SuperrootPodcastsRecommendationsPodcastRecsFeatures

Average fraction of podcast listened per show. Ignores duration < 10s. Computed offline

Probability of a random user from this cluster listening to this podcast randomly (K-means specific feature)

Language of shows in user history and how much they listened to each.

Total duration listened for this episode by all users. Ignores duration < 10s. Computed offline

Per cluster lift where the lift_squashing_factor is set to +2

boost score from the similarity between the candidate and the shows with user negative feedback, currently the score will be applied directly to the final ranking score: go/podcast-reaction-reranking-v1, but can be used as reranker model feature in the future.

Show level trending score percent from Wernicke corpus.

Scaled pagerank score for the feed url in [0..1]. Not to be confused with pagerank above, which measures the pagerank for the feed's homepage.

How frequently are podcasts published by this show. To see publishing frequency calculation go here: http://shortn/_6zzfyEpBRq

Some of all fraction of this show listened by users. Ignores duration < 10s. Computed offline

Ownership verification status for the episode page URL.

This captures the probability that this show could have been listened by the user in lieu of what they have subscribed or listened to.

This captures the fraction of total listening time accounted for by this show.

Score of csai safe search score go/safesearch

category_match is a 0 to 1 score depicting how much of the user's listening history matches the categories of this podcast recommendation.

Per cluster lift where the lift_squashing_factor is set to -3

Match score between user listening mids and episode webref entities, where the listening is limited to last two mohths activity. Values are in a 0.0 to 1.0 range.

This captures the fraction of total listening time accounted for by this subscribed show.

Ordinal rank features like this have been found to be useful in Hermione Recipes, where the top ranked items from a candidate generator is always retained. Ranks are integral values starting with 1 for the highest affinity show.

final reaction boost score computed from positive_reaction_boost_score and negative_reaction_boost_score. The score will be applied as a multiplier on the ranking score to adjust the ranking.

From: indexing/speech/proto/colisten-matrix.proto Absolute number of unique listeners during the past month.

Score of vulgar safe search score go/safesearch

Probability of a random user listening to this podcast randomly

The distance between user's Anima embedding and the centroid of the cluster in k-means.

Show only impressions. Doesn't include shows of episode impressions.

Anima User Embedding based features. Dot product of Anima User Embedding and podcast embeddings from UserEmbeddingBasedSignals (podcasts/quality/proto/ranking_signals.proto)

Ranks are integral values starting with 1 for the highest affinity show.

Number of minutes of podcasts listened by the user

Number of show impressions during the past week.

Number of all episode impressions.

Base quality of the document, used as a multiplier for the query-specific score. Should be in [0, 1] range. See: http://g3doc/indexing/moonshine/generic/g3doc/doc/scoring

Score of porn safe search score go/safesearch

A binary value based on whether this is a canonical source for a cluster.

Average duration listened per episode. Ignores duration < 10s. Computed offline

Some of all fraction of this listened by users. Ignores duration < 10s. Computed offline

Expresses the absolute popularity rank within all documents.

Total duration listened for this show by all users. Ignores duration < 10s. Computed offline

The rank of the show in the subscription candidate generator. Ranks are integral values starting with 1 for the highest subscribed affinity show.

Match score between user listening mids and episode webref entities, where the listening can have happened at any time in the past. Values are in a 0.0 to 1.0 range.

Score of spoof safe search score go/safesearch

Score of fringe safe search score go/safesearch

The total minutes listened to this podcast show by users in this cluster

Feature ID of a location. For more info, see go/feature-id.

Dot product of user embedding and podcast show embedding from the two tower model v2a

Number of podcasts listened by the user

Per cluster lift where the lift_squashing_factor is set to 0.4

rank_percent_contrib = 100 / ( 1 + rank_percent ) Here rank_percent was derived from show level data in Wernicke corpus So if rank_percent of a feed was 3, then the feature value would be: 100 / (1 + 3) = 25. This indicator falls off quickly from 100 to almost 1 as we go from rank_percent 0 to 100.

Number of show episode impressions.

A reranking feature showing the age of the episode that is being recommended. This is computed as (current time - publication_time) publication_time of the episode is defined: (http://shortn/_S46Ouk5ZWW) publication_time is specified as seconds since Unix Epoch.

Language of the show from show data.

Match score between user interest mids and episode salient entities. in a 0.0 to 1.0 range.

boost score from the similarity between the candidate and the shows with user positive feedback, currently the score will be applied directly to the final ranking score: go/podcast-reaction-reranking-v1, but can be used as reranker model feature in the future.

Match score between user listening mids and episode webref entities, where the listening is limited to recent activity. Values are in a 0.0 to 1.0 range.

The number of users in this k-means cluster.

Inferred language preferences of the user with their probabilities.

The duration of a single episode, in seconds. - For show documents, this is a duration of a typical recent episode, or an approximation thereof. - For episode documents, this is the duration of the episode itself.

Score of medical safe search score go/safesearch

Whether the show is marked as explicit by the authors

Average fraction of podcast listened per episode. Ignores duration < 10s. Computed offline

This is the language extracted from the query_language.

The rank of the show in the top listened shows candidate generator. Ranks are integral values starting with 1 for the highest affinity show.

The number of users in the k-means cluster who have listened to this feed.

Number of episode impressions during the past week.

How much of user listening history match the language of this episode

The index in the list of most popular podcasts.

How much of ulp match the language of this episode

From: indexing/speech/proto/colisten-matrix.proto The number of subscribers for this podcast series.

Match score between user interest mids and episode webref entities. in a 0.0 to 1.0 range.

This captures the rank of the podcast show in the colisten candidate generator.

Average duration listened per show. Ignores duration < 10s. Computed offline

Score of violence safe search score go/safesearch

Sigmoid of the score obtained by dot product of user embedding and show embedding. See https://b.corp.google.com/issues/158602034#comment2 for intuition.

Score of offensive safe search score go/safesearch