Article recommendation method and device, equipment, medium and product

Abstract

The invention discloses an article recommendation method, device and equipment, a medium and a product. The method comprises the following steps: associating a terminal behavior data sequence of a user with behavior characteristic attributes of the user under different transaction categories to obtain corresponding behavior data subsequences of the user under the different transaction categories; determining a transaction behavior floating attribute based on the behavior data subsequences corresponding to the user under different transaction categories, and determining a transaction behavior floating range based on the recessive weight coefficient and the behavior data subsequences corresponding to the user under different transaction categories; and determining a transaction trend type corresponding to the user based on the transaction behavior floating attribute and the transaction behavior floating range, and inputting the at least one behavior data sub-sequence into a recommendation model matched with the transaction trend type to obtain a target recommended article. According to the technical scheme, high adaptability between the recommended article and the current behavior state and potential demand of the user is ensured, personalized recommendation is realized, and the user satisfaction is improved.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to a method, apparatus, device, medium, and product for recommending items. Background Technology

[0002] With the widespread adoption of the internet and mobile devices, users generate a vast amount of behavioral data in online services such as e-commerce, social media, and video platforms. This data exhibits clear time-series characteristics, reflecting the continuous changes in users' interests and preferences over time. To alleviate information overload, recommender systems are widely used to provide users with personalized content and services.

[0003] Existing recommendation methods are mainly based on modeling user behavior sequences, using deep neural networks to learn the implicit features of users' historical behavior and predict their next potentially interesting item. However, these methods typically treat user behavior as a linear time series, failing to effectively identify users' periodic behavioral patterns. This leads to inaccurate identification of user interests and affects the relevance between recommendation results and users. Summary of the Invention

[0004] This invention provides a method, apparatus, device, medium, and product for recommending items, in order to improve the accuracy of identifying dynamic user behavior trends, ensure a high degree of adaptability between recommended items and user needs, achieve personalized recommendations, and enhance user satisfaction.

[0005] According to one aspect of the present invention, an item recommendation method is provided, the method comprising:

[0006] Acquire terminal behavior data sequences corresponding to users whose recommended items are to be determined, as well as behavioral feature attributes of the users under different transaction categories; wherein, the behavioral feature attributes are used to characterize the degree of preference of the users in the corresponding transaction categories;

[0007] By associating the behavioral feature attributes with the terminal behavioral data sequence, the user's behavioral data sub-sequences under different transaction categories are obtained;

[0008] Based on the user's behavioral data subsequences under different transaction categories, the transaction behavior fluctuation attribute is determined, and based on the predetermined implicit weight coefficients and the user's behavioral data subsequences under different transaction categories, the transaction behavior fluctuation range is determined.

[0009] Based on the transaction behavior fluctuation attributes and the transaction behavior fluctuation range, the transaction trend type corresponding to the user is determined, and at least one of the behavior data subsequences is input into the recommendation model that matches the transaction trend type to obtain the target recommended item corresponding to the user.

[0010] According to another aspect of the present invention, an item recommendation device is provided, the device comprising:

[0011] The data acquisition module is used to acquire terminal behavior data sequences corresponding to users of items to be recommended, as well as behavioral feature attributes of the users under different transaction categories; wherein, the behavioral feature attributes are used to characterize the degree of preference of the users in the corresponding transaction categories;

[0012] The behavior data subsequence determination module is used to associate the behavior feature attributes and the terminal behavior data sequence to obtain the behavior data subsequence corresponding to the user under different transaction categories;

[0013] The floating information determination module is used to determine the floating attribute of transaction behavior based on the subsequence of behavioral data corresponding to the user under different transaction categories, and to determine the floating range of transaction behavior based on the predetermined implicit weight coefficient and the subsequence of behavioral data corresponding to the user under different transaction categories.

[0014] The target recommended item determination module is used to determine the transaction trend type corresponding to the user based on the transaction behavior fluctuation attribute and the transaction behavior fluctuation range, and input at least one of the behavior data subsequences into the recommendation model that matches the transaction trend type to obtain the target recommended item corresponding to the user.

[0015] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0016] At least one processor; and a memory communicatively connected to said at least one processor; wherein,

[0017] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the item recommendation method according to any embodiment of the present invention.

[0018] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the item recommendation method according to any embodiment of the present invention.

[0019] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the item recommendation method as described in any embodiment of the present invention.

[0020] The technical solution of this invention obtains terminal behavior data sequences corresponding to users of items to be recommended, as well as behavioral feature attributes of users under different transaction categories. The behavioral feature attributes characterize the user's preference level within the corresponding transaction category. The behavioral feature attributes and terminal behavior data sequences are correlated to obtain subsequences of user behavior data under different transaction categories. Based on these subsequences, a transaction behavior fluctuation attribute is determined. A transaction behavior fluctuation range is determined based on pre-determined implicit weight coefficients and the subsequences. A transaction trend type is determined based on the transaction behavior fluctuation attribute and the transaction behavior fluctuation range. At least one subsequence of behavior data is input into a recommendation model matching the transaction trend type to obtain the target recommended item corresponding to the user. This solution addresses the problem in existing technologies that learn implicit features of user historical behavior through deep neural networks, which can lead to issues with understanding user behavior. To address the issue of inaccurate pattern recognition affecting the relevance of recommendations to users, this paper proposes a method to acquire user terminal behavior data sequences and their behavioral characteristic attributes under different transaction categories. Based on the correlation between behavioral characteristic attributes and behavior data sequences, behavioral data subsequences are segmented according to transaction categories, enabling fine-grained separation of differentiated user behavior patterns across multiple transaction scenarios. Furthermore, by analyzing each behavioral data subsequence, the paper determines the fluctuation attributes of user transaction behavior and calculates the fluctuation range of transaction behavior by combining implicit weight coefficients reflecting the influence of each transaction type on the current recommendation scenario. This ensures that the description of dynamic changes in user behavior not only depends on the behavior itself but also incorporates the business scenario. Based on this, the paper identifies the user's transaction trend type and calls a recommendation model matching that trend type for personalized recommendations. This achieves dynamic adaptation and accurate matching of the recommendation model, improving the high adaptability between target recommended items and the user's current behavioral state and potential needs, thus realizing personalized recommendations and improving user satisfaction.

[0021] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1This is a flowchart of an item recommendation method provided in an embodiment of the present invention;

[0024] Figure 2 This is a flowchart of the item recommendation method provided in the embodiments of the present invention;

[0025] Figure 3 This is a flowchart of an item recommendation method provided in an embodiment of the present invention;

[0026] Figure 4 This is a flowchart of the item recommendation method provided in the embodiments of the present invention;

[0027] Figure 5 This is a flowchart of an item recommendation method provided in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of the structure of an item recommendation device provided in an embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram of the structure of an electronic device that implements the item recommendation method of this invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] It should be noted that the collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in this disclosed technical solution all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0033] It should be noted that in the technical solution disclosed herein, all item recommendations to users are made with the individual authorization of the users. The collection, updating, analysis, processing, use, transmission, and storage of user personal information are all conducted with the user's knowledge and consent, and comply with relevant privacy protection regulations.

[0034] Figure 1 This is a flowchart of an item recommendation method provided in an embodiment of the present invention. This embodiment is applicable to situations where items that meet a user's needs are recommended. The items can be of any type, such as music, books, courses, goods, food, loan products, insurance, and various other types of items. This method can be executed by an item recommendation device, which can be implemented in hardware and / or software and can be configured in a computing device. Figure 1 As shown, the method includes:

[0035] S110. Obtain the terminal behavior data sequence corresponding to the user of the recommended item to be determined, as well as the behavioral characteristic attributes of the user under different transaction categories.

[0036] In this system, users can be individuals who require item recommendations. There can be one or more users, and the recommendation method is the same for each user. Terminal behavior data sequences refer to an ordered set of various operations and interactions performed by a user over a period of time using their smart terminal devices (such as mobile phones and tablets), arranged chronologically. Transaction categories can be categorized based on various factors such as the item, the purpose of acquiring the item, its use, the consumption scenario, the content of the consumption, the consumption method, or the nature of the service. Each transaction category represents one or more consumption scenarios or areas of interest. For example, consumption scenarios may include "electronic products," "clothing and footwear," "home furnishings," "education and training," and "travel." Transaction categories may also include "personal consumer loans (such as education, travel, and medical loans)," "housing loans," "auto finance loans," and "startup loans." Furthermore, transaction categories may be based on categories such as product preference, consumption level, behavioral patterns (such as purchasing during discounts, short-term decision-making, planned consumption), and time patterns (such as active nighttime consumption, weekend shopping preferences, and seasonal purchases). Behavioral characteristic attributes can be quantifiable or descriptive indicators of a user's behavior when engaging in transactions within a specific transaction category. For example, behavioral characteristic attributes may include preference intensity, temporal patterns of transaction behavior, frequency of behavior, intensity of behavior, depth of participation, persistence of behavior, or characteristics related to other behaviors. Optionally, behavioral characteristic attributes can characterize the degree of a user's preference within a corresponding transaction category. The degree of preference can refer to the intensity or level of interest a user shows towards a transaction category. This degree of preference can be comprehensively assessed through a comprehensive evaluation of the user's activity level, persistence, and regularity in related applications or services within that transaction category. For example, a user's behavioral characteristic attributes in the "Education and Training" category can reflect the frequency of their attention to content in this field, the duration of their engagement, or the strength of their willingness to consume.

