Big data-based user personalized recommendation method and system

Abstract

The present application relates to big data processing and personalized recommendation technical field, especially in kind of based on big data's user personalized recommendation method and system, the method includes: obtaining user history interaction sequence, separating stable subsequence and clustering generates user interest vector, extracts the burst behavior feature in recent fixed window and obtains interest adjustment vector;With dynamic weight fusion long short-term interest constructs comprehensive model, iteration optimization control short-term interest proportion;Adopt reinforcement learning to evaluate candidate content expected interaction reward, generate optimization recommendation sequence, combine platform real-time available content and replace low matching item.This method realizes the dynamic balance of long short-term interest, solves the interest imbalance problem in traditional recommendation, improves the recommendation accuracy and user experience, and provides technical support for platform to enhance user stickiness.

Description

Technical Field

[0001] This invention relates to the field of big data processing and personalized recommendation technology, and in particular to a user personalized recommendation method and system based on big data. Background Technology

[0002] In the internet age, with the explosive growth of user interaction data, leveraging big data processing technology to mine user interest characteristics has become a key direction for improving the effectiveness of personalized recommendations.

[0003] In existing technologies, personalized recommendation systems often rely on simple statistics or superficial analysis of users' historical interaction data to generate recommendations by matching content tags with user behavior tags. However, current technologies fail to achieve a dynamic balance between users' long-term stable interests and short-term impulsive behaviors through big data processing. User interests have a dual nature: long-term interests determine user loyalty to the platform, while short-term behaviors are influenced by immediate emotions and trending events. Existing systems lack the ability to deeply analyze massive amounts of user interaction data, making it impossible to accurately separate and quantify the weights of these two types of interests through big data processing.

[0004] Therefore, existing technologies are insufficient to meet the need for reasonable coordination of short-term and long-term interests in the field of personalized recommendation. Summary of the Invention

[0005] This invention provides a user-personalized recommendation method and system based on big data, in order to address the difficulty of existing technologies in meeting the need for reasonable coordination of short-term and long-term interests in the field of personalized recommendation.

[0006] In a first aspect, the present invention provides a user personalized recommendation method based on big data, comprising:

[0007] Obtain the user's historical interaction sequence, separate the content feature vector from the historical interaction sequence and perform cluster analysis to obtain the user interest vector;

[0008] Based on the user interest vector, the contribution of content items is calculated using an attention mechanism, and the contribution of the content items is sorted to determine the core interest ranking table.

[0009] The user's interaction sequence within a preset fixed time window is obtained, the sudden behavior features of the interaction sequence are extracted and compared with the user's interest vector, short-term sudden subsequences are marked based on a preset interval threshold, and the short-term sudden subsequences are quantized to obtain the interest adjustment vector.

[0010] The user interest vector and the interest adjustment vector are fused using a weighted summation operation to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is called to correct the fused interest vector to obtain a comprehensive interest model.

[0011] If the proportion of the fluctuation feature segment in the comprehensive interest model exceeds the preset upper limit, the weight coefficient corresponding to the fluctuation feature segment is reduced to obtain a modified weight set. The comprehensive interest model is iteratively optimized based on the modified weight set to obtain a modified interest model.

[0012] Extract a content candidate set from the modified interest model, perform reinforcement learning on the content candidate set, and output an optimized recommendation sequence;

[0013] Obtain item identifiers from the optimized recommendation sequence, and obtain the final recommendation list by matching the vector similarity of the item identifiers with the modified interest model.

[0014] In one optional implementation, the step of obtaining the user's historical interaction sequence, separating stable interaction subsequences and clustering similar content items to obtain a user interest vector containing category weight distribution includes:

[0015] Acquire user historical interaction data to form a historical interaction sequence, and filter out stable interaction sub-sequences based on the time difference of the historical interaction sequence;

[0016] Content feature vectors are extracted from the stable interaction subsequences, and K-means clustering is performed on the content feature vectors to obtain an aggregated content set;

[0017] The interaction frequency of each semantic cluster in the aggregated content set is counted, the preference accumulation value is calculated based on the interaction frequency, and the preference accumulation value is normalized to obtain the category weight distribution.

[0018] The category weight distribution is mapped to a preset interest feature space, and the user interest vector is obtained by weighted calculation in combination with the feature vector of the semantic cluster.

[0019] In one optional implementation, the step of calculating the contribution of content items and determining the core interest ranking table based on the user interest vector using an attention mechanism includes:

[0020] The platform's set of content to be recommended is acquired synchronously, and the set of content to be recommended is mapped to generate a candidate content feature matrix;

[0021] Calculate the original attention scores of the candidate content feature matrix and the user interest vector, and obtain the content item contribution after normalization;

[0022] Based on the contribution of the content items, content exceeding a preset contribution threshold is filtered out to form core interest node features;

[0023] The core interest node features are sorted in descending order according to the contribution of the content items to generate a core interest ranking table.

[0024] In one optional implementation, the steps of obtaining the user's interaction sequence within a preset fixed time window, extracting the burst behavior features of the interaction sequence and comparing them with the user's interest vector, labeling short-term burst subsequences based on a preset interval threshold, and quantizing the short-term burst subsequences to obtain an interest adjustment vector include:

[0025] Map the interaction sequence into an interaction feature matrix;

[0026] Extract continuous behavior segments with time intervals less than a preset interval threshold from the interaction feature matrix, organize the continuous behavior segments, and obtain sudden behavior features;

[0027] Calculate the cosine similarity between the sudden behavioral features and the user interest vector;

[0028] If the cosine similarity value is lower than a preset similarity threshold, the interaction sequence is marked as a short burst subsequence, and the vectors in the short burst subsequence are weighted and summed to calculate the interest adjustment vector.

[0029] In one optional implementation, the user interest vector and the interest adjustment vector are fused using a weighted summation operation to obtain a fused interest vector. If the dimensionality difference of the fused interest vector is lower than a preset difference threshold, a preset balancing factor is invoked to correct the fused interest vector, resulting in a comprehensive interest model, including:

[0030] Obtain user interaction logs and calculate the recent interaction intensity values ​​of the user interaction logs and the interest adjustment vector using a preset time decay function;

[0031] The dynamic weighting coefficient is determined by adjusting the influence ratio of long-term and short-term interests based on the recent interaction intensity values.

[0032] The user interest vector and the interest adjustment vector are weighted and fused using the dynamic weight coefficients to generate a fused interest vector.

[0033] If the dimensional difference of the fused interest vector is lower than a preset difference threshold, the interest differentiation balance factor is invoked for correction, and a comprehensive interest model is constructed.

[0034] In one optional implementation, if the proportion of fluctuation feature segments in the comprehensive interest model exceeds a preset upper limit, the weight coefficients corresponding to the fluctuation feature segments are reduced to obtain a modified weight set. The comprehensive interest model is iteratively optimized based on the modified weight set to obtain a modified interest model. The modified interest model obtained through iterative optimization includes:

[0035] Obtain the time series data of the comprehensive interest model, calculate the characteristic fluctuation amplitude of the time series data, and mark the segments whose characteristic fluctuation amplitude exceeds a preset benchmark threshold as fluctuation feature segments;

[0036] The energy distribution of the wave characteristic segments is statistically analyzed and the proportion value is calculated. If the proportion value exceeds the preset proportion upper limit, a preset attenuation factor is introduced to adjust the weight coefficient corresponding to the wave characteristic segments, and a corrected weight set is generated.

