Personalized display recommendation method based on multi-dimensional behavior characteristics

Abstract

The invention discloses a personalized display recommendation method based on multi-dimensional behavior characteristics. The method comprises the steps that S1, multi-dimensional behavior data of a user is collected and analyzed into a cognitive discharge event; s2, performing time sequence fusion and dynamic attenuation processing on the event stream, and constructing a user cognitive vector; s3, using a ColBERT model to encode display content, and constructing a content semantic field tensor; s4, calculating a semantic association degree based on the tensor set, and constructing a content universe network; s5, mapping the user cognitive vector and the content universe into an improved HMC algorithm, introducing a cognitive inertia factor, simulating a particle dynamics trajectory, and calculating resonance intensity; s6, inputting the resonance intensity list into a prediction model to generate a predictive recommendation list; and S7, selecting and rendering a final display form according to the list. According to the method, deep modeling of the cognitive state of the user and accurate quantification of the content value are achieved, and the recommendation accuracy and timeliness are remarkably improved through dynamic simulation and trend prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine learning and human-computer interaction, and particularly relates to a personalized display recommendation method based on multi-dimensional behavior characteristics. BACKGROUND

[0002] The existing personalized recommendation system still has significant limitations when facing complex and dynamic cognitive states of users and multi-modal semantic contents. First, most recommendation methods have a shallow understanding of user behavior, usually only recording explicit behaviors such as clicks and browsing, and failing to deeply analyze these behavior data into dynamic cognitive vectors reflecting the user's internal cognitive state, emotional tendency and cognitive agility, resulting in insufficient and inaccurate user portrait modeling.

[0003] Secondly, in terms of content understanding, the existing technology mostly uses single-modal feature extraction or simple multi-modal feature splicing, which is difficult to deeply mine the complex and multi-granularity semantic association and emotional polarity between text description and visual elements, and cannot construct a semantic field that can fully represent the intrinsic semantics of the content, limiting the accuracy of content matching. In addition, traditional recommendation algorithms are mostly based on static similarity calculation or collaborative filtering, lacking deep modeling of the dynamic interaction and matching process between user cognitive state and content semantic field, and being difficult to simulate the exploration trajectory and resonance state of users in the content universe.

[0004] Finally, most of the existing recommendation systems make instant recommendations based on current and historical data, lacking effective prediction ability for future user interest evolution and content matching trend, resulting in inaccurate grasp of the recommendation timing and insufficient foresight and initiative of the recommendation strategy. These technical bottlenecks jointly restrict the further development of personalized recommendation systems in terms of intelligence, precision and predictability.

[0005] Therefore, how to provide a personalized display recommendation method based on multi-dimensional behavior characteristics is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to propose a personalized display recommendation method based on multi-dimensional behavior characteristics, and the present application constructs an intelligent recommendation and presentation process with cognitive state vector extraction, content semantic field tensor generation, gravitational relationship networking, resonance intensity dynamics calculation, recommendation opportunity prediction and display form adaptation. The present application realizes the delayed interaction and multi-granularity semantic matching of text and visual elements by introducing the ColBERT model, and simulates the dynamics trajectory of user cognition in the content cosmic potential field by combining the improved HMC algorithm, introduces the cognitive inertia factor, realizes the accurate matching and forward-looking recommendation of the user's internal cognitive state and the deep semantic content. The present application has the advantages of deepening user cognitive modeling, multi-modal content semantic understanding, dynamics of matching process, prediction of recommendation opportunity, etc., which can significantly improve the accuracy, timeliness and user satisfaction of recommended content, thereby effectively solving the problems of shallow user portrait, single content understanding, static matching process and lagging recommendation opportunity in the existing method.

[0007] According to the personalized display recommendation method based on multi-dimensional behavior characteristics, the method comprises the following steps:

[0008] S1, real-time acquisition of multi-dimensional behavior data of users on a digital interface, and parsing the multi-dimensional behavior data into cognitive discharge events, outputting an original cognitive event stream;

[0009] S2, time sequence fusion and dynamic attenuation processing of the original cognitive event stream, construction and real-time updating and output of a user cognitive vector;

[0010] S3, encoding the display content to be recommended into Token level vector and Patch level vector respectively by using the ColBERT model, calculating the multi-granularity matching relationship, identifying and quantifying the semantic core based on the aggregation strength and determining the emotional polarity, constructing a content semantic field tensor, and outputting a content semantic field tensor set;

[0011] S4, based on the content semantic field tensor set, calculating the semantic correlation degree, determining the gravitational relationship between the display contents, constructing a content cosmic network, and outputting a topological content universe;

[0012] S5, mapping the user cognitive vector and the topological content universe into the improved HMC algorithm, introducing the cognitive inertia factor, simulating the dynamics trajectory of particles in each potential field, calculating the stable state, and quantifying the negative potential value of the stable state as the resonance intensity value, and outputting a resonance intensity list;

[0013] S6, inputting the resonance intensity list as a time sequence into the pre-trained ARIMA model, predicting the evolution trend of the resonance intensity value of each display content in the future, and generating a predictive recommendation list;

[0014] S7, according to the predictive recommendation list, when the recommendation opportunity arrives, selecting and rendering the final display form, outputting and presenting the adaptive display content to the user.

