A scene-aware personalized recommendation system and method based on temporal interest evolution

Abstract

The present invention discloses a scene-aware personalized recommendation system and method based on temporal interest evolution, the system comprising a multi-dimensional behavior data collection module, a behavior sequence analysis module, a feature engineering module and a recommendation model module. The system collects the user's explicit and implicit behavior data, scene data and content feature data, uses a time-weighted mechanism to perform pattern mining on the behavior sequence, constructs an interest state transition graph, and generates a behavior-scene association matrix based on this. Through a multi-level feature fusion strategy, a DeepFM model that fuses a factor decomposition machine and a deep neural network is used to generate personalized recommendation results. The present invention solves the technical problems that traditional recommendation systems ignore the temporal evolution characteristics of user interests and lack scene perception capabilities. The innovative introduction of time-weighted sequence pattern mining and multi-dimensional feature fusion strategies effectively captures the dynamic changes of user interests and significantly improves the accuracy and timeliness of recommendation results.

Description

Technical Field

[0001] The present invention relates to the technical field of network information recommendation, and in particular to a scene-aware personalized recommendation system and method based on temporal interest evolution. Background Art

[0002] With the rapid development of Internet technology and the popularization of online services, personalized recommendation systems have become an important tool for solving the problem of information overload. Existing recommendation systems are mainly based on traditional algorithms such as collaborative filtering and content filtering, which predict user preferences and generate recommendation results by analyzing user historical behavior data and project features. In recent years, the introduction of deep learning technology has made significant progress in feature expression and pattern recognition in recommendation systems, and various deep learning models have been widely used in the field of personalized recommendation.

[0003] However, the prior art still has the following significant deficiencies:

[0004] Therefore, how to achieve efficient collection and analysis of multi-dimensional user behavior data, how to deeply explore the temporal evolution patterns in user behavior sequences, and how to build an effective scene-aware feature fusion mechanism are technical problems that need to be solved urgently. Summary of the invention

[0005] In order to solve technical problems in existing recommendation systems such as single dimension of user behavior data, insufficient feature engineering depth, lack of scene perception ability and insufficient feature interaction modeling, the present invention provides a scene-aware personalized recommendation system and method based on temporal interest evolution.

[0006] The system of the present invention comprises:

[0007] Multi-dimensional behavior data collection module, used to collect users' explicit behavior data, implicit behavior data, scenario data and content feature data;

[0008] A behavior sequence analysis module, used to perform sequence pattern mining and interest evolution modeling on the user behavior data, and generate an interest state transition diagram describing the law of user interest evolution; wherein the behavior sequence analysis module includes a sequence pattern mining unit and an interest evolution modeling unit, the sequence pattern mining unit identifies frequent behavior patterns based on time-weighted calculation of subsequence frequency, and the interest evolution modeling unit constructs an interest state transition diagram based on the frequent behavior patterns;

[0009] A feature engineering module, used to construct a behavior-scenario association matrix based on the interest state transition diagram and generate feature representation;

[0010] The recommendation model module is used to process the feature representation and generate recommendation results using a DeepFM model that integrates a factorization machine (FM) and a deep neural network; the DeepFM model captures the second-order interaction relationship between features through the FM component and captures high-order feature combinations through the deep neural network component.

[0011] Furthermore, the data collected by the multi-dimensional behavior data collection module includes:

[0012] Explicit behavior data, which is interaction data actively generated by the user;

[0013] Implicit behavior data, which is interaction data passively generated by users;

[0014] Scenario data, which refers to the environment data when the user behavior occurs;

[0015] Content feature data, wherein the content feature data is attribute data of a user interaction object.

[0016] Furthermore, the processing of the sequential pattern mining unit includes:

[0017] Receive the explicit behavior data and implicit behavior data, the data includes the timestamp information t and behavior identification id of each user behavior, arrange the behaviors of the same user in ascending order according to the timestamp, and generate an initial behavior sequence ,in For a record of behavior;

[0018] Construct each behavior record The eigenvector of , the feature vector includes: behavior identification id, behavior occurrence time t, and user's stay time d on the corresponding content;

[0019] Divide the initial behavior sequence S into time segments of fixed length ;

[0020] In each time segment Within the segment, a sliding window of length m is used to move one position at a time to extract a subsequence consisting of all m consecutive behavior records in the segment, where m is a preset positive integer;

[0021] For each behavior record in a subsequence, the time weight is calculated using the formula:

[0022]

[0023] Where λ is the time attenuation factor with a value range of [0.1, 1.0], The difference between the occurrence time of the behavior record in the subsequence and the end time of the time segment;

[0024] Calculate the weighted frequency F(s) of subsequence s in all user behavior sequences, specifically:

[0025]

[0026] in, It is the accumulation of the time weights of all instances of the subsequence. I(s) indicates whether the subsequence s appears in the user behavior sequence. If it appears, it is 1, and if it does not appear, it is 0.

