Flow control method and system based on e-commerce platform, electronic equipment and storage medium

Abstract

The invention provides a flow control method and system based on an e-commerce platform, electronic equipment and a storage medium. The flow control method comprises the following steps: determining a natural merchant and a flow control merchant based on a search and push general link in response to a user request; determining a merchant set according to the natural merchants and the flow control merchants; determining sorting values of all merchants in the merchant set; performing modeling solution on all the merchants based on a first preset constraint condition and the sorting value to obtain a target merchant set needing to be displayed in the merchant set; performing modeling solution on the target merchant set based on a second preset constraint condition and the sorting value to obtain display positions of merchants in the target merchant set; wherein the first preset constraint condition at least comprises an exposure distribution difference constraint, an exposure total amount constraint and maximum scene efficiency; the second preset constraint condition at least comprises the efficiency loss constraint and the maximum flow control flow.

Description

Technical Field

[0001] The present invention relates to the technical field of merchant traffic support based on an e-commerce platform, and in particular to a traffic control method and system, electronic equipment and storage medium based on an e-commerce platform. Background Art

[0002] An e-commerce platform is an online trading platform based on Internet technology, which allows users and merchants to trade goods or services in a virtual environment. The display location of goods / merchants on the e-commerce platform can be controlled in a variety of ways, thereby optimizing the user experience and improving the sales performance of goods.

[0003] However, the inventors have discovered that the current method of performing traffic control in the form of advertisement insertion cannot take into account differences in traffic distribution of goods / merchants, and cannot ensure that the traffic distribution of goods / merchants remains consistent before and after traffic control, resulting in some goods / merchants with larger traffic gaps always occupying the head area, leading to poor traffic control effects. In addition, the current method of performing traffic control in the form of algorithm modeling only considers the situation where only one product / merchant is displayed at a time, while in actual applications, an indefinite number of products / merchants will be displayed each time a user turns a page, resulting in differences between real scenarios and algorithm modeling, leading to poor traffic control effects.

[0004] Based on this, the inventor believes that the current methods of traffic control in the form of advertisement insertion or algorithm modeling have the problem of poor traffic control effect and inability to accurately reach target users. Summary of the invention

[0005] The present invention provides a traffic control method and system, an electronic device and a storage medium based on an e-commerce platform, which are used to solve the problems that the current methods of traffic control in the form of advertisement insertion or algorithm modeling have poor traffic control effects and cannot accurately reach target users.

[0006] According to one aspect of the present invention, the present invention provides a traffic control method based on an e-commerce platform, including: determining natural merchants and flow control merchants based on a general search and push link in response to user requests; determining a merchant set based on natural merchants and flow control merchants; determining the ranking values ​​of all merchants in the merchant set; modeling and solving all merchants based on a first preset constraint condition and a ranking value to obtain a target merchant set that needs to be displayed in the merchant set; modeling and solving the target merchant set based on a second preset constraint condition and a ranking value to obtain the display position of the merchant in the target merchant set; wherein the first preset constraint condition includes at least an exposure distribution difference constraint, a total exposure constraint and maximizing scene efficiency; the second preset constraint condition includes at least an efficiency loss constraint and maximizing flow control traffic.

[0007] According to some embodiments of the present invention, modeling and solving of all merchants based on the first preset constraint condition and the ranking value to obtain a target merchant set that needs to be displayed in the merchant set includes: determining a modeling formula of a merchant selection algorithm based on exposure distribution difference constraints, total exposure constraints, maximizing scene effectiveness and ranking values; relaxing the modeling formula, introducing regularization terms and simplifying the exposure distribution difference constraints; solving the modeling formula online through the Lagrangian dual algorithm and the gradient boosting algorithm to obtain the target merchants that need to be displayed in the merchant set; and determining a target merchant set based on all the target merchants.

[0008] According to some embodiments of the present invention, in the case of online solution, a log estimation algorithm is used to approximate the Lagrangian dual function of the Lagrangian dual algorithm to perform online solution.

[0009] According to some embodiments of the present invention, modeling and solving the target merchant set based on the second preset constraint condition and the ranking value to obtain the display position of the merchants in the target merchant set includes: determining the modeling formula of the position rearrangement algorithm based on the efficiency loss constraint and the maximum flow control flow and the ranking value; and solving the modeling formula online through the Lagrangian dual algorithm and the gradient boosting algorithm to obtain the display position of the merchants in the target merchant set.

