Tags for candidates

[jace]

To recommend jobs to candidates, we need to link candidates to topic interest tags, based on what jobs they are interested in and what they write in the content of a job application.

1. Interests fade with time, so timestamps and a decay algorithm are required.
2. Interests aren't binary. A candidate may have anything from a strong interest (they mentioned it in the application) to a weak, potentially false positive signal (they clicked a link to something and backtracked immediately).

Available signals (in decreasing order of signal quality):

1. Candidate mentioned a keyword in an application
2. Candidate searched for a keyword
3. Candidate applied to a job that mentions a keyword
4. Candidate opened application form for a job that mentions a keyword
5. Candidate viewed a job that mentions a keyword

We currently track keywords/interests via the Tag model, which is generated by NLP analysis of text looking for named entities. It is imperfect but good enough for a start. An ideal tracker would record every instance of an expression of interest, but this may be impractical to store and process.

1. A continued interest over time, even if weak, is an important signal.
2. A burst of interest in any keyword should be measured relative to session activity. If a particular keyword dominates activity in a brief browsing session, that may be more important than another keyword receiving more absolute attention as a relatively smaller part of a larger browsing session.
3. The relative weights of these signals may need to change over time as we understand their relevance better.