Flask (Python) website for Machine Learning Youtube tags.
- Write topic of your video
- Paste your Youtube channel
- Wait
- Boom! Ready tags to copy, perfect for your videos.
This model is pretty slow. Need probably more than an hour per query to generate a raport.
I know ChatGPT, can generate tags much faster and have much bigger database, but even of small accuracy of this project, from my expirience this project generates a little bit better result in views for me than from ChatGPT.
| Approximation to the 100 views | Logistic Regression | Random Forest |
|---|---|---|
| Precissly predicted amount of views using only tags and channel name by average. (Accuracy) |
~ 20 % | ~ 15 % |
Accuracy is poor, because i dont have enaught information to work with. My data is really limited. Youtube studio have options to load clickrate, impressions etc. but this data is hidden from public.
Because of that I dont think that, there are bad results.
Considering that the "tags" is not most important indicator of video popularity.
Its probably better than nothing. It probably have some impact. And if so, then why we shoudnt use it? Its is still better than random guessing.
I prefere to consider both models during choosing tags to paste into my videos, because i dont know of how exactly Youtube alghoritm works.
- Program asks user for search query. Program will looks for youtube channels with similar topics. (Will add to channels list all thats uploaders info from videos that will appear after searching that text into youtube search in given pages number)
- Ask you for your youtube channel, to better predict your data. It will add your youtube channel to the list of channels.
- Download video info (uploader name, views and list of tags). For each channel will be downloaded given number of videos info (or less, if channel doesnt have that many videos). This is taking the most of loading time, because for better analysis we recommend to download at least 1000 videos infos.
- Preparing data.
- Generate a boolean table (videos in rows and tags (with channel name in considiration because every channel have some sort of regular audiences) as columns).
- Delete rows (videos) with to small amount of tags (rows with too small information to predict results)
- Delete columns bearly used. (To small amout of videos uses this tag, so predicting result from this tags are uncertain)
- Removing outliners points.
- Creating model In this project i use two models (For now. I planning to add more models later)
- Standard Logistic Regression
- Random Forest Classifier
- Cross-Validation (Accuracy) I am using LOOCV (Leave-One-Out Cross-Validation). It is slow method, but I prefer this method because our sort of data is very specific (Variables may occur too infrequently and if we split a set into test and training we can get improper results, even more so for information about the owner of the video)
- Suggested tags These models gives us the coefficient (or feature importances) numbers that we can use to predict tags that will give us the most of views. Sorted from the higher impact.
- Currently im not considering "date of uploading" as a predictor.
- Looking for faster method for downloading Youtube Data. (Currently using yt_dlp, because it is free), but planning Youtube API in the future.