBMRS, Mood-Based Music Recommendation System, that we built to handle exactly that problem. But instead, it plays happy, energetic songs from your old playlist, which doesn’t match how you feel. This paper talks about a system called MBMRS. It looks at your face using your device’s front camera, understands your mood, and then suggests music based on it.

Key Words: Mood Detection; Music Recommendation; Facial Expression Recognition; Deep Learning; Affective Computing; CNN; Personalization;Context-AwareSystems.

1. INTRODUCTION

Our interest in this topic came from a pretty ordinary observation. Music apps have become genuinelygoodattrackingpreferences. Whenyoulisten to many songs over time, the app learns what kind of musicyoulike.Itknowsyourfavouritesingersandwhat type of songs you usually play at different times. That trackingisrealanditworksreasonablywellmostofthe time.

Butpreferencesandpresentstatearetwoseparate things. Your preference for jazz or hip hop or old film songsdoesnot changemuchfromTuesdaytoThursday. Your emotional state absolutely does. But sometimes your mood changes. You may feel tired and don’t want the same energetic songs. The app cannot understand yourmoodatthatmoment,soitstillplaysthesamekind of music. From its perspective, Thursday evening is the sameaseveryothereveningbecauseyourlisteningdata looksthesame.

Whatthatcreatesisagapbetweenwhatthesystem predicts you want and what you actually need in the moment. We kept noticing this personally during the project.Youopentheappwantingsomethingtomatcha

specificfeelingandinsteadyougetasuggestionthatfits your general profile but completely misses where you are right now. The frustration is not dramatic but it is consistent.

The system we put together tries to close that gap by reading something the apps currently ignore completely, which is the emotional information visible on a face. Every phone and most laptops have a front camerasittingthereunusedduringmusicsessions.That camera can capture enough to run a facial expression classifier. If the classifier can reliably say whether someonelookstired oranxious orupbeat,that becomes ausefulinputintopickingwhattoplaynext.

Thegoodnewsisthatthisisnotfuturistic.Emotion classifiers trained on datasets like FER2013 are hitting 75 percent accuracy and above on genuine real-world images right now [1]. Models designed for mobile hardware, MobileNet being the main one, can process a camera frame in real time without the phone getting warm or the battery draining noticeably [2]. The combination of those two things is what made this projectfeeldoableratherthantheoretical.

What follows is structured this way. Section II coverstheresearchwedrewfromacrossthreedifferent fields. Section III walks through how we designed the overall system. Section IV gets into the actual technical build. Section V covers performance during testing. Section VI is an honest account of the problems. Section VIIisaboutwherethisworkcouldgo.SectionVIIIwraps up.

2. LITERATURE REVIEW

Putting this together required reading across three separate bodies of work. How computers detect emotions from facial images. How researchers have characterised the emotional properties of music. And howrecommendationenginesworkandhavedeveloped. The parts below summarise what we found most useful ineacharea.

2.1 Facial emotion recognition

There is a popular dataset called FER2013 that many researchers use for facial emotion recognition. It

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

hasaround36,000blackandwhiteimagesoffaces.Each imageislabelledwithoneofsevenemotions(likehappy, sad,angry,etc).Whatdistinguisheditfromtheacademic datasets that came before it was that the images were not staged. They came from internet photos, which meansimperfectlighting,variedheadpositions,partially visible faces, all the messiness of pictures taken in real situations. Goodfellow and co-authors [3] demonstrated that neural networks trained on this kind of data performed significantly better when tested in practice than models trained on the cleaner, posed collections thatresearchershadpreviouslyreliedon.

In 2016 K.He and colleagues [4] published the paper on residual networks that genuinely altered how people approached building deep models. The difficulty uptothatpointwasthataddinglayerstoa network did not reliably improve it. Gradient signals would weaken as they propagated back through many layers and the earlier parts of the network would fail to update properly. The residual approach inserted bypass connections so that the gradient could travel more directly. That made training very deep architectures stable and predictable. Nearly all the emotion classifiers builtintheyearssincearebasedonthisstructuralidea. The companion practice of starting from a network already trained on a large general image dataset like ImageNet and then adapting it for emotion recognition specifically became standard because it reliably beat buildingfromnothing[5].

