A method and system for personalized big data push

Abstract

The present invention discloses a method and system for personalized push notifications based on big data, comprising the following steps: S1, collecting user behavior data and contextual information to obtain a set of user points of interest; S2, using a neural network model with a NeRF structure to construct a three-dimensional volume of interest function and establish a neural field of interest; S3, obtaining all content to be pushed and encoding it into a set of content voxels; S4, establishing a penetration mapping relationship between the content voxel set and the neural field of interest; S5, using an improved walrus optimization algorithm to perform content path search and generate a set of candidate content paths; S6, obtaining a cumulative distribution curve of interest density and a sequence of direction changes for each candidate content path in the neural field of interest; S7, analyzing and selecting the optimal content path; S8, outputting the content to be pushed corresponding to the optimal content path as the final personalized push result. The present invention integrates NeRF modeling with the improved walrus optimization algorithm to achieve personalized push notifications based on big data.

Description

Technical Field

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

[0002] With the rapid development of mobile internet and the widespread adoption of smart devices, personalized content push technology has become a key means of improving user experience and platform stickiness. By modeling users' historical behavior, interests, preferences, and content features, the system can filter out information that users may be interested in from massive amounts of information and achieve precise push notifications. Existing mainstream methods typically use collaborative filtering, matrix factorization, deep neural networks, graph neural networks, and attention mechanisms to build user-content association models to predict users' ratings or click probabilities for candidate content. However, most of these methods rely on low-dimensional interest space modeling, which 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 embeddings or knowledge graphs for enhanced expression, but these methods are still mainly based on discrete feature spaces or topological structures, and lack modeling of the interaction mechanism between user interests and content features in continuous space. Especially in the push scenario for multimodal content, the interaction between the high-dimensional distribution of unstructured information such as text, images, and videos and user preferences is more complex. Traditional representation learning methods have difficulty capturing the spatial mapping of multimodal information and interest structures. In addition, existing content recommendation methods are mostly based on single-point prediction or ranking optimization, and fail to effectively introduce spatial continuity constraints and path optimization mechanisms. The organization and recommendation of content sequences still tend to be fragmented.

[0004] In recent years, Neural Radiance Field (NeRF), a neural implicit representation model for 3D reconstruction and volume rendering, has demonstrated strong spatial modeling capabilities. Some studies have attempted to use NeRF for behavioral simulation or scene understanding, but it has not yet been used to construct a 3D representation of user interests, thereby enabling personalized content push based on a spatial perspective mechanism. Meanwhile, swarm intelligence optimization algorithms have achieved good results in path planning and combination problems, but in content recommendation scenarios, problems such as unclear optimization objective function construction and coarse search granularity have not been fully addressed. In particular, when dealing with the complex nonlinear 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 optimality, and lack the stability of recommendation results.

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

[0006] One object of the present invention is to propose a method and system for personalized push of big data. The present invention integrates NeRF modeling and an improved walrus optimization algorithm to construct a three-dimensional volume function of user interest, 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. It 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 low user modeling dimensionality, coarse recommendation granularity, and weak path optimization capabilities.

[0007] A method for personalized big data push according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect user behavior data and context information, perform semantic embedding, and obtain the user interest point set;

[0009] S2. Based on the user interest point set, a neural network model with NeRF structure is used to construct a three-dimensional interest volume function and establish a neural interest field;

[0010] S3. Obtain all content to be pushed and encode each content to be pushed into a content voxel set;

[0011] S4. Establishing a penetration mapping relationship between the content voxel set and the neural field of interest, and determining a spatial action path of the content voxel on the three-dimensional volume of interest function;

[0012] S5. Based on the penetration mapping relationship, an improved walrus optimization algorithm is used to perform content path search, where the content path is sequentially composed of multiple content voxel sets to generate a candidate content path set;

[0013] S6. Perform volume rendering on the candidate content path set in the neural interest field to obtain a cumulative distribution curve of interest density and a direction change sequence of each candidate content path in the neural interest field;

[0014] S7. Calculate the density penetration value of each candidate content path based on the interest density cumulative distribution curve, calculate the refraction consistency value of each candidate content path based on the direction change sequence, analyze and select the optimal content path;

[0015] S8. Output the content to be pushed corresponding to the optimal content path as the final personalized push result.

[0016] Optionally, the S1 specifically includes:

[0017] S11. Collect user behavior data, including click behavior, browsing behavior, stay time, search requests, like behavior, and comment behavior, and record the timestamp and content identifier corresponding to each behavior;

[0018] S12. Collect context information when the user behavior occurs, the context information including terminal type, operating system type, geographic location information, access time information, and network connection status, and construct context information features;

[0019] S13. Sort the user behavior data in ascending order according to timestamps to construct a user behavior sequence, where the user behavior sequence consists of a structured triple containing a behavior type, a timestamp, and a content identifier;

[0020] S14. Perform semantic encoding on the text content corresponding to the content identifier, use the BERT-Base pre-trained language model to encode the content title and body, and output semantic features;

[0021] S15, performing feature splicing on the context information feature and the semantic feature to form a user behavior context joint feature;

[0022] S16. Apply a density-based spatial clustering algorithm to the user behavior context joint features, use Euclidean distance as a metric function, and extract the centers of all clusters as a set of user interest points.

