Abstract: A server system increases the diversity of item recommendations provided to a target user by using item similarity data to reorder a ranked list of recommended items for presentation to a user. The reordering is performed such that items identified as similar to each other are spaced apart from each other by at least a minimum number of positions. This minimum number may be selected based, e.g., on how many recommended items will initially be presented on a requesting user device, which may depend on the display size or other attributes of the user device. The server system generates a user interface that displays an initial portion of the reordered list. The user interface includes controls for scrolling through the reordered list.

Abstract: Systems are provided to analyze user behavior with respect to different versions of content, and to identify the behavioral effects that can be attributed to the differences between the versions of content. Systems can determine whether an observed difference in user interactions with different versions of content is caused by the specific changes that have been made to content (e.g., an actual improvement or other substantive change), or whether the observed difference in user interactions is caused by the mere fact that the versions of content are different (e.g., due to the novelty of seeing a different version of content for the first time, regardless of whether the difference is a substantive improvement). In some cases, the observed effect may be caused both by the novelty of the content being changed and by the substance of the changes.

Abstract: A recommendation system increases the diversity of item recommendations provided to a target user by using item similarity data to reorder a recommendation set of items complementary to a source item. In one embodiment, the complementary items in the recommendation set are ranked based on a relevance score that represents a degree of relevance to the source item. The ranked recommendation set is then reordered based on overlap scores that represent degrees of similarity between particular items so that adjacent items and/or groups of items have less than a threshold degree of similarity. The overlap scores may be generated based on an automated comparison of user item viewing behavior, item attributes, item content, or based on another measure of item similarity.