AffectNet [6] moved away from the fixed category approach. Rather than forcing a label from a preset list onto every face, it scored each image on two independent scales. One measured the quality of the emotion, whether it leaned positive or negative. The second measured the intensity dimension, running from energised to calm. This framing draws from theoretical work Russell did on how emotions relate to each other [7]. The reason it matters for our project is that music can be placed on a very similar kind of grid. An upbeat track in a major key and a slow minor key track occupy very different positions on those same two axes. That structural similarity is what lets us build a principled link between a detected emotional state and a class of songs that fits it. Subsequent work extended CNN architectures with attention modules that direct the model to concentrate on specific areas of the face, particularly around the eyes, brows, and corners of the mouth,whereexpressionismostconcentrated[8].

Two papers that came out fairly recently gave us specific confidence in the direction we were taking. One from 2024 combined ResNet50 with a visualisation method called GRAD-CAM, which highlights the image regions the model weighted most heavily in making its decision [9]. That system reached 82 percent accuracy

on emotion classification and showed users exactly why each song recommendation was generated, a transparency feature nothing in commercial music apps currentlyoffers.Apaperpublishedinearly2025showed a live working implementation using the DeepFace libraryreadingfromawebcamandpopulatingaplaylist in near real time [10]. Those results confirmed that the approach holds up outside of controlled research settings.

2.2 Music Emotion Analysis

Audio analysis tools, Librosa [11] being probably the most used in academic work, extract a set of measurable properties from any music file. Tempo, the tonalkey,overallenergylevel,loudness,andthebalance between vocal and instrumental content. Research in music psychology going back decades has shown these measurements correspond to perceived emotional character in consistent ways. Energetic, fast songs in major keys register as positive and lively. Slow, soft songs in minor keys read as heavy or sad. This predictablemappinggivesusastartingframework.Once we know the emotional region a user is in, we can set numerical targets on those audio properties and filter thesonglibraryaccordingly.

Efforts to make emotion labelling more precise havepushed towardcombined systemsthat assign both categorical and continuous scores to music at the same time [12]. The technical difficulty is that different research groups built their datasets using different annotation schemes, which makes direct comparison between them unreliable. Some teams have moved to recurrent architectures that follow the shifting emotional character of a song across time rather than collapsing the whole thing into one summary label [13]. That gives a more granular picture of what a track is doing emotionally and likely improves the precision of matches.

2.3 Recommendation Systems

Collaborative filtering [14] is what drives recommendations on most large platforms. The method works by finding shared patterns across the listening behavior of many users and using those patterns to make predictions about what any individual user would choose next. It scales well and the predictions improve as more data accumulates. The consistent limitation is what happens with someone new. No listening history means no patterns to match against and no basis for a prediction. Content-based filtering sidesteps this by ignoringuserhistoryentirelyandworkingdirectlyfrom thepropertiesofthemusic.Forourpurposesthatmeans filteringbyemotionalaudiofeaturesratherthanbywhat others have chosen. Research consistently found that

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net

combining both approaches outperforms either one runningalone[15].

Layeringininformationaboutcontext,whetherthe userrecentlyexercised,whattimeitis,whethertheyare indoors or out, has been shown to improve how appropriate recommendations feel by a meaningful margin on top of mood alone [16]. Neural collaborative filtering [17] replaced the earlier matrix factorization approach by using a deeper model architecture to capture interaction patterns between users and content that a linear model would miss. It has become the baselineforseriousworkinthisarea.

3. SYSTEM DESIGN AND METHODOLOGY

Before playing any song, the system follows four steps one by one. We kept the overall structure simple intentionally.Systemswith manymoving partsarehard to test properly and when something goes wrong it is hardtotellwhichpartcausedit.

3.1. Reading the Room

Thepipelineopenswiththecamera.Laptopsrunat oneframepersecond.Phonesdroptoaframeeverytwo or three seconds, not for technical reasons but because runningthecamerafasterthanthatstartstoshowupin battery consumption and people notice and resent it quickly. Each frame it captures is like a photo of everythinginfrontofthedevice- itcouldincludechairs, walls,lights,people,oranythingnearby.Thesystemhas no use for any of that. The actual task at this stage is locating the face within the frame and isolating it from everythingsurroundingit.

FordetectionwewentwithMTCNN[18].TheViolaJones approach has been around for a long time and is computationallycheap,butitconsistentlylosesaccuracy when the face is not square to the camera, when part of it is covered, or when the lighting is uneven. In a home setting those conditions are normal. MTCNN was designedwiththatkindofreal-worldmessinessinmind and performed noticeably better in our own testing under poor lighting and natural head positions. After detectingtheface,asmartmodel(MobileNetCNN)looks at it and tries to understand the person’s emotion. It gives the score for each feeling like happy, sad, angry, scared, surprised, disgusted, or normal and decides whichoneismostlikely.

