Abstract: In one embodiment, a method separates subscriber features generated from subscriber interaction with a video delivery service into feature dimensions and inputs the feature dimensions into a respective prediction network. Each prediction network is trained to output a respective dimension score. The method outputs dimension scores using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a de-correlation term to control a correlation of the plurality of feature dimensions. The dimension scores are combined into a retention prediction score and an action is performed on the video delivery service for the subscriber based on the retention score.

Abstract: In one embodiment, a method generates actions for entities found on a video delivery system based on information for user behavior of a user on the video delivery system and generates probabilities for the actions for the entities based on the actions for the entities and the user behavior. A probability for an action indicates the probability the user would select that action for an entity when compared against other actions in the set of actions for the set of entities. The method then selects an action feed based on the probabilities for the set of actions. The action feed includes at least a portion of the actions for the entities. The action feed is outputted to the client for display on an interface where an action on an entity in the action feed is performed by the video delivery system when selected by the user on the interface.

Abstract: A method receives user behavior information at a first system. The user behavior information is determined by user interaction with a first list sent to the user by a first network on a video delivery service. A first state is generated using the received user behavior information and prior user behavior information by the user from cells that store the prior user behavior. The method inputs the first state into a second network with the first recommendation list to generate a value that evaluates a performance of recommending the first recommendation list. An update to parameters is generated for the first network and provided to the first network. The first network generates a second state from the received user behavior information and prior user behavior information derived from cells that store the prior user behavior and outputs a second recommendation list using the second state and updated parameters.

Abstract: In one embodiment, the method renders a first set of objects at a first frame rate and a second set of objects at a second frame rate that is lower than the first frame rate. The second set of objects is stored in a buffer. Then, the method displays the second set of objects from the buffer for a number of frames based on the second frame rate while the rendered first set of objects is displayed at the first frame rate. When a visual change associated with the second set of objects over multiple frames meets a threshold, the method renders the second object at a third frame rate different from the second frame rate based on the visual change meeting the threshold.

Abstract: A method receives an image from a video. The image is split into a first set of first blocks of a first size and then the first blocks are split into a second set of second blocks of a second size. The method tests a first set of down-sampling patterns for the second set of second blocks in a first block in the set of first blocks to determine if a quality of reconstruction of the down-sampled second blocks meets a threshold associated with the respective first set of down sampling patterns. Second blocks satisfying the threshold are down-sampled using the first set of down-sampling patterns. Also, the method tests a second set of down-sampling patterns for the first block using a second set of down-sampling patterns to select one of the second set of down sampling patterns to use to down-sample second blocks that did not satisfy the threshold.

Abstract: In one embodiment, a method receives a portion of a bundle of preview images for a video being played and starts to decode the portion of the bundle of preview images. An input is received for a location in the video. The method determines if a preview image corresponding to the location has been decoded from the bundle of preview images. When the preview image corresponding to the location has been decoded, the preview image corresponding to the location is displayed. When the preview image corresponding to the location has not been decoded, the method switches from decoding the portion of the bundle of preview images at a first position to decoding the portion of the bundle of preview images at a second position corresponding to the location.

Abstract: In one embodiment, a method retrieves mask information for creating a visual effect in a three-dimensional (3D) video. The mask information includes an outline of an object. The video is stored in a first buffer. The mask information is processed in a second buffer to create a mesh that extrudes from the outline of the object. The mesh includes values that describe a shading of the mesh and the mesh being formed around the outline of the object. The method then adjusts the values of the mesh in the second buffer to minimize discontinuities in gradients created in the mesh. When a time to display the visual effect is reached in the video, the method renders the mesh from the second buffer with video from the first buffer to form the visual effect around the object to indicate that an interaction with the object is available in the video.

Abstract: A method receives ratings for videos from a first user that is using a video delivery service. A first model includes connection networks where each connection network corresponds to a rating. The method inputs each rating into a connection network in an order. Also, parameters for the ratings and ratings other than the rating received from the first user are modeled in a respective connection network. Values for the set of parameters are trained such that the plurality of connection networks predict conditional probabilities that the first user would provide the rating corresponding to the each connection network in the order. The conditional probabilities are based on the first user providing ratings that are previously located in the order. The parameters are then used to generate a list of videos to recommend to the first user using the first model.

Abstract: In one embodiment, a method selects a set of accounts related to a video using a first classifier that is trained to identify the set of accounts from a plurality of accounts and selects a token used to identify the video. A second classifier is trained to identify short messages that are related to the video using a first set of short messages from the accounts related to the video and the second set of short messages including the tokens. A third set of short messages is identified that is related to the video from a plurality of short messages. The method normalizes words in the set of short messages using a mixed model including a lexicon model and a translation model. Then, the method generates a set of ratings for the set of short messages and uses the set of ratings in a service for a video delivery service.

