Academia.eduAcademia.edu

Figure 4 – uploaded by Dmitry Bogdanov

See full PDFdownloadDownload figure
Fig. 4. Graphical representation of the proposed recommendation approaches in the semantic space (here reduced to two dimensions and the case of one recommended track for illustration purpose). Solid lines outline recommendation outcomes (items marked by stars) and the respective recommendation sources in the case of the distance-based approaches. The dashed lines indicate regions of equal probability of the respective components of the GMM in the case of the probabilistic approach. In SEM-MEAN, the mean vector of the preference set is used to retrieve the closest track from the music collection using track-to-mean semantic distance. In SEM-ALL, all tracks from the preferences set are considered to retrieve the track closest to the preference set using track-to-set semantic distance. In SEM-GMM, a probabilistic model is built from the preference set. The track from the music collection yielding the highest probability value is returned as recommendation. See text for details.

Figure 4 Graphical representation of the proposed recommendation approaches in the semantic space (here reduced to two dimensions and the case of one recommended track for illustration purpose). Solid lines outline recommendation outcomes (items marked by stars) and the respective recommendation sources in the case of the distance-based approaches. The dashed lines indicate regions of equal probability of the respective components of the GMM in the case of the probabilistic approach. In SEM-MEAN, the mean vector of the preference set is used to retrieve the closest track from the music collection using track-to-mean semantic distance. In SEM-ALL, all tracks from the preferences set are considered to retrieve the track closest to the preference set using track-to-set semantic distance. In SEM-GMM, a probabilistic model is built from the preference set. The track from the music collection yielding the highest probability value is returned as recommendation. See text for details.

Related Figures (13)

