Machine Learning over 1M hotel reviews | Hacker News Reader

Machine Learning over 1M hotel reviews

blog.monkeylearn.com

145 pointsfeconroses41 comments

"Interestingly, ‘Internet’ doesn’t really rise above 70% in any hotel class, indicating that internet access is as bad in 3 star hotels as in 5 star hotels. One would expect that more expensive hotels would provide a better internet service, but If you’ve ever stayed at a 5 star hotel you know that’s not usually the case..."

The more stars the Hotel has the more likely you have to pay for WiFi. If I have to pay for it, I expect it to perform better than if it's free. So I'm going to diss it harder if it is not fast and consistent.

Worst part is that paid Wifi usually works less reliable. I hate it when I stay at a conference hotel where the best I can do is to pay for 12 or 24 hours of internet service and then it brings up the shitty portal page again :-/

Have you tried iodine?

http://code.kryo.se/iodine/

What sort of performance hit does that incur?

From the page: The bandwidth is asymmetrical with limited upstream and up to 1 Mbit/s downstream.

Yes, my experience is that you won't get great performance. But these networks rarely work very well either.

It's good enough for casual surfing and checking email.

I couldn't guess why hotel rating negatively correlates with wifi quality, but the correlation for cost of wifi makes sense. Higher quality hotels in big cities are more likely to get business travelers, and business travelers are more likely to have someone else pay for wifi.

"The more stars the Hotel has the more likely you have to pay for WiFi"

Luckily today, you can see whether Wi-Fi is free on the premises, and give the Hotel its well deserved boot if it's not.

How they ever get this idea is beyond me.

Hotel WiFi in fancy hotels has a much higher expectation of quality / service level than that in a small-ish motel or youth hostel. If the CEO of $BigCorp has to have a Skype call with the client, they won't want package drops and 10 minutes of 'Hello can you hear me'.

I am guessing the overall cost of running a reliable reasonably-high-speed WiFi in the whole hotel is much greater than just having a bunch of routers.

Of course, whether fancy hotels actually end up delivering said quality and service levels, I dunno.

Of course, whether fancy hotels actually end up delivering said quality and service levels, I dunno.

They don't, my experience says your WiFi will be more crappy at a Hilton/Westin/whatever than a Best Western or even a Motel 6.

No, the reason you pay more is because odds are you aren't paying for it, your company is. Motel 6 could never get away with it because that money comes out of my pocket. But when you're putting it on the corporate Amex...

This dynamic cannot be overstated. The economics of services that are heavily business-expensed is ridiculous.

There was an article here years ago (that I'm too lazy to find via my phone) that claimed that many high end hotels got into deals with providers of paid hotel internet early on when offering wifi was unusual and more involved. At the time the service was cutting edge and pricing justified. Contacts were long running though and now hotels are stuck with it.

There's an inversion at ca.$1000/night where the connectivity becomes as good as the personal chef.

It wasn't long ago that telephone charges provided hotels a significant revenue stream. Good summary here--one which also notes that quality wifi installations are expensive for big hotels, and must be constantly maintained and regularly upgraded.

http://www.usatoday.com/story/travel/columnist/2015/02/08/ho...

More stars tend to mean more expensive, usually where most corporate travel would stay. Those wifi? they write it off on their expense rewards or get it for free with hotel reward cards. They rest of normal folks wanting to stay at very nice hotel pay for it. They charge for it because people pay, same way airlines charge for every stupid thing today and people pay.

I think usually "The rest of normal folks wanting to stay at very nice hotel" is just "the folks that were too lazy to register for hotels reward program". At least that's the case with Hilton, Marriott, SPG.

I have consistently and for a long time wondered what "corporate travel" is thinking of when they offer a list of accommodations. Then again, I sort of hate "very nice hotels". They look like TV soap opera sets which is annoying.

It's all "paid"; you are just not given the option to not pay for it if it's "free".

There are some amazing hotels in London. I've had some amazing experiences. My own observation may be anecdotal but i think this comes down to what each dollar gets you for your money. $150 in London gets you an average hotel. $150 in Bangkok gets you something top notch. Chances are most of the analysis has been done at fixed price points, not accounting for relative value. That would explain why London rates poorer.

Berlin, Lisbon, Budapest and Bruxelles are some of those cities where you can stay at a 5-star extremely cheap.

London's the worst for it; I have never been happy with the cheaper 5-star (less than 260 GBP/night) hotels there. Maybe if you stay at Mandarin Oriental...

On the other hand the restaurants are a great value, especially lunch menus: save the 100 GBP on the 5-star hotel and spend 55 GBP (+12.5% "optional" gratuity) on a 3-course lunch menu, with 2 glasses of wine and coffee at an amazing 2-Michelin-star restaurant like "The Square". You'll still have enough for a last-minute good half-price theatre ticket at tkts.co.uk in the evening.

