A Deep Learning-Based Method and System for Constructing Multimodal User Profiles

Abstract

This invention relates to the field of e-commerce data processing technology, and discloses a method and system for constructing multimodal user profiles based on deep learning. The method collects multimodal features from target users' text comments, image browsing, and behavioral logs. It then extracts text sentiment and image focus areas using a deep learning model. Combining these features, it analyzes users' explicit and implicit purchase trends, uncovering conflict characteristics between these trends and actual purchase actions. Subsequently, it determines the conflict type based on the trigger frequency and time distribution of conflict features within a fixed period, analyzing the user's cognitive state. Finally, it analyzes conversion characteristics based on the user's decision-making process, identifies factors hindering purchase conversion, and integrates the cognitive conflict attributes of these factors and conflict types to generate dynamic psychological tags, thereby constructing a user profile for the target user. Its main purpose is to address the technical problems of existing multimodal user profiles that neglect modal semantic conflicts and behavioral discrepancies, lacking psychological insight and dynamic adaptability.

Description

Technical Field

[0001] This invention relates to the field of e-commerce data processing technology, and in particular to a method and system for constructing multimodal user profiles based on deep learning. Background Technology

[0002] Multimodal user profiling technology is a key technology supporting precise marketing, user behavior analysis, and personalized services on platforms. Existing technologies primarily focus on user interaction data such as text reviews, image browsing, and behavior logs. Relying on traditional machine learning or shallow deep learning models, they extract surface features such as explicit consumer preferences, basic browsing behavior, and simple textual sentiment to construct static user profiles. This type of technology only stays at the level of integrating and analyzing external user behavior data, focusing on capturing explicitly expressed consumer tendencies to achieve basic user segmentation and preference matching. It is currently the mainstream technical solution for user operations in the e-commerce industry.

[0003] Existing multimodal user profiling technologies have significant shortcomings. The core problem lies in ignoring semantic conflicts between multimodal data such as text, images, and behavior, as well as the discrepancies between user purchase intentions and actual behavior. Such conflicting data is often discarded as invalid noise, failing to uncover deep cognitive contradictions within users. Furthermore, they lack in-depth insights into user psychological states, with profiling dimensions limited to basic behaviors and explicit preferences, making it difficult to analyze cognitive blind spots, cognitive dissonance, and other psychological characteristics. Moreover, the generated user profiles are static, lacking dynamic adaptability and failing to update dynamically with real-time user behavior. They also cannot accurately identify obstacles to purchase conversion, resulting in insufficient profiling accuracy and an inability to support refined and personalized e-commerce operational decisions. Summary of the Invention

[0004] This invention provides a method and system for constructing multimodal user profiles based on deep learning. Its main purpose is to solve the technical problems of existing multimodal user profiles that ignore modal semantic conflicts and behavioral deviations, and lack psychological insight and dynamic adaptability.

[0005] To achieve the above objectives, the present invention provides a method for constructing multimodal user profiles based on deep learning, comprising: Step 1: Collect multimodal features of the target user, including text comment data, image browsing data, and behavior log data. Extract the target user's sentiment tendency from the text comment data and the target user's focus area from the image browsing data based on a deep learning model. Step 2: Based on the sentiment tendency and the focus area, merge and generate the target user's explicit and implicit purchase trends; based on the actual purchase actions in the behavior log data and the explicit and implicit purchase trends, mine the target user's conflict characteristics. Step 3: Based on the triggering frequency and time distribution of the conflict characteristics within a fixed period, determine the conflict type of the target user, and analyze the cognitive state of the target user based on the conflict type; Step 4: Collect the decision-making process of the target user, and combine the conflict type and the decision-making process to deduce the conversion characteristics of the target user, match the decision-making performance of the cognitive state and the conversion characteristics, and identify the obstacles that prevent the target user from completing the purchase conversion; Step 5: Integrate the cognitive conflict attributes of the hindering factors and the conflict type to generate dynamic psychological tags for the target user, and construct a user profile for the target user based on the dynamic psychological tags.

[0006] Preferably, the step of fusing and generating the explicit and implicit purchase trends of the target user based on the sentiment tendency and the focus area includes: Extract product attribute words corresponding to the sentiment from the text review data to form the explicit preferences of the target user, and extract the visual features of the focal area to form the implicit preferences of the target user; By integrating the explicit and implicit preferences and assessing the cognitive state of the target user, the explicit and implicit purchasing trends of the target user are obtained.

[0007] Preferably, the step of mining the conflict characteristics of the target user based on the actual purchase actions and the explicit / implicit purchase trends in the behavior log data includes: Based on the actual purchase actions in the behavior log data, locate the hesitation behavior segments in the behavior log data that correspond to the explicit and implicit trends, and obtain the decision breakpoint of the target user; By combining the change in sentiment before and after the decision breakpoint with the frequency of revisiting the focal area, the deviation of the target user's intentional behavior is obtained; Filter out abnormal decision points among the decision breakpoints where the deviation of the intended behavior is abnormal, extract the product attribute features corresponding to the abnormal decision points, and obtain the conflict features of the target user.

[0008] Preferably, determining the conflict type of the target user based on the trigger frequency and time distribution of the conflict characteristics within a fixed period includes: Extract the trigger points and trigger frequencies of the conflict features of the target user within a fixed period, and arrange the trigger points in ascending order of time to obtain the trigger time sequence of the conflict features; Calculate the time interval between the trigger points in the trigger time sequence, and compare the differences between the trigger points corresponding to each time interval to obtain the interval difference characteristics between the trigger points; The conflict type of the target user is determined based on the interval difference characteristics and the trigger frequency.

[0009] Preferably, the step of analyzing the cognitive state of the target user based on the conflict type includes: Extract the cognitive component elements of the target user in the conflict type, and identify the cognitive blind spots of the target user based on the cognitive component elements; The cognitive blind spots are mapped to the actual decision points in the behavior log data to verify the degree of influence of the cognitive blind spots on the actual decision points, and a cognitive dissonance signal is generated based on the degree of influence. Based on the intensity of the cognitive dissonance signal and the correlation of the cognitive components, the decision-making tendency bias of the target user is derived, and the decision-making tendency bias is taken as the cognitive state of the target user.

[0010] Preferably, the step of deriving the conversion characteristics of the target user by combining the conflict type and the decision chain includes: Extract the consumer psychology opposition structure of the target user in the conflict type, and identify the conflict injection point in the decision-making link based on the consumer psychology opposition structure; Extract the user behavior intent direction of the nodes before and after the conflict injection point, and analyze the intent deviation tendency of the target user under the switching direction by comparing the switching direction of the user behavior intent direction of the nodes before and after the conflict injection point. By combining the decision compensation orientation of the intention deviation tendency and the decision inhibition attribute of the conflict type, the conversion characteristics of the target user before the actual purchase action are determined.

[0011] Preferably, the step of matching the cognitive state with the decision-making performance of the conversion characteristics to identify the obstacles preventing the target user from completing the purchase conversion includes: Extract the cognitive demand profile corresponding to the conflict type from the cognitive state, and transform the cognitive demand profile into an ideal decision path based on the information carrying type in the decision-making link, and extract the actual decision path from the decision performance of the transformation feature; Filter the decision segments in which the actual decision path appears relative to the ideal decision path, retrieve the cognitive response types that the decision segments are bound to in the ideal decision path, extract the set of cognitively vulnerable areas pointed to by the conflict types, compare the type attribution of the cognitive response types with the set of cognitively vulnerable areas, and obtain the decision obstacle points; By summarizing all the information interface attributes and interaction elements of the decision-making obstacles, the exclusionary content that the target user cannot complete cognitive supplementation is identified, and this exclusionary content is used as an obstacle for the target user to complete the purchase conversion.

[0012] Preferably, the step of fusing the cognitive conflict attributes of the hindering factors and the conflict type to generate the dynamic psychological label of the target user includes: Match the avoidance behavior of the obstacle factors with the cognitive conflict attributes of the conflict type to obtain the conflict obstacle linkage record, and extract the psychological friction identifier from the conflict obstacle linkage record; The psychological friction markers in each of the aforementioned conflict and obstacle linkage records are summarized to obtain the conflict psychological theme of the target user, and the conflict psychological theme is combined with the cognitive state to obtain the dynamic psychological label of the target user.

[0013] Preferably, constructing the user profile of the target user based on the dynamic psychological tags includes: By merging the conflict type identifier in the dynamic psychological tags with the cognitive state, a psychological inhibition distribution record of the target user is generated. By integrating the psychological inhibition distribution records with the multimodal features, a user profile of the target user is obtained.

[0014] To address the aforementioned problems, this invention also provides a deep learning-based multimodal user profile construction system, comprising: Data acquisition module: Collects multimodal features of target users, including text comment data, image browsing data and behavior log data, and extracts the sentiment tendency of the target users in the text comment data and the focus area of ​​the target users in the image browsing data based on a deep learning model; Conflict Analysis Module: Based on the sentiment tendency and the focus area, it merges and generates the explicit and implicit purchase trends of the target user, and mines the conflict characteristics of the target user based on the actual purchase actions in the behavior log data and the explicit and implicit purchase trends; Cognitive analysis module: Based on the frequency and time distribution of the conflict features within a fixed period, determine the conflict type of the target user, and analyze the cognitive state of the target user based on the conflict type; Obstacle identification module: Collects the decision-making process of the target user, and combines the conflict type and the decision-making process to deduce the conversion characteristics of the target user, matches the decision-making performance of the cognitive state with the conversion characteristics, and identifies the obstacles that prevent the target user from completing the purchase conversion; Profile generation module: Integrates the cognitive conflict attributes of the hindering factors and the conflict type to generate dynamic psychological tags for the target user, and constructs a user profile for the target user based on the dynamic psychological tags.

[0015] Beneficial effects This solution collects three types of multimodal data from users: text comments, image browsing, and behavior logs. It then uses a dual-branch deep learning algorithm to extract text sentiment and image focus areas, enabling the simultaneous capture of explicit preferences in user language expression and implicit preferences in visual attention. This allows for accurate prediction of explicit and implicit trends in user purchases and solves the problem of existing technologies relying solely on single-modal data and ignoring potential user needs.

[0016] This solution identifies conflict characteristics by locating user decision-making breakpoints and calculating the deviation between purchase intention and actual behavior. It then automatically determines the conflict type based on the frequency and time distribution of conflict triggers. This solves the problem of existing technologies discarding multimodal semantic conflicts as data noise, and has the benefits of significantly improving the accuracy of user preference identification and efficiently processing massive amounts of real-time user behavior data.

[0017] This solution analyzes the cognitive elements of users corresponding to different conflict types, identifies users' cognitive blind spots, and quantifies their impact on purchasing decisions. It achieves automated and standardized assessment of users' cognitive states, realizing a deep transformation from surface behavioral data to deep psychological states, and solving the problems of existing technologies lacking user psychological insights and having a single user profile dimension.

