Big data personalized pushing method and system

Abstract

The invention discloses a big data personalized pushing method and system, and the method comprises the following steps: S1, collecting user behavior data and context information, and obtaining a user interest point set; s2, adopting a neural network model of a NeRF structure to construct a three-dimensional interest volume function and establish a neural interest field; s3, obtaining all to-be-pushed contents, and encoding the to-be-pushed contents into a content voxel set; s4, establishing a penetration mapping relation according to the content voxel set and the neural interest field; s5, executing content path search by adopting an improved sea elephant optimization algorithm, and generating a candidate content path set; s6, obtaining an interest density cumulative distribution curve and a direction change sequence of each candidate content path in the neural interest field; s7, analyzing and selecting an optimal content path; and S8, outputting the to-be-pushed content corresponding to the optimal content path as a final personalized pushing result. According to the method, NeRF modeling and an improved sea image optimization algorithm are fused, and personalized push of big data is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and in particular to a big data personalized push method and system. Background Art

[0002] With the rapid development of the mobile Internet and the wide popularization of intelligent terminals, personalized content push technology has become a key means to improve user experience and platform stickiness. By modeling the historical behaviors, interest preferences, and content features of users, the system can screen out the information that users may be interested in from a vast amount of information and achieve accurate push. Existing mainstream methods usually adopt methods such as collaborative filtering, matrix factorization, deep neural networks, graph neural networks, attention mechanisms, etc. to construct a user-content association model to predict the scoring or click probability of users for candidate content. However, most of these methods rely on low-dimensional interest space modeling and cannot fully reflect the spatial distribution structure and directional dynamic evolution characteristics of user interests.

[0003] In existing deep learning recommendation systems, some technologies attempt to introduce graph embedding or knowledge graphs for enhanced expression, but these methods still mainly unfold based on discrete feature spaces or topological structures, lacking the modeling of the interaction mechanism between user interests and content features in a continuous space. Especially in the push scenario for multi-modal content, the interaction relationship between the high-dimensional distribution of unstructured information such as text, images, and videos and user preferences is more complex, and traditional representation learning methods are difficult to capture the spatial mapping between multi-modal information and interest structures. In addition, existing content recommendation methods are mostly based on single-point prediction or ranking optimization, and have not effectively introduced spatial continuity constraints and path optimization mechanisms, and the organization and recommendation of content sequences still show a fragmented trend.

[0004] In recent years, the Neural Radiance Field (NeRF), as a neural implicit representation model for 3D reconstruction and volume rendering, has demonstrated extremely strong spatial modeling capabilities. Some studies have attempted to use NeRF for behavior simulation or scene understanding, but have not yet used it to construct a three-dimensional expression body of user interests and then achieve personalized content push based on a spatial perspective mechanism. At the same time, swarm intelligence optimization algorithms have achieved good results in path planning and combinatorial problems, but problems such as unclear construction of the optimization objective function and rough search granularity in the content recommendation scenario have not been fully solved. Especially when dealing with the complex non-linear mapping relationship between user preferences and content semantics, existing optimization algorithms lack the ability to deeply integrate with deep representation models, are prone to falling into local optima, and the stability of the recommendation results is insufficient.

[0005] Therefore, how to provide a big data personalized push method and system is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to provide a big data personalized push method and system. The present invention integrates NeRF modeling and an improved walrus optimization algorithm, constructs a three-dimensional volume function of user interests, encodes the content to be pushed as a set of spatial voxels, and performs penetration mapping and path search to achieve path-level personalized push, which has the advantages of high matching accuracy, strong spatial expression ability, good interpretability of interest paths, and strong semantic coherence of push sequences, effectively breaking through the limitations of traditional recommendation systems in terms of low user modeling dimension, coarse recommendation granularity, and weak path optimization ability.

[0007] A big data personalized push method according to an embodiment of the present invention includes the following steps: S1. Collect user behavior data and context information, perform semantic embedding, and obtain a set of user interest points; S2. Based on the set of user interest points, use a neural network model with a NeRF structure to construct a three-dimensional interest volume function and establish a neural interest field; S3. Obtain all the content to be pushed and encode each piece of content to be pushed as a set of content voxels; S4. Establish a penetration mapping relationship according to the set of content voxels and the neural interest field, and determine the spatial action path of the content voxels on the three-dimensional interest volume function; S5. Based on the penetration mapping relationship, use an improved walrus optimization algorithm to perform content path search. The content path is composed of multiple sets of content voxels in sequence, and a set of candidate content paths is generated; S6. Perform volume rendering of the set of candidate content paths in the neural interest field, and obtain the cumulative distribution curve of interest density and the direction change sequence of each candidate content path in the neural interest field; S7. Calculate the density penetration value of each candidate content path according to the cumulative distribution curve of interest density, calculate the refraction consistency value of each candidate content path according to the direction change sequence, and analyze and select the optimal content path; S8. Output the content to be pushed corresponding to the optimal content path as the final personalized push result.

[0008] Optionally, the S1 specifically includes: S11. Collect user behavior data, where the user behavior data includes click behavior, browsing behavior, stay duration, search requests, like behavior, and comment behavior, and record the time stamp and content identifier corresponding to each behavior; S12. Collect context information when the user behavior occurs. The context information includes terminal type, operating system type, geographical location information, access time information, and network connection status, and construct context information features; S13. Sort the user behavior data in ascending order of the timestamp to construct a user behavior sequence, where the user behavior sequence consists of structured triples containing behavior type, timestamp, and content identifier; S14. Perform semantic encoding on the text content corresponding to the content identifier, and use the BERT - Base pre - trained language model to encode the content title and body to output semantic features; S15. Concatenate the context information features and the semantic features to form joint user behavior context features; S16. Apply a density - based spatial clustering algorithm to the joint user behavior context features, use the Euclidean distance as the metric function, and extract the centers of all clustering clusters as the user interest point set.

