Fusion target perception behavior sequence attention and multi-target information flow sequencing method

Abstract

The present application relates to the technical field of information recommendation, and more particularly to a multi-target information flow sorting method fusing target perception behavior sequence attention, obtaining a candidate content and a user multi-behavior history sequence, extracting a candidate content attribute vector, and constructing a click, interaction and negative feedback query vector, generating click, interaction and negative feedback sequence representations through attention weights, cross-behavior conversion matrix and time decay function, respectively crossing the sequence representations of the three dimensions with the candidate content, and then distinguishing instant interest and persistent preference through orthogonal loss constraint and inputting an estimation network to obtain click, interaction and negative feedback probabilities, generating interaction weights and negative feedback penalty weights according to the candidate content and user feature portrait, and weightedly calculating a final sorting score, thereby effectively improving the accuracy of recommendation sorting and user experience, and optimizing multi-target sorting effect.

Description

Technical Field

[0001] This invention relates to the field of information recommendation technology, and in particular to a method for integrating target-aware behavioral sequence attention with multi-target information flow ranking. Background Technology

[0002] In information recommendation systems, multi-behavior sequence modeling is a key technology for improving ranking performance. Current conventional approaches treat different user behavior types (such as clicks, interactions, and negative feedback) as independent sources, lacking modeling of semantic connections and transformation relationships between behaviors. For example, the representation of the same item in click sequences and interaction sequences is treated in isolation, failing to effectively capture the progressive evolution of user interest from observation to engagement. This fragmentation makes it impossible to distinguish between immediate interests and persistent preferences, making the prediction of interaction probability susceptible to interference from accidental clicks and reducing the robustness of ranking.

[0003] Existing technologies typically use fixed weights or empirical parameters to weight click probabilities and interaction probabilities, ignoring differentiated penalties for negative feedback. Repeated negative feedback from users on certain attributes (such as quickly swiping away or marking as uninterested) may reflect deep-seated aversion, but existing methods do not introduce adaptive weights to strengthen the suppression of negative signals, making it difficult for recommended content to dynamically match the user's true intent at the current stage. These limitations are particularly pronounced in long-term user behavior data, easily leading to homogenized recommendation results or an overemphasis on short-term clicks. Summary of the Invention

[0004] This invention provides a method for integrating target perception behavior sequence attention and multi-target information flow ranking, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for fusing target perception behavior sequence attention and multi-target information flow ranking, comprising:

[0006] Obtain candidate content and user multi-behavior historical sequences, extract attribute vectors from candidate content, and construct click query vectors, interaction query vectors, and negative feedback query vectors based on the different attribute vectors of candidate content;

[0007] Attention weights are calculated for each query vector and the user's multi-behavior history sequence. The representations of the same item in different behavior sequences are fused through a cross-behavior conversion matrix, and a time decay function is introduced to generate click sequence representations, interaction sequence representations, and negative feedback sequence representations.

[0008] The three sequence representations are cross-referenced with candidate content features. The click sequence representation and the interaction sequence representation are constrained by orthogonal loss to maximize the cosine distance between them, thereby capturing immediate interest and persistent preference, respectively. The three feature representations after cross-reference are then input into the corresponding prediction networks, which output click probability, interaction probability, and negative feedback probability.

[0009] Based on the candidate content and user profile, an interaction weight and a negative feedback penalty weight are generated. The click probability is multiplied by the interaction probability weighted by the interaction weight to obtain a positive score, and the negative feedback probability weighted by the negative feedback penalty weight is subtracted to obtain a negative deduction score, thus obtaining the final ranking score.

[0010] Attribute vectors are extracted from candidate content. Click query vectors, interaction query vectors, and negative feedback query vectors are constructed based on the different attribute vectors of the candidate content, including:

[0011] The attribute vectors of candidate content are grouped according to content format attributes and content theme attributes. The attribute vectors corresponding to the content format attributes are concatenated to construct click query vectors, and the attribute vectors corresponding to the content theme attributes are concatenated to construct interactive query vectors.

[0012] Extract the duplication attribute vector of the candidate content, and construct a negative feedback query vector by concatenating the duplication attribute vector.

[0013] Each query vector is weighted with the user's multi-behavior history sequence. Representations of the same item across different behavior sequences are fused using a cross-behavior conversion matrix. A time decay function is introduced to generate click sequence representations, interaction sequence representations, and negative feedback sequence representations, including:

[0014] For each item in a user's multi-behavior history sequence, multiple representation vectors of the item in different behavior type sequences are identified. The click query vector, interaction query vector, or negative feedback query vector is input into a transformation parameter generation network to obtain a cross-behavior transformation matrix corresponding to the query. The cross-behavior transformation matrix is ​​used to project multiple representation vectors onto a unified feature space. The semantic matching degree between the query vector and each behavior type is calculated as the behavior fusion weight. The projected representation vectors are then weighted and fused according to the behavior fusion weight to obtain a query-aware fusion representation vector.

[0015] Based on the characteristics of different behavior types, a differentiated time decay function is constructed, the time decay weight corresponding to the behavior type is calculated, the time decay weight is multiplied by the behavior fusion weight to obtain a comprehensive weight, and the query perception fusion representation vector is adjusted accordingly.

[0016] The attention score is calculated by combining the query vector with the query perception fusion representation vector after comprehensive weight adjustment and normalized to obtain the attention weight. The attention weight is then used to perform weighted summation to generate click sequence representation, interaction sequence representation and negative feedback sequence representation respectively.

[0017] The click query vector, interaction query vector, or negative feedback query vector is input into the conversion parameter generation network to obtain the cross-behavior conversion matrix corresponding to the query, including:

[0018] The query vector is input into the multi-head self-attention module. Each attention head learns the feature representation of the query vector in the semantic dimension, structural dimension and temporal dimension respectively, and the two are concatenated to obtain multi-granular semantic features.

[0019] The multi-granular semantic features are input into a gated fusion unit, which includes a behavior type recognition gate and a behavior intensity modulation gate. The recognition gate generates an activation probability of zero to one for each behavior type, and the modulation gate dynamically adjusts the weight amplitude of each behavior type according to the query semantic intensity. The outputs of the two gates are multiplied to generate a behavior type weight vector.

[0020] Multiple sets of basic conversion matrices for different behavior types are pre-constructed, including four sets of conversion matrices: browsing to click, click to interaction, interaction to deep interaction, and any behavior to negative feedback. Each set of matrices is pre-trained using conversion statistical patterns of user behavior data.

[0021] The behavior type weight vector is used to dynamically weight and fuse multiple basic conversion matrices. The click query vector makes the weight of the click matrix greater than the weight of other matrices. The interaction query vector makes the sum of the weights of the click to the interaction matrix and the interaction to the deep interaction matrix greater than the sum of the weights of other matrices. The negative feedback query vector makes the weight of any behavior to the negative feedback matrix greater than the weight of other matrices.

[0022] Differentiated time decay functions are constructed based on the characteristics of different behavior types, and the time decay weights corresponding to the behavior types are calculated, including:

[0023] An exponential decay function is set for click behavior, and the decay coefficient is adjusted according to the timeliness attribute of the candidate content. The larger the timeliness attribute value, the larger the decay coefficient.

[0024] A power-law decay function is set for interactive behaviors, where the decay rate of interactive behaviors is less than that of click behaviors. The decay exponent is adjusted according to the interaction depth attribute of the candidate content, with a larger interaction depth attribute value resulting in a smaller decay exponent.

[0025] A segmented decay function is set for negative feedback behavior, dividing the time window into a recent window, a medium window, and a long window. The recent window uses a constant weight, the medium window uses linear decay, and the long window uses exponential decay. The decay rate of the recent window is less than that of the medium window, and the decay rate of the medium window is less than that of the long window. The time boundary of each segment window is adaptively adjusted according to the frequency of the user's negative feedback behavior. The higher the frequency of negative feedback behavior, the shorter the time boundary of the recent window.

[0026] A personalized time-aware mechanism is introduced, which extracts user behavior activity features from user profiles to adjust the parameters of the decay function. The larger the user behavior activity feature value, the larger the click decay coefficient, and the smaller the user behavior activity feature value, the smaller the interaction decay index.