[0027] When the weighted frequency F(s) of a subsequence s exceeds a preset frequency threshold θ, the subsequence is identified as a frequent behavior pattern, where the value range of θ is [0.05, 0.5].

[0028] Furthermore, the processing of the interest evolution modeling unit includes:

[0029] The frequent behavior patterns identified by the sequence pattern mining unit are grouped into frequent pattern sets;

[0030] Constructing an interest state vector sequence based on the frequent pattern set, each interest state vector includes a normalized frequency distribution vector b of each behavior type, a cumulative duration distribution vector of each behavior, and a device type and time period mark, wherein the normalized frequency distribution vector b is the ratio of the number of occurrences of each behavior in each frequent pattern to the total number of behaviors, the cumulative duration distribution vector is the ratio of the accumulated duration of each behavior to the total duration, and the device type and time period mark are the device type with the most occurrences in the frequent pattern and its corresponding time period;

[0031] An interest state transition graph is constructed according to the interest state vector sequence, whose nodes are all interest state vectors if and only if adjacent states satisfy the behavior type distribution overlap rate. When the value is greater than a preset threshold and has the same device type and time period tag, a state transfer edge is constructed, wherein the calculation formula for the behavior type distribution overlap rate is:

[0032] in, represents the number of common behavior types, Indicates the total number of non-repeated behavior types;

[0033] Furthermore, the feature engineering module adopts a multi-level feature fusion strategy, including:

[0034] The ratio of the number of continuous edges to the total number of time segments is calculated from the state transition graph as the state stability. The behavior type with the highest frequency in each state is extracted as the dominant behavior type. A two-dimensional behavior-scenario association matrix is ​​constructed, where the rows of the matrix represent the dominant behavior type, the columns represent the combination of device type and time period, and the matrix element value is the number of occurrences of the corresponding dominant behavior type under a specific combination of device type and time period.

[0035] The behavior-scenario association matrix is ​​normalized to the maximum and minimum values, and the normalization formula is:

[0036]

[0037] in, is the element value corresponding to the combination of the i-th behavior type and the j-th scenario in the matrix, M is the entire behavior-scenario association matrix, min(M) represents the minimum element in the matrix M, and max(M) represents the maximum element in the matrix M;

[0038] The normalized element value is weighted and adjusted based on the state stability of time segment k. The weighted formula is:

[0039]

[0040] in, is the final value after weighted adjustment, is the normalized element value, is the state stability of time segment k, α is the adjustment factor in the range of [0, 1];

[0041] The weighted two-dimensional matrix is ​​transformed according to the feature domain mapping rule, the behavior type is mapped to domain id=1, the device type is mapped to domain id=2, and the time period is mapped to domain id=3, and a sparse feature representation in the format of <domain id, unique identifier in the value domain, weighted feature value> is generated.

[0042] Furthermore, the recommendation model module adopts the DeepFM deep learning model, and the processing process includes:

[0043] Constructing a two-way feature interaction network, including an FM component and a Deep component of feature interaction, wherein the two-way feature interaction network receives the sparse feature representation <domain id, unique identifier in the value domain, weighted feature value>;

[0044] For the FM component, the inner product operation of the feature latent vector is used to model the second-order interaction relationship between features, and each feature i is mapped to a k-dimensional latent vector , calculate the latent vector corresponding to any two different features i and j and The inner product of As the interaction strength, the interaction of all feature pairs (i, j) is accumulated to obtain the output result of the second-order feature interaction , the specific calculation formula is:

[0045]

[0046] Where n is the total number of features, and Respectively represent the k-dimensional latent vectors corresponding to the i-th feature and the j-th feature, Represents the latent vector and The inner product operation of Represents the cumulative result of the interaction strength of all feature pairs;

[0047] For the Deep component, the sparse features are mapped into dense vectors of corresponding dimensions through the embedding layer and then concatenated into the input vector , extract high-order feature combinations through L-layer fully connected neural networks, The calculation formula of the layer is:

[0048]

[0049] in, is the input layer, is the output of layer l, is the weight matrix, is the bias vector, is the activation function, , output the last layer As a high-order feature combination output ;

[0050] The second-order feature interaction output of the FM component Combined with the high-order features of the Deep component output Add up to get the final prediction score:

[0051]

[0052] The predicted score is converted into a predicted probability value through the sigmoid function, and the formula is:

[0053]

[0054] Among them, p represents the predicted probability value, and y is the original prediction score of the model;

[0055] The candidate content item set is sorted in descending order based on the predicted probability value, and the first N content items after sorting are intercepted to form a recommendation result set , where N is the number of recommended results preset by the system.

