Heterogeneous content joint distribution method and device, storage medium and electronic equipment

Abstract

The method comprises: in a lower layer of an ascending channel, distributing first content in heterogeneous content to obtain interaction data of the user and the first content, and recalling a target user from the user according to the interaction data; in a middle layer of the ascending channel, constructing a heterogeneous graph according to a single content, the target user and a content album, and determining similarity between the single content and the content album according to the heterogeneous graph; determining whether the single content and the content album can successfully ascend across modes according to the similarity between the single content and the content album; and determining a distribution strategy of the middle layer of the ascending channel and an upper layer of the ascending channel according to a cross-modal ascending result between the single content and the content album. In this way, compared with a single content ascending channel and a distribution method, the cost is saved, and the distribution efficiency of content distribution is improved.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the field of cross-modal recommendation technology, and more specifically, the embodiments of this disclosure relate to a heterogeneous content joint distribution method, a heterogeneous content joint distribution device, a computer-readable storage medium, and an electronic device. Background Technology

[0002] This section is intended to provide background or context for embodiments of this disclosure, and the description herein is not intended to be considered prior art simply because it is included in this section.

[0003] The "Ascension Channel" technology was initially designed to give new content a cold start on the platform, selecting and promoting it through multiple layers until the most popular content is promoted across the entire network, achieving the goals of both users and the platform. Its main processes include: machine review, manual review, recommendations at each level of the channel, and subsequent promotion. These layers work together to screen and ensure new content progresses from a cold start to its final promotion across the entire network. Summary of the Invention

[0004] However, existing growth channels primarily target single-type content (such as videos or songs), and their distribution units are typically of the same type. For business scenarios with diverse content formats, such as podcasts, there is both the distribution of individual content (a single "audio") and the distribution of content collections (a "playlist"). Since different types of content serve different business objectives, the traffic of any single type of content is insufficient to achieve complete content growth and overall business needs on its own.

[0005] Therefore, there is a great need for an improved content recommendation method that allows the upward channel to involve not only the rise of individual content, but also the rise of the content collection to which that individual content belongs, thereby achieving overall business requirements.

[0006] In this context, embodiments of the present disclosure are intended to provide a heterogeneous content joint distribution method, a heterogeneous content joint distribution apparatus, a computer-readable storage medium, and an electronic device.

[0007] According to a first aspect of this disclosure, a method for joint distribution of heterogeneous content is provided, the method comprising:

[0008] In the lower layer of the ascending channel, the first content in the heterogeneous content is distributed to obtain user interaction data with the first content, and target users are recalled from the users based on the interaction data; wherein, the heterogeneous content includes a single content and a content collection to which the single content belongs; the first content is the single content or the content collection to which the single content belongs;

[0009] In the middle layer of the ascending channel, a heterogeneous graph is constructed based on the individual content, the target user, and the content collection, and the similarity between the individual content and the content collection is determined based on the heterogeneous graph;

[0010] Based on the similarity between the individual content and the content collection, determine whether the individual content and the content collection can be successfully promoted across modalities;

[0011] Based on the cross-modal promotion results between the individual content and the content collection, the distribution strategy of the middle layer of the ascending channel is determined; wherein, the distribution strategy of the middle layer of the ascending channel includes single content distribution or content collection distribution;

[0012] In the upper layer of the ascending channel, the distribution strategy of the upper layer of the ascending channel is determined according to the distribution strategies of the lower layer and the middle layer of the ascending channel; wherein, the distribution strategy of the upper layer of the ascending channel includes single content distribution and / or content compilation distribution.

[0013] In one implementation, constructing a heterogeneous graph based on the individual content, the target user, and the content collection includes: using the individual content, the target user, and the content collection as vertices, and using the target user's behavior type towards the individual content and the content collection, and the association relationship between the individual content and the content collection as edges, to construct the heterogeneous graph.

[0014] In one implementation, determining the similarity between the individual content and the content collection based on the heterogeneous graph includes: determining the behavioral similarity between the individual content and the content collection based on the heterogeneous graph; and determining the representational similarity between the individual content and the content collection.

[0015] In one implementation, determining the behavioral similarity between the individual content and the content collection based on the heterogeneous graph includes: constructing a meta-path based on the heterogeneous graph; performing a random walk on the meta-path to obtain semantic and structural relationships between vertices of different types; training a language model to be trained based on the results of the random walk to obtain a trained language model; determining the feature vectors of the individual content and its content collection based on the trained language model, and calculating the similarity between the two to obtain the behavioral similarity between them.

[0016] In one implementation, determining whether a single piece of content and a content collection can successfully advance across modalities based on the similarity between the single piece of content and the content collection includes: determining the advancement score of the heterogeneous content based on the similarity between the single piece of content and the content collection; and determining whether the single piece of content and the content collection can advance across modalities based on the advancement score of the heterogeneous content.

[0017] In one implementation, determining the promotion score of heterogeneous content based on the similarity between the individual content and the content collection includes: using the product of the behavioral similarity, representational similarity, and conversion rate of the current content at the current layer in the ascending channel as the promotion score of the heterogeneous content.

[0018] In one implementation, determining the distribution strategy for the upper layer of the ascending channel based on the distribution strategies of the lower and middle layers of the ascending channel includes: at the upper layer of the ascending channel, if the cross-modal promotion result between the individual content and the content collection is a successful promotion, distributing the individual content and the content collection; if the cross-modal promotion result between the individual content and the content collection is a failed promotion, distributing the content in the form of content from the middle layer of the ascending channel.