[0027] The three sequence representations are cross-referenced with candidate content features. The click sequence representation and the interaction sequence representation are constrained by orthogonal loss to maximize their cosine distance, capturing immediate interest and persistent preference respectively. The three cross-referenced feature representations are then input into their respective prediction networks, outputting click probability, interaction probability, and negative feedback probability, including:

[0028] Multiple attribute vectors of candidate content are concatenated to construct a candidate content feature vector. Three sequence representations are then cross-referenced with the candidate content feature vector to obtain three cross-feature representations.

[0029] In the sequence representation layer and feature cross layer, cosine similarity between click sequence representation and interaction sequence representation is calculated to construct a two-layer orthogonal constraint term. Cosine similarity between click sequence representation and interaction sequence representation and negative feedback sequence representation is calculated to construct a negative feedback reverse association loss term. The two-layer orthogonal constraint term and negative feedback reverse association loss term are jointly optimized with the prediction loss to transform the parameters of the generation network, click prediction network, interaction prediction network and negative feedback prediction network.

[0030] By jointly optimizing the click sequence representation to capture immediate interest patterns, the interaction sequence representation to capture persistent preferences, and the negative feedback sequence representation to capture aversion patterns, the three cross feature representations inherit the characteristics of the corresponding sequence representations and are then fed into the click prediction network, the interaction prediction network, and the negative feedback prediction network, respectively, and output the click probability, interaction probability, and negative feedback probability.

[0031] Based on candidate content and user profiles, interaction weights and negative feedback penalty weights are generated. The click probability is multiplied by the interaction probability weighted by the interaction weights to obtain a positive score, and the negative feedback probability weighted by the negative feedback penalty weights is subtracted to obtain a negative deduction score, resulting in the final ranking score, including:

[0032] A user interaction preference vector is constructed by extracting profile features representing user interaction tendencies from user profiles, and a content interaction adaptation vector is constructed by extracting attribute features representing content interaction potential from candidate content features. The interaction representation between the user interaction preference vector and the content interaction adaptation vector is calculated, and the interaction representation is input into a weight generation network to generate interaction weights.

[0033] User sensitivity vectors are constructed by extracting profile features representing user sensitivity to negative feedback from user profiles. These user sensitivity vectors and candidate content features are then input into a weight generation network to generate negative feedback penalty weights.

[0034] The click probability is used as the base score. The product of the interaction weight and the interaction probability is calculated as an interaction bonus. The base score and the interaction bonus are added together to obtain a positive score. The product of the negative feedback penalty weight and the negative feedback probability is calculated as a negative feedback deduction. The positive score is subtracted from the negative feedback deduction to obtain the final ranking score. The final ranking score comprehensively reflects the immediate click appeal, long-term interaction value, and negative feedback risk of the candidate content.

[0035] A second aspect of the present invention provides an electronic device, comprising:

[0036] processor;

[0037] Memory used to store processor-executable instructions;

[0038] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0039] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0040] This method significantly improves the accuracy and robustness of information feed recommendation ranking. By constructing query vectors specifically for clicks, interactions, and negative feedback, the model can accurately capture the matching degree of candidate content under different user behavioral intentions. Introducing a cross-behavior transformation matrix enables the fusion of representations of the same item across multiple behavioral sequences, effectively eliminating the information fragmentation problem caused by the separation of behavior types. The embedding of a time decay function allows the model to dynamically adapt to the natural changes in interests over time, avoiding interference from outdated interaction data on the current ranking decision, thus ensuring that the recommendation results are more closely aligned with the user's current real needs.

[0041] Click and interaction sequence representations, after being constrained by orthogonal loss, maximize cosine distance, focusing on two dimensions: immediate interest and persistent preference, respectively. This decompositional learning prevents the two types of interests from confounding each other in the feature space, while ensuring that each dimension can autonomously discover the most relevant user behavior patterns. Click sequence representations emphasize short-term impulsive behavior, while interaction sequence representations extract long-term stable value orientations. The two are complementary but not redundant, enabling the model to cover multiple layers of user needs while improving generalization ability, especially suitable for cold-start items and sparse user scenarios.

[0042] The final ranking score is calculated by multiplying the click probability by the weighted interaction probability as a positive score, and then subtracting the weighted negative feedback probability as a negative penalty. The interaction weight and negative feedback penalty weight are dynamically generated based on candidate content and user profiles, enabling adaptive adjustment of positive rewards and negative penalties. This mechanism effectively suppresses the exposure of content that easily triggers negative reactions, while strengthening high-value items that can stimulate long-term engagement, thus achieving a better balance between click-through rate, conversion rate, and user retention. Overall, the ranking results better align with both business objectives and user experience requirements, and the calculation process maintains the efficiency of online inference. Attached Figure Description

[0043] Figure 1 A flowchart illustrating the integration of target-aware behavioral sequence attention and multi-target information flow ranking methods;

[0044] Figure 2 This is a schematic diagram of the differentiated time decay function for the three behavior types. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0047] Figure 1 This is a flowchart illustrating the method for integrating target perception behavior sequence attention and multi-target information flow ranking in an embodiment of the present invention.

[0048] Methods that integrate target-aware behavioral sequence attention with multi-target information flow ranking include:

[0049] Obtain candidate content and user multi-behavior historical sequences, extract attribute vectors from candidate content, and construct click query vectors, interaction query vectors, and negative feedback query vectors based on the different attribute vectors of candidate content;

[0050] Attention weights are calculated for each query vector and the user's multi-behavior history sequence. The representations of the same item in different behavior sequences are fused through a cross-behavior conversion matrix, and a time decay function is introduced to generate click sequence representations, interaction sequence representations, and negative feedback sequence representations.

[0051] The three sequence representations are cross-referenced with candidate content features. The click sequence representation and the interaction sequence representation are constrained by orthogonal loss to maximize the cosine distance between them, thereby capturing immediate interest and persistent preference, respectively. The three feature representations after cross-reference are then input into the corresponding prediction networks, which output click probability, interaction probability, and negative feedback probability.

[0052] Based on the candidate content and user profile, an interaction weight and a negative feedback penalty weight are generated. The click probability is multiplied by the interaction probability weighted by the interaction weight to obtain a positive score, and the negative feedback probability weighted by the negative feedback penalty weight is subtracted to obtain a negative deduction score, thus obtaining the final ranking score.

[0053] Attribute vectors are extracted from candidate content. Click query vectors, interaction query vectors, and negative feedback query vectors are constructed based on the different attribute vectors of the candidate content, including:

[0054] The attribute vectors of candidate content are grouped according to content format attributes and content theme attributes. The attribute vectors corresponding to the content format attributes are concatenated to construct click query vectors, and the attribute vectors corresponding to the content theme attributes are concatenated to construct interactive query vectors.

[0055] Extract the duplication attribute vector of the candidate content, and construct a negative feedback query vector by concatenating the duplication attribute vector.

[0056] Candidate content is typically described by multiple dimensions of attributes, such as content display format, topic category, creator information, content length, and historical exposure frequency. Before constructing the query vector, it is necessary to extract attributes from the candidate content and map each dimension of the attributes into a low-dimensional dense embedding vector. Specifically, for each candidate content, its type tags (such as text, article, post, etc.), content length segmentation tags, and other formal attributes are read from the content metadata, as well as topic tags, domain classifications, keywords, and other thematic attributes. Simultaneously, the repetition attributes, such as the number of times the content has been repeatedly exposed by the same user in historical recommendations and the percentage of similar content consumed by the user, are also read. Each discrete attribute obtains its corresponding embedding vector through a lookup table operation, while continuous attributes are also converted into embedding vectors after binning and discretization, ensuring that all attributes participate in subsequent calculations in a unified vector form.

[0057] The distinction between content format attributes and content theme attributes stems from the fundamental differences in the decision-making criteria users use across different behavioral types. A user's decision to click on content is often strongly influenced by its presentation—the style of the title, the relevance of the content type to the current context, the attractiveness of the accompanying images, etc. These format-level cues can trigger a user's click impulse within a very short exposure time. Therefore, by concatenating the embedding vectors corresponding to the content format attributes along the feature dimension, a click query vector can be constructed. Suppose that the content format attributes include a total of There are n embedding vectors, each vector having a dimension of 1. The concatenated click query vector dimension is Subsequently, it can be mapped to a unified query space dimension through linear transformation, so as to perform attention calculation with user behavior sequences.