[0015] Optionally, the S1 specifically comprises:

[0016] S11, through the data acquisition script deployed in the digital interface front end, real-time capturing the click event coordinates, the continuous coordinate point sequence of the mouse trajectory and the loading and staying time stamp of the page element generated by the user interaction;

[0017] S12, integrating the captured click event coordinates, the continuous coordinate point sequence of the mouse trajectory and the page staying time stamp according to the time sequence, generating an original behavior data stream;

[0018] S13, presetting an energy value for each type of behavior event in the original behavior data stream, setting the energy value of the click event as the highest, mapping the moving speed of the mouse trajectory and the length of the page staying time into the corresponding energy value, and giving energy to each behavior event;

[0019] S14, encapsulating each behavior event in the original behavior data stream and the corresponding energy value and time stamp into an independent data unit, forming an original cognitive event stream and outputting.

[0020] Optionally, the S2 specifically comprises:

[0021] S21, inputting the original cognitive event stream into a preset time window, in the time window, accumulating the energy values of multiple cognitive discharge events occurring at the same time point, generating a time sequence energy sequence, multiplying each energy value in the time sequence energy sequence by a preset decay factor that exponentially decays over time, so that the cognitive discharge events farther away from the current time point have less impact on the current state;

[0022] S22, inputting the time sequence energy sequence after the decay processing into a full connection neural network layer with a preset dimension, mapping and integrating all energy values in the time sequence energy sequence into a vector with a set length, and normalizing the vector with the set length, so that the numerical range is stabilized in a preset interval, constructing and outputting a user cognitive vector.

[0023] Optionally, the S3 specifically comprises:

[0024] S31, for each to-be-recommended display content, input the text description into a pre-trained ColBERT model to generate a context-related Token-level vector for each Token in the text, and input the visual elements of the display content into a pre-trained VisionTransformer model to segment the image into multiple image blocks and generate a context-related Patch-level vector for each image block;

[0025] S32, all generated Token-level vectors and all Patch-level vectors are respectively processed by an independent linear transformation layer for dimension reduction and mapped into the same high-dimensional semantic space to obtain an aligned Token-level vector set and a Patch-level vector set;

[0026] S33, a delayed interaction algorithm is executed to perform vector dot product operation between each aligned Token-level vector and all Patch-level vectors in the set one by one to calculate the matching score of the Token-level vector and each Patch-level vector, and form a matching score vector corresponding to the Token-level vector;

[0027] S34, for each Token-level vector corresponding matching score vector, the maximum value is taken as the strongest matching score of the Token-level vector and the entire visual display content, and all Token strongest matching scores are accumulated to obtain the aggregation strength of the multi-granularity matching relationship;

[0028] S35, the Token-level vector whose aggregation strength exceeds the preset strength threshold is identified as a semantic core vector, the numerical value of the Token-level vector is directly extracted and quantified as the numerical value of the semantic core vector, and the cosine similarity between each semantic core vector and a preset sentiment word vector library containing positive, negative and neutral sentiment is calculated, and the polarity of the sentiment word vector with the highest similarity is selected as the sentiment polarity of the semantic core vector;

[0029] S36, all identified semantic core vectors, corresponding sentiment polarities and calculated aggregation strengths are combined and constructed into a multi-dimensional data structure, i.e. a content semantic field tensor, and the process is repeated for all to-be-recommended display contents to output a set of content semantic field tensors.

[0030] Optionally, the S4 specifically comprises:

[0031] S41, from the content semantic field tensor set, an arbitrary content semantic field tensor is selected as a reference tensor, the cosine similarity between the reference tensor and all other content semantic field tensors in the content semantic field tensor set is calculated, and all calculated cosine similarity values are taken as the semantic correlation between the content semantic field tensors;

[0032] S42, define the calculated semantic correlation degree value as the gravitational relationship between nodes directly, the higher the semantic correlation degree, the stronger the gravitational relationship;

[0033] S43, map each content semantic field tensor as a node in the network graph, and normalize the gravitational relationship between nodes as the weight of the edge connecting the two nodes, to construct the mutually related content universe network;

[0034] S44, use principal component analysis algorithm to reduce the dimension of all node vectors in the high-dimensional content universe network, extract the top two or three principal components with the largest contribution, project each node into a two-dimensional or three-dimensional topological space composed of these principal components, determine the specific position coordinates of each node, and output the topological content universe containing node positions and connection weights.

[0035] Optionally, the S5 specifically includes:

[0036] S51, initialize the user cognitive vector as a particle in the improved HMC algorithm, and define each content semantic field tensor in the topological content universe as an independent potential field, simulate the dynamic trajectory of the particle in each potential field through the improved HMC algorithm, and calculate the stable state of the particle in the potential field according to the lowest potential point to which the trajectory converges;

[0037] S52, multiply the negative potential value of each stable state by a preset positive scaling coefficient to obtain a non-negative resonance intensity value, the larger the resonance intensity value, the higher the matching degree, arrange all corresponding resonance intensity values in descending order to generate and output a resonance intensity list.