[0037] The model is reconstructed based on the modified weight set, and the deviation metric is calculated. The modified interest model is obtained through iterative optimization.

[0038] In one optional implementation, the step of extracting a content candidate set from the modified interest model, performing reinforcement learning on the content candidate set, and outputting an optimized recommendation sequence includes:

[0039] Obtain the multidimensional interest feature vector of the modified interest model, calculate the vector dot product of the multidimensional interest feature vector and the item feature, sort the items according to the size of the vector dot product, and obtain the content candidate set.

[0040] Construct a state vector and a reward function based on the content candidate set;

[0041] A deep Q-network is constructed based on the reward function, the state vector is input into the deep Q-network, and the expected interaction reward estimate is output.

[0042] Based on the estimated expected interaction reward, the content candidate set is sorted in descending order to obtain an optimized recommendation sequence.

[0043] In one optional implementation, obtaining item identifiers from the optimized recommendation sequence and obtaining the final recommendation list based on the vector similarity between the item identifiers and the modified interest model includes:

[0044] Obtain the item identifiers in the optimized recommendation sequence, remove invalid items based on the item identifiers, and obtain the sequence to be verified;

[0045] Item feature vectors are extracted from the sequence to be verified, and the cosine similarity value between the item feature vectors and the multidimensional interest feature vectors is calculated. If the cosine similarity value is lower than a preset correlation threshold, it is marked as a low match; if the cosine similarity value is not lower than the preset correlation threshold, it is marked as a high match.

[0046] The high-matching item is selected from the candidate pool to replace the low-matching item, and the final recommendation list is generated based on the replaced sequence to be verified.

[0047] Secondly, the present invention provides a user personalized recommendation system based on big data, comprising:

[0048] The interest modeling module is used to obtain the user's historical interaction sequence, separate the content feature vector from the historical interaction sequence and perform cluster analysis to obtain the user interest vector;

[0049] The interest ranking module is used to calculate the contribution of content items based on the user interest vector using an attention mechanism, rank the contribution of the content items, and determine the core interest ranking table.

[0050] The interest extraction module is used to obtain the user's interaction sequence within a preset fixed time window, extract the sudden behavior features of the interaction sequence and compare them with the user's interest vector, mark short-term sudden subsequences based on a preset interval threshold, quantize the short-term sudden subsequences, and obtain the interest adjustment vector.

[0051] The interest fusion module is used to fuse the user interest vector and the interest adjustment vector using a weighted summation operation to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is called to correct the fused interest vector to obtain a comprehensive interest model.

[0052] The model optimization module is used to reduce the weight coefficients of the fluctuation feature segments if the proportion of the fluctuation feature segments in the comprehensive interest model exceeds a preset upper limit, thereby obtaining a corrected weight set. Based on the corrected weight set, the comprehensive interest model is iteratively optimized to obtain a corrected interest model. The corrected interest model is obtained through iterative optimization.

[0053] The recommendation sequence module is used to extract a content candidate set from the modified interest model, perform reinforcement learning on the content candidate set, and output an optimized recommendation sequence.

[0054] The final recommendation module is used to obtain item identifiers in the optimized recommendation sequence, and obtain the final recommendation list by matching the vector similarity of the item identifiers with the modified interest model.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] (1) This invention separates stable user interaction subsequences and clusters them using big data processing technology to generate interest vectors that represent long-term preferences. This breaks through the limitations of traditional shallow analysis, achieves accurate capture of core interests, and provides data support for recommendation stability.

[0057] (2) This invention dynamically generates interest adjustment vectors by extracting short-term sudden behavioral features and comparing similarity, and combines the recent interaction intensity to adjust the fusion weights, thereby constructing a comprehensive model that takes into account both short-term and long-term interests, effectively avoiding recommendation imbalance caused by a single interest dominance;

[0058] (3) This invention controls the proportion of short-term interests through iterative optimization, evaluates the value of candidate content by combining reinforcement learning, and finally matches real-time available content to generate a recommendation list, forming a closed-loop mechanism of "interest modeling-fusion optimization-recommendation generation", which improves the accuracy of recommendations and user experience, and enhances the stickiness of platform users. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the user personalized recommendation method based on big data provided in the first embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the structure of a user personalized recommendation system based on big data provided in the second embodiment 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] Reference Figure 1 The first embodiment of the present invention provides a user personalized recommendation method based on big data, including the following steps:

[0063] S1, obtain the user's historical interaction sequence, separate stable interaction subsequences and cluster similar content items to obtain a user interest vector containing category weight distribution;

[0064] S2, Based on the user interest vector, use an attention mechanism to calculate the contribution of content items and determine the core interest ranking table;

[0065] S3, obtain the user's interaction sequence within a preset fixed time window, extract the sudden behavior features of the interaction sequence and compare them with the user's interest vector, mark short-term sudden subsequences based on a preset interval threshold, quantize the short-term sudden subsequences, and obtain the interest adjustment vector;

[0066] S4, use a weighted summation operation to fuse the user interest vector and the interest adjustment vector to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, call a preset balance factor to correct the fused interest vector to obtain a comprehensive interest model.

[0067] S5, determine the proportion of the fluctuation feature segment in the comprehensive interest model. If it exceeds the preset upper limit, reduce the weight coefficient corresponding to the fluctuation feature segment. The corrected interest model is obtained through iterative optimization.

[0068] S6, extract a content candidate set from the modified interest model, perform reinforcement learning on the content candidate set, and output an optimized recommendation sequence;

[0069] S7, obtain the item identifiers in the optimized recommendation sequence, and obtain the final recommendation list by matching the vector similarity of the item identifiers with the modified interest model.

[0070] In step S1, the user's historical interaction sequence is obtained, stable interaction subsequences are separated and clustered to aggregate similar content items, resulting in a user interest vector containing category weight distribution, including:

[0071] S11, acquire user historical interaction data to form a historical interaction sequence, and filter out stable interaction sub-sequences based on the time difference of the historical interaction sequence;

[0072] S12, extract content feature vectors from the stable interaction subsequences, perform K-means clustering on the content feature vectors to obtain an aggregated content set;

[0073] S13, count the interaction frequency of each semantic cluster in the aggregated content set, calculate the preference accumulation value based on the interaction frequency, normalize the preference accumulation value, and obtain the category weight distribution;

[0074] S14, the category weight distribution is mapped to a preset interest feature space, and the user interest vector is obtained by weighted calculation in combination with the feature vector of the semantic cluster.

[0075] In step S11, historical user interaction data is acquired to form a historical interaction sequence, and stable interaction sub-sequences are selected based on the time difference of the historical interaction sequence.

