Methods, systems, and media for collecting precise user preference data

Abstract

This invention discloses a method, system, and medium for collecting precise user preference data, including: collecting multi-source behavioral data, standardizing the data, and generating a standardized behavioral event sequence; constructing a preference perturbation sequence, generating a pseudo-probe chain, and identifying preference gap dimensions; constructing a preference topology tensor, folding the tensor structure, and extracting probe candidate vectors; inputting the data into an InterFormer network, analyzing rating interaction features, and generating probe content vectors; displaying the probe content, collecting user behavior, recording context and displayed information; constructing a causal graph, generating a causal regeneration kernel, and outputting high-confidence preference samples. This invention achieves precise, reliable, and interpretable data collection of users' true preferences by using preference perturbation and uncertainty analysis, generating probe content through topology tensor folding, and combining causal regeneration to construct preference samples.

Description

Technical Field

[0001] This invention relates to the field of data collection technology, and in particular to methods, systems and media for collecting precise user preference data. Background Technology

[0002] In existing internet applications, information services, and intelligent recommendation systems, user preference data is typically used as core foundational data, widely applied in scenarios such as content recommendation, advertising, search ranking, and user profiling. Current technologies primarily acquire user preference data by collecting historical user behavior, such as explicit or implicit actions like clicks, browsing, favorites, and ratings, and then characterize user preferences through statistical analysis or feature modeling. These technical solutions are simple to implement, have low deployment costs, and can reflect user interests to a certain extent, thus they are widely used in practical systems.

[0003] However, with the increasing complexity of business scenarios and the growing diversity of user behaviors, existing methods for collecting user preference data have gradually revealed significant shortcomings. On the one hand, user behavior itself is highly random and context-dependent. Single or small-scale actions are often influenced by external factors such as display location, recommendation strategies, and page layout, resulting in high noise levels in the collected preference signals, making it difficult to accurately reflect users' true preferences. On the other hand, existing technologies mostly adopt passive collection methods, relying solely on naturally generated user behavior data. They cannot proactively identify which preference dimensions have missing information, nor can they conduct targeted preference collection for key preference dimensions, thus causing user preference profiles to be incomplete or unstable across multiple dimensions.

[0004] Existing technologies often overlook the impact of display strategies on user behavior when processing preference data, failing to distinguish between genuine user preferences and recommended exposure results. This can easily lead to misjudging behaviors caused by display order or frequency as user preferences. Traditional methods often directly use behavioral outcomes as preference labels, lacking analysis of the reasons behind the behavior and making it difficult to extract causal evidence of preferences from the behavior. This results in insufficient credibility and interpretability of preference data.

[0005] Therefore, how to provide methods, systems, and media for collecting accurate user preference data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method, system, and medium for collecting precise user preference data. This invention standardizes multi-source user behavior data, constructs preference perturbation sequences, generates pseudo-preference collapse detection chains, proactively identifies information gaps in various preference dimensions, and generates high-value detection content based on preference topology tensors and improved feature interaction networks. On this basis, it collects detection behavior data and constructs a causal relationship between users, content, and display strategies, forming a causal regeneration kernel for preference evidence, thereby generating high-confidence user preference sample data. This invention fully utilizes technologies such as electronic digital data processing, feature interaction analysis, and causal relationship modeling to achieve proactive collection and precise characterization of user preferences, possessing advantages such as high preference data accuracy, strong resistance to behavioral noise, complete preference coverage dimensions, and good interpretability of results.

[0007] A method for collecting precise user preference data according to an embodiment of the present invention includes: Collect multi-source user behavior data, standardize the multi-source behavior data, and form a standardized behavior event sequence; Based on the standardized behavioral event sequence, a preference perturbation sequence is constructed. Continuous perturbation is applied to the user preference-related input feature embedding to generate a pseudo preference collapse detection chain. The multi-order uncertainty structure of the user in each preference dimension is calculated and the target preference dimension with preference gap is determined. For the target preference dimension, user preference features, multi-order uncertainty structure and candidate content features contained in the standardized behavioral event sequence are combined to construct a preference topology tensor. Tensor folding operation is performed on the preference topology tensor to generate fold clusters and gap regions, and probe candidate vectors are extracted. The probe candidate vector and the standardized behavioral event sequence are input into the improved InterFormer network. The probe candidate vector is scored by feature interaction and the probe content vector is determined. The content item corresponding to the probe content vector is retrieved from the content library to form the probe content. Based on the preset detection strategy, display the detection content, collect user detection behavior data for the detection content, and record the display location information, display strategy information and context information corresponding to the detection behavior data; Based on the detection behavior data, a user-content-display strategy causal graph is constructed. The potential preference states corresponding to the detection behavior are inferred to form a potential preference state space. A preference evidence causal regeneration kernel is constructed, and preference sample data is generated to update the user preference profile.

[0008] Optionally, the multi-source behavioral data includes explicit user interaction data with content, implicit user interaction data generated during content browsing, display location information corresponding to user behavior, and contextual feature data corresponding to when user behavior occurs.

[0009] Optionally, the standardization processing of multi-source behavioral data includes converting the collected multi-source behavioral data into standardized behavioral events according to a unified event structure. The standardized behavioral events include user identifier, content identifier, event type, event occurrence time, display order, display position, display duration, and contextual features. The event type is encoded to form an event type code, the event occurrence time is timestamped to form an event timestamp, the display order, display position, and display duration are numerically processed to form a display feature vector, the contextual features are discretized to form a contextual feature vector, and the standardized behavioral events are serialized and arranged according to the user identifier to form a standardized behavioral event sequence.