[0018] This solution combines the complete decision-making process of users to analyze conversion characteristics, accurately pinpoints the specific reasons that prevent users from completing a purchase, integrates the hindering factors and cognitive conflict attributes to generate dynamic psychological tags, and finally constructs a complete user profile that includes the distribution of psychological inhibitions, thus solving the problems of existing user profiles being static and lacking dynamic adaptability. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for constructing multimodal user profiles based on deep learning, provided in an embodiment of the present invention. Figure 2 This is a functional block diagram of a deep learning-based multimodal user profile construction system provided in an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] This application provides a method and system for constructing multimodal user profiles based on deep learning. The execution entity of the method and system includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method and system for constructing multimodal user profiles based on deep learning can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0022] Reference Figure 1 This is a flowchart illustrating a deep learning-based multimodal user profile construction method according to an embodiment of the present invention. In this embodiment, the deep learning-based multimodal user profile construction method includes: Step 1: Collect multimodal features of the target user, including text comment data, image browsing data, and behavior log data. Extract the target user's sentiment tendency from the text comment data and the target user's focus area from the image browsing data based on a deep learning model. Specifically, multimodal features refer to a heterogeneous set of data collected from different interaction channels of an e-commerce platform that represents user consumption behavior and preferences. It consists of feature subsets of three independent modalities: text review data, image browsing data, and behavior log data. Each subset corresponds to a form of user interaction with the platform. The data format is uniformly structured data and stored in a distributed data warehouse. All data collection requires explicit user authorization through the platform's privacy agreement.

[0023] Text comment data refers to natural language text data posted by users on modules such as product detail pages, order review pages, and Q&A communities on e-commerce platforms. Each data entry includes a unique user identifier, posting timestamp, associated product SKU, text content, and interaction statistics fields. The data collection scope is limited to publicly authorized comments and does not include private communication data such as private messages.

[0024] Image browsing data refers to the visual interaction data generated by users when browsing product images, video covers, and details page illustrations on the platform. Each piece of data includes a unique user identifier, browsing timestamp, associated image resource ID, image pixel coordinates, dwell time, scrolling trajectory, and zoom ratio fields, which are collected in real time through front-end tracking technology.

[0025] Behavior log data is a record of user actions throughout their entire lifecycle on the platform. Each data entry includes a unique user identifier, an operation timestamp, an operation type, an operation object ID, a page redirection path, and a session ID. Operation types cover the entire process of product search, favorites, adding to cart, placing an order, payment, order cancellation, and refund. The data is stored on a log server and retained for the longest period authorized by the user.

[0026] This step of the deep learning model adopts a dual-branch parallel deep learning network architecture, which includes a text sentiment extraction branch and a visual focus extraction branch. The two branches share a unique user identifier as the association key. The model training adopts a supervised learning method, and the training set consists of historical user data labeled by the platform. The labeling process undergoes multiple rounds of cross-validation to ensure the labeling quality.

[0027] Sentiment bias is a quantified value output by the text sentiment extraction branch, with a value range of [value range missing]. The value is denoted as -1, which represents extremely negative sentiment, 0 represents neutral sentiment, and 1 represents extremely positive sentiment. The larger the absolute value, the higher the sentiment intensity. Each text comment corresponds to a sentiment tendency value, and the set of product attribute words associated with that sentiment tendency is also output.

[0028] The focal region is the set of image pixel coordinates output by the visual focus extraction branch, in the format of... This represents the rectangular area where the user's gaze lingers for the longest period of time when browsing a single image. It also outputs the visual feature vector corresponding to this area, which is used to characterize the visual attributes of the product within the area.

[0029] Furthermore, after obtaining explicit authorization from the user, the system first collects text comment data, image browsing data, and behavior log data of the target user within a fixed period, completes the cleaning process of filling missing values, deleting duplicate data, and filtering abnormal data, and converts all data into a unified structured format.

[0030] The standardized multimodal features are input into a pre-trained dual-branch deep learning model. The text sentiment extraction branch performs word segmentation, word embedding, and contextual semantic encoding on the input text comment data in sequence. Then, the fully connected layer outputs the sentiment tendency value and associated product attribute words for each comment.

[0031] The visual focus extraction branch performs time-series encoding on the user's gaze trajectory data in the input image browsing data. Combined with the convolutional neural network feature map of the corresponding image, it calculates the gaze dwell weight of each pixel region through an attention mechanism, selects the rectangular region with the highest weight as the focus region, and extracts the visual feature vector of the region.

[0032] Finally, the extracted sentiment, related product attribute words, focal area coordinates, and visual feature vectors are associated with the corresponding user's unique identifier and stored in a structured manner as input data for subsequent steps.

[0033] By employing a dual-branch parallel deep learning architecture, we achieved the correlation extraction between text sentiment and product attributes, as well as the fusion localization of visual focus and image content. This significantly improved the accuracy of text sentiment extraction, greatly enhanced the matching accuracy of associated product attribute words, and effectively improved the accuracy of visual focus region localization, providing reliable input data for subsequent conflict feature mining and cognitive state analysis.

[0034] For example, taking the interaction data of an authorized user A on an e-commerce platform over 7 days as an example, firstly, three text comments from user A were collected: the fabric of a certain dress was very comfortable but the color was too dark; the soles of a certain pair of sneakers were too hard and tiring to wear; and the thermos cup had a very good heat preservation effect. 127 image browsing data entries were collected, including the user's eye trajectory data when browsing the details page images of the above three products and 21 other similar products. 89 behavioral log data entries were collected, including records of product searches, clicks, favorites, and adding to cart.

[0035] The aforementioned multimodal features are input into a two-branch deep learning model. The text sentiment extraction branch outputs sentiment tendency values ​​of 0.2, -0.7, and 0.8 for three reviews, corresponding to the associated product attribute words: fabric, color, sole, hardness, and insulation effect. The visual focus extraction branch outputs the coordinates of the focus area when the user browses the main image of the dress. The visual feature vector of the dress fabric is extracted from the fabric display area corresponding to the dress; the coordinates of the focus area when browsing the main image of the sneakers are... The visual feature vectors of browsing sneakers are extracted from the sole area of ​​the corresponding sneaker. These extraction results are stored in a user feature database for subsequent steps, including analysis of explicit and implicit user purchase trends and mining of conflicting features.

[0036] The specific architecture of deep learning models includes: For the text sentiment extraction branch, a pre-trained language model is used as the basic encoder. Multiple fully connected layers are added after the encoder output, with activation functions including linear rectified functions and hyperbolic tangent functions to output sentiment values. During model training, cross-entropy loss and an adaptive moment estimation optimizer are used. The training set contains a large amount of labeled e-commerce review data, with labels including sentiment values ​​and related product attribute words. Early stopping is employed during training to prevent overfitting, while a learning rate decay mechanism is used to improve model convergence.

[0037] For the visual focus extraction branch, a deep residual network is used as the image feature extractor to output global image features. Simultaneously, the user's gaze trajectory data is converted into a time-series matrix, and the temporal features of the trajectory are extracted through a multi-layer long short-term memory network, outputting a temporal feature vector. The global image features and the temporal feature vector are concatenated, and a pixel-level attention weight map is generated using a multi-head attention mechanism. The pixel region with the highest value in the weight map is the focus region. During model training, a mean squared error loss function is used, and a stochastic gradient descent optimizer with a momentum term is employed. The training set contains a large amount of labeled user gaze trajectories and corresponding focus region data.

[0038] Step 2: Based on the sentiment tendency and the focus area, merge and generate the target user's explicit and implicit purchase trends; based on the actual purchase actions in the behavior log data and the explicit and implicit purchase trends, mine the target user's conflict characteristics. In this embodiment, the step of fusing and generating the explicit and implicit purchase trends of the target user based on the sentiment tendency and the focus area includes: Extract product attribute words corresponding to the sentiment from the text review data to form the explicit preferences of the target user, and extract the visual features of the focal area to form the implicit preferences of the target user; By integrating the explicit and implicit preferences and assessing the cognitive state of the target user, the explicit and implicit purchasing trends of the target user are obtained.

[0039] Specifically, product attribute words are structured lexical units extracted from user-authorized text review data that are directly related to sentiment and describe the physical or functional characteristics of the product. They are stored in the form of standardized word vectors. Each product attribute word is accompanied by a sentiment polarity label and a sentiment weight value. The sentiment weight value is equal to the absolute value of the sentiment of the corresponding review. Words with a high semantic correlation to sentiment are extracted through the attention mechanism of a pre-trained language model.

[0040] Explicit preferences are the consumption preference feature vectors of target users based on textual expression. The dimensions are consistent with the total number of product attribute categories preset by the e-commerce platform. The value of each dimension is the weighted sum of the sentiment weight values ​​of the corresponding product attribute words. The weight is the frequency of the occurrence of the attribute word in all user comments. It is obtained by extracting attribute words and calculating sentiment weights from user-authorized text comment data.

[0041] Visual features are high-dimensional feature vectors representing the visual attributes of products, extracted from the focal area of ​​image browsing data collected with user authorization. They are obtained by performing convolution operations on the cropped image of the focal area through a deep residual network and include quantifiable visual attribute information such as color, texture, material, and shape.

[0042] Implicit preferences are consumption preference feature vectors of target users based on visual attention. The dimensions are completely consistent with explicit preferences. The value of each dimension is the weighted sum of the cosine similarity between the corresponding visual feature and the standard feature vector of that attribute category. The weight is the proportion of the user's dwell time on the focal area to the total browsing time of this session. It is obtained by similarity matching and time-weighted calculation of the visual features of the focal area.

[0043] Cognitive state: This is a set of parameters that characterizes the user's current level of understanding of product information. It includes information completeness score and attribute cognition consistency score. The information completeness score reflects the degree of coverage of attribute categories of information that the user has acquired, while the attribute cognition consistency score reflects the degree of matching between the feature distributions of explicit and implicit preferences. It is calculated by comparing the feature distributions of explicit and implicit preferences.

[0044] Purchase explicit and implicit trends: These are feature vectors that quantify the likelihood of future purchases and the strength of overall preferences. They include two dimensions: purchase probability and preference strength. They are calculated by a multimodal fusion neural network by fusing explicit preferences, implicit preferences, and initial cognitive state parameters.

[0045] Furthermore, the text comment data collected with user authorization is first preprocessed by word segmentation and stop word removal. The preprocessed text is then input into a pre-trained language model for semantic encoding, outputting the sentiment tendency value and contextual semantic vector for each comment. An attention mechanism is used to calculate the correlation between each word and the sentiment tendency semantic vector, and words with high correlation are selected as product attribute words corresponding to the sentiment tendency. The frequency of occurrence and corresponding sentiment weight of each product attribute word are counted, and the categories are classified and summarized according to the platform's preset product attribute categories. The weighted sentiment sum of each category is calculated to form the explicit preference feature vector of the target user.