[0056] The method of the present invention comprises the following steps:

[0057] Step 1: Obtain multi-dimensional behavior data of the user, wherein the multi-dimensional behavior data includes explicit behavior data, implicit behavior data, scenario data, and content feature data;

[0058] Step 2: Performing sequential pattern mining and interest evolution modeling on the multidimensional behavior data, extracting frequent behavior patterns using a time-weighted mechanism and generating an interest state transition diagram;

[0059] Step 3: construct a behavior-scenario two-dimensional association matrix according to the interest state transition diagram, and convert the two-dimensional association matrix into a feature representation;

[0060] Step 4: Input the feature representation into the DeepFM deep learning model and output personalized recommendation results.

[0061] The present invention has the following beneficial effects:

[0062] (1) The present invention designs a multi-dimensional behavior data collection module, which not only collects the user's explicit behavior (such as clicks, favorites), but also includes implicit behavior (such as browsing time, operation sequence) and scenario data (such as time, location, device), thereby realizing a comprehensive perception of user behavior characteristics and overcoming the deficiency of the single dimension of user behavior data in the prior art;

[0063] (2) The present invention innovatively introduces a time-weighted sequence pattern mining mechanism. By calculating the subsequence frequency of time decay, identifying frequent behavior patterns, and constructing an interest state transition graph, the system can effectively capture and predict the dynamic evolution of user interests, greatly improving the ability of the recommendation system to adapt to changes in user interests.

[0064] (3) The present invention achieves accurate matching of behavior patterns and usage scenarios by constructing a two-dimensional behavior-scenario association matrix and combining it with a weighted adjustment mechanism for state stability, significantly improving the adaptability of the recommendation system in different scenarios and making the recommendation results more in line with the actual needs of users in specific scenarios;

[0065] (4) The present invention adopts the DeepFM deep learning model, which efficiently models the second-order interaction relationship between features through the FM component, and uses a deep neural network to capture high-order feature combinations, thereby realizing multi-level modeling of feature interactions. Compared with traditional methods, it has stronger feature expression and combination capabilities, and significantly improves the accuracy of recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the implementation of the present invention or the existing technical solutions, the following briefly introduces the drawings required to be used in the embodiments or the description of the prior art.

[0067] Figure 1 Flowchart of the scene-aware personalized recommendation method based on temporal interest evolution.

[0068] Figure 2 This is the architecture diagram of the scene-aware personalized recommendation system based on temporal interest evolution.

[0069] Figure 3 This is the DeepFM model architecture diagram. DETAILED DESCRIPTION

[0070] The following is a clear and complete description of the technical solutions in the embodiments of the invention in conjunction with the drawings in the embodiments of the invention. Obviously, the described embodiments are only part of the embodiments of the invention, not all of the embodiments. Based on the embodiments of the invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the invention.

[0071] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0072] like Figure 1 As shown, the embodiment of the present invention discloses a scene-aware personalized recommendation method based on temporal interest evolution, comprising the following steps:

[0073] S101: Acquire multi-dimensional behavior data of a user, wherein the multi-dimensional behavior data includes explicit behavior data, implicit behavior data, scenario data, and content feature data;

[0074] S102: performing sequence pattern mining on the multidimensional behavior data, calculating subsequence frequencies based on a time weighting mechanism and identifying frequent behavior patterns;

[0075] S103: performing interest evolution modeling based on the frequent behavior pattern, constructing an interest state vector sequence and generating an interest state transition diagram;

[0076] S104: constructing a behavior-scenario two-dimensional association matrix according to the interest state transition diagram, and converting the two-dimensional association matrix into a feature representation;

[0077] S105: Input the feature representation into the DeepFM deep learning model and output personalized recommendation results.

[0078] In step S101, the collection of multi-dimensional behavior data is specifically implemented as follows:

[0079] The explicit behavior data of users is collected through the client SDK and server API. Explicit behavior data refers to the interactive behavior actively generated by users, including: click behavior (recording the content ID, timestamp, source page, and dwell time of the user's click), collection behavior (recording collection time and collection category), sharing behavior (recording sharing platform and sharing time), comment behavior (recording comment text, comment time, and comment length), and purchase behavior (recording purchase time and purchase quantity), etc. These explicit behavior data are collected and stored in real time through the front-end embedding point and back-end log system.

