A Personalized Music Recommender Service Based on Fuzzy Inference System Md. Saidur Rahman, Md. Saifur Rahman, Shahnewaz Ul Islam Chowdhury, Ashfaq Mahmood, Rashedur M. Rahman Department of Electrical and Computer Engineering, North South University Plot -15, Block -B, Bashundhara, Dhaka 1229, Bangladesh saidurrahman@northsouth.edu, breekal7@gmail.com, muradchowdhury0215@live.com, ashfaq.bd92@gmail.com, rashedur.rahman@northsouth.edu Abstract—In this paper, we are proposing a personalized music recommender service based on Mamdani Fuzzy Interference System (M-FIS). Collection of playlist is used for gathering users’ choice and mood while listening to songs. Similarity between audio files is calculated based on Mel Frequency Cepstral Coefficients (MFCC). We have developed a recommender model based on M- FIS with the aforementioned similarities and playlists. We were able to gain an acceptable accuracy rate using FIS compared to other method reported in literature. Keywords-component; Music Recommender Service; Fuzzy Inference System(FIS); Mel Frequency Cepstral Coefficients (MFCC). I. INTRODUCTION In this modern era people have been much aware of what they listen to. And they also tend to listen to music to be similar to what they have listened to previously. This makes people look for new music every day and thus recommendation systems come in handy. Recommendation systems generally try to find a pattern in people’s listening habits. We are proposing a system that combines "metadata pattern matching" and "analysis of the music tracks" to reach a more accurate point in recommending a song that a person may want to listen to. Similarities among the songs are calculated based on Mel Frequency Cepstral Coefficients (MFCCs) value of the songs and the songs appearances' on the playlists in the sample dataset. The values for the later one is calculated using weighted average value for each song to another song based on the number of appearances and the order they appeared on the playlists. Using the similarity values of MFCC and "appearance on playlists", we have built a Mamdani Fuzzy Inference System which gives us a weighted matrix which represents numerical distances among songs. The lower the distance between two songs is, the more similar they are. For recommending music, we use the sorted output matrix of that FIS for any given song. Section-II briefly describes related works. Section-III presents system design. Distance calculation is depicted in Section-IV. Section-V implements M-FIS with different membership functions. Section-VI shows result analysis. Finally, Section-VII concludes the paper. II. RELATED WORKS Kim et al [1] built recommendation model based on weighted clusters and K-Means algorithm on the values of calculated Hidden Markov Models (HMM) and MFCC considering every sample songs. Aucouturier and Patchet [2] measured similarity based on timbre and used Gaussian Mixed Model (GMM) on those for recommending music. Siddiquee et al [3] recommended songs using rules generated by Association Rule Mining (ARM) from users' playlists and similarity scores based on MFCC of the songs on the sample dataset. Sen and Martha [4] recommended songs based on users' mood detected through a number of Fuzzy Logic models with users' contextual sensor (GPS, Compass etc.) information fused with information from the web. Hoffman et al.[5] on their paper experimented with feature vectors on low- level signal descriptors of music files and used correlation analysis and Principal Component Analysis to optimize them for the music recommendation system. III. SYSTEM DESIGN We propose a music recommender system using a Fuzzy Inference System (FIS) based on similarities amongst the audio files from a defined set. Two different approaches are being used for finding mathematical distances amongst the audio files. In one of our approach, we will use MFCC values for comparing audio tracks with one another. A square matrix with respect to number of unique audio tracks will be generated as distances amongst the tracks using that method. We have developed a simple algorithm for creating a distance matrix based on the weight - each of them carries information due to their appearances in the playlist. This algorithm is well described in Section-IV(B) – “Playlist Based Distance Calculation”. This approach will give us another square matrix as distance matrix. The resultant distances of those approaches will be used as inputs for the Mamdani type Fuzzy Inference System (M-FIS). For any given music file, our recommender system will use the output from the M-FIS, sorted in ascending order, for suggesting music(s) to the user. Figure-1 shows the architecture of the system. In the following sections, distance calculation approaches, M- FIS implementations and music suggestion technique will be described with examples. 978-1-5090-0806-3/16/$31.00 copyright 2016 IEEE ICIS 2016, June 26-29, 2016, Okayama, Japan