clio - a virtual book recommendation system
Shared by bhaprayan · 854d ago · 23 comments
#hackathon

This is an exploration in search of a personalized book recommendation system based on user search history. To design this we're running machine learning (topic models) over past user search history, and then making personalized recommendations. Our goals are to ensure strict privacy, and hence will either be running model inference on the client, or following differential privacy mechanisms. The current service is being prototyped as a Chrome Web Extension.

Team Members:

bhaprayan
aryanvdesh

Webpage: https://aryanvdesh.github.io/clioweb/
URL: https://github.com/bhaprayan/clio
Walkthrough: https://www.loom.com/share/e3a6417afc4c4feda238f73998a85fca

Brainsprays.com · 854d ago

Useful. Looking forward to the release.

bhaprayan · 854d ago

thanks! yes, us too! our next steps our to prioritize and build some features & ideas we had in the process of working on this over the past 24h.

vijpandaturtle · 854d ago

Cool idea ! Need this

john-vandivier · 854d ago

reading into my bookmark history is both a brilliant, unique concept, and also low-key creepy. still upvoted

bhaprayan · 854d ago

haha thanks! agreed, sort of is, but that's why we're trying to ensure strict privacy. we're users of this too ;)

deliverator-140970 · 854d ago

like it!

gabrielenvienta · 854d ago

sound interesting!

josefran · 854d ago

I love the idea, something that can go one step further in replacing goodreadss!

bhaprayan · 854d ago

haha thanks! for sure, that's the goal ;)
it's time someone displaced the incumbent, i for one, think their UI is an eye sore

@olddd · 854d ago

Great idea! My only concern would be trusting the data? Maybe you can elaborate a bit on how do you plan to handle that part in the future?

bhaprayan · 854d ago

thanks! for sure :)
at the moment everything happens on the client side (i.e. no data is leaving your browser). in the future we're planning to move over to running more recent deep learning based topic models, in which case we're thinking of porting over to differential privacy mechanisms to avoid any user association. else, we could run model inference in the browser too (using tensorflow.js for example), but the efficiency of that approach is still unclear. perhaps if we generate recommendations asynchronously (like once an hour or so), that could warrant periodically running model inference in the browser (this would be complicated by reduction in end user performance though if not done right). needs more thought!

Tash · 854d ago

I like this idea. Would I be able to read in books in that I have read on my Kindle, and have books be recommended to me that way? I wonder if Amazon has an API for that.

ender-004958 · 854d ago

Do you want to be recommended books on your previous read books?

bhaprayan · 854d ago

that's a good idea! a quick search yielded that they don't yet have an API for that (https://forums.developer.amazon.com/questions/196052/ability-to-read-kindle-data-via-a-public-api-end-p.html). would need further exploring on how to pull data in a convenient way from the Kindle, does seem like a useful feature to add though!

bhargava25 · 854d ago

The idea is great but i hope it has something accure measures so we can know the why that particular book has been recommended

bhaprayan · 854d ago

this is a good idea :)
we don't current have strong metrics as to why a current book has been recommended. perhaps we'd have better tools for attribution once we move over to separating out the logic for topic models into its' own module. we were originally planning to go the deep learning route, and perform model inference through a dedicated server, but have deferred that plan for now (due to time constraints + the need to put more thought into ensuring search history privacy (i.e. using DP), in the event of data transmission)

hermione-768476 · 854d ago

I totally would use this!

bhaprayan · 854d ago

wow thanks for the encouragement! hopefully we're able to stick to our future goals and build enough features for this to be useful.

princess-leia-251050 · 854d ago

I'd love to see this as my book reccomendation system! Keep up the good work!

bhaprayan · 854d ago

thanks, me too! we're planning to continue building post hackathon (for example a better topic model, and also fixing issues with latency and the current parsing logic). going through the motions of creating a proof of concept was super useful :)

art3mis-208524 · 854d ago

great

nell-182215 · 854d ago

Interesting concept. Would love to see the final version

bhaprayan · 854d ago

thanks! for sure, it's still a work in progress :)