Unsupervised learning on feed

[harshita-srivastava-yral]

• There are lot of hype parameters which is going to effect the feed.
• We will need to use less complex model and then test it.
• This will ensure we have not over fed.

[harshita-srivastava-yral]

• Needs help on prioritiy:
• Komal and Jay working on ML feed improved the first feed
• Duplication of the videos - Fuzzy hashing and use models top detect similarity in embedding
• Moderation flow needs to be chalked out

1. Dislike signal mechanism in the feed (Implementation of negative signal from user)
2. Since HOT or NOT videos are mostly the videos not getting categrised we can look at running the categroisation on these videoas.
3. Soft ban users who are regular offender.
4. NU might be bad actors, who are uploading NSFW but not tagging it. Model to check uploaded video and categorise it to NSFW and then we need human moderation by rewarding team -Candidate generator finalising by EOW to improve recommendation (It will take time to get matured. It will not have immediate effect)

[jay-dhanwant-yral]

Video Processing requirements

• scene change detection
• embeddings generation
• nsfw detection (high prio)
• frame extraction & video decoding -> with gpu
• intelligent watermark placement (not a huge prio)

[harshita-srivastava-yral]

• Some hiccups with Akash and GPU provisioning
• Planning to sync-up to connect with SPOC of google.
• Offering by google on CSM detection however unable to move further on this
• Pricing etc isn't available and involve PM to get this resolved
• Lets try out cloudflare
• @jay-dhanwant to connect with @saikatdas0790 to walkthrough the process and check if we can create some login on cloudflare