[0037] In this embodiment, when it is necessary to recommend items to a user, the user's terminal behavior data sequence generated within a preset time period can be obtained, with the user's knowledge and permission. Optionally, the terminal behavior data sequence includes various types of information such as the user's terminal communication duration, frequency, application usage (launch, dwell time), location changes (base station or GPS positioning point), transaction information between the user and items, browsing information, click information, favorites information, add-to-cart information, and timestamps of the behavior. Simultaneously, behavioral characteristic attributes corresponding to the user under different transaction categories can be obtained. These behavioral characteristic attributes can be indicators extracted through statistical analysis or model calculation based on the user's historical behavior data under the corresponding transaction category. Alternatively, for each transaction category, all related behaviors can be filtered from the user's terminal behavior data sequence to form a behavior set under different transaction categories; behavioral characteristic attributes are then generated for the behavior set. For example, for the "electronic products" category, we can analyze user behavior characteristics by tracking the frequency of browsing related applications, the average time spent on product detail pages, and the number of price comparisons. For the "home furnishings" category, we can analyze the patterns of browsing related products on weekends and the click-through rate of promotional information. For the "travel" category, we can calculate the frequency of destination searches, the depth of travel guide viewing, and the peak activity window before travel. Alternatively, we can identify user behavior data in related applications or services across different categories, statistically analyze key indicators (such as usage frequency, duration, and interaction depth), and generate a numerical behavioral feature vector through normalization, weighting, or model scoring to represent the intensity of user interest in each transaction category.

[0038] For example, with the user's knowledge and permission, a sequence of terminal behavior data can be obtained from behavioral information generated in the user's terminal data. This sequence can be used to analyze the user's dynamic transaction habits, thereby enabling personalized predictions and recommendations. For instance, the terminal behavior data sequence includes multi-dimensional data such as user-related information, location, and signaling data, along with their corresponding timestamps.

[0039] It should be emphasized that in the technical solutions disclosed herein, the collection, updating, analysis, processing, use, transmission, and storage of users' personal information are all carried out with the user's knowledge and permission, and comply with relevant privacy protection regulations.

[0040] S120. Associate the behavioral feature attributes and the terminal behavior data sequence to obtain the behavioral data subsequences corresponding to users under different transaction categories.

[0041] Among them, behavioral data subsequences refer to behavioral sequences filtered and associated with behavioral characteristic attributes based on transaction categories from complete terminal behavioral data sequences. Behavioral data subsequences can reflect changes in user behavioral data within that transaction type domain.

[0042] In this embodiment, each behavior data record in the terminal behavior data sequence can be analyzed to identify the transaction type of one or more behavior data records. All behavior data records belonging to the same transaction type are associated with the user's corresponding behavior feature attributes under that transaction type to obtain a behavior data subsequence corresponding to that transaction type. This behavior data subsequence contains behavior data records and behavior feature attributes related to the transaction category. Accordingly, behavior data subsequences corresponding to the user under different transaction categories can be obtained. For example, behavior data subsequences for the user under multiple independent transaction categories such as "electronic products subsequence" and "home furnishings subsequence" can be generated.

[0043] For example, see Figure 2 This allows us to associate each user's behavioral characteristics across different transaction types with terminal behavior data sequences, resulting in subsequences of behavioral data for different users across different transaction categories. The set of these behavioral data subsequences can be represented as X = .in, Indicates the first Each transaction type corresponds to a subsequence of behavioral data. To improve recommendation accuracy, the terminal behavioral data sequence can be preprocessed, such as imputing missing values, detecting and handling outliers, and converting data types. Feature engineering can also be performed on the terminal behavioral data sequence, such as creating new features (e.g., calculating the user's average preference attribute, the standard deviation of the preference attribute, the time interval of the most recent transaction, etc.), and supplementing the terminal behavioral data sequence with new features to improve its data richness.

[0044] S130. Based on the user's behavioral data subsequences under different transaction categories, determine the floating attributes of transaction behavior.

[0045] Among them, the transaction behavior fluctuation attribute can be used to characterize the dynamic fluctuation of users' overall transaction (such as consumption) behavior, describe the trend and magnitude of changes in user interest intensity, behavioral activity, or engagement depth over time, and reflect the stability of their behavioral patterns, migration frequency, or decision-making activity. The higher the value of the transaction behavior fluctuation attribute, the more unstable the recent behavior or the more frequent the interest migration; the lower the value, the more stable the behavior tends to be.

[0046] In this embodiment, for each behavioral data subsequence corresponding to a transaction category, the temporal distribution of dense behavioral areas and silent periods within the behavioral data subsequence can be identified, continuous behavioral sessions can be divided, and the magnitude of activity changes between adjacent behavioral groups (such as jumps or falls in usage duration and access frequency) can be calculated to extract local fluctuation features. For each behavioral data subsequence, a temporal fluctuation feature quantifying its fluctuation intensity is generated based on the local fluctuation features. Then, by cross-category aggregation (such as weighted aggregation, principal component fusion, or dynamic model integration) of the temporal fluctuation features corresponding to multiple transaction categories, a transaction behavior floating attribute that can comprehensively reflect the overall dynamics of user behavior is generated.

[0047] Specifically, the behavioral data subsequence includes multiple behavioral data groups; based on the behavioral data subsequences corresponding to users under different transaction categories, the floating attributes of transaction behavior are determined, including: for each transaction category, based on the temporally adjacent behavioral data groups in the behavioral data subsequence corresponding to the user under the current transaction category and the number of behavioral information entries in the behavioral data groups, the temporally adjacent behavioral data groups are determined, and the floating characteristic attributes corresponding to the user under the current transaction type are determined based on the floating characteristic attributes corresponding to the user under different transaction types.

[0048] It should be noted that the method for determining the floating characteristic attributes of a user under each transaction type is the same. We can take any transaction type as the current transaction type and use it as an example to introduce how to determine the floating characteristic attributes of a user under the current transaction type.

[0049] In this context, "temporally adjacent behavioral data groups" can refer to adjacent behavioral combinations within a subsequence of behavioral data for a transaction category, divided based on temporal continuity. For example, every two consecutive behaviors (the i-th and i+1-th) can be grouped together, or behavioral segments can be merged based on the interval between behaviors (e.g., less than 30 minutes). The number of behavioral information entries can refer to the number of behavioral records contained within a behavioral data group. For example, the group may contain multiple behavioral records, each collected at a specific moment or within a certain period, with the number of entries equal to the total number of behavioral records. The "temporally fluctuating attribute" can be a numerical value calculated for each pair of temporally adjacent behavioral data groups, considering the time interval between their preceding and following behaviors, the trend of behavior type changes, and the number of behavioral information entries within the group. This value reflects the dynamic fluctuation characteristics of that local behavioral segment. The temporally fluctuating attribute can be used to characterize the "amplitude of change" in a user's behavioral intensity or interest state within a given time period, such as a jump from low activity to high activity, or a decline from high-frequency operation to prolonged inactivity. Floating feature attributes can refer to the comprehensive features obtained by aggregating and analyzing the floating attributes corresponding to all temporally adjacent behavioral data groups under a transaction category. Floating feature attributes are used to characterize the regularity, stability, or activity evolution rhythm of user behavior fluctuations within that transaction type. Transaction behavior floating attributes can also refer to global behavioral features generated by integrating the floating feature attributes of users across all different transaction categories through methods such as weighted fusion, pattern recognition, or vector integration. Transaction behavior floating attributes are used to characterize the frequency of cross-domain interest migration, the instability of behavioral rhythm, or the sensitivity of preference evolution in a user's overall transaction behavior.

[0050] In this embodiment, for each transaction category, the user's behavioral data subsequence corresponding to the current transaction category can be divided based on a preset duration to obtain multiple terminal behavior groups with consecutive time sequences. For one of the terminal behavior groups between any two temporally adjacent terminal behavior groups, the number of data entries of behavioral information within the terminal behavior group can be counted as the first entry number. All behavioral information in the terminal behavior group is summed to obtain a first sum value. The first sum value and the first entry number are divided to obtain a first mean value. The square of the difference between each behavioral information in the terminal behavior group and the first mean value is calculated. All squares are summed to obtain a second sum value. The second sum value and the first entry number are divided to obtain a first intermediate value corresponding to the terminal behavior group. Similarly, for another terminal behavior group within the same temporally adjacent terminal behavior group, the number of data entries for behavior information within the terminal behavior group is counted as the second count; all behavior information in the terminal behavior group is summed to obtain the third sum; the terminal behavior group and the second count are divided to obtain the second mean; the square of the difference between each behavior information in the terminal behavior group and the second mean is calculated; all squares are summed to obtain the fourth sum; the fourth sum is divided by the second count to obtain the second median (i.e., variance) corresponding to the other terminal behavior group. The difference between the first median and the second median is processed to obtain the temporal floating attribute corresponding to the two temporally adjacent terminal behavior groups. Following the above method, the determination of the temporal floating attribute corresponding to the two temporally adjacent terminal behavior groups is repeated to obtain multiple temporal floating attributes. The absolute values ​​of the multiple temporal floating attributes are summed to obtain the floating characteristic attribute corresponding to the user under the current transaction type. Through the above method, the floating characteristic attributes corresponding to the user under different transaction types can be obtained. Furthermore, the floating characteristic attributes corresponding to users under different transaction types can be fused, such as through weighted average, principal component analysis, or neural network fusion, to obtain the floating attributes of users' transaction behavior. These floating attributes, as a representation of the dynamics of users' overall behavior, can be used to determine whether a user is currently in a period of high activity or interest migration.

[0051] For example, the set of behavioral data subsequences can be represented as X= The user's behavioral data subsequence for each transaction category is further divided into k behavioral data groups. The set of these subsequenced behavioral data groups can be represented as x = t represents the index of the row data group, t=1,2,……,k; An index representing a subsequence of behavioral data representing a transaction category. =1,2…… . Indicates the first The kth behavioral data group of a subsequence of behavioral data for a transaction category.

