Recommendation weight optimization method based on group feature aggregation

Abstract

The application relates to the technical field of information processing and the technical field of recommendation system in the computer technology field, and discloses a recommendation weight optimization method based on group feature aggregation. The method faces a video set of a recommendation system, first performs structured expression on various basic attributes of each video, and forms a video attribute matrix in a uniform order; then divides users into multiple groups with different viewing preferences, forms a group average preference vector according to historical viewing scores in the group, and obtains a group weighted attribute matrix by weighting the attribute matrix in the video dimension; subsequently, performs dimension normalization on each group matrix, calculates an attribute covariance matrix, performs eigenvalue decomposition, selects a main projection direction, constructs a main feature space, and obtains a video projection vector by projecting a standardized attribute; in the main space, a scoring rule is constructed according to the similarity between the projection vector and the main preference vector, a group candidate recommendation set is generated, and a personal list is aggregated and output.

Description

Technical Field

[0001] This invention relates to the field of information processing and recommendation system technology in the field of computer technology, specifically to a method for optimizing recommendation weights based on group feature aggregation. Background Technology

[0002] In today's internet environment, recommendation systems are widely used in various content distribution scenarios such as video, news, social media, and shopping. Among them, optimizing recommendation weights based on content attributes, user behavior, and group characteristics has become a core issue in achieving personalized and diversified content delivery. Existing technologies generally employ methods such as collaborative filtering, feature weighting, and deep learning. By modeling user interests, content characteristics, and user historical behavior, different weights are assigned to the objects to be recommended, thereby influencing the final recommendation list.

[0003] However, existing methods often suffer from several prominent problems. First, current mainstream recommendation systems often employ direct weighting or simple feature concatenation when processing video content with highly correlated multiple attributes, failing to effectively identify and eliminate redundant information between attributes. For example, video duration, type, and tags are often highly correlated; directly using them for weighted ranking can lead to some features being counted repeatedly, causing a shift in recommendation weights and reducing the diversity and accuracy of the recommendation system. Furthermore, with the rapid increase in the number of users and the types of features, existing methods often rely on pre-defined feature importance rankings or empirical parameter adjustments. This approach is highly subjective and struggles to adapt to changes in user behavior, easily leading to real-world problems such as cold start, overfitting, or decreased generalization ability. Second, in existing recommendation systems, user feature aggregation typically manifests as statistical analysis of the global mean, weighted sums, or principal component analysis for a specific feature. However, these methods have limited ability to handle intra-user differences, complex feature correlations, and extreme anomalies. Especially in highly heterogeneous and dynamically changing content ecosystems, existing technologies struggle to distinguish which features are highly redundant and which have genuine discriminative power, resulting in weight sequences generated by the recommendation system lacking sufficient information utilization and discriminative ability. Some methods attempt to introduce preprocessing techniques such as feature orthogonalization and normalization, but these often remain at the level of superficial mathematical operations, lacking an adaptive projection mechanism that incorporates group characteristics, making it difficult to truly remove redundant information after group aggregation. Furthermore, traditional personalized recommendation weight allocation mostly focuses on the evolution of individual users' interests or short-term click behavior. However, in real-world content distribution scenarios, complex interest coupling and information transmission exist between users, ignoring the deep feature structure of common group behaviors. Especially when facing large-scale user groups and multi-source attribute interactions, traditional techniques are significantly inadequate in modeling the true contribution of features, distinguishing sparsity and redundancy, and resisting extreme data interference. This makes it difficult for recommendation weight optimization to fundamentally improve the personalization accuracy and diversity coverage of recommended content.

[0004] Therefore, this study aims to propose a recommendation weight optimization method based on group feature aggregation, breaking through the limitations of traditional recommendation systems that rely solely on individual user behavior or simple collaborative filtering. By dividing users into multiple groups according to their viewing preferences, a group average preference vector is constructed and applied to the video attribute matrix, thus integrating group-level behavioral preferences at the model input layer. Subsequently, the weighted attributes are normalized, subjected to covariance analysis, feature decomposition, and projection to form a structured principal feature space. The similarity between the projected vector in the principal feature space and the group preference vector is then calculated to obtain a recommendation score. Finally, the candidate recommendation results from multiple groups are aggregated to achieve accurate recommendations for individual users. Summary of the Invention

[0005] This invention provides a recommendation weight optimization method based on group feature aggregation, which helps to solve the problems mentioned in the background art above.

[0006] This invention provides the following technical solution: a recommendation weight optimization method based on group feature aggregation, comprising:

[0007] Video attribute features are constructed for the video set in the recommendation system. Multiple basic attributes of each video are expressed in a structured manner, and the attribute values ​​of each video are arranged in a unified order to form a video attribute matrix.

[0008] The user set is divided into multiple user groups with different viewing preferences. A group average preference vector is constructed based on the historical viewing behavior ratings of users in each user group. Group weighting is then applied to the video attribute matrix in the video dimension to form a group weighted attribute matrix.

[0009] For each user group, the weighted attribute matrix is ​​weighted and the dimensions of each attribute are normalized to construct a standardized attribute matrix for the user group.

[0010] Based on the standardized attribute matrix, the covariance relationship between each attribute dimension is calculated to construct the attribute covariance matrix of the user group;

[0011] Eigenvalue decomposition is performed on the attribute covariance matrix of the user group, and a predetermined number of principal projection directions are selected to construct the principal feature space.

[0012] The standardized attributes of videos under the user group are projected onto the main feature space to obtain the projection vector of each video in the main feature space;

[0013] In the main feature space, a recommendation rating rule is constructed based on the main preference vector formed by the video projection vector and the average projection component calculated in the main feature space according to the main projection direction, and the recommendation rating value of each video under each user group is obtained.

[0014] Based on the video recommendation rating set of each user group, a corresponding candidate recommendation video set is generated, and the candidate recommendation video sets of multiple user groups are aggregated to form the final recommendation video output list for specific users.

