User preference analysis method and system

Abstract

The invention relates to the technical field of big data processing, in particular to a user preference analysis method and system, and the method comprises the steps: inputting a popularity value sequence of each demand of a target user into a long-short-term memory network, obtaining long-term features and short-term features of each demand, and enabling the long-term features of each demand to form a user portrait of the target user; obtaining a correlation matrix of a user group to which the target user belongs, and inputting the short-time feature of each demand and the correlation matrix into a preference analysis model to obtain a popularity attenuation coefficient of each demand; and inputting the real-time popularity value and the popularity attenuation coefficient of the target user for each demand into a preset attenuation model to predict the popularity duration of each demand, and obtaining an analysis result. Through the technical scheme of the invention, the popularity attenuation of the user for each demand can be considered, and an accurate preference analysis result can be obtained.

Description

Technical Field

[0001] The present application relates to the technical field of big data processing, and in particular to a user preference analysis method and system. Background Art

[0002] With the advent of the big data era, the Internet can provide a variety of services such as goods, financial services, multimedia content, etc. In order to recommend services of interest to users, it is necessary to analyze user preferences to improve the user experience when using the Internet.

[0003] At present, a patent application document with publication number CN114840486A discloses a user behavior data collection method, system and cloud platform. The method includes: obtaining a user behavior data set in an Internet financial business session log through a user behavior capture strategy, separating a behavior event session log corresponding to the user behavior data set, and then performing event interest point mining on the behavior event session log through an interest point mining strategy to obtain user interest knowledge fields. The method can accurately and efficiently parse user behavior data sets from Internet financial business session logs of various forms and collection methods, and can also specifically parse user interest knowledge fields from the behavior event session logs corresponding to the user behavior data set, so as to achieve accurate distinction of interest knowledge categories of massive user behavior events.

[0004] The above method mines event interest points on behavioral event session logs and obtains user interest knowledge fields, thereby determining the knowledge categories that users are interested in. However, the degree of user interest in knowledge categories changes gradually. In the process of user preference analysis, ignoring the changes in interest levels will lead to inaccurate user preference analysis results. Summary of the invention

[0005] In order to solve the technical problem of inaccurate user preference analysis results, the present application provides a user preference analysis method and system, which can take into account the user's heat decay for each demand and obtain accurate preference analysis results.

[0006] In a first aspect, the present application provides a user preference analysis method, the analysis method comprising: inputting a heat value sequence of each demand of a target user into a long short-term memory network to obtain long-term and short-term features of each demand, wherein the long-term features of each demand constitute a user portrait of the target user; clustering the user portraits of all users to obtain the user group to which the target user belongs; obtaining a correlation matrix of the user group, and inputting the short-term features and the correlation matrix of each demand into a preference analysis model to obtain a heat attenuation coefficient of each demand, wherein the correlation matrix includes the correlation of heat value changes between any demands; inputting the real-time heat value and heat attenuation coefficient of each demand of the target user into a preference analysis model; A preset attenuation model is used to predict the duration of the popularity of each demand, and the duration of the popularity of each demand is used as an analysis result. The training method of the preference analysis model includes: obtaining the short-term characteristics of each demand and the correlation matrix of the user group to which the user belongs based on the first historical sequence of the popularity values ​​of each demand of any user as a group of training samples; fitting the preset attenuation model in the second historical sequence of the popularity values ​​of each demand after the first historical sequence of the popularity values ​​of the user, and obtaining the attenuation coefficient label of the training sample; inputting the training sample into the preference analysis model to obtain the output result, and calculating the mean square error loss function of the output result and the attenuation coefficient label to iteratively train the preference analysis model.