[0076] User historical interaction data covers all user interactions on the platform, including content clicks, viewing time, favorites, comments, and sharing. Each log entry contains a unique user identifier, interaction timestamp, content identifier, and behavior type. The historical interaction sequence is constructed by arranging all interaction logs of the same user in chronological order from earliest to latest, forming a continuous behavior sequence. For example, after sorting a user's interaction logs within 3 months by time, the original sequence would contain "2024-01-05 19:30 Clicked Course A → 2024-01-07 19:25 Watched Course A → ...".

[0077] The criteria for judging time difference filtering are as follows: First, a long-term observation window, such as 3 months, is set to capture users' long-term interest patterns. Within this window, for user interaction sequences within the same content category, all consecutive interaction segments are identified. A valid segment must contain at least 3 interactions. For example, assuming a segment with 6 interactions, 5 consecutive interaction time differences can be calculated. The coefficient of variation (i.e., the ratio of standard deviation to mean) of this time difference sequence is calculated. Segments with a coefficient of variation lower than a preset stability threshold (e.g., 0.2) and at least 3 interactions are selected as stable interaction subsequences. Setting a minimum number of interactions ensures that the time difference sequence is statistically significant and avoids misjudgments due to insufficient data. For example, if a user watches programming courses every Monday and Wednesday evening, a segment containing 10 interactions has a time difference sequence of [2 days, 2 days, ...]. The average time difference is 2 days, the standard deviation is close to 0, and the coefficient of variation is far below 0.2. This segment is then judged as a stable interaction subsequence.

[0078] User historical interaction data is stored in a relational database, primarily comprising the following table structures: a basic user information table storing user ID, registration time, and other basic attributes; a content information table storing metadata such as content ID, title, category tags, keywords, and difficulty level; and an interaction behavior table storing interaction records including user ID, content ID, interaction timestamp, behavior type, and dwell time. This table uses a composite index of user ID and content ID to support efficient historical data queries. The interaction behavior table employs a partitioned storage strategy, partitioning by interaction timestamp monthly to ensure query performance under large data volumes. Data is updated in real-time, with high-frequency interaction events cached via a message queue and written to the database in batches to reduce system load.

[0079] In step S12, content feature vectors are extracted from the stable interaction subsequences, and K-means clustering is performed on the content feature vectors to obtain an aggregated content set.

[0080] The content feature vector is constructed by extracting information such as thematic keywords, category tags, difficulty level, and related fields of the content, and converting them into multi-dimensional vectors. For example, the feature vector of the course "Python Loop Statements" includes dimensions such as "Programming → Python → Basics → Syntax". The number of clusters K is determined by first calculating the number N of all content feature vectors in the stable interaction subsequence. When N < 4, K is directly set to 1, that is, all content is regarded as a cluster. When N ≥ 4, the value range of K is set to [2, min(10, N-1)]. The silhouette coefficient method is used to traverse each candidate K value within this range, perform K-means clustering, and calculate the average silhouette coefficient of all samples.

[0081] The silhouette coefficient is calculated as follows: for each sample, the average distance *a* between the sample and other samples in the same cluster, and the average distance *b* between the sample and the nearest sample in a different cluster, are calculated. The silhouette coefficient is (ba) / max(a,b), and its value ranges from [-1,1]. The closer the value is to 1, the better the clustering effect. This method of determining the K value can adaptively select the optimal number of clusters based on the data distribution, ensuring the semantic consistency of the aggregated content set.

[0082] In step S13, the interaction frequency of each semantic cluster in the aggregated content set is counted, the preference accumulation value is calculated based on the interaction frequency, and the preference accumulation value is normalized to obtain the category weight distribution.

[0083] The interaction frequency is calculated by traversing stable interaction subsequences and recording the total number of interactions between the user and all content within each semantic cluster. For example, the "Python Introduction" semantic cluster has 30 interactions, the "Data Structures" semantic cluster has 15, and the "Machine Learning Basics" semantic cluster has 10. These counts represent the accumulated preference values. Normalization is achieved by dividing the accumulated preference value of each semantic cluster by the sum of the accumulated preference values ​​of all semantic clusters, yielding the weight percentage for each semantic cluster. For example, with a total of 55 interactions, the weight of "Python Introduction" is approximately 30 / 55 ≈ 0.55, "Data Structures" is approximately 15 / 55 ≈ 0.27, and "Machine Learning Basics" is approximately 10 / 55 ≈ 0.18. Normalization converts interaction frequencies of different magnitudes into weights between 0 and 1, facilitating subsequent interest vector construction. This calculation method objectively reflects the preference intensity for each interest direction.

[0084] In step S14, the category weight distribution is mapped to a preset interest feature space, and a weighted calculation is performed by combining the feature vectors of the semantic clusters to obtain the user interest vector.

[0085] The interest feature space consists of a core feature label extracted from the platform's content metadata for each dimension. By collecting metadata of all content on the platform (including category labels, topic keywords, difficulty levels, etc.), and removing duplicates, a basic label set containing the corresponding features is formed. Each label corresponds to a dimension of the feature space. For example, Dimension 1: Programming, Dimension 2: Python, Dimension 3: Basic, Dimension 4: Data Structures, ..., Dimension D: Advanced.

[0086] For a given piece of content, its feature vector is a D-dimensional vector. Using a feature label-based vector representation method, if the content contains a feature label corresponding to a certain dimension, that dimension has a value of 1; otherwise, it has a value of 0, resulting in a binary vector. Subsequently, this binary vector is L2 normalized so that its magnitude is 1. This process helps in subsequent similarity calculations. For example, the course "Python Loop Statements" contains labels such as "Programming," "Python," "Basics," and "Syntax." Before normalization, its binary vector has a value of 1 in the corresponding dimension; after normalization, it becomes a set of decimals, such as 0.8, 0.2, and 0.1, with values ​​of 0 in other irrelevant dimensions.

[0087] The central feature vector of each semantic cluster is multiplied by the weight in the corresponding category weight distribution. Then, the weighted vectors of all semantic clusters are summed dimensionally to obtain the final user interest vector. This vector represents the user's long-term interests and is referred to below as the long-term interest vector.

[0088] The central feature vector of the semantic cluster is obtained by calculating the arithmetic mean of the feature vectors of all contents within the semantic cluster after L2 normalization, and then performing L2 normalization again on the mean vector to make its magnitude 1. For example, the feature vector of the semantic cluster "Python Introduction" is (0.8, 0.2, 0.1, ...) with a weight of 0.55, and the feature vector of "Data Structure" is (0.1, 0.7, 0.2, ...) with a weight of 0.27. The vector obtained by superimposing these features is the user's preference distribution across all feature dimensions.

[0089] In step S2, based on the user interest vector, an attention mechanism is used to calculate the contribution of content items and determine the core interest ranking table, including:

[0090] S21, synchronously acquire the set of content to be recommended from the platform, map the set of content to be recommended, and generate a candidate content feature matrix;

[0091] S22, calculate the original attention scores of the candidate content feature matrix and the user interest vector, and obtain the content item contribution after normalization;

[0092] S23, Based on the contribution of the content items, filter out content that exceeds a preset contribution threshold to form core interest node features;

[0093] S24, Sort the core interest node features in descending order according to the contribution of the content items to generate a core interest ranking table.

