The Goal of this project is to find distinct IronMan from Batman which their both have similarity in many aspect Richman save people
- their both have hi-tech car, rich, and be a man
- their both distinct in wearable just iron man wear iron suit, batman wear pvc suit
- its hard for us to distinct these 2 with subreddit post
- so out business is serious about split batman from ironman
via subreddit post only text too since image didn't included in this project concern
Success will be evaluate by we can distinct Batman and IronMan from subreddit post ironman and batman that we mix its and test and we can classify that post is from batman or ironman subreddit
For Ironman's fanclub that need to distinct post so he can follow only ironman not batman
I create sub_reddit_post_gathering to gather post from any subredditwith 3 parameters that can configure
subreddit_namedata_amountoutput_csvso you can collecting data in any sub reddit and any amout you want
but in this case I collected from ironman, batman for approximately 1_000 posts each
with these raw columns
approved_at_utc,subreddit,selftext,author_fullname,saved,mod_reason_title,gilded,clicked,title,link_flair_richtext,subreddit_name_prefixed,hidden,pwls,link_flair_css_class,downs,thumbnail_height,top_awarded_type,hide_score,name,quarantine,link_flair_text_color,upvote_ratio,author_flair_background_color,subreddit_type,ups,total_awards_received,media_embed,thumbnail_width,author_flair_template_id,is_original_content,user_reports,secure_media,is_reddit_media_domain,is_meta,category,secure_media_embed,link_flair_text,can_mod_post,score,approved_by,is_created_from_ads_ui,author_premium,thumbnail,edited,author_flair_css_class,author_flair_richtext,gildings,content_categories,is_self,mod_note,created,link_flair_type,wls,removed_by_category,banned_by,author_flair_type,domain,allow_live_comments,selftext_html,likes,suggested_sort,banned_at_utc,view_count,archived,no_follow,is_crosspostable,pinned,over_18,all_awardings,awarders,media_only,can_gild,spoiler,locked,author_flair_text,treatment_tags,visited,removed_by,num_reports,distinguished,subreddit_id,author_is_blocked,mod_reason_by,removal_reason,link_flair_background_color,id,is_robot_indexable,report_reasons,author,discussion_type,num_comments,send_replies,whitelist_status,contest_mode,mod_reports,author_patreon_flair,author_flair_text_color,permalink,parent_whitelist_status,stickied,url,subreddit_subscribers,created_utc,num_crossposts,media,is_video,post_hint,url_overridden_by_dest,preview,link_flair_template_id,is_gallery,media_metadata,gallery_data,crosspost_parent_list,crosspost_parent,author_cakeday
as number of post is just 1000 so I didn't throttering anythings
after explore on raw data I decide to use only subreddit, selftext, title, and create to use in data modeling and drop other
after check missing value I decide to migrate 2 columns selftext, title in to content
and remove bat man, iron man from each data source for data leaked
also create my_tokenize to cleanup content and improve by remove [, ] from content in tokenize
convert subreddit into binary target column
and merge into 1 data frame
with 2 vectorize and 3 classifier I use Grid Search to loop through each pipeline to modeling and got these score
I considering use f1_score in this case since false_negative and false_positive case didn't much matter but balance is matter more in this case
after run with 1000 posts for each datasource I try to add more data set with 10,000 posts crawling but I also get approx 3,000 post each
and run once again with same methods to check if data source matter found out that for more data set its make model less over fit and can endure variance better
also for some deep down research for each false posts found that some content is clueless with image and meme so if we can filter out meme or image better also can improve accuracy
and with some bag words for each focus topics also can help modeling better or modeling with a lot of data set like chatGPT will also help modeling understand some context muchmore