A user interest mining method and system in a second-hand e-commerce scenario

Abstract

This invention relates to a method and system for mining user interests in a second-hand e-commerce scenario. The method includes: acquiring behavioral data of target users within a preset time period according to product categories; statistically analyzing the behavioral data to determine the target user's behavior type j, the frequency a of each behavior, and the duration d; and calculating the weight w of each behavior of the target user according to a formula based on the frequency a, duration d, and standard weight W corresponding to the behavior type. j The invention also calculates the interest values ​​of target users within the product categories according to a formula. Even when user behavior is sparse, this invention can accurately reveal the degree of user interest in each product category, providing effective reference information for product recall during downstream product recommendation processes.

Description

Technical Field

[0001] This invention relates to the field of internet application technology, and in particular to a method and system for mining user interests in a second-hand e-commerce scenario. Background Technology

[0002] With the rapid development of science and technology, the internet has become an indispensable part of people's lives. More and more merchants and consumers prefer to complete transactions on e-commerce platforms. These platforms aggregate a large number of merchants, ordinary goods, and advertised products. Compared to primary e-commerce platforms, secondary e-commerce platforms have very unique characteristics. For example, most ordinary and advertised goods are unique items with their own distinctive features, making it difficult for buyers to find satisfactory products. Furthermore, user behavior on secondary e-commerce platforms is sparse, and user retention rates are low. Therefore, helping users on secondary platforms find or recommending products that match their interests is an effective way to improve user experience and attract more users. Summary of the Invention

[0003] To address the technical problems existing in the prior art, this invention proposes a user interest mining method and system in the context of second-hand e-commerce. By mining user interests, it provides effective data support for downstream services such as product recall and recommendation.

[0004] To address the aforementioned technical problems, according to one aspect of the present invention, a method for mining user interests in a second-hand e-commerce scenario is provided, comprising the following steps: obtaining behavioral data of target users within a preset time period according to product categories; statistically analyzing the behavioral data to determine the target user's behavior type j, the frequency a of each behavior, and the duration d; and calculating the weight w of each behavior of the target user according to Formula 1-1 based on the frequency a, duration d, and standard weight W corresponding to the behavior type. j :

[0005]

[0006] Among them, w j Let be the weight of the j-th behavior; Ti is the time decay factor of the i-th behavior;

[0007] W j V represents the standard weight corresponding to the j-th behavior. T The time threshold is used; and the interest value of the target user under the product category is calculated according to formula 1-2:

[0008]

[0009] Among them, Interest idenoted as the interest value for the i-th product category; k represents the total number of behavioral categories in the i-th product category.

[0010] To address the aforementioned technical problems, according to one aspect of the present invention, a user interest mining system for a second-hand e-commerce scenario is provided, comprising a data acquisition module, a data statistics module, a single behavior weight calculation module, and an interest value calculation module; wherein, the data acquisition module is configured to acquire behavioral data of target users within a preset time period according to product categories; the data statistics module is connected to the data acquisition module and is configured to statistically analyze the target user's behavior type j, the frequency a of each behavior, and the duration d; the single behavior weight calculation module is connected to the data statistics module and is configured to calculate the weight w of each behavior of the target user according to the aforementioned formula 1-1. j The interest value calculation module is connected to the single behavior weight calculation module, and is configured to calculate the interest value of the target user under the product category according to formula 1-2.

[0011] This invention can still uncover users' interest levels across various product categories even when user behavior is sparse, and uses numerical values ​​to represent the intensity of interest. This provides valuable reference information for product recall during downstream product recommendation processes, ensuring that downstream services recommend products that match user interests. This means providing users with the products or information they need or are interested in, while also increasing the exposure of advertised products. This benefits both advertisers and secondhand e-commerce platforms. Furthermore, high-quality product recommendations provide users with a better shopping experience, effectively improving user retention rates. Attached Figure Description

[0012] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0013] Figure 1 This is a flowchart of a user interest mining method in a second-hand e-commerce scenario according to an embodiment of the present invention;

[0014] Figure 2 This is a partial flowchart of a user interest mining method in a second-hand e-commerce scenario according to another embodiment of the present invention;

[0015] Figure 3 This is a partial flowchart of a user interest mining method in a second-hand e-commerce scenario according to another embodiment of the present invention;

[0016] Figure 4 This is a flowchart illustrating the determination of standard weights according to an embodiment of the present invention;

[0017] Figure 5This is a schematic diagram of a user interest mining system in a second-hand e-commerce scenario according to an embodiment of the present invention.

