share_type
How common the candidate's own type is among all candidates of the stay. If a stay has 20 candidates and 12 of them are street pieces, the street share is 12/20 = 0.6. With \(n_{t,i}\) candidates of type \(t\) out of \(N_i\) candidates for stay \(i\):
\[\text{share}_t(C_i) = \frac{n_{t,i}}{N_i}\]
Each type gets its own weight (share_bldg, share_street, share_os). Each column holds the share for rows of that type and 0 for the other rows.
In the models: adding this term fixed the biggest problem of the first model (Model 2: share-type and sigma). But its weights for street and open space are large and negative. So a candidate with many neighbours of the same type gets a lower value. That is the cause of Candidate density bias toward building.
Part of Assignment framework.