[0038] Optionally, the S51 specifically includes:

[0039] initialize the user cognitive vector as a particle in the improved HMC algorithm, record the user cognitive vector as the current position, and define each content semantic field tensor in the topological content universe as an independent potential field, and the potential function is specifically defined as the squared Euclidean distance between the particle and the content semantic field tensor;

[0040] extract the value from the user cognitive vector, map the value to a cognitive inertia factor between 0 and 1 through a preset S-shaped function, the higher the user cognitive vector, the lower the cognitive inertia factor, representing that the user's interest changes faster, and the weight of the last round momentum is smaller;

[0041] when the simulated particle enters the potential field defined by the first content semantic field tensor, sample a random momentum vector from a preset standard normal distribution as the initial momentum of the particle in this round of simulation;

[0042] When simulating the particle entering any one of the potential fields defined by the content semantic field tensor in the subsequent, reading the final momentum vector of the particle at the end of the last round of simulation, multiplying the final momentum vector with the step cognitive inertia factor element by element to obtain an inherited momentum vector, and adding a new random momentum vector sampled from the preset standard normal distribution element by element to generate the initial momentum of the particle in the new potential field;

[0043] Using the leapfrog integration algorithm, according to the current position, initial momentum and gradient of the potential field of the particle, with a preset step size and iteration number, the following three sub-steps are executed in a loop to simulate the dynamic trajectory of the particle in the current potential field;

[0044] The first sub-step is to update the momentum by half, which is to subtract half of the step size from the product of the current position potential gradient to obtain the momentum vector after half a step, the second sub-step is to update the position by one step, which is to add the step size to the product of the half-step momentum vector to obtain the updated position vector, and the third sub-step is to update the momentum by half, which is to subtract half of the step size from the product of the new position potential gradient from the half-step momentum vector, and the three sub-steps are executed in turn. The final momentum vector is obtained at the end of this iteration;

[0045] When the iteration number of the leapfrog integration algorithm reaches the preset value, traverse each position vector on the dynamic trajectory and subtract the content semantic field tensor to calculate the Euclidean distance square of the difference vector to obtain the corresponding potential value. Select the position vector with the minimum potential value as the stable state of the particle in the potential field defined by the content semantic field tensor.

[0046] Optionally, the S6 specifically comprises:

[0047] S61, for each display content, extracting the resonance intensity value from historical data, constructing an independent time series with time as index and resonance intensity value as numerical value, and inputting it into the pre-trained ARIMA model to learn the historical evolution law of the display content resonance intensity value. Predicting a preset time window in the future, outputting the predicted resonance intensity value of each time point of the display content in the future time window;

[0048] S62, traverse all the predicted resonance intensity values in the future time window, find out the maximum value as the predicted resonance intensity peak, and take the time point as the recommended opportunity. Match the predicted resonance intensity peak of all display contents and the corresponding recommended opportunity, and sort them according to the predicted resonance intensity peak from high to low to generate a predictive recommendation list.

[0049] Optionally, the S7 specifically comprises: according to the predictive recommendation list, when the recommendation opportunity arrives, the preset form selection rule is used to score the plurality of preset presentation forms of the recommended display content, the final display form with the highest score is selected and rendered, and the adaptive display content is output and presented to the user.

[0050] The beneficial effects of the present application are:

[0051] Firstly, by collecting multi-dimensional interaction behavior data of the user in real time and parsing into a time sequence cognitive event stream, a user cognitive vector capable of dynamically reflecting the user's cognitive state, emotional tendency and cognitive agility is constructed, thereby providing a high-precision data basis for deeply understanding the user's internal intention and immediate interest.

[0052] Secondly, the present application innovatively uses the ColBERT model to perform multi-granularity and delayed interactive semantic encoding on the text and visual elements of the recommended content, and by constructing a content semantic field tensor containing semantic cores, emotional polarity and aggregation strength, the deep semantic of the content is accurately and comprehensively represented. This method effectively solves the problem of single content understanding mode and shallow semantic mining in traditional recommendation, and significantly improves the depth and accuracy of content modeling.

[0053] Finally, by constructing a content universe network, the semantic association between contents is topological, and the user cognitive vector is mapped to a particle moving in a potential energy field by using an improved HMC algorithm, a cognitive inertia factor is introduced, and the resonance strength of the user and the content is calculated by simulating the dynamic trajectory. This physical dynamics-based matching model converts static similarity calculation into a dynamic resonance process, greatly enhancing the depth and novelty of the recommendation logic, making the recommendation result more consistent with the user's deep cognition and potential interest.

[0054] In summary, the present application constructs a complete, accurate and forward-looking personalized recommendation system by fusing cognitive behavior modeling, multi-modal semantic understanding, dynamic resonance matching and predictive recommendation, which can significantly improve the accuracy, novelty and user satisfaction of the recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0056] Figure 1 A flowchart of a personalized display recommendation method based on multi-dimensional behavior characteristics is provided for the present application;

[0057] Figure 2A multi-modal content semantic field tensor construction and multi-granularity matching relationship quantization flowchart based on the ColBERT model is provided for the present application.