[0023] Optionally, the S2 specifically includes:

[0024] S21, mapping each user interest point in the user interest point set to a three-dimensional Cartesian coordinate system to obtain a three-dimensional spatial position coordinate;

[0025] S22. Calculate a unit direction vector corresponding to the position of each user interest point based on the spatial change direction of the consecutive user interest points;

[0026] S23, constructing a neural network model of NeRF structure, wherein the neural network model of NeRF structure includes a position encoding module, a direction encoding module, a multi-layer perceptron network module and an output layer module;

[0027] 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 frequency transformation of sine and cosine functions;

[0028] The direction encoding module processes the input unit direction vector row with the same sine-cosine frequency transformation as the position encoding, and outputs a direction encoding feature vector of a preset length;

[0029] The multilayer 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, wherein the first 7 layers input the position coding feature vector and output the intermediate position feature vector, the intermediate position feature vector and the direction coding feature vector are feature-concatenated and input into the 8th layer to generate a comprehensive feature vector;

[0030] 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 a ReLU activation function to perform nonlinear 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;

[0031] S24. Calculate the three-dimensional volume of interest function according to the neural radiation field volume rendering formula:

[0032] ;

[0033] ;

[0034] in, Indicates the The cumulative interest value of the interest sampling path, Indicates the starting position of the sampling path of interest, Indicates the end position of the sampling path of interest, Indicates the Interest sampling paths at location The interest density value on Indicates from the starting position to the position The cumulative transmittance, represents the exponential function, Indicates the Interest sampling paths at location Interest density value on ;

[0035] S25. Aggregate the integral calculation results of the three-dimensional interest volume function in the directions of multiple interest paths to construct a neural interest field to represent the spatial interest distribution of the user.

[0036] Optionally, the S3 specifically includes:

[0037] S31. Acquire all content to be pushed, where the content to be pushed includes structured content and unstructured content, wherein the unstructured content includes text, images, videos, and voice data;

[0038] S32. Analyze each content to be pushed, extract semantic principal components, tag information, and potential topic embedding, and construct a multimodal semantic feature vector;

[0039] S33: Mapping the multimodal semantic feature vector to position coordinates in a three-dimensional content space, where the position coordinates constitute the spatial distribution center of the content to be pushed;

[0040] S34. Based on the semantic intensity gradient of the content to be pushed and the expected push direction, construct multiple voxel units around the spatial distribution center, each voxel unit including spatial position coordinates, a penetration direction vector, and a refraction tolerance value;

[0041] The spatial position coordinates are used to indicate the specific position of the voxel unit in the three-dimensional content space;

[0042] The penetration direction vector is used to represent the directionality of the spatial interaction between the voxel unit and 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 axis and the user's historical preference direction;

[0043] The refraction tolerance value is used to indicate the permissible degree of semantic deviation of the content after it is penetrated by the sampling path of interest in the direction of the voxel unit;

[0044] S35 . Based on all voxel units corresponding to each content to be pushed, obtain a content voxel set of each content to be pushed.

[0045] Optionally, the S4 specifically includes:

[0046] S41. For each voxel unit in the 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;

[0047] S42, inputting the voxel ray into the established neural interest field, performing volume sampling along each voxel ray in the neural interest field at a preset interval, recording the interest density value at the sampling point, and forming a corresponding interest density sequence;

[0048] S43, calculating the cumulative response value of the voxel ray in the neural interest field according to the interest density sequence, wherein the cumulative response value is a weighted integral result of the interest density values ​​of all sampling points;

[0049] S44. Setting interest density response threshold , the cumulative response value is compared with the refraction tolerance value of the corresponding voxel unit Compare, if the cumulative response value is greater than or equal to , then the voxel ray is determined to meet the penetration condition;

[0050] S45. The voxel rays that meet the penetration condition are regarded as effective penetration paths, and are classified according to the relationship between the content voxels, so as to obtain the spatial action path of each content voxel set in the neural field of interest.

[0051] Optionally, the S5 specifically includes:

[0052] S51. Based on the spatial action path of each content voxel set in the neural interest field, combine all 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 time sequence;

[0053] S52: Allocate a set of encoding vectors to each content path individual, where the encoding vectors are used to represent the arrangement order and direction offset of each content voxel set in the content path;

[0054] S53: defining a fitness function for the content path, wherein the fitness function is a weighted combination of the cumulative response value, content semantic relevance, and structural continuity of the content path;

[0055] S54. The improvement of the improved walrus optimization algorithm is to introduce an adaptive leadership mechanism to update the encoding vector of the content path individual. The adaptive leadership mechanism dynamically adjusts the guidance direction according to the optimal content path individual in the current iteration:

[0056] ;

[0057] in, Indicates the Round The encoding vector of each content path individual, Indicates the Round The encoding vector of each content path, represents the convergence control factor, represents the disturbance factor, Indicates the Round The Euclidean distance between the individual content path and the optimal content path individual, Indicates the The encoding 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, Indicates the Diversity information entropy of content path groups in a round:

[0058] ;

[0059] in, represents the dimension of the encoding vector, Indicates the During the round of iteration The proportion of the most frequently occurring code value in each code position;

[0060] S55, executing a preset number of rounds of iterative content path update process, and retaining several content path individuals with the highest fitness values ​​in each round;

[0061] S56: After all iterations are completed, the content path individuals that are finally retained are output as a candidate content path set.