[0058] User engagement with content, such as liking, saving, commenting, and sharing, largely depends on the content's thematic depth and its relevance to their personal interests. Content highly relevant to a user's long-term interests, even in a seemingly ordinary format, can stimulate sustained interaction. Therefore, constructing an interactive query vector by concatenating the embedding vectors corresponding to the content's thematic attributes along the feature dimension is a viable approach. Assume the content theme attributes contain a total of The concatenated interactive query vector has 1 embedded vector and a dimension of 1. Similarly, the vectors are mapped to a unified query space through a linear transformation. The separation of the feature sources of the click query vector and the interaction query vector allows the subsequent two attention mechanisms to activate memories related to form preferences and memories related to topic preferences from the user's historical behavior sequence, respectively, avoiding interference between the two types of signals and thus capturing the user's immediate interests and persistent preferences.

[0059] The triggering mechanisms for negative feedback behaviors (such as marking "not interested" or quickly swiping away) differ significantly from those for positive behaviors. One of the core reasons users generate negative feedback is content repetition: when content in the recommendation feed is highly similar to recently consumed content, users' boredom and aversion increase significantly. Therefore, the construction of negative feedback query vectors specifically targets the repetition attribute of candidate content. Extracting the repetition attribute vector requires comprehensive consideration of multiple dimensions: the overlap ratio of candidate content with recently clicked content in terms of topic tags, the degree of overlap between the creator of the candidate content and the creators of content recently consumed by the user, the thematic similarity between the candidate content and the user's historically consumed content, and the number of times the candidate content has been exposed in the current recommendation request. The negative feedback query vector is constructed by concatenating the above repetition-related embedding vectors along the feature dimensions. Suppose the repeatability attribute contains a total of The concatenated negative feedback query vector has a dimension of _ embedding vectors. After linear transformation, it is used to retrieve the historical memory most relevant to the repetition of the current candidate content in the user's historical negative feedback sequence.

[0060] The concatenation operation of the three types of query vectors offers the dual advantages of simplicity and interpretability. The concatenation operation does not introduce additional information loss and can fully preserve the semantic information carried by each attribute vector; at the same time, different attribute vectors remain independent in the concatenation dimension, allowing the subsequent linear transformation layer to flexibly learn the contribution weights of each attribute to the query vector without requiring pre-specification. In engineering implementation, the dimension of each attribute embedding vector... Typically, a uniform value is set to ensure the neatness of the splicing operation; if the embedding dimensions of different attributes are different, they need to be aligned to the same dimension through their respective projection layers before splicing is performed.

[0061] The process of acquiring candidate content and user multi-behavior historical sequences requires the simultaneous extraction of candidate content attributes and the reading of user behavior sequences, both of which serve as input for subsequent attention calculations. User multi-behavior historical sequences typically contain three independent behavior chains: click sequences, interaction sequences, and negative feedback sequences. Each sequence records the items the user has historically interacted with under the corresponding behavior type and their timestamps. When reading sequences, each sequence is sorted in descending order of timestamp, retaining the most recent records. The attribute embedding vectors of each historical item in the sequence are also extracted, serving as the key-value pair source for attention calculations. The attribute vector extraction for candidate content and the item vector extraction for user sequences share the same embedding table, ensuring semantic consistency between the candidate and historical sides. This gives the similarity calculation between the query vector and the item vectors in the sequence practical semantic meaning.

[0062] In the overall process, the construction of the three types of query vectors is a prerequisite for the subsequent cross-behavior attention mechanism. (Click query vector) Attention weighting will be performed on the click behavior sequence to extract historical click memories most relevant to the current candidate content format; interactive query vector Attention weighting will be applied to the sequence of interactive behaviors to extract historical interaction memories most relevant to the current candidate content topic; negative feedback query vectors will be used. Attention weighting is applied to the negative feedback behavior sequence to extract historical negative feedback memories most relevant to the repetition of the current candidate content. The separate design of the three query vectors allows the entire model to model different user decision dimensions separately within a unified attention framework, ultimately merging them into a precise characterization of the user's comprehensive preferences, providing a sufficient feature foundation for subsequent ranking score calculation.

[0063] Each query vector is weighted with the user's multi-behavior history sequence. Representations of the same item across different behavior sequences are fused using a cross-behavior conversion matrix. A time decay function is introduced to generate click sequence representations, interaction sequence representations, and negative feedback sequence representations, including:

[0064] For each item in a user's multi-behavior history sequence, multiple representation vectors of the item in different behavior type sequences are identified. The click query vector, interaction query vector, or negative feedback query vector is input into a transformation parameter generation network to obtain a cross-behavior transformation matrix corresponding to the query. The cross-behavior transformation matrix is ​​used to project multiple representation vectors onto a unified feature space. The semantic matching degree between the query vector and each behavior type is calculated as the behavior fusion weight. The projected representation vectors are then weighted and fused according to the behavior fusion weight to obtain a query-aware fusion representation vector.

[0065] Based on the characteristics of different behavior types, a differentiated time decay function is constructed, the time decay weight corresponding to the behavior type is calculated, the time decay weight is multiplied by the behavior fusion weight to obtain a comprehensive weight, and the query perception fusion representation vector is adjusted accordingly.

[0066] The attention score is calculated by combining the query vector with the query perception fusion representation vector after comprehensive weight adjustment and normalized to obtain the attention weight. The attention weight is then used to perform weighted summation to generate click sequence representation, interaction sequence representation and negative feedback sequence representation respectively.

[0067] In a user's multi-behavior history sequence, the same item may appear simultaneously in click sequences, interaction sequences, and negative feedback sequences. Its embedded representations under different behavior types often reside in different semantic subspaces, and direct concatenation or averaging can lead to semantic confusion. To address this issue, for each item in the sequence, its representation vector under each behavior type is extracted, forming a set of multi-behavior representation vectors. For example, let the original embedding vectors corresponding to click behavior, interaction behavior, and negative feedback behavior be respectively... , and If an item has no historical records under a certain behavior type, the corresponding vector is filled with a zero vector. This multi-representation structure provides a complete information foundation for subsequent cross-behavior fusion.

[0068] The generation of the cross-behavior transformation matrix relies on a transformation parameter generation network. This network takes the query vector as input and dynamically outputs a matrix for spatial projection, enabling different query targets to drive different cross-behavior alignment methods. (Based on the click query vector...) For example, after inputting the conversion parameters to generate the network, the output is a cross-behavior conversion matrix corresponding to the click. Similarly, interactive query vectors Corresponding transformation matrix Negative feedback query vector Corresponding transformation matrix The structure of the parameter generation network can be implemented using a two-layer fully connected network plus an activation function, where the input is the query vector and the output dimension is... The vectors are then reshaped to obtain a square matrix. Using this transformation matrix, the embedding vectors of items under each behavior type are projected onto a unified feature space. Taking a click scenario as an example, the projected representation vector is... ,in Iterate through all behavior types. Represents items In behavior type The representation below is the result after projection of the click transformation matrix.

[0069] After projection is completed, the semantic matching degree between the query vector and the projected representation of each behavior type is calculated, serving as the behavior fusion weight. Specifically, taking click query as an example, behavior type... Corresponding fusion weights Through calculation and The inner product is obtained by softmax normalization, i.e.

[0070]

[0071] The summation iterates through all behavior types. Based on the behavior fusion weights, the projected representations of each behavior type are weighted and summed to obtain the item. Query-aware fusion representation vector under "Click to query" Query-aware fusion representation vector in interactive and negative feedback scenarios , Calculate in the same way, using the corresponding transformation matrix and query vector respectively.

[0072] The introduction of a time decay function is used to model the impact of the time of an action on a user's current interest. Different action types exhibit fundamentally different sensitivities to time: clicks often reflect immediate content preferences and decay rapidly; interactive actions (such as saving and forwarding) reflect more stable, persistent preferences and decay relatively slowly; negative feedback actions (such as marking as "not interested") decay at a rate between the two, and should maintain a strong punitive effect within a certain time window. Based on these differentiated characteristics, an independent time decay function is constructed for each action type. Let the item... In behavior type The time interval between the occurrence of the following action and the current time is . The corresponding time decay weight is defined as ;in For behavior type The corresponding decay rate hyperparameter corresponds to a larger click behavior. Value, the interaction behavior corresponds to a smaller value Value, negative feedback behavior corresponds to medium Values. Each It can be initialized based on business priors before training and optimized in the form of learnable parameters during training, thereby adaptively fitting the time sensitivity under different scenarios.