[0010] According to another aspect of the present invention, the present invention provides a traffic control system based on an e-commerce platform, including a merchant processing module, a first-stage modeling and solving module, and a second-stage modeling and solving module. The merchant processing module determines natural merchants and flow control merchants based on the search and push general link in response to user requests, determines a merchant set based on the natural merchants and flow control merchants, and determines the ranking values ​​of all merchants in the merchant set; the first-stage modeling and solving module models and solves all merchants based on the first preset constraint and ranking value to obtain the target merchant set that needs to be displayed in the merchant set; the second-stage modeling and solving module models and solves the target merchant set based on the second preset constraint and ranking value to obtain the display position of the merchant in the target merchant set; wherein the first preset constraint includes at least exposure distribution difference constraint, exposure total amount constraint and maximization of scene efficiency; the second preset constraint includes at least efficiency loss constraint and maximization of flow control traffic.

[0011] According to some embodiments of the present invention, the one-stage modeling and solving module determines the modeling formula of the merchant selection algorithm based on the exposure distribution difference constraint, the total exposure constraint, the maximized scene effectiveness and the ranking value; the one-stage modeling and solving module relaxes the modeling formula, introduces regularization terms and simplifies the exposure distribution difference constraint; the one-stage modeling and solving module solves the modeling formula online through the Lagrangian dual algorithm and the gradient boosting algorithm to obtain the target merchants that need to be displayed in the merchant set; the one-stage modeling and solving module determines the target merchant set based on all the target merchants.

[0012] According to some embodiments of the present invention, the one-stage modeling and solving module uses a log estimation algorithm to approximate the Lagrangian dual function of the Lagrangian dual algorithm to perform an online solution.

[0013] According to some embodiments of the present invention, a two-stage modeling and solving module determines a modeling formula of a position rearrangement algorithm based on an efficiency loss constraint and maximizing flow control flow and a sorting value; the two-stage modeling and solving module solves the modeling formula online through a Lagrangian dual algorithm and a gradient boosting algorithm to obtain the display positions of merchants in a target merchant set.

[0014] According to another aspect of the present invention, an electronic device is provided, which includes: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors can implement the flow control method described above.

[0015] According to another aspect of the present invention, a non-volatile computer-readable storage medium is provided, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the flow control method described above can be implemented.

[0016] The present invention responds to user requests, determines a merchant set based on a search and push general link, and determines the ranking values ​​of all merchants in the merchant set, models and solves all merchants based on a first preset constraint condition and the ranking value, and can obtain a target merchant set that needs to be displayed, and models and solves the target merchant set based on a second preset constraint condition and the ranking value, and can obtain the display positions of the merchants in the target merchant set.

[0017] The present invention can determine the merchants that need to be exposed and displayed in response to user needs each time in the first modeling and solving stage, and determine the specific display locations of the merchants that need to be exposed and displayed in the second modeling and solving stage. The present invention can reduce the complexity of the solution through two-stage modeling and solving (i.e., two-stage flow control method).

[0018] The present invention sets exposure distribution difference constraints, exposure total amount constraints and maximized scene performance in the first modeling and solving stage, which can make the difference between merchants supported by traffic before and after flow control smaller, making the merchant traffic distribution more reasonable and stable, thereby improving the user's browsing experience. And the present invention sets performance loss constraints and maximizes flow control flow in the second modeling and solving stage for online solving, which can improve scene performance and reduce online pressure. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 A flow chart showing a flow control method according to an embodiment of the present invention;

[0021] Figure 2 A schematic diagram showing a merchant recall according to an embodiment of the present invention;

[0022] Figure 3 Another flow chart showing the flow control method according to an embodiment of the present invention;

[0023] Figure 4 Another flow chart showing the flow control method according to an embodiment of the present invention;

[0024] Figure 5 A schematic diagram showing the offline experimental results of an embodiment of the present invention;

[0025] Figure 6 Another schematic diagram showing the offline experimental results of an embodiment of the present invention;

[0026] Figure 7 A schematic diagram showing the online experimental results of an embodiment of the present invention;

[0027] Figure 8 A schematic diagram showing a flow control system according to an embodiment of the present invention.

[0028] Description of reference numerals:

[0029] Traffic control system 1; merchant processing module 10; first-stage modeling and solving module 20; second-stage modeling and solving module 30. DETAILED DESCRIPTION

[0030] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0031] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that the present invention will be comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted.

[0032] The described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of these specific details, or other modes, components, materials, devices, etc. may be adopted. In these cases, known structures, methods, devices, implementations, materials or operations will not be shown or described in detail.

[0033] In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0034] The terms "first", "second" and the like in the specification, claims and the above drawings of the present invention are used to distinguish different objects rather than to describe a specific order.

[0035] The technical solution of the present invention is described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0036] According to one aspect of the present invention, the present invention provides a flow control method based on an e-commerce platform. Figure 1As shown, the flow control method may include steps S100-S500. Exemplarily, the flow control method may be executed by a flow control system having computing capability.

[0037] According to an example embodiment, in step S100, the traffic control system determines natural merchants and traffic control merchants based on the search and push universal link in response to a user request.

[0038] For example, a user request may be a search request from a user based on an e-commerce platform, and a search and push general link may be a general link used for data transmission and processing in a search engine or recommendation system. A natural merchant may be a merchant directly associated with a user's search request, and a traffic control merchant may be a traffic control merchant that needs traffic support.