[0015] Optionally, the step of constructing video attribute features for the video set in the recommendation system, structurally representing multiple basic attributes of each video, and arranging the attribute values ​​of each video in a unified order to form a video attribute matrix specifically includes:

[0016] Obtain all video objects in the recommendation system, assign a unique video number to each video, and form a sequence of video numbers;

[0017] Six basic attribute features are set for each video, namely: the total duration of the video; the average daily number of clicks by users within a set time span; the number of tags associated with the video; the number of days between the upload time and the current time; the average playback speed of the video during playback; and the cumulative number of interactions with the video throughout its entire lifecycle, including the number of likes, comments, and shares.

[0018] The six basic attribute features of each video are arranged in a fixed order to form a six-dimensional attribute column vector of the target video;

[0019] Arrange the attribute column vectors of all videos in order of video number to form a video attribute matrix with the number of videos as the number of rows and the number of six basic attribute features as the number of columns. Each row in the matrix contains the six basic attribute feature values ​​of one video.

[0020] Optionally, the step of dividing the user set into multiple user groups with different viewing preferences, constructing a group average preference vector based on the historical viewing behavior ratings of users within each user group, and performing group weighting processing on the video attribute matrix in the video dimension to form a group weighted attribute matrix specifically includes:

[0021] Based on business needs, the user set is divided into a preset number of user groups, which include at least a short video preference group, an in-depth content group, a hot topic focus group, and a niche exploration group, and each user is assigned one or more user group identifiers.

[0022] The viewing behavior of each user in the historical usage process is statistically analyzed, and a behavior score between zero and one is obtained for the target user's viewing result for each video. The behavior score represents the percentage of time the target user completes watching the target video.

[0023] For each user group, select at least one user in the user group, sum the behavioral scores of all users in the user group on the video dimension and divide by the number of users in the user group to obtain the average preference component of the user group for each video, and arrange the average preference components of all videos in order of video number to form the average preference vector of the user group.

[0024] The average preference vector of each user group is used as the weighting coefficient of the video dimension. The six basic attribute features of each video in the video attribute matrix are weighted to obtain the group weighted attribute matrix corresponding to the user group. The weighted attribute value of each video in each attribute dimension is equal to the product of the average preference component of the target video in the corresponding user group and the original attribute value of the target video in the corresponding attribute dimension.

[0025] Optionally, the step of performing dimensional normalization processing on each attribute dimension in the group-weighted attribute matrix corresponding to each user group to construct a standardized attribute matrix for the user group specifically includes:

[0026] For any user group, the weighted attribute matrix is ​​used to calculate the arithmetic mean of all weighted attribute values ​​in each attribute column according to the video dimension, thereby obtaining the weighted attribute mean of the user group in the corresponding attribute dimension.

[0027] For each attribute column, calculate the arithmetic mean of the squares of the differences between each weighted attribute value and the mean of the corresponding weighted attribute by video dimension, and perform a square root operation on the arithmetic mean of the squares of the differences between each weighted attribute value and the mean of the corresponding weighted attribute to obtain the weighted attribute standard deviation of the user group in the corresponding attribute dimension.

[0028] For each attribute column, the weighted attribute values ​​of each video under the user group are standardized: when the weighted attribute standard deviation of the selected attribute dimension is greater than zero, the weighted attribute mean of the selected attribute dimension is subtracted from the weighted attribute value of the target video in the selected attribute dimension, and then divided by the weighted attribute standard deviation of the selected attribute dimension to obtain the standardized attribute value of the target video in the selected attribute dimension; when the weighted attribute standard deviation of the selected attribute dimension is equal to zero, all standardized attribute values ​​of the corresponding attribute column are uniformly set to zero.

[0029] The standardized attribute values ​​of all attribute columns under the user group are arranged in a fixed order according to the video number and attribute dimension to form a standardized attribute matrix of the user group. The standardized attribute matrix corresponds to the video in the row dimension and to the six attribute dimensions in the column dimension.

[0030] Optionally, the step of calculating the covariance relationship between each attribute dimension based on the standardized attribute matrix to construct the attribute covariance matrix of the user group specifically includes:

[0031] For any user group, the standardized attribute matrix is ​​used to multiply and sum the standardized attribute values ​​of each video under the corresponding user group in these two attribute dimensions according to the video dimension, and then divide by the number of videos to obtain the covariance value between the corresponding user group in these two attribute dimensions.

[0032] Arrange the covariance values ​​corresponding to the pairwise combinations of all attribute dimensions according to the row dimension index and column dimension index to construct the attribute covariance matrix for the corresponding user group. The attribute covariance matrix is ​​a square matrix, where each row and each column corresponds to an attribute dimension, and each matrix element corresponds to a pair of attribute dimensions and their covariance relationship.

[0033] Optionally, the step of performing eigenvalue decomposition on the attribute covariance matrix of the user group, selecting a predetermined number of principal projection directions, and constructing a principal feature space specifically includes:

[0034] Eigenvalue decomposition is performed on the attribute covariance matrix corresponding to any user group to obtain multiple eigenvalues ​​of the attribute covariance matrix and eigenvectors corresponding to each eigenvalue. Each component of the eigenvector corresponds to an attribute dimension.

[0035] When all elements of the attribute covariance matrix of the corresponding user group are zero, the three unit direction vectors that are parallel to the first attribute axis, the second attribute axis and the third attribute axis respectively are taken as the three main projection directions of the corresponding user group, and arranged in the order of the attribute axes to form the main projection direction matrix.

[0036] When there are non-zero elements in the attribute covariance matrix of the corresponding user group, all eigenvalues ​​are sorted from largest to smallest. The three eigenvectors corresponding to the three largest eigenvalues ​​are selected as the three principal projection directions of the corresponding user group. When there are eigenvalues ​​with equal values, the corresponding eigenvectors are selected in the column order of the eigenvector matrix. The three principal projection directions are arranged in the selection order to form the principal projection direction matrix of the corresponding user group.