[0080] The user's implicit behavior data is collected through front-end tracking technology. Implicit behavior data refers to the behavior characteristics that users passively generate during use, including browsing tracks (recording the order of page visits and the length of stay), scrolling behavior (recording the scrolling depth and the page pause position), mouse tracks (recording the hovering position and the hovering duration), search behavior (recording the search keywords and the search time), and playback behavior (recording the video playback progress and fast forward and rewind operations).

[0081] Collect scenario data when user behavior occurs. Scenario data describes the user's usage environment from multiple dimensions, including: time dimension (weekdays / weekends, morning, noon and evening hours), geographic dimension (user's city, region type), device dimension (device type, network status) and occasion dimension (work / leisure scenario marker).

[0082] Extract feature data from content objects that users interact with. Content feature data includes: basic attributes (title, description, release time), category tags (first-level categories, second-level categories), content indicators (number of reads, number of likes, number of comments), and quality features (text length, number of pictures, video length), etc.

[0083] After all types of data are collected, these multi-dimensional data are uniformly stored in a distributed database, and an index of user dimensions is established to provide a data basis for subsequent behavior sequence analysis and feature engineering. In the data storage process, a layered storage strategy is adopted: hot data is stored in the memory database to support real-time processing, and cold data is transferred to the distributed file system for archiving. At the same time, data quality monitoring is carried out, including data integrity check, outlier detection and data consistency verification, to ensure the quality of collected data.

[0084] In step S102, the specific method of sequential pattern mining is as follows:

[0085] Preprocess the received user behavior data. To improve processing efficiency, the system filters and cleans the original behavior data: filter out click behaviors with a stay time of less than 3 seconds, merge repeated behaviors with an interval time of less than 5 seconds, and remove obviously abnormal behavior records (such as repeated clicks for more than 100 consecutive times). The preprocessed behavior data contains the timestamp t and the behavior identifier id, and is arranged in ascending order by timestamp to generate the initial behavior sequence ,in For a record of behavior;

[0086] Construct the feature vector of the behavior record. Each behavior record The eigenvector of It includes three key dimensions: the behavior identification id is encoded with integers (such as click behavior = 1, favorite behavior = 2, share behavior = 3, etc.); the time when the behavior occurs t is uniformly converted into a timestamp format; the user's stay time d is recorded in seconds and logarithmically transformed to reduce numerical differences.

[0087] Divide the initial behavior sequence S into time segments of fixed length , taking into account the timeliness characteristics of user behavior, for example, for active users (average number of daily behaviors>50), select 2 hours as a time segment; for moderately active users (10<average number of daily behaviors≤50), select 4 hours as a time segment; for low-active users (average number of daily behaviors≤10), select 8 hours as a time segment.

[0088] Sliding window processing is performed in each time segment. The selection of window length m is related to the content type. For example, m=3 is selected for news information; m=5 is selected for video entertainment; and m=7 is selected for e-commerce shopping. Each time the window moves one position, a subsequence of m consecutive behavior records is extracted. Each time it moves one position, a subsequence of all m consecutive behavior records in the segment is extracted, where m is a preset positive integer.

[0089] For each behavior record in a subsequence, the time weight is calculated using the formula:

[0090]

[0091] in, is the difference between the occurrence time of the behavior record in the subsequence and the end time of the time segment, in hours, and λ is a time decay factor in the range of [0.1, 1.0]. For example, for real-time content, λ=0.8, for daily content, λ=0.5, and for long-term content, λ=0.2.

[0092] Calculate the weighted frequency F(s) of subsequence s in all user behavior sequences, specifically:

[0093]

[0094] in, It is the accumulation of the time weights of all instances of the subsequence. I(s) indicates whether the subsequence s appears in the user behavior sequence. If it appears, it is 1, and if it does not appear, it is 0.

[0095] When the weighted frequency F(s) of a subsequence s exceeds a preset frequency threshold θ, the subsequence is identified as a frequent behavior pattern, where the value range of θ is [0.05, 0.5].