[0010] Optionally, the step of calculating the user's multi-order uncertainty structure across each preference dimension and determining the target preference dimension where a preference gap exists includes: Based on standardized behavioral event sequences, feature encoding is performed on user identifiers, content identifiers, and contextual features. Each behavioral event is mapped to an input feature embedding vector. The input feature embedding vectors are aggregated along the user dimension to obtain the basic preference feature embedding corresponding to the target user. For each preset preference dimension, the perturbation direction is determined based on the feature components associated with the preference dimension in the basic preference feature embedding. The perturbation direction is divided into enhancement direction and weakening direction, and a perturbation intensity sequence from zero to maximum perturbation intensity is set for each direction. For each preference dimension, three-stage perturbation is performed sequentially according to the enhancement direction, the competition dimension suppression direction, and the weakening direction. In each stage, the basic preference feature embedding is perturbed step by step according to the perturbation intensity sequence. Under each perturbation intensity, the perturbed basic preference feature embedding is processed to obtain the corresponding preference output result. All records obtained under the three-stage perturbation of the same preference dimension are connected in the execution order to form a pseudo preference collapse detection chain of the preference dimension. For each preference dimension, based on the preference output results corresponding to the perturbation strengths in the pseudo-preference collapse detection chain, the multi-order uncertainty structure of the preference dimension is determined, where: The first-order uncertainty is the overall dispersion of the preference output results of each preference dimension throughout the entire pseudo-preference collapse detection chain; The second-order uncertainty is the average value of the change in preference output between adjacent perturbation intensities along the pseudo-preference collapse detection chain; The third-order collapse uncertainty is the perturbation intensity that occurs when the preference output result first converges from a multi-peaked distribution to a single-peaked distribution in the pseudo-preference collapse detection chain. For each preference dimension, a comprehensive uncertainty index is obtained based on the first-order uncertainty, the second-order uncertainty, and the third-order collapse uncertainty. Preference dimensions with a comprehensive uncertainty index greater than a preset threshold are identified as target preference dimensions with preference gaps.

[0011] Optionally, the step of performing a tensor folding operation on the preferred topology tensor to generate fold clusters and gap regions, and extracting candidate probe vectors, includes: For each defined target preference dimension, user preference features, multi-order uncertainty structure, and candidate content features related to the target preference dimension are extracted from the standardized behavioral event sequence. User preference features are represented as user preference feature vectors, multi-order uncertainty structure is represented as uncertainty feature vectors, and candidate content features are represented as content feature vectors. The user preference feature vector, uncertainty feature vector, and content feature vector are arranged and combined according to the preset indexing rules in three dimensions: preference dimension index, feature type index, and candidate content index, to construct a preference topology tensor corresponding to the target user. Tensor folding is performed on the preference topology tensor. Folding is performed in the combined space of preference dimension index and feature type index. Tensor elements that are close in the numerical space are aggregated into several fold clusters. Tensor elements that are not aggregated during the folding process and are located between different fold clusters are divided into gap regions, forming a folded structure containing multiple fold clusters and gap regions. In the folded structure, for each target preference dimension, representative tensor elements are extracted from the corresponding gap region based on the distribution of tensor elements in the gap region, and the components of the representative tensor elements in the candidate content index dimension are used as the probe candidate vector.

[0012] Optionally, retrieving content items from the content library corresponding to the probe content vector to form probe content includes: User attribute features, content attribute features, and context attribute features are extracted from standardized behavioral event sequences. The three types of features are encoded into user attribute feature vectors, content attribute feature vectors, and context attribute feature vectors, respectively. The probe candidate vectors are combined with the three types of feature vectors to form the input feature sequence. Embedding processing is performed on each vector in the input feature sequence to generate an embedding vector sequence corresponding to a preset dimension. The embedding vector sequence is then input into the improved InterFormer network. An improved InterFormer network is constructed, comprising a context-sensitive feature filtering layer, a preference difference amplification layer, and a cross-feature contrast interaction layer: Context-sensitive feature filtering layer: Taking the embedded vector sequence as input, it generates the correlation strength based on the correlation between the context attribute feature vector and the other embedded vectors, assigns filtering weights to each embedded vector according to the correlation strength, performs weighted filtering on the embedded vectors according to the filtering weights, and outputs the filtered embedded vector sequence. Preference difference amplification layer: Based on the filtered embedding vector sequence, obtain the preference difference representation between the probe candidate vector and the user attribute feature vector, perform amplification processing on the preference difference representation, and output the preference difference amplified embedding vector sequence; Cross-feature contrast interaction layer: Taking the embedded vector sequence enhanced by preference difference as input, a reference feature combination for contrast is constructed on the probe candidate vector. Interaction processing is performed between the reference feature combination and the user attribute feature vector, content attribute feature vector and context attribute feature vector. The output is an embedded representation containing the contrast interaction results. The outputs of the context-sensitive feature filtering layer, the preference difference amplification layer, and the cross-feature contrast interaction layer are sequentially input into the multi-head feature interaction layer of InterFormer to generate the final interaction representation vector corresponding to the probe candidate vector. The feature interaction score of the probe candidate vector is generated based on the final interaction representation vector. The feature interaction scores of all candidate probe vectors are sorted, and the candidate probe vectors that meet the preset conditions are determined as probe content vectors. Based on the probe content vectors, the corresponding content items are retrieved from the content library to form the probe content.

[0013] Optionally, the display location information, display strategy information, and context information corresponding to the recorded detection behavior data include: Based on the determined probe content vector, the corresponding probe content item is retrieved from the content library and inserted into the content sequence to be displayed according to the preset probe strategy to form a probe display sequence. In the detection display sequence, display parameters are generated for each detection content item. The display parameters include the position index of the detection content item in the display sequence, the display duration of the detection content item, and the exposure order of the detection content item in the display sequence. The detection display sequence is presented to the user, and during the display process, the user's detection behavior data for the detection content items is collected. The detection behavior data includes the user's interaction type with the detection content items, the duration of the user's stay on the detection content items, and the number of times the user performs actions on the detection content items. The detection behavior data is associated with the display parameters and combined with the context features corresponding to the detection behavior to form a detection behavior record. The detection behavior record includes the display parameters, detection behavior data and corresponding context features. The detection behavior record is stored in the detection behavior data set according to the time sequence of the detection behavior.