[0073] The overall weight is the product of the time decay weight and the behavior fusion weight. Taking a click scenario as an example, the item... In behavior type The overall weight is The query-aware fusion representation vector is adjusted using comprehensive weights, and the adjusted representation vector is: ,in For items The scalar coefficients are the combined weights of all behavior types after normalization. This adjustment automatically reduces the contribution of older behaviors to the current sequence representation, while recent behaviors retain a stronger influence, which aligns with the dynamic evolution of user interests.

[0074] After obtaining the query-aware fusion representation vector of each item in the sequence after comprehensive weight adjustment, the relevance of each item to the candidate content is further calculated through an attention mechanism. Taking a click sequence as an example, let the sequence length be... For the first in the sequence For each item, calculate the query vector. Its adjusted fusion representation vector Attention scores between

[0075]

[0076] in This is a scaling factor used to prevent the inner product values ​​from becoming overly concentrated as the dimension increases, thus stabilizing the gradient distribution of the softmax function. The attention scores of all items in the sequence are then softmax normalized to obtain the attention weights.

[0077]

[0078] By using attention weights to perform a weighted summation of the adjusted fused representation vectors of each item in the sequence, the click sequence representation is obtained.

[0079]

[0080] Interactive sequence representation and negative feedback sequence characterization Using the same process, respectively , , and , , Obtained through independent calculation.

[0081] The generation processes of the three sequence representations described above are computationally independent and can be executed in parallel to improve inference efficiency. In engineering implementation, the generation and projection operations of cross-row transformation matrices can be efficiently implemented using batch matrix multiplication, with a time interval of [missing information]. The computation is completed and cached during the feature preprocessing stage to avoid the overhead of repeated computations during online inference. Decay rate. The learnable design enables the entire time decay module to automatically adapt to the interest decay patterns under different business scenarios during end-to-end training, without the need for manual parameter tuning for each scenario. The final generated click sequence representation, interaction sequence representation, and negative feedback sequence representation carry user immediate interest, persistent preference, and aversion signals guided by candidate content, respectively, providing semantically clear and information-complementary sequence-side inputs for subsequent cross-calculation with candidate content features.

[0082] The click query vector, interaction query vector, or negative feedback query vector is input into the conversion parameter generation network to obtain the cross-behavior conversion matrix corresponding to the query, including:

[0083] The query vector is input into the multi-head self-attention module. Each attention head learns the feature representation of the query vector in the semantic dimension, structural dimension and temporal dimension respectively, and the two are concatenated to obtain multi-granular semantic features.

[0084] The multi-granular semantic features are input into a gated fusion unit, which includes a behavior type recognition gate and a behavior intensity modulation gate. The recognition gate generates an activation probability of zero to one for each behavior type, and the modulation gate dynamically adjusts the weight amplitude of each behavior type according to the query semantic intensity. The outputs of the two gates are multiplied to generate a behavior type weight vector.

[0085] Multiple sets of basic conversion matrices for different behavior types are pre-constructed, including four sets of conversion matrices: browsing to click, click to interaction, interaction to deep interaction, and any behavior to negative feedback. Each set of matrices is pre-trained using conversion statistical patterns of user behavior data.

[0086] The behavior type weight vector is used to dynamically weight and fuse multiple basic conversion matrices. The click query vector makes the weight of the click matrix greater than the weight of other matrices. The interaction query vector makes the sum of the weights of the click to the interaction matrix and the interaction to the deep interaction matrix greater than the sum of the weights of other matrices. The negative feedback query vector makes the weight of any behavior to the negative feedback matrix greater than the weight of other matrices.

[0087] The core task of the conversion parameter generation network is to map query vectors into cross-behavioral conversion matrices that adapt to the current behavioral semantics. Different query vectors carry different behavioral intent signals—click query vectors reflect users' immediate preferences for content format, interaction query vectors reflect users' deep interest in content topics, and negative feedback query vectors reflect users' aversion to repetitive or low-quality content. By inputting these three types of query vectors into the conversion parameter generation network, a unique cross-behavioral conversion matrix can be dynamically generated for each behavioral semantic, thus ensuring that the projection direction of the representation of the same item in different behavioral sequences remains consistent with the query intent.

[0088] After the query vector enters the transformation parameter generation network, it first undergoes multi-granularity feature extraction through a multi-head self-attention module. This module processes the query vector across multiple parallel attention heads, each assigned to learn feature representations of the query vector in different semantic dimensions. Specifically, the first type of attention head captures the semantic dimension features of the query vector, i.e., static semantic information such as the content theme and category preferences expressed by the query; the second type captures the structural dimension features, i.e., the relational structure between the components of the query vector, reflecting the combination pattern of behavioral intent; and the third type captures the temporal dimension features, i.e., the time-sensitive information encoded in the query vector, reflecting the evolution trend of user interests over time. The outputs of the three types of attention heads are concatenated along the channel dimension to form a multi-granularity semantic feature vector. This multi-head splitting design avoids the problem of a single attention head processing semantic, structural, and temporal information in the same space, enabling subsequent gating fusion units to accurately identify behavior types and modulate weight amplitudes based on richer features.

[0089] The multi-granularity semantic feature vectors are then input into a gated fusion unit. This unit consists of two parallel gated substructures: a behavior type recognition gate and a behavior intensity modulation gate. The behavior type recognition gate performs a fully connected transformation on the multi-granularity semantic features and then activates them with a Sigmoid function, generating an activation probability between 0 and 1 for each predefined behavior type. A higher activation probability indicates a better semantic match between the current query vector and that behavior type. For example, when the input is an interactive query vector, the activation probabilities for click-to-interaction and interaction-to-deep interaction behaviors should be significantly higher than other behavior types; when the input is a negative feedback query vector, the activation probability for any behavior to negative feedback should be close to 1, while the activation probabilities for other behavior types should be close to 0. The behavior intensity modulation gate performs another fully connected transformation on the multi-granularity semantic features and then activates them with a Tanh function, outputting a modulation coefficient vector of equal length to the number of behavior types. Each coefficient reflects the intensity of the query semantics on the corresponding behavior type—behavior types with higher semantic intensity should have a larger weight in the final weighted fusion. The activation probability vector of the recognition gate is multiplied element-wise with the intensity coefficient vector of the modulation gate to obtain the behavior type weight vector. The meaning of this multiplication operation is that only when a certain behavior type is both determined to be active by the recognition gate (high activation probability) and determined to have significant semantic strength by the modulation gate (large absolute value of modulation coefficient) will that behavior type occupy a large weight component in the final weight vector, thereby achieving fine screening and intensity calibration of behavior types.

[0090] In the parameter system of the conversion parameter generation network, multiple sets of basic conversion matrices for different behavior types are pre-constructed and stored. These basic matrices cover four typical conversion patterns in the user behavior chain: the conversion matrix from browsing to clicking depicts the mapping relationship of the item representation space from passive exposure to active clicking; the conversion matrix from clicking to interaction depicts the representation transformation from clicking to generating deep interactive behaviors such as comments, favorites, and sharing; the conversion matrix from interaction to deep interaction depicts the representation evolution from ordinary interaction to high-frequency, continuous interaction; and the conversion matrix from arbitrary behavior to negative feedback depicts the representation shift direction when a user triggers negative feedback signals such as "dislike" in the context of any behavior sequence. Each set of basic conversion matrices is obtained by pre-training on statistical patterns of various behavior conversions in large-scale user behavior data. The training objective is to minimize the prediction loss between the converted item representation and the user feedback label of the target behavior type. After pre-training, these four sets of basic matrices are fixed in the network as learnable parameters and can be further fine-tuned during the end-to-end training phase.

[0091] After the behavior type weight vector is generated, the four sets of basic transformation matrices are dynamically weighted and fused to obtain the final cross-behavior transformation matrix. Let the four sets of basic transformation matrices be denoted as follows: , , , This corresponds to four conversion modes: browsing to click, click to interaction, interaction to deep interaction, and any behavior to negative feedback; the four components of the behavior type weight vector are denoted as... , , , After Softmax normalization, it satisfies The final cross-behavior transformation matrix obtained through fusion for: .