[0096] In step S103, interest evolution modeling includes the following specific steps:

[0097] Based on the frequent pattern set, an interest state vector sequence is constructed. Each interest state vector includes a normalized frequency distribution vector b of each behavior type, a cumulative duration distribution vector of each behavior, and a device type and time period mark. The normalized frequency distribution vector b is the ratio of the number of occurrences of each behavior in each frequent pattern to the total number of behaviors. For example, if a frequent pattern contains 10 behaviors (6 views, 3 favorites, and 1 share), its normalized frequency distribution vector is [0.6, 0.3, 0.1]. The cumulative duration distribution vector is the ratio of the cumulative duration of each behavior to the total duration. The scene mark is to extract the device type (such as mobile terminal / PC terminal) that appears most frequently in the frequent pattern and its corresponding time period mark (such as weekday morning / weekend evening).

[0098] Construct an interest state transition graph, and use each interest state vector as a node in the graph. The node attributes include state ID, behavior distribution, duration distribution, and scene tag. If and only if adjacent states satisfy the behavior type distribution overlap rate When the value is greater than the preset threshold and has the same device type and time period tag, a state transition edge is constructed, where the behavior type distribution overlap rate calculation formula is:

[0099] in, represents the number of common behavior types, Indicates the total number of non-repetitive behavior types; for example, if the behavior distribution of state i includes [browse, favorite, share], and the behavior distribution of state i+1 includes [browse, favorite, comment], then the number of common behaviors is 2 (browse, favorite), the total number of behavior types is 4 (browse, favorite, share, comment), and the overlap rate is 0.5. When the overlap rate of the behavior distribution of adjacent states exceeds the preset threshold (recommended to be set between 0.3-0.5) and has the same device type tag and continuous or similar time period tags, a transfer edge is established between the two states.

[0100] In step S104, constructing the behavior-scenario two-dimensional association matrix includes the following specific steps:

[0101] The ratio of the number of continuous edges to the total number of time segments is calculated from the state transition graph as the state stability. The specific calculation method is: for each time segment k, the number of continuous edges Ek in the state transition graph in the segment is counted, and the stability sk=Ek / N is calculated, where N is the total number of states in the time segment. If sk≥0.7, the user interest state of the time segment is considered to be relatively stable.

[0102] The behavior type with the highest frequency in each state is extracted as the dominant behavior type. For example, when the frequency of a certain type of behavior accounts for more than 60%, it is directly determined as the dominant behavior; when the highest frequency behavior accounts for between 40% and 60%, the average stay time of the behavior is also considered. If the average stay time exceeds 1.5 times the average value of the state, it is also determined as the dominant behavior; if no behavior meets the above conditions, the first two types of behaviors with the highest frequency are taken together as the dominant behaviors.

[0103] A two-dimensional behavior-scenario association matrix is ​​constructed, in which the rows of the matrix represent the dominant behavior types, the columns represent the combination of device type and time period, and the matrix element values ​​are the number of occurrences of the corresponding dominant behavior type under a specific combination of device type and time period; in order to balance the data sparsity problem under different scenarios, smoothing processing is introduced. When the number of samples under a certain scenario combination is too small (less than 1% of the total sample size), the average number of occurrences of the behavior type in similar scenarios is used to fill in the gaps.

[0104] Taking into account the differences in data magnitudes of different behavior types and scene combinations, the improved maximum and minimum normalization method is used to perform maximum and minimum normalization on the behavior-scene association matrix. The normalization formula is:

[0105]

[0106] in, is the element value corresponding to the i-th behavior type and the j-th scenario combination in the matrix, M is the entire behavior-scenario association matrix, min(M) represents the minimum element in the matrix M, max(M) represents the maximum element in the matrix M, and each column is normalized to ensure a more balanced weight distribution of different behavior types and scenario combinations.

[0107] The normalized element value is weighted and adjusted based on the state stability of time segment k. The weighted formula is:

[0108]

[0109] in, is the final value after weighted adjustment, is the normalized element value, is the state stability of time segment k, and α is an adjustment factor in the range of [0, 1]. For example, for scenarios that require rapid response to changes in user interests, set α=0.8; for scenarios that prefer stable recommendations, set α=0.5; for scenarios that take into account both new and old content, set α=0.3.

[0110] The weighted two-dimensional matrix is ​​transformed according to the feature domain mapping rule, the behavior type is mapped to domain id=1, the device type is mapped to domain id=2, and the time period is mapped to domain id=3, and a sparse feature representation in the format of <domain id, unique identifier in the value domain, weighted feature value> is generated.