But relying on just one photo (one frame) was not very accurate, because emotions can change quickly or the system can make mistakes from a single moment Earlyindevelopmenttheplaylistwouldshiftbasedona squint,abriefgrimacewhilereadingsomething,ayawn. Noneofthosemeantanythingaboutactualmoodbutthe system would respond to each one. Averaging the

classifier output across a rolling five to eight second window fixed that. Only expressions that persist long enoughtooutlastthewindowaffecttheoutput.Thereis a delay baked into that design and we address it in the limitations section, but it was the only realistic way to getstablebehavior.

3.2 From a Feeling to a Song

Getting from a detected emotion to an actual song choice is the harder half of this problem. Knowing that someone appears exhausted or tense or quietly content is one thing. Translating that into a specific track recommendation that will actually land for them is somethingelse.

The first matching layer uses audio features. Both Librosa and the Spotify Audio Features API can return numericaldescriptorsforanytrack.Wesettargetranges on those descriptors for each detected emotion. Exhaustion mapsto trackswitha tempounder 80 beats perminute,lowenergyscores,andasoftoveralltexture. Happinessmapstotracksabove120BPMinmajorkeys. One practical benefit of this approach is that it can recommend music the user has no prior relationship with. Since the filter operates on the properties of the musicratherthanthehistoryoftheperson,itopensthe doortodiscovery.Theemotionalstatebecomesasearch parameterthatcanreturngenuinelynewthings.

Collaborativefilteringrunsinasecondlayerontop of that. We pull data from users whose emotional listening patterns resemble the current user and look at what they chose during comparable emotional states. Userswhoconsistentlyreachforacertainkindofmusic when they are anxious form a useful reference group eveniftheirgeneraltasterunsdifferently.Theoverlapin emotionally specific choices matters more here than overall taste similarity. A final personalization step removes anything the current user has already heard recently or deliberately skipped under similar conditions.

3.3 Learning the Person Over Time

Starting from population averages is unavoidable early on but it is not where you want to stay. The way individual peopleuse musicvariesa lotmore thanmost recommendation systems assume and the system needs to accumulate enough data to reflect a specific person ratherthanastatisticalcenterofmass.

We designed two parallel feedback streams. One is deliberate: skipping a track within the first half minute, savingit,ortappingthecorrectionbuttoniftheemotion readwasclearlyoff.Thesesignalsareunambiguousand immediately useful. The problem is frequency. Most people engage with music passively. They are not

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

thinking about giving the system information; they are just listening. The second stream gathers behavior that happens without any intentional signaling. A song that runs to its end, that comes back in another session, a queue left running untouched for a long stretch. These quiet signals accumulate over time and eventually tell thesystem a lotabout what thisspecific person reaches forindifferentemotionalstates.Wekeepusingthisdata again and again to improve and personalize the recommendationsforeachuser.

4. SYSYTEM ARCHITECTURE

The four stages describedabove map ontofive software components when actually implemented. Table I breaks eachonedown.

TableI:Five-StageMBMRSArchitecturePipeline

Stage Function Technologies

1.Frame Capture Cameratakes framesatintervals OpenCV,V4L2

2.Facical Identification Findsandisolates face MTCNN,SSD

3.Emotion Class. Scoresfaceacross 7categories MobileNet, TFLite

4.Music Select. Rankssongsby emotionalmatch Collaborative+ ContentBasedFilter

5. Playback/Log Playstrack, recordsbehavior SpotifySDK, Flask

The backend is Python, using Flask or FastAPI. OpenCV manages frame acquisition. TensorFlow with Keras handles the classifier. On the browser side, the MediaDevices API provides the camera access through the same permission dialog that appears during the process of initiating a video communication session Once granted, the pipeline runs without any further interaction. Audio output goes through the Spotify Web SDK or the Web Audio API based on the deployment context.

Puttingthisonaphonerequiredmodelconversion. Android uses TensorFlow Lite, iOS uses Core ML [19]. After conversion, a single MobileNet inference on a typical mid-range Android device takes under 50 milliseconds. That is fast enough that no one notices it running.Thesystemsitssilentlyinthebackgroundwhile themusicplaysanddoesnothingvisible.