[0058] Figure 3 A user cognitive vector dynamics trajectory simulation and resonance intensity calculation flowchart based on an improved HMC algorithm is provided for the present application. DETAILED DESCRIPTION

[0059] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0060] REFERENCE Figures 1-3 A personalized display recommendation method based on multi-dimensional behavior characteristics, comprising the following steps:

[0061] S1, real-time collection of multi-dimensional behavior data of a user on a digital interface, the multi-dimensional behavior data including click events, mouse trajectories and page dwell time, and parsing of the multi-dimensional behavior data into cognitive discharge events with time stamps and energy values, output of an original cognitive event stream;

[0062] S2, time series fusion and dynamic attenuation processing of the original cognitive event stream, construction and real-time updating and output of a user cognitive vector reflecting the current cognitive state, emotional tendency and cognitive agility of the user;

[0063] S3, encoding of the display content to be recommended into Token level vectors and Patch level vectors using the ColBERT model, calculation of the multi-granularity matching relationship between the Token level vectors and the Patch level vectors, identification and quantization of the semantic core of the display content based on the aggregation strength of the multi-granularity matching relationship and determination of the emotional polarity, construction of a content semantic field tensor, and output of a content semantic field tensor set;

[0064] S4, calculation of the semantic correlation degree based on the content semantic field tensor set, determination of the gravitational relationship between the display contents, construction of an interrelated content universe network, each content semantic field tensor serving as a node in the network, and output of a topological content universe including node positions and connection weights;

[0065] S5, mapping of the user cognitive vector into a particle in the improved HMC algorithm, and definition of each content semantic field tensor in the topological content universe as an independent potential field, simulation of the dynamics trajectory of the particle in each potential field by the improved HMC algorithm, introduction of a cognitive inertia factor, calculation of the stable state, quantization of the negative potential value of the stable state as a resonance intensity value reflecting the matching degree of the user cognition and the display content, and output of a resonance intensity list;

[0066] S6, input the resonance intensity list as a time sequence into the ARIMA model, predict the evolution trend of the resonance intensity value of each display content in a preset time window in the future, and generate a predictive recommendation list by combining the resonance intensity value at the current time;

[0067] S7, according to the predictive recommendation list, when the recommendation opportunity arrives, select and render the final display form, and output and present to the user the adaptive display content.

[0068] The embodiment can significantly improve the accuracy of personalized recommendation and user satisfaction. By analyzing the multi-dimensional behavior data of the user into a cognitive event stream, the transient cognitive state of the user is deeply described. Using the ColBERT model to perform multi-granularity matching on text and visual elements, a content semantic field tensor is constructed, which can more accurately quantify the deep value and emotional attributes of the display content. The user cognitive vector and the content universe are mapped into the improved HMC algorithm, and the cognitive inertia factor is introduced. The resonance intensity is calculated by simulating the dynamic trajectory of particles in the potential field, which skillfully simulates the psychological process of user decision-making. Combined with the ARIMA model to predict future trends, it realizes the leap from current matching to future prediction, making the recommendation more forward-looking and timely, effectively solving the shortcomings of traditional recommendation methods in understanding the depth of user intent and content relevance.

[0069] In the embodiment, S1 specifically includes:

[0070] S11, through the data acquisition script deployed in the front end of the digital interface, real-time capture the click event coordinates, continuous coordinate point sequence of mouse trajectory and loading and staying time stamp of page elements generated by user interaction;

[0071] S12, integrate the captured click event coordinates, continuous coordinate point sequence of mouse trajectory and page staying time stamp in chronological order to generate an original behavior data stream;

[0072] S13, preset an energy value for each type of behavior event in the original behavior data stream, set the energy value of the click event to the highest, map the moving speed of the mouse trajectory and the length of the page staying time to the corresponding energy value, and give each behavior event energy;

[0073] S14, encapsulate each behavior event in the original behavior data stream and the corresponding energy value and time stamp into an independent data unit to form an original cognitive event stream and output.

[0074] In the embodiment, S2 specifically includes:

[0075] S21, input the original cognitive event stream into a preset time window, accumulate the energy values of multiple cognitive discharge events occurring at the same time point in the time window, generate a time sequence energy sequence, multiply each energy value in the time sequence energy sequence by a preset decay factor that decays exponentially over time, so that the cognitive discharge events farther away from the current time point have less impact on the current state;

[0076] S22, input the decayed time sequence energy sequence into a preset dimension fully connected neural network layer, map and integrate all energy values in the time sequence energy sequence into a set length vector, and normalize the set length vector to stabilize the numerical range within a preset interval, construct and output the user cognitive vector.

[0077] In the embodiment, the S3 specifically comprises:

[0078] S31, for each recommended display content, input the text description into a pre-trained ColBERT model to generate a context-related Token-level vector for each Token in the text, input the visual elements of the display content into a pre-trained VisionTransformer model, segment the image into multiple image blocks, and generate a context-related Patch-level vector for each image block;

[0079] S32, map all generated Token-level vectors and all Patch-level vectors to the same high-dimensional semantic space through an independent linear transformation layer for dimensionality reduction processing, respectively, to obtain an aligned Token-level vector set and a Patch-level vector set;

[0080] S33, perform a delayed interaction algorithm, and perform vector dot product operation on each aligned Token-level vector and all Patch-level vectors in the set one by one, calculate the matching score of the Token-level vector and each Patch-level vector, and form a matching score vector corresponding to the Token-level vector;

[0081] S34, for each Token-level vector, take the maximum value of the matching score vector corresponding to the Token-level vector as the strongest matching score of the Token-level vector and the entire visual display content, and accumulate all Token strongest matching scores to obtain the aggregation strength of the multi-granularity matching relationship;

[0082] S35, the Token level vector with a polymerization strength exceeding a preset strength threshold is identified as a semantic core vector, the numerical value of the Token level vector is directly extracted and quantified as the numerical value of the semantic core vector, and a cosine similarity calculation is performed between each semantic core vector and a preset sentiment word vector library containing positive, negative and neutral sentiments, and the polarity of the sentiment word vector with the highest similarity is selected as the sentiment polarity of the semantic core vector;

[0083] S36, all the identified semantic core vectors, corresponding sentiment polarities and calculated polymerization strengths are combined and constructed into a multi-dimensional data structure, i.e., a content semantic field tensor, and the process is repeated for all the display contents to be recommended to output a set of content semantic field tensors.