[0062] Optionally, S6 specifically includes: performing volume rendering on the candidate content path set in the neural interest field, and for each candidate content path, performing weighted superposition of interest density values ​​on all penetration rays according to the order of content voxels, to obtain a cumulative distribution curve of interest density of the candidate content path in the neural interest field, wherein the cumulative distribution curve of interest density has the path position as the horizontal axis and the interest density integral value as the vertical axis;

[0063] According to the spatial variation relationship of the penetration direction vector in the content voxel set, a sequence of direction angles between consecutive voxels in each candidate content path is calculated to generate a direction variation sequence for each candidate content path. The direction variation sequence is used to characterize the spatial refraction trend of the candidate content path.

[0064] Optionally, the S7 specifically includes:

[0065] S71. Integrate the interest density cumulative distribution curve of each candidate content path, where the integral interval covers the sequentially arranged section of all content voxels in the path, and calculate the density penetration value of the candidate content path in the neural interest field:

[0066] ;

[0067] in, Indicates the density penetration value of the candidate content path. Indicates the cumulative distribution curve of interest density at the path position The integral value of the density of interest at Indicates the total length of the path;

[0068] S72. Calculate the refraction consistency value of each candidate content path according to the direction change sequence:

[0069] ;

[0070] in, Refractive consistency value representing the candidate content path, represents the total number of content voxels in the candidate content path, Indicates the direction change sequence The value of the dimension;

[0071] S73. Normalize the density penetration value and refraction consistency value of each candidate content path, and combine them according to preset weights to form a comprehensive score for each candidate content path;

[0072] S74: Select the candidate content path with the largest comprehensive score as the optimal content path.

[0073] A big data personalized push system according to an embodiment of the present invention includes the following modules:

[0074] User behavior modeling module, used to collect user behavior data and context information, and extract user interest point sets;

[0075] The neural interest field construction module is used to map the user's interest point set to a three-dimensional coordinate system, construct a neural network model of the NeRF structure and a three-dimensional interest volume function, and establish a neural interest field;

[0076] Content voxel encoding module, used to parse the content to be pushed and construct a content voxel set;

[0077] The penetration mapping establishment module is used to generate penetration rays in the neural field of interest based on the content voxel set, perform volume sampling, calculate the cumulative response value, and screen the effective penetration path according to the refraction tolerance value to generate the spatial action path;

[0078] The path search optimization module is used to build a content path population, iteratively update content path individuals based on the improved walrus optimization algorithm that introduces an adaptive leadership mechanism, and output a set of candidate content paths;

[0079] A volume rendering analysis module is used to input the candidate content path set into the neural interest field, perform volume rendering, and generate a cumulative distribution curve of interest density and a direction change sequence for each candidate content path;

[0080] A path evaluation and selection module is used to calculate the density penetration value based on the interest density cumulative distribution curve, calculate the refraction consistency value based on the direction change sequence, form a comprehensive score value, and select the path with the highest comprehensive score value as the optimal content path;

[0081] The push output module is used to output the content to be pushed corresponding to the optimal content path, forming a spatially structured personalized content perspective chain push result.

[0082] The beneficial effects of the present invention are:

[0083] (1) By constructing a neural interest field based on the NeRF structure, user interests are modeled in the form of a continuous spatial volume function for the first time, effectively capturing the distribution characteristics and directional change laws of interest density in space, breaking through the limitation of traditional low-dimensional embedding models that cannot reflect directionality and spatial continuity.

[0084] (2) By encoding the content to be pushed into a set of penetrable voxels and combining the penetration direction and refraction tolerance modeling, a spatial penetration mapping between the content voxels and the neural interest field is established, achieving high coupling modeling of content and interest in three-dimensional space, and enhancing the physical interpretability of the matching.

[0085] (3) A content path search method based on the improved walrus optimization algorithm is proposed, which introduces an adaptive leadership mechanism and diversity control strategy to achieve the global optimal search of multi-content voxel combination paths in the interest field, avoiding the problems of fragmentation and local greedy strategies in traditional recommendation.

[0086] (4) By performing volume rendering on the candidate content path, the cumulative distribution of its interest density and the directional change trend are quantified, and the density penetration value and refraction consistency value are calculated based on this, ensuring that the final recommended path is dually optimized in terms of spatial continuity and interest fit. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The accompanying 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 of the present invention. In the accompanying drawings:

[0088] Figure 1 This is an overall flow chart of a big data personalized push method proposed by the present invention;

[0089] Figure 2 This is a structural diagram of a big data personalized push system proposed by the present invention. DETAILED DESCRIPTION

[0090] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0091] refer to Figure 1 , a big data personalized push method, comprising the following steps:

[0092] S1. Collect user behavior data and context information, perform semantic embedding, and obtain the user interest point set;

[0093] S2. Based on the user interest point set, a neural network model with NeRF structure is used to construct a three-dimensional interest volume function and establish a neural interest field;

[0094] S3. Obtain all content to be pushed and encode each content to be pushed into a content voxel set;

[0095] S4. Establishing a penetration mapping relationship between the content voxel set and the neural field of interest, and determining a spatial action path of the content voxel on the three-dimensional volume of interest function;

[0096] S5. Based on the penetration mapping relationship, an improved walrus optimization algorithm is used to perform content path search, where the content path is sequentially composed of multiple content voxel sets to generate a candidate content path set;

[0097] S6. Perform volume rendering on the candidate content path set in the neural interest field to obtain a cumulative distribution curve of interest density and a direction change sequence of each candidate content path in the neural interest field;

[0098] S7. Calculate the density penetration value of each candidate content path based on the interest density cumulative distribution curve, calculate the refraction consistency value of each candidate content path based on the direction change sequence, analyze and select the optimal content path;

[0099] S8. Output the content to be pushed corresponding to the optimal content path as the final personalized push result.