[0092] When the input is a click query vector, the training objective of the parameter generation network is to... Significantly greater than after normalization , , This means that the click matrix dominates the fusion process, aligning the projection direction of the item representation with the semantic space of the click behavior. When the input is an interaction query vector, the training objective is to... Greater than This means that the sum of the weights of the click-to-interaction matrix and the interaction-to-deep-interaction matrix exceeds the sum of the weights of the other two sets of matrices, ensuring that the projection of the item representation fully covers the continuous behavior space from light to deep interaction. When the input is a negative feedback query vector, the training objective is to make... After normalization, it is significantly larger than the other three components, meaning that the negative feedback matrix of any behavior dominates the fusion, making the projection direction of the item representation precisely aligned with the semantic space of the negative feedback behavior.

[0093] This dynamic weighted fusion mechanism offers significant advantages over maintaining a separate transformation matrix for each query vector. Independent matrix schemes are prone to overfitting on behavior types with sparse training data, while the base matrix-based weighted fusion scheme effectively transfers transformation knowledge accumulated on data-rich behavior types through parameter sharing, improving the quality of the transformation matrix in low-frequency behavior scenarios such as negative feedback. Furthermore, the behavior type weight vectors are dynamically generated from the query vectors rather than fixed presets, ensuring that the same type of query vector produces differentiated fusion weights across different user profiles or content scenarios, enhancing the transformation matrix's adaptability to contextual changes.

[0094] In actual deployment scenarios, the inference computation of the transformation parameter generation network is executed immediately after the query vector of the candidate content is determined. The three types of cross-behavior transformation matrices generated are cached for subsequent item representation projection, avoiding repeated network inference for each sequence item. This ensures the quality of the transformation matrix while keeping the additional computational overhead within an acceptable range.

[0095] Differentiated time decay functions are constructed based on the characteristics of different behavior types, and the time decay weights corresponding to the behavior types are calculated, including:

[0096] An exponential decay function is set for click behavior, and the decay coefficient is adjusted according to the timeliness attribute of the candidate content. The larger the timeliness attribute value, the larger the decay coefficient.

[0097] A power-law decay function is set for interactive behaviors, where the decay rate of interactive behaviors is less than that of click behaviors. The decay exponent is adjusted according to the interaction depth attribute of the candidate content, with a larger interaction depth attribute value resulting in a smaller decay exponent.

[0098] A segmented decay function is set for negative feedback behavior, dividing the time window into a recent window, a medium window, and a long window. The recent window uses a constant weight, the medium window uses linear decay, and the long window uses exponential decay. The decay rate of the recent window is less than that of the medium window, and the decay rate of the medium window is less than that of the long window. The time boundary of each segment window is adaptively adjusted according to the frequency of the user's negative feedback behavior. The higher the frequency of negative feedback behavior, the shorter the time boundary of the recent window.

[0099] A personalized time-aware mechanism is introduced, which extracts user behavior activity features from user profiles to adjust the parameters of the decay function. The larger the user behavior activity feature value, the larger the click decay coefficient, and the smaller the user behavior activity feature value, the smaller the interaction decay index.

[0100] Figure 2 This diagram illustrates the differentiated time decay functions for three behavior types. Click behavior uses an exponential decay function (solid line), with its weight decreasing rapidly over time, reflecting the short-term characteristics of immediate interest. Interaction behavior uses a power-law decay function (long dashed line), with a significantly slower decay rate than click behavior, reflecting the long-term stability of persistent preferences. Negative feedback behavior uses a piecewise decay function (short dashed line), maintaining a constant weight in the near-term window (0-24 hours) to reinforce the penalty effect of recent negative feedback, decaying linearly in the medium-term window (24-72 hours), and decaying exponentially to near zero in the long-term window (after 72 hours). The two dashed lines T1 and T2 in the diagram mark the time window boundaries of negative feedback behavior, which can be adaptively adjusted according to the frequency of users' negative feedback behavior. The decay rates of the three decay functions show significant differences: click behavior decays the fastest, with its weight dropping to around 0.03 after 7 days; interaction behavior decays the slowest, still maintaining a weight of 0.22 after 7 days; negative feedback behavior, due to its piecewise design, decays rapidly after 72 hours but remains higher than click behavior. This differentiated design enables the model to accurately distinguish users' immediate interests, persistent preferences, and aversion patterns, providing a time-aware behavioral weighting basis for multi-objective ranking.

[0101] Different behavior types exhibit significantly different decay characteristics over time. Click behavior typically reflects a user's immediate interest, and its relevance diminishes rapidly over time. Interactive behaviors (such as saving, forwarding, and commenting) reflect a user's lasting preferences, with a relatively slow decay rate. Negative feedback behaviors (such as disinterest) have a phased characteristic; recent negative feedback signals should be given high priority, while the influence of long-term negative feedback needs to be gradually reduced. To address these differences, differentiated time decay functions are designed for the three types of behavior, and the parameters are dynamically adjusted by combining candidate content attributes and user profiles.

[0102] For click behavior, an exponential decay function is used to model time weights. Let the item... The time interval for the click action is The timeliness attribute value of the candidate content is The base decay factor for click behavior is The actual attenuation coefficient It is determined by both the basic attenuation coefficient and the time-dependent properties: ,in This is a timeliness adjustment factor. Timeliness attribute value. The larger the value, the higher the requirement for freshness in the content (such as breaking news or real-time trending topics); the lower the decay coefficient. As this increases, the weight of historical click behavior decays more quickly, thus highlighting the importance of recent behavioral signals. This is especially relevant for content with lower timeliness (such as tutorials and knowledge-based content). The value is small. Even at near baseline values, historical click behavior still retains high reference value. Click behavior time decay weight. Calculate as follows: ,in Normalization is performed on an hourly basis to ensure consistent decay behavior across different time granularities.

[0103] For interactive behavior, a power-law decay function is used to model time weights to reflect the long-tail persistence of interactive preferences. Power-law decay is slow in the initial stage but retains a certain weight in the long term, consistent with the formation mechanism of users' persistent interests. Let items... The time interval between interactive behaviors is The interaction depth attribute value of the candidate content is The basic decay index of interactive behavior is The actual decay index Adjust as follows: ,in This is an interaction depth adjustment factor. Interaction depth attribute value. The larger the value, the more deeply the content can engage users (e.g., in-depth comments, multiple saves and shares), and the lower the decay index. This decreases accordingly, making the power-law curve flatter, and allowing the weight of historical interactions to remain at a high level in the long term. (Time decay weight of interactions) The calculation is as follows: In order to avoid To mitigate the singularity of time, the denominator is increased by 1 for smoothing. In parameter design, the decay rate of interactive behavior is required to be generally less than that of click behavior; that is, within the same time interval... Through the and Initialization constraints are used to ensure that this relationship holds.

[0104] To address negative feedback behavior, a piecewise decay function is employed to finely characterize the weight changes across different time periods. The time axis is divided into recent windows. Mid-term window and long-term window The system is divided into three intervals, each employing a different decay strategy. Within the recent window, negative feedback signals are treated as high-confidence expressions of current aversion, and the time decay weight remains constant. This ensures that recent negative feedback has a stable inhibitory effect on the ranking results. Within the medium-term window, the user's aversion to relevant content begins to decrease linearly over time, and the time decay weight decreases according to a linear function from... Decrease to The decreasing slope is Within the long-term window, the reference value of negative feedback signals decreases significantly. Exponential decay is used to rapidly reduce the weights to zero, with a decay coefficient of [value missing]. A three-segment function at boundary points and To maintain continuity and avoid instability in the sorting caused by sudden changes in weights.

[0105] Time Boundary and The algorithm adaptively adjusts based on the frequency of user negative feedback behavior. Let the frequency of user negative feedback behavior within the statistics window be... (Unit: times / day), then the recent window boundary Determine as follows: ,in Basic boundary duration, This is the frequency adjustment coefficient. The frequency of negative feedback behavior. The higher the value, the more sensitive the user is to content quality; the closer the window boundary... This shortens the recent constant weight range, allowing the negative feedback signal to enter the linear decay phase more quickly, thus avoiding excessive accumulation of historical signals for high-frequency negative feedback users. (Medium-term window boundary) Similarly The length is increased and then appropriately shortened to maintain the rationality of the three-segment ratio.