[0009] Optionally, the specific steps of S2 are as follows: S21. Map each user interest point in the user interest point set to a three - dimensional Cartesian coordinate system to obtain three - dimensional spatial position coordinates; S22. Calculate the unit direction vector corresponding to each user interest point position according to the spatial change direction of consecutive user interest points; S23. Construct a neural network model with a NeRF structure, where the neural network model with a NeRF structure includes a position encoding module, a direction encoding module, a multi - layer perceptron network module, and an output layer module; The position encoding module performs position encoding on the input three - dimensional spatial position coordinates, and generates a position encoding feature vector with a preset length through the frequency transformation of sine and cosine functions; The direction encoding module performs the same sine - cosine frequency transformation process as the position encoding on the input unit direction vector and outputs a direction encoding feature vector with a preset length; The multi - layer perceptron network module includes an 8 - layer fully - connected network. The number of neurons in the first 7 layers is 256 each, and the number of neurons in the 8th layer is 128. Among them, the first 7 layers input the position encoding feature vector and output an intermediate position feature vector. After concatenating the intermediate position feature vector and the direction encoding feature vector and inputting them into the 8th layer, a comprehensive feature vector is generated; The output layer module is used to output the interest density value, and includes a first fully - connected layer, a first activation function layer, and a second fully - connected layer connected in sequence; the first fully - connected layer maps the input comprehensive feature vector to a 64 - dimensional intermediate comprehensive feature vector, the first activation function layer uses the ReLU activation function to perform non - linear processing on the intermediate comprehensive feature vector, and the second fully - connected layer maps the activated intermediate comprehensive feature vector to a single scalar output as the interest density value; S24. Calculate the three - dimensional interest volume function according to the neural radiance field volume rendering formula: ; ; Among them, represents the cumulative interest value of the th interest sampling path, represents the starting position of the interest sampling path, represents the ending position of the interest sampling path, represents the th interest density value of the interest sampling path at position , represents the cumulative transmittance from the starting position to position , represents the exponential function, represents the th interest density value of the interest sampling path at position ; S25. Converge the integral calculation result of the three-dimensional interest volume function in multiple interest path adoption directions to construct a neural interest field, representing the spatial interest distribution of the user.

[0010] Optionally, the S3 specifically includes: S31. Obtain all the content to be pushed, where the content to be pushed includes structured content and unstructured content, and the unstructured content includes text, image, video, and voice data; S32. Parse the content of each piece of content to be pushed, extract semantic principal components, tag information, and latent topic embeddings, and construct a multi-modal semantic feature vector; S33. Map the multi-modal semantic feature vector to the position coordinates in the three-dimensional content space, and the position coordinates form the spatial distribution center of the content to be pushed; S34. Based on the semantic intensity gradient and the push expected direction of the content to be pushed, construct multiple voxel units around the spatial distribution center, and each voxel unit includes spatial position coordinates, penetration direction vectors, and refraction tolerance values; The spatial position coordinates are used to indicate the specific position of the voxel unit in the three-dimensional content space; The penetration direction vector is used to represent the directionality of the voxel unit establishing a spatial interaction with the user's interest path in the neural interest field, and the penetration direction vector is obtained by weighted calculation based on the content semantic main axis and the user's historical preference direction; The refraction tolerance value is used to represent the allowable degree of semantic deviation after the content is penetrated by the interest sampling path in the direction of the voxel unit; S35. Based on all the voxel units corresponding to each piece of content to be pushed, obtain the content voxel set of each piece of content to be pushed.

[0011] Optionally, S4 specifically includes: S41. For each voxel unit in each content voxel set, extract the spatial position coordinates and the penetration direction vector, and construct a voxel ray with the spatial position coordinates as the starting point and the penetration direction as the ray direction; S42. Input the voxel ray into the established neural interest field, perform volume sampling at preset intervals along each voxel ray in the neural interest field, record the interest density values at the sampling points, and form a corresponding interest density sequence; S43. Calculate the cumulative response value of the voxel ray in the neural interest field according to the interest density sequence, where the cumulative response value is the weighted integral result of all sampling point interest density values; S44. Set an interest density response threshold , and compare the cumulative response value with the refraction tolerance value of the corresponding voxel unit . If the cumulative response value is greater than or equal to , it is determined that the voxel ray meets the penetration condition; S45. Use the voxel rays that meet the penetration condition as effective penetration paths, and classify them according to the belonging relationship of the content voxels to obtain the spatial action paths of each content voxel set in the neural interest field.

[0012] Optionally, S5 specifically includes: S51. Based on the spatial action paths of each content voxel set in the neural interest field, combine all the content voxel sets of the content to be pushed to construct an initial content path population, where each content path is composed of several content voxel sets connected in sequence; S52. Assign a set of coding vectors to each content path individual, where the coding vectors are used to represent the arrangement order and direction offset of each content voxel set in the content path; S53. Define a fitness function for the content path, where the fitness function is weighted and combined according to the cumulative response value, content semantic relevance, and structural continuity of the content path; S54. The improvement of the improved walrus optimization algorithm lies in introducing an adaptive leadership mechanism to update the coding vectors of the content path individuals, and the adaptive leadership mechanism dynamically adjusts the guiding direction according to the optimal content path individual in the current iteration: ; Among them, represents the coding vector of the th content path individual in the th round, represents the coding vector of the th content path individual in the th round, represents the convergence control factor, represents the perturbation factor, represents the Euclidean distance between the th content path individual and the optimal content path individual in the th round, represents the coding vector of the optimal content path individual in the th round, represents a Gaussian perturbation vector with a mean of 0 and a variance of 1, represents the diversity information entropy of the content path population in the ; wherein, represents the dimension of the coding vector, represents the proportion of the coding value with the highest frequency of occurrence in the th coding bit during the th round of iteration; S55. Execute the content path iterative update process for a preset number of rounds, and retain several content path individuals with the highest fitness values in each round; S56. After all iterations are completed, output the finally retained content path individuals as the candidate content path set.

