A deep learning-based user behavior modeling and personalized recommendation method

Abstract

The application discloses a kind of based on deep learning user behavior modeling and individualized recommendation method, it is related to deep learning and behavior modeling technical field, including: step one, collect user behavior data and generate user behavior event sequence;Step two, extract behavior object semantic features and generate user behavior semantic trajectory;Step three, calculate semantic change data and generate conversation fragment set;Step four, construct conversation state matrix;Step five, conversation state matrix is input into improved STAMP model, executes behavior fission coding, conversation memory reorganization, end behavior competition, interest inertia construction and memory attention aggregation processing, generates user interest representation data;Step six, generate interest correlation matrix;Step seven, generate candidate object adaptation matrix;Step eight, generate individualized recommendation result.The application realizes the intelligent modeling and individualized recommendation of user interest evolution process.

Description

Technical Field

[0001] This invention relates to the field of deep learning and behavior modeling technology, and in particular to a deep learning-based method for user behavior modeling and personalized recommendation. Background Technology

[0002] With the rapid development of internet platforms, e-commerce platforms, and content service platforms, users generate a large amount of behavioral data during browsing, searching, clicking, collecting, purchasing, and commenting. How to accurately extract user interest features from complex behavioral data and generate personalized recommendation results has become an important research direction in the field of recommendation systems. Currently, existing technologies mainly use collaborative filtering, matrix factorization, sequence recommendation models, and deep learning-based conversational recommendation models to model user behavior. These models construct interest profiles by analyzing users' historical behavior records and then filter and rank recommended objects based on these profiles.

[0003] However, existing technologies still have certain limitations in practical applications. Most recommendation methods mainly focus on local correlations or short-term interest features in behavioral sequences, lacking effective characterization of the changing patterns of user interests during continuous behavior, and struggling to identify complex evolutionary phenomena such as interest switching, interest fission, and interest continuation. Although some deep learning-based conversation recommendation models introduce conversation memory and attention mechanisms, they usually model all historical behaviors uniformly into the same memory structure, lacking segmented modeling mechanisms for interest change states, resulting in insufficient utilization of interest switching boundary information and interest evolution process information. At the same time, most existing methods focus on the impact of end behaviors on recommendation results, failing to adequately consider the continuous interest inertia features and the correlation between different interest directions, making it difficult to form a unified interest expression structure that takes into account short-term interests, long-term interests, and interest transfer features. In addition, during the candidate matching process, existing technologies usually rank based on single interest matching results, lacking comprehensive consideration of interest correlation and interest coverage, causing recommendation results to easily concentrate on a few interest directions, affecting the diversity and adaptability of recommendation results.

[0004] Therefore, how to provide a deep learning-based method for user behavior modeling and personalized recommendation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for user behavior modeling and personalized recommendation. This invention constructs a user behavior semantic trajectory, a session state matrix, and an improved STAMP model. It utilizes behavior fission coding, session memory reorganization, end-behind behavior competition, and interest inertia maintenance mechanisms to model the evolution of user interests. Furthermore, it constructs an interest association matrix and a candidate object adaptation matrix, enabling the recommendation process to simultaneously consider interest matching and interest association relationships. Finally, it adjusts the recommendation results based on interest coverage distribution, improving the accuracy of user interest representation and the effectiveness of personalized recommendations.

[0006] A method for user behavior modeling and personalized recommendation based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Collect user behavior data, sort it according to the time of occurrence of the behavior, and generate a user behavior event sequence; Step 2: Extract semantic features of the behavior objects based on the user behavior event sequence to generate user behavior semantic trajectories; Step 3: Calculate semantic change data between adjacent behaviors based on the user behavior semantic trajectory, divide the user behavior event sequence according to the semantic change data, and generate a set of session fragments; Step 4: Extract session features based on the set of session fragments and construct a session state matrix; Step 5: Input the session state matrix into the improved STAMP model, perform behavioral fission encoding on the session state matrix to generate a fission behavior sequence; reconstruct the session memory based on the fission behavior sequence to generate a reconstructed session memory matrix; calculate the end-behavior competition relationship based on the reconstructed session memory matrix to generate a competitive behavior representation; construct an inertial interest vector based on the competitive behavior representation and the reconstructed session memory matrix; perform memory attention aggregation on the reconstructed session memory matrix, the competitive behavior representation, and the inertial interest vector to generate user interest representation data; Step Six: Calculate interest associations based on the user interest representation data to generate an interest association matrix; Step 7: Calculate the object adaptation relationship based on the interest association matrix to generate a candidate object adaptation matrix; Step 8: Sort the candidate objects based on the adaptation matrix and generate personalized recommendation results.

[0007] Optionally, step one specifically includes: Collect user browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell history generated on the target platform, and extract user identifiers, behavior object identifiers, behavior type identifiers, behavior occurrence time, visited page identifiers, dwell time, interaction location identifiers, as well as device type information, operating system information, and network environment information; The browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell history are collected according to user identifiers to form a user behavior record set; The user behavior record set is sorted by time based on the time the behavior occurred, forming a behavior time series; The time interval between adjacent behaviors is calculated based on the behavior time series, and the forward and backward time intervals corresponding to each behavior are recorded to form a behavior time association sequence; Based on the behavioral object identifier, the number of consecutive views, consecutive clicks, consecutive favorites, and consecutive stays are counted to form a behavioral activity sequence; Based on the behavior time association sequence and behavior activity sequence, consecutive behavior records with the same behavior object identifier, the same behavior type identifier and a time interval less than the average time interval of adjacent behaviors are merged. Data items with missing access page identifier, dwell time and device environment information are supplemented to form a standardized behavior sequence. The user behavior event sequence is generated by combining and arranging the user identifier, behavior occurrence time, behavior type identifier, behavior object identifier, behavior time association sequence, and behavior activity sequence.

