The location of files

The scraped Instagram users are found in /projectnb/llamagrp/davidat/bumc_instascraping/scraped_users/.

Note that scraping is not yet complete because despite writing a new Instagram scraper by reverse engineering, Instagram keeps blocking the scraping after some time.

The trained model is found at: /projectnb/llamagrp/davidat/bumc_instascraping/model_out_dir/finetuned_model.pth.
The predictions for instagram users whose scrapping was finished is found at: /projectnb/llamagrp/davidat/bumc_instascraping/predictions.csv.

Installing the requirements and the repo

This repo uses poetry to handle dependencies. To install poetry if you don't have it, please run:

curl -sSL https://raw.githubusercontent.com/python-poetry/poetry/master/install-poetry.py | python -

After cloning this repo, please run:

poetry install
poetry shell

This will create a virtual environment, install the required python libraries, and activate the virtual environment.

Regarding data collection

Instagram is very hard to scrape because the company has no official API, and they very actively puts up many obstacles to reverse engineering their API, including rate limiting, and requiring face recognition verification if one uses an account to do the scraping.

I tried many scraping libraries on Github, but ended up writing my own. The scraper has some features to help get around Instagram's limitations. This is the output of running python -m bumc_instascraping scrape-users --help:

Options:
  --user-file PATH                A file with one instagram user on each line.
                                  [required]

  --output-dir PATH               The directory where the output JSON files
                                  will be.  [required]

  --posts-per-user INTEGER        How many posts to scrape per user.
                                  [default: 100]

  --backoff-mult FLOAT            If you want to retry failed requests
                                  exponential backoff, set this option as well
                                  as the --backoff-exp option. For example
                                  --backoff-mult 2 --backoff-exp 3 would start
                                  exponential backoff that would initially
                                  wait 2 seconds to retry, and then if the
                                  same request failed 2^3 = 8 seconds, and
                                  then 8^3 = 192 seconds.

  --backoff-exp FLOAT             Look at the documentation for the --backoff-
                                  mult option.

  --ratelimit TEXT                You can specify multiple rate limits. You
                                  specify them in this format:
                                  <number_of_calls>,<number_of_seconds.
                                  Therefore, --ratelimit 5,60 would place a
                                  limitation of a maximum of 5 API calls in
                                  any 60 second block.  [default: ]

  --useproxies / --no-useproxies  If you set this, the code will automatically
                                  fetch proxy server IP addresses and try to
                                  use them. This hasn't proven effective in my
                                  experience because (most likely) Instagram
                                  blocks proxy IP addresses somehow.
                                  [default: False]

  --help                          Show this message and exit.

Below are the recommended arguments to pass while starting / continuing a scraping session.

python -m bumc_instascraping  scrape-users \
    --user-file /projectnb/llamagrp/davidat/bumc_instascraping/valid\ users.csv \
    --output-dir /projectnb/llamagrp/davidat/bumc_instascraping/scraped_users \
    --ratelimit 1,30 --ratelimit 60,3600 \
    --posts-per-user 100

Training

The architecture used was a BERT backbone with two linear layers on top. The output of the last four layers of BERT was used, and only that of the [CLS] token. For each user, the [CLS] output for the last four layers was averaged across all the tweets of the user.

That means that the first linear layer had 3072 input gates (since BERT's hidden size is 768). The second linear layer had 384 input gates. THere was a GELU activation between the first and second linear layer. Binary cross entropy loss was used to train after passing the output of the second layer through the sigmoid function.

The best model trained achieved a validation accuracy of: 73.9%.

It was trained with the following hyperparameters.

learn_rate=2e-5,
epoch_num=10,
max_seq_len=18,
print_step=100,
seed=2021,
accumulation_steps=10,
model_name_or_path='bert-base-uncased'

How to actually train(ie, the command)

One would type in python -m bumc_instascraping train [OPTIONS AS SPECIFIED BELOW] in bash to train a model.

  --max-seq-len INTEGER         [default: 18]
  --print-step INTEGER          [default: 100]
  --seed INTEGER                [default: 2021]
  --accumulation-steps INTEGER  Steps to accumulate loss before doing backward
                                prop.  [default: 10]

  --model-name-or-path TEXT     The HuggingFace model name to use. Usually,
                                'bert-base-uncased'.  [default: bert-base-
                                uncased]

  --finetuned-model-name TEXT   The name of the file that will have the
                                weights of the model.  [default:
                                finetuned_model.pth]

  --age-cutoff INTEGER          [default: 21]
  --device [cpu|cuda]           [default: Device.cpu]
  --help                        Show this message and exit.

Predicting

Similarly, to predict, you'd run python -m bumc_instascraping predict [OPTIONS AS SPECIFIED BELOW].