Hello! I present to you a simple Python project that I've been working on for some time. This is a recommendation system for e621 that works by analyzing tags from your favorites and comparing them with the latest posts.

https://github.com/DestroyingFilms/e6-Recommendation-System

The algorithm itself is not that complex, but the whole application is a GUI app, which tries to automate the whole process. As I've described in my repository, it works by following these steps:
1. Request profile (username, API key)
2. Read and use settings (blacklist, default post rating, etc.)
3. Download favorited & latest posts
4. Compare and give grades to all the latest posts with favorite ones

After that, the recommendations are complete with preview images being displayed on the main window for the top results.

I will briefly write about each of those steps, as more in-depth information can be found in the GitHub repository that I've provided above.

1. Request profile

The application will ask for your e621 username and API key. The API key is optional, as long as you have your favorites NOT hidden, but if you do, then you have to provide it to the app.
The username is kept in the local config.json file, created in the root directory of the application. However, the API key is hidden, only kept in the encrypted form in the OS itself by using keyring backend.

2. Read and use settings

There is an option menu in the GUI, which has configuration that you can modify, such as a blacklist (tags that will not be recommended to you), minimum score threshold (a min score that the post must have to be recommended), default post rating (a rating that the post must have to be recommended), and much more. Those options are saved to config.json.
Content that is commonly considered objectionable is blacklisted by default, some of which are ALWAYS blacklisted. Additionally, the default post rating is set to "Safe", which will only use "Safe" recent posts for recommendations.

3. Download favorited & latest posts

As the application uses the API service of e621, which has some limits, it is best to download the needed data for favorite posts and latest ones to a local folder. So that it can be used later without needing to send the same request to the API twice or even more. The only data which is needed from those posts is an ID and a list of tags (with a url for a preview image and a file extension of the original post additionally being used for latest posts).

4. Compare and give grades

And finally the recommendation system itself, which uses a simple cosine similarity function to find the best match for each post. What might be important to point out is how the app saves the results, as the top 30 recommendations will have preview images downloaded locally and encoded in base64 to keep in the results.json file. This is once again done to not overflow the API service with useless requests every time the preview image needs to be loaded.

This is pretty much it. As I wrote in the repository, this is a hobby project made in my free time, so there is probably some bugs that I've overlooked and haven't fixed yet. But I am open to criticism and would be glad to hear feedback. I don't guarantee that I will regularly fix issues and bugs. But, I will try my best to fix whatever would be possible once I have enough time for it.

Hope you will still enjoy this simple tool. And have a nice day!

I'm currently building an on-site related posts feature, and I was curious to know how exactly you derived relevancy data to enter into the cosine similarity formula? My current approach was to not analyze the image, but instead calculate the Levenshtein distance between a chosen post's tags & a sampling of random posts' tags & order them by shortest distance. Are you doing image analysis, tag analysis, or something else?

aacafah said:
I'm currently building an on-site related posts feature, and I was curious to know how exactly you derived relevancy data to enter into the cosine similarity formula? My current approach was to not analyze the image, but instead calculate the Levenshtein distance between a chosen post's tags & a sampling of random posts' tags & order them by shortest distance. Are you doing image analysis, tag analysis, or something else?

I am using the CountVectorizer from scikit-learn to create a matrix of tokens from all the tags.
It uses a bag-of-words approach, which counts the number of tag occurrences in the tag_string (a string variable which contains all the tags for each post separated by a comma and a space).
There is also TF-IDF Vectorizer, which takes into account the importance of each token. But I didn't use it in this program.
The matrix received from those algorithms can be used in the cosine similarity formula to find matches between posts.
However, as an alternative to all of this, the Levenshtein distance might be a good replacement.

Thank you so much for the resources & insight! I was unfamiliar with this prior, I'll have to look in this.

For context, my project is comparing a single post/tag list to a selection of prospective posts, so direct comparison is viable; cosine similarity seems like a much better fit for relevance to a larger set.

No problem! If you need further assistance, I would be glad to help with whatever I can. Good luck!