[0014] Optionally, the step of constructing a preference evidence causal regeneration kernel to generate preference sample data for updating the user preference profile includes: Based on the probe behavior data set, the user identifier, content identifier, display parameters, probe behavior data and context features of each probe behavior record are extracted, which correspond to user nodes, content nodes, display strategy nodes, behavior nodes and context nodes respectively. A user-content-display strategy causal graph is constructed according to the causal relationship between user nodes pointing to display strategy nodes, display strategy nodes pointing to behavior nodes, content nodes pointing to behavior nodes, context nodes pointing to display strategy nodes and behavior nodes. Based on the cause-effect graph of user-content-display strategy, multiple potential preference states are set for each user in each preference dimension, and preference direction information and preference intensity information are associated with each potential preference state to form a set of potential preference states. For each probe behavior record, a candidate preference generation path is selected from the user node through the display strategy node and content node to the behavior node in the user-content-display strategy causal graph. Combining the display parameters, probe behavior data and context features of the probe behavior record, the consistency between the candidate preference generation path and the potential preference state is evaluated, and path weights are assigned to each candidate preference generation path. For each preference dimension, the candidate preference generation paths associated with the same potential preference state are aggregated to obtain the corresponding preference increment. The preference increments generated by different detection behavior records are time-aligned and consistency-compared. Conflicting preference increments are resolved. Preference increments with consistent direction and similar intensity are merged to construct a preference evidence causal regeneration kernel that includes a preference generation path index, a preference increment set, path weights, and consistency markers. Based on the preference increment set, consistency marker and path weight in the preference evidence causal regeneration kernel, preference sample data is generated for each preference dimension. The preference sample data includes preference direction, preference intensity and confidence information. The preference sample data is written into the preference data storage area to update the user preference profile.

[0015] A system for collecting precise user preference data according to an embodiment of the present invention includes the following modules: The data collection and standardization module is used to collect multi-source behavioral data and standardize it to form a standardized sequence of behavioral events. The perturbation analysis module is used to construct a preference perturbation sequence based on a standardized sequence of behavioral events and generate a pseudo preference collapse detection chain, calculate the multi-order uncertainty structure and determine the target preference dimension. The tensor candidate module is used to construct a preference topology tensor for the target preference dimension and perform tensor folding to extract probe candidate vectors; The probe generation module is used to input the probe candidate vectors and the standardized behavioral event sequence into the improved InterFormer network for scoring, determine the probe content vector, and generate the probe content. The display and acquisition module is used to display the detection content according to the detection strategy and collect detection behavior data and display information. The causal regeneration module is used to construct a causal graph of user-content-display strategy, generate a causal regeneration kernel of preference evidence, and output preference sample data to update user preference profiles.

[0016] A computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform a method for collecting precise user preference data.

[0017] The beneficial effects of this invention are: The proposed method for collecting precise user preference data enables standardized processing of multi-source behaviors during the user behavior data acquisition phase. It introduces preference perturbation and multi-order uncertainty analysis mechanisms to proactively identify information gaps in different preference dimensions, thus avoiding the missing preference dimensions caused by relying solely on passive behavior data in existing technologies. Through precise characterization of preference uncertainty, the system can collect preferences in a targeted manner, effectively reducing invalid data collection and redundant calculations, and improving the relevance and effectiveness of preference data acquisition.

[0018] This invention constructs a preference topology tensor and performs tensor folding operations to extract discriminative probe candidate vectors from a high-dimensional preference feature space. It then combines this with an improved feature interaction network to generate probe content, thereby enabling proactive detection of user preferences. This enhances the ability of the probe content to trigger genuine user preferences, reduces the interference of external factors such as display location and page layout on behavioral data, and obtains preference feedback data with higher information content and lower noise under the same number of interactions, thus improving the accuracy and stability of preference data.

[0019] This invention introduces a preference evidence regeneration mechanism based on causal relationships in the preference data generation stage. It performs causal path analysis and consistency verification on the detection behavior, thereby distinguishing between behaviors caused by the display strategy and behaviors driven by the user's true preferences. It generates preference sample data with causal significance and confidence level, which not only improves the credibility and interpretability of user preference data, but also provides a more stable and reliable data foundation for user profile updates and business applications. Overall, it effectively overcomes the problems of high noise, low credibility and insufficient interpretability of preference data in the prior art. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the method for collecting precise user preference data proposed in this invention; Figure 2 This is a schematic diagram of the structure of the system for collecting precise user preference data proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 Methods for collecting precise user preference data include: Collect multi-source user behavior data, standardize the multi-source behavior data, and form a standardized behavior event sequence; Based on the standardized behavioral event sequence, a preference perturbation sequence is constructed. Continuous perturbation is applied to the user preference-related input feature embedding to generate a pseudo preference collapse detection chain. The multi-order uncertainty structure of the user in each preference dimension is calculated and the target preference dimension with preference gap is determined. For the target preference dimension, user preference features, multi-order uncertainty structure and candidate content features contained in the standardized behavioral event sequence are combined to construct a preference topology tensor. Tensor folding operation is performed on the preference topology tensor to generate fold clusters and gap regions, and probe candidate vectors are extracted. The probe candidate vector and the standardized behavioral event sequence are input into the improved InterFormer network. The probe candidate vector is scored by feature interaction and the probe content vector is determined. The content item corresponding to the probe content vector is retrieved from the content library to form the probe content. Based on the preset detection strategy, display the detection content, collect user detection behavior data for the detection content, and record the display location information, display strategy information and context information corresponding to the detection behavior data; Based on the detection behavior data, a user-content-display strategy causal graph is constructed. The potential preference states corresponding to the detection behavior are inferred to form a potential preference state space. A preference evidence causal regeneration kernel is constructed, and preference sample data is generated to update the user preference profile.

[0023] In this embodiment, the multi-source behavioral data includes explicit user interaction data with content, such as click behavior, collection behavior, like behavior, rating behavior, and negative feedback behavior; implicit user interaction data generated during content browsing, such as content dwell time, scroll depth, number of repeated visits, and exit behavior; display location information corresponding to user behavior, such as the display order, display position, and display duration of content on the page; and contextual feature data corresponding to the occurrence of user behavior, such as the behavior occurrence time, terminal type, network status, and page scene information.

