Danish techies claim they can predict your next move (and your last)

Life's a vector, then you die

Just as language models can predict what phrase might come next in a sentence, Danish researchers claim to have shown human life events can be predicted using similar statistical techniques.

Early mortality and personality nuances were among the target predictions of the team based at the Technical University of Denmark who drew upon education, health, income, occupation, and other life-event data from around 6 million people from a national register.

The study published in Nature Computational Science shows that the foundation model may be able to provide a quantitative understanding of human behavior.

"The findings demonstrate that, by representing the complex linkage between social and health outcomes, accurately predicting life outcomes may be possible," the paper said. However, the authors stress their research is an exploration of what may be possible, and it should only be used in real-world conditions under regulations that protect individual rights.

The research team, led by networks and complexity science professor Sune Lehmann, have named their model "life2vec," presumably after the language processing technique word2vec, which helped develop the idea of representing words as multi-dimensional vectors that can be used to suggest their semantic relationship with other words.

Applying the technique to human life events, the research team said they could establish complex relationships between concepts such as health-related diagnoses and place of residence to income levels, encoding individual lives with a compact vector representation that forms the foundation for the prediction of life outcomes.

For example, the paper says life2vec can forecast the likelihood of individuals from a cohort in the age range 35-65 years surviving the four years after January 1, 2016, outperforming other state-of-the-art statistical models by at least 11 percent.

"Life2vec opens a range of possibilities within the social and health sciences," the paper said. "By means of a rich dataset, we can capture complex patterns and trends in individual lives and represent their stories in a compact vector representation. Event sequences are a common data format in the social sciences, and our work shows how powerful transformer methods can be in unveiling the patterns encoded in such data.

"In our case, the embedding vectors represent a new type of comprehensive linkage between social and health outcomes. The output of our model, coupled with causality tools, shows a path to systematically explore how different data modalities are correlated and interlinked and use these interlinkages to explicitly explore how life impacts our health and vice versa."

The researchers reckon the approach would allow computational social science researchers to "establish comprehensive models of human lives in a single representation" and "open the door to a new and more profound interplay between the social and health sciences." ®