Abstract: A method transmits configuration information including a set of events that define when a media program player should send information regarding playback of an advertisement during an advertisement break in a media program. Also, the method transmits, to the user device, advertising control options configured to be output on the user device. A user selection for a control option in the advertising control options is received from the user device, wherein the control option controls sending of an advertisement to the user device during the advertisement break. The method then receives advertisement viewing data describing a viewing of the advertisement on the user device, wherein sending of the advertisement viewing data by the media program player is triggered based on an event in the set of events occurring at the media program player.

Abstract: In one embodiment, a method generates chunk boundaries that are used to form a plurality of chunks from a video. The chunks are transcoded by transcoding units in parallel. The method selects at least one of a first number of frames before a first boundary of a chunk and a second number of frames after a second boundary of the chunk. A chunk is generated that includes the at least one of the first number of frames and the second number of frames, and a third number of frames between the first boundary and the second boundary. Then, the chunk is transcoded to generate a transcoded chunk using a first transcoding unit. The chunk is sent to a chunk assembler where the chunk assembler removes the at least one of the first number of frames and the second number of frames from the chunk and adds the chunk to a transcoded bitstream.

Abstract: A first node determines that a health of a second node that is the accepted leader is invalid and calculates a third proposal number for a proposal of an election for a leader. The first node sends first messages with the third proposal number to nodes to propose the election and determines that a first quorum for a third variable is reached indicating the election for the leader is accepted. The first node sends second messages with a fourth proposal number based on the first proposal number and an identifier for a proposed leader to the nodes to propose the proposed leader as the leader of the nodes. The second messages reject the proposed leader when the proposed leader is equal to the fourth variable. The first node determines that a second quorum is reached from a set of second responses that include the fourth proposal number for the second variable.

Abstract: In one embodiment, a method receives a request for a channel offered by a video delivery service from a user and generates user information for the user. In response to receiving the request, the method sends a query to a database to retrieve a bundle that includes an audience definition that matches the user information. A bundle identifies a source stream for a program and only one bundle for a time slot is returned for the query from multiple bundles for the channel that exist for the time slot. A manifest is constructed for the channel using the bundle where the manifest identifies content from the source stream of the bundle. The method then provides portions of the manifest to a client device for the user to allow the client device to request content from the source stream for the channel.

Abstract: In one embodiment, the method determines a second location associated with a client device that requested to view an encoded video. Replacement content to insert in the encoded video to replace existing content is selected based on the location of the client device. The method decodes blocks in a frame of the encoded video in which the replacement content will be inserted and analyzes the decoded blocks to determine first blocks of the encoded video that will be affected by insertion of the replacement content and second blocks that will not be affected by the insertion. Motion prediction information is saved for the second blocks. The method then encodes the frame where the replacement content is inserted to replace the existing content. Motion prediction information for the second blocks is reused when performing the encoding of the second blocks.

Abstract: In one embodiment, a method receives information for a video entity where a video delivery system delivers the video entity as a plurality of video assets using different video delivery methods. The method generates bundles for the video assets. Also, dependencies are generated between bundles when two bundles include control data parameters that are the same. The method then stores control data parameters for the bundles for the plurality of video assets based on the dependencies. A second bundle is stored with the dependency on a first bundle and the second bundle includes a second set of control data parameters for the second bundle that is different from the first set of control data parameters included in the first bundle. When the second bundle is selected, the method accesses the first set of control data parameters and the second set of control data parameters for the second bundle.

Abstract: In one embodiment, a method receives a video and information for entities that appear in the video. The video is played in a media player without displaying a queue configured to display one or more entities. When an input is received to display the queue while playing the video, the method performs: determining a set of entities in relation to a time associated with playing the video using the information for the entities that appear in the video and displaying the set of entities in the queue, wherein the video continues to play in the media player while the queue is displayed.

Abstract: In one embodiment, a request is received from a user to open a stream for an upcoming second live media program during a time period in which users request a manifest for the upcoming second live media program that will start after a first live media program ends. The method uses a per device stream limit for the user to determine whether to allow the request. A device identifier database is accessed to determine current device identifiers associated with devices that the user is currently using to receive media programs. The method determines whether a device identifier associated with a device that sent the request is found in the current device identifiers. In response to determining that the device identifier for the request is found in the current device identifiers, the request to open the stream is allowed based on determining that the per device stream limit is not violated.