[0007] Firstly, the heat value sequence of each demand of the target user is input into the long short-term memory network to obtain the long-term and short-term features of each demand, and the long-term features of each demand are used as the user portrait of the target user, which can reflect the long-term trend of the heat values ​​of all demands of the target user; the user portraits of all users are clustered to obtain the user group to which the target user belongs, and the correlation matrix of the user group is obtained. The correlation of the changes in the heat values ​​between any demands within the user group is calculated, which can avoid the correlation differences caused by different user groups and improve the accuracy of the correlation matrix; the short-term features and correlation matrix of each demand are input into the preference analysis model to obtain the heat attenuation coefficient of each demand; the real-time heat value and heat attenuation coefficient of each demand of the target user are input into the preset attenuation model to predict the heat duration of each demand, and the heat duration of each demand is used as the analysis result, which takes into account the heat attenuation of the user to each demand and improves the accuracy of the preference analysis result.

[0008] Preferably, the method for obtaining the heat value includes: obtaining a variety of behavior data of the target user for any demand within a collection cycle, the behavior data including the number of clicks, browsing time and the number of purchases; determining the behavior score of each behavior data according to a preset score range, and weighted summing each behavior score according to a preset weight to obtain the heat value of the target user for the demand during the collection cycle.

[0009] Preferably, the user portraits of all users are clustered to obtain the user group to which the target user belongs, including: clustering the user portraits using the Kmeans algorithm, determining the number of clusters using the elbow method, obtaining multiple clusters and the cluster center of each cluster; calculating the Euclidean distance between the user portrait of the target user and each cluster center, and taking the cluster to which the cluster center corresponding to the minimum Euclidean distance belongs as the user group to which the target user belongs.

[0010] All user portraits are clustered so that the differences between multiple user portraits in a cluster are small, and the user group to which the target user belongs is further determined. The long-term trend of the heat value of each user in this user group is basically the same as that of the target user. The correlation of the heat value changes between any demands in the user group is subsequently calculated, which can avoid the correlation differences caused by different user groups and improve the accuracy of the correlation matrix.

[0011] Preferably, the requirement and demand The calculation method of the correlation of the heat value changes between the requirements includes: for any user in the user group, calculating the heat value difference between the current collection cycle and the previous adjacent collection cycle, obtaining the heat value change sequence of each requirement, and calculating the heat value difference between the user in the requirement and demand The Pearson correlation coefficient of the heat value change sequence between them; calculate the representativeness of the user, which is negatively correlated with the Euclidean distance between the user portrait and the user group cluster center; perform weighted summation of the Pearson correlation coefficients of all users according to the representativeness to obtain the demand and demand The correlation between the changes in heat values.

[0012] The correlation matrix can reflect the correlation between the changes in heat values ​​between any demands, and can provide auxiliary information for predicting the heat decay coefficient of each demand. Using the correlation matrix as the input of the preference analysis model can improve the accuracy of the heat decay coefficient.

[0013] Preferably, the user Representativeness for: , For users in the user group User portraits, is the cluster center of the user group, for and The Euclidean distance of User portraits and The sum of the Euclidean distances.

[0014] There may be differences among multiple user portraits within a user group. The representativeness of each user in the user group is calculated based on the Euclidean distance between the user portrait and the cluster center of the user group. The more representative the user, the greater the impact on the correlation, which improves the accuracy of the correlation.

[0015] Preferably, the preference analysis model includes a convolution sub-model and a regression sub-model; the convolution sub-model is used to extract features from the correlation matrix to obtain correlation features; the correlation features and the short-term features of each demand are concatenated and input into the regression sub-model to output the heat attenuation coefficient of each demand.

[0016] Preferably, the prediction of the heat duration of each demand includes: inputting the real-time heat value and heat attenuation coefficient of the target user for any demand into a preset attenuation model to obtain the predicted heat value of the demand in multiple future collection cycles; taking the first future collection cycle in which the predicted heat value is less than the heat threshold as the cutoff cycle, and taking the number of collection cycles between the current collection cycle and the cutoff cycle as the heat duration of the demand.

[0017] The preset attenuation model is: , is the real-time heat value, is the heat attenuation coefficient, For the A collection cycle, For the The predicted heat value of a collection cycle.

[0018] The calculation method of the mean square error loss function includes: calculating the average variance of the heat value of each demand in the user group to which the user belongs; the mean square error loss function satisfy: , is the quantity of all requirements, Required in the attenuation coefficient label The thermal attenuation coefficient, For the output results The thermal attenuation coefficient, For demand The average variance of the heat value, The sum of the average variance of the heat values ​​of all requirements, It is the sum of the absolute values ​​of the heat attenuation coefficients in the output results.