We debated running the emotion processing remotelyonaserverratherthanonthedevice.Thecase for it is straightforward: better hardware means bigger modelsandpotentiallybetteraccuracy.Thecaseagainst itstoppedthatconversationquickly.Runningitremotely means the device is continuously sending images of the user face, taken at home in private moments, to infrastructure the user knows nothing about. We could notconstructaversionofthatwhichwefeltcomfortable asking people to agree to. On-device keeps every frame on hardware the user controls. It gets processed and immediately discarded. Nothing leaves. That was the onlydesignwetrulytrustandfeltconfidentabout.

5. PERFORMANCE AND RESULT

5.1 Emotion Recognition Accuracy

FER2013 benchmark results for a baseline CNN sit inthe72to78percentaccuracyrange.Addingattention layers, training on augmented versions of the data, or running model ensembles pushes that up toward 83 to 86 percent [20]. AffectNet scores look lower in comparison but the dataset itself is harder. The images aremorenaturalisticandnoisier.A75percentresulton AffectNet reflects a model that will hold up better in actual use than a high score on FER2013, which uses cleaner,morecooperativeimages.

Raw accuracy also obscures how differently the model handles different emotions. Happy is consistently the easiest label for any classifier. Recall rates above 90 percentarecommonacrossthepublishedliterature.The visual signature is strong and it appears frequently in training data. Sad and angry are more challenging but workable, landing somewhere between 65 and 80 percent. Fear and disgust are where reliability breaks down, frequently below 55 percent. Those two states shareasignificantportionofthesamemuscleactivityin the face. Experienced human annotators studying the same images regularly produce conflicting labels, so the modeldifficultyhereisnotasignofarchitecturalfailure.

Neutral created its own separate issue. A face in a genuinely neutral, at-rest state tends to get scored toward mild sadness by most models. In practical recommendation terms, this shows up as a user sitting peacefully at their desk and receiving a queue of slow and heavy music they never wanted. It doesn’t clearly show that something is wrong. Instead, it quietly keeps giving slightly wrong recommendations, which slowly makesuserslosetrustinwhattheyhear.

5.2 Recommendation Quality

Published studies comparing mood-aware systems against history-only approaches consistently find the

International

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

largest performance gap during sessions when the user emotional state diverges from their typical one [21]. That is the precise scenario we built for. When nothing unusual is happening emotionally, both approaches perform similarly. When the user is having an atypical day, the mood-aware system clearly pulls ahead. Running collaborative and content-based filtering in combination produced higher satisfaction scores than either alone across every study we examined. Adding contextual signals like time of day and recent physical activity on top of mood improved appropriateness ratingsbyanother15to20percent.Eachofthesesignals may seem small on their own, but together they help create a clearer and more complete understanding of what’shappeningrightnow.

6. CHALLENGES AND LIMITATIONS

6.1 Individual Variation in How People Use Music

Improving detection accuracy does not touch it. The mapping between emotional states and music categories was derived from aggregate behavior, from what most people in a given state tend to choose. That averagedoesnotdescribeeveryoneandforsomepeople it describes no one they know. Some people specifically want something intense and loud when they are angry because it externalizes the feeling rather than internalizing it. Others cannot tolerate that and need something much quieter. Some people like listening to sad music when they feel low, because it makes them feelunderstoodandlessalone.Butothersdon’tlikethat, they prefer music that cheers them up and helps them move on from the feeling. Both approaches are legitimate.Theysimplybelongtodifferentindividuals.

Asingleuniversal mappingappliedtoall users will produce consistent systematic errors for a meaningful portionofthem.Notoccasionalmissesbutthesamekind of wrong answer, repeatedly, in the same direction. The personalization layer is the right structural response to this and it does get better as it accumulates data. But accumulation takes time. During the early period of use the system is drawing on the population average and that average may fit the individual poorly. People will experience thisasthesystemnotgetting them, whichis accurate, and some of them will stop using it before enoughdataaccumulatestoimprove.

6.2 Privacy

We want to be direct here. Activating the front camera throughout a music listening session is a substantial thing to ask. People listen to music in their mostprivatesettings.Lateatnight,whentheyareupset, in physical and emotional states they would not be in

publicly. The fact that all processing happens on-device and no images leave the phone is genuinely important and we designed the system that way deliberately. But this still doesn’t fullyanswerwhether usersareactually comfortablewiththecamerabeingon

Justshowingapermissionscreendoesn’tmeanreal transparency. True transparency means the user clearly understands what the camera is capturing, how quickly theimages are deleted, whatthesystemcan and cannot understand from those images, and how easily they can turn this feature off. When a system uses something as personalasalivecamerafeed,trust can’tbeassumed,it has to be built over time through honest and clear behaviour. It cannot be established in a single permissiondialog.