[0094] In step S21, the set of content to be recommended on the platform is obtained synchronously, and the set of content to be recommended is mapped to generate a candidate content feature matrix.

[0095] The candidate content feature matrix is ​​constructed from all content to be recommended on the platform. It is built by arranging the feature vectors of each piece of content by row, forming a matrix with dimensions of "content quantity × feature dimension". For example, if the platform has 1000 pieces of content to be recommended, and each piece of content has a 50-dimensional feature vector, then the candidate content feature matrix will be a 1000×50 matrix. During construction, it is essential to ensure that the feature dimensions of the content to be recommended are completely consistent with the feature dimensions of the user interest vectors; for example, both should include 50 dimensions such as topic, difficulty, and type, to guarantee the validity of subsequent score calculations.

[0096] In step S22, the original attention scores of the candidate content feature matrix and the user interest vector are calculated and normalized to obtain the content item contribution.

[0097] The original attention score is calculated based on the scaled dot product attention mechanism. Q is the user interest vector used as the query vector, with a dimension of 1 multiplied by the feature dimension d_k. The candidate content feature matrix is ​​simultaneously used as the key vector P, with a dimension of the number of candidate content n multiplied by the feature dimension d_k. The scaling factor of the attention mechanism is equal to the square root of d_k. This scaling factor is set because when d_k is large, the dot product result may enter the saturation region of the softmax function, causing gradient vanishing. Dividing by the scaling factor scales the dot product result to a suitable range. In this embodiment, the feature vector dimension d_k is set to 50, so the scaling factor is approximately 7.07. This setting effectively balances the sharpness of the attention distribution with the gradient propagation efficiency. First, the dot product of the transpose of the query vector Q and the key vector P is calculated to obtain the initial matching score. Then, the dot product result is divided by the scaling factor to obtain the scaled score vector. The scaled score vector is input into the softmax function for normalization to obtain the attention weight vector. Each element in this weight vector has a value between 0 and 1, and the sum of all elements is 1, representing the attention allocation weight of each candidate content relative to the user's interest. This attention weight vector is directly used as the contribution of each content item. For example, the attention weight of "Python Function Practice" is calculated to be 0.6, the attention weight of "Common Algorithm Problem Analysis" is 0.3, and the attention weight of "Introduction to Front-End Development" is 0.1. These weight values ​​represent the degree of matching between each piece of content and the user's interests.

[0098] In step S23, based on the contribution of the content items, content exceeding a preset contribution threshold is filtered to form core interest node features.

[0099] The preset contribution threshold is set to 0.5. This threshold is based on statistical analysis of a large amount of user recommendation feedback data, which shows that content with a contribution score higher than 0.5 generally has a conversion rate exceeding 30%, while content with a score lower than 0.5 has a conversion rate of less than 5%. 0.5 is the critical value for distinguishing between core interests and peripheral interests. The filtering method involves iterating through the contribution scores of all content to be recommended, retaining content with scores higher than 0.5 as core interest nodes. For example, "Python Function Practice" (1.0) and "Common Algorithm Problem Analysis" (0.93) are retained, while "Introduction to Front-End Development" (0.11) is removed. The sorting method involves arranging the core interest nodes from highest to lowest contribution score to form an ordered list.

[0100] In step S24, the core interest node features are sorted in descending order according to the contribution of the content items to generate a core interest ranking table.

[0101] The core interest ranking table is generated by sequentially numbering the ranked core interest nodes and associating them with detailed content information (such as title, description, resource links, etc.) to form a structured list. The first 30% of the nodes in the list are marked as "high-priority recommendations," and subsequent nodes are marked as "medium-priority," "low-priority," etc., according to their contribution. For example, the first 5 items in the list are high-priority and will be displayed first in the top area of ​​the user recommendation page. This list can clearly distinguish the strength of the association between different content and the user's core interests, providing a basis for subsequent recommendation execution.

[0102] In step S3, the user's interaction sequence within a preset fixed time window is obtained, the sudden behavioral features of the interaction sequence are extracted and compared with the user's interest vector, short-term sudden subsequences are labeled based on a preset interval threshold, and the short-term sudden subsequences are quantized to obtain an interest adjustment vector, including:

[0103] S31, map the interaction sequence into an interaction feature matrix;

[0104] S32, extract continuous behavior segments with time intervals less than a preset interval threshold from the interaction feature matrix, organize the continuous behavior segments, and obtain sudden behavior features;

[0105] S33, Calculate the cosine similarity between the sudden behavior feature and the user interest vector;

[0106] S34, if the cosine similarity value is lower than the preset similarity threshold, then the interaction sequence is marked as a short burst subsequence, and the vectors in the short burst subsequence are weighted and summed to calculate the interest adjustment vector.

[0107] In step S31, the interaction sequence is mapped into an interaction feature matrix.

[0108] The fixed time window is set to 7 days. This threshold is based on the fact that users' short-term bursts of interest typically last 1-7 days, and 7 days can fully capture recent concentrated interactive behaviors while avoiding the mixing of stable interest behaviors over longer periods. The interaction sequence is obtained by extracting all user interaction logs from the past 7 days, sorting them by timestamp to form a sequence, for example, "2024-04-01 20:00 Browsing data analysis course → 2024-04-01 20:15 Watching data analysis course → ...". The interaction feature matrix is ​​mapped so that each row corresponds to one interaction, and the columns correspond to behavioral attributes, such as behavior type, dwell time, content characteristics, and interaction intensity. For example, a row of data might be "Watching → 45 minutes → Data Analysis → 0.8", where the interaction intensity is determined by the ratio of dwell time to the total content duration.

[0109] In step S32, continuous behavior segments with time intervals less than a preset interval threshold are extracted from the interaction feature matrix, and the continuous behavior segments are organized to obtain sudden behavior features.

[0110] The preset interval threshold is set to 30 minutes. This threshold is based on the premise that multiple interactions within a 30-minute period are more likely to be sudden behaviors driven by immediate emotions or trending events, rather than a continuation of stable interests. For example, if a user browses three data analysis courses consecutively between 8:00 PM and 8:25 PM, with each time interval less than 30 minutes, this segment is extracted as a sudden behavior segment. The extraction method involves traversing the interaction feature matrix and calculating the time difference between two adjacent interactions. If the time difference between multiple consecutive interactions is less than 30 minutes, these interaction segments are integrated into a single sudden behavior feature set, which includes all behavioral attributes and content features of that segment.

[0111] In step S33, the cosine similarity value between the sudden behavior feature and the user interest vector is calculated.

[0112] To calculate cosine similarity, the sudden behavioral features first need to be vectorized. The content feature vectors corresponding to all interactions within the behavioral segment are arithmetically averaged to obtain a comprehensive feature vector, which represents the overall interest direction of the sudden behavior. For example, averaging the feature vectors of three data analysis courses yields a comprehensive feature vector for "Data Analysis → Tool Application → Intermediate". The closer the cosine similarity result is to 1, the more closely the sudden behavior aligns with long-term interests; the closer it is to 0, the greater the difference.