[0037] Optionally, the step of projecting the standardized attributes of videos within a user group onto the main feature space to obtain the projection vector of each video in the main feature space specifically includes:

[0038] For any user group, the standardized attribute matrix is ​​used to arrange the standardized attribute values ​​of the target video in a fixed order of the six attribute dimensions, forming a standardized attribute column vector of the target video.

[0039] The standardized attribute column vector of the target video is subjected to inner product operation with the three main projection directions in the main projection direction matrix of the corresponding user group one by one to obtain the projection component of the target video in each main projection direction. The three projection components are arranged in order of the main projection direction to form the projection vector of the target video in the main feature space.

[0040] The projection vectors of all videos under the corresponding user group are arranged in order according to the video number to form the video projection matrix of the corresponding user group. The video projection matrix corresponds to the video number in the row dimension and the projection components in the three main projection directions in the column dimension.

[0041] Optionally, the step of constructing a recommendation rating rule based on the main preference vector formed by the video projection vector and the average projection component calculated along the main projection direction in the main feature space, and obtaining the recommendation rating value for each video under each user group, specifically includes:

[0042] For any user group, calculate the arithmetic mean of all video projection components in each column according to the video dimension, and arrange the average values ​​in the three main projection directions in a fixed order to form the main preference vector of the corresponding user group in the main feature space.

[0043] For each video in the corresponding user group, calculate the cosine similarity between the projection vector of the target video and the main preference vector: if both the projection vector and the main preference vector of the target video are zero vectors, set the cosine similarity to zero; if both the projection vector and the main preference vector of the target video are non-zero vectors, perform the inner product operation and divide the result by the product of the lengths of the two vectors to obtain the recommendation rating value of the target video in the corresponding user group.

[0044] The recommended rating values ​​of all videos under the corresponding user group are recorded sequentially according to the video number, forming a set of recommended video ratings for the corresponding user group.

[0045] Optionally, the step of generating a corresponding candidate recommended video set based on the video recommendation rating set of each user group, and performing aggregation processing on the candidate recommended video sets of multiple user groups to form a final recommended video output list for a specific user, specifically includes:

[0046] For any set of video recommendation ratings for a user group, sort all videos under the corresponding user group from high to low according to the recommendation rating value. If there are videos with the same recommendation rating value, sort them in ascending order of video number.

[0047] Select several videos with the highest recommendation scores from the ranking results to form a candidate recommended video set for the corresponding user group. Each element in the candidate recommended video set is a video number.

[0048] For each user, based on one or more user groups to which the target user belongs, obtain the candidate recommended video set corresponding to each user group, and perform a union operation on these candidate recommended video sets to obtain the merged candidate recommended video set for the target user;

[0049] For each video in the merged candidate recommended video set, if the target video exists in the candidate recommended video sets of two or more user groups, the maximum value of the recommendation score of the target video in each user group is obtained as the final individual recommendation score of the target user for the target video; if the target video exists only in the candidate recommended video set of a single user group, the recommendation score of the target user in the corresponding user group is used as the final individual recommendation score.

[0050] Sort all videos in the merged candidate recommendation video set from highest to lowest according to their final individual recommendation score. If there are videos with the same final individual recommendation score, sort them in ascending order of video number. Output the final recommendation video list for the target user.

[0051] The present invention has the following beneficial effects:

[0052] 1. This solution systematically defines six basic video attributes (duration, average daily clicks, number of tags, upload duration, playback speed, and total interactions) in the recommendation system, and maps each video to a six-dimensional attribute vector in a fixed order, thereby constructing an attribute matrix. Compared with existing practices that only use a few explicit features or require manual attribute selection, this solution ensures a comprehensive portrayal of video content and user behavior characteristics by covering multi-dimensional, full-lifecycle attributes. Furthermore, the structured representation and matrix storage of the attribute matrix facilitate parallel processing, exhibiting good scalability and computational efficiency for thousands to tens of thousands of videos.

[0053] 2. This solution divides all users into several preference groups based on business scenarios, such as short video preference and in-depth content preference. Within each group, the historical behavior scores of members are averaged to form a group average preference vector. This vector is then used as a weighting coefficient for the video attribute matrix to obtain a group weighted attribute matrix. Unlike traditional collaborative filtering based on overall user behavior or single user behavior, this solution distinguishes multiple viewing preferences, reflecting the true preferences of each group for different video types while reducing the negative impact of single-user rating noise on recommendation performance. The weighted attribute matrix directly injects group preferences into the feature layer, making subsequent feature engineering and similarity calculation more targeted and interpretable. Balancing group collaboration and personalization needs, it satisfies common group preferences while preserving diversity during the aggregation stage, addressing the shortcomings of existing technologies such as "monopolization of popular videos" and "difficulty in recommending niche videos."

[0054] 3. This scheme calculates the mean and standard deviation of the weighted attribute matrix for each group according to each attribute dimension, and standardizes each element using the formula "(value – mean) / standard deviation". If the standard deviation is zero, it is assigned zero, thus constructing a standardized attribute matrix. Compared with schemes that only perform simple normalization (such as min – max), this method eliminates the differences in the dimensions and distributions of each attribute through standard deviation scaling, making the influence of different attributes on subsequent covariance and eigenvalue decomposition balanced and stable. In addition, assigning zero to the case where the standard deviation is zero avoids division by zero errors. Through statistical standardization, the numerical stability and outlier resistance of the matrix are enhanced, providing a reliable data foundation for high-quality covariance analysis and projection vector calculation, and solving the common numerical distortion problem when large-scale attribute values ​​have large differences and severe distribution skewness.

[0055] 4. This solution calculates the covariance between each pair of six-dimensional attributes based on a standardized matrix, generating a group attribute covariance matrix that accurately characterizes the correlation structure between attributes. Unlike traditional methods that only focus on the distribution of features themselves or simple correlation analysis, the covariance matrix provides global and systematic information on the coupling between attributes, capturing the comprehensive impact of complex attribute combinations on video performance; this is particularly crucial for subsequent feature decomposition. Systematically incorporating "attribute correlation" into the recommendation model allows the model to retain the direction with the most information and the most significant changes when selecting the principal projection direction, while also improving the ability to identify the interaction effects of attributes. This solves the recommendation accuracy bottleneck caused by the disorder of attribute dimensions or information loss in existing technologies.

