Multilabel learning with label relationships

A system comprising:

• one or more processors;
• and
• a non-transitory computer readable storage medium comprising instructions that when executed by the one or processors cause the one or more processors to perform operations comprising:
• in an online connection network system, maintaining a plurality of member profiles and a plurality of job postings;
• constructing a machine learning model and initially training the machine learning model to takes as an input (1) a first vector representing a member profile from the plurality of member profiles, and (2) a second vector representing a job posting from the plurality of job postings, and to generate as an output respective probabilities for a set of labels representing respective engagement events initiated by a member represented by the member profile, with respect to the posting, wherein the training uses training data generated based on previously tracked engagement events in the online connection network system with respect to a plurality of pairs comprising a training member profile from the plurality of member profiles and a representative job posting from the plurality of job postings, wherein the engagement events for which the respective probabilities are generated, comprising one or more of:
• viewing the job posting, applying to a job described by the job posting, saving the job posting, or dismissing the job posting;
• accessing electronically stored relationships between the set of labels defining how each label is either a subset relation, an exclusion relation, or an overlap relation with other labels comprising a number of probabilities for valid configurations of relationships between the engagement events, representing a positive probability output from the machine learning model, and a number of probabilities for invalid configurations of probabilities of relationships between the engagement events, representing a zero probability output from the machine learning model;
• encoding, by executing one or more computer instructions, the relationships as a regularization parameter in a training objective for retraining the machine learning model, wherein, and wherein the training objective is minimizing a penalty, the penalty being directly related to the number of invalid configurations predicted by the machine learning model;
• and
• retraining the machine learning model using training data to minimize the number of invalid configurations and the training objective by executing a training process in a computer system.

A machine-readable non-transitory storage medium having instruction data executable by a machine to cause the machine to perform operations comprising:

• in an online connection network system, maintaining a plurality of member profiles and a plurality of job postings;
• constructing a machine learning model and initially training the machine learning model to takes as an input (1) a first vector representing a member profile from the plurality of member profiles, and (2) a second vector representing a job posting from the plurality of job postings, and to generate as an output respective probabilities for a set of labels representing respective engagement events initiated by a member represented by the member profile with respect to the job posting, wherein the training uses training data generated based on previously tracked engagement events in the online connection network system with respect to a plurality of pairs comprising a training member profile from the plurality of member profiles and a representative job posting from the plurality of job postings, wherein the engagement events for which the respective probabilities are generated, comprising one or more of:
• viewing the job posting, applying to a job described by the job posting, saving the job posting, or dismissing the job posting;
• accessing electronically stored relationships between the set of labels defining how each label is either a subset relation, an exclusion relation, or an overlap relation with other labels comprising a number of probabilities for valid configuration of relationships between the engagement events, representing a positive probability output from the machine learning model, and a number of probabilities for invalid configuration of probabilities of relationships between the engagement events, representing a zero probability output from the machine learning model;
• encoding, by executing one or more computer instructions, the relationships as a regularization parameter in a training objective for retraining the machine learning model, where the training objective is minimizing a penalty, the penalty being directly related to the number of probabilities for invalid configurations predicted by the machine learning model;
• and
• retraining the machine learning model using training data and the training objective by executing a training process in a computer system.