[0046] Simultaneously, the focus areas in the user-authorized image browsing data are uniformly cropped. The cropped images are then input into a pre-trained deep residual network to extract high-dimensional visual feature vectors. The cosine similarity between each visual feature vector and the corresponding feature in the platform's product attribute standard feature library is calculated. Combined with the user's dwell time weight for the focus area, the weighted similarity sum for each attribute category is calculated to form the latent preference feature vector of the target user.

[0047] Next, the cosine similarity between the feature vectors of explicit and implicit preferences is calculated to obtain the attribute cognition consistency score, and the number of product attribute categories covered by the user is counted to obtain the information completeness score. These are combined to form a preliminary cognitive state parameter set.

[0048] Finally, the explicit preference feature vector, implicit preference feature vector, and preliminary cognitive state parameter set are input into the pre-trained multimodal fusion model. The model achieves cross-modal interaction between text and visual features through a cross-attention mechanism, and outputs the explicit and implicit purchase trend vectors of the target user.

[0049] This solution overcomes the limitations of existing technologies that rely solely on unimodal text data to extract user preferences. By simultaneously capturing explicit needs expressed through user language and implicit needs for visual attention, it effectively addresses the one-sidedness of unimodal preference extraction. The introduction of a preliminary cognitive state assessment enables a quantitative representation of the user's information acquisition level and cognitive consistency, providing a precise benchmark for subsequent conflict feature mining. Employing a multimodal fusion model to integrate explicit and implicit preferences significantly improves the accuracy of purchase trend prediction compared to traditional unimodal prediction models, laying a solid data foundation for the subsequent accurate identification of deviations between user intentions and behaviors.

[0050] For example, taking the interaction data of an authorized user A on an e-commerce platform over 7 days as an example, firstly, the three text comments of user A are processed to extract the product attribute words and sentiment weights corresponding to sentiment tendencies: fabric 0.2, color 0.2, sole 0.7, hardness 0.7, and insulation effect 0.8. The frequency of each attribute word is counted as 1. According to the preset product attribute categories, the data is classified and summarized. The explicit preference feature vector of user A is calculated, in which the fabric dimension is 0.2, the color dimension is 0.2, the sole dimension is 0.7, the hardness dimension is 0.7, the insulation effect dimension is 0.8, and the other dimensions are 0.

[0051] Simultaneously, the focus areas in user A's image browsing data were processed. The visual feature vector of the dress fabric area showed a cosine similarity of 0.85 with the standard features of the fabric category, and the dwell time percentage was 0.6. The visual feature vector of the sneaker sole area showed a cosine similarity of 0.92 with the standard features of the sole category, and the dwell time percentage was 0.7. The visual feature vector of the thermos cup insulation effect display area showed a cosine similarity of 0.78 with the standard features of the insulation effect category, and the dwell time percentage was 0.5. The calculated latent preference feature vector of user A showed a fabric dimension of 0.51, a sole dimension of 0.644, an insulation effect dimension of 0.39, and all other dimensions of 0.

[0052] Next, the cosine similarity between the feature vectors of explicit and implicit preferences was calculated to be 0.72, resulting in an attribute cognition consistency score of 0.72. The number of product attribute categories covered by the user was counted as 5, and the total number of attribute categories was 20, resulting in an information completeness score of 0.25. These parameters were combined to form a preliminary cognitive state parameter set, which included: information completeness: 0.25, attribute cognition consistency: 0.72.

[0053] Finally, the explicit preference feature vector, implicit preference feature vector, and preliminary cognitive state parameter set are input into the multimodal fusion dual-tower model to calculate the explicit and implicit purchase trend vectors of user A. This indicates that user A has a 65% probability of making a purchase within 7 days during this period, and the overall preference strength is 7.2.

[0054] In this embodiment, the step of mining the conflict characteristics of the target user based on the actual purchase actions and the explicit and implicit trends of purchases in the behavior log data includes: Based on the actual purchase actions in the behavior log data, locate the hesitation behavior segments in the behavior log data that correspond to the explicit and implicit trends, and obtain the decision breakpoint of the target user; By combining the change in sentiment before and after the decision breakpoint with the frequency of revisiting the focal area, the deviation of the target user's intentional behavior is obtained; Filter out abnormal decision points among the decision breakpoints where the deviation of the intended behavior is abnormal, extract the product attribute features corresponding to the abnormal decision points, and obtain the conflict features of the target user.

[0055] Specifically, the actual purchase action is a specific operation record in the behavior log data that marks the user's final behavior of completing a product transaction. Each record is associated with a unique product SKU and transaction information, serving as an objective basis for judging the user's purchase behavior.

[0056] Hesitant behavior segments are continuous operation subsequences in the behavior log time series that meet the criteria for hesitant behavior, and the time span of the segments does not exceed the duration of a single user session.

[0057] Decision breakpoints: The operation nodes corresponding to the start and end times of the hesitation behavior segment are stored in the form of structured data, including the node timestamp, associated product ID, and operation type before and after the node, marking the key positions where interruption or hesitation occurs in the user's decision-making process.

[0058] The change in sentiment refers to the degree of difference in the textual sentiment of the same user towards the same related product before and after a single decision breakpoint, reflecting the emotional fluctuations of the user during the decision-making process.

[0059] The frequency of revisiting the focus area is the number of times a user independently browses the same related product in the same focus area before and after the decision breakpoint. Each browsing session ends when the user's gaze leaves the focus area for a preset duration.

[0060] The normalized focus area revisit frequency is a value obtained by standardizing the actual revisit frequency of all product focus areas by users within a fixed period. It is used to eliminate the quantitative bias caused by the difference in browsing behavior frequency among different users.

[0061] The deviation between intention and behavior is a quantitative indicator obtained by integrating emotional tendency change characteristics and visual focus revisit characteristics. It represents the degree of deviation between users' purchase intention and actual behavior.

[0062] Anomaly decision points are decision breakpoints where the degree of deviation from intended behavior is significant, marking that users have obvious multimodal behavioral semantic conflicts at this node.

[0063] Product attribute features are a set of structured attributes of products associated with anomalous decision points, including product text description attributes extracted from text review data and product visual-physical attributes extracted from image focal areas.

[0064] Conflict features are feature vectors composed of multi-dimensional features corresponding to abnormal decision points, used to quantitatively represent semantic conflicts between user text, image, and behavioral multimodal data.

[0065] Furthermore, firstly, retrieve the behavior log data that has been explicitly authorized by the user, extract the operation records marked as actual purchase actions, and based on the pre-calculated explicit and implicit trends of user purchases, traverse and filter out continuous operation segments that meet the criteria for hesitant behavior in the time series of the behavior logs. Mark the start and end nodes of these segments as decision breakpoints and store them in a structured manner.

[0066] Next, all text comments on related products by users before and after each decision breakpoint are extracted. The pre-trained text sentiment extraction branch is called to calculate the corresponding sentiment tendency value and analyze the sentiment changes before and after the breakpoint. At the same time, the number of independent browsings of the focus area of ​​related products by users is counted. The frequency is standardized to eliminate individual behavioral differences. The sentiment change features and visual revisit features are integrated to obtain the degree of deviation of intentional behavior corresponding to each decision breakpoint.

[0067] Next, abnormal decision points with significant deviations are selected, and the textual and visual attributes of the products associated with these abnormal decision points are extracted from the product attribute database. These attributes are then combined with the corresponding deviation degree, emotional changes, and standardized return visit frequency according to a preset data structure to generate conflict feature vectors for the target users and store them in the user feature database as input data for subsequent conflict type determination and cognitive state analysis.

[0068] This step, by constructing a complete multimodal conflict quantification calculation process, automates and structures the extraction of user intention and behavioral deviation features, solving the technical problem of existing technologies that discard semantic conflicts between multimodal data as data noise. Compared to traditional anomaly detection methods based on single-modal behavior, this step integrates dynamic changes in textual sentiment and visual attention revisit features, accurately capturing potential psychological contradictions in the user's decision-making process and effectively improving the reliability of conflict feature identification. Simultaneously, this step employs a standardized quantification calculation method, processing massive amounts of user multimodal behavioral data without manual annotation, significantly improving the processing capacity and real-time performance of the user profiling system, and laying a solid data foundation for subsequent accurate analysis of user cognitive states and identification of factors hindering purchase conversion.

[0069] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit authorization from user A, the behavioral log data of user A over 7 days is retrieved, and it is found that user A has browsed, favorited and added to cart three products: dress, sneakers and thermos cup multiple times, but no actual purchase action was taken.

[0070] Based on user A's explicit and implicit purchase trends, three hesitation behavior segments were selected from the behavioral log time series. These segments correspond to user A's actions of repeatedly entering and exiting the dress details page, repeatedly adding and canceling sneakers to the cart, and repeatedly browsing the thermos cup review section and closing the page. The start and end points of these three segments were marked as decision breakpoints.

[0071] The system extracts the sentiment values ​​of user A towards the corresponding products before and after each decision breakpoint, and calculates the change in sentiment values ​​before and after different product decision breakpoints. At the same time, it counts the actual return frequency of the focus area, and calculates the normalized return frequency of the focus area by combining the extreme values ​​of user A's return frequency in all focus areas within 7 days. The system calculates the deviation of the intention behavior of the three decision breakpoints through a preset fusion method, and the system filters out the decision breakpoints corresponding to sports shoes with deviations exceeding the preset abnormal threshold as abnormal decision points.

[0072] The preset anomaly threshold is not a fixed value, but rather based on the statistical distribution of historical conflict data from all users on the platform. The adaptive threshold for anomaly detection criteria and dynamic calibration of e-commerce business scenarios is set as follows: First, collect massive amounts of labeled historical user decision data from the platform, extract the deviation D value of intentional behavior corresponding to all normal decision points and manually labeled abnormal decision points, and construct a global deviation distribution dataset. Secondly, perform a normality test on the dataset, remove extreme outliers, and then calculate the mean of the dataset. with standard deviation ,in accordance with The criterion is to include those exceeding the deviation distribution. The critical value is used as the initial anomaly threshold to filter out only anomalies that significantly deviate from the normal range; Finally, dynamic calibration is performed in conjunction with e-commerce business scenarios. The threshold offset is adjusted according to product category (such as high-priced durable goods and low-priced fast-moving consumer goods) and user consumption level (such as high-frequency active users and low-frequency dormant users). At the same time, the threshold is recalculated and updated regularly as the platform's user behavior data iterates, ensuring that the threshold is adapted to the user decision-making characteristics of different periods and groups, and avoiding misjudgment or omission caused by static thresholds.