[0018] Figure 6 This is a schematic diagram of a user interest mining system in a second-hand e-commerce scenario according to another embodiment of the present invention;

[0019] Figure 7 This is a principle block diagram of a user interest mining system in a second-hand e-commerce scenario according to another embodiment of the present invention; and

[0020] Figure 8 This is a schematic diagram of a user interest mining system in a second-hand e-commerce scenario according to another embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0023] Second-hand e-commerce platforms typically feature a wide variety of product categories, such as clothing, electronics, and home furnishings, with further subcategories within each category. Most users exhibit preferences for certain categories over a period of time. For example, after moving and wanting to buy furniture, a user might conduct numerous searches in the furniture category on the platform within the past week or month. This invention leverages this phenomenon by analyzing user behavior data on the platform to determine their recent interests. When a user triggers recommendation criteria, such as browsing the platform's main page or visiting a page with advertising space, the system can recommend advertised products in the same category as the user's recent interests. This provides users with the product information they need, increases the exposure of advertised products, and ultimately boosts revenue for both advertisers and the platform.

[0024] Figure 1 This is a flowchart of a user interest mining method in a second-hand e-commerce scenario according to an embodiment of the present invention. In this embodiment, the interest values ​​of all users on the second-hand e-commerce platform in various product categories are mined, and the intensity of interest is expressed in numerical form. The method includes the following steps:

[0025] Step S1a: Determine a target user from the user table. In this embodiment, the platform's database stores the IDs of all users on the platform, recorded in a table. Starting from the table, each user is retrieved sequentially and used as the target user to calculate their interest value in each category. For example, in the first implementation of the method, the first user is designated as the target user. When a new user registers, the new user is designated as the target user.

[0026] Step S2a: Obtain the target user's behavioral data on the platform within a preset time period. This behavioral data includes data generated from various user behaviors on the platform, such as browsing, saving, inquiring, placing orders, and commenting. The platform detects and records user behavior on the platform, recording these behaviors in log form. In addition to recording behavior types, such as browsing and saving, the platform also records the time of occurrence, duration, and corresponding product or seller information. The preset time period may be, for example, the most recent week or the most recent three days.

[0027] Step S3a: Classify and statistically analyze the target user's behavioral data to obtain the target user's behavioral types j in each product category, the number of occurrences a of each behavior, and the duration d of each behavior.

[0028] Step S4a: Retrieve behavioral data for a product category. This includes all behavioral types j under a specific product category, the number of times each behavior occurs a, and the duration d of each behavior.

[0029] Step S5a: Obtain one type of behavioral data, namely, the number of occurrences a for a specific type of behavior and the duration d of each behavior.

[0030] Step S6a: Calculate the weight w of this type of behavior of the target user according to Formula 1-1. j :

[0031]

[0032] Among them, w j Let T be the weight of the j-th behavior, and a be the number of behaviors; i is the time decay factor for the i-th action. W j V represents the standard weight corresponding to the j-th behavior.T This is the time threshold. When the first behavior type is currently selected, this step calculates the weight w1 of the first behavior; the time decay factor T1 of the first behavior; and the standard weight W1 corresponding to the first behavior.

[0033] In this step, behaviors with excessively short durations are filtered out by setting a time threshold, thereby reducing the resulting errors. Furthermore, the further back in time an behavior occurs from the current mining time, the lower its likelihood of being considered an interest. Therefore, a time decay factor is set, which decreases according to a certain pattern based on the time interval between the behavior's occurrence and the current mining time, thus improving the accuracy of the final interest value. The time decay factor Ti can be a linear function, a logarithmic function, etc. In a better embodiment, the time decay factor Ti is determined using the following formulas 1-3:

[0034]

[0035] Wherein, m, n, and l are fixed constants, and Interval i Let l be the time interval between the occurrence of the i-th action and the current time. In one embodiment, l represents the furthest time interval among the actions of this type, m is a constant less than l, and n = 1 - m. By setting the constant m, we can ensure that actions that occur too early also have a certain weight.

