Spotify Ab

Rishabh Mehrotra

Mounia Lalmas-Roelleke

Benjamin Lacker

Hugues Bouchard

Samantha Hansen

James E McInerney

Edward Y Lee

Federico Tomasi

Aloïs Gruson

Adam Latour Piel

Harpreet Singh

Madaline Minuet Kirwin

Brian Christian Peter Brost

Aasish Kumar Pappu

Lucas Maystre

Dmitrii Moor

Niannan Xue

Christian Hansen

Casper Hansen

Hugo Flávio Ventura Galvão

An electronic device generates a score for each objective in a hierarchy of objectives. Generating the score comprises, using a first machine learning algorithm, generating a score for a first objective corresponding to a first level in the hierarchy of the objectives and using an output of the first machine learning algorithm, distinct from the score for the first objective, as an input to a second machine learning algorithm to generate a score for a second objective corresponding to a second level in the hierarchy of objectives. The electronic device generates a combined score using the score for the first objective and the score for the second objective. The electronic device selects, automatically without user input, media content based on the combined scores for the plurality of media content items and streams, using an application of the media-providing service, one or more of the selected media content to a user.

Using a hierarchical machine learning algorithm for providing personalized media content

Methods, systems and computer program products are provided personalizing recommendations of items with associated explanations. The example embodiments described herein use contextual bandits to personalize explainable recommendations ('recsplanations') as treatments (“Bart”). Bart learns and predicts satisfaction (e.g., click-through rate, consumption probability) for any combination of item, explanation, and context and, through logging and contextual bandit retraining, can learn from its mistakes in an online setting.

Personalizing explainable recommendations with bandits

An electronic device, for each media content item of a plurality of media content items, receives a respective score for each a first set of objectives and one or more other objectives and generates a respective score between a user and the media content item. The generating includes applying a first ordered weighted average to the respective scores for the first set of objectives, to produce a first combined score for the first set of objectives, applying a second ordered weighted average to the respective scores for a second set of objectives, wherein the second set of objectives includes (i) a resulting objective corresponding to the first set of objectives and having the first combined score and (ii) the one or more other objectives. The electronic device streams media content to the user selected based on the respective scores between the user and the media content items.

Systems and methods for using hierarchical ordered weighted averaging for providing personalized media content

Methods, systems and computer program products are provided for query understanding. A non-focused query quantifier generates non-focused query features that quantify a non-focused query and a non-focused query predictor generates a prediction associated with the non-focused query based on the non-focused query features.

Query understanding methods and systems

An electronic device stores a plurality of vector representations for respective media content items in a vector space, where each vector represents a media content item. The electronic device receives a first set of input parameters representing a previous session of a user of the media-providing service where the previous session included two or more of the respective media content items. The electronic device then receives a second set of input parameters representing a current context of the user and provides the first set of input parameters and the second set of input parameters to a neural network to generate a prediction vector for a current session. The prediction vector is embedded in the vector space. The electronic device identifies, based on the prediction vector for the current session, a plurality of media content items of the respective media content items in the vector space and provides the plurality of media content items to the user of the media-providing service during the current session.

Systems and methods for providing media recommendations using contextual and sequential user embeddings

Methods, systems and computer program products are provided personalizing recommendations of items with associated explanations. The example embodiments described herein use contextual bandits to personalize explainable recommendations ('recsplanations') as treatments ('Bart'). Bart learns and predicts satisfaction (e.g., click-through rate, consumption probability) for any combination of item, explanation, and context and, through logging and contextual bandit retraining, can learn from its mistakes in an online setting.

An electronic device generates a score for each objective in a hierarchy of objectives. Generating the score comprises using a representation of the media content item and a user as inputs to a first machine learning algorithm, to generate a score for a first objective corresponding to a first level in the hierarchy of the objectives and using an output of the first machine learning algorithm, distinct from the score for the first objective, as an input to a second machine learning algorithm to generate a score for a second objective corresponding to a second level in the hierarchy of objectives. The electronic device generates a respective score between the user and the media content item using the score for the first objective and the score for the second objective and provides media content to the user based on the respective scores.

An electronic device for a first session of a user, for each of a plurality of lists of media content items, determines a respective value for each objective of a first set of objectives and a second set of objectives by accessing contextual data for the first session of the user. The first set of objectives corresponds to the user and the second set of objectives corresponds to a second party distinct from the user. The electronic device, using a multi-arm bandit model, identifies a first list of media content items, from the plurality of lists of media content items, to present to the user, including: calculating a score for each list in the plurality of lists of media items; and probabilistically selecting the first list of media content items according to the respective scores corresponding to the respective lists in the plurality of lists of media items.

Systems and methods for selecting content using a multiple objective, multi-arm bandit model

An electronic device, for each media content item of a plurality of media content items, receives a respective score for each of a first set of objectives and a second set of objectives and generates a respective score between a user and the media content item. The generating includes applying a first ordered weighted average to the respective scores for the first set of objectives, to produce a first combined score for the first set of objectives, applying a second ordered weighted average to the respective scores for the second set of objectives, to produce a second combined score for the second set of objectives and applying a third ordered weighted average to the combined score for the first set of objectives and the second set of objectives. The electronic device provides media content to the user based on the respective scores between the user and the media content items.

A method of recommending media items to a user is provided. The method includes receiving historical data for a user of a media providing service. The historical data indicates past interactions of the user with media items. The method includes generating a model of the user. The model includes a first set of parameters, each of the first set of parameters quantifying a predicted latent preference of the user for a respective media item provided by the media providing service. The method includes evaluating the predicted latent preferences of the user for the respective media items against the historical data indicating the past interactions of the user with the media items provided by the media providing service. The method includes selecting a recommender system from a plurality of recommender systems using the model of the user, including the first set of parameters. The method includes providing a media item to a second user using the selected recommender system.

Growing community of inventors

London, United Kingdom

Rishabh Mehrotra