[0013] Optionally, the S6 specifically includes: performing volume rendering of the candidate content path set in the neural interest field. For each candidate content path, the interest density values on all penetration rays are weighted and superimposed in the order of the content voxels to obtain the interest density cumulative distribution curve of the candidate content path in the neural interest field. The interest density cumulative distribution curve has the path position as the horizontal axis and the interest density integral value as the vertical axis; Calculating the sequence of direction angles between consecutive voxels within each candidate content path according to the spatial variation relationship of the penetration direction vectors in the content voxel set, and generating the direction change sequence of each candidate content path. The direction change sequence is used to characterize the spatial refraction trend of the candidate content path.

[0014] Optionally, the S7 specifically includes: S71. Performing integral calculation on the interest density cumulative distribution curve of each candidate content path. The integral interval covers the sequential arrangement section of all content voxels in the path, and calculating the density penetration value of the candidate content path in the neural interest field: ; wherein, represents the density penetration value of the candidate content path, represents the interest density integral value of the interest density cumulative distribution curve at the path position , represents the total length of the path; S72. Calculate the refraction consistency value of each candidate content path according to the direction change sequence: ; Wherein, represents the refraction consistency value of the candidate content path, represents the total number of content voxels in the candidate content path, represents the value of the th dimension of the direction change sequence; S73. Normalize the density penetration value and refraction consistency value of each candidate content path, and combine them according to a preset weight to form a comprehensive score value for each candidate content path; S74. Select the candidate content path with the largest comprehensive score value as the optimal content path.

[0015] A big data personalized push system according to an embodiment of the present invention includes the following modules: A user behavior modeling module, configured to collect user behavior data and context information, and extract a set of user interest points; A neural interest field construction module, configured to map the set of user interest points to a three-dimensional coordinate system, construct a neural network model with a NeRF structure and a three-dimensional interest volume function, and establish a neural interest field; A content voxel encoding module, configured to parse the content to be pushed and construct a set of content voxels; A penetration mapping establishment module, configured to generate penetration rays in the neural interest field according to the set of content voxels, perform volume sampling, calculate the cumulative response value, and filter out effective penetration paths according to the refraction tolerance value to generate a spatial action path; A path search and optimization module, configured to construct a population of content paths, iteratively update the content path individuals based on an improved walrus optimization algorithm introducing an adaptive leadership mechanism, and output a set of candidate content paths; A volume rendering analysis module, configured to input the set of candidate content paths into the neural interest field, perform volume rendering, and generate an interest density cumulative distribution curve and a direction change sequence for each candidate content path; A path evaluation and selection module, configured to calculate the density penetration value according to the interest density cumulative distribution curve, calculate the refraction consistency value according to the direction change sequence, form a comprehensive score value, and select the highest comprehensive score value as the optimal content path; A push output module, configured to output the content to be pushed corresponding to the optimal content path to form a spatially structured personalized content perspective chain push result.

[0016] The beneficial effects of the present invention are: (1) By constructing a neural interest field based on the NeRF structure, for the first time, user interests are modeled in the form of a continuous spatial volume function, effectively capturing the distribution characteristics of interest density in space and the directional change law, breaking through the limitation that traditional low-dimensional embedding models cannot reflect directionality and spatial continuity.

[0017] (2) By encoding the content to be pushed as a set of penetrable voxels and modeling in combination with the penetration direction and refraction tolerance, a spatial penetration mapping between the content voxels and the neural interest field is established, realizing a high-coupling modeling of content and interest in three-dimensional space and enhancing the physical interpretability of the matching.

[0018] (3) A content path search method based on an improved walrus optimization algorithm is proposed, introducing an adaptive leadership mechanism and a diversity control strategy to achieve a global optimal search for the multi-content voxel combination path in the interest field, avoiding the problems of fragmentation and local greedy strategies in traditional recommendations.

[0019] (4) By performing volume rendering on the candidate content path, quantifying its cumulative distribution of interest density and directional change trend, and calculating the density penetration value and refraction consistency value accordingly, a dual optimization of the final recommended path in terms of spatial continuity and interest fit is ensured. Description of the Drawings

[0020] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is the overall flowchart of a big data personalized push method proposed by the present invention; Figure 2 is the structural schematic diagram of a big data personalized push system proposed by the present invention. Detailed Embodiment

[0021] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0022] Refer to Figure 1 , a big data personalized push method, including the following steps: S1. Collect user behavior data and context information, perform semantic embedding, and obtain the set of user interest points; S2. Based on the set of user interest points, use a neural network model with the NeRF structure to construct a three-dimensional interest volume function and establish a neural interest field; S3. Obtain all the content to be pushed and encode each piece of content to be pushed as a set of content voxels; S4. Establish a penetration mapping relationship based on the content voxel set and the neural interest field, and determine the spatial action path of the content voxel on the three-dimensional interest volume function; S5. Based on the penetration mapping relationship, use the improved walrus optimization algorithm to perform content path search. The content path is sequentially composed of multiple content voxel sets, and a candidate content path set is generated; S6. Perform volume rendering on the candidate content path set in the neural interest field, and obtain the interest density cumulative distribution curve and the direction change sequence of each candidate content path in the neural interest field; S7. Calculate the density penetration value of each candidate content path according to the interest density cumulative distribution curve, calculate the refraction consistency value of each candidate content path according to the direction change sequence, and analyze and select the optimal content path; S8. Output the content to be pushed corresponding to the optimal content path as the final personalized push result.

[0023] This method constructs a three-dimensional interest volume function by introducing a neural network model with a NeRF structure, and introduces a spatial penetration mechanism of content voxels in the neural interest field to establish a highly coupled spatial expression relationship between user interests and content features. On this basis, combined with the improved walrus optimization algorithm for path-level push search, it is no longer limited to the similarity matching method of single content or short window in traditional recommendations. The method has a closed-loop structure from user interest modeling to path planning and then to sequence output, which can effectively improve the semantic fit, sequence coherence and user click conversion rate of the recommendation system. At the same time, technologies such as spatial penetration, volume rendering, and density cumulative analysis are introduced to greatly enhance the structural expressiveness and interpretability of content recommendations.