[0111] In step S105, refer to Figure 3 , outputting personalized recommendation results based on the DeepFM deep learning model includes the following specific steps:

[0112] Constructing a two-way feature interaction network, including an FM component and a Deep component of feature interaction, wherein the two-way feature interaction network receives the sparse feature representation <domain id, unique identifier in the value domain, weighted feature value>;

[0113] For the FM component, the inner product operation of the feature latent vector is used to model the second-order interaction relationship between features, and each feature i is mapped to a k-dimensional latent vector , calculate the latent vector corresponding to any two different features i and j and The inner product of As the interaction strength, the interaction of all feature pairs (i, j) is accumulated to obtain the output result of the second-order feature interaction , the specific calculation formula is:

[0114]

[0115] Where n is the total number of features, and Respectively represent the k-dimensional latent vectors corresponding to the i-th feature and the j-th feature, Represents the latent vector and The inner product operation of Represents the cumulative result of the interaction strength of all feature pairs;

[0116] For the Deep component, the sparse features are mapped into dense vectors of corresponding dimensions through the embedding layer and then concatenated into the input vector , extract high-order feature combinations through L-layer fully connected neural networks, The calculation formula of the layer is:

[0117]

[0118] in, is the input layer, is the output of layer l, is the weight matrix, is the bias vector, is the activation function, , output the last layer As a high-order feature combination output ;

[0119] The second-order feature interaction output of the FM component Combined with the high-order features of the Deep component output Add up to get the final prediction score:

[0120]

[0121] The predicted score is converted into a predicted probability value through the sigmoid function, and the formula is:

[0122]

[0123] Among them, p represents the predicted probability value, and y is the original prediction score of the model;

[0124] The candidate content item set is sorted in descending order based on the predicted probability value, and the first N content items after sorting are intercepted to form a recommendation result set , where N is the number of recommended results preset by the system.

[0125] like Figure 2 As shown, corresponding to the above method, the embodiment of the present invention further discloses an intelligent recommendation system based on user behavior analysis, including:

[0126] Multi-dimensional behavior data collection module 101: used to collect the user's explicit behavior data, implicit behavior data, scenario data and content feature data;

[0127] Behavior sequence analysis module 102: used for performing sequence pattern mining and interest evolution modeling on the user behavior data;

[0128] Feature engineering module 103: used to construct a behavior-scenario association matrix based on the interest state transition diagram and generate feature representation;

[0129] Recommendation model module 104: used to process the feature representation using the DeepFM deep learning model and generate recommendation results.

[0130] Furthermore, the behavior sequence analysis module 102 includes:

[0131] Sequence pattern mining unit 1021: identifying frequent behavior patterns based on time-weighted calculation of subsequence frequencies;

[0132] Interest evolution modeling unit 1022: constructs an interest state transition diagram based on the frequent behavior pattern.

[0133] Furthermore, the feature engineering module 103 includes:

[0134] The association matrix construction unit 1031 is used to extract the dominant behavior type from the state transition diagram and construct a behavior-scenario two-dimensional association matrix;

[0135] Feature representation generating unit 1032: used to perform normalization and weighting processing on the correlation matrix to generate a sparse feature representation.

[0136] Furthermore, the recommendation model module 104 includes:

[0137] FM component 1041: used to capture the second-order feature interaction relationship through the feature latent vector inner product operation;

[0138] Deep component 1042: used to extract high-order feature combinations through a multi-layer neural network.

[0139] The recommendation result generating unit 1043 is used to sort the candidate content item set in descending order based on the predicted probability, and extract the first N content items after sorting to form a recommendation result set.

[0140] The above embodiment is only one of the preferred implementation modes of the present invention. Those skilled in the art may optimize or modify it without violating the spirit of the present invention, but such optimization or modification shall not be regarded as limiting the present invention.

