Social media cocoon reduction method based on heterogeneous information network

Abstract

Embodiments of the present application disclose a kind of social media information cocoon house reduction based on heterogeneous information network;Multiple users in the historical interaction data of social media platform can be acquired in the embodiments of the present application;For each user, corresponding heterogeneous information network is constructed based on its historical interaction data;For any target user, according to the social group node in its heterogeneous information network, the social group node in the heterogeneous information network of other users is matched, and the other users existing at least one common social group are determined as the associated user of target user;Based on the difference between the topic distribution of the heterogeneous information network of target user and the heterogeneous information network of associated user, potential interest content of target user is filtered out from the media content corresponding to the difference topic label;Potential interest content is recommended to target user. Thus, the information cocoon effect is effectively alleviated in the embodiments of the present application, and the information diversity of recommendation result and the effectiveness of user interest expansion are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a method for reducing social media information cocoons based on heterogeneous information networks. Background Technology

[0002] With the rapid development of social media platforms, personalized recommendation algorithms are widely used in content distribution to improve user experience and platform stickiness. Current mainstream recommendation systems typically analyze the relationship between users and media content based on their historical interaction behavior (such as clicks, likes, comments, or content dwell time), pushing content similar to their past preferences.

[0003] However, such mechanisms can easily lead to the "information cocoon" effect in practice. Because the recommendation results are concentrated in a narrow area where users have already shown interest, users are constantly exposed to a homogenized information environment, making it difficult for them to access diverse perspectives or potential new areas of interest. This not only limits the expansion of users' cognitive breadth but also weakens the information diversity value of the platform. Summary of the Invention

[0004] This application provides a method for reducing social media information cocoons based on heterogeneous information networks. It not only integrates users' own interactive behavior and social topology, but also actively discovers users' blind spots of interest through cross-circle topic difference analysis, effectively alleviating the information cocoon effect. This not only improves the information diversity of recommendation results and the effectiveness of expanding users' interests, but also helps to enhance the overall content ecosystem diversity and information dissemination breadth of the platform.

[0005] This application provides a method for reducing social media information cocoons based on heterogeneous information networks. The method includes:

[0006] Acquire historical interaction data of multiple users on social media platforms. The historical interaction data includes the corresponding user's historical media interaction content, the topic tags to which the historical media interaction content belongs, and the communities that the corresponding user participates in.

[0007] For each user, a corresponding heterogeneous information network is constructed based on their historical interaction data. The heterogeneous information network includes multiple types of nodes and edges between nodes. The multiple types of nodes include user nodes, media content nodes, topic nodes, and community nodes. The edges between nodes represent the relationships between the corresponding nodes.

[0008] For any target user, based on the community nodes in its heterogeneous information network, it is matched with the community nodes in the heterogeneous information networks of other users, and other users who have at least one common community are identified as the target user's associated users.

[0009] Based on the differences in topic distribution between the heterogeneous information networks of target users and the heterogeneous information networks of associated users, potential interest content of target users is selected from the media content corresponding to the different topic tags. The different topic tags are different from the topic tags of the target user's historical interactive media content.

[0010] Recommend content of potential interest to target users.

[0011] This application also provides an electronic device, including a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to execute steps in any of the social media information cocoon reduction methods based on heterogeneous information networks provided in this application.

[0012] This application also provides a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute steps in any of the social media information cocoon reduction methods based on heterogeneous information networks provided in this application.

[0013] This application also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps in any of the social media information cocoon reduction methods based on heterogeneous information networks provided in this application.

[0014] This application acquires historical interaction data of multiple users on social media platforms. Each historical interaction data includes the user's historical media content, the topic tags associated with that content, and the communities the user participates in. Based on this historical interaction data, a heterogeneous information network can be constructed. This network includes multiple types of nodes and edges between them, such as user nodes, media content nodes, topic nodes, and community nodes. The edges represent the relationships between these nodes, thus comprehensively and structurally depicting the user's social behavior patterns and interest semantics, accurately reflecting their long-term interaction habits. Furthermore, for any target user, by comparing their community nodes in the heterogeneous information network with those of other users, users sharing at least one common community are identified as related users. This effectively utilizes the social circle structure to mine potential groups with similar interests. Further, by analyzing the differences in topic distribution between the target user and their related users in their respective heterogeneous information networks, frequently interacted topic tags that the related users have not yet encountered are identified. Candidate content is then extracted from the media content corresponding to these different topic tags as potential interest content for the target user. Ultimately, content of this potential interest is recommended to the target users. Thus, this solution not only integrates users' own interactive behaviors and social topology, but also proactively discovers their interest blind spots through cross-circle topic difference analysis. This effectively expands from "known preferences" to "unknown but relevant interests," effectively mitigating the information cocoon effect. It not only improves the information diversity of recommendation results and the effectiveness of expanding user interests, but also helps enhance the overall content ecosystem diversity and information dissemination breadth of the platform. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1a A flowchart illustrating the method for reducing social media information cocoons based on heterogeneous information networks provided in this application embodiment;

[0017] Figure 1b A schematic diagram of a heterogeneous information network provided in the application embodiments;

[0018] Figure 2 This is a schematic diagram of the structure of the social media information cocoon reduction device based on heterogeneous information networks provided in the embodiments of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] This application provides a method for reducing social media information cocoons based on heterogeneous information networks.

[0021] The social media information cocoon reduction method based on heterogeneous information networks can be executed in a social media information cocoon reduction device based on heterogeneous information networks. Specifically, this device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet, smart Bluetooth device, laptop, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0022] In some embodiments, the social media information cocoon reduction device based on heterogeneous information networks can also be integrated into multiple electronic devices. For example, the social media information cocoon reduction device based on heterogeneous information networks can be integrated into multiple servers, and the social media information cocoon reduction method based on heterogeneous information networks of this application can be implemented by multiple servers.

[0023] In some embodiments, the server may also be implemented as a terminal.

[0024] It is understood that in the specific implementation of this application, data such as users' historical interaction data, interaction behavior data, and user profile information are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0025] The following sections provide detailed descriptions of each example. It should be noted that the sequence numbers of the following embodiments are not intended to limit the preferred order of the embodiments.

[0026] In this embodiment, a method for reducing social media information cocoons based on heterogeneous information networks is provided, such as... Figure 1a The specific process of this method for reducing social media information cocoons based on heterogeneous information networks can be as follows:

[0027] 101. Obtain historical interaction data of multiple users on social media platforms. The historical interaction data includes each user's historical media interaction content, the topic tags to which the historical media interaction content belongs, and the communities that each user participates in.

[0028] Users refer to individuals or entities that register on social media platforms and have independent identities, who can publish, browse, like, comment, forward, or otherwise interact with media content on the platform, and can join or participate in one or more communities.

[0029] Social media platforms refer to online service platforms that support user-generated content, social interaction, and information dissemination. Their core functions include media content publishing, the use of hashtags, the establishment of user follow relationships, and the organization of communities (such as groups, circles, and super topics).

[0030] Historical interaction data refers to all user activity records related to media content or communities on social media platforms within a historical time period, including but not limited to structured or semi-structured data formed by behaviors such as browsing content, liking, commenting, forwarding, collecting, posting, and joining communities.

