Representative Comments

• For a given comment $c$, and subset (group) in the conversation $g$, the Representativeness of that comment $c$ for group $g$ (denoted $R_v(g,c)$), is a measure of how much more likely participants in group $g$ are to place vote $v$ on said comment than those outside group $g$.

• Here's how we compute Representativeness :

  • We first estimate the probably that a person in a group $g$ votes $v$ on comment $c$ as follows: $P_v(g,c)=\frac{1+N_v(g,c)}{2+N(g,c)}$. Here:

    • $N_a(g,c)$ is the number of people in group $g$ who vote $v$ ($A$gree or $D$isagree or $P$ass) on comment $c$
    • $N(g,c)$ is the number of people in group $g$ who vote at all on comment $c$.
    • The values of 1 & 2 added to the numerator and denominator of the above fraction are psuedocounts, and are used as a semi-naive prior in frequentist statistics.

  • We then compute $R_v(g,c)=\frac{P_v(g,c)}{P_v(\bar{g},c)}$

    • Here $\bar{g}$ is the complement of $g$, that is, all the participants in the conversation who are not members of group $g$.

  • Selection criteria are a bit more challenging to describe, but this involves looking at the two-property test (basically the fisher test) and multiplying this by the Representativeness to obtain a mashup number reflecting both estimated effect size and confidence.