liveBook · Manning

Impression

As explained in Hidden Influences

A count of how many times content is displayed to users in a distribution or recommendation system.

A post appearing in a feed can be counted each time it is viewed. Each view is an impression: 10 impressions mean the post was seen 10 times, not necessarily by 10 different people. A person who sees it twice contributes two impressions. section 5.5

Impression counts depend on the time window. A post might receive 100 impressions on its first day and another 100 on its second. Comparisons should specify the period measured. If the question is how many different people saw a post, unique impressions can be used instead. section 5.5

Impressions help measure algorithmic amplification: how a recommendation system changes content exposure. From a producer’s side, the question may be whether distribution is fair. From a reader’s side, it may be whether recommendations reflect the reader’s preferences. section 5.5

Impressions can support several comparisons:

  • In an A/B test, compare a post’s exposure under a new recommender with exposure under a baseline, accounting for the different sizes of the test groups. A ratio of 2 means twice the exposure under the new system. section 5.5.2
  • Compare impressions with a producer’s follower count to help account for differences in audience size. Impressions can also be split between followers and people outside the producer’s network. section 5.5.2
  • For a reader’s feed, compare a producer’s share of impressions with its share of the content available for recommendation. This helps distinguish preferential distribution from simply having produced more content. section 5.5.3

A feed with ten available posts

Suppose three available posts come from one producer, three from another, and four from other sources. If the reader sees all ten, the first producer receives 30% of the impressions. If that producer instead receives 50%, its share of views is higher than its 30% share of available content. section 5.5.3

An impression count alone does not settle whether exposure is fair or matches a reader’s preferences. Measuring reader-side amplification also requires knowing what content was available for recommendation, which is cumbersome to record at scale. section 5.5 section 5.5.3

Table 5.1 Types of algorithmic amplification
Metric
Side
Question being answered
A/B tests
Author
How much extra exposure does specific content receive, thanks to the algorithmic recommendations?
Normalized impressions
Author
How can we compare different kinds of users?
Out-of-network ratio
Author
How much exposure comes from direct followers rather than from showing the content to other users?
Feed analysis
Reader
Do the recommendations match the intent of the reader?

Compare the questions that different producer-side and reader-side amplification measures ask. section 5.5

How Hidden Influences covers it

How deep it goes
Defines time-dependent counts and develops exposure comparisons, while noting the difficulty of recording all recommendation candidates.
Examples
Uses a ten-post feed and a recommender test with 1% of traffic assigned to the new model.
What you can do after
Evaluate exposure by comparing impressions across recommenders or against the content available to a reader.
How it connects
Connects impressions to algorithmic amplification, A/B testing, follower counts, and feed analysis.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Hidden Influences. The section links point to the parts of the book it draws on.