[0031] Historical interactive media content refers to the specific media content involved in a user's historical interaction data, including text, images, videos, audio, or combinations thereof, such as a text and image post published by a user, a short video watched, or a long article commented on.

[0032] Hashtags are used to label media content by topic, facilitating content aggregation and retrieval. In this application, hashtags serve as semantic metadata for media content, reflecting the topic or discussion area to which the content belongs. For example, hashtags could specifically be "#Artificial Intelligence" or "#TravelDiary".

[0033] A community refers to a group of users with shared interests, identities, goals, or social relationships on a social media platform. It typically has clear boundaries and membership affiliations. Users can join a community voluntarily or be invited to do so and interact with content within it.

[0034] 102. For each user, a corresponding heterogeneous information network is constructed based on their historical interaction data. The heterogeneous information network includes multiple types of nodes and edges between nodes. The multiple types of nodes include user nodes, media content nodes, topic nodes and community nodes. The edges between nodes represent the corresponding relationships between nodes.

[0035] Among them, heterogeneous information networks are user-centric, integrating multi-source heterogeneous information involved in their historical interactions, and constructing a structured and semantically rich user behavior knowledge graph through different types of nodes (such as users, media content, topics, and communities) and their interconnections.

[0036] Types are used to distinguish entities of different natures, such as "users", "media content", "topics" and "communities" as four different node types.

[0037] A node is a basic unit in a heterogeneous information network, representing a specific entity. In this application, nodes include, but are not limited to, user nodes corresponding to a user on a social media platform, media content nodes corresponding to a piece of historical interactive media content, topic nodes corresponding to a hashtag, community nodes corresponding to a community a user participates in, and so on.

[0038] An edge between nodes refers to a directed or undirected link connecting two nodes, used to indicate a relationship between the two entities. For example, ... Figure 1b As shown, there is an edge between a user node and a media content node, indicating that the user has interacted with the content (such as posting, liking, commenting, etc.). There is an edge between a media content node and a topic node, indicating that the content has been tagged with the corresponding topic. There is an edge between a user node and a community node, indicating that the user has joined the community. There is an edge between a topic node and a community node, indicating that there are users in the community who have posted or interacted with media content belonging to the topic tag. That is, the topic has been actually discussed or disseminated in the community, reflecting the community's thematic tendency or interest focus.

[0039] The relationships between nodes refer to the actual connections between the entities represented by each node on a social media platform caused by user interaction.

[0040] In some embodiments, directed or undirected edges are established between different types of nodes to represent their relationships, based on the user's historical interactive media content, the topic tags to which the historical interactive media content belongs, and the real interactive semantics in the communities in which each user participates.

[0041] In some embodiments, by introducing a community stability assessment mechanism, this application effectively excludes users whose interests drift or whose social relationships are noisy due to frequent community switching, thereby improving the rationality of target user selection. Consequently, related users identified through subsequent matching based on shared communities are more likely to be highly homogeneous with the target user in terms of interests or values, thus enhancing the accuracy of topic difference analysis and the relevance of potential interest content recommendations. This design not only improves the robustness of the recommendation system but also further ensures the technical feasibility of extracting "effective interest expansion" from a "stable social structure."

[0042] Before matching any target user with community nodes in their heterogeneous information network and with community nodes in other users' heterogeneous information networks, and identifying other users who share at least one common community as associated users of the target user, the process also includes:

[0043] Retrieve the community change records of each user's participating communities within a preset time window;

[0044] Users whose community change records are empty are identified as target users.

[0045] The preset time window is a continuous time interval (such as the last 30 days, 90 days, or the active period since user registration) used to limit the scope of historical interaction data and community status collection. This time window can be dynamically configured according to platform policies or application scenarios, aiming to focus on the stability of users' recent social behavior.

[0046] Community change logs refer to the records of changes in a user's community membership within a preset time window, including joining a new community, leaving an existing community, being removed from a community, or having a community disbanded. If a user does not join or leave any communities within the specified time window, their community change log will be empty.

[0047] For example, suppose there are three users on a social media platform: User A, User B, and User C. The preset time window is the last 30 days. We retrieve each user's community change records (within the last 30 days): User A: Currently participating in communities: #PhotographyEnthusiastsGroup, #XXFoodCircle; has not joined or left any communities in the last 30 days → community change record is empty; User B: Currently participating in communities: #FitnessChallenge, #BookClub; left #TravelSharingGroup 20 days ago, joined #BookClub 15 days ago → community change record is not empty; User C: Currently participating in community: #TechFrontiers; joined #AIDiscussionGroup 5 days ago → community change record is not empty. According to the rule "users with empty community change records are identified as target users," only User A meets the condition and is therefore selected as the target user.

[0048] 103. For any target user, match the community nodes in the heterogeneous information network of other users with the community nodes in the heterogeneous information network of the target user, and identify other users who have at least one common community as the target user's associated users.

[0049] In this context, the target user refers to a specific user selected as the object of analysis in the current processing flow. In this application, the target user is a user selected from multiple users to perform subsequent operations such as related user identification, interest difference analysis, and potential content recommendation.

[0050] Other users refer to users other than the target user among multiple users. In the process of identifying associated users, these other users will be traversed, and their community nodes in heterogeneous information networks will be matched with the community nodes of the target user to filter out associated users who meet the common community conditions.

[0051] A shared community refers to community nodes that exist in the heterogeneous information networks of both the target user and other users and have the same identifier (such as a community identifier or name). In other words, if the target user and another user both participate in the same community (e.g., "#PetCareExchangeGroup"), then that community is a "shared community" between the two.

[0052] Associated users refer to other users who share at least one community with the target user. In this application, associated users are not defined based on follow relationships or direct interaction behavior, but are identified through the structural social signal of shared community affiliation, which is used to characterize potential interest similarity or group homogeneity.

[0053] Matching refers to performing an intersection operation or an equivalence comparison operation on the community node set of the target user and the community node sets of other users to determine whether there are at least one identical community node. This process can perform exact matching based on a unique community identifier (such as a community identifier), or it can support fuzzy matching at the semantic level (such as alias normalization). However, in the preferred embodiment of this application, exact matching is used to ensure reliability.

[0054] It is understandable that all community nodes of the target user are extracted from the target user's heterogeneous information network to form the target user's community set. Secondly, iterate through the heterogeneous information networks of other users on the social media platform and extract the community node set for each user in turn. Next, for each pair ( , Perform set intersection calculation, i.e., determine whether the following conditions are met: If the intersection is not empty, it means that the two users share at least one community node, i.e., they have a common community; then, the "other user" is marked as an associated user of the target user and included in the subsequent analysis. This process can be processed in batches in parallel and is suitable for large-scale user scenarios.

[0055] In some embodiments, by introducing a network embedding representation and community embedding feature matching mechanism, this application overcomes the limitations of traditional methods that rely solely on precise matching of community identifiers. It can identify semantically similar but differently named communities, thereby more comprehensively uncovering potential groups with similar interests. It also includes:

[0056] Graph structure features are extracted from heterogeneous information networks to obtain network embedding representations;

[0057] Extract community embedding features corresponding to each community node from the network embedding representation;

[0058] Based on the community nodes in its heterogeneous information network, it matches the community nodes in the heterogeneous information networks of other users, identifying other users who share at least one common community as associated users of the target user, including:

[0059] The community embedding features corresponding to the target user are respectively matched with the community embedding features corresponding to other users.