[0100] This method constructs a three-dimensional volume of interest function by introducing a NeRF-structured neural network model and a spatial penetration mechanism for content voxels within the neural interest field, establishing a highly coupled spatial expression relationship between user interests and content features. Furthermore, it combines this with an improved walrus optimization algorithm for path-level push search, moving beyond the traditional recommendation approach of single-content or short-window similarity matching. This method possesses a closed-loop structure, from user interest modeling to path planning and sequence output, effectively improving the semantic relevance, sequence coherence, and user click-through conversion rate of the recommendation system. Furthermore, the introduction of techniques such as spatial penetration, volume rendering, and density accumulation analysis significantly enhances the structural expressiveness and interpretability of content recommendations.

[0101] In this embodiment, S1 specifically includes:

[0102] S11. Collect user behavior data, including click behavior, browsing behavior, stay time, search requests, like behavior, and comment behavior, and record the timestamp and content identifier corresponding to each behavior;

[0103] S12. Collect context information when the user behavior occurs, the context information including terminal type, operating system type, geographic location information, access time information, and network connection status, and construct context information features;

[0104] S13. Sort the user behavior data in ascending order according to timestamps to construct a user behavior sequence, where the user behavior sequence consists of a structured triple containing a behavior type, a timestamp, and a content identifier;

[0105] S14. Perform semantic encoding on the text content corresponding to the content identifier, use the BERT-Base pre-trained language model to encode the content title and body, and output semantic features;

[0106] S15, performing feature splicing on the context information feature and the semantic feature to form a user behavior context joint feature;

[0107] S16. Apply a density-based spatial clustering algorithm to the user behavior context joint features, use Euclidean distance as a metric function, and extract the centers of all clusters as a set of user interest points.

[0108] By collecting multi-dimensional user behavior data and contextual information, combining it with the BERT language model to deeply encode content semantics, and then integrating this contextual information to construct a joint feature representation, the extraction of user points of interest is semantically sensitive and adaptive to the behavioral context. Furthermore, by spatially partitioning user behavior features through density clustering, it effectively extracts highly representative, non-repetitive points of interest, avoiding the abstractness, static nature, and low update frequency issues of traditional user profiling. This significantly improves the accuracy and real-time nature of user interest modeling, providing more discriminative interest coordinates for spatial modeling and path planning.

[0109] In this embodiment, S2 specifically includes:

[0110] S21, mapping each user interest point in the user interest point set to a three-dimensional Cartesian coordinate system to obtain a three-dimensional spatial position coordinate;

[0111] S22. Calculate a unit direction vector corresponding to the position of each user interest point based on the spatial change direction of the consecutive user interest points;

[0112] S23, constructing a neural network model of NeRF structure, wherein the neural network model of NeRF structure includes a position encoding module, a direction encoding module, a multi-layer perceptron network module and an output layer module;

[0113] 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 frequency transformation of sine and cosine functions;

[0114] The direction encoding module processes the input unit direction vector row with the same sine-cosine frequency transformation as the position encoding, and outputs a direction encoding feature vector of a preset length;

[0115] The multilayer 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, wherein the first 7 layers input the position coding feature vector and output the intermediate position feature vector, the intermediate position feature vector and the direction coding feature vector are feature-concatenated and input into the 8th layer to generate a comprehensive feature vector;

[0116] 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 a ReLU activation function to perform nonlinear 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;

[0117] S24. Calculate the three-dimensional volume of interest function according to the neural radiation field volume rendering formula:

[0118] ;

[0119] ;

[0120] in, Indicates the The cumulative interest value of the interest sampling path, Indicates the starting position of the sampling path of interest, Indicates the end position of the sampling path of interest, Indicates the Interest sampling paths at location The interest density value on Indicates from the starting position to the position The cumulative transmittance, represents the exponential function, Indicates the Interest sampling paths at location Interest density value on ;

[0121] S25. Aggregate the integral calculation results of the three-dimensional interest volume function in the directions of multiple interest paths to construct a neural interest field to represent the spatial interest distribution of the user.

[0122] A neural network model with a NeRF structure is introduced to model user points of interest as differentiable three-dimensional interest volume functions. This is the first time that two types of variables, spatial position and direction vector, are integrated into interest expression, enabling continuous spatial modeling of user interests. By combining position encoding with direction encoding, a multilayer perceptron network accurately models the response changes of the interest density field in different directions. Furthermore, the output layer design supports the precise generation of interest density values, making the interest distribution both high-resolution and adjustable, 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.

[0123] In this embodiment, S3 specifically includes:

[0124] S31. Acquire all content to be pushed, where the content to be pushed includes structured content and unstructured content, wherein the unstructured content includes text, images, videos, and voice data;

[0125] S32. Analyze each content to be pushed, extract semantic principal components, tag information, and potential topic embedding, and construct a multimodal semantic feature vector;

[0126] S33: Mapping the multimodal semantic feature vector to position coordinates in a three-dimensional content space, where the position coordinates constitute the spatial distribution center of the content to be pushed;

[0127] S34. Based on the semantic intensity gradient of the content to be pushed and the expected push direction, construct multiple voxel units around the spatial distribution center, each voxel unit including spatial position coordinates, a penetration direction vector, and a refraction tolerance value;

[0128] The spatial position coordinates are used to indicate the specific position of the voxel unit in the three-dimensional content space;

[0129] The penetration direction vector is used to represent the directionality of the spatial interaction between the voxel unit and 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 axis and the user's historical preference direction;

[0130] The refraction tolerance value is used to indicate the permissible degree of semantic deviation of the content after it is penetrated by the sampling path of interest in the direction of the voxel unit;

[0131] S35 . Based on all voxel units corresponding to each content to be pushed, obtain a content voxel set of each content to be pushed.