[0056] 5. This scheme performs eigenvalue decomposition on the covariance matrix, extracting the most significant eigenvectors as the principal projection directions to construct the principal feature space. Unlike approaches that only select the top three largest directions or fixed dimensions, this scheme provides backup unit vector schemes for all-zero matrices and strictly selects eigenvalues ​​according to vector index order to ensure algorithm consistency when using equal eigenvalues. For non-all-zero matrices, it strictly selects the directions with the most information to ensure the interpretability and discriminative power of the projection space for group preferences. Through rigorous mathematical decomposition and backup scheme design, this scheme ensures the stable generation of an effective feature space regardless of whether the data meets a typical distribution, while maximizing the preservation of collaborative information between attributes and solving the recommendation interference problem caused by feature redundancy or unstable reconstruction.

[0057] 6. This scheme orthogonally projects the six-dimensional standardized attribute vectors of each video along the principal projection direction matrix to obtain a three-dimensional projection vector. Unlike simple dimensionality reduction methods (such as PCA truncation), this scheme explicitly constructs the projection direction as the most representative direction extracted by eigenvalue decomposition, and orthogonal projection ensures the orthogonality between vectors, improving the interpretability of the decomposition. By compressing high-dimensional attributes into a low-dimensional space through projection, not only is the computational cost reduced, but subsequent similarity calculations are also based on a more discriminative and stable vector representation, solving the problems of high-dimensional noise and the curse of dimensionality.

[0058] 7. This scheme calculates the similarity (e.g., cosine similarity) between the group's main preference vector (derived from the mean of the projected vectors) and the projection vectors of each video within the main feature space to obtain the recommendation rating. Unlike traditional similarity calculations based on the user-item matrix product or distance, this scheme directly establishes the rating function in the feature space, resulting in ratings that better reflect group preferences and attribute structures. Furthermore, it avoids computational anomalies caused by empty vectors by segmenting the zero-vector case. This achieves high interpretability of the rating rules—each rating can be traced back to specific attributes and the main direction of group preferences, making the recommendation model transparent and easy to optimize, thus addressing the pain points of black-box models being difficult to interpret and difficult to manually intervene in.

[0059] 8. This solution sorts the recommended rating sets for each group, extracts the top few to generate a candidate recommended video set, and then performs a union operation on the multiple group sets to which the user belongs, taking the maximum score to form a personalized recommendation list. Unlike aggregation strategies that only consider a single group or a simple weighted average, this solution preserves the diversity of each group through the union operation, and takes the maximum score to emphasize the user's most likely preferences, balancing recommendation accuracy and diversity. It aligns with the diverse interests of users across groups, satisfying their diverse needs across different preferences while maximizing the highlighting of content the user is most likely to be interested in through ratings. This solves the problem that single aggregation often leads to a "one-sided" recommendation or reduced recommendation effectiveness when preferences conflict. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example, refer to Figure 1 Recommendation weight optimization methods based on group feature aggregation include:

[0063] Video attribute features are constructed for the video set in the recommendation system. Multiple basic attributes of each video are expressed in a structured manner, and the attribute values ​​of each video are arranged in a unified order to form a video attribute matrix.

[0064] The user set is divided into multiple user groups with different viewing preferences. A group average preference vector is constructed based on the historical viewing behavior ratings of users in each user group. Group weighting is then applied to the video attribute matrix in the video dimension to form a group weighted attribute matrix.

[0065] For each user group, the weighted attribute matrix is ​​weighted and the dimensions of each attribute are normalized to construct a standardized attribute matrix for the user group.

[0066] Based on the standardized attribute matrix, the covariance relationship between each attribute dimension is calculated to construct the attribute covariance matrix of the user group;

[0067] Eigenvalue decomposition is performed on the attribute covariance matrix of the user group, and a predetermined number of principal projection directions are selected to construct the principal feature space.

[0068] The standardized attributes of videos under the user group are projected onto the main feature space to obtain the projection vector of each video in the main feature space;

[0069] In the main feature space, a recommendation rating rule is constructed based on the main preference vector formed by the video projection vector and the average projection component calculated in the main feature space according to the main projection direction, and the recommendation rating value of each video under each user group is obtained.

[0070] Based on the video recommendation rating set of each user group, a corresponding candidate recommendation video set is generated, and the candidate recommendation video sets of multiple user groups are aggregated to form the final recommendation video output list for specific users.

[0071] By sequentially executing a complete process including video attribute matrix construction, user group segmentation and weighting, attribute matrix standardization, covariance analysis, eigenvalue decomposition, projection vector acquisition, and similarity calculation and candidate set aggregation based on projection vectors and group main preference vectors, this method overcomes the problems of existing recommendation systems that rely solely on single-user historical behavior or simple collaborative filtering and are susceptible to data sparsity and preference distortion. First, by structurally representing the multidimensional attributes of videos and constructing a unified matrix, standardized input is provided for feature fusion. Second, users are divided into multiple groups according to viewing preferences and group-level preferences are injected, effectively utilizing group synergy and addressing the pain point of unstable recommendation performance when single-user data is scarce. Then, standardization and covariance matrix characterize the relationships between attributes, overcoming the interference of different attribute scales and distribution differences on the model. Subsequently, eigenvalue decomposition and orthogonal projection are used to compress high-dimensional attribute information into a low-dimensional space, reducing computational complexity while preserving key structures to the greatest extent. Finally, the aggregation strategy of unifying candidate sets from multiple groups and taking the maximum score ensures both recommendation accuracy and diversity. Overall, this method improves the stability, accuracy, and diversity of recommendation systems through complete multi-level feature engineering and group collaborative fusion, and effectively solves the problems of data sparsity, single preference, and monotonous recommendations in existing technologies.