[0060] If there exists at least one pair of community embedding features whose similarity is greater than or equal to a preset similarity, then the corresponding other users will be identified as associated users of the target user.

[0061] Network embedding representation refers to the numerical representation obtained by mapping nodes (such as users, media content, topics, and communities) in heterogeneous information networks to a low-dimensional dense vector space through graph neural networks (GNNs), deep random walks, node 2Vec, or other graph embedding algorithms. This representation preserves the original network's topology, node types, and semantic relationships, making nodes with similar structures or semantics closer together in the embedding space.

[0062] Community embedding features refer to vector features extracted from network embedding representations that specifically correspond to a node in a particular community. Each community node has a unique embedding vector in the heterogeneous information network of its users, which can be used to quantify the semantic or structural similarity between communities.

[0063] Similarity is a numerical metric used to measure the semantic or structural closeness between the embedded features of two communities. It is usually calculated using cosine similarity, normalized Euclidean distance, or dot product. The larger the value, the more similar the two communities are in the embedding space.

[0064] Preset similarity refers to a pre-defined threshold used to determine whether two communities are "sufficiently similar" to be considered equivalent or related. This threshold can be configured or adaptively adjusted based on platform data distribution, validation set performance, or business needs.

[0065] Understandably, in some embodiments, step 103 not only relies on precise matching of community identifiers but also introduces an embedding-based semantic matching mechanism. The specific process is as follows: For each user's heterogeneous information network, an end-to-end learning graph embedding model (such as a heterogeneous graph neural network, HGNN) is applied to generate a network embedding representation containing all nodes (users, media content, topics, and communities). From this network embedding representation, community embedding features corresponding to each community node in which the target user participates are extracted to form the target user's community embedding set. Similarly, extract the community embedding sets of other users. Embedded features of each community of the target user. Each community embedding feature with other users Perform pairwise similarity calculations (such as cosine similarity); if at least one pair exists... If the similarity is greater than or equal to the preset similarity, the two communities are considered to be highly related in semantics or function (even if the IDs are different, such as "#FitnessCheckInCamp" and "#SelfDisciplineTrainingCircle"), and then the "other user" is identified as an associated user of the target user.

[0066] For example, if a target user joins "#coffee lovers" and another user joins "#hand-drip coffee exchange", the two community identifiers are different, but their embedding vectors are highly similar in semantic space (similarity = 0.82 > preset similarity 0.75), they are still identified as related users.

[0067] Understandably, this embodiment is applicable to scenarios such as non-standard community naming or the existence of aliases, users scattered across multiple sub-communities with similar themes, a large number of long-tail or emerging communities on the platform, and a lack of a unified tagging system. As a result, the coverage of associated users is wider and more representative, providing a higher quality benchmark for subsequent interest-based recommendations based on differences in topic distribution. This significantly improves the generalization ability, semantic perception ability, and cross-circle discovery ability of the recommendation system, further enhancing the technical effect of alleviating information cocoons.

[0068] 104. Based on the differences in topic distribution between the heterogeneous information network of the target user and the heterogeneous information network of related users, potential interest content of the target user is selected from the media content corresponding to the different topic tags. The different topic tags are different from the topic tags of the target user's historical interactive media content.

[0069] In this context, topic distribution refers to the statistical representation of the topic tags associated with a user's historical interactive media content in the user's heterogeneous information network. It is usually expressed in terms of the frequency, weight, or probability of the topic tags, and is used to reflect the user's attention to or interest in different topics.

[0070] The difference in topic distribution refers to a significant inconsistency in the topic preferences exhibited by the target user and their associated users within their respective heterogeneous information networks. Specifically, certain topic tags have high weight or frequency in the topic distribution of associated users, but are completely absent or have extremely low weight in the topic tag set of the target user's historical interactive media content. This difference is quantified through set comparison, frequency statistics, probability distribution distance (such as KL divergence, JS divergence), or embedding space distance, and is used to identify topics that the target user has not yet encountered but that are generally of interest to their socially similar groups.

[0071] For example, if the target user's topic distribution is concentrated on "#food" and "#travel", while the topic distribution of their associated users frequently includes "#campinggear" and "#outdoorhiking", and the target user has never interacted with content related to the latter two, then it is considered that there is a significant difference in the topic distribution between the two, where "#campinggear" and "#outdoorhiking" are the different topic tags.

[0072] Differential topic tags reflect content themes that related users are interested in but the target users have not yet encountered. They represent blind spots or potential interest directions that can be explored based on social similarity and are a key basis for screening content of potential interest to the target users.

[0073] Potential interest content refers to media content (such as text, images, videos, posts, etc.) associated with differentiated topic tags. Because the topics to which these content belong are semantically close to the target user's current interests but are not yet covered by them, they are considered effective recommendation candidates that may stimulate the user's new interests.

[0074] In some embodiments, by introducing a community semantic difference degree and a filtering mechanism for differentiated communities, the discovery of topic differences is shifted from "all related user topics" to "semantically novel and structurally extended subsets of communities." This effectively filters out redundant communities that highly overlap with the target user's interests or are only semantically similar but have different names (such as "#hand-drip coffee" and "#coffee tasting"), thereby enhancing the novelty and exploratory value of differentiated topic tags. This design not only enhances the diversity of recommended content but also avoids interest drift caused by generalized matching, making potential interest content both "unknown" and "acceptable," significantly improving the accuracy of interest expansion and user acceptance. Based on the differences in topic distribution between the heterogeneous information network of the target user and the heterogeneous information network of related users, potential interest content of the target user is filtered from the media content corresponding to differentiated topic tags, including:

[0075] Calculate the community semantic difference between the community embedding features corresponding to each community node of the target user and the community embedding features corresponding to each community node of the associated user;

[0076] Based on the semantic differences of the community, filter out the communities that the target user has not participated in from the communities that the associated users participate in;

[0077] Based on the topic tags of the target user's historical interactive media content, identify the different topic tags from the topic tags contained in the different communities, and obtain the media content corresponding to the different topic tags.

[0078] Filter out content that may be of interest to the target users from media content corresponding to different topic tags.

[0079] Among them, community semantic dissimilarity refers to the degree of difference between the target user and the associated user at the semantic or functional level, calculated by comparing the community embedding features (such as low-dimensional vectors learned through graph neural networks) of their respective community nodes. This dissimilarity is usually expressed as vector distance (such as Euclidean distance) or the complement of similarity (such as... The semantic dissimilarity is quantified in forms such as (e.g., [missing information]). A higher value indicates greater semantic dissimilarity between two communities; a lower value indicates greater semantic similarity. In this application, the community semantic dissimilarity score is used to identify "semantically novel" communities where associated users participate but the target users are not covered.

[0080] A differential community is one in which related users participate, but the target user does not participate, and the minimum semantic difference between its community embedding features and the target user's total community embedding features is greater than a preset threshold (or is in a high difference range overall). This indicates that the community has a significant semantic deviation from the target user's existing social circle in terms of theme, function, or user composition, and represents a group space outside the target user's social structure that may contain new interest signals.