Fig. 1. General scheme of the proposed preference elicitation (user modeling) technique and its applications.
Fig. 1. General scheme of the proposed preference elicitation (user modeling) technique and its applications.
Ground truth music collections employed for semantic regression. Source references: (1) Homburg et al. (2005), (2) in-house, (3) Tzanetakis and Cook (2002), (4) Gomez and Herrera (2008), (5) Laurier et al. (2009) + in-house, and (6) Cano et al. (2006). Table 1 For each music track, we calculate a low-level feature representation using an in-house audio analysis tool.® In total, thi tool provides over 60 commonly used low-level audio features, characterizing global properties of the given tracks, related t timbral, temporal, and tonal information. The features include inharmonicity, odd-to-even harmonic energy ratio, tristimuli spectral centroid, spread, skewness, kurtosis, decrease, flatness, crest, and roll-off factors, MFCCs, spectral energy bands zero-crossing rate (Peeters, 2004), spectral complexity (Streich, 2007), transposed and untransposed harmonic pitch class pro files, key strength, tuning, chords (Gémez, 2006), pitch, beats per minute (BPM) and onsets (Brossier, 2007). Most of these fea tures are extracted on a frame-by-frame basis and then summarized by their means and variances across all frames. In the cas¢ of multidimensional features (e.g., MFCCs), covariances between components are also considered.
Ground truth music collections employed for semantic regression. Source references: (1) Homburg et al. (2005), (2) in-house, (3) Tzanetakis and Cook (2002), (4) Gomez and Herrera (2008), (5) Laurier et al. (2009) + in-house, and (6) Cano et al. (2006). Table 1 For each music track, we calculate a low-level feature representation using an in-house audio analysis tool.® In total, thi tool provides over 60 commonly used low-level audio features, characterizing global properties of the given tracks, related t timbral, temporal, and tonal information. The features include inharmonicity, odd-to-even harmonic energy ratio, tristimuli spectral centroid, spread, skewness, kurtosis, decrease, flatness, crest, and roll-off factors, MFCCs, spectral energy bands zero-crossing rate (Peeters, 2004), spectral complexity (Streich, 2007), transposed and untransposed harmonic pitch class pro files, key strength, tuning, chords (Gémez, 2006), pitch, beats per minute (BPM) and onsets (Brossier, 2007). Most of these fea tures are extracted on a frame-by-frame basis and then summarized by their means and variances across all frames. In the cas¢ of multidimensional features (e.g., MFCCs), covariances between components are also considered.
Fig. 2. Box plot of the proportions of provided labels per musical facet, normalized by the total number of labels per participant. Categories from left to right correspond to genre, moods, instruments, rhythm, melody and harmony, musical expression, and other labels respectively. Red crosses stand for extreme outliers. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) The first considered application exploits the computed user model to generate music recommendations based on seman- tic descriptors. For consistency, we focus on the task of retrieving 20 music tracks from a given music collection as recom-
Fig. 2. Box plot of the proportions of provided labels per musical facet, normalized by the total number of labels per participant. Categories from left to right correspond to genre, moods, instruments, rhythm, melody and harmony, musical expression, and other labels respectively. Red crosses stand for extreme outliers. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) The first considered application exploits the computed user model to generate music recommendations based on seman- tic descriptors. For consistency, we focus on the task of retrieving 20 music tracks from a given music collection as recom-
Fig. 3. Tag cloud representing overall music preferences of our participants, based on artist tags found on Last.fm. 5.1. Recommendation approaches
Fig. 3. Tag cloud representing overall music preferences of our participants, based on artist tags found on Last.fm. 5.1. Recommendation approaches
underlying music collection. Table 3
underlying music collection. Table 3
Fig. 5. The percent of fail, trust, hits, and clear categories per recommendation approach (a); the liking, listening intentions, and “give-me-more” meat ratings for recommendation approach (b). The results for the LASTFM approach were obtained on a different underlying music collection. The “give-me. more” rating varies in the [0,1] interval.
Fig. 5. The percent of fail, trust, hits, and clear categories per recommendation approach (a); the liking, listening intentions, and “give-me-more” meat ratings for recommendation approach (b). The results for the LASTFM approach were obtained on a different underlying music collection. The “give-me. more” rating varies in the [0,1] interval.
Selected descriptors, and the corresponding music collections used for regression, per category of semantic descriptors (i.e., genre, moods & instruments, and others) used for visualization. Table 4
Selected descriptors, and the corresponding music collections used for regression, per category of semantic descriptors (i.e., genre, moods & instruments, and others) used for visualization. Table 4
Mapping of the descriptor categories to the graphic groups.
Mapping of the descriptor categories to the graphic groups.
Vector representation example: user profile vs. the instrument graphic group (continuous summarization). A visual element with the minimum distance to the user profile is selected (in this case, the turntable).
Vector representation example: user profile vs. the instrument graphic group (continuous summarization). A visual element with the minimum distance to the user profile is selected (in this case, the turntable).
Fig. 6. Example of the visualization approach. It can be seen how the descriptor values influence the selection of the different graphic elements used to construct the avatar. The values inside the graphic element boxes represent all possible descriptor values that can generate the presented element.
Fig. 6. Example of the visualization approach. It can be seen how the descriptor values influence the selection of the different graphic elements used to construct the avatar. The values inside the graphic element boxes represent all possible descriptor values that can generate the presented element.
Fig. 7. Sample Musical Avatars generated by the three summarization methods (i.e., from left to right, binary, ternary, and continuous) for the same underlying user model. Notice the differences in guitar and headphones sizes among the generated avatars.
Fig. 7. Sample Musical Avatars generated by the three summarization methods (i.e., from left to right, binary, ternary, and continuous) for the same underlying user model. Notice the differences in guitar and headphones sizes among the generated avatars.
Mean ranks and standard deviations for the different visualization methods obtained in the user evaluation. The random column corresponds to the average values of the individual random results (see text for details). from her/his own preference set, two randomly generated avatars, and one neutral avatar. We asked the participants to rank these images assigning the image that best express their musical preferences to the first position in the rank (i.e., rank = 1). Finally, we asked for a written feedback regarding the images, the evaluation procedure, or any other comments.?° Table 7
Mean ranks and standard deviations for the different visualization methods obtained in the user evaluation. The random column corresponds to the average values of the individual random results (see text for details). from her/his own preference set, two randomly generated avatars, and one neutral avatar. We asked the participants to rank these images assigning the image that best express their musical preferences to the first position in the rank (i.e., rank = 1). Finally, we asked for a written feedback regarding the images, the evaluation procedure, or any other comments.?° Table 7
Related topics:
Information SystemsUser ModelingLibrary and Information Studies
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved