Real time fine-tuning of recommendation systems

By Pranjal Timsina

The impact of recommendation systems and content moderation in our everyday life is subtle yet profound - what we read, watch and listen are directly influenced by black-box systems. The extent of the impact cannot be overstated.

This proposal outlines a method addressing a few problems pertaining to recommendation systems and content moderation.

Problem #1: Sometimes the recommendation system needs to react to real world changes

Let’s take for example a time where Swine-flu is rapidly spreading in a particular region. It is imperative that the consumption of pork is controlled. Amongst many things, one thing that might help would be to suppress posts describing pork recipes, and restaurants which primarily sell food based on pork; but this begs the question - how do we suppress such posts or news articles? The answer is real time fine-tuning of recommendation systems.

Problem #2: Content moderation is difficult and expensive

A lot of sites still employ people to moderate the content posted on their media platform. This method is both time consuming and expensive. There has to be a better solution

Problem #3: End-users should be able to customize recommenders

Typical recommenders learn about the user behavior and make sure to give the user the posts which probably have the highest average engagement rates. This works pretty well; however, the user explicitly describing what they want to see should be possible as well.

The solution

Fine-tuning recommendation systems in real time is the solution for all of these perceived problems. While, this sounds like something that requires quite a bit of tinkering, we believe we have found a way - Language Models, and data.

We’ve spent months meticulously creating the dataset, and experimented with smaller language models for vectorizing and ultimately recommending news articles; we even submitted a paper for publication which is under review for the same topic. What we believe is that we can get much better results using better language models, with more and better data and using better hardware.

Improving the results that we obtained from the work would require a language model such as Meta Llama 2 or 3, or other similar language models. However, a big constraint due to which, we are unable to do so is hardware.

Footnote

The purpose of this excerpt is to put forth an idea which we believe is amazing. The contents of this post will be updated over-time, adding further details to what we had done and plan to do.

---