[0072] The process of constructing video attribute features for the video set in the recommendation system, structurally representing multiple basic attributes of each video, and arranging the attribute values ​​of each video in a unified order to form a video attribute matrix specifically includes:

[0073] Obtain all video objects in the recommendation system, assign a unique video number to each video, and form a sequence of video numbers;

[0074] Six basic attribute features are set for each video, namely: the total duration of the video; the average daily number of clicks by users within a set time span; the number of tags associated with the video; the number of days between the upload time and the current time; the average playback speed of the video during playback; and the cumulative number of interactions with the video throughout its entire lifecycle, including the number of likes, comments, and shares.

[0075] The six basic attribute features of each video are arranged in a fixed order to form a six-dimensional attribute column vector of the target video;

[0076] Arrange the attribute column vectors of all videos in order of video number to form a video attribute matrix with the number of videos as the number of rows and the number of six basic attribute features as the number of columns. Each row in the matrix contains the six basic attribute feature values ​​of one video.

[0077] Further specific implementation steps include:

[0078] The total number of videos in the recommendation system is denoted as . , No. The video is recorded as ;in, For video indexing; For the first Video object;

[0079] Each video Six basic attribute features are defined as follows: , , , , , ;in, For the first The total duration of the videos, in seconds; For the first Average daily user views per video, in units of ; For the first The number of tags for each video, expressed in units of individual tags; For the first The upload time of each video, in days; For the first The average playback speed of the videos, expressed as a dimensionless ratio; For the first The total number of interactions with a video, including likes, comments, and shares, measured in times;

[0080] Construct the attribute vector for each video: ;in, For the first A 6-dimensional attribute column vector for each video; Indicates all The set of real number column vectors;

[0081] Construct the attribute matrix for all videos , No. Behavior ;in, Indicates all The set of real number matrices.

[0082] The process involves dividing the user set into multiple user groups with different viewing preferences, constructing a group average preference vector based on the historical viewing behavior ratings of users within each user group, and performing group weighting processing on the video attribute matrix along the video dimension to form a group weighted attribute matrix. Specifically, this includes:

[0083] Based on business needs, the user set is divided into a preset number of user groups, which include at least a short video preference group, an in-depth content group, a hot topic focus group, and a niche exploration group, and each user is assigned one or more user group identifiers.

[0084] The viewing behavior of each user in the historical usage process is statistically analyzed, and a behavior score between zero and one is obtained for the target user's viewing result for each video. The behavior score represents the percentage of time the target user completes watching the target video.

[0085] For each user group, select at least one user in the user group, sum the behavioral scores of all users in the user group on the video dimension and divide by the number of users in the user group to obtain the average preference component of the user group for each video, and arrange the average preference components of all videos in order of video number to form the average preference vector of the user group.

[0086] The average preference vector of each user group is used as the weighting coefficient of the video dimension. The six basic attribute features of each video in the video attribute matrix are weighted to obtain the group weighted attribute matrix corresponding to the user group. The weighted attribute value of each video in each attribute dimension is equal to the product of the average preference component of the target video in the corresponding user group and the original attribute value of the target video in the corresponding attribute dimension.

[0087] Further specific implementation steps include:

[0088] Divide users into Each group is denoted as [group name]. , ;in, The total number of people in the group; For the first A user group; For group indexing, Indicates short video preference groups, Indicates the in-depth content group, Indicates the focus group, This refers to the niche exploration group;

[0089] Each user For each video Has one historical behavior score ;in, Index for users; For users For video The historical behavior score indicates the percentage of videos that were completed and watched.

[0090] For each group Given at least one user, calculate their group average preference vector for all videos:

[0091] , ;in, For set The cardinality; For the group The group's average preference vector; For the group For video The average preference component;

[0092] Using the group average preference vector As a property matrix Video dimensions are weighted to construct a group-weighted attribute matrix. Specifically:

[0093] , , , ;in, For elements Composition 3D matrix; For the first The video is in the Weighted attribute values ​​across each attribute dimension; For attribute indexing; For the first The first video One original attribute value;

[0094] Equivalently, it can be written in matrix multiplication form:

[0095] ;in, For the group The weighted attribute matrix; For diagonalization operations; vector Convert to a diagonal matrix.

[0096] The step of performing dimensional normalization on each attribute dimension in the group-weighted attribute matrix corresponding to each user group to construct a standardized attribute matrix for the user group specifically includes:

[0097] For any user group, the weighted attribute matrix is ​​used to calculate the arithmetic mean of all weighted attribute values ​​in each attribute column according to the video dimension, thereby obtaining the weighted attribute mean of the user group in the corresponding attribute dimension.

[0098] For each attribute column, calculate the arithmetic mean of the squares of the differences between each weighted attribute value and the mean of the corresponding weighted attribute by video dimension, and perform a square root operation on the arithmetic mean of the squares of the differences between each weighted attribute value and the mean of the corresponding weighted attribute to obtain the weighted attribute standard deviation of the user group in the corresponding attribute dimension.

[0099] For each attribute column, the weighted attribute values ​​of each video under the user group are standardized: when the weighted attribute standard deviation of the selected attribute dimension is greater than zero, the weighted attribute mean of the selected attribute dimension is subtracted from the weighted attribute value of the target video in the selected attribute dimension, and then divided by the weighted attribute standard deviation of the selected attribute dimension to obtain the standardized attribute value of the target video in the selected attribute dimension; when the weighted attribute standard deviation of the selected attribute dimension is equal to zero, all standardized attribute values ​​of the corresponding attribute column are uniformly set to zero.

[0100] The standardized attribute values ​​of all attribute columns under the user group are arranged in a fixed order according to the video number and attribute dimension to form a standardized attribute matrix of the user group. The standardized attribute matrix corresponds to the video in the row dimension and to the six attribute dimensions in the column dimension.

[0101] Further specific implementation steps include:

[0102] For the matrix of the th The Calculate the mean and standard deviation for each column:

[0103] , ;in, For the group Below, weighted attribute matrix In the The mean of the column; For the group Below, weighted attribute matrix In the The standard deviation of the column;

[0104] Constructing groups The normalized matrix below The elements of the standardized matrix are:

[0105] ;in, For the group Next video number The standardized value of each attribute.