[0019] According to a second aspect of this disclosure, a heterogeneous content joint distribution apparatus is provided, the apparatus comprising:

[0020] The lower-level distribution module is configured to distribute the first content in the heterogeneous content at the lower level of the ascending channel to obtain user interaction data with the first content, and recall target users from the users based on the interaction data; wherein, the heterogeneous content includes a single content and a content collection to which the single content belongs; the first content is the single content or the content collection to which the single content belongs;

[0021] The mid-layer distribution module is configured to construct a heterogeneous graph based on the individual content, the target user, and the content collection in the middle layer of the ascending channel; determine the similarity between the individual content and the content collection based on the heterogeneous graph; determine whether the individual content and the content collection can successfully advance across modalities based on the similarity between them; and determine the distribution strategy for the middle layer of the ascending channel based on the cross-modal advancement result between the individual content and the content collection. The distribution strategy for the middle layer of the ascending channel includes single-content distribution or content collection distribution.

[0022] The upper-layer distribution module is configured to determine the distribution strategy of the upper layer of the rising channel based on the distribution strategies of the lower layer and the middle layer of the rising channel; wherein, the distribution strategy of the upper layer of the rising channel includes single content distribution and / or content compilation distribution.

[0023] In one implementation, the mid-layer distribution module is configured to construct the heterogeneous graph by using the individual content, the target user, and the content collection as vertices, and the target user's behavior type towards the individual content and the content collection, and the association relationship between the individual content and the content collection as edges.

[0024] In one implementation, the mid-layer distribution module is configured to: determine the behavioral similarity between the individual content and the content collection based on the heterogeneous graph; and determine the representational similarity between the individual content and the content collection.

[0025] In one implementation, the mid-layer distribution module is configured to: construct meta-paths based on the heterogeneous graph; perform random walks on the meta-paths to obtain semantic and structural relationships between vertices of different types; train a language model to be trained based on the results of the random walks to obtain a trained language model; determine the feature vectors of the individual content and its associated content collection based on the trained language model, and calculate the similarity between the two to obtain the behavioral similarity between them.

[0026] In one implementation, the mid-level distribution module is configured to: determine the promotion score of heterogeneous content based on the similarity between the individual content and the content collection; and determine whether the individual content and the content collection can be promoted across modalities based on the promotion score of the heterogeneous content.

[0027] In one implementation, the mid-layer distribution module is configured to use the product of the behavioral similarity and representational similarity between the individual content and the content collection, and the conversion rate of the current content in the current layer of the ascending channel, as the promotion score of the heterogeneous content.

[0028] In one implementation, the upper-layer distribution module is configured to: distribute the individual content and the content collection at the upper layer of the ascending channel if the cross-modal promotion result between the individual content and the content collection is successful; and distribute the content in the form of content at the middle layer of the ascending channel if the cross-modal promotion result between the individual content and the content collection is unsuccessful. This provides an ascending channel and content distribution method that combines the ascending of individual content and content collections. Compared to the ascending channel and distribution method for single content in related technologies, this method saves costs and improves the efficiency of content distribution. Furthermore, in this ascending channel, content is distributed at the lower layer of the ascending channel based on the business scenario, in the middle layer of the ascending channel, cross-modal promotion is performed on the content, and at the upper layer of the ascending channel, the content distribution form is determined based on the cross-modal promotion result. This better achieves the business goals of users, creators, and the platform. Attached Figure Description

[0029] The above and other objects, features, and advantages of this disclosure will become readily apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. Several embodiments of this disclosure are illustrated in the drawings by way of example and not limitation, in which:

[0030] Figure 1 This diagram illustrates a heterogeneous content joint distribution process architecture according to an embodiment of the present disclosure.

[0031] Figure 2 A flowchart illustrating a heterogeneous content joint distribution method according to an embodiment of this disclosure is shown;

[0032] Figure 3 This diagram illustrates a heterogeneous graph in a heterogeneous content joint distribution method according to an embodiment of the present disclosure.

[0033] Figure 4 This diagram illustrates a flowchart of determining similarity in a heterogeneous content joint distribution method according to an embodiment of the present disclosure;

[0034] Figure 5 This diagram illustrates a flowchart of determining behavioral similarity in a heterogeneous content joint distribution method according to an embodiment of the present disclosure;

[0035] Figure 6 This diagram illustrates a flowchart of determining cross-modal promotion in a heterogeneous content joint distribution method according to an embodiment of the present disclosure;

[0036] Figure 7 This diagram illustrates a flowchart of determining the upper-layer distribution strategy of the ascending channel in a heterogeneous content joint distribution method according to an embodiment of this disclosure;

[0037] Figure 8 This diagram illustrates the structure of a heterogeneous content joint distribution device according to an embodiment of the present disclosure.

[0038] Figure 9 A schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure is shown.

[0039] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0040] The principles and spirit of this disclosure will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0041] Those skilled in the art will recognize that embodiments of this disclosure can be implemented as a system, apparatus, device, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0042] According to embodiments of this disclosure, a method for joint distribution of heterogeneous content, an apparatus for joint distribution of heterogeneous content, a computer-readable storage medium, and an electronic device are provided.

[0043] In this document, any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0044] The principles and spirit of this disclosure are explained in detail below with reference to several representative embodiments. Invention Overview

[0046] The "Ascension Channel" technology was initially designed to give new content a cold start on the platform, selecting and promoting it through multiple layers until the most popular content is promoted across the entire network, achieving the goals of both users and the platform. Its main processes include: machine review, manual review, recommendations at each level of the channel, and subsequent promotion. These layers work together to screen and ensure new content progresses from a cold start to its final promotion across the entire network.

[0047] In related technologies, upward channels primarily target a single content type (e.g., video, song), and their distribution units are typically of the same content type. For business scenarios with diverse content formats, such as podcasts, there is both the distribution of individual content (a single "audio") and the distribution of content collections (a "playlist"). Since different types of content serve different business objectives, the traffic of any single type of content is insufficient to achieve complete content upward movement and meet overall business needs independently. Therefore, upward channels need to involve not only the upward movement of individual content but also the upward movement of the content collection to which that individual content belongs—that is, the joint upward movement of heterogeneous content. For this type of channel involving the joint upward movement of individual content and content collections, the following technical challenges exist that are not faced or resolved in single-type upward channels:

[0048] (1) How to organize different types of heterogeneous content in the ascending channel? Is it that each layer distributes two types of content, or that some layers distribute one type of content and other layers distribute the other type of content?