[0008] Optionally, step two specifically includes: Read the behavior object identifier in the user behavior event sequence, associate the behavior object attribute library with the behavior object identifier, and extract object category information, object attribute information, object price information, object theme information, and object scene information; Encode object category information, object attribute information, object price information, object theme information, and object scene information to generate a set of semantic features for behavioral objects; Arrange the semantic feature sets of each behavior object according to the order of the occurrence of the behavior to form a semantic sequence of behavior objects; Calculate category change data, attribute change data, price change data, theme change data, and scene change data between adjacent behavioral objects based on the semantic sequence of behavioral objects; Based on category change data, attribute change data, price change data, theme change data, and scenario change data, semantic relationships between adjacent behavioral objects are constructed to form a set of behavioral semantic relationships; By connecting the semantic feature set of the behavior object and the corresponding semantic relationship of the behavior in the order of the behavior occurrence, a semantic trajectory of user behavior is generated.

[0009] Optionally, step three specifically includes: Read the semantic sequence of behavioral objects and the set of behavioral semantic associations from the user's behavioral semantic trajectory; The data on category changes, attribute changes, price changes, theme changes, and scene changes between adjacent behavioral objects are normalized and then weighted and accumulated to generate semantic change values, forming a semantic change sequence. The distribution of semantic change values ​​in the semantic change sequence is statistically analyzed according to the chronological order of the occurrence of the behavior, and a semantic change distribution sequence is generated. The frequency and magnitude of semantic changes in a continuous behavior window are statistically analyzed to generate a semantic change density sequence. The semantic change levels are generated according to the sorting position of the semantic change values ​​in the semantic change distribution sequence, forming a semantic change level sequence; Based on the semantic change level sequence and semantic change density sequence, determine the semantic transition positions in the user behavior event sequence, and record the behavior object identifier and behavior occurrence time corresponding to each semantic transition position; The user behavior event sequence is divided into multiple session fragments by using each semantic transition point as a segmentation point; Arrange the conversation segments in chronological order of the events to generate a set of conversation segments.

[0010] Optionally, step four specifically involves: Read the sequence of behavioral objects and the time sequence of behavioral occurrences corresponding to each session segment in the set of session segments; Statistically analyze the number of behavioral objects, duration of behavior, distribution of behavior types, and distribution of behavioral object categories in each session segment to generate session scale characteristics; Based on the semantic feature set of the behavioral objects corresponding to each conversation segment, the distribution of object topics, object attributes, object prices, and object scenarios are statistically analyzed to generate conversation semantic features; Based on the semantic change values ​​of adjacent behavioral objects in each conversation segment, the mean semantic change, the range of semantic change, the frequency of semantic change, and the density of semantic change are statistically analyzed to generate conversation change features. Based on the behavior occurrence time series, the average time interval, maximum time interval, and minimum time interval between adjacent behaviors in each session segment are calculated to generate session time features; A session feature vector is generated by combining and arranging session size features, session semantic features, session change features, and session time features. Arrange the session feature vectors according to the time sequence of session fragment formation to construct a session state matrix.

[0011] Optionally, step five specifically includes: Input the session state matrix into the improved STAMP model; The improved STAMP model includes sequentially connected behavioral fission coding units, conversation memory reorganization units, end-behavior competition units, interest inertia maintenance units, and memory attention aggregation units; The behavior fission coding unit reads each session feature vector in the session state matrix, calculates the sum of the absolute values ​​of the differences between the corresponding feature elements of adjacent session feature vectors, and performs cumulative calculations in conjunction with the corresponding semantic change density to generate a session fission intensity sequence. Based on the session fission intensity sequence, each session feature vector is marked for fission, forming a fission-enhanced session matrix; The session memory recombination unit reads the fission enhancement session matrix and statistically analyzes the direction and magnitude of change in the session fission intensity corresponding to adjacent session feature vectors. The consistency of fission trends is generated based on the direction and magnitude of change. Session feature vectors with continuously changing fission trend consistency are divided into the same memory block; session feature vectors with changing direction are divided into different memory blocks to form a recombined session memory matrix. The end-behind behavior contention unit reads the end-behind behavior record corresponding to each memory block; The duration of the behavior, the frequency of the behavior, and the number of behavior conversions corresponding to the terminal behavior records are counted separately, and the proportion of each statistical result to the total amount of the corresponding memory block is calculated. The competitive behavior is represented by a weighted summation based on the proportional values. The interest inertia preservation unit reads and reconstructs the session memory matrix; and counts the number of consecutive occurrences of the same object category, the same topic, and the same scene. Generate an inertial interest vector based on the proportion of consecutive occurrences to the total number of actions in the session; The memory attention aggregation unit reads and reconstructs the session memory matrix, the competitive behavior representation, and the inertial interest vector; Calculate the sum of the element-wise products between the reconstructed session memory matrix and the inertial interest vector, and the sum of the element-wise products between the competitive behavior representation and the inertial interest vector, respectively. Based on the calculation results, a memory attention distribution is generated, and the recombined session memory matrix, competitive behavior representation, and inertial interest vector are weighted and aggregated according to the memory attention distribution to form user interest representation data.

[0012] Optionally, the reorganized session memory matrix is ​​specifically: Read the feature vectors of each session and the corresponding session fission intensity sequence from the fission enhancement session matrix; Arrange the feature vectors of each session in chronological order of their occurrence to form a sequence of sessions to be reconstructed. Calculate the direction and magnitude of change in fission intensity between adjacent sessions; The consistency of fission trend is determined by statistically analyzing the number of times the direction of change corresponds to the fission intensity of multiple consecutive sessions and the fluctuation range of the change amplitude. The session feature vectors with continuously changing fission trend consistency are divided into the same memory block; The session feature vectors whose fission trend consistency changes are divided into different memory blocks; The memory blocks are arranged in the order of their formation time, and the original temporal order of the session feature vectors is maintained within each memory block. The memory blocks are arranged in a matrix to form a reorganized session memory matrix.

[0013] Optionally, step six specifically includes: Read the user interest representation data; User interest representation data is classified and statistically analyzed according to the category of the behavior object, the theme of the behavior object, and the scenario of the behavior object to form multiple interest feature sequences; Calculate the absolute value of the difference between the corresponding positional feature values ​​of any two interest feature sequences, and sum all the absolute values ​​of the differences to form the interest difference value; Construct an interest difference set based on each interest difference value; Sort the interest feature sequences in ascending order of their interest difference values; Establish interest associations between adjacent interest feature sequences after sorting, and record the corresponding interest difference values; Construct an interest association matrix based on all interest relationships.