[0024] In this embodiment, the standardization processing of multi-source behavioral data includes converting the collected multi-source behavioral data into standardized behavioral events according to a unified event structure. The standardized behavioral events include user identifier, content identifier, event type, event occurrence time, display order, display position, display duration, and contextual features. The event type is encoded to form an event type code, the event occurrence time is timestamped to form an event timestamp, the display order, display position, and display duration are numerically processed to form a display feature vector, the contextual features are discretized to form a contextual feature vector, and the standardized behavioral events are serialized and arranged according to the user identifier to form a standardized behavioral event sequence.

[0025] In this embodiment, calculating the user's multi-order uncertainty structure across each preference dimension and determining the target preference dimension with a preference gap includes: Based on standardized behavioral event sequences, feature encoding is performed on user identifiers, content identifiers, and contextual features. Each behavioral event is mapped to an input feature embedding vector. The input feature embedding vectors are aggregated along the user dimension to obtain the basic preference feature embedding corresponding to the target user. For each preset preference dimension, the perturbation direction is determined based on the feature components associated with the preference dimension in the basic preference feature embedding. The perturbation direction is divided into enhancement and weakening directions, and a perturbation intensity sequence from zero to maximum perturbation intensity is set for each direction, where: The preset preference dimensions include: content theme preference dimension, content format preference dimension, content timeliness preference dimension, content depth preference dimension, style and tone preference dimension, interaction intensity preference dimension, novelty preference dimension, and scene context preference dimension. For each preference dimension, three-stage perturbation is performed sequentially according to the enhancement direction, the competition dimension suppression direction, and the weakening direction. In each stage, the basic preference feature embedding is perturbed step by step according to the perturbation intensity sequence. Under each perturbation intensity, the perturbed basic preference feature embedding is processed to obtain the corresponding preference output result. All records obtained under the three-stage perturbation of the same preference dimension are connected in the execution order to form a pseudo preference collapse detection chain of the preference dimension. For each preference dimension, based on the preference output results corresponding to the perturbation strengths in the pseudo-preference collapse detection chain, the multi-order uncertainty structure of the preference dimension is determined, where: The first-order uncertainty is the overall dispersion of the preference output results of each preference dimension throughout the entire pseudo-preference collapse detection chain; The second-order uncertainty is the average value of the change in preference output between adjacent perturbation intensities along the pseudo-preference collapse detection chain; The third-order collapse uncertainty is the perturbation intensity that occurs when the preference output result first converges from a multi-peaked distribution to a single-peaked distribution in the pseudo-preference collapse detection chain. For each preference dimension, a comprehensive uncertainty index is obtained based on the first-order uncertainty, the second-order uncertainty, and the third-order collapse uncertainty. Preference dimensions with a comprehensive uncertainty index greater than a preset threshold are identified as target preference dimensions with preference gaps.

[0026] In this embodiment, performing a tensor folding operation on the preferred topology tensor to generate fold clusters and gap regions, and extracting candidate probe vectors, includes: For each defined target preference dimension, user preference features, multi-order uncertainty structure, and candidate content features related to the target preference dimension are extracted from the standardized behavioral event sequence. User preference features are represented as user preference feature vectors, multi-order uncertainty structure is represented as uncertainty feature vectors, and candidate content features are represented as content feature vectors. User preference feature vectors, uncertainty feature vectors, and content feature vectors are arranged and combined according to a preset indexing rule across three dimensions: preference dimension index, feature type index, and candidate content index. This constructs a preference topology tensor corresponding to the target user. The preference dimension index is used to distinguish different target preference dimensions, the feature type index is used to distinguish between user preference features and uncertainty features, and the candidate content index is used to distinguish between different candidate content. The preset indexing rule is as follows: The first dimension uses the preference dimension index to distinguish the preset preference dimensions and number them in a fixed order; the second dimension uses the feature type index to distinguish the user preference feature vector and the uncertainty feature vector and number them in a fixed order; the third dimension uses the candidate content index to distinguish the candidate content and number them in the sorting order of the candidate content identifiers; during tensor construction, the user preference feature vector and the uncertainty feature vector corresponding to the same preference dimension and the same candidate content are aligned and filled, the missing features are filled with preset zero value vectors or mean vectors, and the feature vectors from different sources are aligned in a unified dimension and written to the corresponding index positions, thereby forming a preference topology tensor corresponding to the target user; Tensor folding is performed on the preference topology tensor. Folding is performed in the combined space of preference dimension index and feature type index. Tensor elements that are close in the numerical space are aggregated into several fold clusters. Tensor elements that are not aggregated during the folding process and are located between different fold clusters are divided into gap regions, forming a folded structure containing multiple fold clusters and gap regions. In the folded structure, for each target preference dimension, representative tensor elements are extracted from the corresponding gap region based on the distribution of tensor elements in the gap region, and the components of the representative tensor elements in the candidate content index dimension are used as the probe candidate vector.

