Participants labeled ads in our dataset of 500 ads using one or more opinion labels, such as "boring", "good design", and "deceptive".
What can we say about ads using these labels? For example, what kinds of ads elicit similar reactions from people? Which opinions towards ads tend go together (e.g. "clickbait" and "deceptive")? What are the combinations of opinions that people have towards high quality ads, versus low quality ads?
To understand these questions, we clustered ads based on the distribution of opinion labels, using a technique called Population Label Distribution Learning. Below, you can explore the clusters of ads that our model produced.
|Cluster||Number of Ads||Avg. User Rating||Description|
|Cluster A||13||5.23||Top quality, well designed ads - mostly for consumer products|
|Cluster C||15||4.86||High quality, simple ads|
|Cluster D||115||4.53||High quality ads, mostly for consumer products|
|Cluster E||3||4.52||Confusing but well-designed ads|
|Cluster G||59||4.07||Average quality ads for consumer products|
|Cluster I||101||3.81||Average quality, low relevance ads (for niche interests or B2B products)|
|Cluster J||4||3.67||Average quality ads with political content|
|Cluster L||46||3.29||Average quality unattractive ads - mostly native ads or B2B ads|
|Cluster M||10||3.13||Ads with political content, e.g. TV programs, political T-shirts|
|Cluster N||3||3.12||Ads for strongly disliked products, e.g. vape pens|
|Cluster P||15||3.04||Poorly designed and confusing ads, i.e. unfamiliar or missing brand names|
|Cluster Q||29||2.95||Poorly designed clickbait ads|
|Cluster R||39||2.8||Clickbait ads with sexual or distasteful content|
|Cluster T||7||2.31||Deceptive and clickbait ads with political content|
|Cluster U||31||2.21||Deceptive ads, e.g. supplements and software downloads|