6.3 Smoothing Lag

The rolling window average that stabilizes predictions creates a fixed delay. If the user emotional statechangesquicklythesystemwillkeeprespondingto the old state for several seconds before catching up. Narrowing the window makes the response faster but also makes single stray expressions capable of disrupting the queue. Widening it produces smoother behavior but the lag becomes perceptible at the moments when responsiveness matters most. We did not land on a configuration that resolves both concerns and we do not believe one exists in the current design. This is a structural tension in the smoothing approach rather than a parameter that can be tuned to a clean solution.

7 FUTURE DIRECTIONS

7.1 Using More Than Just the Camera

The face is where we started but it captures only partofthepicture.Noteveryoneexpresseshowtheyfeel facially. Some people have a naturally flat expression regardless of what is going on internally and a camera pointed at them conveys almost no emotional information. Voice carries a separate set of signals: the speed of speech, the energy in it, flatness versus expressiveness, none of which requires the face at all. Wearables measure physiological signals like heart rate and galvanic skin response that respond to emotional state in ways people cannot voluntarily suppress the way they can control a facial expression. Each of these input channels breaks down under different circumstances. The useful property of using several at once is that they are unlikely to all fail simultaneously, which means the combined estimate is more stable and morereliablethananysinglechannelwouldbe[23].

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net

7.2 Understanding the Situation Better

How someone feels is one factor in determining what music would work for them right now. Simple thingsmatter,like ifapersonissittingstillormoving,if the place is quiet or noisy, and if they are alone or with others. All these help understand the situation better. Mostofthesesignalsareaccessiblethroughsensorsand apps that already exist on the device: a fitness tracker, thedeviceclock,themicrophonereadingambientlevels, the calendar [16]. Using them does not require any new hardware.Itrequiresdecidingtolookatwhatisalready there and building the logic to interpret it. The larger question is what information the user can be honestly toldisbeingusedandwheretodrawlinesoncollection.

7.3 Learning the Individual More Deeply

The current implementation uses a fixed base model with a personalization layer that adjusts the recommendations based on accumulated individual behavior. That is a reasonable approach up to a point. People change though, in ways that go deeper than listening preferences. Someone who used this system during a difficult stretch in their life and then moved throughittoabetterplacehasadifferentrelationshipto music now than the data from that earlier period reflects.Thepersonalizationlayerlearnsthatthisperson historically skips a certain kind of track but it does not understand that the reason has changed. A more adaptivearchitecturewouldallowtheunderlyingmodel toupdateattheindividuallevelovertime,trackingwho a person actually is at present rather than averaging theirentirehistorywiththeapplication.

7.4 Music as Emotional Support

Every design decision in this project was oriented toward matchingmusictohowsomeonecurrentlyfeels. That is useful. But music is frequently used for somethingdifferentfrommatching.Peopleputmusicon to shift how they feel, to work through something, to move from one emotional state toward another one, to feellessaloneinadifficultexperience.Researchinmusic psychology covers this terrain in depth. A system designedaroundthatgoalwouldbehavedifferentlyfrom ours. Instead of asking what fits the current emotional reading, it would ask what the person likely needs in order to feel better and pick music that helps get there. Building that properly requires thinking about the clinicalandpsychologicaldimensionsatalevelthatgoes beyondwhatanengineeringteamcandoindependently. But it is probably the most genuinely useful version of whatthissystemcouldbecome.

8. CONCLUSION

MBMRSisourattemptatthesecondside.Thefront camerareadsfacialexpressioninrealtime.Theemotion classifier maps that to one of seven states. The hybrid filteringlayerusesbothcollaborativeandcontent-based methods to generate a queue that reflects the current emotional reading. Testing showed the system performed most clearly better than history-only approachesinthespecificsituationsitwasdesignedfor, when users were in an emotional state that departed from their baseline and a habit-based system had nothingtooffer.

The limitations are genuine. Benchmark accuracy does not survive contact with actual home lighting. Training datasets under-represent large portions of the world's users. The early experience before personalization has kicked in can feel frustratingly generic. And pointing a camera at someone face throughout a solo listening session is a privacy ask that deserves far more careful handling than a standard consentflowprovides.

9. REFERENCES

[1] I. J. Goodfellow et al., "Challenges in representation learning: A report on three machine learning contests," Neural Networks, vol. 64, pp. 59-63, Apr. 2015.