[0024] In this embodiment, the specific steps of S1 include: S11. Collect user behavior data, which includes click behavior, browsing behavior, stay duration, search requests, like behavior, and comment behavior, and record the timestamp and content identifier corresponding to each behavior; S12. Collect the context information when the user behavior occurs. The context information includes terminal type, operating system type, geographical location information, access time information, and network connection status, and construct context information features; S13. Sort the user behavior data in ascending order of timestamp, and construct a user behavior sequence. The user behavior sequence is composed of structured triples including behavior type, timestamp, and content identifier; S14. Perform semantic encoding on the text content corresponding to the content identifier, and use the BERT-Base pre-trained language model to encode the content title and text, and output semantic features; S15. Concatenate the context information features and the semantic features to form a user behavior context joint feature; S16. Apply a density-based spatial clustering algorithm to the user behavior context joint feature, with the Euclidean distance as the metric function, and extract the centers of all clustering clusters as the user interest point set.

[0025] By collecting the multi-dimensional behavior data and context environment information of users, deeply encoding the content semantics using the BERT language model, and then fusing the context information to construct a joint feature representation, the extraction of user interest points is made semantically sensitive and behavior context adaptable. At the same time, through density clustering, the user behavior characteristics are spatially partitioned, which can effectively refine a highly representative and non-repetitive interest point set, avoiding the problems of abstraction, staticity, and low update frequency existing in traditional user portraits, significantly improving the accuracy and real-time performance of user interest modeling, and providing more discriminative interest coordinates for spatial modeling and path planning.

[0026] In this embodiment, the specific steps of S2 include: S21. Map each user interest point in the user interest point set to a three-dimensional Cartesian coordinate system to obtain three-dimensional spatial position coordinates; S22. Calculate the unit direction vector corresponding to the position of each user interest point according to the spatial change direction of consecutive user interest points; S23. Construct a neural network model with a NeRF structure, and the neural network model with a NeRF structure includes a position encoding module, a direction encoding module, a multi-layer perceptron network module, and an output layer module; The position encoding module performs position encoding on the input three-dimensional spatial position coordinates, and generates a position encoding feature vector with a preset length through the frequency transformation of sine and cosine functions; The direction encoding module performs the same sine and cosine frequency transformation processing on the input unit direction vector as the position encoding, and outputs a direction encoding feature vector with a preset length; The multi-layer perceptron network module includes an 8-layer fully connected network. The number of neurons in the first 7 layers is 256, and the number of neurons in the 8th layer is 128. Among them, the first 7 layers input the position encoding feature vector and output an intermediate position feature vector. After concatenating the intermediate position feature vector and the direction encoding feature vector and inputting them into the 8th layer, a comprehensive feature vector is generated; The output layer module is used to output the interest density value, and includes a first fully-connected layer, a first activation function layer, and a second fully-connected layer connected in sequence; the first fully-connected layer maps the input comprehensive feature vector into a 64-dimensional intermediate comprehensive feature vector, the first activation function layer performs non-linear processing on the intermediate comprehensive feature vector using the ReLU activation function, and the second fully-connected layer maps the activated intermediate comprehensive feature vector into a single scalar output as the interest density value; S24. Calculate the three-dimensional interest volume function according to the neural radiance field volume rendering formula: ; ; where, represents the cumulative interest value of the th interest sampling path, represents the starting position of the interest sampling path, represents the ending position of the interest sampling path, represents the th interest sampling path at the position interest density value, represents the cumulative transmittance from the starting position to the position , represents the exponential function, represents the th interest sampling path at the position interest density value; S25. Converge the integral calculation result of the three-dimensional interest volume function in multiple interest path adoption directions to construct a neural interest field, representing the spatial interest distribution of the user.

[0027] Introduce a neural network model with the NeRF structure, model the user's interest points as a differentiable three-dimensional interest volume function, and for the first time fuse two types of variables, spatial position and direction vector, in interest expression to achieve continuous spatial modeling of user interests. By combining position encoding and direction encoding, the multi-layer perceptron network can accurately model the response changes of the interest density field in different directions. In addition, the output layer design supports the precise generation of interest density values, making the interest distribution both high-resolution and adjustable, thus significantly enhancing the fine-grained and personalized capabilities of interest expression. This mechanism provides a strong spatial foundation for voxel path penetration and path evaluation.

[0028] In this embodiment, the S3 specifically includes: S31. Obtain all the content to be pushed, where the content to be pushed includes structured content and unstructured content, and the unstructured content includes text, image, video, and voice data; S32. Parse the content of each content to be pushed, extract the semantic principal components, tag information, and potential topic embeddings, and construct a multi-modal semantic feature vector; S33. Map the multi-modal semantic feature vector to the position coordinates in the three-dimensional content space, and the position coordinates form the spatial distribution center of the content to be pushed; S34. Based on the semantic intensity gradient and the push expected direction of the content to be pushed, construct a plurality of voxel units around the spatial distribution center, and each voxel unit includes spatial position coordinates, penetration direction vector, and refraction tolerance value; The spatial position coordinates are used to indicate the specific position of the voxel unit in the three-dimensional content space; The penetration direction vector is used to represent the directionality of the spatial interaction between the voxel unit and the user interest path in the neural interest field, and the penetration direction vector is obtained by weighted calculation based on the content semantic main axis and the user historical preference direction; The refraction tolerance value is used to represent the allowable degree of semantic deviation after the content is penetrated by the interest sampling path in the direction of the voxel unit; S35. Based on all the voxel units corresponding to each content to be pushed, obtain the content voxel set of each content to be pushed.

[0029] Unify the multi-modal content to be pushed to the three-dimensional content space, and construct a voxel set with spatial position, penetration direction, and semantic tolerance as features, which effectively solves the problems of lack of spatial position information and single expression dimension in traditional content representation. Through voxelization modeling, spatial testing of content semantics in terms of directionality and tolerance can be realized in the neural interest field, laying a foundation for establishing penetration mapping. In addition, the penetration direction is jointly constructed based on the semantic main axis and user preferences, with stronger personality adaptation ability, enhancing the geometric rationality and semantic relevance of the matching between content and interest.