[0052] The user's position in the first month can be determined based on formula (1). The floating characteristic attributes corresponding to each transaction type. Formula (1) can be expressed as: ;in: This indicates that the user is at the The floating characteristic attributes corresponding to each transaction type; This indicates which time-series adjacent action data group; This represents the total number of temporally adjacent behavioral data groups (the total number is k-1). In representing temporally adjacent behavioral data groups The number of samples (i.e., the number of records) within the behavioral data group. In representing temporally adjacent behavioral data groups The number of samples (i.e., the number of behavioral information entries) within the behavioral data group. express The first in the behavioral data group Individual behavioral information; Indicates the first The first mean of behavioral information within the behavioral data group; Indicates the first The first in the behavioral data group Individual behavioral information; Indicates the first The second mean of behavioral information within the behavioral data group; | represents the absolute value symbol.

[0053] The advantage of this setup is that by constructing behavioral data subsequences for each transaction category and determining the temporal fluctuation attributes based on temporally adjacent behavioral data groups and the number of behavioral information entries within each group, it can effectively capture the local changing trends of user activity in specific consumption areas and reflect the dynamic fluctuation characteristics of behavioral density. Furthermore, by aggregating the temporal fluctuation attributes of adjacent time periods, the fluctuation feature attributes under the current transaction type are extracted, realizing the feature extraction of user behavior stability in the current transaction type scenario. Finally, based on the user's fluctuation feature attributes across multiple transaction types, the transaction behavior fluctuation attributes are comprehensively determined, fully representing the dynamic pattern of the user's overall behavior, enhancing the perception of dynamic behaviors such as user interest migration and activity mutations, and improving the accuracy of user behavior description.

[0054] S140. Based on the predetermined implicit weight coefficients and the corresponding behavioral data subsequences of users under different transaction categories, determine the fluctuation range of transaction behavior.

[0055] The implicit weight coefficient is used to characterize the degree of influence of all transaction types on the demand for item acquisition in the current recommendation business scenario. The transaction behavior fluctuation range can be used to describe the dynamic range or threshold boundary of user behavior fluctuations, and to define whether the user is in a fluctuating or stable state in the current business scenario.

[0056] In this embodiment, the implicit weight coefficient can be determined based on the recommendation importance attributes of different transaction categories and the correlation attributes between different transaction categories and the item acquisition needs of the current recommendation business scenario. Recommendation importance attributes can be used to characterize the degree of importance of a transaction category in the recommendation system. For example, a higher value for a recommendation importance attribute indicates a greater likelihood of prioritizing the recommendation of that transaction category. Correlation attributes refer to the degree of association between a transaction category and the item acquisition needs of the current recommendation business scenario. For example, a higher correlation coefficient between a transaction category (such as "large-scale electronic product consumption") and "loan application scenario" indicates that this type of transaction category is more likely to predict a user's loan needs. In practical applications, for each transaction category i, the Pearson correlation coefficient between transaction category i and the item acquisition needs (such as loan needs) Y of the current recommendation business scenario can be calculated. The Pearson correlation coefficient is used as the correlation attribute between the two. Based on multiple correlation attributes and the recommendation importance attribute corresponding to each transaction category, the implicit weight coefficient Tw of all transaction types in the current recommendation business scenario is obtained. The implicit weight coefficient Tw represents the implicit importance of all transaction types in the current recommendation business scenario, that is, it represents the degree of influence on the demand for item acquisition in the current recommendation business scenario. For example, the implicit weight coefficient can be determined based on formula (2); formula (2) can be expressed as ;in, This can be represented as the recommendation importance attribute of the i-th transaction category. ; This represents the correlation between the i-th transaction category and the item acquisition needs of the current recommendation business scenario.

[0057] In practical applications, neural network models can be used to analyze user behavior data subsequences under different transaction categories based on implicit weight coefficients, thereby determining the fluctuation range of user transaction behavior. For example, high-importance transaction categories have a greater impact on their volatility, while the volatility fluctuation range of low-importance transaction categories is moderately suppressed.

[0058] Specifically, in determining the fluctuation range of transaction behavior based on predetermined implicit weight coefficients and the corresponding behavioral data subsequences of users under different transaction categories, the mean of the fluctuation variance can be determined based on the corresponding behavioral data subsequences of users under different transaction categories; the standard deviation of the fluctuation variance can be determined based on the corresponding behavioral data subsequences of users under different transaction categories and the mean of the fluctuation variance; and the fluctuation range of transaction behavior can be determined based on the mean of the fluctuation variance, the standard deviation of the fluctuation variance, and the implicit weight coefficients.

[0059] The mean of the floating variance can be used to characterize the overall average level of user behavior volatility across different transaction categories. The standard deviation of the floating variance can be used to characterize the deviation of user volatility intensity across various transaction behaviors. In this embodiment, for a certain behavior data subsequence, the total number of terminal behavior groups in the behavior data subsequence can be counted; the sum of all terminal behavior groups in the behavior data subsequence can be obtained as a fifth sum; the quotient of the fifth sum and the total number of groups can be calculated to obtain a third mean; the square of the difference between each terminal behavior group in the behavior data subsequence and the third mean can be calculated; all squares can be summed to obtain a sixth sum; the quotient of the sixth sum and the total number of groups can be calculated to obtain the variance corresponding to the behavior data subsequence. This variance can refer to the degree of dispersion of the numerical distribution of the behavior data subsequence under the transaction category. The larger the variance, the more obvious the difference in activity between the terminal behavior data groups in the behavior data subsequence, and the more intense the behavior volatility. The variances of user behavior data subsequences across all transaction categories are summed to obtain a seventh sum. The seventh sum is then divided by the total number of sequences across all behavior data subsequences to obtain the mean of floating variance for all behavior data subsequences. The squares of the differences between each behavior data subsequence and the mean of floating variance are summed to obtain an eighth sum. The eighth sum is then divided by the total number of sequences to obtain the standard deviation of floating variance for all behavior data subsequences. The standard deviation is then enhanced using implicit weighting coefficients to obtain a first processed value. Based on the first processed value and the mean of floating variance, the range of transaction behavior fluctuations is determined.

[0060] For example, the mean of the floating variance of all behavioral data subsequences can be determined based on formula (3). Formula (3) can be expressed as: ;in, For the first The variance of a subsequence of behavioral data; This represents the total number of subsequences of behavioral data. This is expressed as the mean of the floating variance. The floating standard deviation can be determined based on formula (4); formula (4) can be expressed as: ; This is expressed as the floating standard deviation. The fluctuation range of trading behavior can be determined based on formula (5); formula (5) can be expressed as: ; This represents the fluctuation range of a user's transaction behavior.

[0061] The advantage of this setup is that by calculating the mean of the floating variance based on all behavioral data subsequences, it can accurately uncover the overall average level of user behavior fluctuation intensity across multiple transaction categories, effectively reflecting the global stability of their trading behavior. Furthermore, by combining all behavioral data subsequences with the mean of the floating variance to obtain the floating standard deviation, it quantifies the magnitude of user behavior deviating from the norm. Finally, by dynamically constructing the trading behavior fluctuation range by integrating the mean of the floating variance, the floating standard deviation, and implicit weight coefficients, this range fully considers the actual differences in the impact of different transaction types in the current business scenario. This improves the accuracy of determining the boundaries of user behavior fluctuations, enhances the robustness and rationality of user trend state identification, and thus provides differentiated recommendations for users, improving the accuracy of recommendation results.

[0062] S150. Based on the transaction behavior fluctuation attributes and transaction behavior fluctuation range, determine the transaction trend type corresponding to the user, input at least one behavioral data subsequence into the recommendation model that matches the transaction trend type, and obtain the target recommended item corresponding to the user.

[0063] Transaction trend type can be used to characterize the trend of a user's transaction behavior in the current recommended transaction scenario. For example, transaction trend types include, but are not limited to, stable transaction behavior and volatile transaction behavior. The recommendation model can be a machine learning or deep learning model trained for a specific transaction trend type, used to predict items a user might be interested in based on their behavioral data. Different transaction trend types correspond to different recommendation models. The target recommended item can be a recommendation result generated for the user; the target recommended item is an item within the current recommendation business scenario. The current recommendation business scenario can refer to the specific application context served by the current recommendation system, such as new product recommendations on e-commerce platforms, financial product recommendations on financial platforms, and course recommendations on educational platforms.

[0064] In this embodiment, the fluctuation attribute and fluctuation range of transaction behavior can be compared to determine the user's transaction trend type. For example, if the fluctuation attribute is not greater than the fluctuation range, the transaction behavior is determined to be stable; if the fluctuation attribute exceeds the fluctuation range, the transaction behavior is determined to be volatile. Then, based on the determined transaction trend type, a matching recommendation model is selected. The behavioral data subsequences corresponding to each transaction category are used as input data for the recommendation model. The recommendation model filters target recommended items based on the user's behavioral data subsequences under different categories. For example, for users with stable transaction behavior, upgraded versions, related products, or high repurchase rate products in their frequently purchased categories can be recommended, while avoiding pushing high-risk or highly exploratory new categories of content, thereby improving the accuracy of recommendations and user satisfaction.

[0065] For example, see [link to previous article] Figure 2 The user's trading trend type can be determined based on the floating attribute α of the user's trading habits and the floating difference (Fα=Fi-α) of the floating range Fi of the trading behavior.

