Content recommendation method and apparatus

Abstract

Embodiments of the present specification provide a content recommendation method and device. The content recommendation method comprises: generating a target identification sequence according to interest feature data of a target user; determining a similar user set from an initial user set according to the target identification sequence and a user identification sequence of each initial user in the initial user set; determining historical behavior data of the target user according to user behavior data corresponding to the similar user set; and determining recommended content corresponding to the target user according to the historical behavior data. The historical behavior data of the target user is constructed according to the user behavior data corresponding to the similar user set, which effectively enriches the behavior data of a sparse behavior user and makes up for the modeling deficiency caused by the lack of individual behavior data. The method avoids completely relying on popular recommendations due to insufficient data, significantly improves the matching degree of recommended content and individual real interests, and especially improves the poor recommendation effect in the cold start and sparse behavior scenarios.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a content recommendation method. One or more embodiments of this specification also relate to a content recommendation apparatus, a computing device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] In content application platforms, there are significant differences in the richness of user behavior. Existing recommendation algorithms can make accurate content recommendations for users with rich behavioral data, but for users with relatively sparse behavioral data, the recommended content mostly relies on popular recommendations, which has a low degree of matching with the content that the individual is actually interested in, or the recommendation results are unstable and inaccurate due to the lack of user behavioral data. Summary of the Invention

[0003] In view of this, embodiments of this specification provide a content recommendation method. One or more embodiments of this specification also relate to a content recommendation apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, a content recommendation method is provided, comprising: Generate a target identifier sequence based on the target user's interest feature data; Based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set, a similar user set is determined from the initial user set; Based on the user behavior data corresponding to the set of similar users, determine the historical behavior data of the target user; Recommended content for the target user is determined based on the historical behavior data.

[0005] According to a second aspect of the embodiments of this specification, a content recommendation apparatus is provided, comprising: The sequence generation module is configured to generate target identifier sequences based on the target user's interest feature data; The user determination module is configured to determine a set of similar users from the initial user set based on the target identifier sequence and the user identifier sequence of each initial user in the initial user set; The data determination module is configured to determine the historical behavior data of the target user based on the user behavior data corresponding to the similar user set; The content determination module is configured to determine the recommended content corresponding to the target user based on the historical behavior data.

[0006] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the recommended method described above.

[0007] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0008] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0009] One embodiment of this specification provides a content recommendation method, comprising: generating a target identifier sequence based on the interest feature data of a target user; determining a similar identifier sequence from a set of user identifier sequences based on the target identifier sequence; determining the historical behavior data of the target user based on the user behavior data corresponding to the similar identifier sequence; and determining recommended content corresponding to the target user based on the historical behavior data.

[0010] The content recommendation method provided in this specification generates a target identifier sequence composed of multi-level identifiers based on the target user's interest feature data. This target identifier sequence indicates the target user's interest categories at multiple granular levels, thereby characterizing the user's interest features layer by layer from macro-interest categories to micro-interest preferences. Based on this, the target identifier sequence is matched with the user identifier sequences of each initial user in the initial user set to determine a set of similar users whose interest semantics are similar to the target user's. Historical behavior data of the target user is constructed based on the user behavior data corresponding to the similar user set, effectively enriching the behavior data of sparse behavior users and compensating for insufficient modeling caused by missing individual behavior data. Recommended content is determined to be recommended to the target user based on the historical behavior data, avoiding complete reliance on popular recommendations due to insufficient data, significantly improving the matching degree between recommended content and the individual's true interests, especially improving the poor recommendation effect caused by insufficient data in cold start and sparse behavior scenarios. Attached Figure Description

[0011] Figure 1 This is a schematic diagram illustrating a scenario of a content recommendation method provided in one embodiment of this specification; Figure 2 This is a flowchart illustrating a content recommendation method provided in one embodiment of this specification; Figure 3 This is a flowchart illustrating the vector mapping process in a content recommendation method provided in one embodiment of this specification. Figure 4 This is a flowchart illustrating the interest migration process in a content recommendation method provided in one embodiment of this specification. Figure 5 This is a schematic diagram of the processing stages in a content recommendation method provided in one embodiment of this specification; Figure 6 This is a schematic diagram of the structure of a content recommendation device provided in one embodiment of this specification; Figure 7 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0012] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0013] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0014] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0015] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0016] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0017] CTR: Click-Through Rate, the percentage of content that is clicked after it has been displayed.

[0018] SID: SemanticIdentifier, in this embodiment, is a user interest encoding sequence generated by a multi-level residual quantization algorithm.

[0019] RQ-KMeans: Residual Quantization K-Means is a K-means algorithm for residual quantization. Each level clusters the residuals of the previous level to form a hierarchical structure.

[0020] CGC: Customized Gate Control, used for feature fusion in multi-task learning.

[0021] Embedding: Vector representation, used in the embodiments of this specification to map user multi-source interest signals into dense vectors of fixed dimensions.

[0022] TGI: Target Group Index, measures the strength of a group's preference for a specific interest.

[0023] Last N: A sequence of users' historical behavior, used for personalized recall and ranking in recommendation systems.

[0024] Similar audience segment: A set of users with similar interests to the target user, constructed through precise matching of SID prefixes.

[0025] Residual quantization is a multi-stage, hierarchical data compression technique that achieves progressively more accurate compression of the original data by repeatedly quantizing the errors (residuals) generated in the previous step. It is widely used in large model compression, audio and video processing, and vector retrieval, significantly reducing storage and computational costs while maintaining high accuracy.

[0026] In recommendation platforms, the richness of user behavior varies significantly. Existing systems are accurate in modeling users with intensive behavior, but are ineffective for users with sparse behavior (including new users and long-tail users), mainly facing three major challenges: First, in the cold start phase, due to the lack of individual historical data, recommendations rely on global popular or coarse profiles, which have low matching with real interests, resulting in poor first-time experience and high churn rate; second, the interest estimation of users with sparse behavior is unstable, resulting in large fluctuations and low accuracy in recommendation results, making it difficult to form a positive feedback loop; third, the system is prone to the Matthew effect, continuously tilting resources towards active users, widening the experience gap, and damaging the overall retention and activity of the platform.

[0027] To address the aforementioned issues, existing technical solutions have significant limitations: for example, single-layer K-means coarse-grained clustering results in vague user segmentation and fails to precisely characterize interests; methods based on item substitution or semantic encoding focus on the item side or recall stage, neglecting user cold-start modeling; and strategies relying on building hierarchical trees or cross-domain data migration based on user historical behavior cannot cover scenarios with no behavior or missing cross-domain information. In summary, there is currently a lack of a systematic solution for achieving precise interest migration from similar groups to individual users in scenarios with no or sparse behavior.

[0028] This specification provides a content recommendation method, and also relates to a content recommendation device, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0029] See Figure 1 , Figure 1 This diagram illustrates a scenario of a content recommendation method provided in one embodiment of this specification.

[0030] Specifically, the content recommendation method is implemented using the client-side device 102 and the server 104. The client-side device 102 is used to send content recommendation requests to the server 104.

[0031] When server 104 receives a content recommendation request sent by terminal device 102, it generates a target identifier sequence based on the target user's interest feature data; it determines a similar user set from the initial user set based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set; it determines the target user's historical behavior data based on the user behavior data corresponding to the similar user set; it determines the recommended content corresponding to the target user based on the historical behavior data, and returns the recommended content to terminal device 102.

[0032] The edge device 102 may include a browser, an app (application), or a web application such as an H5 (Hypertext Markup Language 5) application, a lightweight application (also known as a mini-program), or a cloud application. The edge device can be developed based on a software development kit (SDK) provided by the server, such as a real-time communication (RTC) SDK. The edge device can be deployed in an electronic device and depends on the device's operation or certain apps within the device to run. The electronic device may have a display screen and support information browsing, such as a personal mobile terminal like a mobile phone, tablet, or personal computer. Various other types of applications can also be configured in the electronic device, such as human-computer interaction applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social media platform software.

[0033] Server 104 can be understood as a server providing various services, including physical servers and cloud servers. Examples include servers providing communication services to multiple clients, servers supporting backend training of models used on clients, and servers processing data sent by clients. It's important to note that Server 104 can be implemented as a distributed server cluster composed of multiple servers, or as a single server. Server 104 can also be a server in a distributed system, or a server integrated with blockchain. Server 104 can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0034] See Figure 2 , Figure 2 A flowchart of a content recommendation method provided in one embodiment of this specification is shown, which specifically includes the following steps.

[0035] Step 202: Generate a target identifier sequence based on the target user's interest feature data.

[0036] The target users are typically those with sparse or extremely scarce behavioral data, such as new users launching the application for the first time (cold start users), returning users who have been inactive for a long time and then log back in, and users with sparse behavioral data who have only a few scattered interactions within the platform.

[0037] Interest feature data can be understood as various semantic information related to user interests obtained from multiple data sources, including those within the content application platform, outside the content application platform, and third-party channels. This includes, but is not limited to, the basic attribute information filled in by the user during registration, the model and operating system information of the device used, advertising interest data mined in real time through the RTA (Real-Time API) interface, and limited interaction records generated by the user in weakly related scenarios such as H5 pages.

[0038] The target identifier sequence can be understood as a user semantic identifier (SID) consisting of multiple levels of identifiers, obtained by mapping and encoding interest feature data.

[0039] Specifically, in content recommendation scenarios involving new user cold starts, returning users, and users with sparse behavior, the historical behavior data of target users is extremely scarce or even completely missing. It is impossible to directly model interests based on traditional behavior sequences. Therefore, it is necessary to fully explore and utilize the multi-source heterogeneous information available to users, integrate and encode the interest feature data obtained from multiple data sources, and convert the user's interest feature data into a hierarchical semantic structure through mapping encoding. This allows even target users with missing individual behavior data to obtain effective encoding in a unified interest semantic space, providing an indexable and matchable identifier foundation for subsequent user matching.

[0040] In one or more embodiments of this specification, a user interest vector is constructed based on the target user's interest feature data. Furthermore, a target identifier sequence is generated by vector mapping of the user interest vector. Specific implementation methods are described below: Generate a target identifier sequence based on the target user's interest feature data, including: Based on the target user's interest feature data, a user interest vector is constructed, wherein the target user's interest feature data is obtained from multiple data sources; The user interest vector is vector-mapped to generate the target identifier sequence, wherein the hierarchical identifier of each level in the target identifier sequence represents the interest category of the target user at different granularities.