[0030] In this embodiment, the S4 specifically includes: S41. For the voxel units in each content voxel set, extract the spatial position coordinates and the penetration direction vector, and construct a voxel ray with the spatial position coordinates as the starting point and the penetration direction as the ray direction; S42. Input the voxel ray into the established neural interest field, perform volume sampling at a preset interval along each voxel ray in the neural interest field, record the interest density values at the sampling points, and form a corresponding interest density sequence; S43. Calculate the cumulative response value of the voxel ray in the neural interest field according to the interest density sequence, and the cumulative response value is the weighted integral result of the interest density values of all sampling points; S44. Set an interest density response threshold , and compare the cumulative response value with the refraction tolerance value of the corresponding voxel unit Compare. If the cumulative response value is greater than or equal to , it is determined that the voxel ray satisfies the penetration condition; S45: Use the voxel rays that satisfy the penetration condition as effective penetration paths, and classify them according to the ownership relationship of the content voxels to obtain the spatial action paths of each content voxel set in the neural interest field.

[0031] By mapping the content voxels into ray form and inputting them into the neural interest field, and performing point-by-point sampling and integration based on the interest density field, this step can accurately capture the spatial response path between the content and the user's interest. Introducing the comparison mechanism of the cumulative response value and the refraction tolerance enables the system to physically determine whether the content semantics can penetrate the user's interest, providing unprecedented push interpretability. In addition, constructing effective penetration paths helps to improve the initial quality of content path search, narrow the search space, and enhance the structural continuity of the push path.

[0032] In this embodiment, S5 specifically includes: S51: Based on the spatial action paths of each content voxel set in the neural interest field, combine the content voxel sets of all content to be pushed to construct an initial content path population, where each content path is composed of several content voxel sets connected in sequence; S52: Assign a set of coding vectors to each content path individual, and the coding vectors are used to represent the arrangement order and direction offset of each content voxel set in the content path; S53: Define the fitness function of the content path, and the fitness function is weighted and combined according to the cumulative response value, content semantic relevance, and structural continuity of the content path; S54: The improvement of the improved walrus optimization algorithm lies in introducing an adaptive leadership mechanism to update the coding vectors of the content path individuals, and the adaptive leadership mechanism dynamically adjusts the guiding direction according to the optimal content path individual in the current iteration: ; where, represents the coding vector of the th content path individual in the th round, represents the coding vector of the th content path individual in the th round, represents the convergence control factor, represents the perturbation factor, represents the Euclidean distance between the th content path individual and the optimal content path individual in the th round, represents the The coding vector of the optimal content path individual in the round, represents a Gaussian perturbation vector with a mean of 0 and a variance of 1, denotes the diversity information entropy of the content path population in the ; wherein, represents the dimension of the coding vector, denotes the proportion of the coding value with the highest frequency of occurrence in the th coding bit during the th round of iteration; S55. Execute the content path iterative update process for a preset number of rounds, and retain several content path individuals with the highest fitness values in each round;

[0033] The content path search mechanism is constructed based on an improved walrus optimization algorithm. By designing an adaptive leadership mechanism and a path diversity entropy guidance strategy, the population activity is maintained during the search process to prevent falling into local optima. At the same time, the path individual coding integrates the content arrangement order and direction offset information, and the optimization result is more structurally logical. The fitness function jointly considers the interest response value, semantic relevance, and structural continuity to ensure that the search result not only hits the interest but also has semantic coherence. The finally formed candidate path set provides more reasonable multiple interest directions for the recommendation system.

[0034] In this embodiment, the S6 specifically includes: performing volume rendering of the candidate content path set in the neural interest field. For each candidate content path, the interest density values on all penetration rays are weighted and superimposed according to the content voxel arrangement order to obtain the interest density cumulative distribution curve of the candidate content path in the neural interest field. The interest density cumulative distribution curve has the path position as the horizontal axis and the interest density integral value as the vertical axis; According to the spatial variation relationship of the penetration direction vectors in the content voxel set, calculate the direction angle sequence between consecutive voxels in each candidate content path to generate the direction change sequence of each candidate content path, and the direction change sequence is used to characterize the spatial refraction trend of the candidate content path.

[0035] By performing volume rendering on candidate paths in the neural interest field, the cumulative distribution curve of interest density and the sequence of spatial direction changes are extracted, thereby realizing the quantifiable characterization of the interest accumulation and structural refraction of each path. This mechanism not only enhances the evaluation dimension of the content path but also provides a visual basis for path recommendation. Paths with continuous distribution under the interest volume can obtain higher interest density values, and those with stable direction changes have better structural consistency, effectively improving the overall logic of the pushed paths and user acceptance.

[0036] In this embodiment, S7 specifically includes: S71. Perform integral calculation on the cumulative distribution curve of the interest density of each candidate content path. The integral interval covers the sequentially arranged section of all content voxels in the path, and the density penetration value of the candidate content path in the neural interest field is calculated: ; Wherein, represents the density penetration value of the candidate content path, represents the interest density integral value of the cumulative distribution curve of the interest density at the path position , represents the total length of the path; S72. Calculate the refraction consistency value of each candidate content path according to the sequence of direction changes: ; Wherein, represents the refraction consistency value of the candidate content path, represents the total number of content voxels in the candidate content path, represents the value of the th dimension of the sequence of direction changes; S73. Normalize the density penetration value and the refraction consistency value of each candidate content path, and form the comprehensive score value of each candidate content path by combining them according to the preset weight; S74. Select the candidate content path with the largest comprehensive score value as the optimal content path.

[0037] Based on the integral and normalization processing of the quantitative indicators of interest density and direction consistency, and using weighted comprehensive scoring to select the optimal path, it ensures that the pushed results take into account both semantic depth and structural coherence, avoiding the problem in traditional recommendations of independently scoring content while ignoring the logical relationship between its sequences. Through path-level evaluation, semantic-driven structured pushing is realized, with stronger user interest fit, significantly improving the click-through rate and user satisfaction of the pushed content.