[0066] It should be noted that the transaction trend type can also be used to characterize the degree of abnormal transaction risk; for example, the more obvious the behavioral fluctuation, the higher the abnormal risk. Optionally, when the transaction behavior fluctuation attribute exceeds the transaction behavior fluctuation range, it is determined that there is abnormal transaction risk. Here, at least one subsequence of behavioral data can be input into a risk identification model that matches the transaction trend type (i.e., the abnormal type), and the model can output the user's corresponding risk level or suspicious transaction identifier to detect whether the user's transaction is risky, so as to provide accurate risk warning.

[0067] The technical solution provided in this embodiment obtains terminal behavior data sequences corresponding to users of items to be recommended, as well as behavioral feature attributes corresponding to users in different transaction categories. These behavioral feature attributes characterize the user's preference level within the corresponding transaction category. The behavioral feature attributes and terminal behavior data sequences are correlated to obtain subsequences of user behavior data corresponding to different transaction categories. Based on these subsequences, a fluctuation attribute for the user's transaction behavior is determined. A fluctuation range for the user's transaction behavior is determined based on pre-determined implicit weight coefficients and the subsequences. Based on the fluctuation attribute and fluctuation range, a transaction trend type is determined. At least one subsequence of behavior data is input into a recommendation model matching the transaction trend type to obtain the target recommended item corresponding to the user. This solution addresses the problem in existing technologies that use deep neural networks to learn implicit features of user historical behavior, which suffers from limitations in understanding user behavior models. To address the issue of inaccurate pattern recognition affecting the relevance of recommendations to users, this paper addresses this problem by acquiring user terminal behavior data sequences and their behavioral characteristic attributes under different transaction categories. Based on the correlation between behavioral characteristic attributes and behavior data sequences, it divides behavior data into subsequences categorized by transaction category, enabling fine-grained separation of differentiated user behavior patterns across multiple transaction scenarios. Furthermore, by analyzing each behavior data subsequence, it determines the fluctuation attributes of user transaction behavior and, combined with implicit weight coefficients reflecting the influence of each transaction type on the current recommendation scenario, calculates the fluctuation range of transaction behavior. This ensures that the description of dynamic changes in user behavior not only relies on the behavior itself but also incorporates the contextual sensitivity of the business scenario. Based on this, it identifies the user's transaction trend type and calls upon a recommendation model matching that trend type for personalized recommendations. This achieves dynamic adaptation and precise matching of the recommendation model, improving the high adaptability between target recommended items and the user's current behavioral state and potential needs, thus realizing personalized recommendations and enhancing user satisfaction.

[0068] Figure 3 This is a flowchart of an item recommendation method provided in an embodiment of the present invention. Based on the aforementioned embodiments, when the transaction behavior fluctuation attribute is not greater than the transaction behavior fluctuation range, the user's corresponding transaction trend type is determined to be a stable transaction behavior type. At least one subsequence of behavior data can be input into a recommendation model that matches the stable transaction behavior type to obtain the target recommended item. Specific implementation methods can be found in the technical solution of this embodiment. Technical terms that are the same as or corresponding to those in the above embodiments will not be repeated here.

[0069] like Figure 3 As shown, the method specifically includes the following steps:

[0070] S210. When the fluctuation attribute of the transaction behavior is not greater than the fluctuation range of the transaction behavior, the transaction trend type corresponding to the user is determined to be the stable transaction behavior type.

[0071] Among them, the stable transaction behavior type can be used to characterize that the fluctuations in a user's transaction behavior are within a predictable and normal range. Such users exhibit behavioral patterns similar to their historical preferences, with no significant shifts in interest.

[0072] In this embodiment, the fluctuation attribute of a user's trading behavior can be compared with the fluctuation range of the trading behavior. If the value of the fluctuation attribute is not greater than the fluctuation range of the trading behavior, it is determined that the fluctuation level is in a normal and controllable state. The user's trading trend type is then classified as stable trading behavior.

[0073] For example, see [link to previous article] Figure 2 When the difference Fα between the value of the trading behavior fluctuation attribute and the trading behavior fluctuation range is less than or equal to 0, it indicates that the trend of change between the behavior sequences is stable, and the trading trend type is stable trading behavior. The current behavior state is related to the behavior states of the previous few periods, and the state value based on a certain trading type may gradually tend to a more stable range.

[0074] S220. When the trading trend type is stable trading behavior, at least one subsequence of behavioral data is input into a recommendation model that matches the stable trading behavior type to obtain reconstructed trading behavior features.

[0075] Reconstructing transaction behavior features refers to generating higher-level user behavior features after feature extraction from subsequences of behavioral data through a recommendation model. These reconstructed transaction behavior features can be semantic vector representations that integrate behavioral patterns, interest intensity, and potential preferences, used to more accurately describe the deep transaction behavior characteristics of users in a stable state.

[0076] In this embodiment, when the transaction trend type is stable, it indicates that user transaction behavior patterns tend to be consistent, preference structures are clear, and there are no significant migrations or abrupt changes in the short term. Behavioral data subsequences can be used as input data for a recommendation model that matches the stable transaction behavior type. The recommendation model can enhance the accurate modeling of users' long-term interests and the capture of subtle preference features, resulting in reconstructed transaction behavior features. For example, the recommendation model can transform behavioral data subsequences into behavioral feature vectors. Through sequence modeling modules (such as GRU or Attention mechanisms), combined with the user's historical preference distribution and category importance, the behavioral feature vectors are weighted, fused, and nonlinearly transformed to obtain reconstructed transaction behavior features. The recommendation model also includes a decision network layer. The reconstructed transaction behavior features can be fed into the decision network layer to determine the matching degree between users and candidate items, thus obtaining the target recommended item.

[0077] It should be noted that a highly recommended transaction type can be determined based on the important recommendation attributes of different transaction categories and the correlation attributes between different transaction categories and the item acquisition needs of the current recommendation business scenario. For example, a weighted average of the important recommendation attributes and correlation attributes corresponding to the same transaction type can be used to obtain the recommendation score for each transaction type, and the transaction type with the highest recommendation score can be used as the highly recommended transaction type. The behavioral data subsequence of the highly recommended transaction type is then input into a recommendation model that matches the stability of the transaction behavior to obtain the reconstructed transaction behavior features corresponding to that transaction type. These reconstructed transaction behavior features are then input into the decision network layer to output the target recommended item. Alternatively, the behavioral data subsequence of each transaction type can be input into a recommendation model that matches the stability of the transaction behavior to obtain the reconstructed transaction behavior features corresponding to each transaction type. These reconstructed transaction behavior features are then input into the decision network layer, and the target recommended item is output by comprehensively considering the reconstructed transaction behavior features of multiple transaction types, thereby improving recommendation accuracy.

[0078] Optionally, the recommendation model also includes a batch normalization layer, a splitting layer, a feature generation layer, a first feature transformation layer, and a second feature transformation layer. The batch normalization layer transforms the input data by normalizing the mean to zero and the variance to one, outputting the first behavioral feature. The splitting layer decomposes the high-dimensional behavioral feature vector into two independent sub-vectors. The feature generation layer extracts high-order features related to user preferences from local behavioral features based on the input data through nonlinear transformations (such as fully connected networks or attention mechanisms). The second feature transformation layer can be used to perform sequence-level fusion (such as weighted summation, self-attention aggregation, or RNN encoding) of the second behavioral features corresponding to all behavioral data groups to generate reconstructed transaction behavioral features.

[0079] In this embodiment, at least one behavioral data subsequence is input into a recommendation model that matches the stability of transaction behavior to obtain reconstructed transaction behavior features. This includes: for the first behavioral data group in the behavioral data subsequence, inputting the behavioral data group into a batch normalization layer to obtain a first behavioral feature; inputting the first behavioral feature into a splitting layer to obtain a first behavioral sub-feature and a second behavioral sub-feature corresponding to the behavioral data group; using the first behavioral sub-feature as input data for a feature generation layer to obtain a second behavioral feature, and inputting the second behavioral sub-feature into a first feature transformation layer to obtain a third behavioral feature; for behavioral data groups following the first behavioral data group, performing the process of reconstructing the current row... The process involves inputting a data set into a batch normalization layer to obtain a first behavioral feature, and inputting the first behavioral feature into a splitting layer to obtain a first behavioral sub-feature and a second behavioral sub-feature corresponding to the current behavioral data set. The first behavioral sub-feature corresponding to the current behavioral data set and the third behavioral feature corresponding to the previous behavioral data set are used as input data to a feature generation layer to obtain a second behavioral feature corresponding to the current behavioral data set. The process then involves inputting the second behavioral sub-feature corresponding to the current behavioral data set into a first feature transformation layer to obtain a third behavioral feature. Finally, the process involves inputting the second behavioral features corresponding to multiple behavioral data sets into a second feature transformation layer to obtain reconstructed transaction behavioral features.

[0080] For the first behavioral data group in the behavioral data subsequence, this first behavioral data group can be used as input data for a batch normalization layer. The batch normalization layer standardizes the input data, eliminating dimensional differences and distribution shifts, and outputs the first behavioral feature. This first behavioral feature is then used as input data for a splitting layer. The splitting layer performs a linear transformation or vector splitting operation on the first behavioral feature, decomposing it into two independent sub-vectors, namely the first and second behavioral sub-features. The first behavioral sub-feature is then used as input data for a feature generation layer. This feature generation layer can use non-linear activation functions (such as ReLU and Sigmoid) and fully connected networks to perform deep transformations and enhancements on the input data, generating a more expressive second behavioral feature, which serves as a higher-order feature of the first behavioral data group. Finally, the second behavioral sub-feature can be used as input data for a first feature transformation layer, outputting the third behavioral feature corresponding to the behavioral data group.