[0049] (2) In the heterogeneous upward channel, there are distribution forms for individual content and distribution forms for content collections. Users also have different types of interactions with different forms of new content. For example, collections are mainly for content collections, while playback is mainly for individual content. There is also a relationship where an individual content is a content unit within a content collection. Based on these behaviors, the conventional recall methods for new content distribution in the upward channel are mainly new single content → content collection → similar content collection → single content → users who have played it, new single content → content collection → similar content collection → users who have collected it, and content collection → similar content collection → single content in the collection → users who have played it. However, each modal transition results in the loss of multiple pieces of information. All individual content in a content collection is not necessarily strongly correlated, and there is a correlation problem between a single content and its content collection. This situation is particularly unfavorable for the distribution of new content, which has limited related information.

[0050] (3) How should the promotion path and rules of heterogeneous content be involved and judged? What points need to be considered when promoting from one type to another? If only the conversion efficiency of the current layer is considered, there are many problems. Distributing content in the form of a single piece of content is effective, but the distribution effect of its content collection may not be good.

[0051] In view of the above, this disclosure provides a method, apparatus, computer-readable storage medium, and electronic device for the joint distribution of heterogeneous content. The heterogeneous content joint distribution system (such as a server or client) distributes a first piece of heterogeneous content at the lower layer of the ascending channel to obtain user interaction data with the first piece of content, and recalls target users from the users based on the interaction data. The heterogeneous content includes a single piece of content and a content collection to which the single piece of content belongs; the first piece of content is either the single piece of content or a content collection to which the single piece of content belongs; at the middle layer of the ascending channel, a heterogeneous graph is constructed based on the single piece of content, the target user, and the content collection, and... The similarity between the individual content and the content collection is determined based on the heterogeneity graph; based on the similarity between the individual content and the content collection, it is determined whether the individual content and the content collection can successfully achieve cross-modal promotion; based on the cross-modal promotion result between the individual content and the content collection, the distribution strategy of the middle layer of the ascending channel is determined; wherein, the distribution strategy of the middle layer of the ascending channel includes single content distribution or content collection distribution; at the upper layer of the ascending channel, the distribution strategy of the upper layer of the ascending channel is determined based on the distribution strategies of the lower layer of the ascending channel and the middle layer of the ascending channel; wherein, the distribution strategy of the upper layer of the ascending channel includes single content distribution and / or content collection distribution. Thus, a method for promoting both individual content and content collections is provided, which saves costs and improves the efficiency of content distribution compared to the promotion channels and distribution methods for single content in related technologies. In addition, in this promotion channel, content is distributed in the lower layer of the promotion channel based on the business scenario, in a form that accumulates better interaction data. In the middle layer of the promotion channel, content is promoted across modalities. In the upper layer of the promotion channel, the content distribution form is determined based on the cross-modal promotion results, which can better achieve the business goals of users, creators and the platform.

[0052] The following is an explanation of the terms used in this article:

[0053] Single content: This refers to the content itself. When distributing, the title, image, etc. of the content itself are displayed directly, making it easy for users to understand the content and facilitating consumption conversion.

[0054] Content collections are lists of multiple items that share common themes or authors. Examples include playlists in podcasts and song lists in music services. This allows users to see other related content while consuming a single item, and this distribution format is beneficial for users to save and revisit the content.

[0055] Podcast service: This is a rich content service within NetEase Cloud Music that includes long-form audio clips, each containing images, text, and corresponding audio. Distribution includes both playlists (content compilations) and individual audio clips. It's a platform where professionally generated content (PGC) and user-generated content (UGC) coexist; therefore, the initial launch of content, the ecosystem, and creator engagement are crucial to the entire business.

[0056] Ascendancy: This is a distribution framework proposed for the sustainable development of the content ecosystem. Based on algorithmic distribution, a portion of traffic is allocated to select content, and then users are selected through the content. The content is then distributed through a multi-layered screening process. Lower layers have less traffic but more individual content, while higher layers have more traffic but fewer individual content. Each layer needs to advance to the next layer. Through various methods, content that meets the platform's value is selected, thereby realizing an ecosystem framework that integrates platform value such as cold start, anchor support, and hit content.

[0057] Joint Ascension Channel: Designed for business scenarios with diverse distribution units. The ascension channel contains multiple distribution units, and it is necessary to consider information migration under different scenarios in different distribution units and cross-modal promotion of different forms of content.

[0058] Cross-modal promotion refers to promotion from one distribution unit to another in a heterogeneous promotion channel containing multiple distribution units. Because the content modalities before and after promotion are different, it requires more consideration than the promotion of a single type of content to ensure the efficiency of the entire promotion channel.

[0059] Content Ecosystem: A content distribution platform needs to consider not only consumer needs but also the needs of creators and the platform itself. Content creation costs are typically high, and if newly created content doesn't receive distribution and support, creator engagement will decrease, potentially leading to churn. Secondly, without a mechanism for discovering and promoting high-quality content, the platform becomes unsustainable, failing to meet user needs and ultimately causing user churn. The loss of both creators and users also harms the platform's needs. Finally, platforms often require a method to guide the value orientation of their content, all of which are encompassed within the content ecosystem. In conclusion, a sustainable content ecosystem is essential for a successful platform.

[0060] Heterogeneous graph: A graph algorithm model in which the sum of the number of types of nodes and edges is greater than 2. Compared with homogeneous graphs, it distinguishes different nodes and the connection relationships between nodes, so the calculated embedding vectors retain more semantic information.