[0027] In this embodiment, retrieving content items from the content library corresponding to the probe content vector to form probe content includes: User attribute features, content attribute features, and context attribute features are extracted from standardized behavioral event sequences. The three types of features are encoded into user attribute feature vectors, content attribute feature vectors, and context attribute feature vectors, respectively. The probe candidate vectors are combined with the three types of feature vectors to form the input feature sequence. Embedding processing is performed on each vector in the input feature sequence to generate an embedding vector sequence corresponding to a preset dimension. The embedding vector sequence is then input into the improved InterFormer network. The preset dimensions include the embedding vector dimension, the maximum length dimension of the feature sequence, and the number of heads for multi-head feature interaction. An improved InterFormer network is constructed, comprising a context-sensitive feature filtering layer, a preference difference amplification layer, and a cross-feature contrast interaction layer: Context-sensitive feature filtering layer: Taking the embedded vector sequence as input, it generates the correlation strength based on the correlation between the context attribute feature vector and the other embedded vectors, assigns filtering weights to each embedded vector according to the correlation strength, performs weighted filtering on the embedded vectors according to the filtering weights, and outputs the filtered embedded vector sequence. Preference difference amplification layer: Based on the filtered embedding vector sequence, obtain the preference difference representation between the probe candidate vector and the user attribute feature vector, perform amplification processing on the preference difference representation, and output the preference difference amplified embedding vector sequence; Cross-feature contrast interaction layer: Taking the embedded vector sequence enhanced by preference difference as input, a reference feature combination for contrast is constructed on the probe candidate vector. Interaction processing is performed between the reference feature combination and the user attribute feature vector, content attribute feature vector and context attribute feature vector. The output is an embedded representation containing the contrast interaction results. The outputs of the context-sensitive feature filtering layer, the preference difference amplification layer, and the cross-feature contrast interaction layer are sequentially input into the InterFormer multi-head feature interaction layer to generate the final interaction representation vector corresponding to the probe candidate vector. A feature interaction score for the probe candidate vector is then generated based on the final interaction representation vector. Specifically, the generation of the feature interaction score for the probe candidate vector based on the final interaction representation vector is as follows: The final interaction representation vector is input into the scoring mapping layer, which consists of a fully connected transformation and a normalization transformation in sequence. The final interaction representation vector is linearly mapped to obtain an intermediate score, and the intermediate score is normalized to obtain a feature interaction score with a value range of zero to one. When there are multiple candidate probe vectors, the above scoring mapping process is performed on the final interaction representation vector corresponding to each candidate probe vector to obtain a feature interaction score set for each candidate probe vector. The feature interaction scores of all candidate probe vectors are sorted, and the candidate probe vectors that meet the preset conditions are determined as probe content vectors. Based on the probe content vectors, the corresponding content items are retrieved from the content library to form the probe content.

[0028] The improved InterFormer network consists of an input embedding layer, a context-sensitive feature filtering layer, a preference difference amplification layer, a cross-feature contrast interaction layer, and a multi-head feature interaction layer connected in series. The input is a sequence of input features formed by combining probe candidate vectors, user attribute feature vectors, content attribute feature vectors, and context attribute feature vectors. After embedding, the input feature sequence becomes an embedding vector sequence that satisfies preset embedding vector dimensions, maximum feature sequence length dimensions, and the number of multi-head interaction heads. This embedding vector sequence enters the context-sensitive feature filtering layer, which generates filtering weights based on the correlation strength between the context attribute feature vectors and the other embedding vectors, and outputs a weighted and filtered embedding vector. The sequence of embedded vectors after weighted filtering enters the preference difference amplification layer. This layer enhances the preference difference representation between the candidate probe vector and the user attribute feature vector and outputs the preference difference-enhanced embedded vector sequence. The preference difference-enhanced embedded vector sequence enters the cross-feature contrast interaction layer. This layer constructs a reference feature combination based on the candidate probe vector and performs interaction processing with the user attribute feature vector, content attribute feature vector, and context attribute feature vector, outputting an embedded representation containing the contrast interaction results. The embedded representation enters the multi-head feature interaction layer to form the final interaction representation vector corresponding to the candidate probe vector, and the feature interaction score of the candidate probe vector is obtained from the final interaction representation vector.

[0029] In this embodiment, the display location information, display strategy information, and context information corresponding to the recorded detection behavior data include: Based on the determined probe content vector, the corresponding probe content item is retrieved from the content library and inserted into the content sequence to be displayed according to the preset probe strategy to form a probe display sequence. In the detection display sequence, display parameters are generated for each detection content item. The display parameters include the position index of the detection content item in the display sequence, the display duration of the detection content item, and the exposure order of the detection content item in the display sequence. The detection display sequence is presented to the user, and during the display process, the user's detection behavior data for the detection content items is collected. The detection behavior data includes the user's interaction type with the detection content items, the duration of the user's stay on the detection content items, and the number of times the user performs actions on the detection content items. The detection behavior data is associated with the display parameters and combined with the context features corresponding to the detection behavior to form a detection behavior record. The detection behavior record includes the display parameters, detection behavior data and corresponding context features. The detection behavior record is stored in the detection behavior data set according to the time sequence of the detection behavior.

[0030] In this embodiment, the step of constructing a preference evidence causal regeneration kernel and generating preference sample data for updating the user preference profile includes: Based on the probe behavior data set, the user identifier, content identifier, display parameters, probe behavior data and context features of each probe behavior record are extracted, which correspond to user nodes, content nodes, display strategy nodes, behavior nodes and context nodes respectively. A user-content-display strategy causal graph is constructed according to the causal relationship between user nodes pointing to display strategy nodes, display strategy nodes pointing to behavior nodes, content nodes pointing to behavior nodes, context nodes pointing to display strategy nodes and behavior nodes. Based on the cause-and-effect graph of user-content-display strategy, multiple potential preference states are set for each user across various preference dimensions. Preference direction and intensity information are associated with each potential preference state to form a set of potential preference states, where: Preference direction information refers to the directional identifier of a user's preference value on the corresponding preference dimension. It is used to characterize the user's tendency towards the set of candidate values ​​on the preference dimension, including preference for a specific value or avoidance of a specific value. Preference intensity information refers to the quantitative degree of preference direction on the corresponding preference dimension, which is used to characterize the degree of user preference for the preference direction, and is represented by an intensity value within a preset numerical range; For each probe behavior record, a candidate preference generation path is selected from the user node through the display strategy node and content node to the behavior node in the user-content-display strategy causal graph. Combining the display parameters, probe behavior data and context features of the probe behavior record, the consistency between the candidate preference generation path and the potential preference state is evaluated, and path weights are assigned to each candidate preference generation path. For each preference dimension, the candidate preference generation paths associated with the same potential preference state are aggregated to obtain the corresponding preference increment. The preference increments generated by different detection behavior records are time-aligned and consistency-compared. Conflicting preference increments are resolved. Preference increments with consistent direction and similar intensity are merged to construct a preference evidence causal regeneration kernel that includes a preference generation path index, a preference increment set, path weights, and consistency markers. Based on the preference increment set, consistency marker and path weight in the preference evidence causal regeneration kernel, preference sample data is generated for each preference dimension. The preference sample data includes preference direction, preference intensity and confidence information. The preference sample data is written into the preference data storage area to update the user preference profile.