[0081] Understandably, for each associated user, the community embedding features of each community node they participate in are compared pairwise with the community embedding features of all community nodes of the target user. The semantic similarity (e.g., cosine similarity) between each pair of communities is calculated, and the maximum value is taken as the nearest semantic distance between the associated user's community and the target user's community. The complement or inverse mapping of this distance represents the semantic difference between the communities. If the maximum similarity between a community a associated user participates in and all communities of the target user is lower than a preset similarity threshold (equivalent to the semantic difference being higher than a preset difference threshold), then this community is marked as a differential community—a community that the target user has neither joined nor is semantically distant from their existing interest circle. From the set of topic tags contained in all differential communities, topic tags already covered by the target user's historical interactive media content are removed; the remaining tags are the differential topic tags. These tags originate from active discussion forums "outside the target user's social sphere" and have high interest expansion value. The system retrieves media content associated with the aforementioned differentiated topic tags (such as highly interactive posts within differentiated communities), and sorts or filters it based on metrics such as content quality, timeliness, and cross-domain relevance to the target user's existing interests. Finally, it outputs content of potential interest to the target user for recommendation.

[0082] For example, if the target user is active in the "#coffee tasting" and "#baking DIY" communities, the community embedding features focus on the semantic space of "lifestyle aesthetics". The "#wilderness survival skills" community that the related user participates in has an embedding similarity of less than 0.3 with all its communities (the threshold is set to 0.5), and is identified as a differentiated community. The high-frequency topics in this community, "#water purification equipment" and "#emergency rations", have not been encountered by the target user, so they become differentiated topic tags, and related review videos are selected as potential interest content.

[0083] In some embodiments, by introducing user profile information to personalize the filtering of candidate topic tags, this application effectively avoids including generalized but clearly inconsistent "pseudo-potential interests" in the recommendation process. This mechanism, while maintaining topic novelty, significantly improves the user adaptability and acceptability of differentiated topic tags, thereby enhancing the relevance and click-through rate of potential interest content. Furthermore, this design reduces the risk of negative user feedback due to the recommendation of irrelevant content, further optimizing the platform's content distribution efficiency and user experience. Based on the topic tags belonging to the target user's historical interactive media content, differentiated topic tags are determined from the topic tags included in the differentiated communities, including:

[0084] Obtain user profile information of the target users;

[0085] Based on the topic tags of the target user's historical interactive media content, candidate topic tags are selected from the topic tags contained in the different communities.

[0086] Based on user profile information, candidate topic tags are filtered to obtain filtered differentiated topic tags.

[0087] User profile information refers to a multi-dimensional set of features constructed based on the static attributes and dynamic behaviors of target users on social media platforms, including but not limited to: demographic attributes (such as age, gender, and region), interest preference tags, device usage habits, active time periods, spending power levels, and content interaction tendencies (such as preference for videos or text and images). This information is used to characterize the long-term stable features of users and help determine their potential acceptance or relevance to specific topics.

[0088] Candidate topic tags refer to topic tags that have been initially selected from the topic tags included in the different communities and have not appeared in the target user's historical interactive media content. These candidate topic tags constitute the original candidate pool of potential interest topics and have not yet undergone personalized adaptation evaluation.

[0089] Filtered differential topic tags refer to the topic tags retained after filtering based on relevance or suitability information from candidate topic tags and the target user's profile. For example, if the user profile information shows that the target user is a "woman aged 25-35 in a first-tier city," then obviously mismatched candidate tags such as "#agricultural machinery repair" and "#elderly care" might be filtered out, while tags that better match their profile characteristics, such as "#light outdoor wear" and "#workplace productivity tools," might be retained. These tags have greater personalized recommendation value and are the core basis for generating potential interest content.

[0090] Understandably, the platform retrieves structured user profile data from its user data center, covering dimensions such as basic attributes, long-term interest tags, and behavioral preferences. First, several differentiated communities are identified (i.e., communities the target user has not participated in and whose semantics differ significantly from existing communities). Then, all topic tags that have appeared in these differentiated communities are extracted. Next, the topic tags already present in the target user's historical media interactions are removed, leaving candidate topic tags. These candidate topic tags are then matched with user profile information. For example, keyword rules (such as location and age-sensitive words) are used to exclude obviously irrelevant topic tags. A pre-trained tag-profile relevance model (such as a classifier or embedding similarity) is used to calculate the matching score between each candidate tag and the user profile information. Tags with scores higher than a preset threshold are retained, or the Top-K topic tags are selected based on their scores. The final output is the filtered differentiated topic tags. Based on these filtered differentiated topic tags, corresponding high-quality media content is retrieved and recommended as potential interest content for the target user.

[0091] For example, the target user profile is "28 years old, female, Shanghai, interested in fashion and career growth." Differentiated communities include the hashtags "#campinggear," "#fishingtips," and "#commuteroutfit." "#fishingtips" was filtered out due to its low relevance to the user profile's interests; "#campinggear" and "#commuteroutfit" were retained because they fit the "light outdoor" and "workplace" scenarios, becoming the filtered differentiated hashtags.

[0092] In some embodiments, by constructing multi-hop association paths that integrate shared community interactions, cross-community behaviors of associated users, and content semantics, and by introducing an interpretable attention mechanism to quantify path relevance, this application achieves fine-grained, context-aware filtering of potential interest content. The target user's historical interaction data also includes interaction data between the target user and associated users within the shared community.

[0093] Filter out content of potential interest to the target users from media content corresponding to different topic tags, including:

[0094] For each different topic tag corresponding to media content, based on the different communities to which the media content belongs, the participation behavior of related users in the different communities, and the interaction data of target users and related users in the common community, a connection path from target users to media content is constructed.

[0095] The embedded features of nodes on the associated path are aggregated to obtain the semantic representation vector of the associated path;

[0096] Based on the user embedding features and semantic representation vectors of the target user, the attention score of the associated path is calculated;

[0097] Based on the attention score, media content with an attention score higher than the preset attention score is selected from the media content corresponding to the different topic tags, and these are taken as the potential interest content of the target users.

[0098] The participation behavior of associated users in differentiated communities refers to the interaction records generated by associated users in the differentiated communities they participate in (i.e., communities that the target user has not joined but whose semantics are novel), including but not limited to: posting content, commenting, liking, forwarding, initiating topics, and active days. This behavior reflects their influence and depth of interest in the community and is an important basis for evaluating the credibility of their recommendations.

[0099] Interaction data between target users and related users within a shared community refers to direct or indirect interactions between a target user and a related user who share at least one common community. Examples include commenting on / liking the same media content, tagging each other, and replying to each other's posts within the same community. This data reflects the degree of social intimacy or shared interests between the two users.

[0100] The association path refers to a multi-hop heterogeneous path starting from the target user and proceeding through "common community → related users → different communities → media content". Specifically, it takes the form: target user — (common community) — related users — (participation behavior) — different communities — (containment) — media content (belonging to a different topic tag). This path explicitly models the indirect social-semantic connection between the target user and the media content corresponding to the different topic tag.

[0101] Semantic representation vectors are low-dimensional dense vectors obtained by aggregating the embedded features of nodes on a related path (such as common communities, related users, different communities, and media content corresponding to different topic tags) using methods such as mean, weighted summation, and graph attention aggregation. They are used to characterize the overall semantic meaning of the path.

[0102] Attention score is a numerical metric that measures the relevance of a particular connection path to the target user's interests. It is calculated by the matching degree between the target user's user embedding features and the semantic representation vector of the path. The higher the score, the more likely the media content guided by the path is to match the target user's potential interests.