[0061] After introducing the basic principles of this disclosure, various non-limiting embodiments of this disclosure will be described in detail below.

[0062] Application Scenarios Overview

[0063] It should be noted that the following application scenarios are shown only to facilitate understanding of the spirit and principles of this disclosure, and the implementation of this disclosure is not limited in any way. On the contrary, the implementation of this disclosure can be applied to any applicable scenario.

[0064] This disclosure can be applied to any scenario where a client or server needs to distribute content; for example: the server distributes playlists for an app on the client. In the lower layer of the ascending channel, the content is distributed as a single piece of content; in the middle layer of the ascending channel, after it is determined that the single piece of content has successfully ascended to the content collection across modalities, the content is distributed as a content collection; in the upper layer of the ascending channel, the content is distributed simultaneously in the form of a single piece of content and a content collection, so that both the single piece of content and the content collection to which the single piece of content belongs achieve good distribution results.

[0065] Exemplary methods

[0066] The following is combined Figure 1 The system architecture and application scenarios of the operating environment of this exemplary implementation are described in an exemplary manner.

[0067] Figure 1 A schematic diagram of the system architecture is shown. This system architecture 100 may include a client 110 or a server 120. The client 110 may be a terminal, such as a smartphone, tablet, or personal computer, while the server 120 may be a single server or a cluster of multiple servers. At the lower layer of the ascending channel, the client 110 or server 120 distributes the first content in the heterogeneous content to obtain user interaction data with the first content, and recalls target users from among the users based on the interaction data. At the middle layer of the ascending channel, a heterogeneous graph is constructed based on individual content, target users, and content collections, and the similarity between individual content and content collections is determined based on the heterogeneous graph. Based on the similarity between individual content and content collections, it is determined whether cross-modal promotion between individual content and content collections is successful. Based on the cross-modal promotion results between individual content and content collections, the distribution strategies for the middle and upper layers of the ascending channel are determined.

[0068] Exemplary embodiments of this disclosure first provide a method for the joint distribution of heterogeneous content, which may include:

[0069] In the lower layer of the ascending channel, the first content in the heterogeneous content is distributed to obtain user interaction data with the first content, and target users are recalled from the users based on the interaction data; wherein, the heterogeneous content includes single content and content collections to which a single content belongs; the first content is a single content or a content collection to which a single content belongs;

[0070] In the middle layer of the ascending channel, a heterogeneous graph is constructed based on individual content, target users, and content collections, and the similarity between individual content and content collections is determined based on the heterogeneous graph;

[0071] Based on the similarity between individual content and content collections, determine whether an individual content and content collection can be successfully promoted across modalities;

[0072] Based on the cross-modal promotion results between individual content and content collections, the distribution strategy for the middle layer of the upward channel is determined; wherein, the distribution strategy for the middle layer of the upward channel includes single content distribution or content collection distribution.

[0073] Editing and distribution;

[0074] In the upper layer of the ascending channel, the distribution strategy of the upper layer of the ascending channel is determined based on the distribution strategies of the lower and middle layers of the ascending channel; wherein, the distribution strategy of the upper layer of the ascending channel includes single content distribution and / or content collection distribution.

[0075] Figure 2 An exemplary flow of this heterogeneous content joint distribution method is shown below. Figure 2 Each step in the process will be explained in detail.

[0076] refer to Figure 2 In step S210, in the lower layer of the ascending channel, the first content in the heterogeneous content is distributed to obtain the interaction data between the user and the first content, and the target user is recalled from the users based on the interaction data.

[0077] The lower layer of the ascending channel can be understood as the initial stage of content distribution, and the time can be set according to actual needs, such as three days, five days, or one week; generally, the lower layer of the ascending channel is the bottom layer of the ascending channel.

[0078] Heterogeneous content includes single content and content collections to which a single content belongs; the first content is a single piece of content, or a content collection to which a single piece of content belongs.

[0079] In practice, users' consumption preferences generally differ across different business scenarios. For example, in the podcast business, since the text and title of a single piece of content can directly convey the specific content, consuming a single piece of content is more convenient for users. In the audiobook business, since the different pieces of content within an audiobook are interconnected, consuming a collection of content is more convenient for users.

[0080] For new content, without sufficient interaction data, it's difficult to directly distribute it to the right users. Therefore, the lower layer of the ascending channel primarily aims to help new content be distributed quickly, evaluated, and to accumulate interaction data. Furthermore, whether new content is distributed as individual pieces or as a collection depends on the specific business scenario. For example, in podcasting, individual pieces are more convenient for users to consume; therefore, the lower layer of the ascending channel distributes them as individual pieces. Conversely, in audiobooks, collections are more convenient for users to consume; therefore, the lower layer of the ascending channel distributes them as collections.

[0081] Target users can be understood as users who prefer the primary content; in practice, users who have played, liked, favorited, purchased, subscribed to, or commented on the primary content can be identified as target users based on interaction data.

[0082] Continue to refer to Figure 2 In step S220, in the middle layer of the ascending channel, a heterogeneous graph is constructed based on individual content, target user, and content collection, and the similarity between individual content and content collection is determined based on the heterogeneous graph.

[0083] The lower layer of the ascending channel can be understood as the middle stage of content distribution, and the time can be set according to actual needs, such as three days, five days, or one week. Generally, the middle layer of the ascending channel is the layer between the bottom layer and the top layer of the ascending channel.

[0084] The efficiency of the upward channel mainly includes two aspects: first, the efficient distribution of content within each layer; and second, the promotion of content. Compared to the traditional upward channel with only one distribution unit, the combined distribution upward channel has a richer variety of behavior types and corresponding content units. For example, in the podcast business, there are playback, commenting, and liking of audio, as well as collection and clicking of playlists, and there is also the inclusion relationship of audio belonging to a certain playlist; in the audiobook business, there are collection and clicking of a certain audiobook, as well as playback, commenting, and liking of a certain chapter of a certain audiobook.