[2] A. G. Howard et al., "MobileNets: Efficient convolutional neural networks for mobile vision applications,"arXiv:1704.04861,Apr.2017.

[3] I. J. Goodfellow et al., "Challenges in representation learning: A report on three machine learning contests,"NeuralNetworks,vol.64,pp.59-63,2015.

[4] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proc. IEEE CVPR, Jun.2016,pp.770-778.

[5] A. Mollahosseini, D. Chan, and M. H. Mahoor, "Going deeper in facial expression recognition using deep neural networks," in Proc. IEEE WACV, Mar. 2016, pp.1-10.

[6] A. Mollahosseini, B. Hasani, and M. H. Mahoor, "AffectNet: A database for facial expression, valence and arousal computing in the wild," IEEE Trans. AffectiveComput.,vol.10,no.1,pp.18-31,2019.

[7] J. A. Russell, "A circumplex model of affect," J. Personality Social Psychol., vol. 39, no. 6, pp. 11611178,1980.

International Research Journal of Engineering and Technology (IRJET) e-

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

[8] Y. Li, J. Zeng, S. Shan, and X. Chen, "Occlusion aware facial expression recognition using CNN with attention mechanism," IEEE Trans. Image Process., vol.28,no.5,pp.2439-2450,May2019.

[9] S. Kumar and A. Singh, "Explainable emotion-driven music recommendation using ResNet50 and GRADCAM,"arXiv:2404.04654,Apr.2024.

[10] R. Patel et al., "Real-time mood-based music recommendation using DeepFace and webcam input,"arXiv:2503.20739,Mar.2025.

[11] B. McFee et al., "librosa: Audio and music signal analysisinPython,"in Proc. SciPy, Jul.2015, pp. 1825.

[12] Y. Zhang, Y. Yang, and R. Gu, "A unified multitask framework for music emotion recognition," arXiv:2406.08809,Jun.2024.

[13] H. Wang, L. Zhang, and X. Xu, "Fully convolutional recurrent attention networks for emotion-aware music recommendation," Egyptian Informatics J., 2025,doi:10.1016/j.eij.2025.100502.

[14] Y. Koren, R. Bell, and C. Volinsky, "Matrix factorizationtechniquesfor recommendersystems," Computer,vol.42,no.8,pp.30-37,Aug.2009.

[15] M. Schedl et al., "Current challenges and visions in music recommender systems research," Int. J. Multimedia Inf. Retrieval, vol. 7, pp. 95-116, Jun. 2018.

[16] G. Adomavicius et al., "Incorporating contextual information in recommender systems using a multidimensional approach," ACM Trans. Inf. Syst., vol.23,no.1,pp.103-145,Jan.2015.

[17]X.He,L.Liao,H.Zhang,L.Nie,X.Hu,andT.-S.Chua, "Neural collaborative filtering," in Proc. WWW, Apr. 2017,pp.173-182.

[18] K. Zhang, Z. Zhang, Z. Li, and Y. Qiao, "Joint face detection and alignment using multitask cascaded convolutional networks," IEEE Signal Process. Lett., vol.23,no.10,pp.1499-1503,Oct.2016

[19] A. G. Howard et al., "MobileNetV2: Inverted residuals and linear bottlenecks," in Proc. IEEE CVPR,Jun.2018,pp.4510-4520.

[20] J. Li, Y. Jin, and D. Zhou, "Occlusion-aware facial expressionrecognition:Asurveyandanalysis,"IEEE Trans. Affective Comput., vol. 13, no. 3, pp. 15121530,2022.

[16]Y.Deldjooetal.,"Recommendersystemsleveraging multimediacontent,"ACMComput.Surv.,vol.53,no. 5,Art.106,Sep.2021.

[17] M. S. Bartlett et al., "Automatic decoding of facial movements reveals deceptive pain expressions," Curr.Biol.,vol.24,no.7,pp.738-743,2014.

[23] S. Poria, E. Cambria, R. Bajpai, and A. Hussain, "A review of affective computing: From unimodal analysis to multimodal fusion," Inf. Fusion, vol. 37, pp.98-125,Sep.2017.

[24] F. N. Iandola et al., "SqueezeNet: AlexNet-level accuracy with 50x fewer parameters," arXiv:1602.07360,Feb.2016.

[25]X.Zhaoetal.,"Deepreinforcementlearningforlistwise recommendations," arXiv:1801.00209, Jan. 2018.