[0031] refer to Figure 2 A system for collecting precise user preference data includes the following modules: The data collection and standardization module is used to collect multi-source behavioral data and standardize it to form a standardized sequence of behavioral events. The perturbation analysis module is used to construct a preference perturbation sequence based on a standardized sequence of behavioral events and generate a pseudo preference collapse detection chain, calculate the multi-order uncertainty structure and determine the target preference dimension. The tensor candidate module is used to construct a preference topology tensor for the target preference dimension and perform tensor folding to extract probe candidate vectors; The probe generation module is used to input the probe candidate vectors and the standardized behavioral event sequence into the improved InterFormer network for scoring, determine the probe content vector, and generate the probe content. The display and acquisition module is used to display the detection content according to the detection strategy and collect detection behavior data and display information. The causal regeneration module is used to construct a causal graph of user-content-display strategy, generate a causal regeneration kernel of preference evidence, and output preference sample data to update user preference profiles.

[0032] A computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform a method for collecting precise user preference data.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a mobile news and short video content platform. This platform provides users with information services such as news, short videos, and text / image content, with approximately five million users and about 1.5 million daily active users. The platform pushes content to users through a recommendation system, and the recommendation effect is highly dependent on the accuracy and completeness of user preference data.

[0034] In actual operation, the platform found that the traditional method of building user preference profiles based on behavioral data such as historical clicks and browsing time has significant limitations. On the one hand, the behavior of some users is highly concentrated on a few content types, resulting in the long-term lack of other preference dimensions; on the other hand, user behavior is easily affected by display strategies such as recommendation ranking and exposure position, and some click behaviors cannot truly reflect the user's preference intentions, resulting in large noise and insufficient stability in preference data, which affects the recommendation effect.

[0035] In this embodiment, the platform deploys the method for collecting precise user preference data described in this invention. The system first collects multi-source behavioral data generated by users on the client side, including behaviors such as content clicks, browsing pauses, swiping, and returning, as well as the corresponding display positions, display orders, and contextual information. This data undergoes unified structured processing to form standardized behavioral event sequences.

[0036] Based on standardized behavioral event sequences, the system constructs a preference perturbation sequence, applying continuous perturbations to the input feature embeddings related to user preferences to generate a pseudo-preference collapse detection chain. By analyzing the changes in preference output results at different perturbation stages, the system calculates the multi-order uncertainty structure of users across various preference dimensions, thereby identifying which content dimensions have insufficient preference information. For example, in test users, the system found that some users had significantly higher uncertainty in dimensions such as "financial news" and "technology news" compared to other dimensions, indicating that these types of preferences have not been fully collected.

[0037] For the identified target preference dimension, the system constructs a preference topology tensor by combining user preference features, multi-order uncertainty structure, and candidate content features. Tensor folding operations are then used to generate fold clusters and gap regions. Probe candidate vectors with high preference discrimination potential are extracted from the gap regions. The system inputs the probe candidate vectors and standardized behavioral event sequences into an improved InterFormer network, performs feature interaction scoring on the candidate vectors, and selects the probe content vectors with higher scores. Corresponding content is then retrieved from the content library to form the probe content.

[0038] During actual demonstration, the system inserts the detected content into the user's recommendation list at a low frequency according to a preset detection strategy, and controls the display position so that the detected content can be naturally perceived by the user without affecting the normal user experience. The system simultaneously collects user detection behavior data in response to the detected content and records the corresponding display parameters and context information.

[0039] After accumulating detection behavior data, the system constructs a user-content-display strategy causal graph, infers the potential preference state of detection behavior, and aggregates, verifies consistency, and resolves conflicts of preference increments generated by different detection behaviors through a preference evidence causal regeneration kernel. Finally, it generates preference sample data containing preference direction, preference strength, and confidence level, which is used to update user preference profiles.

[0040] Table 1. Comparison of the effects of the method of the present invention and the traditional method. As shown in Table 1, compared to traditional methods, the method of this invention exhibits simultaneous improvements in the coverage, stability, and usability chain of preference data. The average number of preference dimensions increased from 4.3 to 5.6, indicating that active detection and preference gap identification can fill the gaps in preference dimensions that are difficult to cover using traditional passive collection. The preference stability index increased from 0.64 to 0.79, indicating that the preference samples obtained after causal regeneration kernel processing are more consistent and less susceptible to fluctuations from single behaviors. The click-through rate of the detection content increased from 6.9% to 8.4%, and the effectiveness rate of the detection behavior increased from 46% to 61%, indicating that the generation and display strategy of the detection content is more likely to trigger high-quality feedback related to preferences, reducing the proportion of noisy samples that are clicked but do not represent preferences.

[0041] On the business outcome side, the method of this invention brings a relatively mild but stable gain: recommendation accuracy improved from 0.71 to 0.78, and next-day retention rate improved from 42.8% to 45.3%. This level of improvement is consistent with the common effects of iterations in real online systems, typically reflecting the gradual accumulation of ranking and recall effects over multiple session cycles after the improvement in preference data quality, rather than a one-time drastic leap. In particular, the simultaneous increase in the efficiency of the probing behavior and the rise in recommendation accuracy corroborate each other, indicating that this invention not only increases the amount of data but also enhances the effectiveness of the data in characterizing preferences.