[0106] The step of calculating the covariance relationship between each attribute dimension based on the standardized attribute matrix to construct the attribute covariance matrix of the user group specifically includes:

[0107] For any user group, the standardized attribute matrix is ​​used to multiply and sum the standardized attribute values ​​of each video under the corresponding user group in these two attribute dimensions according to the video dimension, and then divide by the number of videos to obtain the covariance value between the corresponding user group in these two attribute dimensions.

[0108] Arrange the covariance values ​​corresponding to the pairwise combinations of all attribute dimensions according to the row dimension index and column dimension index to construct the attribute covariance matrix for the corresponding user group. The attribute covariance matrix is ​​a square matrix, where each row and each column corresponds to an attribute dimension, and each matrix element corresponds to a pair of attribute dimensions and their covariance relationship.

[0109] Further specific implementation steps include:

[0110] Calculate the normalized matrix covariance matrix The elements of the covariance matrix are:

[0111] , ;in, For the group The covariance matrix of the standardized attributes; The first covariance matrix is ​​the first... Line 1 Column elements; , These are the row and column indices of the covariance matrix, respectively. For the first The video is in the Standardized values ​​for each attribute dimension; For the first The video is in the The standardized values ​​of each attribute dimension.

[0112] The process of performing eigenvalue decomposition on the attribute covariance matrix of the user group, selecting a predetermined number of principal projection directions, and constructing the principal feature space specifically includes:

[0113] Eigenvalue decomposition is performed on the attribute covariance matrix corresponding to any user group to obtain multiple eigenvalues ​​of the attribute covariance matrix and eigenvectors corresponding to each eigenvalue. Each component of the eigenvector corresponds to an attribute dimension.

[0114] When all elements of the attribute covariance matrix of the corresponding user group are zero, the three unit direction vectors that are parallel to the first attribute axis, the second attribute axis and the third attribute axis respectively are taken as the three main projection directions of the corresponding user group, and arranged in the order of the attribute axes to form the main projection direction matrix.

[0115] When there are non-zero elements in the attribute covariance matrix of the corresponding user group, all eigenvalues ​​are sorted from largest to smallest. The three eigenvectors corresponding to the three largest eigenvalues ​​are selected as the three principal projection directions of the corresponding user group. When there are eigenvalues ​​with equal values, the corresponding eigenvectors are selected in the column order of the eigenvector matrix. The three principal projection directions are arranged in the selection order to form the principal projection direction matrix of the corresponding user group.

[0116] Further specific implementation steps include:

[0117] For covariance matrix Eigenvalue decomposition is performed as follows:

[0118] ;in, For the group The eigenvector matrix below has column vectors as follows: ; For the group Next 1 eigenvector; The feature vector index; For the group The eigenvalue diagonal matrix below has diagonal elements as follows: ; For the group Next One eigenvalue;

[0119] When it appears season ;in, for A zero matrix; For the group The selected principal projection direction matrix has dimensions of ;

[0120] The unit direction of the first attribute axis, corresponding to the attribute. ; The unit direction of the second attribute axis, corresponding to the attribute. ; The unit direction of the third attribute axis, corresponding to the attribute. ;

[0121] when When selecting the directions corresponding to the first three largest eigenvalues, construct the main projection matrix: When equal eigenvalues ​​exist, the eigenvectors are used to determine the eigenvalues. The column indexes are selected in ascending order; among them... These are the eigenvectors corresponding to the first three eigenvalues.

[0122] The step of projecting the standardized attributes of videos within a user group onto the main feature space to obtain the projection vector of each video in the main feature space specifically includes:

[0123] For any user group, the standardized attribute matrix is ​​used to arrange the standardized attribute values ​​of the target video in a fixed order of the six attribute dimensions, forming a standardized attribute column vector of the target video.

[0124] The standardized attribute column vector of the target video is subjected to inner product operation with the three main projection directions in the main projection direction matrix of the corresponding user group one by one to obtain the projection component of the target video in each main projection direction. The three projection components are arranged in order of the main projection direction to form the projection vector of the target video in the main feature space.

[0125] The projection vectors of all videos under the corresponding user group are arranged in order according to the video number to form the video projection matrix of the corresponding user group. The video projection matrix corresponds to the video number in the row dimension and the projection components in the three main projection directions in the column dimension.

[0126] Further specific implementation steps include:

[0127] For each video Its standardized attribute vector is:

[0128] ;in, For the group Next A 6-dimensional normalized attribute column vector of each video; for The Line 1 Column elements;

[0129] Perform orthogonal projection along the principal direction, specifically as follows:

[0130] ;in, For the group Next The projected coordinates of the video in the main direction space; for The set of column vectors of real numbers;

[0131] Combine the master projection results of all videos to construct , No. Behavior ;in, For the group The set matrix of all video projection vectors, with dimensions of .

[0132] In the main feature space, a recommendation scoring rule is constructed based on the main preference vector formed by the video projection vector and the average projection component calculated along the main projection direction in the main feature space. This process obtains the recommendation score value for each video under each user group, specifically including:

[0133] For any user group, calculate the arithmetic mean of all video projection components in each column according to the video dimension, and arrange the average values ​​in the three main projection directions in a fixed order to form the main preference vector of the corresponding user group in the main feature space.

[0134] For each video in the corresponding user group, calculate the cosine similarity between the projection vector of the target video and the main preference vector: if both the projection vector and the main preference vector of the target video are zero vectors, set the cosine similarity to zero; if both the projection vector and the main preference vector of the target video are non-zero vectors, perform the inner product operation and divide the result by the product of the lengths of the two vectors to obtain the recommendation rating value of the target video in the corresponding user group.

[0135] The recommended rating values ​​of all videos under the corresponding user group are recorded sequentially according to the video number, forming a set of recommended video ratings for the corresponding user group.

