Having used a feed reader for the better part of a year, I wanted to see if RSS could be augmented with a recommender system that could highlight the content I find more interesting. LSS is that solution, just through click-tracking it learns to predict what items you would read by their titles.

Example of LSS feeds in a feed reader

• See which RSS items you would read (⭐), might read (🔵), and probably wouldn't read (🔴). The items you probably wouldn't read only pass through the filter if you toggle filter: false in your config.yml.
• HTTP GET the URLs of all your filtered feeds at /feeds, paths are auto generated to keep the config.yml minimal.
• Clicking an item's default link updates the title as average (🔵) in the database, and additional links for updating the item as either good (⭐) or poor (🔴) are added to the top of the item's description.
• Filter with an ensemble of classical TF-IDF regression and transformer-based DistilBERT classification.
• Toggle filtering for individual feeds or all feeds with the filter and cold_start fields, respectively.
• Set weights for each classifier, their weighted softmaxes are added up to classify items.
• Inference and most of the service runs inside the container, DistilBERT training runs on AWS Sagemaker and requires the AWS CLI to be configured on the host machine.

0. Recommended to be run behind a reverse-proxy with authentication, default port is 5000.
1. Copy over the config.yml and crontab from config/examples to config, make any changes as you'd like.
2. If using DistilBERT, add an iam_role field to the config that you'll have to create on AWS, also need to have an AWS config at ~/.aws for said IAM role.
3. Starting out, I recommend to set cold_start: true at the top of your config.yml until you have enough data for the training cron jobs to work.
4. Run docker compose up production -d

08/12/24 - Has labels for poor/average/good ratings that are associated with red, blue, and star emojis in the generated feeds. TF-IDF classification and DistilBERT had better F1 scores when they just did binary classification instead of ternary, looking into a solution that is good for small datasets. Test coverage needs improvement, alembic mostly unused so migrations are still done manually, needs documentation for setup and probably an init script. Thinking of including confidence scores in the generated feeds next to the emojis.

09/17/24 - Through passive use of the application I'm seeing better results i.e. getting over the cold-start hurdle. Here's a recent evaluation metric from the docker compose logs, note that initially the precision/recall scores were closer to that of a trivial random classifier (~0.33 for ternary classification):

lss-prod  | {'eval_loss': 0.8556984066963196, 'eval_accuracy': 0.671875, 'eval_f1': 0.6756054421768708, 'eval_precision': 0.6959779006737331, 'eval_recall': 0.671875, 'eval_runtime':
        │  0.8686, 'eval_samples_per_second': 221.042, 'eval_steps_per_second': 3.454, 'epoch': 3.0}