I'm trying to find a good starting place (or existing algorithm) to determine the similarity of various items to one another based on subjective assessments of two items' relative similarity to a third.
These assessments are stored in a series of uniquely-identified transactions of the format:
Reference Item, More Similar Item, Less Similar Item
These transactions are generated when a human "teacher" is presented with one random "reference" item and two other random items and then asked to select which of the latter two items best matches the reference item.
The set becomes a series of value judgments in the form: $A$ is more similar to $R$ than $B$ is to $R$. I reckon this approach is more implicit than querying the teacher for a number to represent the distance between items, but I've run into a roadblock when it comes to converting these relative judgments to absolute numbers to represent the distance between items.
What existing machine learning approaches best addresses this problem? Are there modifications that could be made to the data collection methodology which would improve the efficiency/accuracy of the algorithm?