[0041] Among them, the user interest vector can be understood as a fixed-dimensional vector representation generated after mapping the above interest feature data to a unified semantic space, which is used to characterize the position of the target user in the interest semantic space.

[0042] Vector mapping can be understood as the process of converting user interest vectors in a high-dimensional continuous interest vector space into a discretized, structured identifier sequence according to a preset mapping rule. Hierarchical identifiers can be understood as codeword index values ​​of a certain level in the target identifier sequence. Each hierarchical identifier is used to classify users into specific interest categories under that level.

[0043] Specifically, the system acquires multi-source interest feature data of the target user from multiple preset data sources. These multiple data sources include registration information and device information within the site, data tags from third-party DMP (data collected from non-proprietary channels and imported into the data management platform) purchased from external third parties, refined interest signals mined in real-time by second-party RTA, and records of weak interactive behaviors generated by the user in the past in H5 scenarios. Subsequently, the above-mentioned interest feature data from different sources and in different formats are standardized and vectorized preprocessed to generate user interest vectors for the target user.

[0044] Given a target user's corresponding user interest vector, the user interest vector is vector-mapped to transform the continuous user interest vector into a discrete identifier sequence with a clear hierarchical structure, thereby generating a target identifier sequence.

[0045] For example, in one application scenario, when a new user launches the application for the first time, the system can only obtain the interest tags (sports, variety shows) selected by the user during registration and device brand A. This information is mapped to corresponding feature codes to generate the user's interest vector. In another application scenario, when a long-inactive returning user logs back into the platform, the system infers the user's spending power level through device information and simultaneously obtains the user's interest tags from external advertising systems in real time via the RTA interface (e.g., recently viewed maternal and infant advertisements). This information is encoded and fused with historically retained decay features to construct the returning user's interest vector.

[0046] This embodiment integrates multi-source interest feature data that can be obtained from target users in the absence of behavior or in a sparse behavior state, and generates user interest vectors corresponding to the multi-source interest feature data. This enables target users with missing individual behavior data to obtain effective encoding in a unified interest semantic space, providing a reliable data foundation for subsequent multi-level quantization and population matching.

[0047] In the embodiments of this specification, the vector mapping refers to the process of converting each interest vector in a high-dimensional continuous interest vector space into a discretized and structured identifier sequence according to a preset mapping rule. This mapping process aims to map semantically similar interest vectors to the same or similar identifier regions, so that users with similar interests have proximity or share the same prefix identifier in the identifier space, thereby providing a foundation for subsequent hierarchical user segmentation and retrieval.

[0048] The specific implementation of vector mapping is not limited to a fixed algorithm or process. For example, in one or more embodiments, multi-level residual quantization (i.e., residual quantization K-means algorithm) can be used to implement vector mapping: by performing clustering (such as K-means clustering) on ​​the set of interest vectors level by level, each vector is mapped to the cluster center at each level in turn, and the cluster indices at each level are concatenated in order to form the user's semantic identifier sequence SID. In other embodiments, vector mapping can also be implemented based on product quantization (PQ), locality-sensitive hashing (LSH), or neural network-based encoding mapping (such as autoencoders or semantic hash networks), etc., without limitation, as long as it can convert high-dimensional interest vectors into discrete identifier sequences with hierarchical semantic structures.

[0049] The embodiments in this specification employ the Residual Quantization K-Means (RQ-KMeans) algorithm. Each layer continues clustering quantization based on the quantization residuals of the previous layer, forming a hierarchical structure from coarse-grained to fine-grained. Based on this, the target identifier sequence can be understood as a user semantic identifier (SID) generated through multi-level quantization and composed of multiple codeword indices. The earlier the level in the sequence, the coarser the granularity and the wider the coverage of the user group; the later the level, the finer the granularity and the more precise the interest characterization. The hierarchical identifier can be understood as the codeword index value of a certain level in the target identifier sequence. Each hierarchical identifier is used to classify users into specific interest categories under that granularity level.

[0050] Specifically, given the target user's corresponding user interest vector, multiple levels of quantization are performed on the user interest vector. For example, there are 8 quantization levels. The codeword index i obtained at each level is concatenated in hierarchical order to form the target user's 8-level semantic identifier sequence SID=[i1,i2,i3,i4,i5,i6,i7,i8].

[0051] In fact, although the full 8-level quantization is technically supported in this embodiment, according to conflict rate analysis and generalization performance evaluation, the first two levels of identifiers (i1, i2) can be enabled in actual use. In actual application, since the interest feature data of the target user is relatively sparse, the interest feature data of the target user may not be enough to be subdivided into 8 quantization levels. The users can be effectively distinguished by the quantization of the first two levels.

[0052] In the embodiments of this specification, if the target user can be effectively distinguished through the first two levels of quantization, and if continuing to use deeper levels may lead to the risk of overfitting, multi-level quantization can be quantization with two levels. Therefore, the target identifier sequence of the target user is SID=[i1,i2].

[0053] For example, in one application scenario, a new user's interest vector, after the first level of quantization, is mapped to the cluster center with index 43 (i.e., i1=43), indicating that it belongs to the "sports and outdoor" category at the coarsest level. The residual vector between the user's interest vector and the first-level cluster center i1=43 is mapped to the cluster center with index 62 (i.e., i2=43) in the second level of quantization, indicating that it is biased towards "basketball" at a finer level. The target identifier sequence SID=[43,62] of the user is generated, which fully records the interest direction from "sports" to "basketball".

[0054] By performing residual quantization on user interest vectors at multiple levels, the continuous space of user interest vectors is transformed into discrete codeword index sequences with clear semantic hierarchical relationships. This allows the interest location of any user to be quickly indexed and matched at different granular levels through the prefix of its SID sequence. The deeper the level, the more refined the interest characterization, and the subsequent prefix matching method can clearly and efficiently determine the similar groups corresponding to the target user.

[0055] In practical applications, based on different types of interest feature data, a first interest vector related to conversion and a second interest vector related to click-through rate can be determined. When recommending content to target users, the first and / or second interest vectors can be referenced to recommend more comprehensive and interesting content to users. Specific implementation methods are described below: Based on the target user's interest feature data obtained from multiple data sources, a user interest vector is constructed, including: Interest feature data of the target user is obtained from multiple data sources. Different types of interest feature data are encoded to generate a first interest vector and a second interest vector. The first interest vector and / or the second interest vector are determined as the user's interest vector.

[0056] The first interest vector can be understood as a vector generated based on interest feature data related to the user's actual purchase or conversion behavior. It reflects the user's clear consumption preferences and has conversion-oriented semantic features. The second interest vector can be understood as a vector generated based on interest feature data related to the user's click and browsing behavior. It includes non-converted but interest-reflecting behavioral signals such as clicks, and is used to capture the user's broader potential interest distribution.

[0057] In this embodiment, interest feature data from different sources and semantic levels are processed using independent encoding branches, each mapped to a unified vector space, outputting two vector representations that maintain independent semantics. In reality, different types of interest feature data differ fundamentally in their semantic connotations and the user preference dimensions they reflect. For example, explicit consumption preferences expressed through actual purchasing behavior have high confidence and strong conversion orientation; while behavioral signals such as clicks and browsing reflect a broader range of potential user interests, containing rich information for interest exploration even without conversion. By employing a strategy of separately encoding and independently generating two interest vectors, the independent semantic space corresponding to each type of interest feature data is preserved, allowing subsequent quantification and audience mapping to flexibly select or combine different interest feature data according to the needs of the scenario.

[0058] Specifically, the system acquires target user interest feature data from multiple data sources. Based on the semantic type reflected by the interest feature data, it divides the data into conversion-oriented procurement data and interest-oriented ranking data. For the procurement data, a Transformer encoder combined with a CGC network structure is used to encode the data, generating a first interest vector (procurement embedding). For the ranking data, the system independently encodes the data using the same encoding framework, generating a second interest vector (ranking embedding). Finally, the first and second interest vectors are output as independent user interest vectors.

[0059] Of course, in practical applications, the first interest feature and the second interest feature can also be fused together, and the fused feature obtained can be used as the user interest vector of the target user. This is not limited here.

[0060] For example, in one application scenario, for a target user who has generated a few clicks on the platform but has no purchase history, the user's corresponding click behavior sequence (such as browsing several fashion notes or food videos) is encoded as fine-grained data to generate a second interest vector, which is then used as the user's interest vector. Alternatively, for a moderately active user with both historical clicks and conversion records, a first and second interest vector can be generated independently. The first interest vector reflects the user's confirmed consumption preferences verified through purchases, while the second interest vector reflects the exploratory interests shown in their clicks and browsing. These two vectors are stored independently, and different vectors can be selected as quantification inputs based on the recommendation goal (whether to pursue conversion rate or user activity).

[0061] The content recommendation method provided in the embodiments of this specification generates two interest vectors that retain their respective semantic characteristics by independently encoding different types of interest features. This allows conversion-oriented precise preferences and interest-oriented broad exploration to be used independently for subsequent quantification and audience matching, providing a data foundation for the recommendation system to flexibly select under different business objectives.

[0062] Step 204: Determine a set of similar users from the initial user set based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set.

[0063] The target identifier sequence can be understood as an identifier sequence with a hierarchical semantic structure generated for the target user through vector mapping; the initial user set can be understood as a user set composed of active users; and the similar user set can be understood as a set of users in the initial user set whose user identifier sequences meet the preset similarity conditions with the target identifier sequence.

[0064] Specifically, given the target identifier sequence corresponding to the target user, it is matched with the user identifier sequences of each initial user in the initial user set. Based on a preset similarity determination strategy (such as prefix matching, distance metric, or other similarity evaluation methods), similar users with similar interest semantics to the target user are selected from the initial user set. A similar user set is constructed based on these similar users, which serves as the target audience for subsequent content recommendations.