[0042] The advantages are even more pronounced among cold-start users. Traditional cold-start users typically have only 3.1 preference dimensions and a stability index of 0.58, with a recommendation accuracy of 0.66 and a next-day retention rate of 39.5%, reflecting the lack and instability of preferences due to sparse behavior. After introducing this invention, the number of preference dimensions for cold-start users increases to 4.8, the stability index increases to 0.76, the click-through rate of probed content increases to 8.0%, the effectiveness of probed behavior increases to 60%, and ultimately the recommendation accuracy increases to 0.75 and the next-day retention rate increases to 43.1%. This demonstrates that when preference signals are insufficient, this invention can more quickly fill in key preference dimensions through a preference gap identification—probe content generation—causal regeneration link to construct preference samples, reducing biases caused by display strategies, thereby forming usable user preference profiles more quickly and improving the user experience.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for collecting precise user preference data, characterized in that, include: Collect multi-source user behavior data, standardize the multi-source behavior data, and form a standardized behavior event sequence; Based on the standardized behavioral event sequence, a preference perturbation sequence is constructed. Continuous perturbation is applied to the user preference-related input feature embedding to generate a pseudo preference collapse detection chain. The multi-order uncertainty structure of the user in each preference dimension is calculated and the target preference dimension with preference gap is determined. For the target preference dimension, user preference features, multi-order uncertainty structure and candidate content features contained in the standardized behavioral event sequence are combined to construct a preference topology tensor. Tensor folding operation is performed on the preference topology tensor to generate fold clusters and gap regions, and probe candidate vectors are extracted. The probe candidate vector and the standardized behavioral event sequence are input into the improved InterFormer network. The probe candidate vector is scored by feature interaction and the probe content vector is determined. The content item corresponding to the probe content vector is retrieved from the content library to form the probe content. Based on the preset detection strategy, display the detection content, collect user detection behavior data for the detection content, and record the display location information, display strategy information and context information corresponding to the detection behavior data; Based on the detection behavior data, a user-content-display strategy causal graph is constructed. The potential preference states corresponding to the detection behavior are inferred to form a potential preference state space. A preference evidence causal regeneration kernel is constructed, and preference sample data is generated to update the user preference profile.

2. The method for collecting precise user preference data according to claim 1, characterized in that, The multi-source behavioral data includes explicit user interaction data with content, implicit user interaction data generated during content browsing, display location information corresponding to user behavior, and contextual feature data corresponding to when user behavior occurs.

3. The method for collecting precise user preference data according to claim 1, characterized in that, The standardization process for multi-source behavioral data includes converting the collected multi-source behavioral data into standardized behavioral events according to a unified event structure. The standardized behavioral events include user identifier, content identifier, event type, event occurrence time, display order, display position, display duration, and contextual features. The event type is encoded to form an event type code, the event occurrence time is timestamped to form an event timestamp, the display order, display position, and display duration are numerically processed to form a display feature vector, the contextual features are discretized to form a contextual feature vector, and the standardized behavioral events are serialized and arranged according to the user identifier to form a standardized behavioral event sequence.

4. The method for collecting precise user preference data according to claim 1, characterized in that, The calculation of the user's multi-order uncertainty structure across each preference dimension and the determination of the target preference dimension with a preference gap include: Based on standardized behavioral event sequences, feature encoding is performed on user identifiers, content identifiers, and contextual features. Each behavioral event is mapped to an input feature embedding vector. The input feature embedding vectors are aggregated along the user dimension to obtain the basic preference feature embedding corresponding to the target user. For each preset preference dimension, the perturbation direction is determined based on the feature components associated with the preference dimension in the basic preference feature embedding. The perturbation direction is divided into enhancement direction and weakening direction, and a perturbation intensity sequence from zero to maximum perturbation intensity is set for each direction. For each preference dimension, three-stage perturbation is performed sequentially according to the enhancement direction, the competition dimension suppression direction, and the weakening direction. In each stage, the basic preference feature embedding is perturbed step by step according to the perturbation intensity sequence. Under each perturbation intensity, the perturbed basic preference feature embedding is processed to obtain the corresponding preference output result. All records obtained under the three-stage perturbation of the same preference dimension are connected in the execution order to form a pseudo preference collapse detection chain of the preference dimension. For each preference dimension, based on the preference output results corresponding to the perturbation strengths in the pseudo-preference collapse detection chain, the multi-order uncertainty structure of the preference dimension is determined, where: The first-order uncertainty is the overall dispersion of the preference output results of each preference dimension throughout the entire pseudo-preference collapse detection chain; The second-order uncertainty is the average value of the change in preference output between adjacent perturbation intensities along the pseudo-preference collapse detection chain; The third-order collapse uncertainty is the perturbation intensity that occurs when the preference output result first converges from a multi-peaked distribution to a single-peaked distribution in the pseudo-preference collapse detection chain. For each preference dimension, a comprehensive uncertainty index is obtained based on the first-order uncertainty, the second-order uncertainty, and the third-order collapse uncertainty. Preference dimensions with a comprehensive uncertainty index greater than a preset threshold are identified as target preference dimensions with preference gaps.

5. The method for collecting precise user preference data according to claim 1, characterized in that, The step of performing a tensor folding operation on the preferred topology tensor to generate fold clusters and gap regions, and extracting candidate probe vectors, includes: For each defined target preference dimension, user preference features, multi-order uncertainty structure, and candidate content features related to the target preference dimension are extracted from the standardized behavioral event sequence. User preference features are represented as user preference feature vectors, multi-order uncertainty structure is represented as uncertainty feature vectors, and candidate content features are represented as content feature vectors. The user preference feature vector, uncertainty feature vector, and content feature vector are arranged and combined according to the preset indexing rules in three dimensions: preference dimension index, feature type index, and candidate content index, to construct a preference topology tensor corresponding to the target user. Tensor folding is performed on the preference topology tensor. Folding is performed in the combined space of preference dimension index and feature type index. Tensor elements that are close in the numerical space are aggregated into several fold clusters. Tensor elements that are not aggregated during the folding process and are located between different fold clusters are divided into gap regions, forming a folded structure containing multiple fold clusters and gap regions. In the folded structure, for each target preference dimension, representative tensor elements are extracted from the corresponding gap region based on the distribution of tensor elements in the gap region, and the components of the representative tensor elements in the candidate content index dimension are used as the probe candidate vector.