Claims

1. A scene-aware personalized recommendation system based on temporal interest evolution, characterized in that: include: A multi-dimensional behavior data collection module is used to collect user behavior data, wherein the behavior data includes explicit behavior data, implicit behavior data, scenario data and content feature data, wherein the explicit behavior data is the interaction data actively generated by the user, the implicit behavior data is the interaction data passively generated by the user, the scenario data is the environment data when the user behavior occurs, and the content feature data is the attribute data of the user interaction object; A behavior sequence analysis module, used to perform sequence pattern mining and interest evolution modeling on the user behavior data, and generate an interest state transition diagram describing the law of user interest evolution; A feature engineering module, used to calculate the ratio of the number of consecutive appearances of state transition edges to the total number of time segments from the interest state transition graph as state stability, construct a behavior-scenario two-dimensional association matrix, and perform weighted adjustment on the behavior-scenario two-dimensional association matrix based on the state stability to generate feature representation; A recommendation model module, used to process the feature representation and generate a recommendation result by using a DeepFM model that combines a factorization machine and a deep neural network; The behavior sequence analysis module includes a sequence pattern mining unit and an interest evolution modeling unit. The sequence pattern mining unit calculates the time weight of the behavior record in the subsequence through a time decay function, and calculates the weighted frequency of the subsequence based on the time weight. When the weighted frequency exceeds a preset frequency threshold, it is identified as a frequent behavior pattern. The interest evolution modeling unit constructs an interest state vector sequence based on the frequent behavior pattern, wherein each interest state vector includes a normalized frequency distribution vector of the behavior type, a cumulative duration distribution vector, and a device type and time period mark, and calculates the behavior type distribution overlap rate between adjacent states of the interest state vector, and establishes the state transition edge only when the behavior type distribution overlap rate exceeds a preset threshold and has the same device type and time period mark; The DeepFM model captures the second-order interaction relationship between features through the FM component, captures the high-order feature combination through the deep neural network component, maps the behavior type, device type and time period type to different domain identifiers through the feature domain mapping rules, and generates a sparse feature representation including the domain identifier, value range identifier and weighted feature value, and embeds and maps the features of different domains into dense vectors of corresponding dimensions and then splices them.

2. The scene-aware personalized recommendation system based on temporal interest evolution according to claim 1, characterized in that: The processing process of the sequence pattern mining unit includes: Receive the explicit behavior data and implicit behavior data, the data includes the timestamp information t and behavior identification id of each user behavior, arrange the behaviors of the same user in ascending order according to the timestamp, and generate an initial behavior sequence ,in For a record of behavior; Construct each behavior record The eigenvector of , the feature vector includes: behavior identification id, behavior occurrence time t, and user's stay time d on the corresponding content; Divide the initial behavior sequence S into time segments of fixed length ; In each time segment Within the segment, a sliding window of length m is used to move one position at a time to extract a subsequence consisting of all m consecutive behavior records in the segment, where m is a preset positive integer; For the behavior record in each subsequence, the time weight is calculated by the time decay function, and the formula is: Where λ is the time attenuation factor with a value range of [0.1, 1.0], The difference between the occurrence time of the behavior record in the subsequence and the end time of the time segment; Calculate the weighted frequency F(s) of the subsequence s appearing in all user behavior sequences, specifically: in, is the accumulation of the time weights of all occurrences of the subsequence, I(s) indicates whether the subsequence s appears in the user behavior sequence, 1 if it appears, and 0 if it does not appear; When the weighted frequency F(s) of the subsequence s exceeds the preset frequency threshold θ, the subsequence is identified as the frequent behavior pattern, where the value range of θ is [0.05, 0.5].

3. The scene-aware personalized recommendation system based on temporal interest evolution according to claim 1, characterized in that: The processing process of the interest evolution modeling unit includes: The frequent behavior patterns identified by the sequence pattern mining unit are grouped into a frequent pattern set; Constructing the interest state vector sequence based on the frequent pattern set, each interest state vector includes the normalized frequency distribution vector b of each behavior type, the cumulative duration distribution vector of each behavior, and the device type and the time period mark, wherein the normalized frequency distribution vector b is the ratio of the number of occurrences of each behavior in each frequent pattern to the total number of behaviors, the cumulative duration distribution vector is the ratio of the accumulated duration of each behavior to the total duration, and the device type and time period mark are the device type with the most occurrences in the frequent pattern and its corresponding time period; An interest state transition graph is constructed according to the interest state vector sequence, whose nodes are all interest state vectors if and only if adjacent states satisfy the behavior type distribution overlap rate. When the value is greater than the preset threshold and has the same device type and time period tag, the state transition edge is constructed, wherein the behavior type distribution overlap rate calculation formula is: in, represents the number of common behavior types, Indicates the total number of unique behavior types.

4. The scene-aware personalized recommendation system based on temporal interest evolution according to claim 1, characterized in that: The feature engineering module adopts a multi-level feature fusion strategy, specifically including: The ratio of the number of continuous edges to the total number of time segments is calculated from the state transition graph as the state stability, the behavior type with the highest frequency in each state is extracted as the dominant behavior type, and a behavior-scenario two-dimensional association matrix is ​​constructed, wherein the rows of the matrix represent the dominant behavior type, the columns represent the combination of device type and time period, and the matrix element value is the number of occurrences of the corresponding dominant behavior type under a specific device type and time period combination; The behavior-scenario association matrix is ​​normalized to the maximum and minimum values, and the normalization formula is: in, is the element value corresponding to the combination of the i-th behavior type and the j-th scenario in the matrix, M is the entire behavior-scenario association matrix, min(M) represents the minimum element in the matrix M, and max(M) represents the maximum element in the matrix M; The normalized element value is weighted and adjusted based on the state stability of the time segment k, and the weighted formula is: in, is the final value after weighted adjustment, is the normalized element value, is the state stability of time segment k, α is a regulation factor in the range of [0, 1]; The weighted two-dimensional matrix is ​​transformed according to the feature domain mapping rule, the behavior type is mapped to domain id=1, the device type is mapped to domain id=2, and the time period is mapped to domain id=3, and the sparse feature representation in the format of <domain id, unique identifier in the value domain, weighted feature value> is generated.