[0106] Based on the aforementioned differentiated decay functions, a user-personalized time-aware mechanism is introduced to extract behavioral activity features from user profiles and uniformly adjust the parameters of various decay functions. Behavioral activity features This reflects the frequency of user interaction on the platform, and is typically calculated using a combination of multiple metrics, including recent login days, average daily activity frequency, and content consumption duration, with values ​​normalized to [value missing]. The range is defined as follows: a larger value indicates a more active user. For users with higher behavioral activity characteristic values, their behavioral data is updated frequently, and the timeliness of historical behavior is relatively short. Therefore, the click decay coefficient should be increased accordingly to give more weight to recent click signals. Specifically, the personalized decay coefficient for click behavior is... Adjustments were made based on the existing foundation: ,in This is an activity adjustment factor for click behavior. For users with low activity characteristic values, their interaction behavior is sparse, making interaction behavior data particularly valuable. Therefore, the effective time window of the interaction signal should be extended, and the interaction decay index should be reduced accordingly. The personalized decay index for interaction behavior is adjusted as follows: ,in As a moderating factor for the activity level of interactive behavior, when When it approaches 0, Approaching The decay curve is flatter, and historical interaction behavior can still contribute effective weight over a longer time span.

[0107] In actual engineering deployment, behavioral activity characteristics Retrieved in real time from user profiling services, along with the timeliness attributes of candidate content. Interaction depth attributes and frequency of negative feedback behavior Together, they constitute the dynamic input of the attenuation parameter. Each adjustment factor... , , , , Parameter tuning was performed offline on the validation set using ranking metrics (such as NDCG and AUC) as the optimization objective, and the parameters were periodically recalibrated as user behavior distribution changed. Finally, personalized time-decay weights for the three types of behavior were determined. , , Each value is multiplied by the attention weight corresponding to the behavior type to obtain the overall weight. The time decay component is fed into the subsequent sequence representation aggregation stage, providing a time-aware behavioral weight basis for the generation of click sequence representation, interaction sequence representation and negative feedback sequence representation.

[0108] The three sequence representations are cross-referenced with candidate content features. The click sequence representation and the interaction sequence representation are constrained by orthogonal loss to maximize their cosine distance, capturing immediate interest and persistent preference respectively. The three cross-referenced feature representations are then input into their respective prediction networks, outputting click probability, interaction probability, and negative feedback probability, including:

[0109] Multiple attribute vectors of candidate content are concatenated to construct a candidate content feature vector. Three sequence representations are then cross-referenced with the candidate content feature vector to obtain three cross-feature representations.

[0110] In the sequence representation layer and feature cross layer, cosine similarity between click sequence representation and interaction sequence representation is calculated to construct a two-layer orthogonal constraint term. Cosine similarity between click sequence representation and interaction sequence representation and negative feedback sequence representation is calculated to construct a negative feedback reverse association loss term. The two-layer orthogonal constraint term and negative feedback reverse association loss term are jointly optimized with the prediction loss to transform the parameters of the generation network, click prediction network, interaction prediction network and negative feedback prediction network.

[0111] By jointly optimizing the click sequence representation to capture immediate interest patterns, the interaction sequence representation to capture persistent preferences, and the negative feedback sequence representation to capture aversion patterns, the three cross feature representations inherit the characteristics of the corresponding sequence representations and are then fed into the click prediction network, the interaction prediction network, and the negative feedback prediction network, respectively, and output the click probability, interaction probability, and negative feedback probability.

[0112] The construction of candidate content feature vectors begins with the concatenation of multiple attributes. For candidate content to be ranked, the attribute dimensions cover multiple dimensions such as content format (article, image and text, post, etc.), topic category (e.g., technology, sports, food), publication time, content duration, and creator tags. These attributes are mapped into dense vectors through an embedding layer, and then concatenated in dimensional order to form a unified candidate content feature vector. This vector carries a complete description of the candidate content in multiple semantic spaces, providing a unified alignment basis for the feature intersection of the subsequent three-way sequence representations and the candidate content.

[0113] Feature crossover operations represent click sequences respectively Interactive sequence representation and negative feedback sequence characterization Execute independently. Taking click sequence representation as an example, With candidate content feature vectors After concatenation, the data is fed into a multi-layer fully connected network. The network explicitly captures the matching relationship between the two data points through element-wise multiplication and difference operations, and outputs a click-cross feature representation. Interaction sequence representation Characterization with negative feedback sequences Using cross-networks with the same structure but independent parameters, interactive cross-feature representations are generated separately. Cross-feedback feature representation The parameters of the three-way cross-network are completely decoupled, ensuring that each feature maintains its independent expressive power in the semantic space and will not cause semantic confusion due to parameter sharing.

[0114] The introduction of orthogonality constraints is the core mechanism for differentiating the three-way sequence representations into different semantic spaces. At the sequence representation layer, computation... and The cosine similarity is denoted as . Construct orthogonal constraint terms for sequence representation layers This is achieved by minimizing this term to maximize the cosine distance between the two elements, meaning they are as orthogonal as possible in the vector space. At the feature intersection layer, the calculation... and The cosine similarity is denoted as . Construct orthogonal constraint terms for cross-feature layers By weighted summing of the two levels of constraint terms, we obtain the two-level orthogonal constraint terms. ,in and These are the orthogonal constraint weight hyperparameters for the sequence representation layer and the cross-feature layer, respectively. The dual-layer constraint design ensures that orthogonality is strengthened in both the representation generation and feature fusion stages, preventing semantic overlap degradation that occurs in the other layer when constraints are applied only in one layer.

[0115] The design logic for the negative feedback reverse correlation loss term differs from that of orthogonal constraints. Negative feedback sequence representation. The captured aversion patterns should semantically contrast with the user's positive interests (i.e., the immediate interests and persistent preferences represented by click sequence representations and interaction sequence representations). To this end, separate calculations are performed. and cosine similarity as well as and cosine similarity Construct a negative feedback inverse correlation loss term By minimizing during training This drives the negative feedback sequence representation to move away from the positive sequence representation in the vector space, thereby making the negative feedback signal's contribution to the ranking score deduction more accurate and preventing confusion with the positive interest signal.

[0116] The joint optimization objective integrates the predicted loss with the losses associated with orthogonal constraints and negative feedback. Predicted Loss The loss function is a weighted average of the cross-entropy losses of three prediction networks, corresponding to the supervision signals of click, interaction, and negative feedback labels, respectively. The final joint loss function is... ,in and This represents the global balancing coefficient for each auxiliary loss term. Through end-to-end backpropagation, joint optimization simultaneously applies to all parameters of the query vector generation network (i.e., the transformation parameter generation network that generates click query vectors, interaction query vectors, and negative feedback query vectors from candidate content attribute vectors), the click prediction network, the interaction prediction network, and the negative feedback prediction network. In each gradient update, the prediction loss drives each network to improve prediction accuracy, the orthogonality constraint drives the semantic differentiation between click sequence representation and interaction sequence representation, and the negative feedback association loss drives the semantic opposition between aversion pattern representation and positive interest representation. These three factors mutually restrain each other, jointly shaping the distribution structure of the three-way representation in the semantic space.

[0117] The click prediction network, interaction prediction network, and negative feedback prediction network all employ a multi-layer fully connected network with sigmoid activation, but their input dimensions and hidden layer parameters are independent. The click prediction network uses cross-feature representation. Input: Click probability Interactive prediction network with Given the input, output the interaction probability. Negative feedback prediction network Input: negative feedback probability Because the three-way cross-feature representations inherit the semantic characteristics of the corresponding sequence representations and integrate the matching information of the candidate content, the click prediction network captures the probability that a user will have an immediate impulse to click on the candidate content, the interaction prediction network captures the probability that a user will have a deep interaction with the candidate content (such as saving, sharing, or commenting), and the negative feedback prediction network captures the probability that a user will have a clear aversion to the candidate content (such as disinterest). Semantically, the three outputs naturally correspond to the two components of positive score and negative deduction in the ranking score calculation.