[0132] The multimodal content to be pushed is uniformly mapped to a three-dimensional content space, and a voxel set characterized by spatial position, penetration direction, and semantic tolerance is constructed. This effectively addresses the lack of spatial position information and the single expression dimension in traditional content representation. Voxelized modeling enables spatial testing of the directionality and tolerance of content semantics in the neural interest field, laying the foundation for establishing penetration mapping. In addition, the penetration direction is jointly constructed based on the semantic axis and user preferences, which has stronger personalized adaptability and enhances the geometric rationality and semantic relevance of the match between content and interests.

[0133] In this embodiment, the S4 specifically includes:

[0134] S41. For each voxel unit in the 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;

[0135] S42, inputting the voxel ray into the established neural interest field, performing volume sampling along each voxel ray in the neural interest field at a preset interval, recording the interest density value at the sampling point, and forming a corresponding interest density sequence;

[0136] S43, calculating the cumulative response value of the voxel ray in the neural interest field according to the interest density sequence, wherein the cumulative response value is a weighted integral result of the interest density values ​​of all sampling points;

[0137] S44. Setting interest density response threshold , the cumulative response value is compared with the refraction tolerance value of the corresponding voxel unit Compare, if the cumulative response value is greater than or equal to , then the voxel ray is determined to meet the penetration condition;

[0138] S45. The voxel rays that meet the penetration condition are regarded as effective penetration paths, and are classified according to the relationship between the content voxels, so as to obtain the spatial action path of each content voxel set in the neural field of interest.

[0139] By mapping content voxels into rays and inputting them into the neural interest field, and performing point-by-point sampling and integration based on the interest density field, this step accurately captures the spatial response path between content and user interests. By introducing a comparison mechanism between cumulative response values ​​and refraction tolerance, the system physically determines whether content semantics can penetrate user interests, providing unprecedented push interpretability. Furthermore, constructing effective penetration paths helps improve the initial quality of content path searches, narrowing the search space and enhancing the structural continuity of push paths.

[0140] In this embodiment, the S5 specifically includes:

[0141] S51. Based on the spatial action path of each content voxel set in the neural interest field, combine all 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 time sequence;

[0142] S52: Allocate a set of encoding vectors to each content path individual, where the encoding vectors are used to represent the arrangement order and direction offset of each content voxel set in the content path;

[0143] S53: defining a fitness function for the content path, wherein the fitness function is a weighted combination of the cumulative response value, content semantic relevance, and structural continuity of the content path;

[0144] S54. The improvement of the improved walrus optimization algorithm is to introduce an adaptive leadership mechanism to update the encoding vector of the content path individual. The adaptive leadership mechanism dynamically adjusts the guidance direction according to the optimal content path individual in the current iteration:

[0145] ;

[0146] in, Indicates the Round The encoding vector of each content path, Indicates the Round The encoding vector of each content path individual, represents the convergence control factor, represents the disturbance factor, Indicates the Round The Euclidean distance between the individual content path and the optimal content path individual, Indicates the The encoding 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, Indicates the Diversity information entropy of content path groups in a round:

[0147] ;

[0148] in, represents the dimension of the encoding vector, Indicates the During the round of iteration The proportion of the most frequently occurring code value in each code position;

[0149] S55, executing a preset number of rounds of iterative content path update process, and retaining several content path individuals with the highest fitness values ​​in each round;

[0150] S56: After all iterations are completed, the content path individuals that are finally retained are output as a candidate content path set.

[0151] The content path search mechanism is based on an improved walrus optimization algorithm. By designing an adaptive leadership mechanism and a path diversity entropy guidance strategy, it maintains population activity during the search process and prevents regression into local optima. Furthermore, individual path encodings incorporate information about content order and directional offsets, resulting in a more structured and logically optimized optimization result. The fitness function combines interest response value, semantic relevance, and structural continuity to ensure that search results are not only relevant but also semantically coherent. The resulting set of candidate paths provides the recommendation system with a more rational set of multiple interesting paths.

[0152] In this embodiment, S6 specifically includes: performing volume rendering on the candidate content path set in the neural interest field, and for each candidate content path, weightedly superimposing the interest density values ​​on all penetration rays according to the order of content voxels, to obtain a cumulative distribution curve of interest density of the candidate content path in the neural interest field, wherein the cumulative distribution curve of interest density has the path position as the horizontal axis and the interest density integral value as the vertical axis;

[0153] According to the spatial variation relationship of the penetration direction vector in the content voxel set, a sequence of direction angles between consecutive voxels in each candidate content path is calculated to generate a direction variation sequence for each candidate content path. The direction variation sequence is used to characterize the spatial refraction trend of the candidate content path.

[0154] By performing volume rendering on candidate paths within the neural interest field, we extract the cumulative distribution curve of interest density and the sequence of spatial directional changes, thereby achieving a quantifiable characterization of the accumulated interest and structural refraction of each path. This mechanism not only enhances the evaluation dimensionality of content paths but also provides a visual basis for path recommendations. Paths with continuous distribution within the interest volume achieve higher interest density values, while those with stable directional changes exhibit greater structural consistency, effectively improving the overall logic and user acceptance of the recommended paths.

