# Glossary¶

Song

In libmunin’s Context a Song is a set of attributes that have a name and a value. For example a Song might have an artist attribute with the value Amon Amarth.

Apart from the Attributes, every Song has a unique ID.

Distance

A distance is the similarity of two songs or attributes a and b expressed in a number between 0.0 and 1.0, where 1.0 means maximal unsimilarity. Imagining a point space, two points are identical when their geometric distance is 0.0.

The Distance is calculated by the DistanceFunction.

DistanceFunction

A DF is a function that takes two songs and calculates the Distance between them.

More specifically, the DF looks at all Common Attributes of two songs a and b and calls a special DF attribute-wise. These results are weighted, so that e.g. genre gets a higher precedence, and summed up to one number.

The following must be true for a valid DF, when $$D$$ is the database:

$$D(i, j) = D(j, i) \forall i,j \in D$$

$$D(i, i) = 0.0 \forall i \in D$$

$$D(i, j) \leq D(i, x) + (x, j)$$

Session
A Session is the usage of libmunin over the time on one music collection. It can be saved to disk and later resumed.
Every Session requires a Mapping where the possible keys are defined that a single song may have. The AM stores this information as dictionary, the keys being the names of the possible attributes and the values being a tuple, conisting of the Provider for this Attribute, a fitting Distance Function and a weight.
Attribute
An Attribute is a key in the Mask to which a Provider, a DistanceFunction and a certain weight is attached. The name of the Attribute will be used in Song s to reference the individual values.
Provider
Normalizes a Value by certain characteristics. The resulting value should be optimized for comparasion by areturn DistanceFunction.
Rule

A Rule associates certain songs, or one single song with other songs or another single song. The strenght of the association is given by the rating of the rule, which is technically calculated as:

$$(1.0 - Kulczynski) \cdot Imbalance$$
Recommendation
A Recommendation is a Song that is outputted as a request by the user. The outputed Song s should have a low Distance to the previously listened or to the seed song.
Graph
The Graph models the relation between all Song in the database. Each song has, in the best case, the most similar Song s to it as neighbors. Since this would require calculating the Distance from one song to all others, which in turn requires quadratic complexity. Therefore an approximation of the Graph is built that might contain small errors.