[0103] The preset attention score refers to a pre-defined threshold (e.g., 0.65) used to filter highly relevant association paths. Only when the attention score of a path corresponding to a certain media content is greater than or equal to this threshold is it considered potential content of interest.

[0104] Understandably, in some embodiments, to improve the accuracy of recommending content of potential interest, the system not only relies on the differences in topic tags, but also constructs a connection path based on the social propagation chain, and introduces an attention mechanism to evaluate the importance of the path. Specifically, for each media content m corresponding to a different topic tag, if its respective different community is... And there are associated users Simultaneously satisfy: Participated (i.e., there is participation behavior) With target users Share at least one common community Construct a related path: "in" refers to the target user. In the community inside, Community There are associated users. Associated users Participating in the community of differences Differentiated communities The discussion focused on the media content *m* corresponding to the different topic tags. Pre-trained embedding features of each node along the association path were obtained, including the node embedding features of the target user. Node embedding features of a common community Node embedding features of associated users Node embedding features of differentiated communities Node embedding features of media content The path semantic representation vector is obtained by using weighted aggregation (such as weighting based on the intensity of participation behavior) or simple mean. More complex graph neural network aggregators can also be used here. Employing bilinear attention or dot product attention, the relevance between the target user and the associated path is calculated, resulting in an attention score. ,in, For learnable weight matrix, For bias terms, The Sigmoid activation function maps the output to the [0,1] interval. If If the value is greater than or equal to τ (where τ is the preset attention score), then the media content m is determined as the potential interest content of the target user.

[0105] The embedding features of each node are obtained by pre-training the graph embedding model of the user's corresponding heterogeneous information network (such as a random walk model based on metapath or a heterogeneous graph neural network), which can effectively capture the structural and semantic characteristics of the node in the social-content joint semantic space.

[0106] In some embodiments, by introducing a clustering mechanism based on topic semantic embedding and an optimal attention selection strategy within clusters, this application effectively solves the recommendation redundancy problem caused by highly relevant content being concentrated in a single sub-topic. This design, while retaining content with high attention scores, forcibly covers multiple semantic sub-directions (e.g., "camping equipment" is subdivided into "tents," "cooking utensils," and "lighting"), significantly improving the topic diversity and information breadth of the recommendation results. Simultaneously, because each recommended content is the most relevant representative within its cluster, relevance is not sacrificed, achieving high-quality interest expansion that is "diverse yet cohesive, novel yet relevant," further enhancing users' willingness to explore and the vitality of the platform's content ecosystem. After selecting media content with attention scores higher than the preset attention score from media content corresponding to different topic tags as potential interest content for the target user, the following steps are also included:

[0107] Based on the semantic embedding vectors of the topic tags corresponding to the potential interest content, the potential interest content is clustered to obtain multiple content clusters;

[0108] The media content with the highest attention score in each content cluster is selected as the representative media content;

[0109] Identify representative media content as potential target interest content;

[0110] Recommend content of potential interest to target users, including:

[0111] Recommend content that aligns with the target user's potential interests.

[0112] The semantic embedding vector of the topic tags corresponding to the potential interest content refers to the vector representation obtained by mapping the different topic tags (such as "#campinggear") to a low-dimensional dense semantic space through a pre-trained language model (such as a neural network language model that maps words to dense vectors (Word2Vec), a bidirectional encoder representation model (BERT), or a platform-developed topic embedding model) to the topic tags to which each potential interest content belongs. This vector can capture the semantic similarity between topics (for example, "#hiking" and "#outdoorexploration" are close in the vector space) and is used to support subsequent semantic clustering.

[0113] Content clusters refer to several groups formed by clustering all potential interest content according to the semantic embedding vectors of their corresponding topic tags (such as K-Means clustering, density-based noise applied space clustering (DBSCAN), or hierarchical clustering). The content in each group is highly similar in topic semantics and represents a sub-field of interest (such as sub-fields such as "outdoor equipment", "lightweight camping", and "wild cooking").

[0114] Representative media content refers to the highest-scoring piece of media content (such as text, video, or post) selected from each content cluster based on its corresponding attention score. This content represents both the semantic theme of the cluster and has the strongest relevance to the target user along the social connection path.

[0115] Target potential interest content refers to the collection of all representative media content selected from all content clusters. While maintaining thematic diversity, this collection ensures that each piece of content is the most relevant and recommendable candidate in its semantic cluster, serving as the refined recommendation result ultimately pushed to the target user.

[0116] Understandably, in some embodiments, to avoid recommending a large amount of semantically repetitive or thematically overlapping potential interest content to target users, after initially screening media content with high attention scores, the system further introduces a clustering and representativeness selection mechanism based on topic semantics. Specifically, this involves processing each piece of potential interest content... Extract the difference topic tags to which they belong (e.g., "#portable water purifier"), and obtain its semantic embedding vector through a pre-trained topic embedding model. ∈ All As input, the data is divided into K clusters using a clustering algorithm (such as K-Means), resulting in multiple content clusters { , ,…, Each cluster contains semantically similar content of potential interest. For each content cluster... It iterates through all media content, compares their attention scores, and selects the highest-scoring content as the representative media content of that cluster. All (k=1,…,K) form a set of potential interest content for the target user, and recommend it to the target user, thereby achieving a recommendation effect that balances diversity and relevance.

[0117] 105. Recommend content of potential interest to target users.

[0118] In some embodiments, by fusing potential interest content mined from heterogeneous social networks with target media content matched based on user profiles, this application achieves a "reliable yet innovative" recommendation strategy: on the one hand, the target media content ensures the basic relevance of the recommendation results and user satisfaction; on the other hand, the potential interest content introduces novel, highly reliable signals across different user groups, effectively stimulating the expansion of user interest boundaries. The synergistic effect of these two approaches avoids information cocoons caused by purely interest-driven behavior and prevents experience disruptions due to over-exploration, significantly improving the long-term user retention rate, content diversity metrics, and platform ecosystem vitality of the recommendation system. Recommending potential interest content to target users includes:

[0119] The network embedding representation of the target user is compared with the preset interest tag vector to calculate the similarity and determine the target interest tag that matches the target user.

[0120] Filter out target media content that has not yet been recommended to the target users from the media content library corresponding to the target interest tags;

[0121] The system integrates and sorts potential interest content with target media content to generate a recommendation list.

[0122] Push recommendation lists to target users.

[0123] The preset interest tag vectors refer to a set of standardized semantic embedding representations of interest topics pre-constructed by the platform. Each vector corresponds to a high-level interest category (such as "outdoor sports", "technology and digital", "career growth", "food exploration", etc.). These vectors can be generated by encoding interest keywords through large-scale user behavior clustering, expert annotation, or pre-trained language models (such as BERT), and uniformly mapped to the same vector space as the user's network embedding representation to support cross-modal similarity calculation.

[0124] Target interest tags refer to one or more interest tags selected by calculating the similarity (such as cosine similarity) between the target user's network embedding representation and various preset interest tag vectors. This tag represents a structured discrimination result of the target user's current core interest area, used to guide the recall of high-quality content.

[0125] The media content library refers to the collection of all media content (such as text, video, and posts) on a social media platform that has been tagged with interest tags. Each piece of content is associated with one or more interest tags (which may come from topic tags or manual / automatic classification systems) and supports efficient retrieval by interest tags.

