Loading...
Please wait, while we are loading the content...
Similar Documents
A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions
| Content Provider | Semantic Scholar |
|---|---|
| Author | Friedler, Sorelle A. Wilson, Christo |
| Copyright Year | 2017 |
| Abstract | In this section we describe how we partition the observed data for training and testing purposes. The basic strategy is to form a graph whose connected components correspond to children who have been observed together on at least one referral. By randomly partitioning the set of connected components in this graph we obtain a split that is guaranteed to satisfy two properties: (1) No child will appear in both the train and test test; (2) All children from a given referral will appear together in either the test or the train set. Let ci, i = 1, . . . , Nc denote the distinct children in the data, and let rk, j = 1, . . . , Nr denote the distinct referrals. We will write ci ∈ rk to mean that child i was involved in referral k. Consider the graph G formed by connecting vertices ci and cj whenever there exists an rk such that ci, cj ∈ rk. Let {G`}`=1 denote the (maximal) connected components of G. Randomly partition the set of connected components into two sets, GTrain and GTest. Define Train to be all the referral records of children ci that are vertices of GTrain. Similarly, defined Test to be all of the referral records of children ci that are vertices of GTest. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://proceedings.mlr.press/v81/chouldechova18a/chouldechova18a-supp.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