[0065] In one or more embodiments of this specification, a sequence knowledge graph is constructed using the user identifier sequences of each initial user, thereby enabling hierarchical and detailed searching based on the sequence knowledge graph to obtain a set of similar users. Specific implementation methods are described below: The step of determining a similar user set from the initial user set based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set includes: A sequence knowledge graph is constructed based on the user identifier sequence of each initial user in the initial user set, wherein the sequence node in the sequence knowledge graph corresponds to at least one user identifier sequence; Based on the target identifier sequence, similar sequence nodes are determined from the sequence knowledge graph; A set of similar users is determined based on the user identifier sequences corresponding to the similar sequence nodes.

[0066] In the embodiments of this specification, a pre-constructed hierarchical sequence knowledge graph is used to locate the set of similar users to which the target user belongs by matching the target user's target identifier sequence in the sequence knowledge graph. Thus, the target user's interest content is determined based on the behavioral statistical characteristics and interest distribution of each similar user.

[0067] The sequence knowledge graph can be understood as a user profile knowledge base with a coarse-to-fine hierarchical structure built based on the user identifier sequences of active users. Each level in the sequence knowledge graph corresponds to the granularity of the user group defined by a certain prefix length in the identifier sequence. Each node contains the distribution of interest tags and the TGI index for the user corresponding to that node. It should be noted that the TGI index in the sequence knowledge graph is used to measure whether the preference of a certain level of user group for a specific interest tag deviates significantly from the average level of all users, thereby verifying the accuracy and rationality of the segmented user group.

[0068] The initial users in the initial user set are active users. Active users can be understood as a set of users with sufficient behavioral data (such as clicks, browsing, interactions, etc.) within the content application platform. The richness of the behavior of active users is sufficient to support the stable generation and multi-level quantization of their interest vectors, so it can serve as the data foundation for constructing a sequence knowledge graph. The user identifier sequence can be understood as the user semantic identifier sequence (SID) corresponding to an active user, which is composed of multiple codeword indices. This user identifier sequence of active users is also generated by vector mapping of the user interest vectors of active users.

[0069] A hierarchical structure can be understood as a tree-like or hierarchical structure of a sequence knowledge graph organized according to the prefix depth of the SID sequence. For example, the first layer divides the population into the coarsest-grained categories according to the first-level index in the SID sequence, the second layer further subdivides according to the first two-level indexes, and so on. The deeper the layer, the finer the granularity of the population and the more accurate the interest characterization. A similar sequence node can be understood as a node in the sequence knowledge graph that matches the prefix of the target identifier sequence corresponding to the target user. The population represented by this node has the same prefix identifier as the target user, and is therefore considered to have similar interests to the target user at the corresponding level.

[0070] Specifically, a pre-built sequence knowledge graph is obtained. The construction process of this sequence knowledge graph is as follows: The system obtains user interest vectors from multiple active users and generates user identifier sequences for each active user using the vector mapping algorithm described above. Then, active users are grouped according to the hierarchical prefix of their user identifier sequences. For example, at the first level, users with the same first-level index i1 form the first-level sequence nodes; at the second level, users with the same indices (i1, i2) from the previous two levels form the second-level sequence nodes, and so on, increasing the SID prefix depth layer by layer to form a multi-level user segmentation from coarse-grained to fine-grained. For each level of sequence node, the system collects statistical information such as the distribution of interest tags and the TGI index of each tag for the users covered by that node. This statistical information is stored in association with the node, thus forming a complete sequence knowledge graph.

[0071] After the sequence knowledge graph is constructed, when it is necessary to locate similar sequence nodes for a target user, the target identifier sequence of the target user is obtained, and the node at the corresponding level that completely matches the prefix of the target identifier sequence is located in the sequence knowledge graph. The people represented by this node are the similar sequence nodes of the target user.

[0072] The node identifier corresponding to the similar sequence node is the user identifier sequence that constitutes the node. At least one user covered by the user identifier sequence is identified as a similar user of the target user, and a set of similar users is formed.

[0073] In practical applications, when locating the set of similar users corresponding to a target user from a sequence knowledge graph, the prefix depth can be flexibly adjusted according to the interest feature data available to the target user. That is, the richer the interest feature data, the greater the prefix depth, the finer the hierarchical granularity, and the more accurate the interest matching. Therefore, it can adapt to the granularity of interest positioning for users with different sparsity levels.

[0074] In the embodiments of this specification, a multi-level sequence knowledge graph is constructed in advance based on the user identifier sequences of all active users. This allows for the rapid location of corresponding similar sequence nodes in the sequence knowledge graph based on the prefix of the corresponding target identifier sequence for any target user, thereby obtaining statistical information such as the interest tag distribution and TGI index of the corresponding similar user set.

[0075] In one or more embodiments of this specification, a sequence knowledge graph is constructed based on a large number of active users with sufficient behavioral data. Specific implementation methods are described below: Constructing a sequence knowledge graph based on the user identifier sequences of each initial user in the initial user set, including: Based on the user behavior data of each initial user in the initial user set, construct the active interest vector corresponding to each initial user; Perform vector mapping on each active interest vector to obtain the user identifier sequence corresponding to each initial user; Based on the level identifiers of each level in each user identifier sequence, the sequence nodes of the corresponding level on the sequence knowledge graph are determined. In the case of level N, the level identifier of the Nth level is determined by the level identifiers of the previous N levels, where N≥1. The sequence knowledge graph is constructed based on sequence nodes at at least one level.

[0076] The initial user set can be understood as a group of users with sufficient behavioral data (such as clicks, browsing, interactions, etc.) within the content application platform (initial users are active users, and the initial user set is the active user set); the active interest vector can be understood as an interest vector representation constructed based on the user behavior data of active users. The active interest vector and the target user's user interest vector have the same vector dimension and semantic space.

[0077] Specifically, within the content application platform, if a user's user behavior data meets the activity criteria (such as the total number of interaction behaviors within a month ≥ 1000, the effective browsing time within 7 days ≥ 120 minutes, the number of consecutive active days within 7 days ≥ 5 days, etc., without any restrictions here), the user is marked as an active user (i.e., an initial user).

[0078] The process involves acquiring user behavior data from each active user in the active user set. This data includes historical records of user actions such as clicks, browsing, interactions, and conversions within the content application platform. Based on this user behavior data, an active interest vector is constructed for each active user. Subsequently, vector mapping is performed on the active interest vectors of each active user.

[0079] This embodiment uses a multi-level residual quantization approach for vector mapping as an example to explain the vector mapping process in detail.

[0080] In the first quantization level, K-means clustering is performed on the set of active interest vectors of all active users. In this embodiment, the number of clusters K1=1024. The active interest vector of each active user is mapped to the nearest cluster center to obtain the first-level codeword index i1, and the residual vector r1 between the active interest vector and its cluster center is calculated. In the second quantization level, K-means clustering is performed again on the set of all residual vectors r1 produced in the previous level. The number of clusters K2=1024 to obtain the second-level codeword index i2 and a new residual vector r2. This process is repeated until multiple levels of quantization are completed, with each level having 1024 clusters. Finally, the codeword indices obtained from each level are concatenated in hierarchical order to form the semantic identifier sequence SID=[i1,i2,…,i...] of the active user. n ], where n is the number of levels, which can be set according to the actual situation. After the above vector mapping, each initial user obtains a user identifier sequence.

[0081] Having obtained the user identifier sequence corresponding to each initial user, a coarse-to-fine sequence knowledge graph is constructed based on the user identifier sequence. This allows all active users within the content application platform to be systematically organized into a queryable and matchable hierarchical structure according to their interest semantics, providing a matching basis for target user group matching.

[0082] Specifically, hierarchical identifiers of different levels are extracted from the user identifier sequence of the initial user, and a sequence knowledge graph with a hierarchical structure is constructed based on these identifiers.

[0083] In this context, "level" can be understood as different scales of division in a sequence knowledge graph, ranging from coarse to fine. When N=1, it's the coarsest granularity; with K=1024, active users are divided into 1024 broad categories based on the first-level index, according to the user identifier sequence. When N=2, the granularity is further refined, dividing into 1024×1024 more refined subcategories based on the combination of the first two levels of indexes. And so on; the larger the value of N, the deeper the level, the finer the granularity of the user set's categories, and the more precise the interest characterization. The level identifier can be understood as the combination of codeword indices corresponding to the first N levels extracted from each user identifier sequence when constructing the sequence node of the Nth level.

[0084] A sequence node can be understood as a set of user combinations and associated statistical information of users with the same level identifier at a certain level. The value of N can be flexibly set according to business needs in practical applications. Usually, N≥1. Considering that the conflict rate has become stable and the generalization performance is good after the first two levels of quantization, N=2 is usually taken in practical applications.

[0085] Specifically, obtain the user identifier sequence of all active users. The user identifier sequence of each active user is SID=[i1,i2,…,i...]. n Then, for each level N, extract the first N level index combination from each user identifier sequence as the target level identifier for that user at that granular level. That is, when N=1, extract [i1], when N=2, extract [i1,i2], when N=3, extract [i1,i2,i3], and so on.

[0086] At each level N, users with the same level identifier are grouped into the same sequence node, forming a user segmentation at that level. For example, at level N=2, all users whose first two level indices are "43_602" form one sequence node, all users whose first two level indices are "43_603" form another sequence node, and so on. After completing the sequence node segmentation at each level, statistical information such as the distribution of group tag preferences and the TGI index of each interest tag for each sequence node is collected. The node identifier (the identifier sequence corresponding to the sequence node), the level to which it belongs, the user set, and the statistical information are stored together.

[0087] All sequence nodes at all levels are organized into a tree-like or hierarchical structure with a progressive relationship, in order from coarse to fine. That is, a sequence node at one level is subdivided into multiple child nodes at the next level. Parent nodes and child nodes are associated through the inclusion relationship of the SID prefix, thus forming a complete sequence knowledge graph.

[0088] For example, in one application scenario, based on the user identifier sequence of all active users, two levels of sequence nodes are constructed, from N=1 to N=2. At the N=1 level, the first-level index i1 of each user is extracted as the level identifier, and all users with i1=43 are assigned to the "43" sequence node. For example, this sequence node represents the coarsest-grained interest category of "sports and outdoor", covering all active users with a first-level index of 43.

