Multi-task equidistant embedding

At least one processing device including a processing system and at least one computer-readable storage medium, the at least one processing device comprising:

• equidistant embedding layers of a neural network, the equidistant embedding layers configured to enforce, for each respective category of a plurality of categories described in a dataset, an equidistant relationship defined using a distance measurement among a plurality of features formatted as one-hot encoding vectors corresponding to tasks included within the respective category in a reduced dimension dataset in a reduced dimension space from a dimension space that is maintained from the dataset, the reduced dimension space generated by:
• applying constraints to the dataset using an embedding loss function based on a comparison between a dense embedding matrix and a binary feature matrix;
• concatenating embeddings of the reduced embedding space;
• performing batch normalization on the reduced dimension dataset;
• and
• reducing overfitting of the reduced dimension dataset by randomly dropping units from the reduced dimension dataset;
• shared layers of the neural network, the shared layers configured to extract feature interactions between the features described in the dataset corresponding to at least one of a first task and a second task;
• exclusive layers of the neural network corresponding to the first task, the exclusive layers corresponding to the first task configured to utilize the extracted feature interactions to generate a first machine learning model corresponding to the first task by:
• segmenting the first task from the plurality of tasks of the dataset using k-means clustering;
• and
• comparing the first task to the plurality of tasks of the dataset using a joint loss function that includes an embedding loss and a segmentation loss based on the k-means clustering;
• and
• exclusive layers of the neural network corresponding to the second task, the exclusive layers corresponding to the second task configured to utilize the extracted feature interactions to generate a second machine learning model corresponding to the second task.

A computer-readable storage medium storing executable instructions, which when executed by a processing device, cause the processing device to perform operations comprising:

• receiving a dataset corresponding to at least one task, the first dataset describing a three-dimensional Euclidean space having three categories and a plurality of features formatted as one-hot encoding vectors corresponding to tasks included within respective said categories, the plurality of features having equidistant relationships, one to another, within respective said categories in the three-dimensional space;
• generating training data that enforces the equidistant relationships of the plurality of features corresponding to the dataset in the three-dimensional space within respective said categories in a two-dimensional space by:
• applying constraints to the dataset using an embedding loss function based on a comparison between a dense embedding matrix and a binary feature matrix;
• extracting complementary information including feature interactions by:
• segmenting a particular feature from the plurality of features of the dataset using k-means clustering;
• and
• comparing the particular feature to the plurality of features of the dataset using a joint loss function that includes an embedding loss and a segmentation loss based on the k-means clustering;
• concatenating the training data with the complementary information;
• training a machine learning model using the training data;
• performing batch normalization on the training data;
• and
• reducing overfitting of the training data by randomly dropping units from the training data;
• receiving an input corresponding to the at least one task;
• generating a prediction of an outcome of the at least one task based on the input and the trained machine learning model;
• and
• outputting the generated prediction of the outcome of the at least one task.