[0073] Extract the product attribute features of the sports shoes associated with the abnormal decision point, including text attribute word vectors and corresponding visual feature vectors. Combine these with the corresponding deviation of intentional behavior, change value of sentiment tendency, and frequency of revisiting the normalized focus area according to the preset data structure to generate the conflict feature vector of user A, and store it in the user feature database for subsequent conflict type determination and cognitive state analysis.

[0074] The formula for calculating the deviation of voluntary behavior is as follows: It is the deviation of intentional behavior, and its value range is: The larger the value, the more significant the deviation between the user's purchase intention and actual behavior, and it is the core quantitative indicator for identifying abnormal decision points.

[0075] It is the exponential time decay factor, with a value range of 100%. This allows for dynamic weighting of actions occurring at different times, with more recent actions contributing more to the current deviation calculation.

[0076] It is the half-life of user behavior memory, a global constant obtained by the platform based on the statistical analysis of all users' historical data, representing the average effective memory duration of users' consumption decisions.

[0077] It is the time interval between the decision breakpoint and the current calculation time, in hours, reflecting the timeliness of the behavior.

[0078] It is the emotional feature weighting coefficient. It is the visual focus feature weight coefficient, which satisfies The algorithm is adaptively determined through gradient descent training on all historical conflict data of all users, without the need for manual pre-setting.

[0079] It is the normalized asymmetric correction coefficient for sentiment polarity, obtained through the sign function and the sentiment polarity correction coefficient. The combination of these factors amplifies the contribution of negative emotional fluctuations to the deviation of voluntary behavior while weakening the impact of positive emotional fluctuations; simultaneously, by dividing by... To achieve range normalization, ensure that the corrected coefficients take values ​​within the specified range. The dimensions match those of the emotional change magnitude term.

[0080] For the asymmetric correction of emotional polarity: through the sign function The combination with the sentiment polarity correction coefficient γ applies differentiated weights to sentiment changes in different directions, with the weights weakening for positive sentiment changes and amplifying for negative sentiment changes, reflecting the stronger driving effect of negative sentiment on decision deviation.

[0081] This is the theoretical maximum value of the numerator term, used to standardize the corrected coefficients to... The interval eliminates the shift in the sentiment value range caused by asymmetric correction, ensuring that the weights of the sentiment and visual items are on the same order of magnitude, thus guaranteeing the rigor and comparability of the overall deviation calculation.

[0082] It is the change in sentiment before and after the decision breakpoint, reflecting the intensity of the user's emotional fluctuations.

[0083] It is a sign function, and the value changes when the sentiment tendency changes. hour, Take 1, hour, Take -1, when hour, Take 0.

[0084] It is the emotional polarity correction coefficient, with a value range of [value range missing]. The algorithm, learned from the training set, is used to distinguish the asymmetric effects of positive and negative sentiment changes on decision deviation. The amplification effect of negative sentiment changes on deviation is significantly higher than that of positive sentiment changes.

[0085] It is a globally normalized focus area revisit frequency coefficient, used to standardize the number of unique user views of a specific product's focus area to a global normalized value. The interval eliminates the lack of differentiation caused by individual differences in browsing habits among different users and the extreme values ​​of users themselves, and quantifies the degree of influence of the product attributes corresponding to the interval on user decisions; among which This represents the maximum number of unique views of all focused areas by all users on the platform within a fixed period, and is a global statistical constant.

[0086] It refers to the number of times a user independently browses the same related product and the same focal area within a preset time window before and after the decision breakpoint, reflecting the user's attention to the product attributes corresponding to that area.

[0087] It represents the maximum number of unique views a target user makes on all product focus areas within a fixed period, used to normalize the frequency of return visits.

[0088] It is the adaptive weighting coefficient for the focal region, and its value range is... The weight of a focal area is determined by the importance score of the product attribute corresponding to that focal area in historical purchase decisions. Different focal areas of different product categories have different weights. For example, the price area has a higher weight than the packaging area, and the core function display area has a higher weight than the auxiliary function display area.

[0089] For example, taking user A, who has obtained explicit authorization, as an example, a decision interruption occurred while browsing a certain pair of sneakers on an e-commerce platform. The decision interruption point was quantitatively calculated based on the formula for deviation of intentional behavior: First, extract the relevant parameters of the decision breakpoint, including the distance between the decision breakpoint and the current calculation time. The value is set to 2 hours. This value is directly derived from the difference between the timestamp generated by the decision breakpoint in the behavior log and the current calculation timestamp. The platform's statistical user behavior memory half-life global constant T is 24 hours. User A's sentiment tendency towards sneakers was 0.7 before the decision breakpoint, and dropped to 0.3 after the breakpoint due to browsing negative reviews. Therefore, the change in sentiment tendency value is... It is -0.4; because Negative, sign function The value is -1, which is the preset emotional polarity correction coefficient. The emotional feature weight coefficient is 0.5. The visual focus feature weight coefficient is 0.6. The value is 0.4; User A's independent revisit frequency F for the sole area of ​​the athletic shoes is 3 times, which is the maximum revisit frequency for all focus areas within 7 days. The adaptive weighting of the sole as the core functional area of ​​the athletic shoe is calculated to be 10 times. It is 0.8.

[0090] Then, the calculation is completed step by step: First, the time decay factor is calculated. =2 hours Substituting the formula into the small value, the calculation yields... The results show that recent decision-making breakpoints have a higher weight in influencing the degree of deviation from intended behavior.

[0091] The second step is to calculate the sentiment feature weighting term, and... =0.6、 , =0.5、 Substituting into the formula, we can calculate... The numerical value reflects the driving effect of users' negative emotional fluctuations on decision deviation.

[0092] The third step is to calculate the weighted term of the visual focus feature, and... Substituting into the formula, the result can be obtained; This result demonstrates the impact of users' repeated focus on the core area of ​​the shoe sole on the deviation between intention and behavior.

[0093] The fourth step is to calculate the deviation from the intended behavior. Substituting into the formula, we obtain the deviation of user A's intended behavior at this decision breakpoint. , Finally, a large amount of historical deviation data of user intentions in this scenario was collected from the platform, and obviously abnormal invalid data was removed; then the mean and standard deviation of the valid data were calculated, and the initial threshold was calculated according to the 3σ criterion; finally, the initial threshold was fine-tuned in combination with product category and user consumption level, and updated regularly with user data iteration, and finally the adaptive anomaly threshold was determined to be 0.25.

[0094] The calculated deviation value The value is greater than the platform's preset adaptive anomaly threshold of 0.25. Therefore, user A's decision point regarding the sneakers is determined to be an abnormal decision point, corresponding to the occasional single-attribute conflict feature. Based on this, the user's cognitive state can be further analyzed and factors hindering purchase conversion can be identified.

[0095] Step 3: Based on the triggering frequency and time distribution of the conflict characteristics within a fixed period, determine the conflict type of the target user, and analyze the cognitive state of the target user based on the conflict type; In this embodiment, determining the conflict type of the target user based on the trigger frequency and time distribution of the conflict characteristics within a fixed period includes: Extract the trigger points and trigger frequencies of the conflict features of the target user within a fixed period, and arrange the trigger points in ascending order of time to obtain the trigger time sequence of the conflict features; Calculate the time interval between the trigger points in the trigger time sequence, and compare the differences between the trigger points corresponding to each time interval to obtain the interval difference characteristics between the trigger points; The conflict type of the target user is determined based on the interval difference characteristics and the trigger frequency.

[0096] Specifically, the fixed period is a continuous time window preset by the e-commerce platform's user behavior analysis system for statistical analysis of user multimodal behavioral conflicts. It is quantified in hours and determined by the platform based on the consumption decision cycle statistics of different product categories. It is dynamically adjusted through system configuration files to limit the time range of conflict feature statistics and ensure the timeliness of analysis results.

[0097] The trigger point is the generation timestamp of the user conflict feature vector in the behavior log time series, accurate to the second, and corresponds one-to-one with the timestamp of the abnormal decision point. It is obtained by associating the user's unique identifier and product ID field with the conflict feature table and the behavior log table, marking the specific time when the user's multimodal semantic conflict occurred.

[0098] Trigger frequency is the total number of conflict feature vectors generated by the target user within a fixed period. It is obtained by counting and statistically analyzing the conflict feature records corresponding to the unique identifier of the same user within a fixed period, reflecting the intensity of multimodal semantic conflicts of the user.

[0099] The trigger time sequence is a one-dimensional time series array formed by arranging all trigger points of the same user within a fixed period in ascending order of time. The array elements are the second-level timestamps of the trigger points. It is generated by sorting the trigger time field of the conflict feature table using a database sorting function, and is used to analyze the time distribution pattern of conflict occurrences.

[0100] The time interval is the difference in timestamps between two adjacent trigger points in the trigger time sequence, in hours. It is obtained by performing a difference operation on the trigger time sequence and reflects the time distance between two adjacent multimodal semantic conflicts.

[0101] The interval difference feature is a feature vector composed of the statistical characteristics of all time intervals in the trigger time sequence. The data dimensions include the mean, variance, coefficient of variation, and ratio of the maximum interval to the minimum interval of the time intervals. It is obtained by performing descriptive statistical calculations on the time interval array and characterizes the dispersion and regularity of the conflict occurrence time distribution.

[0102] Conflict type is a multimodal behavioral conflict category automatically classified by clustering algorithm based on interval difference features and trigger frequency. It is stored in the form of integer codes, with different codes corresponding to different conflict time distribution patterns. It is the core input parameter for subsequent analysis of the user's cognitive state.

[0103] Furthermore, firstly, extract all conflict feature records of the target user who has obtained explicit user authorization within a fixed period, extract the trigger point timestamp corresponding to each record and count the trigger frequency, and arrange all trigger points in ascending order of time to generate a trigger time sequence.

[0104] Next, a difference operation is performed on the trigger time sequence to obtain the time interval between adjacent trigger points. Descriptive statistical calculations are then performed on all time intervals to obtain the interval difference characteristics, including mean, variance, coefficient of variation, and extreme value ratio.

[0105] Finally, the interval difference feature and the trigger frequency are combined into a conflict type determination feature vector, which is then input into a pre-trained K-means clustering model for classification calculation, and the conflict type code corresponding to the target user is output.

[0106] By transforming the temporal distribution characteristics of conflicts into quantifiable interval difference characteristics, the system achieves automated and standardized determination of conflict types, solving the technical problem that existing technologies cannot perform structured classification of multimodal semantic conflicts. It can accurately identify the temporal patterns of conflicts occurring among different users, avoiding the subjectivity and inefficiency of manual classification. This provides an objective and unified technical basis for subsequent analysis of user cognitive states based on conflict types, while also improving the system's efficiency in processing massive amounts of user behavior data.