[0089] At level N=2, the first two levels of index combination (i1, i2) for each user are extracted as level identifiers. Users with the first two levels index "43_602" are assigned to a finer-grained node (e.g., representing "Sports & Outdoors → Basketball"), while users with the index "43_603" are assigned to another node (e.g., representing "Sports & Outdoors → Football"). In other words, the parent node "43" is subdivided into multiple child nodes at level N=2, each corresponding to a more specific interest direction. This constructs a sequence knowledge graph with two levels of structure based on the sequence nodes at both levels.

[0090] The content recommendation method provided in the embodiments of this specification extracts corresponding hierarchical identifiers from a large number of user identifier sequences according to different levels of N values, and generates multi-level sequence nodes from coarse-grained to fine-grained, so that the sequence knowledge graph has hierarchical indexing capabilities, thereby adapting to richer matching scenarios.

[0091] In one or more embodiments of this specification, in a sequence knowledge graph with a multi-level structure, similar sequence nodes corresponding to the target identifier sequence are determined from the sequence nodes of the corresponding level based on prefix matching. Specific implementation methods are described below: Based on the target identifier sequence, similar sequence nodes are determined from the sequence knowledge graph, including: The target level corresponding to the target identifier sequence is determined, and similar sequence nodes corresponding to the target identifier sequence are determined from the sequence nodes corresponding to the target level in the sequence knowledge graph based on prefix matching.

[0092] In this context, the target level can be understood as the level corresponding to the SID prefix depth L used for prefix matching. This prefix depth L determines at which level of the sequence knowledge graph similar sequence nodes are located. Prefix matching can be understood as comparing the first L levels of the target user's target identifier sequence with the sequence node identifiers at the corresponding L levels in the sequence knowledge graph. When the combination of the target user's first L levels of indices is exactly the same as the sequence node identifier of a certain sequence node, that node is a matched similar sequence node. A similar sequence node can be understood as a sequence node in the target level L of the sequence knowledge graph whose node identifier is exactly the same as the first L levels of the target identifier sequence. The set of users represented by this node is the set of similar users of the target user.

[0093] Specifically, the sequence knowledge graph is constructed based on the user identifier sequence of active users. The sequence knowledge graph has N levels, while the target identifier sequence corresponds to target users with sparse behavior. The number of levels of the target identifier sequence is usually less than N. The target level L corresponding to the target identifier sequence is determined for prefix matching.

[0094] In practice, the value of L can be dynamically determined based on the richness of behavioral data available to the target user. For example, for target users with relatively abundant behavioral data, L can be set to a deeper value (such as L=2 or L=3) to locate similar users at a finer granular level and obtain more accurate interest matching. For target users with extremely scarce behavioral data, L can be set to a shallower value (such as L=1) to ensure that a set of similar users can still be located at a coarser granular level, avoiding no matching results due to overly strict matching conditions.

[0095] After determining the value of L, the first L level index combination of the target identifier sequence is extracted as the matching index. Then, the target level N=L is located in the sequence knowledge graph, and all sequence nodes under the target level are traversed. The node identifier of each sequence node (also a combination of the first L level SID index) is compared with the matching index one by one. When a node identifier is found to be completely consistent with the matching index, the sequence node is identified as a similar sequence node of the target user. All active users covered by this similar sequence node constitute the similar user set of the target user. That is, these active users have the same prefix identifier as the target user at the selected L level granularity, and are therefore considered to have similar interests to the target user at this target level.

[0096] For example, in one application scenario, the target identifier sequence for a user with sparse behavior is [43,602]. The target level corresponding to this target identifier sequence is determined to be 2. Therefore, the first two level index combination "43_602" of the target identifier sequence is extracted. Then, the node with the node identifier "43_602" is searched in the second-level sequence nodes (N=2) of the sequence knowledge graph. If the sequence node is found, it is identified as a similar sequence node, and the active users covered by the sequence node are identified as similar users of the target user.

[0097] For example, in another application scenario, for a new user, behavioral data is extremely scarce, so its target identifier sequence is

[12] . At this time, the target level is 1. The first-level index 12 is used to locate similar sequence nodes on the first level of the sequence knowledge graph. The similar sequence nodes cover all active users with the first-level index "12". That is, although the granularity is relatively coarse and the interest characterization is not refined enough, it can still provide a similar sequence node on a generalized interest category for new users, obtain the corresponding similar representation sequence, and avoid completely missing the support of similar users due to matching failure.

[0098] The content recommendation method provided in the embodiments of this specification determines the prefix matching depth L of the target identifier sequence corresponding to the target user, and locates similar sequence nodes at the corresponding level in the sequence knowledge graph through precise prefix matching. This enables the target user to determine the set of similar users at a level that is commensurate with the richness of their behavioral data. The system can flexibly adjust the value of L according to the actual behavioral data that the target user can obtain, thereby achieving a dynamic balance between matching accuracy and matching coverage.

[0099] Step 206: Determine the historical behavior data of the target user based on the user behavior data corresponding to the set of similar users.

[0100] Among them, the user behavior data corresponding to the similar user set can be understood as the set of historical behavior records generated by each similar user within the content application platform, including behavioral data such as content exposure, click behavior, browsing duration, and interactive operations.

[0101] In this embodiment, the target user's historical behavior data does not refer to the target user's actual historical behavior, but rather to virtual historical behavior data generated by injecting user behavior data corresponding to a set of similar users into the target user's behavior sequence.

[0102] Specifically, when a similar sequence node corresponding to the target user is located from the sequence knowledge graph, a set of similar users based on the active users covered by the similar sequence node can be obtained. The user behavior data of each similar user in the set of similar users is then inserted into the historical behavior sequence of the target user in order to generate the historical behavior data of the target user.

[0103] By generating historical behavior data of target users, target users who originally had no behavior or sparse behavior now have historical behavior records in form. Downstream recommendation systems can also obtain the historical behavior data of the target users and determine the recommended content to be recommended to the target users.

[0104] By injecting user behavior data of active users corresponding to similar sequence nodes into the historical behavior sequence of target users, users who originally lacked behavioral data during cold starts or with sparse behavior have usable historical behavioral data in the downstream recommendation system. Subsequent recommendation links make personalized recommendations based on this historical behavioral data, thereby effectively avoiding the common problem of low matching degree with individual's true interests in the cold start stage.

[0105] In one or more embodiments of this specification, a set of exposed content is selected from user behavior data corresponding to a set of similar users, and a set of candidate content is further obtained from the set of exposed content. This allows for the injection of highly interesting content into the target user's historical behavior sequence, achieving more relevant content recommendations. Specific implementation methods are described below: Based on the user behavior data corresponding to the set of similar users, the historical behavior data of the target user is determined, including: Obtain user behavior data corresponding to each similar user in the set of similar users; Obtain an exposure content set from the user behavior data, and determine a candidate content set from the exposure content set; Historical behavior data of the target user is generated based on the set of candidate content.

[0106] The exposed content set can be understood as the set of content that was pushed and displayed to the initial user within a preset time window; the candidate content set can be understood as multiple exposed contents selected from the exposed content set according to a preset filtering strategy, and the content in the candidate content set is used to inject into the target user's historical behavior sequence.

[0107] Specifically, after locating the similar sequence node corresponding to the target user from the sequence knowledge graph, the user identifier sequence of all active users covered by the similar sequence node is obtained, and then the user behavior data corresponding to these active users is obtained, including exposure records, click records, interaction records and other behavior data; subsequently, the set of exposure content within a preset time window is extracted from these user behavior data, the set of exposure content is filtered, and a set of candidate content is determined from the set of exposure content.

[0108] After the candidate content set is determined, the candidate content in the candidate content set is injected into the target user's last N historical behavior sequence to generate the target user's historical behavior data. Specifically, the insertion can be done by inserting it at the beginning of the sequence to emphasize its importance, or by mixing and sorting it with the target user's existing behavior according to a preset insertion strategy. No restrictions are imposed here.

[0109] In specific implementation, a candidate content set is determined from the exposed content set, including: The click-through rate of each exposed content in the exposed content set is sorted, and the candidate content set is obtained based on the sorting results.

[0110] Click-through rate (CTR) ranking can be understood as sorting each exposed content from highest to lowest CTR among the similar user set. The candidate content set can be understood as selecting the top-ranked exposed content based on the CTR ranking results; the content in the candidate content set represents high-interest content among the similar user set.

[0111] Specifically, for each piece of content in the exposure content set, the number of exposures and clicks in the similar user set are counted, and the click-through rate of each piece of content is calculated. The content is then sorted from high to low click-through rate, and a preset number of exposures with the highest click-through rate are selected from the sorted results to form a candidate content set.

[0112] For example, the similar sequence node corresponding to the target user was identified as the node "43_602". This node covers approximately 50,000 active users (i.e., a set of similar users). The content exposure of these 50,000 users in the past month was obtained, for example, involving approximately 8,000 different pieces of content, and the click rate of each piece of content was calculated.

[0113] Since this embodiment obtains exposure content sets corresponding to multiple active users, these sets contain exposure content of multiple types and dimensions. For example, the exposure content set includes equipment-related content, technical content, and news content. The top 100 articles with the highest click-through rates are selected from the exposure content set as a candidate content set. The candidate content in this set also covers multiple dimensions such as equipment reviews, skill tutorials, and event information, rather than being concentrated on one type or dimension. These 100 articles are then injected into the target user's last N sequence in descending order of click-through rate.

[0114] In this embodiment, by acquiring user behavior data of similar user sets, the diversified interest distribution corresponding to similar users is transferred to individual target users, avoiding the problem of single recommendation caused by relying solely on a single interest tag or single category of content. This effectively ensures that users in the initial stage can obtain diverse and widely covered recommended content.

[0115] The content recommendation method provided in this specification obtains active users covered by similar sequence nodes and their corresponding user behavior data, and then extracts an exposure content set from the user behavior data. This exposure content set contains multiple types of content, and a candidate content set is obtained based on click-through rate. This ensures that the content injected into the historical behavior sequence corresponding to the target user is highly recognized content that has been verified by the group, thereby improving the recommendation effect of the recommended content.

[0116] Step 208: Determine the recommended content corresponding to the target user based on the historical behavior data.

[0117] Recommended content can be understood as personalized content that is ultimately displayed to the target user after being processed through the recommendation process (including recall and ranking).

