Explaining Machine Learning Models by Tracked Behavioral Latent Features

A computer program product comprising a non-transitory machine-readable medium storing instructions that, when executed by at least one programmable processor, cause the at least one programmable processor to perform operations comprising:

• receive a time series of input variables representing a time-ordered sequence of observations associated with an entity;
• generate a word dictionary for the input variables, the word dictionary representing a unique sequence of event words across the entity and a group of other entities
• generate a set of archetypes from the unique sequence of event words, where topic modeling associates a loading of archetypes associated with each event word;
• generate, with a topic modeling module, a vector of weights representing a distribution of the time series of input variables aligning with each of a given number of latent topic variables, the given number of topic variables corresponding to latent features of behavior based on variable values;
• produce, according to a machine learning model, a detection score associated with the entity for each input event in the time-ordered sequence of observations;
• generate one or more explanations, each explanation describing a reason corresponding to the produced detection score;
• filter, through comparison against profiled average values of latent topic variables, the one or more explanations to generate a filtered list of explanations according to archetype, the filtered list providing a list of archetypes associated with the detection score and increasing in relevance of behavior;
• and
• rank, according to a closeness of a bin of historical latent topic variables' expected score to the current score on the event, the filtered list of archetypes and associated reasons that most affect the detection score.

A system comprising:

• a programmable processor;
• and
• a non-transitory, machine-readable medium storing instructions that, when executed by the programmable processor, cause the at least one programmable processor to perform operations comprising:
• receive a time series of input variables representing a time-ordered sequence of observations associated with an entity;
• generate a word dictionary for the input variables, the word dictionary representing a unique sequence of event words across the entity and a group of other entities
• generate a set of archetypes from the unique sequence of event words, where topic modeling associates a loading of archetypes associated with each event word;
• generate, with a topic modeling module, a vector of weights representing a distribution of the time series of input variables aligning with each of a given number of latent topic variables, the given number of topic variables corresponding to latent features of behavior based on variable values;
• produce, according to a machine learning model, a detection score associated with the entity for each input event in the time-ordered sequence of observations;
• generate one or more explanations, each explanation describing a reason corresponding to the produced detection score;
• filter, through comparison against profiled average values of latent topic variables, the one or more explanations to generate a filtered list of explanations according to archetype, the filtered list providing a list of archetypes associated with the detection score and increasing in relevance of behavior;
• and
• rank, according to a closeness of a bin of historical latent topic variables' expected score to the current score on the event, the filtered list of archetypes and associated reasons that most affect the detection score.