soft-404: a classifier for detecting soft 404 pages

A "soft" 404 page is a page that is served with 200 status, but is really a page that says that content is not available.

Installation

pip install soft-404

Usage

The easiest way is to use the soft404.probability function:

>>> import soft404
>>> soft404.probability('<h1>Page not found</h1>')
0.9736860086882132

You can also create a classifier explicitly:

>>> from soft404 import Soft404Classifier
>>> clf = Soft404Classifier()
>>> clf.predict('<h1>Page not found</h1>')
0.9736860086882132

Development

Classifier is trained on 198801 pages from 35995 domains, with 404 page ratio of about 1/3. With 10-fold cross-validation, PR AUC (average precision) is 0.991 ± 0.002, and ROC AUC is 0.995 ± 0.002.

Getting data for training

Install dev requirements:

pip install -r requirements_dev.txt

Run the crawler for a while (results will appear in pages.jl.gz file):

cd crawler
scrapy crawl spider -o pages.jl.gz -t jl.gz -s JOBDIR=job

Training

First, extract text and structure from html:

./soft404/convert_to_text.py pages.jl.gz items

This will produce two files, items.meta.jl.gz and items.items.jl.gz. Next, train the classifier:

./soft404/train.py items

Vectorizer takes a while to run, but it's result is cached (the filename where it is cached will be printed on the next run). If you are happy with results, save the classifier:

./soft404/train.py items --save soft404/clf.joblib

License

License is MIT.

---