[0113] In step S34, if the cosine similarity value is lower than a preset similarity threshold, the interaction sequence is marked as a short burst subsequence, and the vectors in the short burst subsequence are weighted and summed to calculate the interest adjustment vector.

[0114] The preset similarity threshold is set to 0.6. This threshold is based on the statistical finding that sudden behaviors with similarity higher than 0.6 are mostly extensions of long-term interests, while those lower than 0.6 are independent short-term interests and need to be marked separately. If the similarity is lower than 0.6, the behavior segment is marked as a short-term sudden subsequence. The feature vectors of all content within the sudden subsequence are weighted and summed using the interaction dwell time as the weight. For example, if the dwell times of three interactions within a segment are 45 minutes, 35 minutes, and 28 minutes, the weights are 45 / 108, 35 / 108, and 28 / 108 respectively. The weighted sum yields the interest adjustment vector, which is then L2 normalized to a magnitude of 1. This vector represents the user's recent short-term interests and is referred to below as the short-term interest vector.

[0115] In step S4, a weighted summation operation is used to fuse the user interest vector and the interest adjustment vector to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is invoked to correct the fused interest vector, resulting in a comprehensive interest model, including:

[0116] S41, Obtain user interaction logs, and calculate the recent interaction intensity values ​​of the user interaction logs and the interest adjustment vector by combining the preset time decay function;

[0117] S42, Adjust the influence ratio of long-term and short-term interests according to the recent interaction intensity value, and determine the dynamic weight coefficient;

[0118] S43, the user interest vector and the interest adjustment vector are weighted and fused using the dynamic weight coefficients to generate a fused interest vector;

[0119] S44. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is called to correct the fused interest vector to obtain a corrected interest vector, and the corrected interest vector is used to construct a comprehensive interest model.

[0120] In step S41, user interaction logs are obtained, and the recent interaction intensity values ​​of the user interaction logs and the interest adjustment vector are calculated by combining the preset time decay function.

[0121] The calculation of recent interaction intensity considers all user interactions related to short-term interests within the past 14 days, including clicks, views, favorites, and comments. An exponential decay function is used, specifically the decay weight equal to exp(−λ×t), where the decay rate λ is fixed at 0.1, and t is the number of days since the interaction occurred. The intensity of each interaction (e.g., viewing duration percentage, comment word count weight, etc.) is multiplied by the decay weight calculated using the exponential decay function, and all results are summed to obtain the recent interaction intensity value.

[0122] In step S42, the influence ratio of long-term and short-term interests is adjusted according to the recent interaction intensity value to determine the dynamic weight coefficient.

[0123] The dynamic weighting coefficients are determined as follows: the decay weight output by the decay function is directly related to the recent interaction intensity; the higher the intensity, the higher the proportion of short-term weight and the lower the proportion of long-term weight. The mapping logic is that the weight fluctuates linearly between 0.3 and 0.7 as the decay weight increases. The ratio of the recent interaction intensity value to the long-term average interest intensity value is calculated by selecting a long-term time window (e.g., the past 6 months), statistically analyzing the intensity of all user interactions within that time period, and calculating their arithmetic mean as the long-term average interest intensity value. The calculation method for behavior intensity is consistent with the definition in the calculation of recent interaction intensity. If the ratio is greater than or equal to 2.5, the short-term interest weight coefficient is set to 0.55, and the long-term interest weight coefficient is set to 0.45; if the ratio is between 1.5 and 2.5 (inclusive of 1.5, exclusive of 2.5), the short-term weight is 0.4, and the long-term weight is 0.6; if the ratio is less than 1.5, the short-term weight is 0.3, and the long-term weight is 0.7. The basis for this rule is that when the short-term interaction intensity is significantly higher than the long-term average level, the influence ratio of short-term interest is appropriately increased, and vice versa, long-term interest is emphasized to ensure the rationality of weight adjustment. For example, if the recent interaction intensity ratio is 3.2, the short-term weight is set to 0.55, and the long-term weight is 0.45.

[0124] In step S43, the user interest vector and the interest adjustment vector are weighted and fused using the dynamic weight coefficients to generate a fused interest vector.

[0125] The weighted fusion calculation method involves multiplying the user interest vector by a long-term weight coefficient, multiplying the interest adjustment vector by a short-term weight coefficient, and then element-wise summing the two result vectors to obtain the fused interest vector. For example, the long-term vector (0.6, 0.3, 0.1, ...) multiplied by 0.45 becomes (0.27, 0.135, 0.045, ...); the short-term vector (0.1, 0.2, 0.7, ...) multiplied by 0.55 becomes (0.055, 0.11, 0.385, ...); and the sum yields the fused interest vector (0.325, 0.245, 0.43, ...), which simultaneously contains features of both long-term and short-term interests.

[0126] In step S44, if the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is invoked to correct the fused interest vector to obtain a corrected interest vector, and the corrected interest vector is used to construct a comprehensive interest model.

[0127] The dimensional difference is calculated by finding the maximum and minimum values ​​of all dimensions in the fused interest vector and calculating the difference between them. A preset difference threshold of 0.35 is set, based on the premise that a difference below 0.35 indicates an overly even distribution of interests, failing to highlight the core direction. The correction logic for the interest differentiation balance factor is as follows: First, an interest significance threshold is set, for example, 0.1. For each dimension of the interest vector, if the value of the user interest vector (long-term interest vector) in that dimension is greater than the interest significance threshold, then that dimension is determined to be a long-term interest-dominated dimension. If the value of the user interest vector in that dimension is not greater than the interest significance threshold, and the value of the interest adjustment vector (short-term interest vector) in that dimension is greater than the interest significance threshold, then that dimension is determined to be a short-term interest-dominated dimension. The values ​​of long-term interest-dominated dimensions are uniformly increased by 15%, and the values ​​of short-term interest-dominated dimensions are uniformly decreased by 12%. The corrected vector is the comprehensive interest model.

[0128] In step S5, the proportion of fluctuation feature segments in the comprehensive interest model is determined. If it exceeds a preset upper limit, the weight coefficient corresponding to the fluctuation feature segment is reduced. A modified interest model is obtained through iterative optimization, including:

[0129] S51, acquire the time series data of the comprehensive interest model, calculate the characteristic fluctuation amplitude of the time series data, and mark the segment whose characteristic fluctuation amplitude exceeds a preset benchmark threshold as a fluctuation feature segment;

[0130] S52, Statistically analyze the energy distribution of the wave characteristic segment and calculate the proportion value. If the proportion value exceeds the preset proportion upper limit, introduce a preset attenuation factor to adjust the weight coefficient corresponding to the wave characteristic segment and generate a corrected weight set.

[0131] S53, reconstruct the model based on the modified weight set and calculate the deviation metric, and obtain the modified interest model through iterative optimization.

[0132] In step S51, the time series data of the integrated interest model is obtained, the characteristic fluctuation amplitude of the time series data is calculated, and segments whose characteristic fluctuation amplitude exceeds a preset benchmark threshold are marked as fluctuation feature segments.