[0036] Step S7a: Determine if there are any uncalculated behavior types under the product category. If there are, return to step S5a; otherwise, proceed to step S8a.

[0037] Step S8a: Calculate and record the interest value of the target user under the product category according to formula 1-2:

[0038]

[0039] Among them, Interest i The interest value for the i-th product category is calculated when the first product category is currently selected; k represents the total number of behavioral categories in the i-th product category. After calculating the interest value for the product category, it is recorded in the user interest value table. This user interest value table includes all product categories on the platform for which user interest values ​​can be calculated, along with the user IDs. When the interest value for a user in a product category is obtained, it is entered into the corresponding product category column in the table.

[0040] Step S9a: Determine if there are any uncalculated product categories. If so, return to step 4a. If not, it means that the interest values ​​of all possible product categories of the target user have been mined. Then, in step S10a, improve the user interest value table by adding 0 to the product category column where the target user's interest value does not exist. Then, execute step S11a.

[0041] Step S11a: Determine if there are any users whose interests have not yet been explored. If so, return to step S1a; otherwise, end the exploration process.

[0042] Figure 2 This is a partial flowchart of a user interest mining method in a second-hand e-commerce scenario according to another embodiment of the present invention. In this embodiment, refer to... Figure 1 When according to Figure 1 After the process reaches step S4a, it obtains all the behavior data of a target product category, such as the number of times each behavior occurs (a) and the duration of each behavior (d), and then executes step S1b.

[0043] Step S1b: Compare the differences Da between the frequency of occurrence of various behaviors. For example, calculate the difference a between the frequency of occurrence of each pair of behaviors, and obtain a frequency difference Da for each pair of different types of behaviors.

[0044] Step S2b: Determine whether at least one occurrence difference Da is greater than or equal to the first occurrence threshold V. a1 If at least one occurrence difference Da is greater than or equal to the first occurrence threshold V a1 Then proceed to step S3b. If the difference in the number of occurrences Da for all occurrences is less than the threshold V for the first occurrence, then proceed to step S3b. a1 Then proceed to step S5a. And follow the instructions... Figure 1 The process shown completes the entire excavation process.

[0045] Step S3b: Take the behavioral data of one of the behaviors.

[0046] Step S4b: Calculate the weight w for each behavior according to formulas 1-4. j :

[0047]

[0048] By taking the logarithm of the weights to the base e, the weights of behaviors that occur too frequently can be reduced, thereby reducing the large numerical differences in the final interest values. Then, step S7a is executed, and proceeds according to... Figure 1 The process shown completes the entire excavation process.

[0049] In this embodiment, by analyzing whether the frequency of different types of behaviors differs too much, when the difference is not significant, Formula 1-1 is used to calculate the weight of each behavior, and when the difference is too significant, Formula 1-4 is used to calculate the weight of each behavior. This ensures that the final interest value will not differ too much due to the difference in the number of behaviors.

[0050] Figure 3 This is a partial flowchart of a user interest mining method in a second-hand e-commerce scenario according to another embodiment of the present invention. In this embodiment, refer to... Figure 1 When according to Figure 1 After the process reaches step S4a, all behavioral data for a target product category is obtained, such as the number of times each behavior occurs (a) and the duration of each behavior (d). Then, step S1c is executed.

[0051] Step S1c: Compare the number of occurrences 'a' for each behavior with the second threshold value 'V'. a1 The size of . Wherein, the second threshold value V a1 This is used to measure whether the number of behaviors is too low. When the number of behaviors of a certain type is too low, it will cause the interest mining to go astray. Therefore, in this embodiment, the occurrence of each behavior is monitored to ensure that it reaches a sufficient number.

