Niche topic centroid targeting blog #485
Conversation
I plan to generate a graphic for this post from real embedding classification clusters just to illustrate its effectiveness.
There was a problem hiding this comment.
This seems like a good post, but I think it mixes technical content and product-related content in a way that won't really get much traction. I think if we want to talk about the product, we should probably not include code, and if we want to include tech/code, we should talk mostly about that, instead of burying it at the bottom. This is currently a mix of the 2, which doesn't really explain effectively on either count.
It's probably fine to post this, but I think as is, it likely won't get much engagement.
If we wanted a product-focused post, I think we should talk much more about how effective it is, instead of burying that fact in the copy. If we want to talk about tech, we should talk more about how we're choosing pages to classify and the overall workflow. In general it's also not really talking about the wins here in a way that would make people excited. The only mention of performance is on the ~4th paragraph as a throwaway line.
Co-authored-by: Eric Holscher <[email protected]>
Co-authored-by: Eric Holscher <[email protected]>
Co-authored-by: Eric Holscher <[email protected]>
Co-authored-by: Eric Holscher <[email protected]>
Co-authored-by: Eric Holscher <[email protected]>
Co-authored-by: Eric Holscher <[email protected]>
- Give a snappier title - Shorten the background
|
I reworked the post quite a bit to try to shorten the background and get to some of the meat a little faster. |
| @@ -1,60 +1,62 @@ | |||
| Title: Contextual Topic Targeting with Embedding Centroids | |||
| Title: Improving AI Ad Targeting with Embeddings | |||
There was a problem hiding this comment.
This is still not super likely to get engagement, but seems you're hesitant to make it more attention grabbing :)
There was a problem hiding this comment.
How about "Improving Ad Targeting with a Vibes Based Topic Classifier"
There was a problem hiding this comment.
I think talking about why LLM isn't the right option is probably still the most likely to get attention -- but leading with Ad Targeting is probably going to make sure it doesn't get engagement on HN anyway 🙃
There was a problem hiding this comment.
That's fair. We could add a section based on how we tried just sending content to an LLM and how the results weren't as good.
How about "Improving Classification with a Vibes Based Topic Classifier Built with Embeddings"?
There was a problem hiding this comment.
That is still not mentioning LLMs in the title. I think the current title is probably fine if we aren't going to make it more controversial. That is what I'm trying to get at.
| Also, while the results were great, explaining embeddings and page similarity to marketers proved difficult, | ||
| making the approach harder to sell despite its effectiveness. |
There was a problem hiding this comment.
We arguably still have to explain these concepts to people, just filtered through "topics", but I'm pretty sure we could explain niche targeting in a similar way by just saying "content similarity" or something. I think we could also just sell niche targeting as "custom topic" or similar, but it's still the same logic underneath.
| This essentially answers the question of "how DevOps-ey is this content" or "how Frontend-ey is this content" | ||
| for all possible topics. | ||
| for an arbitrary number of topics. | ||
| To steal a term, it's a vibes-based classifier. |
There was a problem hiding this comment.
This could be a catchy title :)
I plan to generate a graphic for this post from real embedding classification clusters just to illustrate its effectiveness.