[0084] In this embodiment, the ColBERT model and the Vision Transformer model are used to respectively perform deep coding on the text and the visual elements, so as to realize accurate alignment and fusion of cross-modal features. By performing a delayed interaction algorithm, vector dot product operations are performed on the Token level vectors and the Patch level vectors one by one, so as to finely quantify the multi-granularity matching relationship between each region in the text description and the visual content, and effectively capture the deep semantic association between the text and the visual content. The semantic core is identified based on the polymerization strength, and the sentiment polarity is determined in combination with the sentiment word vector library, and the constructed content semantic field tensor not only represents the core value of the content, but also incorporates the sentiment attribute, thereby providing a richer and more three-dimensional data basis for subsequent user-content matching. This method significantly improves the depth and accuracy of the system in understanding multi-modal content, and solves the problems of split text and visual information and shallow semantic understanding in traditional methods.

[0085] In this embodiment, the S4 specifically includes:

[0086] S41, from the set of content semantic field tensors, a content semantic field tensor is randomly selected as a reference tensor, the cosine similarity between the reference tensor and all other content semantic field tensors in the set of content semantic field tensors is calculated, and all the calculated cosine similarity values are used as the semantic association degrees between the content semantic field tensors;

[0087] S42, the numerical value of the calculated semantic association degree is directly defined as the gravitational relationship between nodes, and the higher the semantic association degree, the stronger the gravitational relationship;

[0088] S43, each content semantic field tensor is mapped to a node in a network graph, and the gravitational relationship between the nodes is normalized and used as the weight of the edge connecting the two nodes, thereby constructing a content universe network with mutual association;

[0089] S44, using a principal component analysis algorithm, dimensionality reduction is performed on all node vectors in the high-dimensional content universe network, the top two or three principal components with the largest contribution are extracted, each node is projected into a two-dimensional or three-dimensional topology space composed of the principal components, the specific position coordinates of each node are determined, and a topological content universe containing node positions and connection weights is output.

[0090] In the embodiment, the S5 specifically includes:

[0091] S51, initializing the user cognition vector as a particle in the improved HMC algorithm, and defining each content semantic field tensor in the topological content universe as an independent potential energy field, simulating the dynamic trajectory of the particle in each potential energy field through the improved HMC algorithm, and calculating the stable state of the particle in the potential energy field according to the lowest potential energy point to which the trajectory converges;

[0092] S52, multiplying the negative potential energy value of each stable state by a preset positive scaling coefficient to obtain a non-negative resonance intensity value, the greater the resonance intensity value, the higher the matching degree, and arranging all corresponding resonance intensity values in descending order to generate and output a resonance intensity list.

[0093] In the embodiment, the S51 specifically includes:

[0094] The user cognition vector is initialized as a particle in the improved HMC algorithm, and the user cognition vector is recorded as the current position, and each content semantic field tensor in the topological content universe is defined as an independent potential energy field, and the potential energy function is specifically defined as the squared Euclidean distance between the particle and the content semantic field tensor;

[0095] The numerical value is extracted from the user cognition vector, and the numerical value is mapped to a cognition inertia factor between 0 and 1 through a preset S-shaped function, the higher the user cognition vector, the lower the cognition inertia factor, representing that the user interest changes faster, and the weight of the last round momentum is smaller;

[0096] When the simulated particle enters the potential energy field defined by the first content semantic field tensor, a random momentum vector is sampled from a preset standard normal distribution as the initial momentum of the particle in this round of simulation;

[0097] When the simulated particle enters any subsequent potential energy field defined by a content semantic field tensor, the final momentum vector of the particle at the end of the last round of simulation is read, the final momentum vector is multiplied by the step cognition inertia factor element by element to obtain an inherited momentum vector, and the inherited momentum vector is added to a new random momentum vector sampled from the preset standard normal distribution element by element to generate the initial momentum of the particle in the new potential energy field.

[0098] The leapfrog integration algorithm is used to simulate the dynamic trajectory of the particle in the current potential energy field according to the current position, the initial momentum and the gradient of the potential energy field, with a preset step size and iteration number, and the following three sub-steps are cyclically executed:

[0099] The first sub-step is to update the momentum by half, that is, to subtract the product of one-half step size and the current position potential gradient from the current momentum vector to obtain the momentum vector after half-step; the second sub-step is to update the position by one step, that is, to add the product of one-step size and the momentum vector after half-step to the current position vector to obtain the updated position vector; and the third sub-step is to update the momentum by half, that is, to subtract the product of one-half step size and the new position potential gradient from the momentum vector after half-step, and the three sub-steps are sequentially executed, and the final momentum vector is obtained at the end of the current iteration;

[0100] When the iteration number of the leapfrog integration algorithm reaches the preset value, each position vector on the dynamic trajectory is traversed and subtracted from the content semantic field tensor, the Euclidean distance square of the difference vector is calculated to obtain the corresponding potential energy value, and the position vector with the minimum potential energy value is selected as the stable state of the particle in the potential energy field defined by the content semantic field tensor.

