3 ms·
Hi Potatolicious, I work with the Machine Learning team at Ness. Thanks for all the feedback! Please email us at support@likeness.com as you have more though
by schiff 15y ago
Hi Potatolicious,
I work with the Machine Learning team at Ness. Thanks for all the feedback! Please email us at support@likeness.com as you have more thoughts, especially about where Ness does and does not work for you.
With regard to only teaching Ness by rating restaurants 4-5 stars, this is a great point, especially for the first 10 ratings. There are a couple particular ways we address this today:
1) Which restaurants you pick tells us a lot. Did you rate hole in the walls vs. expensive places? We have spent a lot of time working on our collaborative filter to address this. Wikipedia has a good article for a starting ground on this, and the papers that came out of the Netflix Prize are great [http://www.netflixprize.com/assets/GrandPrize2009_BPC_BellKor.pdf http://www.netflixprize.com/assets/GrandPrize2009_BPC_BellKo...].
2) On the personalize page, there are multiple stages. One of them is to confirm that the system is correct about places you don't like, which is obviously also very informative.
3) Sometimes, as with your Chipotle example, we won't be right. Telling us where we are wrong is particularly helpful in correcting the system, which in some sense gives us more information that confirming what we already know. Think of it of correcting a person, who misunderstood what you like from examples you gave. In a metaphorical sense, this is like providing results that have more entropy.
-Jeremy
VP of Product & Machine Learning
- guimarin 15y agoI'm really happy that you have addressed this problem. My question comes down to, in the rating process, I rated a fairly limited amount of places, in terms of categorical scope. Then when I moved into the main application, there were many many more categories available, like ice-cream and bbq. I found myself not finding restaurants that I liked loading in the initial 10 choices, so I kept skipping a lot. You might consider a continuous scroll feature for this initial rating set. Otherwise, integrate with open table, and this will rank as the best vertical search/discovery app I've ever come across.
- schiff 15y agoGuimarin, Thanks for the feedback. We made a slightly different mechanism for rating on the personalize page (the list of places you saw when you first came to the app allowing you to quickly rate). Either a user can skip a place manually by clicking the X, or can scroll down to the bottom and skip all the places that you don't like. You are saying you would rather just scroll past the ones you don't know? -Jeremy
- perfmode 15y agoMany apps made these days make use of infinite scroll. Users are getting used to this style. If the purpose of the Skip All button is to compile a list of restaurants the user has not yet visited, and if this information is sufficiently useful such that you don't want to get rid of the button altogether, perhaps consider a happy medium: for instance, a Skip All of the Above button placed between every ten restaurants on an infinitely scrolling list. I, for one, think the scroll down -> skip all -> scroll down -> skip all design is clunky--especially since I've visited less than 5% of the restaurants presented at any given time.
- guimarin 15y agoYes, that's exactly what I'm saying. I don't know if I agree with the other commenter about interspersed skip alls. For me, scrolling past seems less final then hitting skip. If I hit skip, I was afraid that a negative mark would be given to the restaurant. I know that's not the case, but it was a very strong incentive for me not to want to hit skip, or the x. I didn't feel like harshly judging a place that I'd never been to.