feat: Machine Leaning Spam Detection #666
Conversation
feat: rewrite UserPipeline to include user id feat: correct user pipeline for user id feat: fix user id column in dataframes
… 'email', 'affiliations', and 'bio' of MemberProfile
fix: small fix in all_users_df(). - Convert df.value from markup to string - Fix the name of df.columns
This spam classifier only looks at the MemberProfile's bio and nothing else.
feat:retrieve data from database
Aiko had the idea of using None instead of an extra field which specified if a model has been labelled before. So, we are going to switch to that. chore:cleaned up unused code chore:cleaned up unused code
chore:migrations for altering SpamRecommendation
- load_labels() will take filepath of dataset and laod a dataframe consist of user__id and is_spam columns. Then it will initialize the SpamRecommendation table.
… by an actual dataset with correct labels add 'TODO' comments on the parts to be fixed.
…e functions names in SpamClassifier chore:removed print statements
chore:todo noel chore:organized imports
chore:fixed messy imports
…r, change filenames, UserPipeline to UserSpamStatusProcessor, and SpamRecommendation to UserSpamStatus
feat: added model validation in TextClassifier
manual tests on curator_spam_detection management command passed only for UserMetadataSpamClassifier.
… spam detection related files
…m spam_detection_models.py to spam.py, dataset.csv added under django/curator/
|
|
||
| # Create a new UserSpamStatus whenever a new MemberProfile is created | ||
| @receiver(post_save, sender=MemberProfile) | ||
| def sync_member_profile_spam(sender, instance, **kwargs): |
There was a problem hiding this comment.
I believe we need to check the created argument https://docs.djangoproject.com/en/4.2/ref/signals/#post-save otherwise this will try and fail to create a second object when a memberprofile gets updated
| self.processor = self.detection.processor | ||
| self.user_meta_classifier = self.detection.user_meta_classifier | ||
| self.text_classifier = self.detection.text_classifier | ||
|
|
There was a problem hiding this comment.
I think this may be better done with subcommands/subparsers (https://docs.python.org/3/library/argparse.html#sub-commands) since it seems the intent is to just run one action. Also may want to consider combining common workflows
There was a problem hiding this comment.
I'm still trying to figure out what to fix for this suggestion. I'll ask you again today at our meeting.
| predict_text = options["predict_text"] | ||
|
|
||
| load_directory = pathlib.Path(DATASET_FILE_PATH) | ||
|
|
There was a problem hiding this comment.
could potentially shorten this to use dynamic method calls e.g. getattr(self, f"handle_{action})()"
refactor: create new file 'spam_processor.py' for UserSpamStatusProcessor. change name from dataset.csv to spam_detaset.csv
|
apologies for the pivot, but after starting to go through the https://course.fast.ai I'm leaning towards using fast.ai for much of this (and pytorch as well) instead of tensorflow + scikit... I think it should be a straightforward replacement or additional strategy tbh, and perhaps we can compare the results from both? |
set SPAM_DIR_PATH as a pathlib.Path Co-authored-by: Aiko Muraishi <[email protected]>
tests should use SpamDetector entrypoint instead of instantiating individual components to ensure proper initialization move initial training dataset path into settings move UserSpamStatusProcessor from detected file into curator/models.py as a collaborating class of UserSpamStatus. Should consider integrating more tightly into the UserSpamStatus objects manager
could also convert to sets because there shouldn't be any duplicates
alee
force-pushed
the
spam_aiko__clean_commits
branch
from
5c2d38a
to
83c6830
Compare
No description provided.