[0038] Refer to Figure 2 , a big data personalized push system includes the following modules: A user behavior modeling module, which is used to collect user behavior data and context information, and extract a set of user interest points; A neural interest field construction module, which is used to map the set of user interest points to a three-dimensional coordinate system, construct a neural network model with a NeRF structure and a three-dimensional interest volume function, and establish a neural interest field; A content voxel encoding module, which is used to parse the content to be pushed and construct a set of content voxels; A penetration mapping establishment module, which is used to generate penetration rays in the neural interest field according to the set of content voxels, perform volume sampling, calculate the cumulative response value, and screen out effective penetration paths according to the refraction tolerance value to generate a spatial action path; A path search and optimization module, which is used to construct a population of content paths, iteratively update the content path individuals based on an improved walrus optimization algorithm introducing an adaptive leadership mechanism, and output a set of candidate content paths; A volume rendering analysis module, which is used to input the set of candidate content paths into the neural interest field, perform volume rendering, and generate an interest density cumulative distribution curve and a direction change sequence for each candidate content path; A path evaluation and selection module, which is used to calculate the density penetration value according to the interest density cumulative distribution curve, calculate the refraction consistency value according to the direction change sequence, form a comprehensive score value, and select the one with the highest comprehensive score value as the optimal content path; A push output module, which is used to output the content to be pushed corresponding to the optimal content path, and form a personalized content perspective chain push result with spatial structure.

[0039] By modularizing the personalized push process, this system constructs a full-link recommendation architecture from user behavior collection, interest field modeling, content voxel encoding, penetration relationship establishment, path search and optimization, volume rendering analysis to final path screening and content push. The logic coupling between modules is tight, the data flow is smooth, and it has high scalability and operability for engineering implementation. By introducing the NeRF neural field expression model on the user side, the three-dimensional spatial expression of interest modeling is realized; by introducing the voxel construction and path optimization mechanism on the content side, the structured encoding of content expression is realized. The overall system can dynamically respond to changes in user behavior, accurately push a sequence of content paths that meet the current preferences, and improve the personality matching degree and structural coherence of the recommendation results. At the same time, the system has high model interpretability and computational efficiency optimization capabilities, is suitable for large-scale and high-concurrency recommendation application scenarios, and significantly improves the response speed and user experience of the push system.

[0040] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the content recommendation system of a provincial news and information mobile application platform. The platform has over 1.2 million daily active users, and its content covers government affairs, people's livelihood, finance, education, technology, and culture channels, with about 15,000 pieces of content updated daily. The original system uses a combination of collaborative filtering and traditional click-through rate prediction models for personalized recommendation, suffering from prominent problems such as insufficient dimensions of user interest expression, rough content recommendation granularity, lack of path correlation, high similarity and lack of coherence in recommended content, seriously affecting user retention rate and content click-through conversion rate.

[0041] The present invention is deployed and tested on this platform for 30 days, mainly pushing content to users of three content channels: news and current affairs, cutting-edge technology, and life services. The system generates joint features of user behavior context by collecting users' click, browse, dwell time, like, comment and other behavior data, as well as context data such as device type, network status, time period, and geographical location during user access. The BERT-Base model is used to perform semantic embedding on the content text, and a density-based clustering algorithm is used to extract user interest points from the behavior sequence. In three-dimensional space, the interest points are expressed in coordinate form, and a spatial interest volume function is established through the NeRF neural network to form a neural interest field. Subsequently, all content to be pushed is encoded as a voxel set. The voxel includes the spatial distribution center, penetration direction, and refraction tolerance value of the content, and a penetration path is established between the voxel ray and the interest field to form a spatial interaction mapping between the content and the user interest.

[0042] In the path search stage, an improved walrus optimization algorithm with an adaptive leadership mechanism is introduced to improve the diversity and fitness search ability of the recommended path. After 500 rounds of iteration, a candidate path set is generated. Subsequently, volume rendering is performed on each path in the neural interest field to obtain the cumulative distribution curve of interest density and the direction change sequence. Then, the path quality is evaluated by calculating the combination of density penetration value and refraction consistency value. Finally, the content sequence corresponding to the optimal content path is pushed to achieve spatial structured content perspective chain recommendation.

[0043] Through comparative experiments, it is found that the recommendation system adopting the method of the present invention significantly outperforms the original system in multiple key indicators. After implementing the system of the present invention on the technology front page of the platform, the average number of content clicks per user increased from 4.2 to 6.8, a growth of 61.9%; the average content stay time increased from 17 seconds to 29 seconds, a growth of 70.6%; the click-through rate of daily active users for content increased from 11.3% to 18.5%; the average scrolling reading depth of users increased from the original 1.9 screens to 3.1 screens, and the average semantic similarity between adjacent contents in the recommended content sequence decreased by 18.7%, indicating that the diversity and sequence fluency of the pushed content are significantly improved. In addition, the overall time-consuming of volume rendering and path evaluation on the server side is controlled within 890 ms, and the system has the ability to be stably launched.

[0044] Taking the current affairs news page of the platform as an example, after implementing the system of the present invention, the average number of content clicks per user increased from 3.7 to 5.9, a growth of 59.5%; the average content stay time increased from 14 seconds to 25 seconds, a growth of 78.6%; the click-through rate of daily active users for content increased from 9.6% to 15.2%; the average scrolling reading depth of users increased from the original 1.6 screens to 2.8 screens, and the average semantic similarity between adjacent contents in the recommended content sequence decreased by 17.4%, indicating that in the distribution process of government affairs information content, the system can more effectively capture the preference fluctuations of users under specific topics, and the pushed results are more in line with the focus of user attention and have a reasonable rhythm. In addition, after optimizing the volume sampling density of the neural interest field, it is maintained at the level of 1.5 × ray points per pixel in this page. Combined with the optimized path search, the overall push response time of the server side remains within 870 ms, and the system can stably handle the content request pressure of more than one million per day.

