An Algorithm That Solves Crimes

Investigative Genetic Genealogy (IGG) is a novel approach to solving violent criminal cases and identifying unidentified human remains. Genotyped DNA samples from a crime scene are submitted to a third-party service (e.g., GEDmatch Pro or FamilyTreeDNA) who output a list of biological relatives (or matches) to the unknown criminal (called the unsub, which is short for unidentified subject). Genealogists then construct a speculative family tree for the unsub using these matches, with the aim of identifying the unsub. A mathematical model of the genealogy process in IGG was constructed and analyzed in Erturk, Fitzpatrick, Press and Wein, Journal of Forensic Sciences, 2022, which resulted in a proposed strategy that attempts to identify the criminal suspect or unidentified human remains as quickly as possible. This strategy, which determines the next action for a genealogist to take, was tested on randomly generated variants of 17 cases from the DNA Doe Project and the simulation results suggested that the strategy could solve cases much more quickly than a standard benchmark strategy.

This talk describes recent work that builds on Erturk’s model. Whereas Erturk’s model works with randomly simulated cases, here we develop an interactive algorithm that can be used for actual (rather than simulated) cases. In this sequential interaction, the algorithm tells the genealogist what action to take (e.g., investigate a particular match and find its ancestors for a specific number of generations), and the genealogist reports the results of this action back to the algorithm, which then updates the state of the system and tells the genealogist the next action.

The interactive algorithm requires several new capabilities relative to Erturk’s model, and incorporates several refinements and generalizations, some of which were inspired by using the algorithm to try and solve actual criminal cases. In the first portion of this talk, Wein will briefly describe these new capabilities and refinements, which include:

• Hierarchical clustering that assigns each match to an ancestral couple of the unsub in each generation;
• After successfully identifying a match, updating the probability distribution of the relationship between the match and the unsub using the dates of birth of the match and the unsub;
• Community ascent, which – when ascending a match – also simultaneously ascends matches that are more closely related to the match than to the unsub;
• Dual descent, which descends a specific number of generations from two lists of possible ancestral couples of the unsub, with the aim of finding a connecting couple (e.g., marriage) at a specific generation among the descendants;
• A new benefit metric for descents, which leads to the ultimate goal of finding the unsub’s parents, which is the connecting couple between the two sides (maternal and paternal) of the speculative family tree;
• Assessing the reliability of each hierarchical clustering result, and prioritizing actions that employ a reliable clustering result;
• When inconsistencies (e.g., finding a duplicate ancestor in the wrong generation) — due to assuming the wrong relationship between the unsub and a match, an inaccurate clustering result, or endogamy — are detected, we update the relationships, the system state, the clustering and past actions.

In the second portion of the talk, Rae-Venter will describe her experience using the algorithm to solve several difficult criminal cases, some of which she had already been working on for three years. Thus far, the algorithm has helped her solve five of these cases. Some issues that arose in these cases include the emancipation wall when the unsub is Black, familial testing, endogamy and reference testing.