[0019] In the user group to which the user belongs, if the demand If the long-term trend is stable, it means that the user group is able to meet the demand. Maintain long-term interest, that is, users' demand The heat attenuation coefficient should be close to 0. On the contrary, if the demand If the long-term trend is large fluctuations, it means that the user group is unable to meet the demand Maintain long-term interest, that is, users' demand The heat decay coefficient should be a larger value. Therefore, the stability of the heat values ​​of each demand in the user group to which the user belongs can reflect the relative size of the heat decay coefficients of each demand, provide auxiliary information for the training of the preference analysis model, and ensure the accuracy of the output results of the preference analysis model.

[0020] In a second aspect of the present application, a user preference analysis system is provided, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a user preference analysis method according to the first aspect of the present application is implemented.

[0021] The technical solution of this application has the following beneficial technical effects: Firstly, the heat value sequence of each demand of the target user is input into the long short-term memory network to obtain the long-term and short-term features of each demand, and the long-term features of each demand are used as the user portrait of the target user, which can reflect the long-term trend of the heat values ​​of all demands of the target user; the user portraits of all users are clustered to obtain the user group to which the target user belongs, and the correlation matrix of the user group is obtained. The correlation of the changes in the heat values ​​between any demands within the user group is calculated, which can avoid the correlation differences caused by different user groups and improve the accuracy of the correlation matrix; the short-term features and correlation matrix of each demand are input into the preference analysis model to obtain the heat attenuation coefficient of each demand; the real-time heat value and heat attenuation coefficient of each demand of the target user are input into the preset attenuation model to predict the heat duration of each demand, and the heat duration of each demand is used as the analysis result, which takes into account the heat attenuation of the user to each demand and improves the accuracy of the preference analysis result. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present application will become easy to understand. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein: Figure 1 is a flow chart of a user preference analysis method according to an embodiment of the present application; Figure 2 is a structural diagram of a preference analysis model according to an embodiment of the present application; Figure 3 is a flowchart of a method for training a preference analysis model according to an embodiment of the present application; Figure 4It is a structural block diagram of a user preference analysis system according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0024] It should be understood that when the terms "first", "second", etc. are used in the claims, specification and drawings of the present application, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the specification and claims of the present application indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections.

[0025] According to the first aspect of the present application, the present application provides a user preference analysis method for performing preference analysis on all users within an Internet platform, where the Internet platform can be a commodity sales platform or a financial service platform.

[0026] Figure 1 FIG. 1 is a flow chart of a method for analyzing user preferences according to an embodiment of the present application. Figure 1 As shown, the user preference analysis method includes steps S101 to S104, which are described in detail below.

[0027] S101, input the heat value sequence of each demand of the target user into the long short-term memory network to obtain the long-term features and short-term features of each demand. The long-term features of each demand constitute the user profile of the target user.

[0028] In one embodiment, the target user is any registered user of the Internet platform, and the method for obtaining the heat value includes: obtaining a variety of behavior data of the target user for any demand within a collection period, and the behavior data includes the number of clicks, browsing time and purchase times; determining the behavior score of each behavior data according to a preset score range, and weighted summing each behavior score according to a preset weight to obtain the heat value of the target user for the demand during the collection period.

[0029] The collection period can be 1 day, that is, the heat value of a target user for a demand can be collected every day, and all heat values ​​can be arranged in the order of the collection period to obtain the heat value sequence of the target user for the demand. The demand can be a commodity, a financial product or a multimedia video, and this application does not limit it.

[0030] The preset score interval for each behavior data is pre-set; taking the number of clicks as an example, the preset score interval for a behavior score of 0 is ; The preset score range for a behavior score of 1 is ; The preset score range for a behavior score of 2 is ; The preset score range for a behavior score of 3 is Therefore, when the number of clicks is 7, the behavioral score of the number of clicks is 1.