[0118] In practical engineering deployments, the construction methods for training samples in the three prediction networks differ. Click and interaction labels are derived from users' positive behavior logs, while negative feedback labels are derived from users' explicit negative feedback logs. Since negative feedback behavior accounts for a much smaller proportion of total exposure than click behavior, negative feedback samples are upsampled during training, or a higher weight is assigned to the loss term of the negative feedback prediction network to alleviate the sample imbalance problem. Simultaneously, the weight hyperparameters of the orthogonal constraint and the negative feedback association loss are also considered. and Optimization on the validation set through offline experiments is necessary. Excessively large constraint weights can reduce the predictive power of sequence representations, while excessively small weights cannot effectively achieve semantic differentiation. In the experiments, [the following is a partial sentence and needs context:] Set within the range of 0.01 to 0.1. Setting it in the range of 0.005 to 0.05 usually achieves a good balance between prediction accuracy and semantic differentiation.

[0119] Through the aforementioned joint training mechanism, click sequence representations ultimately cluster in the vector space to reflect the user's recent browsing preferences, interaction sequence representations cluster to reflect the user's long-term stable interests, and negative feedback sequence representations cluster to reflect the user's explicitly rejected content types. The three cross-feature representations obtained by intersecting these three sequence representations with the candidate content feature vectors describe the relationship between candidate content and the user from three dimensions: immediate interest matching, persistent preference matching, and aversion avoidance. This provides semantically clear and unambiguous probabilistic inputs for subsequent ranking score calculations.

[0120] Based on candidate content and user profiles, interaction weights and negative feedback penalty weights are generated. The click probability is multiplied by the interaction probability weighted by the interaction weights to obtain a positive score, and the negative feedback probability weighted by the negative feedback penalty weights is subtracted to obtain a negative deduction score, resulting in the final ranking score, including:

[0121] A user interaction preference vector is constructed by extracting profile features representing user interaction tendencies from user profiles, and a content interaction adaptation vector is constructed by extracting attribute features representing content interaction potential from candidate content features. The interaction representation between the user interaction preference vector and the content interaction adaptation vector is calculated, and the interaction representation is input into a weight generation network to generate interaction weights.

[0122] User sensitivity vectors are constructed by extracting profile features representing user sensitivity to negative feedback from user profiles. These user sensitivity vectors and candidate content features are then input into a weight generation network to generate negative feedback penalty weights.

[0123] The click probability is used as the base score. The product of the interaction weight and the interaction probability is calculated as an interaction bonus. The base score and the interaction bonus are added together to obtain a positive score. The product of the negative feedback penalty weight and the negative feedback probability is calculated as a negative feedback deduction. The positive score is subtracted from the negative feedback deduction to obtain the final ranking score. The final ranking score comprehensively reflects the immediate click appeal, long-term interaction value, and negative feedback risk of the candidate content.

[0124] After estimating the click probability, interaction probability, and negative feedback probability, it is necessary to further generate personalized interaction weights and negative feedback penalty weights to integrate the three probabilities into a unified ranking score. The generation of interaction weights depends on the two-way matching relationship between users and candidate content, rather than fixed global coefficients. Therefore, it is necessary to extract corresponding subset features from user profiles and candidate content features respectively, and then obtain dynamic weight values ​​for the current user-content pair through interaction calculation.

[0125] User profiles extract profile features representing user interaction tendencies, specifically including historical interaction rates, distribution of interaction behavior types (e.g., the proportion of likes, favorites, shares, and comments), interaction depth preferences (preferring shallow likes or deep comments), and temporal regularity of interaction behavior. These multi-dimensional profile features are then concatenated after embedding layer mapping to obtain the user interaction preference vector. The algorithm extracts attribute features representing the content's interaction potential from the candidate content features. These features include content format (article, image and text, post, etc.), the level of interaction incentive for the content topic, the correlation between content duration and interaction threshold, and the average interaction rate of similar historical content. These attribute features are then mapped through an embedding layer and concatenated to obtain the content interaction adaptation vector. .

[0126] Calculate user interaction preference vector Content interaction adaptation vector When representing interactive elements, a combination of element-wise multiplication and concatenation is used: and element-wise product And the splicing of the two They are concatenated together to form an interactive representation vector. ,Right now Element-wise multiplication captures the degree of synergy between two vectors in the same semantic dimension, while concatenation preserves their independent information. This complementarity makes the interactive representation more complete. The interactive representation vectors... The input weight generation network consists of two fully connected layers with ReLU activation and a final Sigmoid activation, and outputs scalar interaction weights. The range of values ​​is This reflects the strength of a user's personalized tendency to interact with the current candidate content.

[0127] The generation path for negative feedback penalty weights differs from that of interaction weights. The triggering of negative feedback behaviors (such as disinterest or blocking) depends not only on the user's aversion to the content but also on the user's overall negative feedback sensitivity—some users habitually mark content they dislike, while others will not trigger negative feedback even if they are not interested. Therefore, specific user profile features representing negative feedback sensitivity are extracted from user profiles, including historical negative feedback rates, estimated trigger thresholds for negative feedback behaviors, differences in negative feedback distribution across different content categories, and recent trends in the frequency of negative feedback behaviors. These features are then embedded and mapped to form a user sensitivity vector. .

[0128] User sensitivity vector With candidate content feature vector (That is, the candidate content features obtained by concatenating multiple attribute embeddings, which have been defined in the previous steps.) After concatenation, they are input into another weight generation network, which also consists of two fully connected layers and an activation function. Finally, the output scalar negative feedback penalizes the weights through Sigmoid activation. The range of values ​​is . The larger the value, the higher the risk of the current user generating negative feedback to the current candidate content, and the greater the penalty should be applied in the final ranking score. Here, the generation network for the negative feedback penalty weight is independent of the generation network for the interaction weight, and the parameters are not shared, to ensure that the two paths focus on modeling positive interaction value and negative risk respectively.

[0129] Get the click probability Interaction probability Negative feedback probability Interaction weight and negative feedback penalty weights Then, the final ranking score is calculated according to the following logic. (Click probability) As a base score, it directly reflects the estimated immediate click-through rate of candidate content. Interaction weight is calculated. With interaction probability The product of these terms serves as an interactive additive term, i.e. This item reflects the long-term interactive value contribution of candidate content in the current user context. The positive score is obtained by adding the base score to the interaction bonus item. ,Right now .

[0130] The negative deduction item is determined by the negative feedback penalty weight. With negative feedback probability The product of , i.e. This item reflects the expected loss from negative user feedback triggered by candidate content. The final ranking score is obtained by subtracting the negative feedback deduction from the positive score. ,Right now Final ranking score This also comprehensively reflects the immediate click appeal of the candidate content (by... Contribution), long-term interactive value (by) Contribution) and negative feedback risk (by (Deduction), all three factors are weighed within a unified score space, avoiding the problem of low-quality, highly misleading content being over-recommended due to simply ranking by click-through rate.

[0131] From an engineering implementation perspective, both the interactive weight generation network and the negative feedback penalty weight generation network are trained end-to-end along with the preceding prediction network. During the training phase, the interactive weights... and negative feedback penalty weights Supervision signals are obtained from the joint loss function via gradient backpropagation, eliminating the need for separate labeled data. During the online inference phase, user profile features and candidate content features have already been calculated in the preceding processes, resulting in minimal forward computation overhead for the weight generation network, which does not significantly impact the overall ranking latency. For multiple candidate content items from the same user in the same request, the user interaction preference vector... and user sensitivity vector It only needs to be calculated once and can be reused among multiple candidate contents, further reducing computational costs.

[0132] In practical applications, interaction weight Dynamic generation allows the ranking strategy to adapt to the behavioral differences among different user groups. For users with a high tendency to interact, interaction weights are used. Overall, the score is relatively high, with interaction-related factors contributing more to the ranking score. The ranking results tend to push content with deep interaction potential. For users with low interaction tendencies, the interaction weight is lower. Overall low, click probability Ranking is dominated by specific factors, with content recommendation focusing more on immediate appeal. Negative feedback is penalized with a specific weight. Personalized generation also enables the ranking strategy to differentiate between highly sensitive and insensitive users: for users who have historically frequently flagged negative feedback, Higher values ​​result in stronger penalties for negative deductions, which can more effectively filter out candidate content that poses a risk of negative feedback to the user, thereby improving the user experience and reducing the risk of user churn.