[0101] In the improved HMC algorithm, the cognitive inertia factor is introduced to realize dynamic adaptive simulation of user interest drift. The factor is obtained by mapping the user cognitive vector through an S-shaped function, which can quantify the stability of the current interest of the user: the higher the cognitive vector value, the lower the inertia factor, indicating that the user's interest changes faster, and the algorithm should focus more on exploring new possibilities when simulating. In the momentum fusion link, the final momentum of the last round is multiplied by the inertia factor and then added to the new random momentum, realizing the weighted inheritance of historical experience. When the user's interest is stable, the algorithm tends to maintain the original search direction; when the user's interest is variable, the algorithm is more likely to turn to a new exploration path. This improvement makes the particle dynamics simulation more realistically reflect the user's decision-making psychology, significantly improves the accuracy and timeliness of the resonance strength calculation in the dynamic interest scenario, and solves the problem of response lag when the traditional HMC algorithm simulates user interest drift.

[0102] In the embodiment, the S6 specifically includes:

[0103] S61, for each display content, extracting the resonance strength value from the historical data, constructing an independent time series indexed by time and the resonance strength value as a numerical value, and inputting it into a pre-trained ARIMA model to learn the historical evolution law of the display content resonance strength value, predicting a future preset time window, and outputting the predicted resonance strength value of the display content at each time point in the future time window;

[0104] S62, traverse all predicted resonance intensity values in the future time window, find the maximum value as the predicted resonance intensity peak, and the time point of occurrence as the recommended opportunity, match all the predicted resonance intensity peaks of the display content and the corresponding recommended opportunity, and sort them according to the predicted resonance intensity peak from high to low to generate a predictive recommendation list.

[0105] In the embodiment, S7 specifically comprises: according to the predictive recommendation list, when the recommended opportunity arrives, the preset form selection rule is used to weight score the multiple preset presentation forms of the recommended display content, the final display form with the highest score is selected and rendered, and the adaptive display content is output and presented to the user.

[0106] Embodiment 1:

[0107] In order to verify the actual application effect of the application in the field of personalized recommendation of large e-commerce platforms, the application is deployed in the guess you like core recommendation module of a domestic head comprehensive e-commerce platform. The platform has more than 800 million active users, and the number of online commodity SKUs exceeds 1 billion. The data level of user behaviors such as clicks, views, searches, and adds generated by users every day reaches the PB level. The traditional recommendation system mainly relies on collaborative filtering and content-based recommendation algorithms. When facing user interest drift, cold start, and information cocoon, the recommendation accuracy and user satisfaction are both bottlenecked, especially in deeply understanding the instantaneous cognitive intention of the user and the deep emotional association of the content.

[0108] In actual deployment, the method of the application first captures the multi-dimensional behavior data of the user in real time through a data acquisition script, and parses it into cognitive discharge events with timestamps and energy values. Then, the ColBERT model is used to deeply encode the text and visual elements of a large number of commodities, calculate the multi-granularity matching relationship through the delayed interaction algorithm, and finally construct a multi-modal content semantic field tensor for each commodity containing semantic cores and emotional polarity. Based on this, a content universe network is constructed, in which commodities are nodes and gravitational relationships are determined by semantic correlation. Subsequently, the core improved HMC algorithm of the application maps the user cognitive vector into a particle, defines each commodity tensor as a potential field, simulates the dynamics trajectory of the particle in a large number of potential fields, calculates the matching degree between the user and the commodity according to the trajectory convergence point, and quantizes it as a resonance intensity value. Finally, the system inputs the resonance intensity list into the ARIMA model to predict its evolution trend and generate a predictive recommendation list. The most likely to resonate commodities are pushed to the user at the best opportunity. The following is the key performance indicator comparison data of the method of the application and the original hybrid recommendation model based on deep learning of the platform in a month-long A / B test:

[0109] Table 1 Performance comparison data table of the application and the traditional hybrid recommendation model

[0110] Performance indicators Traditional hybrid recommendation model Method of the present application Performance improvement range Click-through rate (CTR) 3.85% 5.42% Increased by 40.8% Conversion rate (CVR) 1.12% 1.89% Increased by 68.8% GMV contribution per person 125.6 yuan 178.3 yuan Increased by 41.9% Recommendation diversity index 0.62 0.81 Increased by 30.6% User bounce rate 48.5% 39.2% Reduced by 19.2% Cold start commodity click rate 0.95% 2.18% Increased by 129.5% Long-tail commodity exposure ratio 15.3% 24.7% Increased by 61.4%

[0111] According to the comparative data in Table 1, the application has a significant performance advantage in the field of e-commerce recommendation. In terms of user engagement, the click-through rate has increased by more than 40%, the conversion rate has increased by nearly 70%, and the GMV contribution per person has increased by more than 40%. This is due to the accurate capture of the real-time cognitive state of users, the excavation of the deep value of goods by content semantic field tensors, and the accurate quantification of the resonance degree between users and goods through improved HMC algorithms, making recommendations more accurate and effective.