[0118] Specifically, by generating historical behavior data of target users, target users who originally had no behavior or sparse behavior now have historical behavior records in form. The downstream recommendation system performs content recall and ranking based on the historical behavior data of the target users, and determines the recommended content to be recommended to the target users based on the ranked recommended content.

[0119] In this implementation, by generating the historical behavior sequence of the target user, users who originally lacked behavioral data or had sparse behavior have usable historical behavioral data in the downstream recommendation system. Subsequent recommendation links make personalized recommendations based on this historical behavioral data, thereby effectively avoiding the common problem of low matching degree with the individual's true interests in the cold start stage.

[0120] In one or more embodiments of this specification, determining the recommended content corresponding to the target user based on the historical behavior data includes: Based on the historical behavior data, content is retrieved and sorted to determine the recommended content to be recommended to the target user.

[0121] Content recall and ranking can be understood as the process of retrieving relevant candidate content from the full content library based on historical behavioral data, scoring and ranking the recall results to determine the final recommended content.

[0122] The target user's historical behavior data is sent to the downstream recommendation chain. For example, the recommendation chain uses a content recommendation model to perform content recall (including item-to-item vector recall and user-to-item vector recall, which are used to recall more relevant content based on candidate content) and personalized ranking, and determines the ranked content list as the recommended content to be recommended to the target user.

[0123] In practical applications, by using item-to-item vector recall and user-to-item vector recall, the recommended content is not limited to content that similar users have already been exposed to, but can cover more semantically related content through vector similarity, effectively enriching the diversity of recommendation results.

[0124] Item-to-item vector recall refers to recommending similar items based on similar items. In a recommendation system, each piece of content (such as notes, products, or videos) is represented as a fixed-dimensional content vector. This vector is obtained by encoding the text, image, and tag features of the content through a deep learning model, reflecting the content's position in the semantic space. In this embodiment, content vectors corresponding to multiple exposed content can be obtained. Then, in a pre-built content vector index library, similar content vectors are found by calculating vector similarity (such as cosine similarity). The similar content corresponding to these similar content vectors is used as the recall result, realizing recommendation expansion by utilizing the semantic relationships between content itself.

[0125] User-to-item vector recall refers to matching items based on user preferences. In a recommendation system, each user is represented as a fixed-dimensional user vector. In this embodiment, the user vector is determined based on the user's behavior data (for the target user, the user vector is determined through generated historical behavior data), reflecting the user's overall position in the interest semantic space. When the system needs to recommend content to a target user, it generates the target user's user vector and then searches for similar content vectors that are close to the user vector in the content vector index. The similar content corresponding to the similar content vectors is used as the recall result, realizing a global search from the entire content library for content that matches the user's overall interest distribution. Item-to-item vector recall ensures the accuracy and relevance of the recommended content, while user-to-item vector recall ensures the coverage and exploratory nature of the recommended content.

[0126] For example, following the previous example, the downstream recommendation chain can obtain recalled content based on the target user's last N sequence through item-to-item vector recall and user-to-item vector recall. Among them, item-to-item recall can further recall other similar content based on the content vectors corresponding to 100 pieces of content, such as expanding from basketball shoe reviews to sports protective gear recommendations. User-to-item recall can calculate user vectors based on the historical behavior sequence of the target user after injection, and match more candidate content according to the user vectors.

[0127] The retrieved content is sorted and the resulting recommended content can include various types of content such as equipment reviews, sports protective gear recommendations, skill tutorials, event information, and training equipment, rather than being limited to a single type of content.

[0128] By injecting historical behavioral data of similar user groups into target users and using this data for content retrieval and ranking, the recommended content is not limited to content already exposed by similar groups. Instead, it can expand the retrieval scope from two dimensions: content semantic association and overall user interest distribution, through item-to-item vector retrieval and user-to-item vector retrieval, respectively. This effectively enriches the diversity and coverage of recommendation results and avoids the problem of limited recommendation results caused by relying solely on limited candidate content for retrieval.

[0129] The content recommendation method provided in the embodiments of this specification, through a vector recall mechanism, enables the recommended content to not only ensure its relevance to verified high-interest content of similar users, but also to cover more semantically related content through vector similarity matching. This achieves a balance between accuracy and diversity, effectively improving the richness of recommended content and personalized matching effect in cold start and sparse behavior scenarios.

[0130] In one or more embodiments of this specification, while historical behavioral data effectively compensates for the lack of individual behavioral data during the cold start phase, this historical behavioral data actually indicates the actual preferences of similar groups. However, as target users continuously generate real clicks, browsing, and interactions on the content application platform, their individual behavioral data gradually accumulates and can reflect their true interests. Therefore, this embodiment uses a dynamic transition mechanism to gradually migrate the recommended behavioral basis from historical behavioral data to the actual individual behavioral data generated by the target user. Specific implementation methods are described below: After determining the recommended content for the target user based on the historical behavior data, the method further includes: Obtain individual behavioral data of the target user based on the recommended content; Upon receiving a content recommendation request from the target user, the system determines the target content to recommend to the target user based on the target user's historical behavior data and individual behavior data.

[0131] Individual behavior data can be understood as the actual behavior records of target users after receiving recommended content, including behavioral signals that reflect an individual's true interests, such as clicks, browsing time, interactive operations, and conversions; content recommendation requests can be understood as recommendation requests initiated by target users when they trigger content refresh or enter a page within a content application platform.

[0132] Specifically, after determining the recommended content to be recommended to the target user through historical behavior data, the individual behavior data of the target user based on the recommended content can be continuously monitored and acquired within the content application platform. This individual behavior data includes the target user's click behavior, browsing duration, interaction operations, conversion behavior, etc., and this individual behavior data will be recorded in the target user's behavior sequence.

[0133] When a user enters the content application platform or clicks refresh, a content recommendation request is triggered. Each time a content recommendation request is triggered, the system determines the target content to recommend to the target user based on the target user's historical behavior data and individual behavior data. By simultaneously referring to historical behavior data and individual behavior data, the system compensates for the lack of individual behavior data during the cold start phase by using historical behavior data, while gradually incorporating real individual behavior signals into the recommendation criteria.

[0134] In practice, as more and more individual behavior data is acquired, the weights corresponding to historical behavior data and individual behavior data can be dynamically adjusted. The specific implementation method is as follows: Based on the target user's historical behavior data and individual behavior data, the target content to be recommended to the target user is determined, including: Determine a first initial weight for the historical behavioral data and a second initial weight for the individual behavioral data; The first initial weight and the second initial weight are dynamically adjusted according to the time decay factor to obtain the first updated weight of the historical behavior data and the second updated weight of the individual behavior data. The time decay factor increases monotonically with the accumulation time of the target user's behavior. The first updated weight is negatively correlated with the time decay factor, and the second updated weight is positively correlated with the time decay factor. Based on the historical behavior data, the first update weight, the individual behavior data, and the second update weight, the target content to be recommended to the target user is determined.

[0135] The first initial weight can be understood as the initial weight of historical behavioral data in the recommendation strategy. This weight is higher in the cold start phase to achieve content recommendation based on historical behavioral data. The second initial weight can be understood as the initial weight of individual behavioral data in the recommendation strategy. In the cold start phase, because individual behavioral data has not yet been fully accumulated, the second initial weight is lower.

[0136] The time decay factor is a control parameter that monotonically increases with the duration of accumulated user behavior. It characterizes the degree to which a user evolves from the cold start phase to the behavior enrichment phase, and this time decay factor gradually increases over time. The first update weight can be understood as the weight of historical behavior data adjusted by the time decay factor. This first update weight is negatively correlated with the time decay factor, meaning it gradually decreases as the user's behavior accumulation time increases. The second update weight can be understood as the weight of individual behavior data adjusted by the time decay factor. This second update weight is positively correlated with the time decay factor, meaning it gradually increases as the user's behavior accumulation time increases.

[0137] Specifically, when a user enters the content application platform or clicks refresh, a content recommendation request is triggered. Each time a content recommendation request is triggered, the system determines a time decay factor *t* for the current moment. This time decay factor monotonically increases with the accumulated behavior time of the target user since their first entry into the platform; that is, the more active days the user has and the richer their accumulated behavior, the larger the value of *t*. Simultaneously, a first initial weight (usually set to a relatively high value, such as 0.8 or 0.9) and a second initial weight (usually set to a relatively low value, such as 0.1 or 0.2) for individual behavior data are determined, and their sum is 1. The two initial weights are dynamically adjusted based on the time decay factor *t*. In practice, as the time decay factor *t* increases, the value of the first initial weight is decreased to obtain the first updated weight, and the value of the second initial weight is increased to obtain the second updated weight.

[0138] By weighting and fusing historical behavioral data with the first update weight and individual behavioral data with the second update weight, reference behavioral data is obtained for content recommendation. The reference behavioral data is comprehensively considered in the content recall and ranking stages to generate the final target content recommended to the target users.

[0139] In the early stages of a cold start, the initial weights are relatively high, and the recommendation results are mainly dominated by the injected historical behavioral data. However, as users' real behaviors gradually accumulate, the target content recommended to users is determined by the actual behavior of individual users. In practical applications, when the time decay factor t gradually increases from 0 to a preset threshold (e.g., t=1), the first update weight can monotonically decrease to 0, and the second update weight can monotonically increase to 1.

[0140] For example, a new user with no individual behavioral data has no data when they enter the content application platform on the first day. At this time, the time decay factor t=0, the system sets the first initial weight to 0.9 and the second initial weight to 0.1, and the recommended content is almost entirely determined by the aforementioned historical behavioral data.

[0141] After a week of use, the user generated approximately 50 clicks. At this point, the time decay factor t=0.3, and the system adjusted the first update weight to 0.6 and the second update weight to 0.4. This means that the recommendation strategy began to place more emphasis on the types of content that the user had clicked (such as the user frequently clicking on fashion-related notes within a week). The proportion of fashion-related content in the recommendation list gradually increased, while the initially injected sports-related content gradually decreased.