[0133] A second aspect of the present invention provides an electronic device, comprising:

[0134] processor;

[0135] Memory used to store processor-executable instructions;

[0136] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0137] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0138] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method integrating target perception behavior sequence attention and multi-target information flow ranking, characterized in that, include: Obtain candidate content and user multi-behavior historical sequences, extract attribute vectors from candidate content, and construct click query vectors, interaction query vectors, and negative feedback query vectors based on the different attribute vectors of candidate content; Attention weights are calculated for each query vector and the user's multi-behavior history sequence. The representations of the same item in different behavior sequences are fused through a cross-behavior conversion matrix, and a time decay function is introduced to generate click sequence representations, interaction sequence representations, and negative feedback sequence representations. The three sequence representations are cross-referenced with candidate content features. The click sequence representation and the interaction sequence representation are constrained by orthogonal loss to maximize the cosine distance between them, thereby capturing immediate interest and persistent preference, respectively. The three feature representations after cross-reference are then input into the corresponding prediction networks, which output click probability, interaction probability, and negative feedback probability. Based on the candidate content and user profile, an interaction weight and a negative feedback penalty weight are generated. The click probability is multiplied by the interaction probability weighted by the interaction weight to obtain a positive score, and the negative feedback probability weighted by the negative feedback penalty weight is subtracted to obtain a negative deduction score, thus obtaining the final ranking score.

2. The method according to claim 1, characterized in that, Attribute vectors are extracted from candidate content. Click query vectors, interaction query vectors, and negative feedback query vectors are constructed based on the different attribute vectors of the candidate content, including: The attribute vectors of candidate content are grouped according to content format attributes and content theme attributes. The attribute vectors corresponding to the content format attributes are concatenated to construct click query vectors, and the attribute vectors corresponding to the content theme attributes are concatenated to construct interactive query vectors. Extract the duplication attribute vector of the candidate content, and construct a negative feedback query vector by concatenating the duplication attribute vector.

3. The method according to claim 1, characterized in that, Each query vector is weighted with the user's multi-behavior history sequence. Representations of the same item across different behavior sequences are fused using a cross-behavior conversion matrix. A time decay function is introduced to generate click sequence representations, interaction sequence representations, and negative feedback sequence representations, including: For each item in a user's multi-behavior history sequence, multiple representation vectors of the item in different behavior type sequences are identified. The click query vector, interaction query vector, or negative feedback query vector is input into a transformation parameter generation network to obtain a cross-behavior transformation matrix corresponding to the query. The cross-behavior transformation matrix is ​​used to project multiple representation vectors onto a unified feature space. The semantic matching degree between the query vector and each behavior type is calculated as the behavior fusion weight. The projected representation vectors are then weighted and fused according to the behavior fusion weight to obtain a query-aware fusion representation vector. Based on the characteristics of different behavior types, a differentiated time decay function is constructed, the time decay weight corresponding to the behavior type is calculated, the time decay weight is multiplied by the behavior fusion weight to obtain a comprehensive weight, and the query perception fusion representation vector is adjusted accordingly. The attention score is calculated by combining the query vector with the query perception fusion representation vector after comprehensive weight adjustment and normalized to obtain the attention weight. The attention weight is then used to perform weighted summation to generate click sequence representation, interaction sequence representation and negative feedback sequence representation respectively.

4. The method according to claim 3, characterized in that, The click query vector, interaction query vector, or negative feedback query vector is input into the conversion parameter generation network to obtain the cross-behavior conversion matrix corresponding to the query, including: The query vector is input into the multi-head self-attention module. Each attention head learns the feature representation of the query vector in the semantic dimension, structural dimension and temporal dimension respectively, and the two are concatenated to obtain multi-granular semantic features. The multi-granular semantic features are input into a gated fusion unit, which includes a behavior type recognition gate and a behavior intensity modulation gate. The recognition gate generates an activation probability of zero to one for each behavior type, and the modulation gate dynamically adjusts the weight amplitude of each behavior type according to the query semantic intensity. The outputs of the two gates are multiplied to generate a behavior type weight vector. Multiple sets of basic conversion matrices for different behavior types are pre-constructed, including four sets of conversion matrices: browsing to click, click to interaction, interaction to deep interaction, and any behavior to negative feedback. Each set of matrices is pre-trained using conversion statistical patterns of user behavior data. The behavior type weight vector is used to dynamically weight and fuse multiple basic conversion matrices. The click query vector makes the weight of the click matrix greater than the weight of other matrices. The interaction query vector makes the sum of the weights of the click to the interaction matrix and the interaction to the deep interaction matrix greater than the sum of the weights of other matrices. The negative feedback query vector makes the weight of any behavior to the negative feedback matrix greater than the weight of other matrices.

5. The method according to claim 3, characterized in that, Differentiated time decay functions are constructed based on the characteristics of different behavior types, and the time decay weights corresponding to the behavior types are calculated, including: An exponential decay function is set for click behavior, and the decay coefficient is adjusted according to the timeliness attribute of the candidate content. The larger the timeliness attribute value, the larger the decay coefficient. A power-law decay function is set for interactive behaviors, where the decay rate of interactive behaviors is less than that of click behaviors. The decay exponent is adjusted according to the interaction depth attribute of the candidate content, with a larger interaction depth attribute value resulting in a smaller decay exponent. A segmented decay function is set for negative feedback behavior, dividing the time window into a recent window, a medium window, and a long window. The recent window uses a constant weight, the medium window uses linear decay, and the long window uses exponential decay. The decay rate of the recent window is less than that of the medium window, and the decay rate of the medium window is less than that of the long window. The time boundary of each segment window is adaptively adjusted according to the frequency of the user's negative feedback behavior. The higher the frequency of negative feedback behavior, the shorter the time boundary of the recent window. A personalized time-aware mechanism is introduced, which extracts user behavior activity features from user profiles to adjust the parameters of the decay function. The larger the user behavior activity feature value, the larger the click decay coefficient, and the smaller the user behavior activity feature value, the smaller the interaction decay index.

6. The method according to claim 1, characterized in that, The three sequence representations are cross-referenced with candidate content features. The click sequence representation and the interaction sequence representation are constrained by orthogonal loss to maximize their cosine distance, capturing immediate interest and persistent preference respectively. The three cross-referenced feature representations are then input into their respective prediction networks, outputting click probability, interaction probability, and negative feedback probability, including: Multiple attribute vectors of candidate content are concatenated to construct a candidate content feature vector. Three sequence representations are then cross-referenced with the candidate content feature vector to obtain three cross-feature representations. In the sequence representation layer and feature cross layer, cosine similarity between click sequence representation and interaction sequence representation is calculated to construct a two-layer orthogonal constraint term. Cosine similarity between click sequence representation and interaction sequence representation and negative feedback sequence representation is calculated to construct a negative feedback reverse association loss term. The two-layer orthogonal constraint term and negative feedback reverse association loss term are jointly optimized with the prediction loss to transform the parameters of the generation network, click prediction network, interaction prediction network and negative feedback prediction network. By jointly optimizing the click sequence representation to capture immediate interest patterns, the interaction sequence representation to capture persistent preferences, and the negative feedback sequence representation to capture aversion patterns, the three cross feature representations inherit the characteristics of the corresponding sequence representations and are then fed into the click prediction network, the interaction prediction network, and the negative feedback prediction network, respectively, and output the click probability, interaction probability, and negative feedback probability.

7. The method according to claim 1, characterized in that, Based on candidate content and user profiles, interaction weights and negative feedback penalty weights are generated. The click probability is multiplied by the interaction probability weighted by the interaction weights to obtain a positive score, and the negative feedback probability weighted by the negative feedback penalty weights is subtracted to obtain a negative deduction score, resulting in the final ranking score, including: A user interaction preference vector is constructed by extracting profile features representing user interaction tendencies from user profiles, and a content interaction adaptation vector is constructed by extracting attribute features representing content interaction potential from candidate content features. The interaction representation between the user interaction preference vector and the content interaction adaptation vector is calculated, and the interaction representation is input into a weight generation network to generate interaction weights. User sensitivity vectors are constructed by extracting profile features representing user sensitivity to negative feedback from user profiles. These user sensitivity vectors and candidate content features are then input into a weight generation network to generate negative feedback penalty weights. The click probability is used as the base score. The product of the interaction weight and the interaction probability is calculated as the interaction bonus. The base score and the interaction bonus are added together to obtain the positive score. The product of the negative feedback penalty weight and the negative feedback probability is calculated as the negative feedback deduction. The positive score is subtracted from the negative feedback deduction to obtain the final ranking score.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images