[0107] The K-means clustering algorithm is pre-trained based on the platform's full set of historical user conflict data. During training, the platform first extracts the accumulated labeled user conflict feature records, calculates the interval difference features and trigger frequency of each record, and constructs the training dataset.

[0108] The elbow rule is used to determine the optimal number of clusters, corresponding to various typical conflict types, including high-frequency dense conflict, occasional single-attribute conflict, periodic fluctuating conflict, and decaying conflict. Euclidean distance is used as the similarity metric for model training, and reasonable iteration counts and convergence thresholds are set to ensure model convergence. After training, the model can effectively distinguish different conflict temporal distribution patterns.

[0109] The K-means algorithm was chosen because of its low computational complexity and fast convergence speed, making it suitable for handling the real-time conflict type determination needs of a large number of users on e-commerce platforms. At the same time, the clustering results have good interpretability and can form a stable technical connection with the subsequent cognitive state analysis steps.

[0110] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit authorization from user A, the conflict feature records of user A within a fixed period of 7 days are extracted to obtain a conflict feature record corresponding to the sports shoes. The trigger point is then arranged in ascending order of time to generate a trigger time sequence containing only a single element.

[0111] Since the trigger time sequence length is 1, the automatically calculated time interval array is empty, and the mean, variance, coefficient of variation, and extreme value ratio in the corresponding interval difference feature vector are all assigned to 0. The interval difference feature vector is combined with the trigger frequency of 1 to form a conflict type determination feature vector, which is then input into a K-means clustering model for calculation. The output conflict type code is 1, corresponding to an occasional single-attribute conflict type. In this embodiment, the step of analyzing the cognitive state of the target user based on the conflict type includes: Extract the cognitive component elements of the target user in the conflict type, and identify the cognitive blind spots of the target user based on the cognitive component elements; The cognitive blind spots are mapped to the actual decision points in the behavior log data to verify the degree of influence of the cognitive blind spots on the actual decision points, and a cognitive dissonance signal is generated based on the degree of influence. Based on the intensity of the cognitive dissonance signal and the correlation of the cognitive components, the decision-making tendency bias of the target user is derived, and the decision-making tendency bias is taken as the cognitive state of the target user.

[0112] Specifically, the conflict type is obtained by using integer encoded data output by the mean clustering model, which corresponds to high-frequency dense conflict, occasional single-attribute conflict, periodic fluctuation conflict, and decaying conflict, respectively. It is obtained by clustering the trigger frequency and interval difference features of conflict features and stored in the conflict type field of the user conflict feature database.

[0113] The cognitive component elements are structured cognitive dimension sets that are strongly related to the conflict type, extracted from the predefined cognitive dimension library corresponding to the conflict type. They are stored in the form of a one-dimensional string array. Each element corresponds to a quantifiable user cognitive dimension, including product attribute cognitive dimension, price cognitive dimension, brand cognitive dimension, and function cognitive dimension. They are automatically matched and obtained through the mapping relationship between conflict type encoding and cognitive dimension library.

[0114] Cognitive blind spots are cognitive dimension identifiers in which users have not generated effective information acquisition behavior in multimodal interaction data. They are stored in the form of a Boolean array with the same array length as the cognitive component element array. A true value indicates that there is a cognitive blind spot in that dimension, while a false value indicates that the cognitive dimension is complete. They are obtained by comparing the attribute coverage of the cognitive component elements with the information already acquired by the user.

[0115] The actual decision point is the time node in the behavior log time series where the user performs an operation directly related to the purchase decision. It is stored in the form of a second-level timestamp and associated with the corresponding operation type and product identifier. It is obtained by filtering the operation records marked as favorites, add to cart, place an order, and cancel an order in the behavior log.

[0116] The degree of influence is a numerical value that quantifies the degree of interference of cognitive blind spots on users' actual decision-making behavior. The larger the value, the higher the degree of interference. It is obtained by calculating the proportion of decision points with cognitive blind spots to all actual decision points, and combining the weighted sum of the deviation of the intentional behavior corresponding to the decision points.

[0117] Cognitive dissonance signals are structured signal data that characterize the degree of inconsistency between a user's cognition and behavior. They include three fields: signal generation timestamp, associated cognitive blind spot identifier, and impact level value. They are automatically generated when the impact level exceeds a preset threshold and stored in the user's cognitive state database.

[0118] The intensity of a cognitive dissonance signal is a normalized result of the influence value of the cognitive dissonance signal. It is obtained by dividing the influence value by the maximum influence value of all cognitive dissonance signals of the user, and is used to unify the quantitative standard of cognitive dissonance degree of different users.

[0119] The correlation degree of cognitive component elements is a numerical matrix that quantifies the degree of mutual influence between different cognitive component elements. The matrix dimension is consistent with the number of cognitive component elements. The larger the value, the higher the correlation degree between the two cognitive component elements. It is pre-generated by calculating the Pearson correlation coefficient of the platform's full user historical cognitive data and stored in the system parameter library.

[0120] Decision bias is a feature vector that quantitatively represents the direction and degree of deviation between a user's actual decision-making behavior and ideal rational decision-making behavior. Its dimensions are consistent with the cognitive components. The value of each dimension represents the degree of decision bias under that cognitive dimension. Positive values ​​indicate positive bias, and negative values ​​indicate negative bias. It is obtained by matrix multiplication of the cognitive dissonance signal intensity and the correlation matrix of cognitive components.

[0121] Cognitive state is a set of parameters that characterizes the user's current level of understanding of product information and decision-making tendency. It is stored in a structured key-value pair format and includes three core fields: cognitive blind spot array, cognitive dissonance signal list, and decision tendency deviation vector. These fields serve as the core input parameters for subsequent conversion feature analysis and obstacle factor identification and are automatically generated through the complete calculation process of this step.

[0122] Furthermore, based on the identified conflict type encoding, the corresponding cognitive component element set is extracted from the pre-built cognitive dimension library. Each cognitive component element is compared with the attribute coverage of the multimodal information already acquired by the user to identify cognitive blind spots where the user has not generated effective information acquisition behavior.

[0123] Next, the identified cognitive blind spots are mapped to all actual decision points in the user-authorized collected behavior log data. The number of decision points with cognitive blind spots is counted, and the degree of influence of the cognitive blind spots is calculated by combining the deviation of the intentional behavior of the corresponding decision points. When the degree of influence exceeds the system's preset threshold, a corresponding cognitive dissonance signal is generated.

[0124] Finally, the intensity of all cognitive dissonance signals is normalized, and a pre-generated cognitive component element correlation matrix is ​​called. The user's decision tendency bias under each cognitive dimension is derived through matrix multiplication. Cognitive blind spots, cognitive dissonance signals and decision tendency bias are integrated into a structured set of cognitive state parameters.

[0125] This step transforms the abstract user cognitive process into a computer-executable quantitative calculation process, solving the core problem that existing technologies cannot establish a technical correlation between multimodal conflict information and user decision-making cognition. By automating the identification and quantification of the degree of impact of cognitive blind spots, it replaces the traditional manual subjective analysis mode, significantly improving the accuracy and efficiency of locating user cognitive defects.

[0126] The decision bias derivation mechanism based on matrix operations enables traceable calculation of the transmission path of cognitive conflict to decision-making behavior, providing an objective and unified technical basis for subsequent accurate identification of factors hindering purchase conversion. At the same time, it effectively reduces the system latency of processing massive user data and improves the real-time response capability of the user profiling system.

[0127] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit authorization from user A, based on the determined conflict type code one of user A, the corresponding set of cognitive component elements is matched and extracted from the cognitive dimension library as the functional cognitive dimension.

[0128] By comparing the cognitive component elements with the multimodal information attribute coverage already acquired by User A, it was found that User A only acquired textual description information of the hardness of the shoe sole, and did not acquire related functional information such as sole material, wear resistance, and cushioning effect. Therefore, the cognitive blind spot array was identified as true.

[0129] Next, the cognitive blind spot is mapped to the actual decision points in User A's behavior log data. It is found that User A has six actual decision points related to sports shoes in seven days, all of which have the cognitive blind spot of this function. The influence degree of the cognitive blind spot is calculated by combining the deviation degree of the intention behavior corresponding to each decision point. If it exceeds the system's preset threshold, a corresponding cognitive dissonance signal is generated, which includes the signal generation timestamp, the cognitive dimension of the associated cognitive blind spot identification function, and the influence degree value.

[0130] Finally, the intensity of the cognitive dissonance signal is normalized, and a pre-generated single-dimensional cognitive component element correlation matrix is ​​called. The decision tendency bias vector of user A is derived through matrix multiplication, indicating that user A has a negative decision tendency bias in the functional cognitive dimension.

[0131] The cognitive blind spot array, the cognitive dissonance signal list, and the decision tendency bias vector are integrated into a structured cognitive state parameter set and stored in User A's cognitive state database for subsequent conversion feature analysis and purchase conversion obstacle identification.

[0132] Step 4: Collect the decision-making process of the target user, and combine the conflict type and the decision-making process to deduce the conversion characteristics of the target user, match the decision-making performance of the cognitive state and the conversion characteristics, and identify the obstacles that prevent the target user from completing the purchase conversion; In this embodiment, the step of deriving the conversion characteristics of the target user by combining the conflict type and the decision chain includes: Extract the consumer psychology opposition structure of the target user in the conflict type, and identify the conflict injection point in the decision-making link based on the consumer psychology opposition structure; Extract the user behavior intent direction of the nodes before and after the conflict injection point, and analyze the intent deviation tendency of the target user under the switching direction by comparing the switching direction of the user behavior intent direction of the nodes before and after the conflict injection point. By combining the decision compensation orientation of the intention deviation tendency and the decision inhibition attribute of the conflict type, the conversion characteristics of the target user before the actual purchase action are determined.

[0133] Specifically, the consumer psychology opposition structure is a binary data structure based on conflict type encoding mapping, consisting of two sets of behavioral feature vectors that represent the mutual constraints in the user's decision-making process. Each set of feature vectors corresponds to a decision-driving dimension and a decision-inhibiting dimension, with the dimension values ​​ranging from [value range missing]. It is stored in the user's psychological characteristics database and used to locate the sources of conflict in the decision-making process.

[0134] The decision chain is a directed acyclic graph data structure formed by arranging all operation behavior nodes in chronological order during a user's journey from product access to final purchase or exit within a single session. Each node contains fields such as operation type code, operation timestamp, associated product ID, and page dwell time. It is generated by grouping behavior log data by session ID and sorting by timestamp. The directed edges between nodes represent page jump relationships.

