Using polling results as discrete metrics for content quality prediction model

Abstract

A social networking system presents content items to users, who then provide feedback regarding pairs of content items. The feedback includes a selection of a content item of the pair of content items that was preferred by the user over the other content item. The social networking system uses this information to train a predictive model that scores content items based on quality. The content items may be advertisements. The social networking system uses the pair-wise comparisons of the advertisements to determine feedback coefficients in an advertising quality score prediction model using regression analysis of the pair-wise comparisons for each predictive factor in the model. In this way, the pair-wise comparisons are used to train the prediction model to understand which advertisements are more enjoyable than others. A feedback coefficient for each predictive factor may be computed based on the preferences received from the group of users.