What’s new?

We’ve been adding a lot to the site lately – here’s a rundown of what’s new:

- When you add a new question to a topic, we now (finally!) let you train it on all the results in the topic (instead of having to go to each result and “fix” it which was super cumbersome).

Here’s what it looks like:

Additionally, at the end of the add question process you can now set Advanced Question Properties. This exposes a lot of functionality that was previously only available to Hunch staff:

On the first tab:

- You can specify how important this question is – High, Medium, or Low – relative to the other questions in the topic. The higher the importance, the more the user’s answer is weighted in the final results ranking.

- You can specify whether you want Hunch to exclusively use human training for this question or whether you want to let Hunch make adjustments to the training via statistical inference. For questions where there’s a clear “right answer” (e.g. the price of a laptop) we recommend using manual settings. For questions that are more a matter of taste (e.g. whether the laptop is good looking) we recommend letting Hunch learn it.

On the second tab:

You can set “dependencies” or “prerequisites” for that question. For example, suppose you are creating a restaurants topic and want to ask “What kind of Italian food do you like?” but only want this question to be asked if the user answered a previous question “I want Italian food.” This tab lets you do that. (Note that Hunch’s question selection algorithm is normally smart enough to figure this out by itself but the dependency feature is a way to tell Hunch that directly and not make it learn.)

The third tab lets you tell Hunch whether any results are irrelevant to this question.

Advanced Question Properties is a big new feature which I’m sure will cause some confusion so I’ve started a forum thread where we’ll answer any questions here.

Other things we’ve added recently:

- Improvements to the workshop. You can now vote separately on a topic’s “concept” (is it a topic that people would find useful? Is it too similar to an existing Hunch topic?) versus “completeness” (does it have enough quality questions and results to be promoted out of the workshop?). You can also have discussions about the topic right on the workshop page.

2) This is a bit subtler, but we have a fresh new algorithm for ranking results according to how you answered the Teach Hunch About You (THAY) answers. For example, if you press Skip to End in a topic and then See All you’ll see the rankings of the results purely according to your THAY answers (e.g. click here to see which magazine’s you might like).

3) When you add a result, we also have a new training flow that is smart enough to take into account question dependencies (it asks you to train dependent questions only when appropriate) and question importances (it asks you to train important questions first).

4) We now clearly label all sponsored links. Here’s an example. As we’ve said before, sponsored links have absolutely no impact on Hunch’s decisions.

5) We now have lots of notifications by email and RSS. When you make a forum post you can subscribe to responses. You can also subscribe to changes to a topic by going to the the Topic Info page (example – see the Subscribe dropdown to the right of Activity Feed).

6) The Hunch learning bot has been unleashed – see our earlier blog post here.

As to upcoming features, we have interesting new features coming out soon including a new widget to complement Teach Hunch About You called Rate Hunch Results which we think will be a lot of fun and also improve Hunch’s data a great deal.

We are also spending lots of time trying to improve the core algorithm. Currently, Hunch has an overall 77% “success rate” which we want to get up to 90-95%. We are building tools to help analyze and improve this number, many of which we’ll be sharing with you in the near future.

Finally, and most importantly, we are working on a version of Hunch that doesn’t require users to login, so that you can embed Hunch topics on other websites and also reach them via search engines. We think this will be important so that your topics can be seen by a broader audience and Hunch is more like a public resource a la Wikipedia.

Thanks again for all your feedback and feel free to make suggestions in the forums.