[0081] For each behavior data group following the first behavior data group, each behavior data group can be sequentially treated as the current behavior data group. The current behavior data group is input into the batch normalization layer to obtain the first behavior feature. This first behavior feature is then input into the splitting layer to obtain the first and second behavior sub-features corresponding to the current behavior data group. Using a context-aware mechanism, the first behavior sub-feature corresponding to the current behavior data group is concatenated or jointly processed with the third behavior feature corresponding to the previous behavior data group to obtain a fused feature. This fused feature is used as input data for the feature generation layer to obtain the second behavior feature corresponding to the current behavior data group. In this way, the model can combine historical behaviors to enhance its understanding and feature representation capabilities for the current behavior. The second behavior sub-feature corresponding to the current behavior data group is then input into the first feature transformation layer to obtain the third behavior feature. The above process can be executed progressively group by group according to the temporal order of the behavior data groups. Once all behavioral data groups have been processed, the corresponding second behavioral features can be input into the second feature transformation layer. The second feature transformation layer can use sequence aggregation mechanisms (such as self-attention, LSTM, or average pooling) to perform weighted fusion and structural compression on all second behavioral features, extract global behavioral patterns, and output reconstructed transaction behavioral features.

[0082] For example, assuming the current recommendation business scenario is a loan scenario, when the transaction trend type is stable, a recommendation model used to predict users with stable transaction behavior regarding their product demand in the loan scenario can be used. See [link to relevant documentation]. Figure 4 The recommendation model recommends items in the following way: For the first set of behavioral data, the behavioral data set is input into a batch normalization layer, which outputs the first behavioral feature. A split layer then divides the first behavioral feature into a first behavioral sub-feature and a second behavioral sub-feature. The first behavioral sub-feature is input into a feature generation layer (such as a Temporal Transformer) to obtain the second behavioral feature. The second behavioral sub-feature is then input into a first feature generation layer to obtain the third behavioral feature. For subsequent sets of behavioral data, the current set of behavioral data is used as input to the batch normalization layer, and the steps to obtain the first and second behavioral sub-features are repeated. The first behavioral sub-feature corresponding to the current set of behavioral data and the third behavioral feature corresponding to the previous set of behavioral data are used as input to the feature generation layer to obtain the second behavioral feature. The second behavioral sub-feature is then input into the first feature generation layer to obtain the third behavioral feature. Finally, the second behavioral feature output from each feature generation layer is input into an encoder and transformed into a high-weight feature matrix by the second feature generation layer, outputting the reconstructed transaction behavioral feature.

[0083] The advantage of this setup is that by independently initializing the first behavior data group in the behavior data subsequence, it is input into a batch normalization layer to eliminate input distribution bias, outputting a standardized first behavior feature. This feature is then decomposed into first and second behavior sub-features by a splitting layer. A feature generation layer then performs a non-linear transformation on the first behavior sub-features to generate semantically expressive second behavior features, effectively extracting the deep patterns of the initial behavior. The other half of the second behavior sub-features is input into a first feature transformation layer to obtain the third behavior feature corresponding to the behavior data group. For each subsequent behavior data group, the batch normalization layer and splitting operation are repeated. The first behavior sub-feature of the current behavior data group is combined with the third behavior feature of the previous behavior data group and input into the feature generation layer, introducing state dependencies across time steps. This allows historical behavior states to guide the current behavior context, improving the accuracy of capturing the temporal features of user behavior. Finally, the second behavior features corresponding to all behavior data groups are aggregated and input into the second feature transformation layer to generate unified reconstructed transaction behavior features. This technical solution achieves layer-by-layer decomposition, context awareness, and dynamic reconstruction of user behavior sequences, enhancing the model's ability to express behavioral trends and the accuracy of identifying temporal dependencies. This enables the reconstructed transaction behavior features to accurately represent the deep preference features of users in stable transaction behavior states, improving recommendation accuracy and ensuring high adaptability between recommended items and user needs in stable preference scenarios.

[0084] To further improve recommendation accuracy, the feature generation layer may include an attention unit and a feature generation unit. The first behavior sub-feature corresponding to the current behavior data group and the third behavior feature corresponding to the previous behavior data group are used as input data to obtain the second behavior feature corresponding to the current behavior data group. This includes: determining the current attention weight based on the attention unit, according to the first behavior sub-feature corresponding to the current behavior data group, the third behavior feature corresponding to the previous behavior data group, historical attention weights, a pre-acquired weight matrix, and implicit weight coefficients; and determining the second behavior feature based on the feature generation unit, according to the first behavior sub-feature and the current attention weight.

[0085] The Attention Network Layer can be a neural network structure in a recommendation model used to implement the attention mechanism, dynamically allocating attention levels based on the contextual importance of the input data. For example, the Attention Network Layer can highlight behavioral data groups that have a greater impact on the current recommendation task by calculating the relative weights of behavioral data groups across different time periods. The weight matrix includes transaction demand weights corresponding to different transaction categories under different recommended business scenarios and time periods. Transaction demand weights characterize the relative importance or influence of transaction categories on users' actual transaction needs. For example, the weight matrix records transaction demand weights under different recommended business scenarios (such as e-commerce promotions, financial management, and educational course recommendations), different time periods (such as weekdays / weekends, daytime / nighttimes, and holidays), and different transaction categories (such as "electronic products" and "education and training"). For instance, in a certain recommendation scenario and time interval (such as "8:00 PM - 10:00 PM"), which of the different transaction categories (A, B, and C) is more likely to stimulate a user's actual acquisition intention? The implicit weight coefficients include the degree of influence of all transaction types on the demand for items in the current recommended business scenario across different time periods.

[0086] It should be noted that the attention weight corresponding to the first behavior data group can be 1. The current attention weight of each behavior data group can be the historical attention weight used to determine the current attention weight of its next behavior data group. For example, if the current attention weight of behavior data group A is 0.5, then when determining the current attention weight of the next behavior data group B, the current attention weight of behavior data group A can be called the historical attention weight of behavior data group A. Attention weights can be used to characterize the contribution ratio of the first behavior sub-feature in generating the second behavior feature corresponding to the behavior data group; a higher value indicates that the feature is more critical.

[0087] In this embodiment, the transaction demand weights matching the collection time period, current recommendation business scenario, and transaction category of the current behavior data group can be obtained from the weight matrix and used as the target transaction demand weights. The influence degree matching the collection time period and current recommendation business scenario of the current behavior data group can be obtained from the implicit weight coefficients and used as the target influence degree. The intermediate vector of the first behavior sub-feature corresponding to the current behavior data group is extracted to obtain the first intermediate vector, and the intermediate vector of the third behavior feature corresponding to the previous behavior data group is extracted to obtain the second intermediate vector. Based on the first evaluation function in the attention unit, the target transaction demand weight, target influence degree, first intermediate vector, second intermediate vector, and historical attention weights corresponding to the previous behavior data group are processed to obtain the current attention weight corresponding to the current behavior data group.

[0088] For example, the first evaluation function can be expressed as: ;in, This indicates the current attention weight corresponding to the current behavior data group; This indicates the historical attention weight corresponding to the previous behavioral data group; This represents the intermediate vector of the sub-features in the first row; The intermediate vector representing the third row's features; Indicates the target transaction demand weight; Indicates the degree of impact of the target. This represents the dot product operation; Represents the normalization function; The attention unit represents the weight parameters learned at the current time step of the current action data set.

[0089] Alternatively, the current behavior data group can be further subdivided into finer-grained segments, encompassing multiple behavioral information segments. In this case, the first behavior sub-feature includes smaller first behavior features corresponding to the multiple behavioral information segments within the behavior data group. Based on the second evaluation function in the attention unit, the transaction demand weight, influence level, and the third behavior feature and historical attention weights corresponding to the smaller first behavior feature are processed to obtain the attention weight of the smaller first behavior feature. A weighted average of these attention weights is then applied to obtain the current attention weight of the current behavior data group.

[0090] For example, the second evaluation function can be expressed as: ;in, This indicates the total number of behavioral information entries in the current behavioral data group; Indicates the first action in the previous action data set. Historical attention weights corresponding to each behavioral information; This indicates the first [item] in the current behavior data set. The first row is the intermediate vector of the small features; Indicating the third behavioral characteristic The intermediate vector corresponding to each behavioral information; Indicates the first [number]th [item] in the current behavior data set. Each first row represents the transaction demand weight corresponding to a minor feature; This indicates the first [item] in the current behavior data set. The degree of influence corresponding to each minor feature in the first row.

[0091] S230. Based on the decision network layer, the reconstructed transaction behavior features are processed to obtain the target recommended items corresponding to the user.

[0092] In this embodiment, the reconstructed transaction behavior features can be input into the decision network layer. The decision network layer can calculate the matching degree based on the reconstructed transaction behavior features and the item feature information of each candidate item in the current recommendation business scenario, and output the final target recommended item.

[0093] For example, see [link to previous article] Figure 4 The decision network layer can be a decision tree. Based on the decision tree, the characteristics of reconstructed transaction behavior can be analyzed to obtain the probability distribution of each candidate item in the current recommendation business scenario. Then, the top 3 items in descending order are selected as the target recommended items for users with stable transaction types.

[0094] Based on the above technical solutions, a recommendation model that matches the stable type of transaction behavior can be pre-trained. When the transaction trend type corresponding to a user is the stable type of transaction behavior, one or more subsequences of the user's behavior data can be input into the recommendation model that matches the transaction trend type to obtain the target recommended items corresponding to the user.