[0052] Step 2c: Determine whether the number of times the current type of behavior occurs, 'a', is less than the second threshold value, V. a1 If it is less than, proceed to step S3c; if it is greater than or equal to, proceed to step S3c. Figure 1 5a or Figure 2 1b.

[0053] Step S3c: Obtain the similar user group of the user. In one embodiment, based on user characteristics, such as clicked, viewed, or purchased items, a dual-tower model is used to calculate the distance between users. Users whose distance is less than a threshold are identified as similar users of the target user, thereby obtaining the similar user group.

[0054] Step S4c: Obtain the frequency of similar behaviors among similar users. For example, read behavioral data of similar users from user logs and count the frequency of each type of behavior in each product category.

[0055] Step S5c: Calculate the average number of occurrences of the current type of behavior for all similar users.

[0056] Step S6c, using the average number of occurrences The number of times 'a' represents the type of behavior of the target user occurs.

[0057] Then execute Figure 15a or Figure 2 1b in the middle.

[0058] In this embodiment, considering that sparse user behavior in the second-hand scenario may lead to inaccurate interest mining when a certain behavior occurs too infrequently, a threshold is set. When the frequency of a user's certain behavior is less than the threshold, the average frequency of that behavior of similar users is used to replace the frequency of that user's behavior, thereby obtaining the user's interest value more accurately.

[0059] In the above embodiments, the standard weights used when calculating user interest values ​​can be set manually, or they can be automatically learned using a deep learning model based on user behavior data to obtain the standard weights for each behavior. In one embodiment, such as Figure 4 The diagram shown is a flowchart for determining the standard weights.

[0060] Step S1d: Obtain training samples. For example, obtain multiple user behavior data within a preset time period, classify and statistically analyze them to obtain the frequency and duration of each behavior for each product category.

[0061] Step 2d involves labeling the samples. The label is determined by whether a user placed an order. For example, a sample labeled with an order is labeled 1, and a sample labeled with no order is labeled 0.

[0062] Step S3d: Train the logistic regression model. This embodiment uses a logistic regression model, but other deep learning models, such as support vector machines, gradient boosting trees, etc., can also be used. The behavior type is used as a feature, and the frequency of that behavior is used as the feature value, which is then input into the logistic regression model for training. If a user does not exhibit a certain behavior, its feature value is set to 0. After repeated training and adjustments, a model that meets the evaluation criteria is obtained. The objective function f of this model is then obtained as shown in Formula 1-5:

[0063] f=sigmoid(W1·x1+W2·x2+W3·x3+…) 1-5

[0064] In this embodiment, x1, x2, x3... are features, corresponding to behavior types, and W1, W2, W3... are weights corresponding to each feature.

[0065] In step S4d, the weight W corresponding to feature x in formulas 1-5 is determined as the standard weight of the corresponding type of behavior.

[0066] The above methods and steps can yield standard weights for a user's various behaviors across different product categories over a given period. Considering that user interests may change over time—for example, a user might be interested in furniture one period and then in cosmetics the next—in one embodiment of this invention, to recommend products matching a user's interests when they log in, the method further includes the following steps:

[0067] Step S5d: Timing. After obtaining the standard weights from the model's objective function, timing is initiated. The timing period can be, for example, 1 day, 3 days, 1 week, etc. Alternatively, the timing period can be flexibly varied based on factors such as the user's login time, number of logins, and frequency.

[0068] Step 6d: Determine if the timing period has been reached. If the timing period has been reached, proceed to step S7d. Otherwise, return to this step and monitor whether the timing period has been reached.

[0069] Step S7d: Determine whether to stop updating. For example, if the user cancels registration and leaves the platform, it is no longer necessary to monitor their interests, so updates should be stopped and the process ends. If the user has not left the platform and it is necessary to continue monitoring their interests, return to step S1d, obtain new training samples, and retrain the model to obtain new standard weights.

[0070] After the standard weights are updated, the user's interest values ​​can be updated simultaneously, thus ensuring that the data in the user interest value table stored in the current system is the user's latest interest value.