[0031] After obtaining the behavior score of each behavior data, the behavior scores are weighted and summed according to the preset weights. The preset weights of click count, browsing time and purchase count are: , and ; At this point, the heat value of any demand of the target user within a collection cycle can be obtained.

[0032] In one embodiment, the collection period is set to 1 day, and the heat value sequence of any demand of the target user includes the heat value of each day in the period from the registration date to the current day, wherein the heat value of the current day is the real-time heat value; the heat value sequence of any demand of the target user is input into the long short-term memory network to obtain the long-term characteristics and short-term characteristics of the demand, wherein the long-term characteristics are used to characterize the long-term trend of the heat value of the target user for the demand since the registration date, and the short-term characteristics are used to characterize the short-term trend of the target user for the demand. Since the long-term characteristics can characterize the long-term trend of the heat value of each demand of the target user, the user portrait of the target user can be obtained by splicing the long-term characteristics of each demand together.

[0033] It can be understood that both long-term features and short-term features are vectors with A rows and 1 columns. The value of A is related to the long short-term memory network. The long short-term memory network here can use a pre-trained long short-term memory network to extract features from the heat value sequence. If the number of all demand types is B, the user profile of the target user is a matrix with A rows and B columns.

[0034] In this way, the long short-term memory network is used to extract features of the heat value sequence of each demand of the target user, and a user profile of the target user is constructed based on the extracted long-term features of each demand. The user profile can reflect the long-term trend of the heat values ​​of all demands of the target user.

[0035] S102, clustering the user portraits of all users to obtain the user group to which the target user belongs.

[0036] In one embodiment, since there are many users on the Internet platform, each user corresponds to a user portrait. By clustering all user portraits, the user group to which the target user belongs can be determined. The long-term trend of the popularity value of each user in the user group is basically the same.

[0037] Specifically, the user portraits of all users are clustered to obtain the user group to which the target user belongs, including: clustering the user portraits using the Kmeans algorithm, determining the number of clusters using the elbow method, obtaining multiple clusters and the cluster center of each cluster; calculating the Euclidean distance between the user portrait of the target user and each cluster center, and taking the cluster to which the cluster center corresponding to the minimum Euclidean distance belongs as the user group to which the target user belongs.

[0038] The elbow method is a well-known technique for those skilled in the art and will not be described in detail here. A cluster includes at least one user profile, and the cluster center is the average value of all user profiles in the cluster.

[0039] S103, obtaining a correlation matrix of the user group, and inputting the short-term characteristics and correlation matrix of each demand into a preference analysis model to obtain a heat decay coefficient of each demand, wherein the correlation matrix includes the correlation of heat value changes between any demands.

[0040] In one embodiment, multiple users in a user group have the same interest preference. Calculating the correlation between changes in heat values ​​between any requirements in the user group can avoid correlation differences caused by different user groups and improve the accuracy of the correlation matrix.

[0041] Specifically, the demand and demand The calculation method of the correlation of the heat value changes between the requirements includes: for any user in the user group, calculating the heat value difference between the current collection cycle and the previous adjacent collection cycle, obtaining the heat value change sequence of each requirement, and calculating the heat value difference between the user in the requirement and demand The Pearson correlation coefficient of the heat value change sequence between them; calculate the representativeness of the user, which is negatively correlated with the Euclidean distance between the user portrait and the user group cluster center; perform weighted summation of the Pearson correlation coefficients of all users according to the representativeness to obtain the demand and demand The correlation between the changes in heat values.

[0042] Among them, users Representativeness for: , For users in the user group User portraits, is the cluster center of the user group, for and The Euclidean distance of User portraits and The sum of the Euclidean distances.

[0043] Among them, demand and demand satisfy: , is the number of users in the user group, For users representativeness, For users In demand and demand Pearson correlation coefficient of the heat value change series.

[0044] The correlation matrix can reflect the correlation between changes in heat values ​​between any demands, can provide auxiliary information for predicting the heat decay coefficient of each demand, and using the correlation matrix as the input of the preference analysis model can improve the accuracy of the heat decay coefficient.