[0095] The method for training a recommendation model that matches the stability of transaction behavior can be as follows: obtain multiple first training samples, which include terminal behavior sequences corresponding to the transaction type; input the first training samples into the recommendation model to be trained to obtain predicted reconstructed transaction behavior features; use the reconstruction residual loss function to determine the first loss value based on the predicted reconstructed transaction behavior features and the first training samples; correct the model parameters in the recommendation model based on the first loss value; take the convergence of the reconstruction residual loss function as the initial training objective to obtain the recommendation model to be validated.

[0096] For example, the reconstruction residual loss function can be expressed as: ;in, This represents the terminal behavior sequence in the first training sample. In the terminal behavior subsequence, the first The actual transaction behavior characteristics corresponding to each behavioral data group; This indicates that the recommendation model reconstructs transaction behavior features based on the predictions made from the first training sample. This represents the first loss value.

[0097] To improve the accuracy of the model's recommendations, the model parameters can be further adjusted using multiple labeled second training samples to adapt them to the prediction task. The mean squared error loss function is used to optimize the model parameters. The second training samples can be theoretically reconstructed transaction behavior features corresponding to the terminal behavior sequences under different transaction types, or they can be the target recommendation list corresponding to the terminal behavior sequences.

[0098] It should also be noted that the hold-out method can be used. Randomly sample 10M and 20M datasets according to the distribution ratio of the training sample data, respectively. Then, divide the two datasets into training and test sets in a ratio of 6:4 and 8:2. Train the recommendation model based on the training set and validate the model based on the test set. To ensure a balanced label distribution, the proportion of each transaction type can be kept consistent in both the training and test sets, reducing errors caused by data differences.

[0099] The technical solution provided in this embodiment, when identifying a user's transaction trend type as stable, inputs a subsequence of behavioral data categorized by transaction type into a recommendation model matching the stable transaction behavior type, generating reconstructed transaction behavior features. The decision network layer within the recommendation model then refines these reconstructed features, outputting target recommended items that match the user's behavior. Since stable users' behavioral patterns are characterized by concentrated preferences and slow changes, this solution uses a recommendation model matching the stable transaction behavior type to specifically model their behavioral data, avoiding overfitting or feature dilution problems caused by introducing too many dynamic parameters when processing stable behavior. The decision network layer, based on the reconstructed high-fidelity behavioral features, can fully explore the user's long-term stable consumption tendencies and category preferences, improving the interpretability and decision accuracy of the recommendation logic. This achieves accurate identification of stable user behavioral features, effectively enhancing the consistency between recommendation results and users' long-term needs, and improving the personalization of recommendations and user satisfaction.

[0100] Figure 5 This is a flowchart of an item recommendation method provided in an embodiment of the present invention. Based on the aforementioned embodiments, when the transaction behavior fluctuation attribute is greater than the transaction behavior fluctuation range, the user's corresponding transaction trend type is determined to be a transaction behavior fluctuation type. The behavior data subsequence can be input into a recommendation model matching the transaction behavior fluctuation type to obtain the target recommended item. Specific implementation methods can be found in the technical solution of this embodiment. Technical terms that are the same as or corresponding to those in the above embodiments will not be repeated here.

[0101] like Figure 5 As shown, the method specifically includes the following steps:

[0102] S310. When the floating attribute of the transaction behavior is greater than the floating range of the transaction behavior, determine that the transaction trend type corresponding to the user is the floating type of transaction behavior.

[0103] Among them, the floating transaction behavior type can be used to characterize users' transaction behavior as exhibiting obvious dynamic changes, such as frequently switching transaction categories in a short period of time, being active in multiple consumption areas, and experiencing dramatic fluctuations in the intensity of behavior. Such user behavior patterns may indicate that the user is in a state of exploring new needs, interests, or making important decisions.

[0104] In this embodiment, the user's transaction behavior fluctuation attribute can be compared with the transaction behavior fluctuation range. If the value of the transaction behavior fluctuation attribute is greater than the transaction behavior fluctuation range, it indicates that the transaction behavior change trend is relatively large, and the user's transaction trend type is classified as transaction behavior fluctuation type.

[0105] For example, see [link to previous article] Figure 2 When the difference Fα between the value of the transaction behavior fluctuation attribute and the transaction behavior fluctuation range is greater than 0, it indicates that the user's transaction trend type is classified as transaction behavior fluctuation type. Due to the high degree of data dispersion, user behavior dependencies can be learned through the hidden state of the user using a graph neural network (GNN) to achieve predictive recommendation. For example, for predictive recommendation of transaction behavior fluctuation type, the recommendation model can learn the hidden state of user behavior and its dependencies. This may include two max pooling operations to learn the hidden state of the user in different time periods, and then use this as input data to feed into the GNN; the GNN learns the dependencies and then predicts the hidden state of the user in the next time period to generate a recommendation list.

[0106] S320. When the trading trend type is floating trading behavior, obtain a recommendation model that matches the floating trading behavior type; the recommendation model includes a first pooling layer, a second pooling layer, and a graph neural network layer.

[0107] In this embodiment, when the transaction trend type is floating transaction behavior, it indicates that the user's transaction behavior pattern changes significantly, exhibiting characteristics such as frequent interest migration, multi-domain activity, and unstable behavior rhythm. At this time, a recommendation model matching the floating transaction behavior type can be invoked. Based on the first pooling layer, the second pooling layer, and the graph neural network layer in the recommendation model, feature mining is performed on the behavioral data subsequences to capture cross-category interest associations, short-term behavioral mutations, and potential intent changes, thereby improving the recommendation accuracy in highly dynamic scenarios.

[0108] S330. From at least one subsequence of behavioral data, determine the highly relevant behavioral features of the user in the current recommendation business scenario.

[0109] Highly relevant behavioral features can refer to the set of features in a subsequence of behavioral data that are relevant to the current recommendation scenario. For example, highly relevant behavioral features can be represented in matrix form, where each row corresponds to an item and each column represents a feature dimension, such as item category, price range, brand, popularity, user rating, sales volume, whether it is a new product, or whether it is on promotion. The first pooling layer can refer to the features in the recommendation model used to aggregate user features relevant to the current recommendation scenario at different times.

[0110] In this embodiment, multiple behavioral data subsequences can be identified based on similarity algorithms or recognition algorithms, and highly relevant behavioral features of users in the current recommendation business scenario can be determined from one or more behavioral data subsequences.

[0111] For example, assuming the current recommendation business scenario is a loan business scenario, the feature vectors in the behavioral data subsequence that are highly correlated with the loan category can be used as highly correlated behavioral features.

[0112] S340. Input the highly correlated behavioral features into the first pooling layer to obtain the first transaction behavior features.

[0113] In this embodiment, highly relevant features can be used as input data for the first pooling layer. The first pooling layer can extract and aggregate features that are potentially related to the current recommendation business scenario in different periods, and output the first transaction behavior features. Optionally, the first pooling layer can be a max pooling layer.

[0114] S350. Input the first transaction behavior feature and multiple behavior data subsequences into the second pooling layer to obtain the second transaction behavior feature.

[0115] In this embodiment, the first transaction behavior feature and multiple behavior data subsequences can be used as input data for the second pooling layer. The second pooling layer can extract and aggregate the user's long-term behavior feature vector and the current business scenario-based features, and output the second transaction behavior feature. Optionally, the second pooling layer can be a max pooling layer.

[0116] For example, multiple behavioral data subsequences are represented as Representing highly correlated behavioral features with high loan-related correlation in multiple behavioral data subsequences as ,and , d is the vector length, and N is the number of transaction types. Multiple behavioral data subsequences are then used. Highly correlated behavioral characteristics By concatenating the matrices, we obtain a concatenated matrix, which can be represented as: The concatenated matrix is ​​input into the recommendation model, and the second transaction behavior feature is determined based on the first processing function in the recommendation model.

[0117] The first processing function can be expressed as: ; Represented as the first Individual users Highly relevant behavioral characteristics in the j-th sub-period of time period 1; This is represented as the first pooling layer; This is represented as the second pooling layer. Represented as the first Individual users The characteristics of the second transaction behavior in the j-th sub-period of time period 1.

[0118] S360, based on the graph neural network layer, processes the second transaction behavior features according to the implicit weight coefficients to obtain the target recommended items corresponding to the user.

[0119] The Graph Neural Network Layer (GNN Layer) can be a deep learning component in the recommendation model that utilizes graph structures to model relationships. Nodes in the network graph corresponding to the GNN layer can represent transaction categories, user behaviors, or items. Edges in the network graph can represent the strength of associations between nodes. For example, edges can be represented by co-occurrence, migration, semantic similarity, or interaction relationships between transaction categories, user behaviors, and items. It should be noted that the behavioral data subsequences have a temporal relationship, and correspondingly, the second transaction behavior features also have a temporal relationship. The second transaction behavior features include second transaction behavior features from multiple sub-time periods.

[0120] Based on the dependency weight determination function in a graph neural network, the data dependencies between the second transaction behavior features of each sub-period adjacent to the current sub-period (such as all past sub-periods related to the current sub-period) can be analyzed. After summarizing the data dependencies, the dependency weight corresponding to the current sub-period can be determined by combining implicit weight factors. Furthermore, based on the state update function, the hidden state of the next sub-period corresponding to the current sub-period can be predicted according to the dependency weight and the second transaction behavior features of the current sub-period. Based on multiple hidden states, the user's preference strength on different transaction types can be inferred, thereby determining the target recommended item corresponding to the user.

[0121] Alternatively, the sub-period closest to the current recommendation time can be used as the target sub-period; based on the hidden state corresponding to the predicted target sub-period, the user's preference strength in different transaction types can be inferred, thereby determining the target recommended item corresponding to the user.

