Predicting Intimate Partner Femicide Using AI

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Oct 29 2023

In an article recently published in the journal Scientific Reports, researchers demonstrated the effectiveness of using artificial intelligence (AI) to extract critical insights from legal documents for predicting intimate partner femicide.

Background

The application of AI in legal document analysis can improve different aspects of law enforcement and criminology. Analyzing legal documents can lead to a better understanding of the patterns and dynamics of specific criminal behaviors, as these documents contain crucial criminal offense-related information.

Several studies have displayed the effectiveness of data extraction from legal documents to investigate the profiles and characteristics of offenders and criminal behavior-related risk factors. For instance, a study determined a substantial association between crime perpetration and substance use by offenders from court judgments.

The use of AI can assist in identifying important characteristics related to criminal activities, such as modus operandi and criminal records, automatically, which improves the understanding of criminal incidents and assists in anticipating potential crimes in similar conditions.

The major advantage of using AI is the ability of this technology to handle large-scale data efficiently and effectively. AI can precisely identify non-linear and complex patterns within data, which makes it exceptionally suitable for the comprehensive analysis of extensive legal documents. However, no study has been performed focusing on examining legal documents to understand crime using an AI-based approach until now.

The proposed approach

In this study, researchers used AI, specifically machine learning (ML) and natural language processing (NLP), to identify relevant information available in legal documents/court judgments for crime prediction. Researchers applied this innovative approach to the specific crime of intimate partner femicide owing to its severe consequences and prevalence.

The sample used in this study contained cases of non-lethal violence (NLV) and lethal violence (LV and intimate partner femicide) against women by their former or current male intimate partners determined by court judgments. Overall, 491 legal documents/cases related to NLV and LV by male-to-female intimate partners, including 161 LV cases and 330 NLV cases, were extracted from the Vlex legal database using NLP.

The 330 NLV cases involved crimes against privacy, sexual integrity, psychological and physical integrity, and personal freedom, while the 161 LV cases involved murders and homicides. All crimes were perpetrated against women by men who were their ex-partners, partners, ex-husbands, or husbands.

A total of 33 independent variables organized into three groups were obtained from the 491 penal court judgments, with the first group including variables about past criminal behaviors and the sanctions imposed for such behaviors, the second group including variables related to situation and environment where the NLV and LV crimes occur, and the third group including variables referred to characteristics of violence perpetrated by the intimate partners of women.

Experimental evaluation and findings

The information in the legal documents was analyzed using 14 AI algorithms belonging to Bayesian, functions-based, tree-based, rule-based, and instance-based classifiers, including BayesNet, NaivesBayes, logistic regression, SimpleLogistic, multilayer perceptron, radial basis function (RBF) classifier, sequential minimal optimization (SMO), Ibk, KStar, logistic model trees (LMT), RandomTree, RandomForest (RF), J48, and JRip.

Researchers implemented these classifiers in Weka, a popular ML toolkit. They applied a hold-out as a specific resampling technique form to the final sample of 491 cases, which divided the data into two sets to ensure the reliability of the outcomes.

The first set containing 324 cases/66% of the total sample was utilized for a training classification model, while the second set containing 167 cases/34% of the total sample was employed for the model testing. The splitting was performed 30 times, randomizing by altering the set partitioner seed. Sensitivity of the first class/NLV class (S1), sensitivity of the second class/LV class (S2), correct classification rate (CCR), F-score, average accuracy (AA), and geometric mean (GM) were the specific metrics considered in this study.

Researchers performed three analyses that focused on the classification of LV and NLV against women by their intimate partners, including the first analysis using only the first group of variables, the second analysis using the first and second group of variables, and the third analysis using the first, second, and third group of variables.

The CCR, F-score, S1, AA, GM, and S2 metrics displayed a significant improvement in the results obtained after the second analysis compared to the results obtained after the first analysis. Additionally, all metrics showed their best performance after the third analysis, which demonstrated that incorporating a larger number of variables leads to an enhanced identification of both LV and NLV cases.

In the third analysis, the RF algorithm outperformed other algorithms in S1, CCR, F-score, and AA metrics, while the BayesNet and RBF classifier outperformed other algorithms in S2 and GM metrics, respectively. The RF algorithm accurately identified NLV cases with a detection rate of 87.04%, classified both the LV and NLV classes with a CCR score of 83.16%, identified classes and captured the majority of cases per class with an F-score of 87.40%, and detected both classes with an AA score of 81.14%, while the BayesNet effectively detected the lethal class with a detection rate of 82.36%.

To summarize, the findings of this study demonstrated the feasibility of using AI to highlight the crucial/specific information present in legal documents to predict crime, specifically male-to-female intimate partner violence. These findings can enhance the prevention and detection of intimate partner femicide cases.