Can x2vec Save Lives? Integrating Graph and Language Embeddings for Automatic Mental Health Classification

  • 2020-01-04 20:56:21
  • Alexander Ruch
  • 0

Abstract

Graph and language embedding models are becoming commonplace in large scaleanalyses given their ability to represent complex sparse data densely inlow-dimensional space. Integrating these models' complementary relational andcommunicative data may be especially helpful if predicting rare events orclassifying members of hidden populations - tasks requiring huge and sparsedatasets for generalizable analyses. For example, due to social stigma andcomorbidities, mental health support groups often form in amorphous onlinegroups. Predicting suicidality among individuals in these settings usingstandard network analyses is prohibitive due to resource limits (e.g., memory),and adding auxiliary data like text to such models exacerbates complexity- andsparsity-related issues. Here, I show how merging graph and language embeddingmodels (metapath2vec and doc2vec) avoids these limits and extracts unsupervisedclustering data without domain expertise or feature engineering. Graph andlanguage distances to a suicide support group have little correlation (\r{ho} <0.23), implying the two models are not embedding redundant information. Whenused separately to predict suicidality among individuals, graph and languagedata generate relatively accurate results (69% and 76%, respectively); however,when integrated, both data produce highly accurate predictions (90%, with 10%false-positives and 12% false-negatives). Visualizing graph embeddingsannotated with predictions of potentially suicidal individuals shows theintegrated model could classify such individuals even if they are positionedfar from the support group. These results extend research on the importance ofsimultaneously analyzing behavior and language in massive networks and effortsto integrate embedding models for different kinds of data when predicting andclassifying, particularly when they involve rare events.

 

Quick Read (beta)

loading the full paper ...