[0135] The node order is the sequential arrangement of each operational behavior node in the decision-making chain in the time dimension. It is marked as an integer sequence number on each decision-making chain node, with the sequence number increasing sequentially from the session start node. It is automatically generated by sorting the behavior log records within the same session in ascending order of timestamps.

[0136] Conflict injection points are nodes in the decision-making chain where user behavior intentions undergo unexpected changes. They are stored as unique identifiers of decision-making chain nodes and are identified by comparing the activation threshold of the decision inhibition dimension in the consumer psychology opposition structure with the similarity of the behavioral feature vector of the corresponding node. Nodes with similarity exceeding a preset threshold are conflict injection points.

[0137] User behavior intent is a categorized encoding data that represents the core behavioral goals of a user at a single decision-making node. It is obtained by classifying and predicting the operation type, page dwell time, and interaction trajectory characteristics of the node, with each node corresponding to a unique intent encoding.

[0138] The switching direction refers to the direction of change in the user behavior intent code between two adjacent nodes before and after the conflict injection point. It is divided into three categories: forward switching, reverse switching, and no switching. Forward switching means that the intent code changes from a low conversion stage to a high conversion stage, reverse switching means that the intent code changes from a high conversion stage to a low conversion stage, and no switching means that the intent code of the nodes before and after is the same. It is determined by calculating the sign of the difference in the conversion stage of the intent code of the nodes before and after.

[0139] Intent deflection tendency is a numerical value that quantifies the degree to which a user's behavioral intent deviates from the expected conversion path, with a value range of [value missing]. The larger the value, the higher the degree of deviation. It is obtained by calculating the cosine similarity between the switching direction of the intention of the nodes before and after the conflict injection point and the switching direction of the corresponding node of the ideal decision path. The lower the similarity, the higher the intention deviation tendency value.

[0140] Decision compensation orientation is a feature vector of specific behavioral patterns that users exhibit to compensate for decision uncertainty after their intentions have shifted. It includes three dimensions: frequency of information supplementation behavior, number of comparisons with similar products, and percentage of time spent in the evaluation area. It is calculated by statistically analyzing user behavior data within a preset time window after the conflict injection point and is used to characterize the behavioral features of users attempting to resolve cognitive conflicts.

[0141] The decision inhibition attribute is a set of quantitative parameters that are bound to the conflict type encoding and characterize the strength of the obstacle to user purchase conversion behavior of that type of conflict. It includes three dimensions: conversion delay coefficient, abandonment probability coefficient, and decision breakpoint generation probability. It is pre-generated through the correlation analysis of the platform's full user historical conflict data and conversion results and stored in the system parameter library. Different conflict types correspond to different decision inhibition attribute parameter values.

[0142] Conversion features are feature vectors that comprehensively characterize the conversion probability and conversion obstacle patterns of users in the purchase decision-making process. They include three core dimensions: intent deflection intensity, decision compensation degree, and conversion inhibition probability. They are calculated by multilayer perceptron by fusing intent deflection tendency, decision compensation orientation, and decision inhibition attributes, and are used to identify subsequent purchase conversion obstacle factors.

[0143] Furthermore, firstly, the consumer psychology opposition structure corresponding to the identified conflict type is extracted, all nodes in the target user's decision-making chain are traversed, the similarity between the behavioral feature vector of each node and the decision inhibition dimension in the consumer psychology opposition structure is calculated, and nodes with similarity exceeding a preset threshold are marked as conflict injection points.

[0144] Then, the user behavior intent direction code of the conflict injection point and its adjacent nodes before and after it is extracted, the switching direction of the intent direction of the nodes before and after is calculated, and the switching direction is compared with the expected switching direction of the node corresponding to the ideal decision path to calculate the user's intent deviation tendency.

[0145] Finally, after statistically analyzing the user behavior data within a preset time window after the conflict injection point, a decision compensation orientation feature vector is generated. The decision inhibition attribute parameter set corresponding to the conflict type is retrieved, and the intention deviation tendency, decision compensation orientation, and decision inhibition attribute are integrated to output the conversion characteristics of the target user before the actual purchase action.

[0146] By transforming abstract consumer psychological contradictions into quantifiable consumer psychological opposition structures and intention shift characteristics, this method achieves automated and precise location of conflict occurrences during the user decision-making process, solving the technical problem that existing technologies cannot track the transmission path of multimodal conflicts in the decision-making chain. By integrating decision-making compensation behavior characteristics with the inherent inhibitory properties of conflict types, it can objectively quantify the intensity of internal obstacles in the user conversion process, avoiding the uncertainty brought about by subjective psychological analysis and significantly improving the accuracy and consistency of conversion feature extraction.

[0147] Meanwhile, this step adopts a standardized machine learning computing process, which can process real-time decision-making data of massive users on e-commerce platforms in parallel, meeting the system's requirements for high concurrency and low latency, and providing reliable technical support for subsequent accurate identification of factors hindering purchase conversion.

[0148] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit authorization from user A, based on the determined occasional single-attribute conflict type code of user A, the corresponding consumer psychology opposition structure is extracted, where the decision-driven dimension is the product functional requirements and the decision-inhibition dimension is the risk of missing functional information.

[0149] Retrieve the decision-making process of user A in this session. This process includes six nodes: product search, clicking on the sneaker list page, browsing the sneaker details page, zooming in on the sole area, browsing the review section, and exiting the details page. The nodes are numbered 1 to 6 in chronological order.

[0150] The similarity between the behavioral feature vector and the decision inhibition dimension of each node is calculated. It is found that the similarity of the browsing of the evaluation area of ​​node 5 exceeds the preset threshold, and it is marked as a conflict injection point.

[0151] Extract the user behavior intent codes of the nodes before and after the conflict injection point. The intent of zooming in on the shoe sole area at node 4 is to query information. The intent of browsing the review area at node 5 is to query information. The intent of exiting the details page at node 6 is to exit. The comparison shows that the switching direction is the reverse of the switch from querying information to exiting.

[0152] The switching direction is compared with the expected switching direction information query of the corresponding node in the ideal decision path to calculate the intention bias tendency of user A. The behavioral data of user A within a preset time window after the conflict injection point are statistically analyzed to generate a decision compensation orientation feature vector, in which the frequency of information supplementation behavior, the number of similar products compared, and the proportion of time spent in the evaluation area are all obtained according to actual behavior statistics.

[0153] Retrieve the decision inhibition attribute parameter set corresponding to the occasional single-attribute conflict type, input the intention deviation tendency, decision compensation orientation and decision inhibition attribute into the multilayer perceptron model, and output the conversion feature vector of user A, which includes the specific values ​​of the three dimensions of intention deviation intensity, decision compensation degree and conversion inhibition probability.

[0154] In this embodiment, the step of matching the cognitive state with the decision-making performance of the conversion characteristics to identify the obstacles preventing the target user from completing the purchase conversion includes: Extract the cognitive demand profile corresponding to the conflict type from the cognitive state, and transform the cognitive demand profile into an ideal decision path based on the information carrying type in the decision-making link, and extract the actual decision path from the decision performance of the transformation feature; Filter the decision segments in which the actual decision path appears relative to the ideal decision path, retrieve the cognitive response types that the decision segments are bound to in the ideal decision path, extract the set of cognitively vulnerable areas pointed to by the conflict types, compare the type attribution of the cognitive response types with the set of cognitively vulnerable areas, and obtain the decision obstacle points; By summarizing all the information interface attributes and interaction elements of the decision-making obstacles, the exclusionary content that the target user cannot complete cognitive supplementation is identified, and this exclusionary content is used as an obstacle for the target user to complete the purchase conversion.

[0155] Specifically, decision performance refers to the set of actual decision-making behaviors and cognitive response results presented by target users who have signed privacy authorization agreements on e-commerce platforms, based on their own cognitive state and conversion characteristics throughout the entire product purchase decision-making process. These behaviors can be objectively collected and quantified by front-end tracking and back-end logging systems. This is the core technical basis for matching cognitive state and conversion characteristics and locating obstacles to purchase conversion in this step. It is mainly reflected in the actual decision-making path, decision-making segments, decision-making obstacles, information interface attributes, and interactive elements.

[0156] The cognitive demand profile is a feature vector of user information demand that is strongly correlated with the type of conflict and is extracted from the cognitive state. The dimensions are consistent with the total number of product information categories preset by the platform. The value of each dimension represents the intensity of the user's demand for that category of information. It is generated by calculating the mapping relationship between cognitive state and conflict type.

[0157] The decision-making chain is a directed graph structure formed by arranging all operation nodes in chronological order from the user's first contact with the product to completing the purchase or abandoning the purchase. Each node contains attributes such as operation type, operation time, associated page ID, and dwell time, and is generated by splicing behavioral log data in chronological order.

[0158] Information carrying type is a classification identifier for the information carried by each node page in the decision-making link. It includes four types: text information, image information, video information, and interactive information. It is stored in the page information database in the form of enumeration values ​​and corresponds one-to-one with the page ID.

[0159] The ideal decision path is an optimal sequence of decision nodes that can satisfy all of the user's cognitive needs, generated based on the user's cognitive needs profile and information carrying type. It is stored in an ordered array, with each element corresponding to a decision node ID.

[0160] The actual decision path is the sequence of decision nodes actually executed by the user during real interaction. It is stored in the form of an ordered array, with each element corresponding to an actual operation node ID. It is generated by extracting the time sequence of behavior log data.

[0161] A decision segment is a continuous subsequence in the actual decision path that differs from the ideal decision path at certain nodes, identified by the start node index and the end node index.

[0162] Cognitive response types are classifications of cognitive processing methods required by the user to complete the decision in each decision segment of the ideal decision-making path. They include four types: information acquisition, comparative analysis, risk assessment, and decision confirmation, and are stored in the cognitive model database as enumerated values.

[0163] The set of cognitively vulnerable areas is a set of cognitive dimensions that correspond to the types of conflict and that users are prone to cognitive dissonance. It is stored in the form of a one-dimensional array, with each element corresponding to a cognitive dimension identifier. It is obtained through a predefined mapping relationship between conflict types and cognitively vulnerable areas.

[0164] Type attribution is the matching relationship between cognitive response type and cognitive dimension in the set of cognitive vulnerability areas, represented by a Boolean value: true for a match and false for a mismatch.

[0165] Decision-making obstacles are nodes in the decision-making process that match the cognitive response type and the set type of cognitive vulnerability area. They are stored with node IDs and associated cognitive dimensions and are the core data nodes that cause user decision-making to be interrupted.

[0166] Information interface attributes are a set of visual and content attributes of the page corresponding to decision-making obstacles, including page layout, information density, content organization, font size, color matching, etc., which are stored in the page information database in the form of structured data.

[0167] Interactive elements are the collection of user-operable interface elements on the page corresponding to decision-making obstacles, including buttons, links, input boxes, drop-down menus, sliders, etc., which are stored in the page interaction database with element ID and operation type.