[0045] After deploying the three-dimensional interest push system proposed by the present invention on the life service page of the platform, the average number of content clicks per user per day increased from 4.0 to 6.1, a growth of 52.5%; the average content stay time increased from 15 seconds to 26 seconds, a growth of 73.3%; the click-through rate of daily active users increased from 10.1% to 14.8%, indicating that the adaptability and guiding ability of the system for pushing fragmented and scenario-based content are significantly enhanced. The scrolling reading depth increased from the original 1.8 screens to 2.9 screens, and the content sliding behavior of users tended to be continuous and in-depth, enhancing the immersive experience of life-related content; at the same time, the adjacent semantic similarity in the content recommendation sequence decreased by 19.2%, further reducing the cognitive fatigue caused by repetitive content. In the actual deployment of the platform, the daily generated scale of the voxel set of this page reached 420,000 groups. Through the parallel rendering scheduling mechanism, the system controlled the time-consuming of the content path search and evaluation process within 920 ms, ensuring that users can also obtain a smooth personalized recommendation experience under mobile network conditions.

[0046] This embodiment shows that the big data personalized push method based on NeRF neural interest field and improved walrus optimization algorithm can significantly improve the interaction depth and click conversion efficiency of users for different content channels, enhance the diversity and coherence of the push sequence, and at the same time, the system response is stable and the computing resources are controllable, having the feasibility and practical value for large-scale deployment in real business scenarios.

[0047] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent replacements or changes, should be covered within the protection scope of the present invention.

Claims

1. A big data personalized push method, characterized in that, It includes the following steps: S1. Collect user behavior data and context information, perform semantic embedding, and obtain a set of user interest points; S2. Based on the set of user interest points, use a neural network model with a NeRF structure to construct a three-dimensional interest volume function and establish a neural interest field; S3. Obtain all content to be pushed and encode each piece of content to be pushed into a content voxel set; S4. Establish a penetration mapping relationship between the content voxel set and the neural interest field, and determine the spatial action path of the content voxel on the three-dimensional interest volume function; S5. Based on the penetration mapping relationship, use an improved walrus optimization algorithm to perform content path search. The content path is sequentially composed of multiple content voxel sets, and a candidate content path set is generated; S6. Perform volume rendering on the candidate content path set in the neural interest field, and obtain the cumulative distribution curve of interest density and the direction change sequence of each candidate content path in the neural interest field; S7. Calculate the density penetration value of each candidate content path according to the cumulative distribution curve of interest density, calculate the refraction consistency value of each candidate content path according to the direction change sequence, and analyze and select the optimal content path; S8. Output the content to be pushed corresponding to the optimal content path as the final personalized push result.

2. The method for personalized push of big data according to claim 1, wherein The specific steps of S1 include: S11. Collect user behavior data, where the user behavior data includes click behavior, browsing behavior, stay duration, search requests, like behavior, and comment behavior, and record the timestamp and content identifier corresponding to each behavior; S12. Collect context information when the user behavior occurs. The context information includes terminal type, operating system type, geographical location information, access time information, and network connection status, and construct context information features; S13. Sort the user behavior data in ascending order of timestamp to construct a user behavior sequence, where the user behavior sequence is composed of structured triples containing behavior type, timestamp, and content identifier; S14. Perform semantic encoding on the text content corresponding to the content identifier, and use the BERT-Base pre-trained language model to encode the content title and body text to output semantic features; S15. Concatenate the context information features and the semantic features to form a user behavior context joint feature; S16. Apply a density-based spatial clustering algorithm to the user behavior context joint feature, use the Euclidean distance as the metric function, and extract the center of all clustering clusters as the set of user interest points.

3. A big data personalized push method according to claim 1, characterized in that The specific steps of S2 include: S21. Map each user interest point in the set of user interest points to a three-dimensional Cartesian coordinate system to obtain three-dimensional spatial position coordinates; S22. Calculate the unit direction vector corresponding to the position of each user interest point according to the spatial change direction of consecutive user interest points; S23. Construct a neural network model with a NeRF structure. The neural network model with a NeRF structure includes a position encoding module, a direction encoding module, a multi-layer perceptron network module, and an output layer module; The position encoding module performs position encoding on the input three-dimensional spatial position coordinates, and generates a position encoding feature vector of a preset length through the frequency transformation of sine and cosine functions; The direction encoding module performs the same sine and cosine frequency transformation processing on the input unit direction vector as the position encoding, and outputs a direction encoding feature vector of a preset length; The multi-layer perceptron network module includes an 8-layer fully connected network. The number of neurons in the first 7 layers is 256 each, and the number of neurons in the 8th layer is 128. Among them, the first 7 layers input the position encoding feature vector and output an intermediate position feature vector. After feature concatenation of the intermediate position feature vector and the direction encoding feature vector, it is input into the 8th layer to generate a comprehensive feature vector; The output layer module is used to output the interest density value, including a first fully connected layer, a first activation function layer, and a second fully connected layer connected in sequence; the first fully connected layer maps the input comprehensive feature vector to a 64-dimensional intermediate comprehensive feature vector, the first activation function layer uses the ReLU activation function to perform non-linear processing on the intermediate comprehensive feature vector, and the second fully connected layer maps the activated intermediate comprehensive feature vector to a single scalar output as the interest density value; S24. Calculate the three-dimensional interest volume function according to the neural radiance field volume rendering formula: ； ； Among them, represents the cumulative interest value of the th interest sampling path, represents the starting position of the interest sampling path, represents the th interest density value of the interest sampling path at position represents the cumulative transmittance from the starting position to position , represents the exponential function, represents the th interest density value of the interest sampling path at position S25. Converge the integral calculation results of the three-dimensional interest volume function in multiple directions of the interest path to construct a neural interest field, representing the spatial interest distribution of the user.

4. A big data personalized push method according to claim 1, characterized in that The specific steps of S3 are as follows: S31. Obtain all the content to be pushed, where the content to be pushed includes structured content and unstructured content, and the unstructured content includes text, image, video, and voice data; S32. Perform content parsing on each piece of content to be pushed, extract semantic principal components, tag information, and latent topic embeddings, and construct a multi-modal semantic feature vector; S33. Map the multi-modal semantic feature vector to the position coordinates in the three-dimensional content space, and the position coordinates form the spatial distribution center of the content to be pushed; S34. Based on the semantic intensity gradient and the push expected direction of the content to be pushed, construct a plurality of voxel units around the spatial distribution center. Each voxel unit includes spatial position coordinates, a penetration direction vector, and a refraction tolerance value; The spatial position coordinates are used to indicate the specific position of the voxel unit in the three-dimensional content space; The penetration direction vector is used to represent the directionality of the voxel unit in establishing a spatial interaction with the user interest path in the neural interest field. The penetration direction vector is obtained by weighted calculation based on the content semantic main axis and the user's historical preference direction; The refraction tolerance value is used to represent the allowable degree of semantic deviation after the content is penetrated by the interest sampling path in the direction of the voxel unit; S35. Based on all the voxel units corresponding to each piece of content to be pushed, obtain the content voxel set of each piece of content to be pushed.

