search

euagendas / m3inference

A deep learning system for demographic inference (gender, age, and individual/person) that was trained on massive Twitter dataset using profile images, screen names, names, and biographies
http://www.euagendas.org
GNU Affero General Public License v3.0
145 stars 57 forks source link

Question about training procedure #1

Closed StanislasBertrand closed 5 years ago

StanislasBertrand commented 5 years ago

Hi, First of all thank you for your great work.

I was wondering what you used as ground truth label for age and gender when user profiles are organizations. You wouldn't want the model to train to recognize any gender / age on an organization profile. I believe this is not mentioned in the article, or maybe I misunderstood something about the training procedure ?

davidjurgens commented 5 years ago

Organizational accounts are excluded from the training sets for age and gender. Each of the tasks (age/gender/is-org) use separate training sets, as not all accounts have all the labels.

StanislasBertrand commented 5 years ago

This makes sense for the pre-training of the separate models, but for the fine-tuning of the entire M3 model, wouldn't you need input users that have a ground truth for every attribute ? Maybe you could define a loss function that only calculates the error on organization (cutting out softmax for age / gender) when an organization user is used as training input ?

davidjurgens commented 5 years ago

The fused models are fine-tuned using one attribute at a time. (Single-task fine-tuning for multi-modal input) I don't think we looked into a joint loss because so few accounts have both age and gender.

StanislasBertrand commented 5 years ago

Ok, thank you for clarifying.