[0155] In this embodiment, the S7 specifically includes:

[0156] S71. Integrate the interest density cumulative distribution curve of each candidate content path, where the integral interval covers the sequentially arranged section of all content voxels in the path, and calculate the density penetration value of the candidate content path in the neural interest field:

[0157] ;

[0158] in, Indicates the density penetration value of the candidate content path. Indicates the cumulative distribution curve of interest density at the path position The integral value of the density of interest at Indicates the total length of the path;

[0159] S72. Calculate the refraction consistency value of each candidate content path according to the direction change sequence:

[0160] ;

[0161] in, Refractive consistency value representing the candidate content path, represents the total number of content voxels in the candidate content path, Indicates the direction change sequence The value of the dimension;

[0162] S73. Normalize the density penetration value and refraction consistency value of each candidate content path, and combine them according to preset weights to form a comprehensive score for each candidate content path;

[0163] S74: Select the candidate content path with the largest comprehensive score as the optimal content path.

[0164] Based on the integration and normalization of quantitative indicators of interest density and direction consistency, and the use of weighted comprehensive scoring to select the optimal path, it ensures that the push results take into account both semantic depth and structural coherence, avoiding the problem of independent scoring of content while ignoring the logical relationship between its sequences in traditional recommendations. Through path-level evaluation, semantically driven structured push is achieved, which has a stronger fit with user interests and significantly improves the click-through rate and user satisfaction of pushed content.

[0165] refer to Figure 2 , a big data personalized push system, including the following modules:

[0166] User behavior modeling module, used to collect user behavior data and context information, and extract user interest point sets;

[0167] The neural interest field construction module is used to map the user's interest point set to a three-dimensional coordinate system, construct a neural network model of the NeRF structure and a three-dimensional interest volume function, and establish a neural interest field;

[0168] Content voxel encoding module, used to parse the content to be pushed and construct a content voxel set;

[0169] The penetration mapping establishment module is used to generate penetration rays in the neural field of interest based on the content voxel set, perform volume sampling, calculate the cumulative response value, and screen the effective penetration path according to the refraction tolerance value to generate the spatial action path;

[0170] The path search optimization module is used to build a content path population, iteratively update content path individuals based on the improved walrus optimization algorithm that introduces an adaptive leadership mechanism, and output a set of candidate content paths;

[0171] A volume rendering analysis module is used to input the candidate content path set into the neural interest field, perform volume rendering, and generate a cumulative distribution curve of interest density and a direction change sequence for each candidate content path;

[0172] A path evaluation and selection module is used to calculate the density penetration value based on the interest density cumulative distribution curve, calculate the refraction consistency value based on the direction change sequence, form a comprehensive score value, and select the path with the highest comprehensive score value as the optimal content path;

[0173] The push output module is used to output the content to be pushed corresponding to the optimal content path, forming a spatially structured personalized content perspective chain push result.

[0174] This system modularizes the personalized push process and builds a full-link recommendation architecture from user behavior collection, interest field modeling, content voxel encoding, penetration relationship establishment, path search optimization, volume rendering analysis to final path screening and content push. The modules are tightly logically coupled and data flows smoothly, with high scalability and engineering feasibility. 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 content path sequences that meet current preferences, and improve the personalized matching and structural coherence of recommendation results. At the same time, the system has a high degree of model interpretability and computational efficiency optimization capabilities, which is suitable for large-scale, high-concurrency recommendation application scenarios, significantly improving the response speed of the push system and user experience.

[0175] Example 1:

[0176] To verify the feasibility of this invention, we applied it to the content recommendation system of a provincial news and information mobile application platform. The platform boasts over 1.2 million daily active users, covering government affairs, livelihoods, finance, education, science and technology, and culture, with approximately 15,000 pieces of content updated daily. The original system, which combined collaborative filtering with traditional click-through rate prediction models for personalized recommendations, suffered from insufficient dimensionality in user interest expression, coarse content recommendation granularity, missing path associations, and a high degree of similarity and lack of coherence in recommended content, severely impacting both user retention and content click-through rates.

[0177] The present invention was deployed in the platform and tested for 30 days, focusing on pushing content to users of the three content channels of news and current affairs, science and technology, and life services. The system generates user behavior context joint features by collecting user behavior data such as clicks, browsing, dwell time, likes, comments, and context data such as device type, network status, time period, and geographic location when the user visits. The BERT-Base model is used to semantically embed the content text, and a density-based clustering algorithm is used to extract user interest points from the behavior sequence. In three-dimensional space, points of interest are expressed in the form of coordinates, 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 into a set of voxels, which includes the spatial distribution center of the content, the penetration direction, and the refraction tolerance value. A penetration path is established through voxel rays and the field of interest to form a spatial interactive mapping of content and user interests.

[0178] During the path search phase, an improved walrus optimization algorithm with an adaptive leadership mechanism was introduced to enhance the diversity and fitness of recommended paths. After 500 iterations, a set of candidate paths was generated. Each path was then volume-rendered within the neural interest field to generate a cumulative distribution curve of interest density and a sequence of directional changes. Path quality was then evaluated by combining density penetration and refraction consistency values. Finally, the corresponding content sequence in the optimal content path was pushed, enabling spatially structured content perspective chain recommendations.