[0085] Compared to finding a content collection to which a single piece of content belongs, and then finding similar content collections within that collection, or vice versa, it's best to further associate it with corresponding user behavior. Furthermore, a heterogeneous graph can be constructed based on users, single pieces of content, content collections, and the interactions between users and single pieces of content or content collections. Specifically, for example... Figure 3As shown, a heterogeneous graph is constructed by taking individual content, target user, and content collection as vertices, and the behavior types of target users towards individual content and content collection, and the relationship between individual content and content collection as edges.

[0086] In practice, users, individual content, and content collections can be placed in the same vector space using methods from related technologies to generate corresponding feature vectors. Then, the similarity between individual content and content collections can be calculated using similarity calculation methods such as Jaccard similarity coefficient, cosine similarity, Euclidean distance, Manhattan distance, and Pearson correlation coefficient.

[0087] Continue to refer to Figure 2 In step S230, based on the similarity between a single piece of content and a content collection, it is determined whether a single piece of content and a content collection can be successfully promoted across modalities.

[0088] Content determines the upper limit of distribution. The advancement of each level of content in the upward channel determines the ceiling of the next level of distribution, and also determines the content that will ultimately be pushed to its peak in the upward channel. The upward channel of joint distribution presents a challenge not found in traditional upward channels: cross-modal promotion. The good distribution effect of an individual piece of content does not necessarily mean a good distribution effect for the content collection to which it belongs. This is because the content and quality level of each individual piece of content in the collection may not be the same, and the content may not be completely related.

[0089] When a single piece of content is deemed to have good distribution performance at the lower levels of the upward mobility channel, it's necessary to determine whether the content collection to which that single piece of content belongs is similar to it in all aspects. This ensures that the distribution performance of the content collection to which the single piece of content belongs is also good. Similarly, when a content collection is deemed to have good distribution performance at the lower levels of the upward mobility channel, it's necessary to determine whether the individual pieces of content within that content collection are similar to the content collection in all aspects. This ensures that the distribution performance of the individual pieces of content within that content collection is also good. Therefore, when a single piece of content and a content collection are very similar in all aspects, cross-modal promotion between the single piece of content and the content collection is generally possible.

[0090] In practice, the similarity between an individual piece of content and a collection of content can be determined by the representational similarity, or by the behavioral similarity, or by a combination of representational and behavioral similarity. No specific limit is imposed here.

[0091] Continue to refer to Figure 2 In step S240, the distribution strategy of the middle layer of the ascending channel is determined based on the cross-modal promotion results between individual content and content collection.

[0092] The distribution strategies for the middle layer of the upward channel include single content distribution or content compilation distribution.

[0093] In practice, when a single piece of content is successfully promoted across modalities to a content collection, the content is distributed in the middle of the promotion channel in the form of the promoted content. For example, after a single piece of content is successfully promoted across modalities to a content collection, the content is distributed in the form of a content collection; after a content collection is successfully promoted across modalities to a single piece of content, the content is distributed in the form of a single piece of content.

[0094] Continue to refer to Figure 2 In step S250, the distribution strategy of the upper layer of the rising channel is determined based on the distribution strategies of the lower and middle layers of the rising channel.

[0095] The upper layer of the ascending channel can be understood as the later stage of content distribution, and the time can be set according to actual needs, such as three days, five days, or one week; generally, the upper layer of the ascending channel is the top layer of the ascending channel.

[0096] The distribution strategies at the upper levels of the ascending channel include single content distribution and / or content compilation distribution.

[0097] When a single piece of content or a content collection successfully advances across modalities at the upper or middle levels of the ascending channel, it can be distributed as a single piece of content, a content collection, or a combination of both; there are no restrictions here. If a single piece of content does not advance across modalities with a content collection, the form in which the content is distributed at the lower or middle levels of the ascending channel will determine the form it is distributed at the upper level of the ascending channel.

[0098] In one implementation, the similarity between individual content and a content collection can be determined by combining behavioral similarity with representational similarity; specifically, refer to... Figure 4 The step S220 above, "determining the similarity between individual content and content collection based on the heterogeneous graph," can further include the following steps S410 to S430:

[0099] Step S410: Determine the behavioral similarity between individual content and content collections based on the heterogeneous graph.

[0100] Behavioral similarity is primarily calculated through user interactions with content. Heterogeneous graphs illustrate user interactions with individual pieces of content and content collections; therefore, the behavioral similarity between a single piece of content and its associated content collection can be determined based on the heterogeneous graph.

[0101] In practice, a meta-path can be constructed, and then a random walk can be performed based on the meta-path. The language model can then be trained based on the results of the random walk to obtain a trained language model that places users, individual content, and content collections in the same vector space. Subsequently, the vector of an individual content and the vector of its content collection can be found in this vector space. The similarity between the two can then be calculated using similarity calculation methods to obtain the behavioral similarity between an individual content and its content collection.

[0102] Step S420: Determine the representational similarity between individual content and content collections.

[0103] The representation includes both images and text of the content; therefore, representation similarity can be calculated using the CLIP model and CBDSSM model in related technologies to calculate the representation vector of the content, and then the similarity between individual content and the collection of content can be calculated to obtain the representation similarity.

[0104] In one implementation, reference Figure 5 The above step S410 may further include the following steps S510 to S540:

[0105] Step S510: Construct meta-paths based on the heterogeneous graph.

[0106] In a heterogeneous graph, a meta-path is the shortest path among all paths connecting two vertices. Meta-paths are used to analyze network topology and predict relationships between nodes. Meta-path analysis has wide applications in recommender systems, social network analysis, and bioinformatics. The definition of meta-paths provides important methods and tools for studying network structure and function.