5. A big data personalized push method according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. For the voxel units in each content voxel set, extract the spatial position coordinates and the penetration direction vector, and construct a voxel ray with the spatial position coordinates as the starting point and the penetration direction as the ray direction; S42. Input the voxel ray into the established neural interest field, perform volume sampling at preset intervals along each voxel ray in the neural interest field, record the interest density values at the sampling points, and form a corresponding interest density sequence; S43. Calculate the cumulative response value of the voxel ray in the neural interest field according to the interest density sequence, where the cumulative response value is the weighted integral result of the interest density values of all sampling points; S44. Set the interest density response threshold , compare the cumulative response value with the refraction tolerance value of the corresponding voxel unit . If the cumulative response value is greater than or equal to , it is determined that the voxel ray meets the penetration condition; S45. Take the voxel rays that meet the penetration condition as effective penetration paths, and classify them according to the ownership relationship of the content voxels to obtain the spatial action paths of each content voxel set in the neural interest field.

6. A big data personalized push method according to claim 1, characterized in that The specific steps of S5 are as follows: S51. Based on the spatial action paths of each content voxel set in the neural interest field, combine the content voxel sets of all content to be pushed to construct an initial content path population, where each content path is composed of several content voxel sets connected in sequence; S52. Assign a set of coding vectors to each content path individual, and the coding vectors are used to represent the arrangement order and direction offset of each content voxel set in the content path; S53. Define the fitness function of the content path, and the fitness function is weighted and combined according to the cumulative response value, content semantic relevance, and structural continuity of the content path; S54. The improvement of the improved walrus optimization algorithm lies in introducing an adaptive leadership mechanism to update the coding vectors of the content path individuals, and the adaptive leadership mechanism dynamically adjusts the guiding direction according to the optimal content path individual in the current iteration: ； Among them, represents the encoding vector of the th content path individual in the th round, represents the encoding vector of the th content path individual in the th round, represents the convergence control factor, represents the perturbation factor, represents the Euclidean distance between the th content path individual and the optimal content path individual in the th round, represents the encoding vector of the optimal content path individual in the th round, represents the diversity information entropy of the content path population in the th round: ； Among them, represents the dimension of the encoding vector, represents the proportion of the encoding value with the highest frequency of occurrence on the th encoding bit during the S55. Execute the content path iterative update process for a preset number of rounds, and retain several content path individuals with the highest fitness values in each round; S56. After all iterations are completed, output the finally retained content path individuals as a candidate content path set.

7. A big data personalized push method according to claim 1, characterized in that, The specific steps of S6 are as follows: Perform volume rendering on the candidate content path set in the neural interest field. For each candidate content path, weight and superimpose the interest density values on all penetration rays in the order of content voxels to obtain the cumulative interest density distribution curve of the candidate content path in the neural interest field. The cumulative interest density distribution curve takes the path position as the horizontal axis and the interest density integral value as the vertical axis; According to the spatial change relationship of the penetration direction vectors in the content voxel set, calculate the direction angle sequence between consecutive voxels in each candidate content path, and generate the direction change sequence of each candidate content path. The direction change sequence is used to describe the spatial refraction trend of the candidate content path.

8. A big data personalized push method according to claim 1, characterized in that The specific steps of S7 are as follows: S71. Perform integral calculation on the cumulative interest density distribution curve of each candidate content path, and the integral interval covers the sequential arrangement section of all content voxels in the path to calculate the density penetration value of the candidate content path in the neural interest field: ； Among them, represents the density penetration value of the candidate content path, represents the interest density integral value of the cumulative distribution curve of interest density at the path position ; represents the total path length. S72. Calculate the refraction consistency value of each candidate content path according to the direction change sequence; ； Among them, represents the refraction consistency value of the candidate content path, represents the total number of content voxels in the candidate content path, represents the th dimension value of the direction change sequence; S73. Normalize the density penetration value and refraction consistency value of each candidate content path, and form the comprehensive score value of each candidate content path according to the preset weight combination; S74. Select the candidate content path with the largest comprehensive score value as the optimal content path.

9. A big data personalized push system, which executes a big data personalized push method according to any one of claims 1 to 8, characterized in that It includes the following modules: User behavior modeling module, which is used to collect user behavior data and context information, and extract the set of user interest points; Neural interest field construction module, which is used to map the set of user interest points to a three-dimensional coordinate system, construct a neural network model with a NeRF structure and a three-dimensional interest volume function, and establish a neural interest field; Content voxel encoding module, which is used to parse the content to be pushed and construct a content voxel set; Penetration mapping establishment module, which is used to generate penetration rays in the neural interest field according to the content voxel set, perform volume sampling, calculate the cumulative response value, and screen out effective penetration paths based on the refraction tolerance value to generate a spatial action path; Path search and optimization module, which is used to construct a content path population, iteratively update the content path individuals based on an improved walrus optimization algorithm introducing an adaptive leadership mechanism, and output a set of candidate content paths; Volume rendering analysis module, which is used to input the set of candidate content paths into the neural interest field, perform volume rendering, and generate the cumulative distribution curve of interest density and the direction change sequence of each candidate content path; Path evaluation and selection module, which is used to calculate the density penetration value according to the cumulative distribution curve of interest density, calculate the refraction consistency value according to the direction change sequence, form a comprehensive score value, and select the highest comprehensive score value as the optimal content path; Push output module, which is used to output the content to be pushed corresponding to the optimal content path, and form a personalized content perspective chain push result with spatial structure.

Images