Trip advisor requires jumping through a few hoops to see a price so I doubt they have that data at all. And with hotels getting into the dynamic pricing game it will only get more complicated.

It's hard not to get a little pissed when I'm paying twice as much and things are not done well so I suspect the value factor does not fully capture the effect of London being so damn expensive. Even when I've moved to a more expensive city, it talked time for the sticker shock to where off.

I can corroborate some of the high level factoids in this. I work for a company with a large set of hotel sentiment data that should be a superset of the data in the linked post. It's probably around 35M reviews over several major sites.

82% of sentiments being positive sounds plausible. Ours is somewhat lower but still well over 50%. I think it's in the 70s. Practically speaking, I've had some really frustrating conversations with people in the industry trying to explain that a hotel can have positive sentiment and still be way below average for a given cohort.

Wifi sucks - that matches our data. Within the most common entities we extract, the only thing worse than wifi is "maintenance", which has a much lower occurrence rate. Unsurprisingly, nobody really talks about maintenance unless something's broken. Location is generally highly rated. Surprisingly to me, so is "service" and equivalents.

Anecdotally, the croissant thing sounds correct. A lot of reviewers use the example of cold croissants to describe an unimpressive continental breakfast.

Ah, hotel Wi-Fi... maybe it's not as good as the one at home because... "in theory" you don't need it as much?

Like, if you are a family on vacation @ beautiful Aruba, maybe management thinks that you will spend time on the beach, or golfing, or in spa or using any other of their 60+ carefully tended-to amenities?

Perhaps they can't imagine the whole family sitting in their hotel room for 14 hours a day with 4 iDevices spread across Netflix/Youtube/Twitch, everyone complaining that they can't get more than 720p.

Lol, who knows.

The cross-group results are close enough that you really need to use confidence intervals for the proportions, even with the large sample size. Especially because sentiment analysis does not perfectly classify positivity/negativity.

On a semi-related note, TripAdvisor will likely not appreciate you scraping 1M reviews, especially since even their officially-sanctioned API explicitly says "you can't use this for data analysis." (https://developer-tripadvisor.com/content-api/request-api-ac...)

Isn't it an overreach to make absolute statements from sentiment analysis? There are some bold claims in this article that can only stand if the authors corrected for anchoring, confirmation bias, and herd behaviour.

I also believe that people with bad experiences tend to review hotels more often than those with good experiences. That would bias the entire data set in one way or the other (obviously this has nothing to do with the quality of the tool being presented).

A more accurate conclusion would be that London's bad hotels are a popular option.

During my tour around Europe, I quickly realized that all the hotels in the City of London (other than extremely expensive ones) are very cheap but also awful.

How do we feel that they stole that data against the TOS?

While I don't have any particular opinion on TOS violations here, I have to point out that it's definitely not stealing. At worst, it's an unauthorized copying.

When I was in second grade, there was a rule in the school library that grades 1-3 were only allowed in section A, grades 4-6 were allowed in section B. I regularly crawled around the barricade so I could read the Hardy Boys books and the Great Brain stories that were squirrelled away there. I got caught several times and bawled out, but I never felt bad.

But I still feel worse about that than this TOS violation.

Is this a really great example of Machine Learning? It seems most of the work could have been done with common data stores and a bit of coding.

They could have provided richer graphs, more of them, or at least a link towards a firehose of results. Maybe I'm missing something?

As is the trend in 2016, the article serves as promotional content marketing for the startup OP works for, by using data that cannot be questioned.

It is a personal pet peeve and one of the reasons I make all my statistical analyses open source. (That being said, in this case, the OP does have a notebook available on GitHub)

Out of how many cities? They only present 7.

I found this line interesting:

"Some trial and error followed for getting the best model for the data"

Unlike AI a human can do that at a glance, with no prior training.

"with no prior training"

is a false statement, the prior training is life and learning to read and having human experiences. For example if all reviews were in english and the human only spoke Mandarin they would need "prior training" before being able to preform the task.

I don't mean training on the dataset. I mean training on the model, i.e. knowing what to pick to train on.

Based on my own personal experiences... fuck London. People are rude and indifferent to you. The hostels are all retarded, overpriced pubs. British food is terrible. Everything is overpriced, even by NY/SF/European standards. Even the UK hates Londoners

Gee, I wonder why people are rude to you?

It is my experience that the people are cold and indifferent, but once they warm up to you, you have a fantastic relationship with them that is above superficial and stands the test of time.

Well, when your UK friends from other parts of the commonwealth tear into Londoners worse than you do, perhaps the problem is indeed with them?

When you're passing through it is hard to get to know anyone.

London is by far the worst treatment I have seen from anyone in the western world, rich or poor. Even city Dutch are nicer and more kind.

No, the problem is with you.

Reply on news.ycombinator.com