[0142] After a month of use, the user generated hundreds of clicks, the time decay factor t=0.9, the first update weight dropped to 0.1, and the second update weight rose to 0.9. At this point, the recommended content was almost entirely dominated by the user's actual behavior. When the richness of user behavior continued to increase and the time decay factor reached the preset threshold t=1, the first update weight dropped to 0. At this point, the main recommendation logic of the content application platform recommended target content to the user. The main recommendation logic is based entirely on the user's own rich behavioral data for personalized recommendations.

[0143] The content recommendation method provided in the embodiments of this specification achieves automatic adjustment of the weight of historical behavior data and the user's actual individual behavior data through a dynamic adjustment mechanism in which the time decay factor increases monotonically with the accumulation time of user behavior. This ensures that the recommendation system makes recommendations based on the most appropriate behavior at different stages of user activity, and realizes a seamless transition from "prior recommendation based on similar groups" to "personalized recommendation based on individual real preferences".

[0144] See Figure 3 , Figure 3 A flowchart illustrating the vector mapping process in a content recommendation method provided in one embodiment of this specification is shown.

[0145] This content recommendation method is applicable to recommendation optimization scenarios for cold-start users and users with sparse behavior. Specifically, the embodiments in this specification can be applied to scenarios such as new user cold start, interest recovery for returning users, and recommendation optimization for users with sparse behavior. In the new user cold start scenario, for newly registered users with no behavior, interest profiles are quickly established using limited signals such as registration information and device characteristics. In the returning user interest recovery scenario, for returning users who have been inactive for a long time and whose historical behavior has expired, interest representations are re-established using multi-source signals. In the sparse behavior user recommendation optimization scenario, for low-activity users with insufficient behavioral data and unstable model interest estimation, interest content is supplemented by similar user groups.

[0146] The vector mapping process in content recommendation methods includes the following steps: S1: Fusion of multi-source interest features.

[0147] Specifically, multi-source interest feature data of users is collected from multiple data sources such as registration information, device characteristics, third-party procurement signals, RTA mining, and H5 scenarios.

[0148] The registration information includes the attribute information provided by the user when registering for the content application platform. For example, when a new user downloads and opens a content application, the registration process will require them to select the "content categories they are interested in". Suppose a user actively selects the "technology" and "automobile" interest tags. At this time, "technology" and "automobile" are included in the registration information. Therefore, "technology" and "automobile" can be used as the user's interest feature data.

[0149] Device characteristics include the device model, operating system, and manufacturer information of the user's corresponding client, which are not limited here; when the system captures that the device used by the user is "Device A" and the operating system is the latest version, combined with the manufacturer information platform, it can be inferred that the user is likely to have high spending power and a willingness to try new things. Therefore, high-end device accessories can be used as user interest characteristic data.

[0150] Third-party procurement signals include third-party points of interest, application lists, and manufacturer interests from external DMP data. For example, if a platform obtains a user's application list from a third-party DMP (data management platform) and finds that the user has "K Fitness," "H Health," and "Sports" apps installed on their smart device, even if the user's registration information on the platform does not include "sports," it can still determine that the user is a "potential sports enthusiast" and therefore use "sports" as an interest feature data.

[0151] RTA refers to user interest data mined using Real-Time APIs (Application Programming Interfaces). For example, when an e-commerce platform sends an advertising bidding request to a content platform via the RTA interface, the interface will send back real-time data showing that the user recently "searched for cameras and browsed 3 reviews" on the e-commerce platform. Upon receiving this information, the content platform can then use "camera" and "shooting equipment" as interest feature data for that user.

[0152] The H5 scenario includes records of weakly correlated user behaviors within the H5 context. For example, a user clicks on a news article titled "Job Interview Skills" (H5 page) in the W application, but stays for a very short time (weak correlation) and does not like or comment. However, if the user has recently clicked on three similar "interview skills" H5 articles, the system will aggregate these signals and infer that the user is looking for a job. Therefore, when this user visits a content platform, content related to work and interviews can be used as the user's interest feature data.

[0153] Before encoding the aforementioned multi-source interest feature data, it is necessary to classify and integrate it according to business applications. Purchase signals can be obtained from third-party purchase signals and RTAs. These signals originate from external platform purchases and real-time interfaces, focusing on characterizing users' external preferences and immediate intentions, and are used for purchase scenarios such as ad targeting and audience expansion. Fine-grained ranking signals cover registration information, device characteristics, and weakly correlated H5 behavioral data. This data comes from the platform's own channels or weak behavioral traces of users within the platform, focusing more on reflecting users' static attributes and potential interests within the platform, and can be used in the fine-grained ranking stage of the recommendation system.

[0154] Based on the above classification, the categorized purchasing signals and ranking signals are input into the Transformer encoder and the CGC (Customized Gate Control) network structure for independent encoding processing, and finally generate two types of user representations: purchasing interest vector embedding and ranking interest vector embedding.

[0155] Purchased interest vectors or finely ranked interest vectors are used as user interest features for multi-level quantification.

[0156] S2: Quantization based on residual quantization K-means algorithm.

[0157] In this embodiment, the 8-level residual quantization K-means algorithm (RQ-KMeans) is used to perform vector mapping on the user interest vector.

[0158] In the first quantization level (Layer 0), K-means clustering is performed on the user interest vectors to construct the first-level codebook, with a codebook size of K1=1024, obtaining the codeword index i1 and residual vector r1 corresponding to each user interest vector in this level; in the subsequent j-th quantization level (j=2,3,...,8), the residual vector set r1 of the previous level is clustered... (j-1) Perform K-means clustering to construct the j-th level codebook, with a codebook size of K. j =1024, thus obtaining the codeword index i of this level. j and residual vector r j r2.

[0159] Repeat the above steps until all 8 levels of quantization are completed (i.e., Layer 0 to Layer 7, a total of 8 levels). Figure 3 The intermediate layers are represented by “Layer2...Layer7”. Ultimately, each user can be represented as a semantic identifier sequence SID consisting of 8 levels of codeword indexes, where the value of each codeword index ranges from 0 to 1023.

[0160] The above quantification process is executed independently for the purchase vector and the fine-ranking vector, meaning that a user can have two sets of identification systems at the same time: the purchase SID sequence and the fine-ranking SID sequence.

[0161] S3: Construct a hierarchical knowledge graph of the population.

[0162] Based on user behavior data of all active users and corresponding user identifier sequences (SIDs), active users are grouped according to hierarchical prefixes. In this embodiment, the SID sequence consists of an 8-level codeword index, with each level representing the user's interest at a specific semantic granularity. Therefore, using different prefix lengths of the SIDs as the grouping basis naturally forms a multi-level user segmentation system ranging from broad to fine.

[0163] Specifically, the first-level index i1 of the SID is used as the coarsest level of user segmentation. Since the first-level index is generated by the first round of K-means clustering, it corresponds to the broad category of user interests at the highest semantic level. For example, all users are roughly divided into top-level categories such as "tech enthusiasts," "fashion enthusiasts," and "sports and fitness enthusiasts." The first two levels of the SID index (i1, i2) are used as the second level for further subdivision. After superimposing the second-level index on the first-level broad categories, the user group is split into more specific subcategories, such as "tech enthusiasts" being further subdivided into subgroups such as "digital product enthusiasts" and "software developers."

[0164] By analogy, the index depth of SIDs is increased layer by layer. With each additional level of indexing, the user group is further broken down according to the semantic dimensions of the next level, thus forming a multi-level user structure ranging from broad interest categories to fine-grained interest tags. As the levels deepen, the user scale covered by each sequence node gradually decreases, but the consistency of interest characteristics within the user group increases accordingly. That is, the deeper the level, the more concentrated the distribution of the interest TGI (Target Group Index) of the user group at that node, and the higher the accuracy of characterization.

[0165] For each level of sequence node, statistical information such as its group tags and interest TGI distribution is collected to form a queryable and matchable population knowledge graph (i.e., the sequence knowledge graph in the above embodiment). Specifically, for the first-level node, the distribution of the most representative content categories and user profiles in this coarse-grained population can be collected; for the second-level and deeper nodes, more refined features such as tag preferences, content consumption time distribution, and active time periods can be collected.

[0166] When the online service receives a matching request from a target user, it only needs to extract the user's SID prefix to directly index the corresponding level of sequence node in the sequence knowledge graph and instantly obtain all statistical information under that node, thereby efficiently supporting subsequent similar user content mining and interest migration operations.

[0167] See Figure 4 , Figure 4 This document illustrates a flowchart of the interest migration process in a content recommendation method provided in one embodiment of this specification.

[0168] This process undertakes Figure 3 The completed population knowledge graph (i.e., sequence knowledge graph) is used to perform interest transfer and recommendation optimization for the target users to be recommended (users with no behavior or sparse behavior).

[0169] S4: Semantic identifier mapping and similar group packaging construction.

[0170] For target users to be recommended (users with no behavior or sparse behavior), limited semantic signals such as the target user's registration information, device characteristics, and initial interactions are utilized to... Figure 3 Steps S1 and S2 generate the semantic identifier sequence SID_target of the target user (i.e., the target identifier sequence in the above embodiment).

[0171] Extract the first L-level index of SID_target (L≥2, usually L=2), locate the corresponding level node in the sequence knowledge graph through precise prefix matching, and use the matched level node as the similar sequence node corresponding to the target user. Use the users covered by the similar sequence node as the similar user group of the target user (i.e., the similar user set in the above embodiment). That is, select the user set that is exactly the same as the first L-level index of SID_target to form the similar user group of the target user.

[0172] The core idea of ​​prefix matching is that the higher the prefix level of SID (i.e., the smaller the L value), the coarser the semantic granularity, the larger the user group that can be matched, but the accuracy is relatively low. When L is 2, a better balance can be achieved between coverage and matching accuracy. It can ensure that a sufficient number of similar users are matched to form a statistically effective similar group, and also ensure that these users are highly consistent with the target users in the core semantic dimension.

[0173] In this embodiment, since K in the sequence knowledge graph is 1024, the two-layer index matching can accurately locate similar groups of people in a space of 1024×1024=1,048,576 combinations. The codebook conflict rate is as low as 0.0035% (purchased SID) and 3.19% (fine-sorted SID), which can ensure the accuracy and interpretability of the group matching.