[0071] Figure 5 This is a schematic diagram of a user interest mining system in a second-hand e-commerce scenario according to an embodiment of the present invention. The system includes a data acquisition module 1, a data statistics module 2, a single-behavior weight calculation module 3, and an interest value calculation module 4. The data acquisition module 1 determines target users from a user table and obtains behavioral data of the target users for each product category within a preset time period from user logs. The data statistics module 2 is connected to the data acquisition module 1 and is configured to statistically analyze the target users' behavior types j, the frequency a of each behavior, and the duration d in different product categories. The single-behavior weight calculation module 3 is connected to the data statistics module 2 and obtains the required parameters and data from preset parameters, calculating the weight w of each behavior of the target users according to Formula 1-1. j :

[0072]

[0073] Among them, w j Let be the weight of the j-th behavior; Ti is the time decay factor of the i-th behavior;

[0074] W j V represents the standard weight corresponding to the j-th behavior. T This is the time threshold.

[0075] The single behavior weight calculation module 3 can be one or more. When there is one, it calculates the behavior weight of each behavior type in turn. If there are multiple modules, the behavior weight of each behavior type can be calculated simultaneously.

[0076] The interest value calculation module 4 is connected to the single behavior weight calculation module 3. It obtains the standard weight of each behavior from preset parameters and calculates the interest value of the target user under the product category according to formula 1-2:

[0077]

[0078] Among them, Interest i denoted as the interest value for the i-th product category; k represents the total number of behavioral categories in the i-th product category.

[0079] Figure 6 This is a schematic diagram of a user interest mining system in a second-hand e-commerce scenario according to another embodiment of the present invention. In addition to a data acquisition module 1, a data statistics module 2, a single behavior weight calculation module 3, and an interest value calculation module 4, the system also includes a first behavior frequency comparison module 5, which is connected to the data statistics module 2 and the single behavior weight calculation module 3. The first behavior frequency comparison module compares the difference between the occurrence frequencies 'a' of each behavior and sends the comparison result to the single behavior weight calculation module 3. The single behavior weight calculation module 3 receives the comparison result and determines whether the difference between the occurrence frequencies 'a' of at least two behaviors is greater than or equal to the first frequency threshold V. a1 At that time, the weight w of each behavior of the target user is calculated according to formulas 1-4. j The difference between the number of occurrences (a) of all types of behavior is less than the first occurrence threshold (V). a1 At that time, the weight w of each behavior of the target user is calculated according to Formula 1-1. j .

[0080]

[0081]

[0082] Figure 7This is a schematic diagram of a user interest mining system in a second-hand e-commerce scenario according to another embodiment of the present invention. In this embodiment, in addition to the data acquisition module 1, data statistics module 2, single behavior weight calculation module 3, interest value calculation module 4, and first behavior frequency comparison module 5, the system also includes a first behavior frequency comparison module 6 and a behavior frequency determination module 7. The second behavior frequency comparison module 6 is connected to the data statistics module 2 and is used to compare the occurrence frequency 'a' of each behavior with the second frequency threshold V. a2 When the number of occurrences of behavior 'a' is less than the threshold V for the second occurrence. a2 When the number of occurrences of the behavior a is greater than or equal to the second threshold V, a notification is sent to the behavior count determination module 7. a2 At that time, a notification is sent to the first behavior frequency comparison module 5. The behavior frequency determination module 7 calculates the average occurrence frequency of the same type of behavior among all similar user groups of the target user. The frequency of the aforementioned type of behavior of the target user is denoted as 'a'. If the system has already calculated and maintained a similar user group for each user, the behavior frequency determination module 7 queries the database to obtain the similar user group for the target user. If the system does not maintain a similar user group for the user, the similar user acquisition module 8 calculates the similar user group for the target user. For example, using clicked, viewed, and purchased items as user characteristics, a dual-tower model is used to calculate the distance between users, and multiple users whose distance is less than a threshold are identified as similar users of the target user, thus obtaining the similar user group.

