About our dataset

Raw data

We gathered reviews about several e-shops from Yandex.Market. In order to do this, we derived the following criteria:

• e-shop is "reliable" when it has about the same count of negative and positive reviews (judging by grade);
• e-shop is "unreliable" when it has mostly (> 90%) positive reviews + review count is low (< 200). We expect reviews for "unreliable" shops to be mostly paid, because real people tend to express more opinions than "excellent".

Then we manually created three lists of e-markets: reliable, unreliable and everything else.

Reliable:

• МВидео
• МТС
• Мегафон
• Эльдорадо

Unreliable:

• Landcom
• Opt-Device
• Telebar.ru
• SLK-Service.ru
• BONCH.PRO
• ProAnima.RU

Everything else:

• Cifrovoi.com
• Apple-Zone

We gather 100 reviews per market using Y.Market API (1200 in total), then shuffle those groups and split them into 4 equal parts (300 reviews per part).

Statistics

See initial data folder.

data.json

• Raw reviews count: 1200
• Number of negative reviews: 186

data1.json

• Number of entries: 300
• Number of negative reviews: 23

data2.json

• Number of entries: 300
• Number of negative reviews: 100

data3.json

• Number of entries: 300
• Number of negative reviews: 60

data4.json

• Number of entries: 300
• Number of negative reviews: 3