[0014] Optionally, step seven specifically includes: Read all object records in the object repository; Based on the interest feature sequence in the interest association matrix, corresponding object records are filtered to form candidate recommendation object data; Extract the object category information, object attribute information, object price information, object theme information, and object scenario information corresponding to each candidate recommended object; A candidate object feature sequence is constructed based on object category information, object attribute information, object price information, object theme information, and object scene information; Calculate the absolute value of the difference between the corresponding position feature values ​​between the candidate object feature sequence and each interest feature sequence in the interest association matrix, and sum all the absolute values ​​of the differences to form the initial interest difference value; Read the degree of interest association between the corresponding interest feature sequences in the interest association matrix, and correct the initial interest difference value according to the degree of interest association to form the object interest difference value; Construct a set of object adaptation relationships based on the differences in interest values ​​among the objects; Establish an object adaptation relationship between the candidate recommendation objects and the corresponding interest feature sequences, and record the corresponding object interest difference values; Construct a candidate object adaptation matrix based on the adaptation relationships of all objects.

[0015] Optionally, step eight specifically includes: Read the candidate object adaptation matrix; Statistically analyze the overall fit degree of each candidate recommended object in the candidate object fit matrix; The overall fit value of the object is calculated by summing up all the fit scores corresponding to the same candidate recommendation object. Construct an object adaptation value set based on the object's comprehensive adaptation value; Candidate recommended objects are sorted in descending order of their comprehensive suitability scores to form an object sorting sequence; Read the interest feature sequence identifiers corresponding to each candidate recommendation object in the object sorting sequence; The number of candidate recommendation objects corresponding to each interest feature sequence is counted to form an interest coverage distribution sequence; The object sorting sequence is adjusted based on the interest coverage distribution sequence to form a recommended object sequence; Output personalized recommendation results based on the recommended object sequence.

[0016] The beneficial effects of this invention are: This invention addresses the shortcomings of existing recommendation methods, such as difficulty in characterizing user interest fission processes, insufficient utilization of interest continuity relationships, and limited interest coverage of recommendation results. It proposes a session segmentation strategy based on semantic change density and semantic transition position identification to achieve a structured representation of the user interest evolution process. This is achieved by constructing user behavior event sequences, user behavior semantic trajectories, session fragment sets, and session state matrices, combined with an improved STAMP model and an interest association analysis mechanism. In the interest modeling stage, a behavior fission encoding unit and a session memory reconstructing unit are introduced. The session fission intensity sequence and fission trend consistency are used to reconstruct and model user historical sessions, enabling interest switching information and interest continuity information to participate in memory construction. The process involves introducing end-behavior competition units and interest inertia maintenance units in the user interest representation stage to construct competitive behavior representations and inertial interest vectors, respectively. These are then combined with a memory attention aggregation mechanism to form a multi-dimensional interest representation structure that includes the current interest state, historical interest continuation state, and interest transition state. In the recommendation matching stage, an interest association matrix and a candidate object adaptation matrix are constructed, enabling candidate recommendation objects to simultaneously obtain direct interest matching information and related interest matching information. Finally, the recommendation results are optimized and adjusted based on the interest coverage distribution sequence, outputting personalized recommendation results that take into account both the degree of interest matching and the scope of interest coverage. This achieves refined modeling of user interest evolution patterns, in-depth mining of interest association relationships, and intelligent generation of recommendation results. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of a deep learning-based user behavior modeling and personalized recommendation method proposed in this invention; Figure 2 This is an overall structural diagram of the improved STAMP model in the deep learning-based user behavior modeling and personalized recommendation method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-2 A deep learning-based method for user behavior modeling and personalized recommendation includes the following steps: Step 1: Collect user behavior data, sort it according to the time of occurrence of the behavior, and generate a user behavior event sequence; Step 2: Extract semantic features of behavioral objects based on user behavior event sequences to generate user behavior semantic trajectories; Step 3: Calculate semantic change data between adjacent behaviors based on the semantic trajectory of user behavior, divide the user behavior event sequence according to the semantic change data, and generate a set of session fragments; Step 4: Extract session features based on the set of session fragments and construct a session state matrix; Step 5: Input the session state matrix into the improved STAMP model, perform behavioral fission encoding on the session state matrix to generate fission behavior sequences; reconstruct the session memory based on the fission behavior sequences to generate a reconstructed session memory matrix; calculate the end-behavior competition relationship based on the reconstructed session memory matrix to generate a competitive behavior representation; construct an inertial interest vector based on the competitive behavior representation and the reconstructed session memory matrix; perform memory attention aggregation on the reconstructed session memory matrix, competitive behavior representation, and inertial interest vector to generate user interest representation data; Step Six: Calculate interest associations based on user interest representation data and generate an interest association matrix; Step 7: Calculate the object adaptation relationship based on the interest association matrix and generate a candidate object adaptation matrix; Step 8: Sort the candidate objects based on the adaptation matrix and generate personalized recommendation results.

[0020] In this embodiment, step one specifically includes: Collect user browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell history generated on the target platform, and extract user identifiers, behavior object identifiers, behavior type identifiers, behavior occurrence time, visited page identifiers, dwell time, interaction location identifiers, as well as device type information, operating system information, and network environment information; Browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell history are aggregated according to user identifiers to form a user behavior record set; The user behavior record set is sorted by time based on the time the behavior occurred, forming a behavior time series; The time interval between adjacent behaviors is calculated based on the behavior time series, and the forward and backward time intervals corresponding to each behavior are recorded to form a behavior time association sequence; Based on the behavioral object identifier, the number of consecutive views, consecutive clicks, consecutive favorites, and consecutive stays are counted to form a behavioral activity sequence; Based on the behavior time association sequence and behavior activity sequence, consecutive behavior records with the same behavior object identifier, the same behavior type identifier and a time interval less than the average time interval of adjacent behaviors are merged. Data items with missing access page identifier, dwell time and device environment information are supplemented to form a standardized behavior sequence. The user behavior event sequence is generated by combining and arranging the user identifier, behavior occurrence time, behavior type identifier, behavior object identifier, behavior time association sequence, and behavior activity sequence.