[0133] Time series data records the changes in feature values ​​of each dimension in the comprehensive interest model over time, with a time granularity of 1 day. For example, the feature value change sequence for the "Data Analysis" dimension over the past 30 days is (0.1, 0.12, 0.43, ...). For each dimension of the comprehensive interest model, the daily feature fluctuation amplitude is calculated in chronological order. The calculation method is the absolute value of the difference between the feature value of the current day and the feature value of the same dimension of the previous day, divided by the feature value of the previous day. A fluctuation feature segment is defined in the time series as a continuous data segment in which the feature fluctuation amplitude of N consecutive time steps (N≥3) under the same dimension exceeds a preset benchmark threshold. For a time segment, such as 5 consecutive days, the number of days in which the daily fluctuation ratio of each dimension in the segment continuously exceeds the preset benchmark threshold is counted. The preset benchmark threshold is set to 1.5. When the feature fluctuation amplitude exceeds 1.5, it indicates that the feature value changes drastically, which is more likely caused by short-term sudden interest. The time steps that continuously meet the threshold are marked as fluctuation points. The complete data segment composed of consecutive fluctuation points is the identified and marked fluctuation feature segment.

[0134] In step S52, the energy distribution of the wave characteristic segment is statistically analyzed and the percentage value is calculated. If the percentage value exceeds the preset percentage upper limit, a preset attenuation factor is introduced to adjust the weight coefficient corresponding to the wave characteristic segment, and a corrected weight set is generated.

[0135] The statistical method for energy distribution is to calculate the sum of the squares of all dimensional feature values ​​of the comprehensive interest model for each day within the segment, and then sum these daily energy values ​​to obtain the total energy of the segment. The total model energy is the sum of the daily energy values ​​of the comprehensive interest model for all days within the same time period (e.g., the last 30 days). The calculation method for the proportion of fluctuating feature segments is to divide the total energy of the fluctuating feature segments by the total energy of the time series model, i.e., the energy integral proportion. A preset upper limit of 60% is set after verification with a large amount of historical user interest data. This threshold is the critical value for distinguishing between normal small fluctuations and short-term sudden excessive fluctuations. Exceeding this value indicates that short-term fluctuations excessively interfere with long-term interests. If the proportion value exceeds the preset upper limit, a preset attenuation factor is introduced to adjust the weight coefficients corresponding to the fluctuating feature segments, generating a corrected weight set.

[0136] In step S53, the model is reconstructed based on the modified weight set and the deviation metric is calculated. The modified interest model is obtained through iterative optimization.

[0137] The model reconstruction employs an element-wise weighted fusion method. Each dimension weight in the corrected weight set is multiplied element-wise with the corresponding dimension of the original feature vector of the comprehensive interest model, and then summed dimensionally to generate the reconstructed interest model. The deviation metric is calculated using a cosine distance deviation index, calculated as 1 minus the cosine similarity between the reconstructed interest model and the original comprehensive interest model. A larger deviation value indicates a greater model difference, while a smaller deviation value indicates a higher matching degree. During iterative optimization, dual termination conditions are set: the deviation metric value ≤ 0.05 (convergence threshold) or the number of iterations reaches 5 (maximum iteration limit). Iteration stops when either condition is met. The adjustment rule for the attenuation factor γ is as follows: the initial value is set to a preset attenuation factor; if the current deviation metric value > 0.1, γ is increased by 5%; if the deviation metric value ≤ 0.1, γ is decreased by 5%. After each adjustment, a new corrected weight set is generated, and the next iteration begins. The corrected interest model is obtained through iterative optimization.

[0138] In step S6, a content candidate set is extracted from the modified interest model, reinforcement learning is performed on the content candidate set, and an optimized recommendation sequence is output, including:

[0139] S61, obtain the multidimensional interest feature vector of the modified interest model, calculate the vector dot product of the multidimensional interest feature vector and the item feature, sort according to the size of the vector dot product, and obtain the content candidate set;

[0140] S62, construct a state vector and a reward function based on the content candidate set;

[0141] S63, Construct a deep Q-network based on the reward function, input the state vector into the deep Q-network, and output the expected interaction reward estimate;

[0142] S64, based on the expected interaction reward estimate, sort the content candidate set in descending order to obtain an optimized recommendation sequence.

[0143] In step S61, the multidimensional interest feature vector of the modified interest model is obtained, the vector dot product of the multidimensional interest feature vector and the item feature is calculated, and the vector dot product is sorted according to the size of the vector dot product to obtain the content candidate set.

[0144] The multidimensional interest feature vector is the core vector for refining the interest model, containing the user's preference distribution across all content feature dimensions. The item feature vector is the feature vector of all content in the platform's content library, with dimensions consistent with the interest feature vector. The vector dot product is calculated by multiplying the interest vector by the corresponding dimension of each item vector and then summing the results. A larger dot product indicates a higher degree of matching between the content and the user's interests. The content candidate set is extracted by selecting the top 200 items based on the dot product results. This number is chosen because 200 items cover sufficient potential matches while avoiding an excessive number that would increase the pressure on subsequent evaluations, thus forming the content candidate set.

[0145] In step S62, a state vector and a reward function are constructed based on the content candidate set.

[0146] Construct a state vector representing the current recommendation environment. This is a global vector and contains the following parts:

[0147] The user's current interest state, i.e., the multidimensional interest feature vector of the modified interest model; recent interaction history, i.e., the encoding of the user's most recent interaction behaviors, such as behavior type, content feature summary of the interaction, etc.; contextual information, such as current time, device, and other contextual features. The instantaneous feedback obtained after the user performs a recommendation action (i.e., clicking or viewing recommended content) is defined, and the reward function is as follows:

[0148] User clicks on content: +1 reward; User watches content for more than 70% of the total time: +2 bonus reward; User provides positive feedback on content (favorites, likes, shares): +3 bonus reward; User ignores recommendations or provides negative feedback (dislikes, reports): 0 or -1 reward.

[0149] In step S63, a deep Q-network is constructed based on the reward function, the state vector is input into the deep Q-network, and the expected interaction reward estimate is output.

[0150] The reinforcement learning evaluation network employs a deep Q-network (DQN), and the state vector has been constructed in S62. The action is to select a piece of content from the platform's content library for recommendation. To achieve efficient learning, the action space is defined as the identifier corresponding to each piece of content in the content candidate set (containing 200 pieces of content). Therefore, the action space size is 200.

[0151] The network input is a state vector, and the output is a 200-dimensional vector, with each output node corresponding to the Q-value of a content in the action space. The number of neurons in the input layer is the same as the dimension of the state vector; the first hidden layer contains 128 neurons using the ReLU activation function; the second hidden layer contains 64 neurons using the ReLU activation function; the third hidden layer contains 32 neurons using the ReLU activation function; and the output layer contains 200 neurons using a linear activation function, each corresponding to one of the 200 candidate content's Q-values.

