Objective
For this demo, a SOM was generated using responses to a questionnaire about social demographic variables and hobbies, including music preferences. Respondents were ordered in the SOM according to their responses about their attributes and hobbies, and the completed map was used to display their music interests over the generated clusters.
The attributes and hobbies from a new person can now be used to determine his or her position in the SOM. The music preferences of the respondents at any position are likely to be similar, allowing the music preferences of the new person to be read off from his or her position in the SOM.
The map below shows the Viscovery model (SOM) that was created from a subset of questionnaire data collected by a dating service. From this map, the relationship between social demographic variables and hobbies and 8 predefined music preferences can be determined.
Advantages and benefits
The evaluation of personal interests, preferences and tastes of individuals or groups is particularly difficult when limited data are available. The ability to rate the interesting subjective characteristics from partial data is extremely valuable. Many diverse applications are available for such evaluation, including target-group specific design of products, individualized approaches to address customers, product promotion and placement, as well as analysis of social networks, psychological research and dating services.
Test your music style
Check out where your personal profile is located in the map of all respondents and what kinds of music the people with a similar profile are typically interested in. The table below shows the prediction of your affinities to 8 styles of music (expressed as a score) as well as a comparison of your affinities with the average from all respondents. The right bar chart graphically shows the deviations of your affinities from the average for each style.
marks your position
Please note that this demo can provide only a coarse assignment of your affinities to the predefined music styles. By using more variables (in this case, responses from the questionnaire) to generate the model, the SOM will be more selective, allowing more sophisticated prediction.
Data source and acknowledgement
The online PARSHIP dating service provided approximately 120,000 anonymous records which were used for the initial data set. The records primarily contained 24 attributes for describing the hobbies and interests of European members over 18 years as well as their social demographic variables.
PARSHIP regularly validates their matching algorithms to provide their members with the best possible partner suggestions. You can find the PARSHIP dating service here.
Data preprocessing
Variables were standardized and scaled according to defaults by Viscovery. Correlation compensation was automatically applied for the variables. The Age attribute was transformed sigmoidally.
The values of the Education attribute were grouped into an ordinal variable with 3 levels. Active and passive hobbies were combined independently for each variable.
The presented model was calculated only from complete records that contained no missing values.
SOM creation
The resulting self-organizing map with 500 microclusters (“nodes”) was created form the final data set with Viscovery SOMine using standard settings, “Normal” training schedule, and a “Tension” of 0.6.
To generate the displayed SOM from an initially larger number of attributes, only those attributes contributing to the ordering of music preferences were chosen and prioritized iteratively until the music preferences – which were not prioritized – were differentiable. The final priorities for the attributes were as follows: 1.0 for Age, 0.9 for Education, 0.6 for Gender, 0.7 for Theater, Literature, and Art, 0.6 for Film/Video and for Dance.
For the online classification, respondent records were grouped into approximately 400 clusters to ensure sufficient granularity to differentiate the entire set.