[0126] Targeted media content refers to media content that has not yet been recommended to the target user, further filtered from the media content library corresponding to the target interest tags. This filtering is achieved by comparing the user's historical recommendation logs to ensure the freshness of the content and avoid duplicate exposure.

[0127] The recommendation list refers to the final sequence of content to be recommended generated by merging and sorting potential interest content (novel content mined based on social differences) with target media content (highly relevant content based on interest matching). The sorting can comprehensively consider multiple factors such as attention score, content popularity, timeliness, and diversity constraints to balance the dual goals of "interest consolidation" and "interest expansion".

[0128] It is understandable that, in some embodiments, to balance interest relevance and novelty, the system introduces a dual-channel fusion recommendation mechanism when recommending potentially interesting content to the target user. Specifically, this involves the target user's network embedding representation... With all preset interest tag vectors { , …, Perform similarity calculations (such as cosine similarity): The tags with the highest similarity are selected as target interest tags (or Top-K tags). All content belonging to the target interest tags is retrieved from the media content library, excluding content already appearing in the target user's historical recommendation records, to obtain the target media content. The potential interest content obtained in the previous steps (emphasizing novelty and social credibility) is merged with the target media content (emphasizing interest matching) and input into a unified ranking model. This model can be a weighted scoring function, for example: α·AttentionScore(m) + (1-α)·RelevanceScore(m), where α∈[0,1] controls the balance between exploration and utilization, AttentionScore(m) is the attention score, and RelevanceScore(m) is the relevance score, where RelevanceScore(m) represents the degree of direct matching between media content m and the target user's currently known interests and preferences. This score is calculated based on user profiles, historical behavior, or interest tag alignment, reflecting whether the content belongs to the user's "proven interest" domain. The final recommendation list is generated by sorting the scores in descending order and pushed to the target user through client news feeds, personalized notifications, etc.

[0129] In some embodiments, by introducing a dynamic weight adjustment mechanism based on community change records, this application achieves context-aware adaptive capabilities for the recommendation strategy: for users with stable social networks, the system proactively enhances interest expansion efforts, effectively mitigating information narrowing caused by long-term use; for users with active or frequently changing social networks, the system prioritizes content relevance and experience consistency, avoiding the introduction of excessive noise during periods of interest drift. This mechanism organically combines macro-level user social behavior signals (community stability) with micro-level content recommendation strategies (exploration / utilization balance), significantly improving the robustness, personalization level, and long-term user satisfaction of the recommendation system, while also considering the diversity of the platform's content ecosystem and user retention goals.

[0130] The system merges and sorts potential interest content with target media content to generate a recommendation list, including:

[0131] Obtain the community change records of the target user's communities within a preset time window;

[0132] If the community change record is empty, increase the recommendation weight of potential interest content and decrease the recommendation weight of target media content.

[0133] If the community change record is not empty, increase the recommendation weight of the target media content and decrease the recommendation weight of the potential interest content.

[0134] Based on the adjusted recommendation weights, potential interest content and target media content are weighted and sorted to generate a recommendation list.

[0135] The preset time window refers to a defined continuous time interval (such as the last 7 days, 30 days, or user activity cycles) used to limit the observation range of user community behavior stability. This window can be dynamically configured according to platform policies, aiming to capture recent trends in users' social structure changes.

[0136] Community change logs refer to the logs of changes in community membership that occur to a target user within a preset time window, including events such as joining a new community, leaving an existing community, being removed from a community, or the disbandment of a community. If a user does not perform any community additions or deletions within this window, their community change logs will be empty, indicating that their social circle is in a stable state.

[0137] Recommendation weight refers to the numerical coefficient assigned to different categories of candidate content (such as content of potential interest or target media content) during the fusion and ranking stage, used to adjust their priority in the final recommendation list. The higher the weight, the higher the content of that category will be ranked; the weight can be dynamically adjusted to respond to changes in user status.

[0138] Understandably, the system first queries the activity logs of the target user's communities within a preset time window (e.g., the last 30 days) to generate a community change record. If the community change record is empty (i.e., the user has not recently joined or left any communities), it indicates that their social circle is stable and their interests are becoming fixed. In this case, the system increases the recommendation weight of potential interest content (e.g., increasing α from 0.5 to 0.7) and decreases the recommendation weight of target media content to proactively guide them to break out of their information cocoon and explore adjacent interest areas. If the community change record is not empty (e.g., frequent switching of communities), it indicates that the user is in a period of interest migration or exploration, and their social behavior is unstable. In this case, the system increases the recommendation weight of target media content (e.g., increasing 1−α to 0.8) and decreases the recommendation weight of potential interest content to prioritize providing reliable content that highly matches their current profile, avoiding user confusion or negative feedback due to excessive recommendations of novel content. Based on the adjusted weights, a weighted score is applied to the two types of content. For example, for any candidate content m: if m belongs to potential interest content, its final score is calculated as follows: If m belongs to the target media content, its final score is calculated as follows: ;in + =1, and the weight is dynamically allocated based on community change records. Potential interest content and target media content are merged and sorted to generate a recommendation list.

[0139] In some embodiments, by introducing an interest confirmation mechanism based on interactive feedback and a secondary mining strategy for content on the same topic, this application achieves an upgrade from "single interest exploration" to "continuous interest cultivation." This mechanism not only verifies the effectiveness of potential interest content but also efficiently recalls high-quality related content using confirmed new interest tags, significantly improving user dwell time, interaction depth, and conversion rate in new interest areas. Simultaneously, this closed-loop design enhances the adaptive learning capability of the recommendation system, enabling the platform to dynamically capture the evolution trajectory of user interests and further optimize long-term content distribution efficiency and user experience. After recommending potential interest content to target users, the system also includes:

[0140] Acquire data on the interaction behavior of target users with potentially interesting content;

[0141] If the interaction behavior data reaches a preset threshold, it is determined that the potential interest content has been accepted by the target user;

[0142] Obtain target topic tags corresponding to potential interest content, and filter out other media content that has not been recommended from the media content corresponding to the target topic tags;

[0143] Recommend other media content to target users.

[0144] Interactive behavior data refers to the records of actions taken by target users on social media platforms after receiving content of potential interest. These actions include, but are not limited to, positive interactions such as clicking, watching (e.g., video playback duration ≥ a threshold), liking, commenting, forwarding, saving, and following the content publisher. This data is used to quantify the user's actual interest in the content.

[0145] Target hashtags refer to differentiated hashtags associated with potential interest content (e.g., "#lightweightcamping"). Once it is determined that the content has been accepted by the user, the hashtag is identified as the user's newly accepted interest direction and serves as an anchor point for subsequent content expansion.

[0146] Other media content refers to other media content (such as text, short videos, posts, etc.) that belongs to the target topic tag but has not yet been recommended to the target user. This content constitutes a secondary recommendation candidate pool based on newly confirmed interests, used to deepen the user's content consumption under this topic.

[0147] Understandably, potential interest content (or further refined target potential interest content) obtained after multiple stages of screening and optimization is pushed to target users through the recommendation interface of social media platforms. This recommendation can appear in news feeds, interest exploration zones, personalized push notifications, or "You may be interested in" recommendation slots.