[0021] In this embodiment, step two specifically involves: Read the behavior object identifier in the user behavior event sequence, associate the behavior object attribute library with the behavior object identifier, and extract object category information, object attribute information, object price information, object theme information, and object scene information; Encode object category information, object attribute information, object price information, object theme information, and object scene information to generate a set of semantic features for behavioral objects; Arrange the semantic feature sets of each behavior object according to the order of the occurrence of the behavior to form a semantic sequence of behavior objects; Calculate category change data, attribute change data, price change data, theme change data, and scene change data between adjacent behavioral objects based on the semantic sequence of behavioral objects; Based on category change data, attribute change data, price change data, theme change data, and scenario change data, semantic relationships between adjacent behavioral objects are constructed to form a set of behavioral semantic relationships; By connecting the semantic feature set of the behavior object and the corresponding semantic relationship of the behavior in the order of the behavior occurrence, a semantic trajectory of user behavior is generated.

[0022] In this implementation, the behavior object attribute library is constructed from platform business data. It records the correspondence between behavior object identifiers and object category information, object attribute information, object price information, object theme information, and object scene information. The encoding process employs a unified field mapping method, converting attribute data from different sources and in different formats into structured semantic data, ensuring that the semantic features of each behavior object have a consistent data expression form. Category change data, attribute change data, price change data, theme change data, and scene change data are all generated based on the degree of difference in the corresponding semantic features of adjacent behavior objects and are associated and recorded according to the chronological order of behavior occurrence. The behavior semantic association set stores the semantic association records between adjacent behavior objects and their corresponding change data, thus forming a continuous semantic structure that reflects the evolution of user interests, providing a unified data foundation for subsequent session segmentation, behavior fission encoding, and interest representation construction.

[0023] In this embodiment, step three specifically includes: Read the semantic sequence of behavioral objects and the set of semantic associations of behavioral objects from the semantic trajectory of user behavior; The data on category changes, attribute changes, price changes, theme changes, and scene changes between adjacent behavioral objects are normalized and then weighted and accumulated to generate semantic change values, forming a semantic change sequence. The distribution of semantic change values ​​in the semantic change sequence is statistically analyzed according to the chronological order of the occurrence of the behavior, and a semantic change distribution sequence is generated. The frequency and magnitude of semantic changes in a continuous behavior window are statistically analyzed to generate a semantic change density sequence. The semantic change levels are generated according to the sorting position of the semantic change values ​​in the semantic change distribution sequence, forming a semantic change level sequence; Based on the semantic change level sequence and semantic change density sequence, determine the semantic transition positions in the user behavior event sequence, and record the behavior object identifier and behavior occurrence time corresponding to each semantic transition position; The user behavior event sequence is divided into multiple session fragments by using each semantic transition point as a segmentation point; Arrange the conversation segments in chronological order of the events to generate a set of conversation segments.

[0024] In this implementation, semantic change values ​​correspond to the degree of change in the semantic state of a behavioral object during continuous behavior. Category change data, attribute change data, price change data, topic change data, and scene change data are all converted to a uniform scale before being included in the calculation. The continuous behavior window uses sliding analysis units composed of adjacent behaviors in the user behavior event sequence. The semantic change frequency represents the number of significant fluctuations in semantic change values ​​within the window, and the semantic change amplitude represents the difference between the maximum and minimum semantic change values ​​within the window. The semantic change density sequence is used to characterize the degree of clustering of user interest changes, and the semantic transition position corresponds to the position where the semantic structure of user behavior undergoes continuous change. High semantic consistency is maintained within conversational segments, while significant semantic differences are maintained between adjacent conversational segments. This ensures that the generated conversational segment set simultaneously retains both continuous interest features and interest switching features, providing structured input data for subsequent behavior fission encoding and conversation memory reorganization.

[0025] In this embodiment, step four specifically includes: Read the sequence of action objects and the time sequence of action occurrence for each session segment in the session segment collection; Statistically analyze the number of behavioral objects, duration of behavior, distribution of behavior types, and distribution of behavioral object categories in each session segment to generate session scale characteristics; Based on the semantic feature set of the behavioral objects corresponding to each conversation segment, the distribution of object topics, object attributes, object prices, and object scenarios are statistically analyzed to generate conversation semantic features; Based on the semantic change values ​​of adjacent behavioral objects in each conversation segment, the mean semantic change, the range of semantic change, the frequency of semantic change, and the density of semantic change are statistically analyzed to generate conversation change features. Based on the behavior occurrence time series, the average time interval, maximum time interval, and minimum time interval between adjacent behaviors in each session segment are calculated to generate session time features; A session feature vector is generated by combining and arranging session size features, session semantic features, session change features, and session time features. Arrange the session feature vectors according to the time sequence of session fragment formation to construct a session state matrix.

[0026] In this implementation, session scale features describe the behavioral activities within a session segment, session semantic features describe the semantic distribution of objects within a session segment, session change features describe the semantic evolution within a session segment, and session time features describe the temporal structure of behavior within a session segment. The session feature vector is formed by arranging these features in a uniform order, with different session segments corresponding to the same feature dimension. Semantic difference data is further calculated between adjacent session segments. This semantic difference data is generated by the degree of difference between the corresponding semantic features of adjacent session segments and is recorded as part of the session state matrix. Each row in the session state matrix corresponds to a session segment, and each column corresponds to a session feature, thus forming structured data that can simultaneously characterize the behavioral state within a session and the evolutionary relationship between sessions, providing unified input data for subsequent behavior fission coding, session memory reorganization, and interest inertia construction.