[0083] Figure 8 This is a principle block diagram of a user interest mining system in a second-hand e-commerce scenario according to another embodiment of the present invention. In this embodiment, in Figure 5-7 Each embodiment includes a standard weight determination module 9 and a timing module 10. To clearly illustrate the key points of this embodiment, Figure 8 Other related terms are omitted. Figure 5-7 Similar structures or connections are found within the data. The standard weight determination module 9 includes a sample acquisition unit 91, a statistics unit 92, a model training unit 93, and a weight extraction unit 94. The sample acquisition unit 91 acquires behavioral data of multiple users within a preset time period for various behaviors in each product category as training samples for deep learning models, such as logistic regression models and support vector machine models, and labels the corresponding samples based on whether the user placed an order. For example, behavioral data from the last 3 days or the last week is acquired as samples. Each sample is labeled according to whether the user placed an order during that period; for example, samples from users who placed orders are labeled with 1, and samples from users who did not place orders are labeled with 0.

[0084] The statistics unit 92 is connected to the sample acquisition unit 91 and is used to count the number of times each type of behavior is performed for each user. The model training unit 93 is connected to the statistics unit 92 and calls the logistic regression model, using the user behavior type as a feature and the number of times that type of behavior is performed as a feature value, and inputs it into the logistic regression model for training. After training, the objective function f is obtained as shown in Formulas 1-5.

[0085] f=sigmoid(W1·x1+W2·x2+W3·x3+…) 1-5

[0086] Wherein, x1, x2, x3... are features corresponding to behavior types, and W1, W2, W3... are the weights corresponding to the aforementioned features.

[0087] The weight extraction unit 94 is connected to the model training unit 93. After the model training is completed, the weights corresponding to the features x in formulas 1-5 are determined as the standard weights for the corresponding types of behaviors. Furthermore, after obtaining the standard weights, the timing module 10 is triggered to start timing. When the timing period is reached, the sample acquisition unit 91 of the standard weight determination module 9 acquires new samples to redetermine the standard weights.

[0088] This invention determines a user's interest values ​​across multiple product categories based on user behavior data over a period of time, ultimately resulting in a user interest table. The numerical values ​​representing strong interests can provide reference information for downstream product recommendations, ensuring that recommended products align with the user's interests. This provides users with the information they need or are interested in, increases the exposure of advertised products, and improves both advertiser and secondhand e-commerce platform revenue. Furthermore, high-quality product recommendations provide users with a better shopping experience, effectively improving user retention rates.

[0089] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. A method for mining user interests in a second-hand e-commerce scenario, including: Obtain behavioral data of target users within a preset time period based on product category; Based on the behavioral data, the target user's behavior type j, the number of times each behavior occurs a, and the duration d are statistically analyzed; The weight of each behavior of the target user is calculated according to Formula 1-1 based on the occurrence frequency 'a', duration 'd', and standard weight W corresponding to the behavior type. : 1-1 in, Let be the weight of the j-th behavior; Ti is the time decay factor of the i-th behavior; , Let the standard weight be the weight corresponding to the j-th behavior. For time threshold; and Calculate the interest value of the target user under the aforementioned product category using formula 1-2: 1-2 in, Let be the interest value for the i-th product category; k is the total number of behavioral categories in the i-th product category. This further includes comparing the differences in the frequency of occurrence of multiple behaviors, where the difference Da between the frequency of occurrence of one or more pairs of behaviors is greater than or equal to the first threshold V. a1 Then, calculate the weight of each behavior of the target user according to formulas 1-4. : 1-4。 2. The method according to claim 1, wherein the time decay factor Ti is: 1-3 in, m, n, and l are fixed constants, and the Interval i Let be the time interval between the occurrence of the i-th action and the current time.

3. The method according to claim 1, wherein when counting the target user's behavior type j and the number of occurrences a of each behavior a, it further includes: The number of times the behavior occurs, 'a', is compared with the second threshold value, V. a2 Compare; In response to the behavior occurring a less than the second threshold V a2 Obtain the similar user group of the target user; Calculate the average number of occurrences of the aforementioned type of behavior among similar user groups. ; as well as With the average number of occurrences The number of times 'a' represents the type of behavior of the target user occurs.