[0148] The recommendation process can dynamically sort or adapt the display format (such as prioritizing short videos, image and text card layouts, etc.) based on the target user's real-time context (e.g., current active time, device type, geographical location) to improve click-through rates and engagement. Furthermore, the system can record subsequent user feedback on recommended content (e.g., whether they clicked, liked, saved, or skipped), which is used to optimize the construction of heterogeneous information networks, attention score calculation, and clustering strategies in a closed loop, enabling continuous iteration of the recommendation model.

[0149] This recommendation mechanism not only breaks through the traditional "interest echo chamber" based on users' own historical behavior, but also achieves a paradigm upgrade from "passively catering to known interests" to "actively guiding reasonable interest expansion" by integrating social structure (common community), cross-circle semantic differences (different topics), path credibility (attention score), and content diversity control (semantic clustering). This effectively alleviates information cocoons, improves the information diversity of recommendation results and the effectiveness of expanding user interests, and also helps to enhance the overall content ecosystem diversity and information dissemination breadth of the platform.

[0150] To better implement the above methods, this application also provides a social media information cocoon reduction device based on a heterogeneous information network. This device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet, smart Bluetooth device, laptop, or personal computer; the server can be a single server or a server cluster composed of multiple servers.

[0151] For example, in this embodiment, a social media information cocoon reduction device based on a heterogeneous information network will be specifically integrated into an electronic device to illustrate the method of this application embodiment in detail.

[0152] For example, such as Figure 2 As shown, the social media information cocoon reduction device based on heterogeneous information networks may include a data acquisition unit 201, a network construction unit 202, a user determination unit 203, a content filtering unit 204, and a content recommendation unit 205, as follows:

[0153] (I) Data acquisition unit 201.

[0154] The data acquisition unit 201 is used to acquire historical interaction data of multiple users on social media platforms. The historical interaction data includes the corresponding user's historical interactive media content, the topic tags to which the historical interactive media content belongs, and the communities that the corresponding user participates in.

[0155] (ii) Network building unit 202.

[0156] Network building unit 202 is used to build a corresponding heterogeneous information network for each user based on their historical interaction data. The heterogeneous information network includes multiple types of nodes and edges between nodes. The multiple types of nodes include user nodes, media content nodes, topic nodes and community nodes. The edges between nodes represent the association relationship between the corresponding nodes.

[0157] (III) User determination unit 203.

[0158] User determination unit 203 is used to match, for any target user, community nodes in its heterogeneous information network with community nodes in the heterogeneous information network of other users, and determine other users who have at least one common community as associated users of the target user.

[0159] In some embodiments, it also includes:

[0160] Graph structure features are extracted from heterogeneous information networks to obtain network embedding representations;

[0161] Extract community embedding features corresponding to each community node from the network embedding representation;

[0162] Based on the community nodes in its heterogeneous information network, it matches the community nodes in the heterogeneous information networks of other users, identifying other users who share at least one common community as associated users of the target user, including:

[0163] The community embedding features corresponding to the target user are respectively matched with the community embedding features corresponding to other users.

[0164] If there exists at least one pair of community embedding features whose similarity is greater than or equal to a preset similarity, then the corresponding other users will be identified as associated users of the target user.

[0165] In some embodiments, before matching other users with community nodes in the heterogeneous information networks of any target user based on their community nodes in the heterogeneous information network, and identifying other users who share at least one common community as associated users of the target user, the method further includes:

[0166] Retrieve the community change records of each user's participating communities within a preset time window;

[0167] Users whose community change records are empty are identified as target users.

[0168] (iv) Content filtering unit 204.

[0169] The content filtering unit 204 is used to filter out potential interest content of the target user from the media content corresponding to the different topic tags based on the differences in topic distribution between the target user's heterogeneous information network and the heterogeneous information network of related users. The different topic tags are different from the topic tags of the target user's historical interactive media content.

[0170] In some embodiments, based on the differences in topic distribution between the heterogeneous information networks of the target user and the heterogeneous information networks of associated users, potential interest content of the target user is filtered from media content corresponding to differing topic tags, including:

[0171] Calculate the community semantic difference between the community embedding features corresponding to each community node of the target user and the community embedding features corresponding to each community node of the associated user;

[0172] Based on the semantic differences of the community, filter out the communities that the target user has not participated in from the communities that the associated users participate in;

[0173] Based on the topic tags of the target user's historical interactive media content, identify the different topic tags from the topic tags contained in the different communities, and obtain the media content corresponding to the different topic tags.

[0174] Filter out content that may be of interest to the target users from media content corresponding to different topic tags.

[0175] In some embodiments, the target user's historical interaction data may also include interaction data between the target user and associated users in a common community;

[0176] Filter out content of potential interest to the target users from media content corresponding to different topic tags, including:

[0177] For each different topic tag corresponding to media content, based on the different communities to which the media content belongs, the participation behavior of related users in the different communities, and the interaction data of target users and related users in the common community, a connection path from target users to media content is constructed.

[0178] The embedded features of nodes on the associated path are aggregated to obtain the semantic representation vector of the associated path;

[0179] Based on the user embedding features and semantic representation vectors of the target user, the attention score of the associated path is calculated;

[0180] Based on the attention score, media content with an attention score higher than the preset attention score is selected from the media content corresponding to the different topic tags, and these are taken as the potential interest content of the target users.

[0181] In some embodiments, based on the topic tags to which the target user's historical interactive media content belongs, differentiating topic tags are determined from the topic tags contained in the differentiating communities, including:

[0182] Obtain user profile information of the target users;

[0183] Based on the topic tags of the target user's historical interactive media content, candidate topic tags are selected from the topic tags contained in the different communities.

[0184] Based on user profile information, candidate topic tags are filtered to obtain filtered differentiated topic tags.

[0185] (V) Content Recommendation Unit 205.

[0186] Content recommendation unit 205 is used to recommend potentially interesting content to target users.

[0187] In some embodiments, recommending content of potential interest to a target user includes:

[0188] The network embedding representation of the target user is compared with the preset interest tag vector to calculate the similarity and determine the target interest tag that matches the target user.

[0189] Filter out target media content that has not yet been recommended to the target users from the media content library corresponding to the target interest tags;

[0190] The system integrates and sorts potential interest content with target media content to generate a recommendation list.

[0191] Push recommendation lists to target users.

[0192] In some embodiments, potential interest content is merged and sorted with target media content to generate a recommendation list, including:

[0193] Obtain the community change records of the target user's communities within a preset time window;

[0194] If the community change record is empty, increase the recommendation weight of potential interest content and decrease the recommendation weight of target media content.

[0195] If the community change record is not empty, increase the recommendation weight of the target media content and decrease the recommendation weight of the potential interest content.

[0196] Based on the adjusted recommendation weights, potential interest content and target media content are weighted and sorted to generate a recommendation list.

[0197] In some embodiments, after recommending potentially interesting content to the target user, the method further includes:

[0198] Acquire data on the interaction behavior of target users with potentially interesting content;

[0199] If the interaction behavior data reaches a preset threshold, it is determined that the potential interest content has been accepted by the target user;

[0200] Obtain target topic tags corresponding to potential interest content, and filter out other media content that has not been recommended from the media content corresponding to the target topic tags;

[0201] Recommend other media content to target users.