[0168] Exclusionary content is a combination of information interface attributes and interactive elements that users cannot fully understand, stored in the form of structured data, and is a specific technical manifestation that hinders users from completing purchase conversions.

[0169] Obstacles are structured data sets that prevent users from completing a purchase conversion. They include three fields: information interface attributes, interaction elements, and related cognitive blind spots. These are automatically extracted by the decision obstacle identification module and used to guide the generation of personalized conversion intervention strategies for the platform.

[0170] Furthermore, firstly, the cognitive need profile corresponding to the current conflict type is extracted, and the complete decision-making chain of the user's current session and the information carrying type of each node are obtained. The cognitive need profile is then transformed into an ideal decision-making path that covers all cognitive needs. At the same time, the actual decision-making path of the user's current interaction is parsed from the user behavior feature database.

[0171] Next, a global sequence comparison is performed between the ideal decision path and the actual decision path to locate continuous decision segments in the actual path that deviate from the ideal path. The cognitive response types of these decision segments bound to the ideal path are retrieved from the cognitive model database. At the same time, the set of cognitively vulnerable areas corresponding to the current conflict type is extracted from the conflict feature database. By comparing the cognitive response type with the type attribution of the cognitively vulnerable area one by one, all nodes that match successfully are marked as decision-making obstacles.

[0172] Finally, the page information database and page interaction database are queried in batches to summarize the information interface attributes and interaction elements corresponding to all decision-making obstacles. Content combinations that users have not interacted with effectively and cannot obtain corresponding cognitive information are filtered out, and these are structured and written into the user profile database as purchase conversion obstacles for target users.

[0173] This step transforms abstract psychological barriers to user decision-making into quantifiable and traceable data nodes in a computer system, solving the problem that existing technologies can only identify user conversion results but cannot pinpoint the root causes of conversion failures. This enables automated and precise identification of factors hindering purchase conversion.

[0174] By comparing and analyzing the ideal decision-making path with the actual decision-making path, the conflict at the user's cognitive level can be directly mapped to the page interface and interaction design of the e-commerce platform. This provides objective data support for the platform's personalized page optimization and real-time conversion intervention, significantly improving the ability to gain deeper insights into user profiles and the conversion efficiency of the e-commerce platform's precision marketing, while reducing subjective bias and processing delays caused by manual analysis.

[0175] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit privacy authorization from user A, the cognitive needs profile of user A is extracted based on the cognitive state parameter set of user A and the determined occasional single attribute conflict type, among which the functional cognitive dimension has the highest demand intensity.

[0176] Based on the information types carried by each node in User A's decision-making process for purchasing athletic shoes, the ideal decision path is generated as follows: browsing the main product image, viewing the product details page's functional parameters, viewing user feedback in the user review section, adding to cart, and making payment.

[0177] Simultaneously, user A's actual decision path was extracted as: browsing the product main image, viewing the product details page, viewing negative reviews in the user review section, and exiting the page. A dynamic time warping algorithm was used to compare the two paths, filtering out the two decision segments: viewing the product details page and viewing negative reviews in the user review section.

[0178] The cognitive response types bound to the corresponding segments in the ideal decision-making path were identified as information acquisition and risk assessment. The set of cognitively vulnerable areas pointed to by occasional single-attribute conflict types was extracted as the functional cognitive dimension. Comparison revealed that the cognitive response type of the product details page appearance display viewing segment matched the functional cognitively vulnerable area type, thus marking this node as a decision-making obstacle.

[0179] The information interface attributes corresponding to this decision-making obstacle are: the functional parameters are hidden in the second-level menu at the bottom of the page, the information density is low, and the interactive elements are: the button to view the functional parameters is not obvious and there is no quick jump entry.

[0180] The aforementioned exclusionary content that fails to supplement functional understanding will be considered as an obstacle preventing user A from completing the purchase of sneakers. This information will be stored in user A's profile database and used to subsequently push personalized product pages that prominently display functional parameters to user A.

[0181] Step 5: Integrate the cognitive conflict attributes of the hindering factors and the conflict type to generate dynamic psychological tags for the target user, and construct a user profile for the target user based on the dynamic psychological tags.

[0182] In this embodiment, the process of fusing the cognitive conflict attributes of the hindering factors and the conflict type to generate the dynamic psychological label of the target user includes: Match the avoidance behavior of the obstacle factors with the cognitive conflict attributes of the conflict type to obtain the conflict obstacle linkage record, and extract the psychological friction identifier from the conflict obstacle linkage record; The psychological friction markers in each of the aforementioned conflict and obstacle linkage records are summarized to obtain the conflict psychological theme of the target user, and the conflict psychological theme is combined with the cognitive state to obtain the dynamic psychological label of the target user.

[0183] Specifically, avoidance behavior orientation is a classification and encoding data representing the information dimension to which a user performs avoidance operations such as exiting, jumping, or closing when encountering obstacles. It is obtained by classifying and predicting the user operation sequence after the occurrence of obstacles in the behavior log, and each avoidance behavior corresponds to a unique dimension code.

[0184] Cognitive conflict attributes are a set of structured attributes that are bound to the conflict type code and characterize the core contradiction source of the conflict. They include conflict dimension field, contradiction intensity field, and transmission path field. They are pre-generated and stored in the system parameter library through correlation analysis of all users' historical conflict data on the platform.

[0185] The conflict and obstacle linkage record is a structured association record generated by associating and matching the obstacle factor record with the corresponding conflict type record within the same decision session of the same user. It includes a user unique identifier field, a session ID field, a conflict type code field, an obstacle factor ID field, and a time association parameter field, and is stored in the user conflict and obstacle association database.

[0186] Psychological friction identifiers are feature codes extracted from conflict and obstacle linkage records, representing the psychological resistance users experience in making decisions under the combined effect of conflict and obstacle. Each psychological friction identifier corresponds to a specific conflict-obstacle combination pattern.

[0187] The conflict psychology theme is a high-level psychological feature classification code obtained by clustering and summarizing all psychological friction indicators of the same user within a fixed period. It is used to summarize the user's core decision-making psychological contradictions within that period.

[0188] Dynamic psychological tags are a classification tag data structure with time decay factors, confidence scores, and behavioral triggering conditions. They include tag encoding fields, conflict psychological topic fields, cognitive state parameter fields, confidence fields, effective time window fields, and triggering condition fields. They are automatically generated by the computer system and stored in the user's dynamic tag library to represent the user's decision-making psychological state within a specific time window.

[0189] Furthermore, firstly, retrieve the records of the identified conflict types and the records of the identified obstacles, extract the timestamp corresponding to each record, and calculate the time correlation parameter between the two.

[0190] Record pairs within a preset strong correlation threshold range are selected, and the avoidance behavior of the obstacle factor is matched with the cognitive conflict attribute of the corresponding conflict type to generate a structured conflict obstacle linkage record and write it into the correlation database.

[0191] Subsequently, all valid conflict-obstacle linkage records are traversed. Using a pre-loaded conflict-obstacle combination and identifier mapping table, the corresponding psychological friction identifier is extracted from each record. Hierarchical clustering is then performed on all extracted psychological friction identifiers to merge semantically similar identifiers and summarize them to obtain the target user's conflict psychological theme.

[0192] Finally, the conflict psychology topic is combined with the corresponding cognitive state parameters stored in the user's cognitive state database to generate dynamic psychological tags containing complete fields and update the user's dynamic tag library.

[0193] This step achieves cross-dimensional correlation and fusion of user multimodal conflict information and conversion obstacle information, transforming abstract user decision-making psychological resistance into a standardized tag data structure that can be processed by computers, thus solving the technical problems of static user tags and lack of deep psychological insight generated by existing technologies.

[0194] By introducing cognitive conflict attributes, the timeliness and causal rationality of dynamic psychological tags are effectively guaranteed, improving the quantitative accuracy of user psychological state assessment. The generated dynamic psychological tags can directly provide data support for the personalized recommendation system and real-time conversion intervention system of e-commerce platforms, realizing automated tracing and precise intervention of decision-making obstacles, significantly improving the marketing conversion efficiency of e-commerce platforms. Simultaneously, this step adopts a fully automated calculation process, capable of processing massive amounts of multimodal user data in parallel, meeting the high-concurrency, low-latency system operation requirements of e-commerce platforms.

[0195] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit privacy authorization from user A, the conflict type record of user A's current sports shoe purchase session is retrieved. The conflict type code is 1, which corresponds to occasional single-attribute conflict. The corresponding cognitive conflict attribute is decision uncertainty caused by the lack of functional information.

[0196] At the same time, the obstacle factor record of user A is retrieved. The obstacle factor is that the function parameters of the product details page are hidden in the second-level menu at the bottom of the page and the position of the function parameter viewing button is not obvious. The difference between the generation timestamp of the obstacle factor and the trigger timestamp of the corresponding conflict feature is within the preset strong correlation time window, and it is determined that the two have a strong time correlation.

[0197] The avoidance behavior matching this obstacle is directed to the functional information query dimension, which aligns with the functional information deficiency dimension of the cognitive conflict attribute, generating a corresponding conflict-obstacle linkage record. The system extracts the psychological friction marker from this linkage record as a functional information acquisition obstruction marker. Since user A only has one conflict-obstacle linkage record in this period, the system directly summarizes the conflict psychological theme as the functional cognitive insufficiency theme.

[0198] Finally, the conflict psychological theme was fused with user A's cognitive state parameters to generate a dynamic psychological label, coded as P001. The conflict psychological theme was functional cognitive insufficiency, the cognitive state was negative decision-making bias in the functional cognitive dimension, the confidence level was 0.82, the effective time window was 7 days, and the trigger condition was that functional parameter information was not seen on the first screen of the page when browsing sports footwear products.

[0199] Update the dynamic psychological tag to user A's dynamic tag library for use in subsequent push notifications of personalized product pages that prioritize user A's functional parameters.

[0200] In this embodiment, constructing the user profile of the target user based on the dynamic psychological tags includes: By merging the conflict type identifier in the dynamic psychological tags with the cognitive state, a psychological inhibition distribution record of the target user is generated. By integrating the psychological inhibition distribution records with the multimodal features, a user profile of the target user is obtained.

[0201] Specifically, the conflict type identifier is an integer encoded data that corresponds one-to-one with the conflict type code. It is stored in the label encoding field of the dynamic psychological label and is used to uniquely identify the category of user multimodal behavioral conflict, corresponding to the classification results of high-frequency dense conflict, occasional single attribute conflict, periodic fluctuation conflict, and decaying conflict.

[0202] The psychological inhibition distribution record is a structured data record that integrates users' conflict types and cognitive states. It is stored in the user psychological characteristic database and includes conflict type identification field, cognitive blind spot distribution field, decision tendency bias distribution field, and timestamp field. It is used to quantitatively characterize the intensity and distribution of users' psychological inhibition in different cognitive dimensions.