4. The method of claim 1, further comprising setting a standard weight W for each behavior under each product category.

5. The method of claim 4, wherein the step of setting the standard weight W for each behavior under each product category further comprises: The behavioral data of multiple users in each product category within a preset time period are used as training samples for a deep learning model. Whether a user places an order is used as the label for the training samples; Count the number of times each user performs each type of action; The behavior type is used as a feature, and the number of times that behavior is used as a feature value is input into the deep learning model for training. After training, the objective function f shown in Formula 1-5 is obtained. 1-5 Wherein, x1, x2, x3... are features corresponding to behavior types; W1, W2, W3... are the weights corresponding to the aforementioned features; as well as The weight W corresponding to feature x in Formula 1-5 is determined as the standard weight of the corresponding type of behavior.

6. The method according to claim 4 or 5, further comprising: Timing begins after the standard weights are obtained; as well as In response to the completion of the timer period, reset the standard weight W for each behavior under each product category.

7. A user interest mining system for a second-hand e-commerce scenario, comprising: The data acquisition module is configured to acquire behavioral data of target users within a preset time period according to product category; The data statistics module, which is connected to the data acquisition module, is configured to count the target user's behavior type j, the number of occurrences a of each behavior, and the duration d. A single-behavior weight calculation module, connected to the data statistics module, is configured to calculate the weight of each behavior of the target user according to formula 1-1. : 1-1 in, Let be the weight of the j-th behavior; Ti is the time decay factor of the i-th behavior; , Let the standard weight be the weight corresponding to the j-th behavior. For time threshold; and The interest value calculation module, connected to the single behavior weight calculation module, is configured to calculate the interest value of the target user under the product category according to formula 1-2: 1-2 in, Let be the interest value for the i-th product category; k is the total number of behavioral categories in the i-th product category. The system includes a first behavior frequency comparison module, which is connected to both the data statistics module and the single behavior weight calculation module. This first behavior frequency comparison module is configured to compare the difference between the occurrence frequencies of two behaviors. When the difference Da between the occurrence frequencies of one or more pairs of behaviors is greater than or equal to a first threshold V, the comparison is successful. a1 When a notification is sent to the single-behavior weight calculation module, the single-behavior weight calculation module calculates the weight of each behavior of the target user according to formulas 1-4 after receiving the notification. : 1-4。 8. The system according to claim 7, further comprising a second behavior frequency comparison module and a behavior frequency determination module, wherein the second behavior frequency comparison module is connected to the data statistics module and configured to compare the occurrence frequency 'a' of the behavior with a second threshold value 'V'. a2 When the number of occurrences of behavior 'a' is less than the threshold V for the second occurrence. a2 A notification is sent to the behavior frequency determination module, and after receiving the notification, the behavior frequency determination module calculates the average occurrence frequency of the same type of behavior among similar user groups of the target user. The number of times 'a' represents the type of behavior of the target user occurs.

9. The system of claim 7, further comprising a standard weight determination module, which includes: The sample acquisition unit is configured to acquire behavioral data of multiple users in each product category within a preset time period as training samples for the deep learning model, and use whether the user places an order as the label of the training sample. A statistics unit, connected to the sample acquisition unit, is configured to count the number of times each user performs each type of behavior. The model training unit, which is connected to the statistical unit, is configured to use user behavior types as features and the number of such behaviors as feature values ​​as inputs to the deep learning model for training. After training, the objective function f shown in Formulas 1-5 is obtained. 1-5 Wherein, x1, x2, x3... are features corresponding to behavior types; W1, W2, W3... are the weights corresponding to the aforementioned features; as well as The weight extraction unit, after configuration, determines the weights corresponding to features x in formulas 1-5 as the standard weights of the corresponding types of behaviors.

10. The system according to claim 9, further comprising a timing module connected to the standard weight determination module, wherein the timing module starts timing when the standard weight determination module obtains the standard weight, and the standard weight determination module re-determines the standard weight when the timing period is reached.

Images