[0027] In this embodiment, step five specifically includes: Input the session state matrix into the improved STAMP model; The improved STAMP model includes sequentially connected behavioral fission coding units, conversation memory reorganization units, end-behavior competition units, interest inertia retention units, and memory attention aggregation units; The behavior fission coding unit reads each session feature vector in the session state matrix, calculates the sum of the absolute values ​​of the differences between corresponding feature elements of adjacent session feature vectors, and accumulates them in combination with the corresponding semantic change density to generate a session fission intensity sequence. Based on the session fission intensity sequence, each session feature vector is marked for fission, forming a fission-enhanced session matrix; The session memory recombination unit reads the fission enhancement session matrix and statistically analyzes the direction and magnitude of change in session fission intensity corresponding to adjacent session feature vectors. The consistency of fission trends is generated based on the direction and magnitude of change. Session feature vectors with continuously changing fission trend consistency are divided into the same memory block; session feature vectors with changing direction are divided into different memory blocks to form a recombined session memory matrix. The end-point behavior contention unit reads the end-point behavior record corresponding to each memory block; The duration of the behavior, the frequency of the behavior, and the number of behavior conversions corresponding to the terminal behavior records are counted separately, and the proportion of each statistical result to the total amount of the corresponding memory block is calculated. The competitive behavior is represented by a weighted summation based on the proportional values. The interest inertia retention unit reads and reassembles the session memory matrix; it counts the number of consecutive occurrences of the same object category, the same topic, and the same scene; Generate an inertial interest vector based on the proportion of consecutive occurrences to the total number of actions in the session; The memory attention aggregation unit reads and reconstructs the session memory matrix, competitive behavior representation, and inertial interest vector; Calculate the sum of the element-wise products between the reconstructed session memory matrix and the inertial interest vector, and the sum of the element-wise products between the competitive behavior representation and the inertial interest vector, respectively. Based on the calculation results, a memory attention distribution is generated, and the recombined session memory matrix, competitive behavior representation, and inertial interest vector are weighted and aggregated according to the memory attention distribution to form user interest representation data.

[0028] In this implementation, the improved STAMP model adopts a three-way memory structure: "fission-enhanced memory—competitive behavior memory—inertial interest memory." The fission-enhanced session matrix records the interest changes between sessions, the competitive behavior representation records behavioral patterns with high activity and conversion contribution in the current stage, and the inertial interest vector records interest evolution information that persists across multiple sessions. The memory attention distribution is not generated based on a single end behavior but is formed by comprehensively considering the correlation between the three memory data, ensuring that the user interest representation data simultaneously retains the current interest state, the historical interest continuation state, and the interest switching state. The output results of each unit maintain a unified feature dimension and are passed and updated according to the session order, thus forming a continuous memory representation structure oriented towards the user's interest evolution process, providing unified input data for subsequent interest association construction and candidate object adaptation analysis. The improved STAMP model introduces a behavior fission coding unit, a conversation memory reorganization unit, an end-behavior competition unit, and an interest inertia maintenance unit on the basis of the original STAMP model. The behavior fission coding unit uses the conversation fission intensity sequence to characterize the interest change state between adjacent conversations; the conversation memory reorganization unit constructs a reorganized conversation memory matrix based on the consistency of fission trends; the end-behavior competition unit constructs a competitive behavior representation using multiple end-behavior records; and the interest inertia maintenance unit constructs an inertial interest vector using continuous interest evolution relationships, which, together with the reorganized conversation memory matrix, participates in the memory attention aggregation calculation. By introducing a behavioral fission coding mechanism and a session memory reorganization mechanism, the changing state of interest participates in the session memory construction process; by introducing an end-behavior competition mechanism, multiple end behaviors jointly participate in the interest expression process; and by introducing an interest inertia maintenance mechanism, continuous interest evolution information participates in the user interest representation construction process. This forms a multi-dimensional interest representation structure that includes interest switching information, interest continuation information, and current behavior information, improving the continuity and stability of user interest representation.

[0029] In this embodiment, the reorganization of the session memory matrix specifically involves: Read the feature vectors of each session and the corresponding session fission intensity sequence from the fission enhancement session matrix; Arrange the feature vectors of each session in chronological order of their occurrence to form a sequence of sessions to be reconstructed. Calculate the direction and magnitude of change in fission intensity between adjacent sessions; The consistency of fission trend is determined by statistically analyzing the number of times the direction of change corresponds to the fission intensity of multiple consecutive sessions and the fluctuation range of the change amplitude. The session feature vectors with continuously changing fission trend consistency are divided into the same memory block; The session feature vectors whose fission trend consistency changes are divided into different memory blocks; The memory blocks are arranged in the order of their formation time, and the original temporal order of the session feature vectors is maintained within each memory block. The memory blocks are arranged in a matrix to form a reorganized session memory matrix.

[0030] In this implementation, the fission trend consistency is used to characterize the stability of the interest evolution state between consecutive sessions. The closer the fission trend consistency, the more consistent the corresponding session segments are in the direction and rhythm of interest change. Memory blocks are not directly divided based on the time of behavior occurrence, but are formed based on the fission evolution relationship between consecutive sessions, so that session segments in the same memory block maintain a high degree of interest continuity. The reconstructed session memory matrix uses a segmented storage method to record the session feature vectors corresponding to each memory block and the memory block order information. The original behavior evolution order is retained within the memory block, and the interest switching boundary information is retained between memory blocks, thereby simultaneously retaining the continuous evolution characteristics and interest transfer characteristics of user interests, providing structured memory data for subsequent end-behavior competition analysis, interest inertia construction, and memory attention aggregation.

[0031] In this embodiment, step six specifically includes: Read user interest representation data; User interest representation data is classified and statistically analyzed according to the category of the behavior object, the theme of the behavior object, and the scenario of the behavior object to form multiple interest feature sequences; Calculate the absolute value of the difference between the corresponding positional feature values ​​of any two interest feature sequences, and sum all the absolute values ​​of the differences to form the interest difference value; Construct an interest difference set based on each interest difference value; Sort the interest feature sequences in ascending order of their interest difference values; Establish interest associations between adjacent interest feature sequences after sorting, and record the corresponding interest difference values; Construct an interest association matrix based on all interest relationships.

[0032] In this implementation, the interest feature sequence consists of feature elements with the same interest attributes from the user interest representation data. Different interest feature sequences correspond to the user's interest evolution state in the dimensions of object category, object theme, or object scene, respectively. The interest difference value is used to characterize the degree of similarity in changes between different interest feature sequences; the smaller the interest difference value, the closer the evolutionary patterns between the corresponding interest feature sequences. The interest association matrix is ​​stored using a symmetric matrix structure, with matrix rows and columns corresponding to interest feature sequence identifiers, and matrix elements recording the degree of association between corresponding interest feature sequences. The interest association matrix not only preserves the distribution of user interests but also retains the association and evolutionary relationships between different interests, thus forming a structured interest expression result and providing unified interest association basic data for subsequent candidate object adaptation relationship calculation.