[0136] Further specific implementation steps include:

[0137] The principal preference vector is obtained by averaging all projection vectors.

[0138] ;in, For the group The principal preference vector in the 3D projection space;

[0139] A cosine similarity scoring function is constructed, and the score for each video is calculated in the following segmented manner:

[0140] ;in, For the group For the first Recommended ratings for each video; It is the Euclidean norm.

[0141] The process of generating a corresponding candidate recommended video set based on the video recommendation rating set of each user group, and performing aggregation processing on the candidate recommended video sets of multiple user groups to form a final recommended video output list for a specific user, specifically includes:

[0142] For any set of video recommendation ratings for a user group, sort all videos under the corresponding user group from high to low according to the recommendation rating value. If there are videos with the same recommendation rating value, sort them in ascending order of video number.

[0143] Select several videos with the highest recommendation scores from the ranking results to form a candidate recommended video set for the corresponding user group. Each element in the candidate recommended video set is a video number.

[0144] For each user, based on one or more user groups to which the target user belongs, obtain the candidate recommended video set corresponding to each user group, and perform a union operation on these candidate recommended video sets to obtain the merged candidate recommended video set for the target user;

[0145] For each video in the merged candidate recommended video set, if the target video exists in the candidate recommended video sets of two or more user groups, the maximum value of the recommendation score of the target video in each user group is obtained as the final individual recommendation score of the target user for the target video; if the target video exists only in the candidate recommended video set of a single user group, the recommendation score of the target user in the corresponding user group is used as the final individual recommendation score.

[0146] Sort all videos in the merged candidate recommendation video set from highest to lowest according to their final individual recommendation score. If there are videos with the same final individual recommendation score, sort them in ascending order of video number. Output the final recommendation video list for the target user.

[0147] Further specific implementation steps include:

[0148] By rating For all videos Sort in descending order;

[0149] Extract the top-rated The recommended collection consists of 10 videos:

[0150] When there are tied scores, the video index is used. Sort and truncate according to ascending order criterion; among which... The number of candidate recommendations output for each group; For the group A collection of recommended video indexes;

[0151] If user If a user belongs to multiple groups, their recommendation set is as follows: ;in, For users The set of indexes of the group to which it belongs, for example ; For user-oriented The final set of candidate recommended video indexes; Union operator;

[0152] For sets For each video in the series, the final individual rating is: ;in, For users For video The final individual score;

[0153] according to Sort the results in descending order and output the final recommendation list for each user; when there are ties in scores, sort by video index. Sort by ascending order.