[0174] Specifically, K=1024 means that each layer of the sequence knowledge graph has 1024 independent semantic clusters, while the extremely low codebook conflict rate indicates that the probability of two semantically dissimilar SIDs being incorrectly mapped to the same sequence node is extremely low.

[0175] S5: High click-through rate content mining.

[0176] In a similar audience segment, extract the content that was exposed to that audience within a preset time window (such as the most recent month) to form an exposure content set; calculate the click-through rate (CTR) of each piece of content in the exposure content set, and sort them in descending order of CTR from high to low; select the top-ranked content (i.e., high CTR content) as a candidate content set.

[0177] In practical applications, if content achieves a higher click-through rate (CTR) within a similar audience segment, it indicates that the content is more attractive to that group, meaning that the content's theme, style, and expression better align with the group's collective preferences. Therefore, candidate content selected in descending CTR order is essentially content with high general appeal, validated by the actual clicking behavior of this similar audience. This content can serve as an effective representative of the audience's interests and preferences, and can be used in subsequent personalized recommendations.

[0178] S6: Interest Transfer and Last N Sequence Injection.

[0179] High-CTR content from the candidate content set is inserted sequentially into the target user's last N historical behavior sequence, completing the migration of interests from similar groups to the target individual.

[0180] Specifically, each piece of content in the candidate content set has already been prioritized by CTR sorting in the previous step. When inserting into the last N sequence, it can be added sequentially from high to low priority, so that the position of the migrated content in the behavior sequence reflects the relative magnitude of interest intensity. For users with no behavior whose original behavior sequence is empty or extremely short, this operation constructs an initial behavior sequence for them, supplementing their user behavior data. For users with some behavior but sparse behavior, this operation is equivalent to expanding their actual behavior sequence with reliable supplementary interest data to compensate for the information sparsity problem caused by insufficient individual behavior.

[0181] The downstream personalized ranking strategy is based on the last N sequence. When candidate content migrated from similar users is successfully injected into the last N sequence, it participates in model inference along with the user's original behavior records, thereby indirectly affecting the downstream scoring results. It should be noted that the insertion operation does not disrupt the temporal logic of the original last N sequence. Migrated content is usually inserted sequentially or appended uniformly to ensure the overall usability of the sequence structure.

[0182] In this embodiment, item-to-item vector recall and user-to-item vector recall can cover more candidate content, effectively expanding the breadth and relevance of the recommendation pool. Item-to-item recall uses each item in the last N sequence as a seed, retrieving other semantically similar content from the global content library based on vector similarity between items (such as cosine distance or inner product). User-to-item recall, on the other hand, uses the target user's user vector as the query condition, retrieving content from the content vector index that best matches the user's overall interests.

[0183] The two recall methods complement each other in their retrieval logic: item-to-item focuses on "content similar to content that the user is already interested in," while user-to-item focuses on "content that matches the user's overall profile." Combining the two ensures that the recalled content aligns with the user's existing interests while also broadening the boundaries of interests by including content that is not directly similar to a single seed content but aligns with the user's overall interests in the recommendation pool. Ultimately, both methods contribute to improving the overall coverage and effectiveness of the recommendation system.

[0184] S7: Dynamic interest decay and individualized transition.

[0185] As subsequent user behavior continues to accumulate, the system gradually transitions from prior content of the target audience (i.e., the set of candidate content in the aforementioned implementation) to precise individual modeling. In the early stages of cold start, the target user's own behavioral records are extremely limited. At this time, the recommendation strategy relies heavily on prior content of the target audience (i.e., the set of candidate content) migrated from similar audience groups to fill the information gaps. However, as users continuously generate interactions such as browsing, clicking, collecting, and commenting within the platform, their individual behavioral data becomes increasingly rich, and the individual behavioral signals themselves gradually acquire sufficient statistical significance. Therefore, it is natural to shift the weight of recommendation decisions from "similar audience groups" to "individual target users."

[0186] Specifically, in this embodiment, the time decay factor gradually increases over time. Through dynamic adjustment of the time decay factor, the weight of the injected prior content will gradually decrease over time, while the weight of individual behaviors will gradually increase as individual behavior data accumulates. The time decay factor t represents the time elapsed since the migration behavior occurred. It should be noted that after the target user completes the initial interaction, the impact of the injected prior content on subsequent recommendations gradually diminishes, avoiding recommendation effect fluctuations caused by a sudden drop in weight.

[0187] The weight of individual behaviors also gradually increases based on the number of user cumulative behaviors. Whenever a user generates a new effective interaction within the platform (such as playing a full video or clicking), the content features corresponding to that interaction behavior are given a higher confidence level.

[0188] Through the aforementioned dynamic adjustment mechanism of dual weights, a smooth transition from group prior content to precise individual modeling is achieved. The core advantages of this mechanism lie in its continuity and adaptability. Continuity is reflected in the gradual rather than abrupt changes in weights, ensuring a natural transition in the user experience between the cold start, transition, and maturity stages, without any sudden changes in recommendation style perceived at any given moment. Adaptability is reflected in the transition rate being entirely determined by the user's own activity level. Active users accumulate individual behavior quickly, and the group prior weights also fade quickly, allowing for a faster transition to the individual customization stage. Inactive users accumulate individual behavior slowly, and the group prior weights fade slowly, causing the system to rely on similar user groups for a relatively longer period to maintain recommendation quality. This ensures that each type of user receives the optimal recommendation effect commensurate with the current data completeness at any stage of their lifecycle.

[0189] In this embodiment, the recommendation strategy can evolve seamlessly from a cold start phase with no behavior, to a transition phase with sparse behavior, and then to a mature phase with rich behavior, without manual intervention or version switching.

[0190] The content recommendation method provided in this specification enriches the behavioral representation of target users by acquiring user behavior data from multiple data sources. Even when the target user's own behavioral data is sparse, it can construct a relatively comprehensive set of user interest features, avoiding complete reliance on popular recommendations due to insufficient data and effectively improving the matching degree between recommended content and individual's true interests. By quantifying user interest features at multiple levels, a target identifier sequence containing identifiers of different granular levels is generated, enabling the recommendation system to understand users layer by layer from macro-interest categories to micro-interest preferences. This balances the breadth and specificity of user interests during the recommendation process, improving the granularity of recommendations. The sequence knowledge graph constructed based on active users has a hierarchical structure and can... By mapping the target user's target identifier sequence to similar sequence nodes and supplementing the target user's historical behavior information with user behavior data from similar sequence nodes, the problem of unstable recommendation results for sparse users is effectively solved, enhancing the robustness of the recommendation. Since the recommended content no longer relies solely on the target user's limited behavior records but incorporates the behavior patterns of similar groups, the recommendation results remain relatively stable even when user behavior data fluctuates or is updated, reducing randomness and bias caused by data sparsity. For new users or low-activity users, effective recommendations can be achieved through multi-source behavior collection and sequence knowledge graph mapping without relying on a large amount of historical interaction data, significantly improving the performance of traditional recommendation algorithms in cold start scenarios.

[0191] See Figure 5 , Figure 5 The diagram illustrates the processing stages of a content recommendation method provided in one embodiment of this specification.

[0192] In this embodiment, the offline stage and the online stage, as well as the connection between the two, are clearly divided from the perspective of the overall architecture.

[0193] like Figure 5 As shown on the left, the offline stage starts with multi-source data collection, acquiring multi-source interest feature data of users from multiple sources such as registration information, device characteristics, third-party purchase signals, and second-party RTA; then, the purchase signals and fine-ranking signals are encoded and processed by Transformer and CGC feature fusion network respectively to generate two types of user interest vector representations: purchase and fine-ranking.

[0194] The RQ-KMeans algorithm is used to perform 8-level residual quantization on user interest vectors to generate semantic identifier sequences (SIDs) for each user. A hierarchical sequence knowledge graph is constructed based on the SID sequences of all active users (i.e., the user identifier sequences in the above embodiment). Each level in the graph corresponds to the population granularity divided by a certain prefix depth of the SID sequence. Each level node contains statistical information such as the group label, interest TGI distribution, and demographic characteristics of the corresponding population.

[0195] The audience knowledge graph (i.e., sequence knowledge graph) built in the offline phase can be used in the online phase, currently for content recommendation for target users with no behavioral data or sparse behavioral data. In the online phase, a SID sequence (i.e., the target identifier sequence in the above embodiment) is first generated for the target user. Similar sequence nodes are located in the sequence knowledge graph by matching SID prefixes, thereby determining the similar audience packages (i.e., similar user sets) corresponding to the similar sequence nodes.

[0196] High CTR content is mined based on user behavior data from similar audience groups. Specifically, a preset number of content items are selected from the exposure content of similar audience groups over the past month, sorted by click-through rate, to form a candidate content set. The content in the candidate content set is the high CTR content. Then, the mined candidate content set is injected into the last N sequence of the target user to realize interest transfer and trigger the recall and ranking of downstream recommendation strategies to generate personalized recommendation content for the target user.

[0197] When the recommendation is effective, as the target user's own behavioral data continues to accumulate, the system dynamically adjusts the weight of historical behavioral data and individual behavioral data through a time decay factor, gradually transitioning from group prior content to individual precise modeling, and realizing a seamless evolution of the recommendation strategy from the cold start stage to the behavior rich stage.

[0198] The content recommendation method provided in this specification applies a multi-level residual quantization system to the user side. It quantifies user interest vectors using RQ-KMeans to generate user semantic identifiers (SIDs), achieving a paradigm shift from "item semantic retrieval" to "user semantic matching," effectively filling the technical gap in cold-start user interest modeling. An 8-level residual quantization replaces single-layer clustering, forming a hierarchical user structure from coarse to fine. As the SID level deepens, the interest granularity and TGI significantly improve, achieving a finer characterization capability far exceeding that of single-layer clustering. The precise matching mechanism based on SID prefixes achieves low-conflict-rate user localization in a million-level combination space, ensuring the accuracy and interpretability of the matching. Simultaneously, this method integrates multi-source heterogeneous signals from third-party procurement, second-party RTA, H5 weak associations, and registration devices. Through end-to-end fusion using Transformer and CGC, it generates dual-path representations for procurement and fine-ranking, fully utilizing all semantic signals available to users without behavioral data, without relying on cross-domain behavioral data migration, and thus having a wider range of applicable scenarios.