[0203] User profiles are structured user data sets that integrate multimodal behavioral characteristics and dynamic psychological characteristics. They are stored in a user profile database and include four core modules: basic attribute tags, behavioral preference tags, dynamic psychological tags, and conversion barrier tags. They are used to provide data support for personalized recommendation, precision marketing, and conversion intervention systems of e-commerce platforms.

[0204] Furthermore, the system first retrieves the target user's dynamic psychological tags from the user dynamic tag library, extracts the conflict type identifier and corresponding cognitive state parameters, and then structurally merges the two according to the preset field mapping rules to generate a psychological inhibition distribution record containing conflict type, cognitive blind spot distribution and decision tendency bias distribution, and writes it into the user psychological feature database.

[0205] Subsequently, the user's multimodal feature set is retrieved from the distributed data warehouse, and the psychological inhibition distribution record and multimodal features are subjected to cross-dimensional feature splicing and normalization processing to generate a complete user profile containing basic behavioral features and dynamic psychological features, which is then updated to the user profile database.

[0206] For example, taking the interaction data of an authorized user A in an e-commerce platform over 7 days as an example, after obtaining explicit privacy authorization from user A, the computer system retrieves user A's dynamic psychological tags from the user dynamic tag library, extracts the conflict type identifier corresponding to occasional single-attribute conflicts and the cognitive state parameters containing functional cognitive blind spots and negative functional decision-making tendency biases, merges them according to preset field mapping rules to generate user A's psychological inhibition distribution record and writes it into the user psychological characteristic database.

[0207] Subsequently, the multimodal feature set of user A is retrieved from the distributed data warehouse, including sentiment tendencies and product attribute words extracted from text review data, focal areas and visual features extracted from image browsing data, and operation records and decision links extracted from behavior log data. The psychological inhibition distribution records are then cross-dimensionally spliced ​​and normalized with the above multimodal features to generate a complete user profile containing user A's basic behavioral preference tags, dynamic psychological tags, and conversion obstacle tags, and updated to the user profile database.

[0208] like Figure 2 The diagram shown is a functional block diagram of a deep learning-based multimodal user profile construction system provided in an embodiment of the present invention.

[0209] The deep learning-based multimodal user profile construction system 100 of this invention can be installed in an electronic device. Depending on the functions implemented, the deep learning-based multimodal user profile construction system 100 may include a data acquisition module 101, a conflict analysis module 102, a cognitive analysis module 103, an obstacle identification module 104, and a profile generation module 105. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0210] In this embodiment, the functions of each module / unit are as follows: Data acquisition module 101: Collects multimodal features of the target user, including text comment data, image browsing data and behavior log data, extracts the target user's sentiment tendency from the text comment data based on a deep learning model, and extracts the target user's focus area from the image browsing data; Conflict Analysis Module 102: Based on the sentiment tendency and the focus area, it merges and generates the target user's explicit and implicit purchase trends, and mines the target user's conflict characteristics based on the actual purchase actions in the behavior log data and the explicit and implicit purchase trends; Cognitive analysis module 103: Based on the frequency and time distribution of the conflict features within a fixed period, determine the conflict type of the target user, and analyze the cognitive state of the target user based on the conflict type; Obstacle identification module 104: Collects the decision-making process of the target user, and combines the conflict type and the decision-making process to deduce the conversion characteristics of the target user, matches the decision-making performance of the cognitive state and the conversion characteristics, and identifies the obstacles to the target user completing the purchase conversion; Profile generation module 105: Integrates the cognitive conflict attributes of the hindering factors and the conflict type to generate dynamic psychological tags for the target user, and constructs a user profile for the target user based on the dynamic psychological tags.

[0211] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0212] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0213] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0214] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0215] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0216] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for constructing multimodal user profiles based on deep learning, characterized in that, The method includes: Step 1: Collect multimodal features of the target user, including text comment data, image browsing data, and behavior log data. Extract the target user's sentiment tendency from the text comment data and the target user's focus area from the image browsing data based on a deep learning model. Step 2: Based on the sentiment tendency and the focus area, merge and generate the target user's explicit and implicit purchase trends; based on the actual purchase actions in the behavior log data and the explicit and implicit purchase trends, mine the target user's conflict characteristics. Step 3: Based on the triggering frequency and time distribution of the conflict characteristics within a fixed period, determine the conflict type of the target user, and analyze the cognitive state of the target user based on the conflict type; Step 4: Collect the decision-making process of the target user, and combine the conflict type and the decision-making process to deduce the conversion characteristics of the target user, match the decision-making performance of the cognitive state and the conversion characteristics, and identify the obstacles that prevent the target user from completing the purchase conversion; Step 5: Integrate the cognitive conflict attributes of the hindering factors and the conflict type to generate dynamic psychological tags for the target user, and construct a user profile for the target user based on the dynamic psychological tags.

2. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The step of fusing and generating the explicit and implicit purchase trends of the target user based on the sentiment tendency and the focus area includes: Extract product attribute words corresponding to the sentiment from the text review data to form the explicit preferences of the target user, and extract the visual features of the focal area to form the implicit preferences of the target user; By integrating the explicit and implicit preferences and assessing the cognitive state of the target user, the explicit and implicit purchasing trends of the target user are obtained.

3. The method for constructing multimodal user profiles based on deep learning as described in claim 2, characterized in that, The step of mining conflict characteristics of the target user based on actual purchase actions and explicit / implicit purchase trends in the behavior log data includes: Based on the actual purchase actions in the behavior log data, locate the hesitation behavior segments in the behavior log data that correspond to the explicit and implicit trends, and obtain the decision breakpoint of the target user; By combining the change in sentiment before and after the decision breakpoint with the frequency of revisiting the focal area, the deviation of the target user's intentional behavior is obtained; Filter out abnormal decision points among the decision breakpoints where the deviation of the intended behavior is abnormal, extract the product attribute features corresponding to the abnormal decision points, and obtain the conflict features of the target user.

4. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The step of determining the conflict type of the target user based on the trigger frequency and time distribution of the conflict characteristics within a fixed period includes: Extract the trigger points and trigger frequencies of the conflict features of the target user within a fixed period, and arrange the trigger points in ascending order of time to obtain the trigger time sequence of the conflict features; Calculate the time interval between the trigger points in the trigger time sequence, and compare the differences between the trigger points corresponding to each time interval to obtain the interval difference characteristics between the trigger points; The conflict type of the target user is determined based on the interval difference characteristics and the trigger frequency.

5. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The process of analyzing the cognitive state of the target user based on the conflict type includes: Extract the cognitive component elements of the target user in the conflict type, and identify the cognitive blind spots of the target user based on the cognitive component elements; The cognitive blind spots are mapped to the actual decision points in the behavior log data to verify the degree of influence of the cognitive blind spots on the actual decision points, and a cognitive dissonance signal is generated based on the degree of influence. Based on the intensity of the cognitive dissonance signal and the correlation of the cognitive components, the decision-making tendency bias of the target user is derived, and the decision-making tendency bias is taken as the cognitive state of the target user.

6. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The conversion characteristics of the target user are derived by combining the conflict type and the decision chain, including: Extract the consumer psychology opposition structure of the target user in the conflict type, and identify the conflict injection point in the decision-making link based on the consumer psychology opposition structure; Extract the user behavior intent direction of the nodes before and after the conflict injection point, and analyze the intent deviation tendency of the target user under the switching direction by comparing the switching direction of the user behavior intent direction of the nodes before and after the conflict injection point. By combining the decision compensation orientation of the intention deviation tendency and the decision inhibition attribute of the conflict type, the conversion characteristics of the target user before the actual purchase action are determined.

7. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The decision-making performance that matches the cognitive state with the conversion characteristics, and identifies the obstacles preventing the target user from completing the purchase conversion, includes: Extract the cognitive demand profile corresponding to the conflict type from the cognitive state, and transform the cognitive demand profile into an ideal decision path based on the information carrying type in the decision-making link, and extract the actual decision path from the decision performance of the transformation feature; Filter the decision segments in which the actual decision path appears relative to the ideal decision path, retrieve the cognitive response types that the decision segments are bound to in the ideal decision path, extract the set of cognitively vulnerable areas pointed to by the conflict types, compare the type attribution of the cognitive response types with the set of cognitively vulnerable areas, and obtain the decision obstacle points; By summarizing all the information interface attributes and interaction elements of the decision-making obstacles, the exclusionary content that the target user cannot complete cognitive supplementation is identified, and this exclusionary content is used as an obstacle for the target user to complete the purchase conversion.

8. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The process of integrating the cognitive conflict attributes of the hindering factors and the conflict types to generate dynamic psychological labels for the target user includes: Match the avoidance behavior of the obstacle factors with the cognitive conflict attributes of the conflict type to obtain the conflict obstacle linkage record, and extract the psychological friction identifier from the conflict obstacle linkage record; The psychological friction markers in each of the aforementioned conflict and obstacle linkage records are summarized to obtain the conflict psychological theme of the target user, and the conflict psychological theme is combined with the cognitive state to obtain the dynamic psychological label of the target user.

9. The method for constructing multimodal user profiles based on deep learning as described in claim 1, characterized in that, The process of constructing a user profile for the target user based on the dynamic psychological tags includes: By merging the conflict type identifier in the dynamic psychological tags with the cognitive state, a psychological inhibition distribution record of the target user is generated. By integrating the psychological inhibition distribution records with the multimodal features, a user profile of the target user is obtained.

10. A multimodal user profile construction system based on deep learning, characterized in that, The system includes: Data acquisition module: Collects multimodal features of target users, including text comment data, image browsing data and behavior log data, and extracts the sentiment tendency of the target users in the text comment data and the focus area of ​​the target users in the image browsing data based on a deep learning model; Conflict Analysis Module: Based on the sentiment tendency and the focus area, it merges and generates the explicit and implicit purchase trends of the target user, and mines the conflict characteristics of the target user based on the actual purchase actions in the behavior log data and the explicit and implicit purchase trends; Cognitive analysis module: Based on the frequency and time distribution of the conflict features within a fixed period, determine the conflict type of the target user, and analyze the cognitive state of the target user based on the conflict type; Obstacle identification module: Collects the decision-making process of the target user, and combines the conflict type and the decision-making process to deduce the conversion characteristics of the target user, matches the decision-making performance of the cognitive state with the conversion characteristics, and identifies the obstacles that prevent the target user from completing the purchase conversion; Profile generation module: Integrates the cognitive conflict attributes of the hindering factors and the conflict type to generate dynamic psychological tags for the target user, and constructs a user profile for the target user based on the dynamic psychological tags.

Images