[0179] Through comparative experiments, it was found that the recommendation system using the method of the present invention is significantly better than the original system in many key indicators. After the implementation of the system of the present invention in the platform's Science and Technology Frontier Channel, the average number of content clicks per user increased from 4.2 to 6.8, an increase of 61.9%; the average content dwell time increased from 17 seconds to 29 seconds, an increase of 70.6%; the daily active user content click-through rate increased from 11.3% to 18.5%; the average user scrolling reading depth increased from the original 1.9 screens to 3.1 screens, and the average semantic similarity of adjacent content in the recommended content sequence decreased by 18.7%, indicating that the diversity of pushed content and the fluency of the sequence have been significantly improved. In addition, the overall time consumption of server-side volume rendering and path evaluation is controlled within 890ms, and the system has the ability to be stably launched.

[0180] Taking the platform's current affairs news channel 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 59.5% increase; the average dwell time on content increased from 14 seconds to 25 seconds, a 78.6% increase; the daily active user content click-through rate increased from 9.6% to 15.2%; the average user scrolling reading depth increased from 1.6 screens to 2.8 screens, and the average semantic similarity of adjacent content in the recommended content sequence decreased by 17.4%, indicating that in the distribution process of government information content, the system can more effectively capture the fluctuations in user preferences on specific topics, and the push results are more in line with the user's focus and at a reasonable rhythm. In addition, after optimization, the volume sampling density of the neural interest field is maintained at a level of 1.5× per pixel ray points in this channel. Combined with the optimized path search, the overall push response time of the server is kept within 870ms, and the system can stably cope with the pressure of over one million content requests per day.

[0181] After the platform's life service channel deployed the three-dimensional interest push system proposed in this invention, the average daily number of clicks on content by users increased from 4.0 to 6.1, an increase of 52.5%; the average content dwell time increased from 15 seconds to 26 seconds, an increase of 73.3%; and the daily active click-through rate increased from 10.1% to 14.8%, indicating that the system's adaptability and guidance capabilities for fragmented and scenario-based content push have been significantly enhanced. The scrolling reading depth has increased from the original 1.8 screens to 2.9 screens, and the user's content sliding behavior has 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 has decreased by 19.2%, further reducing the cognitive fatigue caused by repetitive content. In the actual deployment of the platform, the average daily generation scale of the channel voxel set reached 420,000 groups. Through the parallel rendering scheduling mechanism, the system controls the content path search and evaluation process time to within 920ms, ensuring that users can also obtain a smooth personalized recommendation experience under mobile network conditions.

[0182] This embodiment shows that the big data personalized push method based on NeRF neural interest fields and improved walrus optimization algorithm can significantly improve the user's interaction depth and click-through conversion efficiency for different content channels, enhance the diversity and coherence of push sequences, and at the same time, the system response is stable and computing resources are controllable, which has the feasibility and practical value of large-scale deployment in real business scenarios.

[0183] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A big data personalized push method, characterized in that: The steps include: S1. Collect user behavior data and context information, perform semantic embedding, and obtain the user interest point set; S2. Based on the user interest point set, a neural network model with NeRF structure is used to construct a three-dimensional interest volume function and establish a neural interest field; S3. Obtain all content to be pushed and encode each content to be pushed into a content voxel set; S4. Establishing a penetration mapping relationship between the content voxel set and the neural field of interest, and determining a spatial action path of the content voxel on the three-dimensional volume of interest function; S5. Based on the penetration mapping relationship, an improved walrus optimization algorithm is used to perform content path search, where the content path is sequentially composed of multiple content voxel sets to generate a candidate content path set; S6. Perform volume rendering on the candidate content path set in the neural interest field to obtain a cumulative distribution curve of interest density and a direction change sequence of each candidate content path in the neural interest field; S7. Calculate the density penetration value of each candidate content path based on the interest density cumulative distribution curve, calculate the refraction consistency value of each candidate content path based on the direction change sequence, 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. A big data personalized push method according to claim 1, characterized in that: Said S1 specifically includes: S11. Collect user behavior data, including click behavior, browsing behavior, stay time, 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 including terminal type, operating system type, geographic location information, access time information, and network connection status, and construct context information features; S13. Sort the user behavior data in ascending order according to timestamps to construct a user behavior sequence, where the user behavior sequence consists of a structured triple containing a behavior type, a timestamp, and a content identifier; S14. Perform semantic encoding on the text content corresponding to the content identifier, use the BERT-Base pre-trained language model to encode the content title and body, and output semantic features; S15, performing feature splicing on the context information feature and the semantic feature to form a user behavior context joint feature; S16. Apply a density-based spatial clustering algorithm to the user behavior context joint features, use Euclidean distance as a metric function, and extract the centers of all clusters as a set of user interest points.

3. A big data personalized push method according to claim 1, characterized in that: The S2 specifically includes: S21, mapping each user interest point in the user interest point set to a three-dimensional Cartesian coordinate system to obtain a three-dimensional spatial position coordinate; S22. Calculate a unit direction vector corresponding to the position of each user interest point based on the spatial change direction of the consecutive user interest points; S23, constructing a neural network model of NeRF structure, wherein the neural network model of 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 frequency transformation of sine and cosine functions; The direction encoding module processes the input unit direction vector row with the same sine-cosine frequency transformation as the position encoding, and outputs a direction encoding feature vector of a preset length; The multilayer 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, wherein the first 7 layers input the position coding feature vector and output the intermediate position feature vector, the intermediate position feature vector and the direction coding feature vector are feature-concatenated and 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 a ReLU activation function to perform nonlinear 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 volume of interest function according to the neural radiation field volume rendering formula: ； ； in, Indicates the The cumulative interest value of the interest sampling path, Indicates the starting position of the sampling path of interest, Indicates the end position of the sampling path of interest, Indicates the Interest sampling paths at location The interest density value on Indicates from the starting position to the position The cumulative transmittance, represents the exponential function, Indicates the Interest sampling paths at location Interest density value on ; S25. Aggregate the integral calculation results of the three-dimensional interest volume function in the directions of multiple interest paths to construct a neural interest field to represent the spatial interest distribution of the user.