[0199] Based on this, this method uses the interest migration path of "high CTR content mining → last N sequence injection" to migrate the group's prior content to the downstream ranking and recall links at low cost. Combined with the dynamic adjustment mechanism of time decay and individual behavior weight, it achieves a smooth transition from group prior to individual customization, ensuring continuous effectiveness throughout the entire lifecycle of users from cold start, sparse behavior to mature behavior.

[0200] Corresponding to the above method embodiments, this specification also provides embodiments of a content recommendation device. Figure 6 A schematic diagram of a content recommendation device according to one embodiment of this specification is shown. Figure 6 As shown, the device includes: Sequence generation module 602 is configured to generate a target identifier sequence based on the target user's interest feature data; User determination module 604 is configured to determine a set of similar users from the initial user set based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set; The data determination module 606 is configured to determine the historical behavior data of the target user based on the user behavior data corresponding to the similar user set; The content determination module 608 is configured to determine the recommended content corresponding to the target user based on the historical behavior data.

[0201] Optionally, the sequence generation module 602 is further configured to: Based on the target user's interest feature data, a user interest vector is constructed, wherein the target user's interest feature data is obtained from multiple data sources; The user interest vector is vector-mapped to generate the target identifier sequence, wherein the hierarchical identifier of each level in the target identifier sequence represents the interest category of the target user at different granularities.

[0202] Optionally, the user determination module 604 is further configured to: A sequence knowledge graph is constructed based on the user identifier sequence of each initial user in the initial user set, wherein the sequence node in the sequence knowledge graph corresponds to at least one user identifier sequence; Based on the target identifier sequence, similar sequence nodes are determined from the sequence knowledge graph; A set of similar users is determined based on the user identifier sequences corresponding to the similar sequence nodes.

[0203] Optionally, the user determination module 604 is further configured to: Based on the user behavior data of each initial user in the initial user set, construct the active interest vector corresponding to each initial user; Perform vector mapping on each active interest vector to obtain the user identifier sequence corresponding to each initial user; Based on the level identifiers of each level in each user identifier sequence, the sequence nodes of the corresponding level on the sequence knowledge graph are determined. In the case of level N, the level identifier of the Nth level is determined by the level identifiers of the previous N levels, where N≥1. The sequence knowledge graph is constructed based on sequence nodes at at least one level.

[0204] Optionally, the user determination module 604 is further configured to: The target level corresponding to the target identifier sequence is determined, and similar sequence nodes corresponding to the target identifier sequence are determined from the sequence nodes corresponding to the target level in the sequence knowledge graph based on prefix matching.

[0205] Optionally, the data determination module 606 is further configured to: Obtain user behavior data corresponding to each similar user in the set of similar users; Obtain an exposure content set from the user behavior data, and determine a candidate content set from the exposure content set; Historical behavior data of the target user is generated based on the set of candidate content.

[0206] Optionally, the data determination module 606 is further configured to: The click-through rate of each exposed content in the exposed content set is sorted, and the candidate content set is obtained based on the sorting results.

[0207] Optionally, the content determination module 608 is further configured to: Based on the historical behavior data, content is retrieved and sorted to determine the recommended content to be recommended to the target user.

[0208] Optionally, the content determination module 608 is further configured to: Obtain individual behavioral data of the target user based on the recommended content; Upon receiving a content recommendation request for the target user, the system determines the target content to recommend to the target user based on the target user's historical behavior data and individual behavior data.

[0209] Optionally, the content determination module 608 is further configured to: Determine a first initial weight for the historical behavioral data and a second initial weight for the individual behavioral data; The first initial weight and the second initial weight are dynamically adjusted according to the time decay factor to obtain the first updated weight of the historical behavior data and the second updated weight of the individual behavior data. The time decay factor increases monotonically with the accumulation time of the target user's behavior. The first updated weight is negatively correlated with the time decay factor, and the second updated weight is positively correlated with the time decay factor. Based on the historical behavior data, the first update weight, the individual behavior data, and the second update weight, the target content to be recommended to the target user is determined.

[0210] The content recommendation device provided in this specification generates a target identifier sequence composed of multi-level identifiers based on the target user's interest feature data. This target identifier sequence indicates the target user's interest categories at multiple granular levels, thereby characterizing the user's interest features layer by layer from macro-interest categories to micro-interest preferences. Based on this, the device matches the target identifier sequence with the user identifier sequences of each initial user in the initial user set, determining a set of similar users from the initial user set whose interest semantics are similar to the target user's. The device then constructs the target user's historical behavior data based on the user behavior data corresponding to the similar user set, effectively enriching the behavior data of sparse behavior users and compensating for insufficient modeling caused by missing individual behavior data. The device determines the recommended content to be recommended to the target user based on the historical behavior data, avoiding complete reliance on popular recommendations due to insufficient data, significantly improving the matching degree between recommended content and the individual's true interests, especially improving the poor recommendation effect caused by insufficient data in cold start and sparse behavior scenarios.

[0211] The above is an illustrative scheme of a content recommendation device according to this embodiment. It should be noted that the technical solution of this content recommendation device and the technical solution of the content recommendation method described above belong to the same concept. For details not described in detail in the technical solution of the content recommendation device, please refer to the description of the technical solution of the content recommendation method described above.

[0212] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0213] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0214] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0215] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.

[0216] The processor 720 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the recommended method described above.

[0217] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computing device embodiments are basically similar to the content recommendation method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the content recommendation method embodiments.

[0218] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0219] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computer-readable storage medium embodiments are relatively simple in description because they are fundamentally similar to the content recommendation method embodiments; relevant parts can be referred to in the description of the content recommendation method embodiments.

[0220] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0221] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the content recommendation method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the content recommendation method described above.

[0222] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0223] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0224] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0225] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0226] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A content recommendation method, characterized in that, include: Generate a target identifier sequence based on the target user's interest feature data; Based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set, a similar user set is determined from the initial user set; Based on the user behavior data corresponding to the set of similar users, determine the historical behavior data of the target user; Recommended content for the target user is determined based on the historical behavior data.

2. The method as described in claim 1, characterized in that, Generate a target identifier sequence based on the target user's interest feature data, including: Based on the target user's interest feature data, a user interest vector is constructed, wherein the target user's interest feature data is obtained from multiple data sources; The user interest vector is vector-mapped to generate the target identifier sequence, wherein the hierarchical identifier of each level in the target identifier sequence represents the interest category of the target user at different granularities.

3. The method as described in claim 1, characterized in that, Based on the target identifier sequence and the user identifier sequences of each initial user in the initial user set, a similar user set is determined from the initial user set, including: A sequence knowledge graph is constructed based on the user identifier sequence of each initial user in the initial user set, wherein the sequence node in the sequence knowledge graph corresponds to at least one user identifier sequence; Based on the target identifier sequence, similar sequence nodes are determined from the sequence knowledge graph; A set of similar users is determined based on the user identifier sequences corresponding to the similar sequence nodes.

4. The method as described in claim 3, characterized in that, Constructing a sequence knowledge graph based on the user identifier sequence of each initial user in the initial user set, including: Based on the user behavior data of each initial user in the initial user set, construct the active interest vector corresponding to each initial user; Perform vector mapping on each active interest vector to obtain the user identifier sequence corresponding to each initial user; Based on the level identifiers of each level in each user identifier sequence, the sequence nodes of the corresponding level on the sequence knowledge graph are determined. In the case of level N, the level identifier of the Nth level is determined by the level identifiers of the previous N levels, where N≥1. The sequence knowledge graph is constructed based on sequence nodes at at least one level.

5. The method as described in claim 3, characterized in that, Based on the target identifier sequence, similar sequence nodes are determined from the sequence knowledge graph, including: The target level corresponding to the target identifier sequence is determined, and similar sequence nodes corresponding to the target identifier sequence are determined from the sequence nodes corresponding to the target level in the sequence knowledge graph based on prefix matching.

6. The method as described in claim 1, characterized in that, Based on the user behavior data corresponding to the set of similar users, the historical behavior data of the target user is determined, including: Obtain user behavior data corresponding to each similar user in the set of similar users; Obtain an exposure content set from the user behavior data, and determine a candidate content set from the exposure content set; Historical behavior data of the target user is generated based on the set of candidate content.

7. The method as described in claim 6, characterized in that, Determining a candidate content set from the exposed content set includes: The click-through rate of each exposed content in the exposed content set is sorted, and the candidate content set is obtained based on the sorting results.

8. The method as described in claim 1, characterized in that, Based on the historical behavior data, the recommended content for the target user is determined, including: Based on the historical behavior data, content is retrieved and sorted to determine the recommended content to be recommended to the target user.

9. The method as described in claim 1, characterized in that, After determining the recommended content for the target user based on the historical behavior data, the method further includes: Obtain individual behavioral data of the target user based on the recommended content; Upon receiving a content recommendation request for the target user, the system determines the target content to recommend to the target user based on the target user's historical behavior data and individual behavior data.

10. The method as described in claim 9, characterized in that, Based on the target user's historical behavior data and individual behavior data, the target content to be recommended to the target user is determined, including: Determine a first initial weight for the historical behavioral data and a second initial weight for the individual behavioral data; The first initial weight and the second initial weight are dynamically adjusted according to the time decay factor to obtain the first updated weight of the historical behavior data and the second updated weight of the individual behavior data. The time decay factor increases monotonically with the accumulation time of the target user's behavior. The first updated weight is negatively correlated with the time decay factor, and the second updated weight is positively correlated with the time decay factor. Based on the historical behavior data, the first update weight, the individual behavior data, and the second update weight, the target content to be recommended to the target user is determined.

11. A content recommendation device, characterized in that, include: The sequence generation module is configured to generate target identifier sequences based on the target user's interest feature data. The user determination module is configured to determine a set of similar users from the initial user set based on the target identifier sequence and the user identifier sequence of each initial user in the initial user set; The data determination module is configured to determine the historical behavior data of the target user based on the user behavior data corresponding to the similar user set; The content determination module is configured to determine the recommended content corresponding to the target user based on the historical behavior data.

12. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, characterized in that, when the computer programs / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 10.

13. A computer-readable storage medium storing a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 10.

14. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 10.

Images