5. The scene-aware personalized recommendation system based on temporal interest evolution according to claim 1, characterized in that: The recommendation model module adopts the DeepFM deep learning model, and the processing process includes: Constructing a two-way feature interaction network, including an FM component and a Deep component of feature interaction, wherein the two-way feature interaction network receives the sparse feature representation <domain id, unique identifier in the value domain, weighted feature value>; For the FM component, the inner product operation of the feature latent vector is used to model the second-order interaction relationship between features, and each feature i is mapped to a k-dimensional latent vector , calculate the latent vector corresponding to any two different features i and j and The inner product of As the interaction strength, the interaction of all feature pairs (i, j) is accumulated to obtain the output result of the second-order feature interaction , the specific calculation formula is: Where n is the total number of features, and Respectively represent the k-dimensional latent vectors corresponding to the i-th feature and the j-th feature, Represents the latent vector and The inner product operation of Represents the cumulative result of the interaction strength of all feature pairs; For the Deep component, the sparse features are mapped to the dense vectors of the corresponding dimensions through the embedding layer and then concatenated into the input vector , extract the high-order feature combination through L-layer fully connected neural network, The calculation formula of the layer is: in, is the input layer, is the output of layer l, is the weight matrix, is the bias vector, is the activation function, , output the last layer As the high-order feature combination output ; The second-order feature interaction output of the FM component Combined with the high-order features of the Deep component and output Add up to get the final prediction score: The predicted score is converted into a predicted probability value through the sigmoid function, and the formula is: Among them, p represents the predicted probability value, and y is the original prediction score of the model; The candidate content item set is sorted in descending order based on the predicted probability value, and the first N content items after sorting are intercepted to form a recommendation result set , where N is the number of recommended results preset by the system.

6. A scene-aware personalized recommendation method based on temporal interest evolution, characterized in that: The following steps are involved: Step 1: Obtain multi-dimensional behavior data of the user, wherein the multi-dimensional behavior data includes explicit behavior data, implicit behavior data, scenario data and content feature data, wherein the explicit behavior data is the interaction data actively generated by the user, the implicit behavior data is the interaction data passively generated by the user, the scenario data is the environment data when the user behavior occurs, and the content feature data is the attribute data of the user interaction object; Step 2: Perform sequence pattern mining and interest evolution modeling on the multidimensional behavior data, calculate the time weight of the behavior record in the subsequence through the time decay function, and calculate the weighted frequency of the subsequence based on the time weight. When the weighted frequency exceeds a preset frequency threshold, it is identified as a frequent behavior pattern. Based on the frequent behavior pattern, a sequence of interest state vectors is constructed, wherein each interest state vector includes a normalized frequency distribution vector of the behavior type, a cumulative duration distribution vector, and a device type and time period mark. The behavior type distribution overlap rate between adjacent states of the interest state vector is calculated. A state transition edge is established only when the behavior type distribution overlap rate exceeds a preset threshold and has the same device type and time period mark, and an interest state transition graph is generated; Step 3: Calculate the ratio of the number of consecutive appearances of state transition edges to the total number of time segments from the interest state transition graph as the state stability, construct a behavior-scene two-dimensional association matrix, and perform weighted adjustment on the behavior-scene two-dimensional association matrix based on the state stability to generate feature representation; Step 4: Use the DeepFM model that integrates a factorization machine and a deep neural network to process the feature representation. Use the feature domain mapping rule to map the behavior type, device type, and time period type to different domain identifiers, and generate a sparse feature representation containing domain identifiers, value range identifiers, and weighted feature values ​​to distinguish the interaction relationship between same-domain and cross-domain features. By embedding, mapping, and concatenating features from different domains, high-dimensional sparse features are converted into dense vector representations of uniform dimension. The second-order interaction relationship between features is captured by the FM component, and the high-order feature combination is captured by the deep neural network component to output personalized recommendation results.

Images