6. The method for collecting precise user preference data according to claim 1, characterized in that, The step of retrieving content items from the content library that correspond to the probe content vector to form probe content includes: User attribute features, content attribute features, and context attribute features are extracted from standardized behavioral event sequences. The three types of features are encoded into user attribute feature vectors, content attribute feature vectors, and context attribute feature vectors, respectively. The probe candidate vectors are combined with the three types of feature vectors to form the input feature sequence. Embedding processing is performed on each vector in the input feature sequence to generate an embedding vector sequence corresponding to a preset dimension. The embedding vector sequence is then input into the improved InterFormer network. An improved InterFormer network is constructed, comprising a context-sensitive feature filtering layer, a preference difference amplification layer, and a cross-feature contrast interaction layer: Context-sensitive feature filtering layer: Taking the embedded vector sequence as input, it generates the correlation strength based on the correlation between the context attribute feature vector and the other embedded vectors, assigns filtering weights to each embedded vector according to the correlation strength, performs weighted filtering on the embedded vectors according to the filtering weights, and outputs the filtered embedded vector sequence. Preference difference amplification layer: Based on the filtered embedding vector sequence, obtain the preference difference representation between the probe candidate vector and the user attribute feature vector, perform amplification processing on the preference difference representation, and output the preference difference amplified embedding vector sequence; Cross-feature contrast interaction layer: Taking the embedded vector sequence enhanced by preference difference as input, a reference feature combination for contrast is constructed on the probe candidate vector. Interaction processing is performed between the reference feature combination and the user attribute feature vector, content attribute feature vector and context attribute feature vector. The output is an embedded representation containing the contrast interaction results. The outputs of the context-sensitive feature filtering layer, the preference difference amplification layer, and the cross-feature contrast interaction layer are sequentially input into the multi-head feature interaction layer of InterFormer to generate the final interaction representation vector corresponding to the probe candidate vector. The feature interaction score of the probe candidate vector is generated based on the final interaction representation vector. The feature interaction scores of all candidate probe vectors are sorted, and the candidate probe vectors that meet the preset conditions are determined as probe content vectors. Based on the probe content vectors, the corresponding content items are retrieved from the content library to form the probe content.

7. The method for collecting precise user preference data according to claim 1, characterized in that, The recorded detection behavior data includes the following display location information, display strategy information, and context information: Based on the determined probe content vector, the corresponding probe content item is retrieved from the content library and inserted into the content sequence to be displayed according to the preset probe strategy to form a probe display sequence. In the detection display sequence, display parameters are generated for each detection content item. The display parameters include the position index of the detection content item in the display sequence, the display duration of the detection content item, and the exposure order of the detection content item in the display sequence. The detection display sequence is presented to the user, and during the display process, the user's detection behavior data for the detection content items is collected. The detection behavior data includes the user's interaction type with the detection content items, the duration of the user's stay on the detection content items, and the number of times the user performs actions on the detection content items. The detection behavior data is associated with the display parameters and combined with the context features corresponding to the detection behavior to form a detection behavior record. The detection behavior record includes the display parameters, detection behavior data and corresponding context features. The detection behavior record is stored in the detection behavior data set according to the time sequence of the detection behavior.

8. The method for collecting precise user preference data according to claim 1, characterized in that, The construction of the preference evidence causal regeneration kernel, generating preference sample data for updating user preference profiles, includes: Based on the probe behavior data set, the user identifier, content identifier, display parameters, probe behavior data and context features of each probe behavior record are extracted, which correspond to user nodes, content nodes, display strategy nodes, behavior nodes and context nodes respectively. A user-content-display strategy causal graph is constructed according to the causal relationship between user nodes pointing to display strategy nodes, display strategy nodes pointing to behavior nodes, content nodes pointing to behavior nodes, context nodes pointing to display strategy nodes and behavior nodes. Based on the cause-effect graph of user-content-display strategy, multiple potential preference states are set for each user in each preference dimension, and preference direction information and preference intensity information are associated with each potential preference state to form a set of potential preference states. For each probe behavior record, a candidate preference generation path is selected from the user node through the display strategy node and content node to the behavior node in the user-content-display strategy causal graph. Combining the display parameters, probe behavior data and context features of the probe behavior record, the consistency between the candidate preference generation path and the potential preference state is evaluated, and path weights are assigned to each candidate preference generation path. For each preference dimension, the candidate preference generation paths associated with the same potential preference state are aggregated to obtain the corresponding preference increment. The preference increments generated by different detection behavior records are time-aligned and consistency-compared. Conflicting preference increments are resolved. Preference increments with consistent direction and similar intensity are merged to construct a preference evidence causal regeneration kernel that includes a preference generation path index, a preference increment set, path weights, and consistency markers. Based on the preference increment set, consistency marker and path weight in the preference evidence causal regeneration kernel, preference sample data is generated for each preference dimension. The preference sample data includes preference direction, preference intensity and confidence information. The preference sample data is written into the preference data storage area to update the user preference profile.

9. A system for collecting precise user preference data, comprising the method for collecting precise user preference data as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data collection and standardization module is used to collect multi-source behavioral data and standardize it to form a standardized sequence of behavioral events. The perturbation analysis module is used to construct a preference perturbation sequence based on a standardized sequence of behavioral events and generate a pseudo preference collapse detection chain, calculate the multi-order uncertainty structure and determine the target preference dimension. The tensor candidate module is used to construct a preference topology tensor for the target preference dimension and perform tensor folding to extract probe candidate vectors; The probe generation module is used to input the probe candidate vectors and the standardized behavioral event sequence into the improved InterFormer network for scoring, determine the probe content vector, and generate the probe content. The display and acquisition module is used to display the detection content according to the detection strategy and collect detection behavior data and display information. The causal regeneration module is used to construct a causal graph of user-content-display strategy, generate a causal regeneration kernel of preference evidence, and output preference sample data to update user preference profiles.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed on a computer device, cause the computer device to perform a method for collecting precise user preference data according to any one of claims 1 to 8.

Images