[0122] For example, the weight-dependent determination function can be expressed as:

[0123] N(j) represents the set of sub-time periods that have a temporal relationship with the j-th sub-time period; [,] represents the vector concatenation operation; Indicates a fully connected layer; Let represent the dependency weight corresponding to the j-th sub-time period; Tw represents the implicit weight factor. The state update function can be expressed as: ;in, ; Indicates the first The user in the first The hidden state at time j of the time period; w represents the dependency weight of the j-th sub-time period. Based on the predicted value at time j... Hidden state of time period We infer the recommendation probability for different transaction types. Items of the top three transaction types with the highest recommendation probabilities, sorted in descending order, are selected as target recommended items for users with fluctuating transaction behavior.

[0124] Based on the above technical solution, a recommendation model matching the floating pattern of transaction behavior can also be pre-trained. When the transaction trend type is a floating pattern, the target recommended item can be determined for the user's behavioral data subsequence based on the recommendation model matching the floating pattern of transaction behavior. The method for training the recommendation model matching the floating pattern of transaction behavior can be: acquiring multiple third samples, wherein the third samples include the user's terminal behavior sequence and theoretical recommendation tags; and training the recommendation model based on the multiple third samples.

[0125] The technical solution provided in this embodiment, when identifying a user's transaction trend type as a fluctuating transaction behavior type, invokes a dedicated recommendation model matching this fluctuating transaction behavior type. This model integrates a first pooling layer, a second pooling layer, and a graph neural network layer to achieve efficient modeling and accurate recommendation of users' changing behaviors. Specifically, firstly, behavioral features highly relevant to the current recommendation business scenario are identified from multiple behavioral data subsequences divided by transaction category; the highly relevant behavioral features are compressed and extracted through the first pooling layer to obtain the first transaction behavior feature focusing on core behaviors; this first transaction behavior feature and all behavioral data subsequences are input together into the second pooling layer to further integrate local salient information and global behavioral context, generating a more expressive second transaction behavior feature; using a graph neural network layer combined with pre-set implicit weight coefficients, the second transaction behavior feature is structurally modeled to extract the dependencies and influence strengths between different transaction categories, thereby accurately capturing the dynamic migration characteristics of user interests across multiple transaction categories, enhancing the perception of rapidly changing user preferences and cross-category behavioral associations, and achieving high timeliness and high relevance of recommendation results.

[0126] Figure 6 This is a schematic diagram of the structure of an item recommendation device provided in an embodiment of the present invention. Figure 6 As shown, the device includes: a data acquisition module 410, a behavior data subsequence determination module 420, a floating information determination module 430, and a target recommended item determination module 440.

[0127] The system includes a data acquisition module 410, which acquires terminal behavior data sequences corresponding to users whose recommended items are to be determined, as well as behavioral feature attributes corresponding to the users in different transaction categories. The behavioral feature attributes characterize the user's preference level in the corresponding transaction category. A behavior data subsequence determination module 420 associates the behavioral feature attributes and the terminal behavior data sequences to obtain behavior data subsequences corresponding to the users in different transaction categories. A fluctuation information determination module 430 determines transaction behavior fluctuation attributes based on the behavior data subsequences corresponding to the users in different transaction categories, and determines the transaction behavior fluctuation range based on pre-determined implicit weight coefficients and the behavior data subsequences corresponding to the users in different transaction categories. A target recommended item determination module 440 determines the transaction trend type corresponding to the user based on the transaction behavior fluctuation attributes and the transaction behavior fluctuation range, and inputs at least one behavior data subsequence into a recommendation model matching the transaction trend type to obtain a target recommended item corresponding to the user.

[0128] The technical solution of this embodiment obtains terminal behavior data sequences corresponding to users of the recommended items to be determined, as well as behavioral feature attributes corresponding to users under different transaction categories. The behavioral feature attributes characterize the user's preference level in the corresponding transaction category. The behavioral feature attributes and terminal behavior data sequences are correlated to obtain subsequences of user behavior data corresponding to different transaction categories. Based on these subsequences, a transaction behavior fluctuation attribute is determined. Based on pre-determined implicit weight coefficients and the subsequences, a transaction behavior fluctuation range is determined. Based on the transaction behavior fluctuation attribute and range, a transaction trend type is determined. At least one subsequence of behavior data is input into a recommendation model matching the transaction trend type to obtain the target recommended item corresponding to the user. This solution addresses the problem in existing technologies that use deep neural networks to learn the implicit features of users' historical behavior, which suffers from limitations in identifying user behavior patterns. To address the issue of inaccurate recommendations affecting their relevance to users, this approach acquires user terminal behavior data sequences and their behavioral characteristic attributes across different transaction categories. Based on the correlation between these behavioral characteristic attributes and the behavior data sequences, it divides the behavior data into sub-sequences categorized by transaction type, enabling fine-grained separation of differentiated user behavior patterns across multiple transaction scenarios. Furthermore, by analyzing each sub-sequence, it determines the fluctuation attributes of user transaction behavior and, combined with implicit weighting coefficients reflecting the influence of each transaction type on the current recommendation scenario, calculates the fluctuation range of transaction behavior. This ensures that the description of dynamic changes in user behavior not only relies on the behavior itself but also incorporates the contextual sensitivity of the business scenario. Based on this, it identifies the user's transaction trend type and calls upon a recommendation model matching that trend type for personalized recommendations. This achieves dynamic adaptation and precise matching of the recommendation model, improving the high adaptability between target recommended items and the user's current behavioral state and potential needs, thus realizing personalized recommendations and enhancing user satisfaction.

[0129] Based on the above-mentioned device, optionally, the behavior data subsequence includes multiple behavior data groups; the floating information determination module 430 includes:

[0130] The time-series floating attribute determination unit is used to determine the time-series floating attribute for each transaction category based on the time-series adjacent behavior data groups in the behavior data subsequence corresponding to the user in the current transaction category and the number of behavior information items in the behavior data groups;

[0131] A floating feature attribute determination unit is used to determine the floating feature attribute corresponding to the user under the current transaction type based on multiple time-series floating attributes;

[0132] The transaction behavior floating attribute determination unit is used to determine the transaction behavior floating attribute based on the floating feature attributes corresponding to the user under different transaction types.

[0133] Based on the above-mentioned device, optionally, the floating information determination module 430 includes:

[0134] The floating variance mean determination unit is used to determine the floating variance mean based on the behavioral data subsequences corresponding to the user under different transaction categories;

[0135] The floating standard deviation determination unit is used to determine the floating standard deviation based on the behavioral data subsequences corresponding to the user under different transaction categories and the mean of the floating variance;

[0136] The transaction behavior fluctuation range determination unit is used to determine the transaction behavior fluctuation range based on the mean of the fluctuation variance, the standard deviation of the fluctuation, and the implicit weight coefficient.

[0137] Based on the above-described device, the target recommended item determination module 440 optionally includes:

[0138] The first unit is used to determine that the transaction trend type corresponding to the user is stable when the transaction behavior fluctuation attribute is not greater than the transaction behavior fluctuation range;

[0139] The second unit is used to determine that the transaction trend type corresponding to the user is a transaction behavior floating type when the transaction behavior floating attribute is greater than the transaction behavior floating range.

[0140] Based on the above-described device, the target recommended item determination module 440 optionally includes:

[0141] A reconstructed transaction behavior feature determination unit is used to input the behavior data subsequence into a recommendation model that matches the stable transaction behavior type when the transaction trend type is a stable transaction behavior type, thereby obtaining reconstructed transaction behavior features; wherein, the recommendation model includes a decision network layer;

[0142] The target recommended item determination unit is used to process the reconstructed transaction behavior features based on the decision network layer to obtain the target recommended item corresponding to the user.

[0143] Based on the above-mentioned device, optionally, the recommendation model further includes a batch normalization layer, a splitting layer, a feature generation layer, a first feature transformation layer, and a second feature transformation layer; the reconstructed transaction behavior feature determination unit includes:

[0144] The first feature determination unit is used to input the first behavior data group in the behavior data subsequence into the batch normalization layer to obtain the first behavior feature;

[0145] A splitting unit is used to input the first behavioral feature into the splitting layer to obtain a first behavioral sub-feature and a second behavioral sub-feature corresponding to the behavioral data group;

[0146] The second feature determination unit is used to take the first behavioral sub-feature as input data of the feature generation layer to obtain the second behavioral feature, and input the second behavioral sub-feature into the first feature transformation layer to obtain the third behavioral feature;

[0147] The third feature determination unit is used to perform the following operations for behavior data groups located after the first behavior data group: inputting the current behavior data group into the batch normalization layer to obtain a first behavior feature, and inputting the first behavior feature into the splitting layer to obtain a first behavior sub-feature and a second behavior sub-feature corresponding to the current behavior data group.

[0148] The fourth feature determination unit is used to take the first behavior sub-feature corresponding to the current behavior data group and the third behavior feature corresponding to the previous behavior data group as input data of the feature generation layer to obtain the second behavior feature corresponding to the current behavior data group, and to perform the operation of inputting the second behavior sub-feature corresponding to the current behavior data group into the first feature transformation layer to obtain the third behavior feature.

[0149] An aggregation unit is used to input the second behavioral features corresponding to multiple behavioral data groups into the second feature transformation layer to obtain reconstructed transaction behavioral features.

[0150] Based on the above-described apparatus, optionally, the feature generation layer includes an attention unit and a feature generation unit; the fourth feature determination unit includes:

[0151] The current attention weight determination unit is used to determine the current attention weight based on the attention unit, according to the first behavior sub-feature corresponding to the current behavior data group, the third behavior feature corresponding to the previous behavior data group, the historical attention weight, the pre-acquired weight matrix, and the implicit weight coefficient; wherein, the weight matrix includes the transaction demand weight corresponding to different transaction categories under different business scenarios and time periods to be recommended;

[0152] The fifth feature determination unit is used to determine the second behavior feature based on the feature generation unit, according to the first behavior sub-feature and the current attention weight.