[0107] In practice, metapaths can be used symmetrically, such as VUV, VUPUV, and PUP; or asymmetrically, such as VUP and PUV, without limitation. Here, V represents a sound in the heterogeneous graph, U represents a user in the heterogeneous graph, and P represents a playlist in the heterogeneous graph; VUV indicates which other sounds a user who prefers a particular sound also prefers; VUPUV indicates which other sounds a user who prefers a particular sound and a user who prefers a particular playlist, and which other sounds a similar user prefers; PUP indicates which other playlists a user who prefers a particular playlist also prefers; VUP indicates that a user who prefers a particular sound also prefers a particular playlist; and PUV indicates that a user who prefers a particular playlist also prefers a particular sound.

[0108] Step S520: Perform a random walk on the metapath to obtain the semantic and structural relationships between vertices of different types.

[0109] Random walks can be performed using models such as node2vec, deepwalk, struct2vec, and metapath2vec.

[0110] If node types are ignored during random walks, the results will be biased, with more node types appearing more frequently. Meta-path random walks define a walk type path, such as VUV, VUPUV, PUP, etc., and then walk along this path, meaning the next node only samples node types that meet the requirements; generally, meta-paths are symmetrical.

[0111] In practice, the metapath2vec model can be used for random walks, for example: Figure 3 The heterogeneous graph in the diagram is defined as:

[0112]

[0113] Where Ri represents the relationship between different types of nodes Vi and Vi+1, and its transition probability at step i is defined as:

[0114]

[0115] The first line indicates that there is an edge between two points, and the next point belongs to the next type of node on the predefined metapath; E represents the above heterogeneous graph; t represents the t-th content on the metapath; ρ represents the conditional probability.

[0116] The second line indicates that there is an edge between two points, but the next point does not belong to the next type of node on the predefined metapath;

[0117] The third line indicates that there is no edge between the two points.

[0118] Step S530: Train the language model to be trained based on the results of the random walk to obtain the trained language model.

[0119] The language model to be trained can be a statistical language model or a neural network language model, such as: NNLM model, RNNLM model, Word2Vec model, GloVe model, ELMo model, GPT model, BERT model, GPT-2 model, ERNIE model, GPT-3 model, etc., without any limitation here.

[0120] In practice, the Skip-Gram model can be used. The Skip-Gram model is also called the continuous skipping word model. It comes from the Word2Vec word embedding algorithm in the field of NLP (Natural Language Processing). This algorithm can embed words into a vector space and make words with similar meanings closer together in the vector space.

[0121] Specifically, a multi-methpath2vec++ model structure can be adopted, setting multiple meta-paths and specifying a set of multinomial distributions for each type of domain in the output layer of the Skip-Gram model, ultimately yielding the following loss function:

[0122]

[0123] Where σ represents the probability function of the existence of the context node ct given the known node V; Given node V, this represents the probability that the context node ct exists, i.e., a positive sample. P(u) represents the probability that the context is not the sampled node ut, given that m negative samples are sampled and the node is v; P(u) is the predefined distribution of the samples in the negative sampling.

[0124] Step S540: Determine the feature vectors of individual content and their respective content collections based on the trained language model, and calculate the similarity between the two to obtain the behavioral similarity between them.

[0125] The trained language model obtained in step S530 can map each vertex in the heterogeneous graph to the same vector space. Therefore, the feature vector of each vertex in the heterogeneous graph can be determined by the trained language model. Then, the similarity between different vertices in the heterogeneous graph can be determined by similarity calculation.

[0126] In one implementation, the success of cross-modal promotion can be determined by calculating a promotion score; specifically, refer to... Figure 6 The above step S230 may further include the following steps S610 and S620:

[0127] Step S610: Determine the promotion score of heterogeneous content based on the similarity between individual content and content collection.

[0128] The formula for calculating the promotion score can be determined based on experience. For example, the product of behavioral similarity and representational similarity can be used as the promotion score, or the product of behavioral similarity, representational similarity, and the conversion rate of the current content in the current layer can be used as the promotion score. No specific limit is set here.

[0129] Step S620: Based on the promotion scores of heterogeneous content, determine whether individual content and content collections can be promoted across modalities.

[0130] Among these, a threshold can be set based on experience. If the promotion score is greater than the threshold, the cross-modal promotion is successful; otherwise, the cross-modal promotion fails.

[0131] In one implementation, to more comprehensively consider cross-modal promotion, the conversion rate of the current content at the current layer in the ascending channel can be incorporated; specifically, step S610 above may further include the following steps:

[0132] The promotion score for heterogeneous content is calculated by multiplying the behavioral similarity and representational similarity between individual content and content collections, as well as the conversion rate of the current content at the current level in the upward channel.

[0133] The promotion score is calculated as S1 * S2 * R, where R represents the current content's business metrics in the promotion channel, such as click-through rate, play rate, and exposure duration; S1 represents behavioral similarity; and S2 represents representational similarity. The content preceding the current content's cross-modal promotion can be a single piece of content or a collection of content, depending on the specific business scenario.

[0134] In one implementation, to improve distribution performance, the distribution strategy of the upper layer of the ascending channel is determined based on the cross-modal promotion results; specifically, refer to... Figure 7 The above step S250 may further include the following steps S710 and S720:

[0135] Step S710: In the upper layer of the ascending channel, if the cross-modal promotion result between a single piece of content and a content collection is a successful promotion, distribute the single piece of content and the content collection.

[0136] Successful cross-modal promotion indicates that individual content and content collections are quite similar in all aspects. Therefore, distributing both individual content and content collections simultaneously in the later stages of the promotion channel will yield better results.

[0137] Step S720: If the cross-modal promotion result between a single piece of content and a content collection is a promotion failure, distribute the content in the form of content in the middle layer of the ascending channel.