[0154] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0155] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing recommendation weights based on population feature aggregation, characterized in that, include: For the video set in the recommendation system, video attribute features are constructed. Multiple basic attributes of each video are structurally represented, and the attribute values ​​of each video are arranged in a unified order to form a video attribute matrix. Specifically, this includes: Obtain all video objects in the recommendation system, assign a unique video number to each video, and form a sequence of video numbers; Six basic attribute features are set for each video, namely: the total duration of the video; the average daily number of clicks by users within a set time span; the number of tags associated with the video; the number of days between the upload time and the current time; the average playback speed of the video during playback; and the cumulative number of interactions with the video throughout its entire lifecycle, including the number of likes, comments, and shares. The six basic attribute features of each video are arranged in a fixed order to form a six-dimensional attribute column vector of the target video; Arrange the attribute column vectors of all videos in order of video number to form a video attribute matrix with the number of videos as the number of rows and the number of six basic attribute features as the number of columns. Each row in the matrix contains the six basic attribute feature values ​​of one video. The user set is divided into multiple user groups with different viewing preferences. A group average preference vector is constructed based on the historical viewing behavior ratings of users within each user group. Furthermore, a group-weighted attribute matrix is ​​applied to the video attribute matrix along the video dimension, forming a group-weighted attribute matrix, which specifically includes: Based on business needs, the user set is divided into a preset number of user groups, which include at least a short video preference group, an in-depth content group, a hot topic focus group, and a niche exploration group, and each user is assigned one or more user group identifiers. The viewing behavior of each user in the historical usage process is statistically analyzed, and a behavior score between zero and one is obtained for the target user's viewing result for each video. The behavior score represents the percentage of time the target user completes watching the target video. For each user group, select at least one user in the user group, sum the behavioral scores of all users in the user group on the video dimension and divide by the number of users in the user group to obtain the average preference component of the user group for each video, and arrange the average preference components of all videos in order of video number to form the average preference vector of the user group. The average preference vector of each user group is used as the weighting coefficient of the video dimension. The six basic attribute features of each video in the video attribute matrix are weighted to obtain the group weighted attribute matrix corresponding to the user group. The weighted attribute value of each video in each attribute dimension is equal to the product of the average preference component of the target video in the corresponding user group and the original attribute value of the target video in the corresponding attribute dimension. For each user group, the weighted attribute matrix is ​​used to perform dimensional normalization on each attribute dimension, thus constructing a standardized attribute matrix for the user group. Specifically, this includes: For any user group, the weighted attribute matrix is ​​used to calculate the arithmetic mean of all weighted attribute values ​​in each attribute column according to the video dimension, thereby obtaining the weighted attribute mean of the user group in the corresponding attribute dimension. For each attribute column, calculate the arithmetic mean of the squares of the differences between each weighted attribute value and the mean of the corresponding weighted attribute by video dimension, and perform a square root operation on the arithmetic mean of the squares of the differences between each weighted attribute value and the mean of the corresponding weighted attribute to obtain the weighted attribute standard deviation of the user group in the corresponding attribute dimension. For each attribute column, the weighted attribute values ​​of each video under the user group are standardized: when the weighted attribute standard deviation of the selected attribute dimension is greater than zero, the weighted attribute mean of the selected attribute dimension is subtracted from the weighted attribute value of the target video in the selected attribute dimension, and then divided by the weighted attribute standard deviation of the selected attribute dimension to obtain the standardized attribute value of the target video in the selected attribute dimension; when the weighted attribute standard deviation of the selected attribute dimension is equal to zero, all standardized attribute values ​​in the corresponding attribute column are uniformly set to zero. The standardized attribute values ​​of all attribute columns under the user group are arranged in a fixed order according to the video number and attribute dimension to form a standardized attribute matrix of the user group. The standardized attribute matrix corresponds to the video in the row dimension and to the six attribute dimensions in the column dimension. Based on the standardized attribute matrix, the covariance relationship between each attribute dimension is calculated to construct the attribute covariance matrix of the user group, specifically including: For any user group, the standardized attribute matrix is ​​used to multiply and sum the standardized attribute values ​​of each video under the corresponding user group in these two attribute dimensions according to the video dimension, and then divide by the number of videos to obtain the covariance value between the corresponding user group in these two attribute dimensions. Arrange the covariance values ​​corresponding to the pairwise combinations of all attribute dimensions according to the row dimension index and column dimension index to construct the attribute covariance matrix for the corresponding user group. The attribute covariance matrix is ​​a square matrix, where each row and each column corresponds to an attribute dimension, and each matrix element corresponds to a pair of attribute dimensions and their covariance relationship. Eigenvalue decomposition is performed on the attribute covariance matrix of the user group, and a predetermined number of principal projection directions are selected to construct the principal feature space, specifically including: Eigenvalue decomposition is performed on the attribute covariance matrix corresponding to any user group to obtain multiple eigenvalues ​​of the attribute covariance matrix and eigenvectors corresponding to each eigenvalue. Each component of the eigenvector corresponds to an attribute dimension. When all elements of the attribute covariance matrix of the corresponding user group are zero, the three unit direction vectors that are parallel to the first attribute axis, the second attribute axis and the third attribute axis respectively are taken as the three main projection directions of the corresponding user group, and arranged in the order of the attribute axes to form the main projection direction matrix. When there are non-zero elements in the attribute covariance matrix of the corresponding user group, all eigenvalues ​​are sorted from largest to smallest. The three eigenvectors corresponding to the three largest eigenvalues ​​are selected as the three main projection directions of the corresponding user group. When there are eigenvalues ​​with equal values, the corresponding eigenvectors are selected in the column order of the eigenvector matrix. The three main projection directions are arranged in the selection order to form the main projection direction matrix of the corresponding user group. The standardized attributes of videos under the user group are projected onto the main feature space to obtain the projection vector of each video in the main feature space; In the main feature space, a recommendation rating rule is constructed based on the main preference vector formed by the video projection vector and the average projection component calculated in the main feature space according to the main projection direction, and the recommendation rating value of each video under each user group is obtained. Based on the video recommendation rating set of each user group, a corresponding candidate recommendation video set is generated, and the candidate recommendation video sets of multiple user groups are aggregated to form the final recommendation video output list for specific users. 2.The method of claim 1, wherein, The step of projecting the standardized attributes of videos within a user group onto the main feature space to obtain the projection vector of each video in the main feature space specifically includes: For any user group, the standardized attribute matrix is ​​used to arrange the standardized attribute values ​​of the target video in a fixed order of the six attribute dimensions, forming a standardized attribute column vector of the target video. The standardized attribute column vector of the target video is subjected to inner product operation with the three main projection directions in the main projection direction matrix of the corresponding user group one by one to obtain the projection component of the target video in each main projection direction. The three projection components are arranged in order of the main projection direction to form the projection vector of the target video in the main feature space. The projection vectors of all videos under the corresponding user group are arranged in order according to the video number to form the video projection matrix of the corresponding user group. The video projection matrix corresponds to the video number in the row dimension and the projection components in the three main projection directions in the column dimension.

3. The recommendation weight optimization method based on group feature aggregation according to claim 2, characterized in that, In the main feature space, a recommendation scoring rule is constructed based on the main preference vector formed by the video projection vector and the average projection component calculated along the main projection direction in the main feature space. This process obtains the recommendation score value for each video under each user group, specifically including: For any user group, calculate the arithmetic mean of all video projection components in each column according to the video dimension, and arrange the average values ​​in the three main projection directions in a fixed order to form the main preference vector of the corresponding user group in the main feature space. For each video in the corresponding user group, calculate the cosine similarity between the projection vector of the target video and the main preference vector: if both the projection vector and the main preference vector of the target video are zero vectors, set the cosine similarity to zero; if both the projection vector and the main preference vector of the target video are non-zero vectors, perform the inner product operation and divide the result by the product of the lengths of the two vectors to obtain the recommendation rating value of the target video in the corresponding user group. The recommended rating values ​​of all videos under the corresponding user group are recorded sequentially according to the video number, forming a set of recommended video ratings for the corresponding user group.

4. The recommendation weight optimization method based on group feature aggregation according to claim 3, characterized in that, The process of generating a corresponding candidate recommended video set based on the video recommendation rating set of each user group, and performing aggregation processing on the candidate recommended video sets of multiple user groups to form a final recommended video output list for a specific user, specifically includes: For any set of video recommendation ratings for a user group, sort all videos under the corresponding user group from high to low according to the recommendation rating value. If there are videos with the same recommendation rating value, sort them in ascending order of video number. Select several videos with the highest recommendation scores from the ranking results to form a candidate recommended video set for the corresponding user group. Each element in the candidate recommended video set is a video number. For each user, based on one or more user groups to which the target user belongs, obtain the candidate recommended video set corresponding to each user group, and perform a union operation on these candidate recommended video sets to obtain the merged candidate recommended video set for the target user; For each video in the merged candidate recommended video set, if the target video exists in the candidate recommended video sets of two or more user groups, the maximum value of the recommendation score of the target video in each user group is obtained as the final individual recommendation score of the target user for the target video; if the target video exists only in the candidate recommended video set of a single user group, the recommendation score of the target user in the corresponding user group is used as the final individual recommendation score. Sort all videos in the merged candidate recommendation video set from highest to lowest according to their final individual recommendation score. If there are videos with the same final individual recommendation score, sort them in ascending order of video number. Output the final recommendation video list for the target user.

Images