I wonder if you could utilize this information. Perhaps negative reviews could get more weight if they're from people who have also left positive reviews.
Something I try to do, if I want to write an unfavourable review, is write a few positive reviews for other companies I've had recent positive experiences with first.
I figure that this gives my negative review some credibility, and it also helps the good businesses.
I wonder if actually enforcing this might help things. If someone can't find anything positive to say about five other companies, I don't think I'm interested in their negative review.
Negative reviews can generally be tracked easily by registered user. It seems like a good enough idea that the force of negative reviews should fall off exponentially with respect to frequency per user that I'm surprised its either not done more, or more apparent.
If someone complains all the time about everything, the people around them quickly learn to ignore it. Our systems should too.
I would think it would be beneficial for the marketplace to see negative reviews, because they will direct customers to the reliable dogwalkers, leading to greater customer satisfaction and retention. Badly reviewed dogwalkers could be bumped off the platform.
> One of the big hurdles to overcome on a user ratings site is that people are most motivated to write reviews when they are outraged. There is a massive selection bias, and if nothing is done to alleviate this, your site runs the risk of becoming a soap box for ranting.
This is the risk we take listening to ANYTHING depending on voluntary responses. Burying negative reviews to attempt some intrinsic rating dismisses the value in actual user input, with the paternalistic view that the review aggregator somehow knows better. Who decides which and how many negative reviews to hide? You? How about letting the user figure it out, knowing that trusting any opinion is a risk and learning from experience to calibrate reviews he reads to himself? It's the best we can hope for without resorting to the review site, in effect, clobbering consensus views to get something it "knows to be more accurate." In that case the users no longer review, but the site, which is not a user, does.
> If you receive 2 bad reviews and 2 good ones, that does not mean 50% of people are dissatisfied.
You absolutely do not know this. It could be that 70% were dissatisfied. You can't assume this without taking the role of a reviewer, which you cannot legitimately do without experiencing the item being reviewed.
Yeah and I have to wonder if company's are reaching out to purchasers who leave negative reviews and offering them incentives to "update" their negative product reviews. I came across a review from a person who bought a Jackery Bolt phone charger who claimed the company did this after she wrote a negative review on Amazon of their product.
No. I know of companies who have management post good reviews to minimize the impact of a real negative reviews. One bad review from a real former employee? Post five good reviews!
Also, I've seen evidence of one employee posting multiple bad reviews. It's a bad system all around.
It seems like maybe the solution is not to remove reviews, but weight them somehow. 100k reviews left in 2 days probably shouldn't have the same weight as 100k reviews left over 2 years.
Or a more sophisticated version of this would be to somehow cluster reviews based on the particular issue they are reporting.
Yes. Here is an example of the evil after I posted a negative review:
> Thanks for your review on Trustpilot.
>
>Based on your review, Curve would like a little more information about your experience. This will help them write a more useful >reply to you. It'll also help them verify that you’ve had a genuine experience with their business.
>Of course, it’s totally up to you what you share.
"verify" vs. "totally up to me" are different things.
When I labeled the comments, I didn't label books that were criticized. So in theory the model should filter out negative reviews. But currently the training dataset is pretty limited in size so you still can see some negative ones. I suspect that with more training data this problem will go away.
reply