[0138] In this case, a failure to advance across modalities indicates that the individual content and the content collection are not very similar in many aspects. Therefore, in the later stages of the advancement channel, the content is distributed in its original form. For example, if an individual content fails to advance across modalities to a content collection, it is distributed as an individual content; if a content collection fails to advance across modalities to an individual content, it is distributed as a content collection. In other words, in this case, the advancement channel is the same as the traditional advancement channel, and the content is distributed in the form of a single content. The content distribution forms for the lower, middle, and upper levels of the advancement channel are the same.

[0139] Exemplary device

[0140] Having introduced the heterogeneous content joint distribution method according to exemplary embodiments of this disclosure, the following refers to... Figure 8 An apparatus for jointly distributing heterogeneous content according to an exemplary embodiment of this disclosure will be described.

[0141] refer to Figure 8 As shown, the heterogeneous content joint distribution device 800 includes:

[0142] The lower-level distribution module 810 is configured to distribute the first content in the heterogeneous content at the lower level of the ascending channel to obtain the interaction data between the user and the first content, and to recall the target user from the users based on the interaction data; wherein, the heterogeneous content includes a single content and a content collection to which a single content belongs; the first content is a single content or a content collection to which a single content belongs.

[0143] The mid-layer distribution module 820 is configured to construct a heterogeneous graph based on individual content, target users, and content collections in the middle layer of the ascending channel, and determine the similarity between individual content and content collections based on the heterogeneous graph; determine whether individual content and content collections can successfully advance across modalities based on the similarity between them; and determine the distribution strategy for the middle layer of the ascending channel based on the cross-modal advancement results between individual content and content collections; wherein, the distribution strategy for the middle layer of the ascending channel includes single content distribution or content collection distribution.

[0144] The upper-layer distribution module 830 is configured to determine the distribution strategy of the upper layer of the ascending channel based on the distribution strategies of the lower and middle layers of the ascending channel; wherein, the distribution strategy of the upper layer of the ascending channel includes single content distribution and / or content collection distribution.

[0145] In one implementation, the mid-layer distribution module 820 is configured to construct a heterogeneous graph by using individual content, target users, and content collections as vertices, and the behavior types of target users towards individual content and content collections, and the association between individual content and content collections as edges.

[0146] In one implementation, the mid-layer distribution module 820 is configured to: determine the behavioral similarity between individual content and content collections based on the heterogeneous graph; and determine the representational similarity between individual content and content collections.

[0147] In one implementation, the mid-layer distribution module 820 is configured to: construct meta-paths based on the heterogeneous graph; perform random walks on the meta-paths to obtain semantic and structural relationships between vertices of different types; train a language model to be trained based on the results of the random walks to obtain a trained language model; determine the feature vectors of a single piece of content and its associated content collection based on the trained language model, and calculate the similarity between the two to obtain the behavioral similarity between them.

[0148] In one implementation, the mid-level distribution module 820 is configured to: determine the promotion score of heterogeneous content based on the similarity between individual content and content collection; and determine whether individual content and content collection can be promoted across modalities based on the promotion score of heterogeneous content.

[0149] In one implementation, the mid-level distribution module 820 is configured to use the product of the behavioral similarity and representational similarity of a single piece of content with a content collection, and the conversion rate of the current content in the current layer of the ascending channel, as the promotion score of the heterogeneous content.

[0150] In one implementation, the upper-layer distribution module 830 is configured to: in the upper layer of the ascending channel, if the cross-modal promotion result between a single piece of content and a content collection is a successful promotion, distribute the single piece of content and the content collection; if the cross-modal promotion result between a single piece of content and the content collection is a failed promotion, distribute the content in the form of content in the middle layer of the ascending channel.

[0151] Exemplary storage media

[0152] The storage medium of the exemplary embodiments of this disclosure will now be described.

[0153] In this exemplary embodiment, the above method can be implemented by a program product, such as a portable compact disc read-only memory (CD-ROM) containing program code, which can run on a device, such as a personal computer. However, the program product disclosed herein is not limited thereto. In this document, a readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0154] The program product may employ any combination of one or more readable media. A readable media may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0155] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0156] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RE, etc., or any suitable combination thereof.

[0157] Program code for performing the operations of this disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0158] Exemplary electronic devices

[0159] refer to Figure 9 An electronic device according to an exemplary embodiment of the present disclosure will be described.

[0160] Figure 9 The electronic device 900 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0161] like Figure 9 As shown, the electronic device 900 is presented in the form of a general-purpose computing device. The components of the electronic device 900 may include, but are not limited to: at least one processing unit 910, at least one storage unit 920, a bus 930 connecting different system components (including the storage unit 920 and the processing unit 910), and a display unit 940.

[0162] The storage unit stores program code, which can be executed by the processing unit 910 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure. For example, the processing unit 910 can perform actions such as... Figure 1 The methods and steps shown are as follows.

[0163] Storage unit 920 may include volatile storage units, such as random access memory (RAM) 921 and / or cache memory 922, and may further include read-only memory (ROM) 923.

[0164] The storage unit 920 may also include a program / utility 924 having a set (at least one) of program modules 925, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0165] Bus 930 may include a data bus, an address bus, and a control bus.