[0033] In this embodiment, step seven specifically includes: Read all object records in the object repository; Based on the interest feature sequence in the interest association matrix, corresponding object records are filtered to form candidate recommendation object data; Extract the object category information, object attribute information, object price information, object theme information, and object scenario information corresponding to each candidate recommended object; A candidate object feature sequence is constructed based on object category information, object attribute information, object price information, object theme information, and object scene information; Calculate the absolute value of the difference between the corresponding position feature values ​​between the candidate object feature sequence and each interest feature sequence in the interest association matrix, and sum all the absolute values ​​of the differences to form the initial interest difference value; Read the degree of interest association between the corresponding interest feature sequences in the interest association matrix, and correct the initial interest difference value according to the degree of interest association to form the object interest difference value; Construct a set of object adaptation relationships based on the differences in interest values ​​among the objects; Establish an object adaptation relationship between the candidate recommendation objects and the corresponding interest feature sequences, and record the corresponding object interest difference values; Construct a candidate object adaptation matrix based on the adaptation relationships of all objects.

[0034] In this implementation, the interest feature sequence corresponds to different interest directions in the user's interest representation. The object records in the object library and the behavior object attribute library are managed using a unified attribute field system, ensuring a consistent data structure between the candidate object feature sequence and the interest feature sequence. The degree of interest correlation is used to characterize the strength of the correlation between different interest directions. When the interest direction corresponding to a candidate recommendation object is correlated with the target interest direction, the corresponding interest correlation degree participates in the object interest difference value correction process, thus enabling the candidate recommendation object to simultaneously obtain both direct interest matching information and correlated interest matching information. The candidate object adaptation matrix uses a two-dimensional matrix structure to store object adaptation relationships. Matrix rows correspond to candidate recommendation object identifiers, matrix columns correspond to interest feature sequence identifiers, and matrix elements record the corresponding object interest difference value and adaptation degree, thereby forming structured adaptation data for recommending object screening.

[0035] In this embodiment, step eight specifically includes: Read the candidate object adaptation matrix; Statistically analyze the overall fit degree of each candidate recommended object in the candidate object fit matrix; The overall fit value of the object is calculated by summing up all the fit scores corresponding to the same candidate recommendation object. Construct an object adaptation value set based on the object's comprehensive adaptation value; Candidate recommended objects are sorted in descending order of their comprehensive suitability scores to form an object sorting sequence; Read the interest feature sequence identifiers corresponding to each candidate recommendation object in the object sorting sequence; The number of candidate recommendation objects corresponding to each interest feature sequence is counted to form an interest coverage distribution sequence; The object sorting sequence is adjusted based on the interest coverage distribution sequence to form a recommended object sequence; Output personalized recommendation results based on the recommended object sequence.

[0036] Interest feature sequence identifiers are used to characterize the interest directions corresponding to candidate recommendation objects, while the interest coverage distribution sequence records the distribution of each interest direction in the object ranking sequence. The overall object fit value reflects the overall matching degree between the candidate recommendation object and the user's interest structure. Based on the interest coverage distribution sequence, the distribution number of candidate recommendation objects corresponding to each interest direction is counted, and the position of the object ranking sequence is adjusted according to the overall object fit value to maintain a balanced distribution of candidate recommendation objects corresponding to different interest directions in the recommendation object sequence. The recommendation object sequence simultaneously retains highly fit objects and related interest objects, thus forming a recommendation result that balances the degree of interest matching and the scope of interest coverage, improving the ability of the recommendation result to express the user's multi-dimensional interest structure.

[0037] Example 1: To verify the feasibility of this invention in practice, it was applied to a personalized product recommendation scenario on a large-scale comprehensive e-commerce platform. This platform offers multiple product categories, including clothing, digital products, home furnishings, beauty products, sports equipment, and food and beverages. It has over 5 million registered users and approximately 1.2 million daily active users. As the number of products on the platform continues to grow and user interests constantly change, traditional recommendation systems are increasingly exhibiting problems such as insufficient ability to identify interest shifts, severe homogenization of recommended content, and narrow coverage of recommendation results. For example, some users may switch from browsing sneakers to browsing smartwatches, and then to browsing outdoor gear within a short period. Traditional recommendation models often only focus on the most recent behavior or behavioral information within a single session, resulting in recommendation results that are consistently concentrated in a single interest direction and fail to accurately reflect the evolution of user interests.

[0038] In this embodiment, the platform continuously collects user behavior data for 30 days, including browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell time. The system first constructs a user behavior event sequence based on the time of occurrence of the behavior, and then combines this with product category information, product attribute information, price information, theme information, and scenario information to construct a semantic trajectory of user behavior. Subsequently, based on the semantic changes in behavior, the system identifies the turning points in user interest, divides continuous behavior into multiple session segments, and generates a session state matrix.

[0039] To verify the technical effectiveness of this invention, it was compared with traditional collaborative filtering recommendation methods, LSTM sequence recommendation models, and the original STAMP model. The test involved 50,000 users and 300,000 products over a 30-day period. Click-through rate, add-to-cart rate, purchase conversion rate, recommendation coverage, interest matching accuracy, and user dwell time were used as evaluation metrics. The experimental results are shown in Table 1.

[0040] Table 1. Comparison of the effects of the present invention and different recommended methods.

[0041] As shown in Table 1, this invention outperforms traditional methods in several key metrics. Specifically, the click-through rate increased from 11.26% in the original STAMP model to 14.85%, an improvement of 31.88%; the purchase conversion rate increased from 2.97% to 4.63%, an improvement of 55.89%; the recommendation coverage increased from 53.12% to 72.46%, an improvement of 36.41%; the interest matching accuracy increased from 81.47% to 91.26%, an improvement of 12.02%; and the interest switching recognition accuracy increased from 78.44% to 90.83%, an improvement of 15.80%.