[0112] In terms of recommendation quality and ecological health, the recommendation diversity index has increased by more than 30%, and the user bounce rate has decreased by nearly 20%, effectively alleviating the information cocoon problem. Especially in the cold start scenario, the click rate of new goods has achieved an explosive growth of more than 129%, and the exposure proportion of long-tail goods has increased by more than 61%, indicating that the application can better tap potential markets and promote the balanced development of platform ecology. Overall, the application shows great commercial value and technical leadership.

[0113] Overall, in the actual application of large e-commerce platforms, the application successfully builds an intelligent recommendation system that can deeply understand users, accurately match content, and predict future trends by integrating multi-dimensional behavior data, constructing content semantic fields, and using improved HMC algorithms for cognitive dynamics simulation. The system shows great commercial value and technical leadership in improving core business indicators, optimizing user experience, and promoting ecological health, providing a strong empirical basis for the development direction of the next generation of personalized recommendation technology.

[0114] The above is only the preferred specific implementation of the application, but the protection scope of the application is not limited to this. Any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A personalized presentation recommendation method based on multi-dimensional behavioral characteristics, characterized in that, It comprises the following steps: S1, real-time acquisition of multi-dimensional behavior data of users on a digital interface, and parsing the multi-dimensional behavior data into cognitive discharge events, outputting an original cognitive event stream; S2, time sequence fusion and dynamic attenuation processing of the original cognitive event stream, constructing and real-time updating and outputting a user cognitive vector; S3, encoding the display content to be recommended into Token level vector and Patch level vector respectively using ColBERT model, calculating multi-granularity matching relationship, identifying and quantifying semantic core based on aggregation strength and determining sentiment polarity, constructing content semantic field tensor, and outputting content semantic field tensor set; S4, based on the content semantic field tensor set, calculating the semantic correlation degree, determining the gravitational relationship between the display contents, constructing the content universe network, and outputting the topological content universe; S5, mapping the user cognitive vector and the topological content universe into the improved HMC algorithm, introducing the cognitive inertia factor, simulating the dynamic trajectory of particles in each potential field, calculating the stable state, and quantifying the negative potential value of the stable state as the resonance strength value, and outputting the resonance strength list; S6, inputting the resonance strength list as a time sequence into the pre-trained ARIMA model, predicting the evolution trend of the resonance strength value of each display content in the future, and generating a predictive recommendation list; S7, according to the predictive recommendation list, when the recommendation opportunity arrives, selecting and rendering the final display form, and outputting and presenting to the user adaptive display content. 2.The method of claim 1, wherein, The S1 specifically comprises: S11, through the data acquisition script deployed in the front end of the digital interface, real-time capture of click event coordinates, continuous coordinate point sequence of mouse trajectory and loading and staying time stamp of page elements generated by user interaction; S12, integrating the captured click event coordinates, continuous coordinate point sequence of mouse trajectory and page staying time stamp in chronological order to generate an original behavior data stream; S13, presetting an energy value for each type of behavior event in the original behavior data stream, setting the energy value of the click event as the highest, mapping the moving speed of the mouse trajectory and the length of the page staying time into the corresponding energy value, and giving each behavior event an energy value; S14, encapsulating each behavior event in the original behavior data stream and the corresponding energy value and time stamp into an independent data unit to form an original cognitive event stream and output. 3.The method of claim 1, wherein, The S2 specifically comprises: S21, inputting the original cognitive event stream into a preset time window, accumulating the energy values of multiple cognitive discharge events occurring at the same time point in the time window to generate a time sequence energy sequence, multiplying each energy value in the time sequence energy sequence by a preset attenuation factor that exponentially attenuates over time, so that the influence of cognitive discharge events farther away from the current time point on the current state is smaller; S22, inputting the time sequence energy sequence after attenuation processing into a preset dimension fully connected neural network layer, mapping and integrating all energy values in the time sequence energy sequence into a set length vector, and performing normalization processing on the set length vector to stabilize the numerical range within a preset interval, constructing and outputting a user cognitive vector. 4.The method of claim 1, wherein, The S3 specifically comprises: S31, for each to-be-recommended display content, inputting a text description into a pre-trained ColBERT model to generate a context-related Token-level vector for each Token in the text, inputting a visual element of the display content into a pre-trained VisionTransformer model, segmenting an image into multiple image blocks, and generating a context-related Patch-level vector for each image block; S32, performing dimensionality reduction processing and mapping all generated Token-level vectors and all Patch-level vectors into the same high-dimensional semantic space through an independent linear transformation layer, to obtain an aligned Token-level vector set and a Patch-level vector set; S33, performing a delayed interaction algorithm, performing vector dot product operation between each aligned Token-level vector and all Patch-level vectors in the set one by one, calculating the matching degree score of the Token-level vector and each Patch-level vector, and forming a matching degree score vector corresponding to the Token-level vector; S34, taking the maximum value of the matching degree score vector corresponding to each Token-level vector as the strongest matching degree score of the Token-level vector and the entire visual display content, and accumulating all Token strongest matching degree scores to obtain the aggregation strength of the multi-granularity matching relationship; S35, identifying the Token-level vector with an aggregation strength exceeding a preset strength threshold as a semantic core vector, directly extracting the numerical value of the Token-level vector and quantifying it as the numerical value of the semantic core vector, and simultaneously performing cosine similarity calculation between each semantic core vector and a preset sentiment word vector library containing positive, negative and neutral sentiment, and selecting the polarity of the sentiment word vector with the highest similarity as the sentiment polarity of the semantic core vector; S36, combining and constructing all identified semantic core vectors, corresponding sentiment polarities and calculated aggregation strengths into a multi-dimensional data structure, i.e., a content semantic field tensor, repeating the process for all to-be-recommended display contents, and outputting a set of content semantic field tensors. 