[0045] In one embodiment, the preference analysis model is the short-term characteristics and correlation matrix of each demand, and the output is the heat decay coefficient of each demand; Figure 2 This is a structural diagram of a preference analysis model according to an embodiment of the present application. The preference analysis model includes a convolution sub-model and a regression sub-model; the convolution sub-model is used to extract features from the correlation matrix to obtain correlation features; the correlation features and the short-term features of each demand are spliced ​​and input into the regression sub-model to output the heat attenuation coefficient of each demand.

[0046] Among them, the convolution sub-model can adopt existing convolutional neural networks such as ResNet or VGGNet, and the regression sub-model can adopt a fully connected neural network.

[0047] S104, input the real-time popularity value and popularity decay coefficient of each demand of the target user into a preset decay model to predict the popularity duration of each demand, and take the popularity duration of each demand as the analysis result.

[0048] In one embodiment, after obtaining the heat decay coefficient of each demand using the preference analysis model, the duration of the target user's heat for each demand can be predicted based on the target user's real-time heat value for each demand and a preset decay model.

[0049] Specifically, the prediction of the heat duration of each demand includes: inputting the real-time heat value and heat attenuation coefficient of the target user for any demand into a preset attenuation model to obtain the predicted heat value of the demand in multiple future collection cycles; taking the first future collection cycle in which the predicted heat value is less than the heat threshold as the cutoff cycle, and taking the number of collection cycles between the current collection cycle and the cutoff cycle as the heat duration of the demand.

[0050] Wherein, the preset attenuation model is: , is the real-time heat value, is the heat attenuation coefficient, For the A collection cycle, For the The predicted heat value of a collection cycle.

[0051] The value of the heat threshold is 1.5. After the real-time heat value and heat attenuation coefficient of any demand are input into the preset attenuation model, the heat attenuation curve of the demand can be determined, and the predicted heat value of the demand in multiple future collection cycles can be obtained. The first future collection cycle in which the predicted heat value is less than the heat threshold is taken as the cutoff cycle, thereby obtaining the duration of the target user's heat for the demand.

[0052] The duration of popularity of each demand is used as the analysis result. If the duration of popularity of a demand is longer, it means that the target user will maintain interest in the demand for a long time, and the recommendation degree of the demand is greater. Based on the analysis results, the demands of interest can be pushed to the target users.

[0053] It should be noted that the average value of all predicted heat values ​​of any demand between the current collection period and the cut-off period may also be calculated, and the product of the average value and the heat duration may be used as the analysis result.

[0054] In one embodiment, in order for the preference analysis model to output accurate heat decay coefficients, the preference analysis model needs to be trained. Figure 3 , is a flow chart of a method for training a preference analysis model according to an embodiment of the present application. Figure 3 As shown, the user preference analysis method includes steps S201 to S203, which are described in detail below.

[0055] S201 , according to a first historical sequence of heat values ​​of each demand of any user, a short-term feature of each demand and a correlation matrix of the user group to which the user belongs are obtained as a group of training samples.

[0056] The method for obtaining the short-term characteristics of each demand and the correlation matrix of the user group to which the user belongs has been described in detail in steps S101 to S103, and will not be repeated here.

[0057] S202, fitting a preset attenuation model in a second historical sequence of heat values ​​of each demand after the first historical sequence of user heat values, and obtaining an attenuation coefficient label of a training sample.

[0058] Among them, for any demand, the last heat value in the first historical sequence of heat values ​​is used as the real-time heat value in the preset attenuation model; the preset attenuation model is fitted in the second historical sequence of heat values ​​to determine the heat attenuation coefficient of the demand; the heat attenuation coefficients of all demands constitute the attenuation coefficient label of the training sample.

[0059] For example, the first historical sequence of heat values ​​is the heat value sequence of each demand collected from the 1st collection cycle to the 15th collection cycle; the second historical sequence of heat values ​​is the heat value sequence of each demand collected from the 16th collection cycle to the 25th collection cycle after the first historical sequence of heat values; , take the heat value of the 15th collection cycle as the real-time heat value , the demand can be determined by fitting the preset decay model in the second historical sequence of heat value using the least square method The heat attenuation coefficient.

