MDS For Feature/Properties & Points

[josh]

For example –

Start with huge unwieldy data set with many features/properties associated with some combination of verbal, textural, measuremen, and ML/PR forms and cases, probably with a lot of missing data & lack of common experimental frame.

Create linkage measures for the feature/properties based on things like some measure of similarity or mutual information in different frames of reference with some kind of hacked importance sampling.

Alternate techniques like: Minimum Spanning Tree, Spectral Clustering, & MDS to create vector spaces of manageable size/complexity for the application

Now do multi-dimensional scaling of the point cases in this new product vector space. Being able to visualize results of semantic claims is an important guide to getting it right, along with making correct predictions or “Turing/behavioral appropriateness” tests.

[josh]

One of the goals is to capture “context of interpretation” in mechanizable/quantifiable ways. The verbal & physical & information context of a given speech act should map to weighted subspaces of the larger property space.

[josh]

Idea for a potentially cool application of MDS to music marketing:

There are many potential dimensions of music choice that can be recognized as significant:
Old vs. New
Punchy Rhythm vs. languid (and or danceable vs. not)
Fast vs. Slow Tempo
Electronic vs. Acoustic
Popular vs. Esoteric
Familiar to me vs. Unfamiliar to me
Large Volume Dynamic range vs. not
Well-tempered vs. Microtonal
Very Harmonious vs. Dissonant
Lots of Words vs Instrumental
Light vs. Dark tonalities
etc.

An application could help users to create their own MDS for music they liked with various examples, while also referencing the dimensions that others notice & care about. The task of building it could be visually interesting & engaging for a music fan, while the end goal is to help them find new music that fits the context of their search for a given listening application.

Their are many ways to view more than 2/3 dimensions at a time, including Trellis panels. It could become something that users share with others. In my experience, the ability to share music interests is a key driver of commerce.

[josh]

Statistical models can be used to evaluate the similarity of a user’s likes & chance of familiarity other users, conditional on which part of the music space is in focus (multivariate non-linear modeling with sparse data).

What are constructive roles of the expert or critic?

Spaces of individuals, the user group, & the broader market will change over time. New materials coming into focus is just one part of that. There’s a chance for dynamics of influence & focus which are a lot different than Billboard charts for airplay charts of yesteryear. Modeling how vertbal descriptions/participation influences the dynamics in the user community & satisfaction with the commercial experience can help find a good synergy for custom framing of light recs/reviews. The tastemakers who agree to give something new a spin & then rate it can be part of a kind of social critic army. The data will be better & more predictive if reviews secretly include honest info about how many times something was listened to & over what stretch of time. That should be part of the modeling.

[Author]

Posts