5.The method of claim 1, wherein, The S4 specifically comprises: S41, selecting a content semantic field tensor as a reference tensor from the set of content semantic field tensors, calculating the cosine similarity between the reference tensor and all other content semantic field tensors in the set of content semantic field tensors, and taking all calculated cosine similarity values as the semantic correlation degrees between the content semantic field tensors; S42, directly defining the numerical value of the calculated semantic correlation degree as the gravitational relationship between nodes, and the higher the semantic correlation degree, the stronger the gravitational relationship; S43, mapping each content semantic field tensor to a node in a network graph, and normalizing the gravitational relationship between nodes as the weight of the edge connecting two nodes, to construct a mutually related content universe network. S44, using principal component analysis algorithm, dimensionality reduction is carried out to all node vectors in high-dimensional content universe network, the first two or the first three principal components with the largest contribution degree are extracted, each node is projected into two-dimensional or three-dimensional topology space formed by the principal components, the specific position coordinates of each node are determined, and the topological content universe containing node position and connection weight is output. 6.The method of claim 1, wherein, The S5 specifically comprises: S51, the user cognitive vector is initialized as a particle in the improved HMC algorithm, each content semantic field tensor in the topological content universe is defined as an independent potential energy field, the dynamics trajectory of the particle in each potential energy field is simulated by the improved HMC algorithm, and the stable state of the particle in the potential energy field is calculated according to the lowest potential energy point converged by the trajectory; S52, the negative potential energy value of each stable state is multiplied by a preset positive scaling coefficient to obtain a non-negative resonance intensity value, the greater the resonance intensity value, the higher the matching degree, and all corresponding resonance intensity values are arranged in descending order to generate and output a resonance intensity list. 7.The method of claim 6, wherein, The S51 specifically comprises: The user cognitive vector is initialized as a particle in the improved HMC algorithm, and the user cognitive vector is recorded as the current position, and each content semantic field tensor in the topological content universe is defined as an independent potential energy field, and the potential energy function is specifically defined as the square of the Euclidean distance between the particle and the content semantic field tensor; The numerical value is extracted from the user cognitive vector, and the numerical value is mapped to a cognitive inertia factor between 0 and 1 by a preset S-shaped function, the higher the user cognitive vector, the lower the cognitive inertia factor, representing that the user interest changes faster, and the weight of the last round momentum is smaller; When the simulated particle enters the potential energy field defined by the first content semantic field tensor, a random momentum vector is sampled from a preset standard normal distribution as the initial momentum of the particle in this round simulation; When the simulated particle enters any subsequent potential energy field defined by the content semantic field tensor, the final momentum vector of the particle at the end of the last simulation is read, the final momentum vector is multiplied by the step cognitive inertia factor element by element to obtain an inherited momentum vector, and the inherited momentum vector is added to a new random momentum vector sampled from the preset standard normal distribution element by element to generate the initial momentum of the particle in the new potential energy field; By using the leapfrog integration algorithm, according to the current position, the initial momentum of the particle and the gradient of the potential energy field, the following three sub-steps are executed in a loop with a preset step size and iteration number to simulate the dynamics trajectory of the particle in the current potential energy field; The first sub-step is to update the momentum by half, the current momentum vector is subtracted by the product of half step size and current position potential gradient to obtain the momentum vector after half step, the second sub-step is to update the position by one step, the current position vector is added to the product of one step size and the momentum vector after half step to obtain the updated position vector, and the third sub-step is to update the momentum by half, the momentum vector after half step is subtracted by the product of half step size and new position potential gradient, the three sub-steps are executed in turn, and the final momentum vector is obtained at the end of this round iteration; When the number of iterations of the leapfrog integration algorithm reaches a preset value, each position vector on the dynamic trajectory is traversed and subtracted from the content semantic field tensor, the Euclidean distance square of the difference vector is calculated to obtain the corresponding potential energy value, and the position vector with the minimum potential energy value is selected as the stable state of the particle in the potential energy field defined by the content semantic field tensor. 8.The method of claim 1, wherein, The S6 specifically includes: S61, for each display content, extracting the resonance intensity value from the historical data, constructing an independent time series with time as the index and resonance intensity value as the numerical value, and inputting it into the pre-trained ARIMA model to learn the historical evolution law of the display content resonance intensity value, predicting the future within a preset time window, and outputting the predicted resonance intensity value of each time point of the display content within the future time window; S62, traversing all predicted resonance intensity values within the future time window, finding the maximum value as the predicted resonance intensity peak, and taking the time point as the recommended opportunity, matching the predicted resonance intensity peak of all display contents and the corresponding recommended opportunity, and sorting according to the predicted resonance intensity peak from high to low to generate a predictive recommendation list. 9.The method of claim 1, wherein, The S7 specifically includes: according to the predictive recommendation list, when the recommended opportunity arrives, a preset form selection rule is used to weight score the multiple preset presentation forms of the recommended display content, select and render the final display form with the highest score, and output and present it to the user adaptive display content.

Images