[0060] S203, inputting the training sample into the preference analysis model to obtain an output result, and calculating the mean square error loss function of the output result and the attenuation coefficient label to iteratively train the preference analysis model.

[0061] Among them, iterative training of the preference analysis model using the gradient descent method is a well-known technology for those skilled in the art and will not be described in detail here. The preference analysis model is iteratively trained multiple times until the value of the mean square error loss function is less than the preset loss, and the training of the preference analysis model is completed. Among them, the preset loss value is 0.01.

[0062] Mean square error loss function satisfy: , is the quantity of all requirements, Required in the attenuation coefficient label The thermal attenuation coefficient, For the output results The heat attenuation coefficient.

[0063] It should be noted that different user groups have different long-term trends in the heat values ​​of each demand. If the long-term trend is stable, it means that the user group is able to meet the demand. Maintain long-term interest, that is, the user's demand The heat attenuation coefficient should be close to 0. On the contrary, if the demand If the long-term trend is large fluctuations, it means that the user group is unable to meet the demand Maintain long-term interest, that is, the user's demand The heat decay coefficient should be a larger value. Therefore, the heat value stability of each demand in the user group to which the user belongs can be calculated. The heat value stability can reflect the relative size of the heat decay coefficients of each demand, provide auxiliary information for the training of the preference analysis model, and ensure the accuracy of the output results of the preference analysis model.

[0064] Specifically, the calculation method of the mean square error loss function includes: calculating the average variance of the heat value of each demand in the user group to which the user belongs; the mean square error loss function satisfy: , is the quantity of all requirements, Required in the attenuation coefficient label The thermal attenuation coefficient, For the output results The thermal attenuation coefficient, For demand The average variance of the heat value, The sum of the average variance of the heat values ​​of all requirements, It is the sum of the absolute values ​​of the heat attenuation coefficients in the output results.

[0065] Among them, demand The calculation process of the average variance of the heat value is as follows: in the user group to which the user belongs, calculate each user demand The variance of the heat value sequence, and the average of all variances is used as the requirement The average square deviation of the heat values.

[0066] The above introduces the technical principles and implementation details of a user preference analysis method of the present application through specific embodiments. First, the heat value sequence of each demand of the target user is input into the long short-term memory network to obtain the long-term characteristics and short-term characteristics of each demand, and the long-term characteristics of each demand are used as the user portrait of the target user, which can reflect the long-term trend of the heat value of all demands of the target user; the user portraits of all users are clustered to obtain the user group to which the target user belongs, and the correlation matrix of the user group is obtained. The correlation of the heat value changes between any demands within the user group is calculated, which can avoid the correlation differences caused by different user groups and improve the accuracy of the correlation matrix; the short-term characteristics and correlation matrix of each demand are input into the preference analysis model to obtain the heat attenuation coefficient of each demand; the real-time heat value and heat attenuation coefficient of each demand of the target user are input into the preset attenuation model to predict the heat duration of each demand, and the heat duration of each demand is used as the analysis result, which takes into account the heat attenuation of the user to each demand and improves the accuracy of the preference analysis result.

[0067] According to the second aspect of the present application, the present application also provides a user preference analysis system. Figure 4 is a structural block diagram of a user preference analysis system according to an embodiment of the present application. Figure 4 As shown, the system 50 includes a processor and a memory, the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a user preference analysis method according to the first aspect of the present application is implemented. The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface, whose settings and functions are known in the art, and therefore will not be described in detail here.

[0068] In the present application, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. For example, a computer-readable storage medium may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in this application may be implemented by computer-readable / executable instructions stored or otherwise maintained by such a computer-readable medium.