[0042] Further analysis reveals that this invention can accurately identify key turning points in the process of user interest changes through a behavior fission coding mechanism; it can uniformly model conversation fragments with similar interest evolution patterns through a conversation memory recombination mechanism, thereby improving interest expression capabilities; the end-behavior competition mechanism can avoid the excessive influence of a single end-behavior on recommendation results, allowing multiple high-value behaviors to participate in interest construction; the interest inertia maintenance mechanism can preserve users' long-term stable interest directions and reduce interference from short-term abnormal behaviors; the interest association matrix can establish associations between different interest directions, enabling candidate recommendations to simultaneously obtain direct interest matching information and related interest matching information; and the interest coverage distribution adjustment mechanism can enhance the diversity and coverage of recommendation results.

[0043] In summary, this invention can accurately describe the evolution of user interests, improve the accuracy of user interest modeling, enhance the matching degree between recommendation results and users' real needs, and improve the coverage and diversity of recommendation results. It has good application and promotion value in e-commerce platforms, content recommendation platforms, and Internet information service platforms.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for user behavior modeling and personalized recommendation based on deep learning, characterized in that, Includes the following steps: Step 1: Collect user behavior data, sort it according to the time of occurrence of the behavior, and generate a user behavior event sequence; Step 2: Extract semantic features of the behavior objects based on the user behavior event sequence to generate user behavior semantic trajectories; Step 3: Calculate semantic change data between adjacent behaviors based on the user behavior semantic trajectory, divide the user behavior event sequence according to the semantic change data, and generate a set of session fragments; Step 4: Extract session features based on the set of session fragments and construct a session state matrix; Step 5: Input the session state matrix into the improved STAMP model, perform behavioral fission encoding on the session state matrix, and generate fission behavior sequences; Reorganize the session memory according to the fission behavior sequence to generate a reorganized session memory matrix; Based on the recombined session memory matrix, the end-user behavior competition relationship is calculated, and a competition behavior representation is generated; Based on the competitive behavior representation and the reconstructed session memory matrix, an inertial interest vector is constructed; memory attention aggregation is performed on the reconstructed session memory matrix, the competitive behavior representation, and the inertial interest vector to generate user interest representation data; Step Six: Calculate interest associations based on the user interest representation data to generate an interest association matrix; Step 7: Calculate the object adaptation relationship based on the interest association matrix to generate a candidate object adaptation matrix; Step 8: Sort the candidate objects based on the adaptation matrix and generate personalized recommendation results.

2. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step one specifically involves: Collect user browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell history generated on the target platform, and extract user identifiers, behavior object identifiers, behavior type identifiers, behavior occurrence time, visited page identifiers, dwell time, interaction location identifiers, as well as device type information, operating system information, and network environment information; The browsing history, click history, search history, favorites history, add-to-cart history, purchase history, sharing history, comment history, and page dwell history are collected according to user identifiers to form a user behavior record set; The user behavior record set is sorted by time based on the time the behavior occurred, forming a behavior time series; The time interval between adjacent behaviors is calculated based on the behavior time series, and the forward and backward time intervals corresponding to each behavior are recorded to form a behavior time association sequence; Based on the behavioral object identifier, the number of consecutive views, consecutive clicks, consecutive favorites, and consecutive stays are counted to form a behavioral activity sequence; Based on the behavior time association sequence and behavior activity sequence, consecutive behavior records with the same behavior object identifier, the same behavior type identifier and a time interval less than the average time interval of adjacent behaviors are merged. Data items with missing access page identifier, dwell time and device environment information are supplemented to form a standardized behavior sequence. The user behavior event sequence is generated by combining and arranging the user identifier, behavior occurrence time, behavior type identifier, behavior object identifier, behavior time association sequence, and behavior activity sequence.

3. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step two specifically involves: Read the behavior object identifier in the user behavior event sequence, associate the behavior object attribute library with the behavior object identifier, and extract object category information, object attribute information, object price information, object theme information, and object scene information; Encode object category information, object attribute information, object price information, object theme information, and object scene information to generate a set of semantic features for behavioral objects; Arrange the semantic feature sets of each behavior object according to the order of the occurrence of the behavior to form a semantic sequence of behavior objects; Calculate category change data, attribute change data, price change data, theme change data, and scene change data between adjacent behavioral objects based on the semantic sequence of behavioral objects; Based on category change data, attribute change data, price change data, theme change data, and scenario change data, semantic relationships between adjacent behavioral objects are constructed to form a set of behavioral semantic relationships; By connecting the semantic feature set of the behavior object and the corresponding semantic relationship of the behavior in the order of the behavior occurrence, a semantic trajectory of user behavior is generated.

4. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step three specifically involves: Read the semantic sequence of behavioral objects and the set of behavioral semantic associations from the user's behavioral semantic trajectory; The data on category changes, attribute changes, price changes, theme changes, and scene changes between adjacent behavioral objects are normalized and then weighted and accumulated to generate semantic change values, forming a semantic change sequence. The distribution of semantic change values ​​in the semantic change sequence is statistically analyzed according to the chronological order of the occurrence of the behavior, and a semantic change distribution sequence is generated. The frequency and magnitude of semantic changes in a continuous behavior window are statistically analyzed to generate a semantic change density sequence. The semantic change levels are generated according to the sorting position of the semantic change values ​​in the semantic change distribution sequence, forming a semantic change level sequence; Based on the semantic change level sequence and semantic change density sequence, determine the semantic transition positions in the user behavior event sequence, and record the behavior object identifier and behavior occurrence time corresponding to each semantic transition position; The user behavior event sequence is divided into multiple session fragments by using each semantic transition point as a segmentation point; Arrange the conversation segments in chronological order of the events to generate a set of conversation segments.

5. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step four specifically involves: Read the sequence of behavioral objects and the time sequence of behavioral occurrences corresponding to each session segment in the set of session segments; Statistically analyze the number of behavioral objects, duration of behavior, distribution of behavior types, and distribution of behavioral object categories in each session segment to generate session scale characteristics; Based on the semantic feature set of the behavioral objects corresponding to each conversation segment, the distribution of object topics, object attributes, object prices, and object scenarios are statistically analyzed to generate conversation semantic features; Based on the semantic change values ​​of adjacent behavioral objects in each conversation segment, the mean semantic change, the range of semantic change, the frequency of semantic change, and the density of semantic change are statistically analyzed to generate conversation change features. Based on the behavior occurrence time series, calculate the average time interval, maximum time interval, and minimum time interval between adjacent behaviors in each session segment to generate session time features; A session feature vector is generated by combining and arranging session size features, session semantic features, session change features, and session time features. Arrange the session feature vectors according to the time sequence of session fragment formation to construct a session state matrix.

6. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step five specifically involves: Input the session state matrix into the improved STAMP model; The improved STAMP model includes sequentially connected behavioral fission coding units, conversation memory reorganization units, end-behavior competition units, interest inertia maintenance units, and memory attention aggregation units; The behavior fission coding unit reads each session feature vector in the session state matrix, calculates the sum of the absolute values ​​of the differences between the corresponding feature elements of adjacent session feature vectors, and performs cumulative calculations in conjunction with the corresponding semantic change density to generate a session fission intensity sequence. Based on the session fission intensity sequence, each session feature vector is marked for fission, forming a fission-enhanced session matrix; The session memory recombination unit reads the fission enhancement session matrix and statistically analyzes the direction and magnitude of change in the session fission intensity corresponding to adjacent session feature vectors. The consistency of fission trends is generated based on the direction and magnitude of change. The session feature vectors with continuously changing fission trend consistency are divided into the same memory block; The session feature vectors whose direction of change has changed are divided into different memory blocks to form a recombined session memory matrix; The end-behind behavior contention unit reads the end-behind behavior record corresponding to each memory block; The duration of the behavior, the frequency of the behavior, and the number of behavior conversions corresponding to the terminal behavior records are counted separately, and the proportion of each statistical result to the total amount of the corresponding memory block is calculated. The competitive behavior is represented by a weighted summation based on the proportional values. The interest inertia preservation unit reads and reassembles the session memory matrix; Count the number of consecutive occurrences of the same object category, the same theme, and the same scene; Generate an inertial interest vector based on the proportion of consecutive occurrences to the total number of actions in the session; The memory attention aggregation unit reads and reconstructs the session memory matrix, the competitive behavior representation, and the inertial interest vector; Calculate the sum of the element-wise products between the reconstructed session memory matrix and the inertial interest vector, and the sum of the element-wise products between the competitive behavior representation and the inertial interest vector, respectively. Based on the calculation results, a memory attention distribution is generated, and the recombined session memory matrix, competitive behavior representation, and inertial interest vector are weighted and aggregated according to the memory attention distribution to form user interest representation data.

7. The method for user behavior modeling and personalized recommendation based on deep learning according to claim 6, characterized in that, The recombined session memory matrix is ​​specifically as follows: Read the feature vectors of each session and the corresponding session fission intensity sequence from the fission enhancement session matrix; Arrange the feature vectors of each session in chronological order of their occurrence to form a sequence of sessions to be reconstructed. Calculate the direction and magnitude of change in fission intensity between adjacent sessions; The consistency of fission trend is determined by statistically analyzing the number of times the direction of change corresponds to the fission intensity of multiple consecutive sessions and the fluctuation range of the change amplitude. The session feature vectors with continuously changing fission trend consistency are divided into the same memory block; The session feature vectors whose fission trend consistency changes are divided into different memory blocks; The memory blocks are arranged in the order of their formation time, and the original temporal order of the session feature vectors is maintained within each memory block. The memory blocks are arranged in a matrix to form a reorganized session memory matrix.

8. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step six specifically involves: Read the user interest representation data; User interest representation data is classified and statistically analyzed according to the category of the behavior object, the theme of the behavior object, and the scenario of the behavior object to form multiple interest feature sequences; Calculate the absolute value of the difference between the corresponding positional feature values ​​of any two interest feature sequences, and sum all the absolute values ​​of the differences to form the interest difference value; Construct an interest difference set based on each interest difference value; Sort the interest feature sequences in ascending order of their interest difference values; Establish interest associations between adjacent interest feature sequences after sorting, and record the corresponding interest difference values; Construct an interest association matrix based on all interest relationships.

9. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step seven specifically involves: Read all object records in the object repository; Based on the interest feature sequence in the interest association matrix, corresponding object records are filtered to form candidate recommendation object data; Extract the object category information, object attribute information, object price information, object theme information, and object scenario information corresponding to each candidate recommended object; A candidate object feature sequence is constructed based on object category information, object attribute information, object price information, object theme information, and object scene information; Calculate the absolute value of the difference between the corresponding position feature values ​​between the candidate object feature sequence and each interest feature sequence in the interest association matrix, and sum all the absolute values ​​of the differences to form the initial interest difference value; Read the degree of interest association between the corresponding interest feature sequences in the interest association matrix, and correct the initial interest difference value according to the degree of interest association to form the object interest difference value; Construct a set of object adaptation relationships based on the differences in interest values ​​among the objects; Establish an object adaptation relationship between the candidate recommendation objects and the corresponding interest feature sequences, and record the corresponding object interest difference values; Construct a candidate object adaptation matrix based on the adaptation relationships of all objects.

10. The user behavior modeling and personalized recommendation method based on deep learning according to claim 1, characterized in that, Step eight specifically involves: Read the candidate object adaptation matrix; Statistically analyze the overall fit degree of each candidate recommended object in the candidate object fit matrix; The overall fit value of the object is calculated by summing up all the fit scores corresponding to the same candidate recommendation object. Construct an object adaptation value set based on the object's comprehensive adaptation value; Candidate recommended objects are sorted in descending order of their comprehensive suitability scores to form an object sorting sequence; Read the interest feature sequence identifiers corresponding to each candidate recommendation object in the object sorting sequence; The number of candidate recommendation objects corresponding to each interest feature sequence is counted to form an interest coverage distribution sequence; The object sorting sequence is adjusted based on the interest coverage distribution sequence to form a recommended object sequence; Output personalized recommendation results based on the recommended object sequence.

Images