[0153] Based on the above-described device, the target recommended item determination module 440 optionally includes:

[0154] The recommendation model determination unit is used to obtain a recommendation model that matches the floating trading behavior type when the trading trend type is floating trading behavior type; wherein, the recommendation model includes a first pooling layer, a second pooling layer and a graph neural network layer;

[0155] A highly relevant behavioral feature determination unit is used to determine highly relevant behavioral features of the user in the current recommendation business scenario from at least one of the behavioral data subsequences.

[0156] The first transaction behavior feature determination unit is used to input the highly relevant behavior features into the first pooling layer to obtain the first transaction behavior features;

[0157] The second transaction behavior feature determination unit is used to input the first transaction behavior feature and multiple behavior data subsequences into the second pooling layer to obtain the second transaction behavior feature;

[0158] The target recommended item determination unit is used to process the second transaction behavior features based on the graph neural network layer and the implicit weight coefficients to obtain the target recommended item corresponding to the user.

[0159] The item recommendation device provided in the embodiments of the present invention can execute the item recommendation method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of executing the method.

[0160] Figure 7 This is a schematic diagram of the structure of an electronic device implementing the item recommendation method of an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0161] like Figure 7As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory 12 or a random access memory 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the read-only memory 12 or a computer program loaded from storage unit 18 into the random access memory 13. The random access memory 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, read-only memory 12, and random access memory 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0162] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0163] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as item recommendation methods.

[0164] In some embodiments, the item recommendation method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via read-only memory 12 and / or communication unit 19. When the computer program is loaded into random access memory 13 and executed by processor 11, one or more steps of the item recommendation method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the item recommendation method by any other suitable means (e.g., by means of firmware).

[0165] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0166] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0167] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0168] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0169] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0170] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0171] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication unit 19, or installed from storage unit 18, or installed from read-only memory 12. When the computer program is executed by processor 11, it performs the functions defined in the methods of the embodiments of the present invention.

[0172] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the item recommendation method as provided in any embodiment of this invention.

[0173] In implementing the computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0174] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0175] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for recommending items, characterized in that, include: Acquire terminal behavior data sequences corresponding to users whose recommended items are to be determined, as well as behavioral feature attributes of the users under different transaction categories; wherein, the behavioral feature attributes are used to characterize the degree of preference of the users in the corresponding transaction categories; By associating the behavioral feature attributes with the terminal behavioral data sequence, the user's behavioral data sub-sequences under different transaction categories are obtained; Based on the user's behavioral data subsequences under different transaction categories, the transaction behavior fluctuation attribute is determined, and based on the predetermined implicit weight coefficients and the user's behavioral data subsequences under different transaction categories, the transaction behavior fluctuation range is determined. Based on the transaction behavior fluctuation attributes and the transaction behavior fluctuation range, the transaction trend type corresponding to the user is determined, and at least one of the behavior data subsequences is input into the recommendation model that matches the transaction trend type to obtain the target recommended item corresponding to the user.

2. The method according to claim 1, characterized in that, The behavioral data subsequence includes multiple behavioral data groups; determining the transaction behavior floating attribute based on the behavioral data subsequence corresponding to the user under different transaction categories includes: For each transaction category, the time-series floating attribute is determined based on the temporally adjacent groups of behavioral data in the behavioral data subsequence corresponding to the current transaction category and the number of behavioral information entries in the behavioral data groups. Based on multiple time-series floating attributes, determine the floating characteristic attributes corresponding to the user under the current transaction type; The floating attribute of the transaction behavior is determined based on the floating characteristic attributes of the user under different transaction types.

3. The method according to claim 1, characterized in that, The determination of the transaction behavior fluctuation range based on pre-determined implicit weight coefficients and the user's corresponding behavioral data subsequences under different transaction categories includes: Based on the user's behavioral data subsequences under different transaction categories, determine the mean of the floating variance; Based on the user's behavioral data subsequences under different transaction categories and the mean of the floating variance, the floating standard deviation is determined; The fluctuation range of the trading behavior is determined based on the mean of the floating variance, the standard deviation of the floating variance, and the implicit weight coefficient.

4. The method according to claim 1, characterized in that, The step of determining the transaction trend type corresponding to the user based on the transaction behavior fluctuation attribute and the transaction behavior fluctuation range includes: When the fluctuation attribute of the transaction behavior is not greater than the fluctuation range of the transaction behavior, the transaction trend type corresponding to the user is determined to be the stable transaction behavior type. When the transaction behavior fluctuation attribute is greater than the transaction behavior fluctuation range, the transaction trend type corresponding to the user is determined to be transaction behavior fluctuation type.

5. The method according to claim 1, characterized in that, The step of inputting at least one of the behavioral data subsequences into a recommendation model that matches the transaction trend type to obtain target recommended items corresponding to the user includes: When the trading trend type is a stable trading behavior type, at least one subsequence of the trading behavior data is input into a recommendation model that matches the stable trading behavior type to obtain reconstructed trading behavior features; wherein, the recommendation model includes a decision network layer; The reconstructed transaction behavior features are processed based on the decision network layer to obtain target recommended items corresponding to the user.

6. The method according to claim 5, characterized in that, The recommendation model also includes a batch normalization layer, a splitting layer, a feature generation layer, a first feature transformation layer, and a second feature transformation layer. The step of inputting at least one of the behavioral data subsequences into a recommendation model that matches the stability of the transaction behavior to obtain reconstructed transaction behavior features includes: For the first behavior data group in the behavior data subsequence, the behavior data group is input into the batch normalization layer to obtain the first behavior feature; The first behavioral feature is input into the splitting layer to obtain the first behavioral sub-feature and the second behavioral sub-feature corresponding to the behavioral data group; The first behavioral sub-feature is used as input data for the feature generation layer to obtain the second behavioral feature, and the second behavioral sub-feature is input into the first feature transformation layer to obtain the third behavioral feature; For the behavior data group located after the first behavior data group, the following operations are performed: inputting the current behavior data group into the batch normalization layer to obtain the first behavior feature, and inputting the first behavior feature into the splitting layer to obtain the first behavior sub-feature and the second behavior sub-feature corresponding to the current behavior data group; The first behavior sub-feature corresponding to the current behavior data group and the third behavior feature corresponding to the previous behavior data group are used as input data for the feature generation layer to obtain the second behavior feature corresponding to the current behavior data group. Then, the operation of inputting the second behavior sub-feature corresponding to the current behavior data group into the first feature transformation layer is performed to obtain the third behavior feature. The second behavioral features corresponding to multiple behavioral data groups are input into the second feature transformation layer to obtain reconstructed transaction behavior features.

7. The method according to claim 6, characterized in that, The feature generation layer includes an attention unit and a feature generation unit; the step of using the first behavior sub-feature corresponding to the current behavior data group and the third behavior feature corresponding to the previous behavior data group as input data to the feature generation layer to obtain the second behavior feature corresponding to the current behavior data group includes: Based on the attention unit, the current attention weight is determined according to the first behavior sub-feature corresponding to the current behavior data group, the third behavior feature corresponding to the previous behavior data group, the historical attention weight, the pre-acquired weight matrix, and the implicit weight coefficient; wherein, the weight matrix includes the transaction demand weight corresponding to different transaction categories under different business scenarios and time periods to be recommended. Based on the feature generation unit, the second behavior feature is determined according to the first behavior sub-feature and the current attention weight.

8. The method according to claim 1, characterized in that, The step of inputting at least one of the behavioral data subsequences into a recommendation model that matches the transaction trend type to obtain target recommended items corresponding to the user includes: When the trading trend type is a floating trading behavior type, a recommendation model matching the floating trading behavior type is obtained; wherein, the recommendation model includes a first pooling layer, a second pooling layer, and a graph neural network layer; From at least one of the behavioral data subsequences, determine the highly relevant behavioral characteristics of the user in the current recommendation business scenario; The highly correlated behavioral features are input into the first pooling layer to obtain the first transaction behavior features; The first transaction behavior feature and multiple subsequences of the behavior data are input into the second pooling layer to obtain the second transaction behavior feature; Based on the graph neural network layer, the second transaction behavior features are processed according to the implicit weight coefficients to obtain the target recommended items corresponding to the user.

9. An item recommendation device, characterized in that, include: The data acquisition module is used to acquire terminal behavior data sequences corresponding to users of items to be recommended, as well as behavioral feature attributes of the users under different transaction categories; wherein, the behavioral feature attributes are used to characterize the degree of preference of the users in the corresponding transaction categories; The behavior data subsequence determination module is used to associate the behavior feature attributes and the terminal behavior data sequence to obtain the behavior data subsequences corresponding to the user under different transaction categories; The floating information determination module is used to determine the floating attribute of transaction behavior based on the subsequence of behavioral data corresponding to the user under different transaction categories, and to determine the floating range of transaction behavior based on the predetermined implicit weight coefficient and the subsequence of behavioral data corresponding to the user under different transaction categories. The target recommended item determination module is used to determine the transaction trend type corresponding to the user based on the transaction behavior fluctuation attribute and the transaction behavior fluctuation range, and input at least one of the behavior data subsequences into the recommendation model that matches the transaction trend type to obtain the target recommended item corresponding to the user.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to said at least one processor; wherein, The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the article recommendation method according to any one of claims 1-8.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the item recommendation method as described in any one of claims 1-8.

Images