[0152] During DQN training, the four-tuple experiences generated from each recommendation interaction are stored in an experience replay buffer. During training, the learning objective of DQN is to minimize the temporal difference error, and a small batch of experiences is randomly sampled from the buffer. For each sampled experience, the predicted Q-value and the target Q-value of the current network are calculated. The target Q-value is determined by the reward and the maximum Q-value of the target network for the next state, with a discount factor set to 0.9. The loss function is the mean squared error. By minimizing the loss, the parameters of the main network are updated using the Adam optimizer with a learning rate of 0.001, and the parameters of the target network are periodically copied from the main network. The trained DQN is used to generate recommendations. For the current state, the network outputs the Q-values ​​of all 200 actions; these Q-values ​​represent the estimated reward for the expected interaction.

[0153] In step S64, the content candidate set is sorted in descending order according to the expected interaction reward estimate to obtain an optimized recommendation sequence.

[0154] The sorting method involves ranking the 200 candidate items from highest to lowest based on their expected interaction reward valuation. Items with the same valuation are then sorted in reverse chronological order (latest update first). The optimized output format for the recommendation sequence is a structured list containing information such as content ID, recommendation priority, and expected interaction probability, for example, "Content ID: 1001 → Priority 1 → Expected interaction reward valuation 0.85; Content ID: 1023 → Priority 2 → Expected interaction reward valuation 0.82; ...". This sequence clearly defines the display order of the recommended content, providing a basis for generating the final recommendation list.

[0155] In step S7, item identifiers in the optimized recommendation sequence are obtained, and the final recommendation list is obtained by matching the vector similarity of the item identifiers with the modified interest model, including:

[0156] S71, Obtain the item identifiers in the optimized recommendation sequence, and remove invalid items according to the item identifiers to obtain the sequence to be verified;

[0157] S72, extract the item feature vector from the sequence to be verified, calculate the cosine similarity value between the item feature vector and the multidimensional interest feature vector, if the cosine similarity value is lower than the preset correlation threshold, it is marked as a low matching item, if the cosine similarity value is not lower than the preset correlation threshold, it is marked as a high matching item.

[0158] S73, Select the high-matching item from the candidate pool to replace the low-matching item, and generate the final recommendation list based on the replaced sequence to be verified.

[0159] In step S71, the item identifiers in the optimized recommended sequence are obtained, and invalid items are removed based on the item identifiers to obtain the sequence to be verified.

[0160] The item identifier is a unique ID assigned to each piece of content by the platform. This ID allows users to query the real-time status of the content, such as whether it has been removed, whether the copyright is valid, and whether it violates regulations. The criteria for determining invalid items are those whose content status is marked as "removed," "invalid," or "violation." The removal method involves iterating through all item identifiers in the optimization recommendation sequence, querying their real-time status, deleting invalid items, and retaining valid items to form a verification sequence. For example, after removing 5 invalid items from 200 items in the optimization sequence, the verification sequence contains 195 items.

[0161] In step S72, the item feature vector is extracted from the sequence to be verified, and the cosine similarity value between the item feature vector and the multidimensional interest feature vector is calculated. If the cosine similarity value is lower than the preset correlation threshold, it is marked as a low matching item.

[0162] The calculation method for cosine similarity is the same as in step S33, and it is used for secondary verification of the matching degree between the content in the sequence to be verified and the user's current interests. The preset relevance threshold is set to 0.55. The basis for setting this threshold is that content with a similarity value lower than 0.55 has a low degree of matching with the user's interests, and the probability of user interaction after display is low, which may even reduce the user experience. Therefore, it is marked as a low matching item. For example, the similarity value of a certain content is 0.42, and it is marked as a low matching item.

[0163] In step S73, the high-matching item is selected from the candidate pool to replace the low-matching item, and a final recommendation list is generated based on the replaced sequence to be verified.

[0164] The candidate pool is constructed by pre-screening content from the platform's content library that has a similarity score higher than 0.7 but is not included in the initial candidate set. This content is then stored in descending order of similarity, with a pool size of 500 entries to ensure sufficient replacement resources. The replacement rule is as follows: for each low-matching item, the content with the highest similarity within the same content category as that low-matching item is selected from the candidate pool for replacement. For example, if the low-matching item is "Introduction to Data Analysis Course," then "Practical Data Analysis Course" with a similarity score of 0.85 is selected from the candidate pool. The final recommendation list is generated by retaining the non-low-matching items in the verification sequence, adding the replaced high-matching items, adjusting them according to their original priority, and selecting the top 50 items as the final recommendation list. This list includes content titles, descriptions, cover links, and recommendation reasons, and can be directly used for user recommendation displays.

[0165] Reference Figure 2 The second embodiment of the present invention provides a user personalized recommendation system based on big data, comprising:

[0166] The interest modeling module is used to obtain the user's historical interaction sequence, separate the content feature vector from the historical interaction sequence and perform cluster analysis to obtain the user interest vector;

[0167] The interest ranking module is used to calculate the contribution of content items based on the user interest vector using an attention mechanism, rank the contribution of the content items, and determine the core interest ranking table.

[0168] The interest extraction module is used to obtain the user's interaction sequence within a preset fixed time window, extract the sudden behavior features of the interaction sequence and compare them with the user's interest vector, mark short-term sudden subsequences based on a preset interval threshold, quantize the short-term sudden subsequences, and obtain the interest adjustment vector.

[0169] The interest fusion module is used to fuse the user interest vector and the interest adjustment vector using a weighted summation operation to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is called to correct the fused interest vector to obtain a comprehensive interest model.

[0170] The model optimization module is used to reduce the weight coefficients of the fluctuation feature segments if the proportion of the fluctuation feature segments in the comprehensive interest model exceeds a preset upper limit, thereby obtaining a corrected weight set. Based on the corrected weight set, the comprehensive interest model is iteratively optimized to obtain a corrected interest model. The corrected interest model is obtained through iterative optimization.

[0171] The recommendation sequence module is used to extract a content candidate set from the modified interest model, perform reinforcement learning on the content candidate set, and output an optimized recommendation sequence.

[0172] The final recommendation module is used to obtain item identifiers in the optimized recommendation sequence, and obtain the final recommendation list by matching the vector similarity of the item identifiers with the modified interest model.

[0173] The system provided in this embodiment of the invention is used to execute all the process steps of the above method embodiments. The two correspond one-to-one in terms of module function, working logic and technical effect, and have the same personalized recommendation capability. Therefore, the specific working details and beneficial effects of the system can be referred to the foregoing method embodiments, and will not be repeated here.