[0069] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A user preference analysis method, characterized in that: The analysis method comprises: Input the heat value sequence of each demand of the target user into the long short-term memory network to obtain the long-term and short-term features of each demand. The long-term features of each demand constitute the user profile of the target user. Cluster the user portraits of all users to obtain the user group to which the target user belongs; Obtain a correlation matrix of the user group, and input the short-term characteristics and correlation matrix of each demand into the preference analysis model to obtain the heat decay coefficient of each demand, wherein the correlation matrix includes the correlation of heat value changes between any demands; Input the real-time popularity value and popularity decay coefficient of each demand of the target user into the preset decay model to predict the popularity duration of each demand, and take the popularity duration of each demand as the analysis result; The training method of the preference analysis model includes: obtaining the short-term characteristics of each demand and the correlation matrix of the user group to which the user belongs based on the first historical sequence of the heat values ​​of each demand of any user as a set of training samples; fitting a preset attenuation model in the second historical sequence of the heat values ​​of each demand after the first historical sequence of the user's heat values ​​to obtain the attenuation coefficient label of the training sample; inputting the training sample into the preference analysis model to obtain the output result, and calculating the mean square error loss function of the output result and the attenuation coefficient label to iteratively train the preference analysis model.

2. A user preference analysis method according to claim 1, characterized in that: The method for obtaining the heat value includes: Obtaining multiple behavioral data of target users for any needs within a collection cycle, the behavioral data including the number of clicks, browsing time and purchase times; The behavior score of each behavior data is determined according to a preset score range, and each behavior score is weighted and summed according to a preset weight to obtain the heat value of the target user for the demand during the collection period.

3. A user preference analysis method according to claim 1, characterized in that: Cluster the user portraits of all users to obtain the user group to which the target user belongs, including: The Kmeans algorithm is used to cluster user portraits, and the elbow method is used to determine the number of clusters, and multiple clusters and the cluster centers of each cluster are obtained; The Euclidean distance between the user profile of the target user and each cluster center is calculated, and the cluster to which the cluster center corresponding to the minimum Euclidean distance belongs is taken as the user group to which the target user belongs.

4. A user preference analysis method according to claim 1, characterized in that: need and demand The calculation method of the correlation of the change of the heat value includes: For any user in the user group, calculate the difference between the heat value of the current collection cycle and the previous adjacent collection cycle, obtain the heat value change sequence of each demand, and calculate the heat value change sequence of the user in the demand. and demand Pearson correlation coefficient of the heat value change sequence between them; Calculate the representativeness of the user, where the representativeness is negatively correlated with the Euclidean distance between the user portrait and the user group cluster center; The Pearson correlation coefficients of all users are weighted and summed according to representativeness to obtain the demand and demand The correlation between the changes in heat values.

5. A user preference analysis method according to claim 4, characterized in that: user Representativeness for: , For users in the user group User portraits, is the cluster center of the user group, for and The Euclidean distance of User portraits and The sum of the Euclidean distances.

6. A user preference analysis method according to claim 1, characterized in that: The preference analysis model includes a convolution sub-model and a regression sub-model; the convolution sub-model is used to extract features from the correlation matrix to obtain correlation features; the correlation features and the short-term features of each demand are spliced ​​and input into the regression sub-model to output the heat attenuation coefficient of each demand.

7. A user preference analysis method according to claim 1, characterized in that: The predicted heat duration of each demand includes: Input the real-time heat value and heat attenuation coefficient of any demand of the target user into a preset attenuation model to obtain the predicted heat value of the demand in multiple future collection cycles; The first future collection cycle in which the predicted heat value is less than the heat threshold is taken as the cutoff cycle, and the number of collection cycles between the current collection cycle and the cutoff cycle is taken as the heat duration of the demand.

8. A user preference analysis method according to claim 1, characterized in that: The preset attenuation model is: , is the real-time heat value, is the heat attenuation coefficient, For the A collection cycle, For the The predicted heat value of a collection cycle.

9. A user preference analysis method according to claim 1, characterized in that: The calculation method of the mean square error loss function includes: Calculate the average variance of the heat value of each demand in the user group to which the user belongs; mean square error loss function satisfy: , is the quantity of all requirements, Required in the attenuation coefficient label The thermal attenuation coefficient, For the output results The thermal attenuation coefficient, For demand The average variance of the heat value, The sum of the average variance of the heat values ​​of all requirements, It is the sum of the absolute values ​​of the heat attenuation coefficients in the output results.

10. A user preference analysis system, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a user preference analysis method according to any one of claims 1 to 9 is implemented.

Images