4. A big data personalized push method according to claim 1, characterized in that: The S3 specifically includes: S31. Acquire all content to be pushed, where the content to be pushed includes structured content and unstructured content, wherein the unstructured content includes text, images, videos, and voice data; S32. Analyze each content to be pushed, extract semantic principal components, tag information, and potential topic embedding, and construct a multimodal semantic feature vector; S33: Mapping the multimodal semantic feature vector to position coordinates in a three-dimensional content space, where the position coordinates constitute the spatial distribution center of the content to be pushed; S34. Based on the semantic intensity gradient of the content to be pushed and the expected push direction, construct multiple voxel units around the spatial distribution center, each voxel unit including 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 spatial interaction between the voxel unit and 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 axis and the user's historical preference direction; The refraction tolerance value is used to indicate the permissible degree of semantic deviation of the content after it is penetrated by the sampling path of interest in the direction of the voxel unit; S35 . Based on all voxel units corresponding to each content to be pushed, obtain a content voxel set of each content to be pushed.

5. A big data personalized push method according to claim 1, characterized in that: The S4 specifically includes: S41. For each voxel unit in the 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, inputting the voxel ray into the established neural interest field, performing volume sampling along each voxel ray in the neural interest field at a preset interval, recording the interest density value at the sampling point, and forming a corresponding interest density sequence; S43, calculating the cumulative response value of the voxel ray in the neural interest field according to the interest density sequence, wherein the cumulative response value is a weighted integral result of the interest density values ​​of all sampling points; S44. Setting interest density response threshold , the cumulative response value is compared with the refraction tolerance value of the corresponding voxel unit Compare, if the cumulative response value is greater than or equal to , then the voxel ray is determined to meet the penetration condition; S45. The voxel rays that meet the penetration condition are regarded as effective penetration paths, and are classified according to the relationship between the content voxels, so as to obtain the spatial action path of each content voxel set in the neural field of interest.

6. A big data personalized push method according to claim 1, characterized in that: The S5 specifically includes: S51. Based on the spatial action path of each content voxel set in the neural interest field, combine all 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 time sequence; S52: Allocate a set of encoding vectors to each content path individual, where the encoding vectors are used to represent the arrangement order and direction offset of each content voxel set in the content path; S53: defining a fitness function for the content path, wherein the fitness function is a weighted combination of the cumulative response value, content semantic relevance, and structural continuity of the content path; S54. The improvement of the improved walrus optimization algorithm is to introduce an adaptive leadership mechanism to update the encoding vector of the content path individual. The adaptive leadership mechanism dynamically adjusts the guidance direction according to the optimal content path individual in the current iteration: ； in, Indicates the Round The encoding vector of each content path, Indicates the Round The encoding vector of each content path, represents the convergence control factor, represents the disturbance factor, Indicates the Round The Euclidean distance between the individual content path and the optimal content path individual, Indicates the The encoding 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, Indicates the Diversity information entropy of content path groups in a round: ； in, represents the dimension of the encoding vector, Indicates the During the round of iteration The proportion of the most frequently occurring code value in each code position; S55, executing a preset number of rounds of iterative content path update process, and retaining several content path individuals with the highest fitness values ​​in each round; S56: After all iterations are completed, the content path individuals that are finally retained are output as a candidate content path set.

7. A big data personalized push method according to claim 1, characterized in that: S6 specifically includes: performing volume rendering on the candidate content path set in the neural interest field, and for each candidate content path, weightedly superimposing the interest density values ​​on all penetration rays according to the order of content voxels, to obtain a cumulative distribution curve of interest density of the candidate content path in the neural interest field, wherein the cumulative distribution curve of interest density 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 vector in the content voxel set, a sequence of direction angles between consecutive voxels in each candidate content path is calculated to generate a direction variation sequence for each candidate content path. The direction variation sequence is used to characterize the spatial refraction trend of the candidate content path.

8. A big data personalized push method according to claim 1, characterized in that: The S7 specifically includes: S71. Integrate the interest density cumulative distribution curve of each candidate content path, where the integral interval covers the sequentially arranged section of all content voxels in the path, and calculate the density penetration value of the candidate content path in the neural interest field: ； in, Indicates the density penetration value of the candidate content path. Indicates the cumulative distribution curve of interest density at the path position The integral value of the density of interest at Indicates the total length of the path; S72. Calculate the refraction consistency value of each candidate content path according to the direction change sequence: ； in, Refractive consistency value representing the candidate content path, represents the total number of content voxels in the candidate content path, Indicates the direction change sequence The value of the dimension; S73. Normalize the density penetration value and refraction consistency value of each candidate content path, and combine them according to preset weights to form a comprehensive score for each candidate content path; S74: Select the candidate content path with the largest comprehensive score as the optimal content path.

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

Images