[0166] Electronic device 900 can also communicate with one or more external devices 2000 (e.g., keyboard, pointing device, Bluetooth device, etc.) via input / output (I / O) interface 950. Electronic device 900 also includes a display unit 940 connected to input / output (I / O) interface 950 for display purposes. Furthermore, electronic device 900 can communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 960. As shown, network adapter 960 communicates with other modules of electronic device 900 via bus 930. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0167] It should be noted that although several modules or sub-modules of the apparatus have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0168] Furthermore, although the operations of the methods disclosed herein are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0169] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for joint distribution of heterogeneous content, characterized in that, The method includes: In the lower layer of the ascending channel, the first content in the heterogeneous content is distributed to obtain user interaction data with the first content, and target users are recalled from the users based on the interaction data; wherein, the heterogeneous content includes a single content and a content collection to which the single content belongs; the first content is the single content or the content collection to which the single content belongs; In the middle layer of the ascending channel, a heterogeneous graph is constructed based on the individual content, the target user, and the content collection, and the similarity between the individual content and the content collection is determined based on the heterogeneous graph; Based on the similarity between the individual content and the content collection, determine whether the individual content and the content collection can be successfully promoted across modalities; Based on the cross-modal promotion results between the individual content and the content collection, the distribution strategy of the middle layer of the ascending channel is determined; wherein, the distribution strategy of the middle layer of the ascending channel includes single content distribution or content collection distribution; In the upper layer of the ascending channel, the distribution strategy of the upper layer of the ascending channel is determined according to the distribution strategies of the lower layer and the middle layer of the ascending channel; wherein, the distribution strategy of the upper layer of the ascending channel includes single content distribution and / or content compilation distribution; The step of determining the similarity between the individual content and the content collection based on the heterogeneous graph includes: Meta-paths are constructed based on the heterogeneous graph; random walks are performed on the meta-paths to obtain semantic and structural relationships between vertices of different types; a language model to be trained is trained based on the results of the random walks to obtain a trained language model; feature vectors of the individual content and its associated content collection are determined based on the trained language model, and the similarity between the two is calculated to obtain the behavioral similarity between them. Determine the representational similarity between the individual content and the content collection.

2. The method according to claim 1, characterized in that, The construction of a heterogeneous graph based on the individual content, the target user, and the content collection includes: The heterogeneous graph is constructed by taking the individual content, the target user, and the content collection as vertices, and taking the target user's behavior type towards the individual content and the content collection, and the association relationship between the individual content and the content collection as edges.

3. The method according to claim 1, characterized in that, The step of determining whether a single piece of content can successfully advance across modalities with the content collection based on the similarity between the individual content and the content collection includes: The promotion score of heterogeneous content is determined based on the similarity between the individual content and the content collection; Based on the promotion scores of the heterogeneous content, determine whether the individual content and the content collection can be promoted across modalities.

4. The method according to claim 3, characterized in that, The step of determining the promotion score of heterogeneous content based on the similarity between the individual content and the content collection includes: The product of the behavioral similarity and representational similarity between the individual content and the content collection, and the conversion rate of the current content in the current layer of the ascending channel, is used as the promotion score of the heterogeneous content.

5. The method according to claim 1, characterized in that, The step of determining the distribution strategy of the upper layer of the ascending channel based on the distribution strategies of the lower and middle layers of the ascending channel includes: In the upper layer of the ascending channel, if the cross-modal promotion result between the individual content and the content collection is a successful promotion, the individual content and the content collection are distributed. In the event that the cross-modal promotion result between the individual content and the content collection is a promotion failure, the content is distributed as content in the middle layer of the ascending channel.

6. A device for jointly distributing heterogeneous content, characterized in that, The device includes: The lower-level distribution module is configured to distribute the first content in the heterogeneous content at the lower level of the ascending channel to obtain user interaction data with the first content, and recall target users from the users based on the interaction data; wherein, the heterogeneous content includes a single content and a content collection to which the single content belongs; the first content is the single content or the content collection to which the single content belongs; The mid-layer distribution module is configured to construct a heterogeneous graph based on the individual content, the target user, and the content collection in the middle layer of the ascending channel; determine the similarity between the individual content and the content collection based on the heterogeneous graph; determine whether the individual content and the content collection can successfully advance across modalities based on the similarity between them; and determine the distribution strategy for the middle layer of the ascending channel based on the cross-modal advancement result between the individual content and the content collection. The distribution strategy for the middle layer of the ascending channel includes single-content distribution or content collection distribution. The upper-layer distribution module is configured to determine the distribution strategy of the upper layer of the rising channel based on the distribution strategies of the lower layer and the middle layer of the rising channel; wherein, the distribution strategy of the upper layer of the rising channel includes single content distribution and / or content compilation distribution; The middle-layer distribution module is configured as follows: Meta-paths are constructed based on the heterogeneous graph; random walks are performed on the meta-paths to obtain semantic and structural relationships between vertices of different types; a language model to be trained is trained based on the results of the random walks to obtain a trained language model; feature vectors of the individual content and its associated content collection are determined based on the trained language model, and the similarity between the two is calculated to obtain the behavioral similarity between them. Determine the representational similarity between the individual content and the content collection.

7. The apparatus according to claim 6, characterized in that, The middle-layer distribution module is configured as follows: The heterogeneous graph is constructed by taking the individual content, the target user, and the content collection as vertices, and taking the target user's behavior type towards the individual content and the content collection, and the association relationship between the individual content and the content collection as edges.

8. The apparatus according to claim 6, characterized in that, The middle-layer distribution module is configured as follows: The promotion score of heterogeneous content is determined based on the similarity between the individual content and the content collection; Based on the promotion scores of the heterogeneous content, determine whether the individual content and the content collection can be promoted across modalities.

9. The apparatus according to claim 8, characterized in that, The middle-layer distribution module is configured as follows: The product of the behavioral similarity and representational similarity between the individual content and the content collection, and the conversion rate of the current content in the current layer of the ascending channel, is used as the promotion score of the heterogeneous content.

10. The apparatus according to claim 6, characterized in that, The upper-layer distribution module is configured as follows: In the upper layer of the ascending channel, if the cross-modal promotion result between the individual content and the content collection is a successful promotion, the individual content and the content collection are distributed. In the event that the cross-modal promotion result between the individual content and the content collection is a promotion failure, the content is distributed as content in the middle layer of the ascending channel.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 5.

12. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the method of any one of claims 1 to 5 by executing the executable instructions.

Images