[0202] In some embodiments, after selecting media content with attention scores higher than a preset attention score from the media content corresponding to the different topic tags based on attention scores as potential interest content for the target user, the method further includes:

[0203] Based on the semantic embedding vectors of the topic tags corresponding to the potential interest content, the potential interest content is clustered to obtain multiple content clusters;

[0204] The media content with the highest attention score in each content cluster is selected as the representative media content;

[0205] Identify representative media content as potential target interest content;

[0206] Recommend content of potential interest to target users, including:

[0207] Recommend content that aligns with the target user's potential interests.

[0208] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0209] Therefore, the embodiments of this application not only integrate users' own interactive behavior and social topology, but also actively discover their interest blind spots through cross-circle topic difference analysis, effectively alleviating the information cocoon effect. This not only improves the information diversity of recommendation results and the effectiveness of expanding user interests, but also helps to enhance the overall content ecosystem diversity and information dissemination breadth of the platform.

[0210] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0211] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the social media information cocoon reduction methods based on heterogeneous information networks provided in embodiments of this application.

[0212] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0213] Since the instructions stored in the storage medium can execute the steps in any of the social media information cocoon reduction methods based on heterogeneous information networks provided in the embodiments of this application, the beneficial effects that any of the social media information cocoon reduction methods based on heterogeneous information networks provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0214] According to one aspect of this application, a computer program product or computer program is provided, comprising a computer program / instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program / instructions from the computer-readable storage medium and executes the computer program / instructions, causing the electronic device to perform the method provided in the above embodiments regarding the reduction of social media information cocoons based on heterogeneous information networks.

[0215] The above provides a detailed description of a method for reducing social media information cocoons based on heterogeneous information networks, as provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for reducing social media information cocoons based on heterogeneous information networks, characterized in that, The method includes: Acquire historical interaction data of multiple users on social media platforms. The historical interaction data includes the corresponding user's historical interactive media content, the topic tags to which the historical interactive media content belongs, and the communities that the corresponding user participates in. For each user, a corresponding heterogeneous information network is constructed based on their historical interaction data. The heterogeneous information network includes multiple types of nodes and edges between nodes. The multiple types of nodes include user nodes, media content nodes, topic nodes, and community nodes. The edges between nodes represent the association relationships between corresponding nodes. For any target user, based on the community nodes in its heterogeneous information network, it is matched with the community nodes in the heterogeneous information networks of other users, and other users who have at least one common community are identified as the associated users of the target user. Based on the differences in topic distribution between the heterogeneous information network of the target user and the heterogeneous information network of the associated user, potential interest content of the target user is filtered from the media content corresponding to the different topic tags. The different topic tags are different from the topic tags to which the target user's historical interactive media content belongs. Recommend the potential interest content to the target user.

2. The method as described in claim 1, characterized in that, Also includes: Graph structure features are extracted from the heterogeneous information network to obtain the network embedding representation; Extract community embedding features corresponding to each community node from the network embedding representation; The step of matching community nodes in its heterogeneous information network with community nodes in other users' heterogeneous information networks, and identifying other users who share at least one common community as associated users of the target user, includes: The community embedding features corresponding to the target user are respectively matched with the community embedding features corresponding to other users. If there exists at least one pair of community embedding features whose similarity is greater than or equal to a preset similarity, then the corresponding other users are identified as associated users of the target user.

3. The method as described in claim 2, characterized in that, The method of filtering potential interest content of the target user from media content corresponding to differing topic tags based on the differences in topic distribution between the heterogeneous information network of the target user and the heterogeneous information network of the associated users includes: Calculate the community semantic difference between the community embedding features corresponding to each community node of the target user and the community embedding features corresponding to each community node of the associated user; Based on the semantic difference of the community, filter out the differential communities that the target user has not participated in from the communities that the associated user has participated in; Based on the topic tags of the target user's historical interactive media content, different topic tags are determined from the topic tags included in the different communities, and the media content corresponding to the different topic tags is obtained. The target user's potential interest content is selected from the media content corresponding to the differentiated topic tags.

4. The method as described in claim 3, characterized in that, The target user's historical interaction data also includes the interaction data between the target user and the associated user in the common community; The step of filtering the potential interest content of the target user from the media content corresponding to the differentiated topic tags includes: For each media content corresponding to a different topic tag, based on the different community to which the media content belongs, the participation behavior of the associated user in the different community, and the interaction data between the target user and the associated user in the common community, a connection path from the target user to the media content is constructed. The node embedding features on the associated path are aggregated to obtain the semantic representation vector of the associated path; Based on the user embedding features of the target user and the semantic representation vector, the attention score of the associated path is calculated; Based on the attention score, media content with an attention score higher than a preset attention score is selected from the media content corresponding to the differential topic tags, and these are taken as the potential interest content of the target user.

5. The method as described in claim 1, characterized in that, The step of recommending the potential interest content to the target user includes: The network embedding representation of the target user is compared with the preset interest tag vector to calculate the similarity and determine the target interest tag that matches the target user. From the media content library corresponding to the target interest tags, filter out target media content that has not yet been recommended to the target user; The potential interest content is merged and sorted with the target media content to generate a recommendation list; The recommendation list is pushed to the target user.

6. The method as described in claim 5, characterized in that, The step of fusing and sorting the potential interest content with the target media content to generate a recommendation list includes: Obtain the community change records of the target user's communities within a preset time window; If the community change record is empty, the recommendation weight of the potential interest content is increased, and the recommendation weight of the target media content is decreased. If the community change record is not empty, then increase the recommendation weight of the target media content and decrease the recommendation weight of the potential interest content. Based on the adjusted recommendation weights, the potential interest content and the target media content are weighted and sorted to generate the recommendation list.

7. The method as described in claim 1, characterized in that, After recommending the potential interest content to the target user, the method further includes: Obtain the interaction behavior data of the target user with the potential interest content; If the interactive behavior data reaches a preset threshold, it is determined that the potential interest content has been accepted by the target user; Obtain the target topic tags corresponding to the potential interest content, and filter out other media content that has not been recommended from the media content corresponding to the target topic tags; Recommend the other media content to the target user.

8. The method as described in claim 3, characterized in that, The step of determining the differential topic tags from the topic tags included in the differential community, based on the topic tags belonging to the target user's historical interactive media content, includes: Obtain the user profile information of the target user; Based on the topic tags of the target user's historical interactive media content, candidate topic tags are selected from the topic tags included in the differentiated community. Based on the user profile information, the candidate topic tags are filtered to obtain filtered differentiated topic tags.

9. The method as described in claim 4, characterized in that, After selecting media content with an attention score higher than a preset attention score from the media content corresponding to the difference topic tags based on the attention score, as the potential interest content of the target user, the method further includes: Based on the semantic embedding vectors of the topic tags corresponding to the potential interest content, the potential interest content is clustered to obtain multiple content clusters; The media content with the highest attention score in each content cluster is selected as the representative media content; The representative media content was identified as target potential interest content; The step of recommending the potential interest content to the target user includes: Recommend content of potential interest to the target user.

10. The method as described in claim 1, characterized in that, Before matching any target user with community nodes in its heterogeneous information network and other users' heterogeneous information networks, and identifying other users who share at least one common community as associated users of the target user, the method further includes: Retrieve the community change records of each user's participating communities within a preset time window; Users whose community change records are empty are identified as target users.

Images