[0174] The system provided in this embodiment of the invention is used to execute all the process steps of the above method embodiments. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0175] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0176] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A user-personalized recommendation method based on big data, characterized in that, include: Obtain the user's historical interaction sequence, separate the content feature vector from the historical interaction sequence and perform cluster analysis to obtain the user interest vector; Based on the user interest vector, the contribution of content items is calculated using an attention mechanism, and the contribution of the content items is sorted to determine the core interest ranking table. The user's interaction sequence within a preset fixed time window is obtained, the sudden behavior features of the interaction sequence are extracted and compared with the user's interest vector, short-term sudden subsequences are marked based on a preset interval threshold, and the short-term sudden subsequences are quantized to obtain the interest adjustment vector. The user interest vector and the interest adjustment vector are fused using a weighted summation operation to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is called to correct the fused interest vector to obtain a comprehensive interest model. If the proportion of the fluctuation feature segment in the comprehensive interest model exceeds the preset upper limit, the weight coefficient corresponding to the fluctuation feature segment is reduced to obtain a modified weight set. The comprehensive interest model is iteratively optimized based on the modified weight set to obtain a modified interest model. Extract a content candidate set from the modified interest model, perform reinforcement learning on the content candidate set, and output an optimized recommendation sequence; Obtain item identifiers from the optimized recommendation sequence, and match the vector similarity of the item identifiers with the modified interest model to obtain the final recommendation list; The step of fusing the user interest vector and the interest adjustment vector using a weighted summation operation to obtain the fused interest vector includes: Obtain user interaction logs and calculate the recent interaction intensity values ​​of the user interaction logs and the interest adjustment vector using a preset time decay function; The dynamic weighting coefficient is determined by adjusting the influence ratio of long-term and short-term interests based on the recent interaction intensity values. The user interest vector and the interest adjustment vector are weighted and fused using the dynamic weight coefficients to generate a fused interest vector. Wherein, if the proportion of the fluctuation feature segment in the comprehensive interest model exceeds a preset upper limit, the weight coefficient corresponding to the fluctuation feature segment is reduced to obtain a corrected weight set, including: Obtain the time series data of the comprehensive interest model, calculate the characteristic fluctuation amplitude of the time series data, and mark the segments whose characteristic fluctuation amplitude exceeds a preset benchmark threshold as fluctuation feature segments; The energy distribution of the wave characteristic segments is statistically analyzed and the proportion value is calculated. If the proportion value exceeds the preset proportion upper limit, a preset attenuation factor is introduced to adjust the weight coefficient corresponding to the wave characteristic segments, and a corrected weight set is generated. The model is reconstructed based on the modified weight set, and the deviation metric is calculated. The modified interest model is obtained through iterative optimization.

2. The method according to claim 1, characterized in that, The process of obtaining a user's historical interaction sequence, separating content feature vectors from the historical interaction sequence, and performing cluster analysis on the content feature vectors to obtain a user interest vector includes: The user's historical interaction data is acquired, the historical interaction data is organized to form a historical interaction sequence, and stable interaction subsequences are selected based on the time difference of the historical interaction sequence. Content feature vectors are extracted from the stable interaction subsequences, and K-means clustering is performed on the content feature vectors to obtain an aggregated content set; The interaction frequency of each semantic cluster in the aggregated content set is counted, the preference accumulation value is calculated based on the interaction frequency, and the preference accumulation value is normalized to obtain the category weight distribution. The category weight distribution is mapped to a preset interest feature space, and the user interest vector is obtained by weighted calculation in combination with the feature vector of the semantic cluster.

3. The method according to claim 1, characterized in that, The step of calculating the contribution of content items using an attention mechanism based on the user interest vector, ranking the contribution of the content items, and determining the core interest ranking table includes: The platform's set of content to be recommended is acquired synchronously, and the set of content to be recommended is mapped to generate a candidate content feature matrix; Calculate the original attention scores of the candidate content feature matrix and the user interest vector, and obtain the content item contribution after normalizing the original attention scores; Based on the contribution of the content items, content exceeding a preset contribution threshold is filtered out to form core interest node features; The core interest node features are sorted in descending order according to the contribution of the content items to generate a core interest ranking table.

4. The method according to claim 1, characterized in that, The process of extracting burst behavior features from the interaction sequence and comparing them with the user interest vector, labeling short-term burst subsequences based on a preset interval threshold, quantizing the short-term burst subsequences to obtain an interest adjustment vector includes: Map the interaction sequence into an interaction feature matrix; Extract continuous behavior segments with time intervals less than a preset interval threshold from the interaction feature matrix, organize the continuous behavior segments, and obtain sudden behavior features; Calculate the cosine similarity between the sudden behavioral features and the user interest vector; If the cosine similarity value is lower than a preset similarity threshold, the interaction sequence is marked as a short burst subsequence, and the vectors in the short burst subsequence are weighted and averaged to calculate the interest adjustment vector.

5. The method according to claim 1, characterized in that, The step of extracting a content candidate set from the modified interest model, performing reinforcement learning on the content candidate set, and outputting an optimized recommendation sequence includes: Obtain the multidimensional interest feature vector of the modified interest model, calculate the vector dot product of the multidimensional interest feature vector and the item feature, sort the items according to the size of the vector dot product, and obtain the content candidate set. Construct a state vector and a reward function based on the content candidate set; A deep Q-network is constructed based on the reward function, the state vector is input into the deep Q-network, and the expected interaction reward estimate is output. Based on the estimated expected interaction reward, the content candidate set is sorted in descending order to obtain an optimized recommendation sequence.

6. The method according to claim 5, characterized in that, The step of obtaining item identifiers from the optimized recommendation sequence and matching the vector similarity of the item identifiers with the modified interest model to obtain the final recommendation list includes: Obtain the item identifiers in the optimized recommendation sequence, remove invalid items based on the item identifiers, and obtain the sequence to be verified; Item feature vectors are extracted from the sequence to be verified, and the cosine similarity value between the item feature vectors and the multidimensional interest feature vectors is calculated. If the cosine similarity value is lower than a preset correlation threshold, it is marked as a low match; if the cosine similarity value is not lower than the preset correlation threshold, it is marked as a high match. The high-matching item is selected from the candidate pool to replace the low-matching item, and the final recommendation list is generated based on the replaced sequence to be verified.

7. A user-personalized recommendation system based on big data, characterized in that, For implementing the method as described in any one of claims 1-6, comprising: The interest modeling module is used to obtain the user's historical interaction sequence, separate the content feature vector from the historical interaction sequence and perform cluster analysis to obtain the user interest vector; The interest ranking module is used to calculate the contribution of content items based on the user interest vector using an attention mechanism, rank the contribution of the content items, and determine the core interest ranking table. The interest extraction module is used to obtain the user's interaction sequence within a preset fixed time window, extract the sudden behavior features of the interaction sequence and compare them with the user's interest vector, mark short-term sudden subsequences based on a preset interval threshold, quantize the short-term sudden subsequences, and obtain the interest adjustment vector. The interest fusion module is used to fuse the user interest vector and the interest adjustment vector using a weighted summation operation to obtain a fused interest vector. If the dimensional difference of the fused interest vector is lower than a preset difference threshold, a preset balance factor is called to correct the fused interest vector to obtain a comprehensive interest model. The model optimization module is used to reduce the weight coefficient of the fluctuation feature segment if the proportion of the fluctuation feature segment in the comprehensive interest model exceeds the preset upper limit, so as to obtain a corrected weight set, and iteratively optimize the comprehensive interest model based on the corrected weight set to obtain a corrected interest model. The recommendation sequence module is used to extract a content candidate set from the modified interest model, perform reinforcement learning on the content candidate set, and output an optimized recommendation sequence. The final recommendation module is used to obtain item identifiers in the optimized recommendation sequence, and obtain the final recommendation list by matching the vector similarity of the item identifiers with the modified interest model.

Images