[0039] like Figure 2 As shown, the flow control system responds to the user's search request and recalls the merchants corresponding to the search request through the search and push general link. Based on the user's search request, the present invention adds an additional recall of flow-controlled merchants on the basis of the original recall of natural merchants, so that the recall results including natural merchants and flow-controlled merchants can be obtained.

[0040] In step S200, the flow control system determines a merchant set based on natural merchants and flow control merchants.

[0041] In step S300, the traffic control system determines the ranking values ​​of all merchants in the merchant set.

[0042] For example, Figure 2 As shown, natural merchants and flow control merchants are combined to form a merchant set. The flow control system inputs the merchant set into the deep model sorting module based on the deep model sorting algorithm, sorts all merchants in the merchant set, and thus obtains the ranking values ​​(ranking scores) of all merchants.

[0043] In step S400, the traffic control system models and solves all merchants based on the first preset constraint condition and the ranking value to obtain a target merchant set that needs to be displayed in the merchant set.

[0044] The first preset constraint condition includes at least an exposure distribution difference constraint, an exposure total amount constraint and maximizing scene efficiency.

[0045] For example, the target merchant set may be a set of merchants that need to be exposed and displayed in a merchant set obtained by the traffic control system through modeling.

[0046] The exposure distribution difference constraint may be an exposure distribution difference constraint of each page of the merchant before and after traffic control, for example, it may be a KL divergence difference (Kullback-Leibler divergence, an asymmetric measure of the difference between two probability distributions) of the exposure distribution of each page of the merchant before and after traffic control.

[0047] It can be understood here that the current modeling method for the flow control algorithm is to constrain the flow of the flow control merchant to be greater than a given value, and to maximize the ranking points of all merchants. However, the inventors found that when this modeling method is adopted, some merchants with larger flow target setting values ​​will occupy the user's homepage exposure for a long time, resulting in a poor user experience. Therefore, in the modeling process, the present invention adds the exposure distribution difference of each page of the merchant before and after flow control (such as KL divergence difference constraint), which can make the exposure distribution of the merchant before and after flow control remain unchanged as much as possible, thereby making the merchant flow distribution more stable.

[0048] The total exposure constraint can be used to constrain the exposure of all merchants during the modeling and solving process. The total exposure constraint is not a fixed value (such as a fixed value of 1), but the number of merchants that need to be returned on each page.

[0049] The inventors also found that in real scenarios, when users make requests, several merchants are generally returned per page. However, the current modeling methods for flow control algorithms all assume that only one merchant is returned each time, which deviates from the real scenario. Therefore, in the modeling process, the present invention constrains the exposure of all merchants and the number of merchants that need to be returned per page. This setting can align offline modeling with users' online requests.

[0050] Maximizing scene effectiveness can be the click-through rate of the scene.

[0051] Alternatively, if Figure 3 As shown, in step S400, steps S410-S440 may also be included.

[0052] In step S410, the traffic control system determines a modeling formula for a merchant selection algorithm based on exposure distribution difference constraints, exposure total amount constraints, maximized scene effectiveness, and ranking values.

[0053] For example, the modeling formula for the merchant selection algorithm can be:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Among them, x ij is the exposure probability of merchant j in the i-th request; r ij is the ranking value of merchant j in the i-th request; T j is the traffic value required by merchant j; is the KL divergence difference of the exposure distribution of each page of merchant j before and after flow control; δ is the manually set flow control coefficient, which is used to control the distribution difference before and after flow control; is the exposure ratio of merchant j on page k; I k For all user requests for page k; n i is the number of merchants that need to be returned in the i-th request.

[0061] In step S420, the traffic control system relaxes the modeling formula, introduces regularization terms and simplifies the exposure distribution difference constraints.

[0062] For example, to speed up the solution, we can ij Relax and introduce the regular term x into the modeling formula ij lnx ij , and at the same time simplify the KL divergence difference and convert it into standard form:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] Among them, U is the objective function; λ is the regularization term coefficient; p ik Indicates whether the i-th request requests the k-th page; D is the intermediate item.

[0070] In step S430, the traffic control system solves the modeling formula online through the Lagrange dual algorithm and the gradient boosting algorithm to obtain the target merchants that need to be displayed in the merchant set.

[0071] In step S440, the traffic control system determines a target merchant set based on all target merchants.

[0072] The Lagrange dual algorithm is a method for solving constrained optimization problems. By introducing Lagrange multipliers, the original optimization problem is converted into a dual problem, thereby simplifying the solution process.

[0073] For example, the flow control system is solved using the Lagrangian dual algorithm, where the Lagrangian dual function is:

[0074]

[0075] Among them, α and γ are Lagrange multipliers.

[0076] Using the L-BFGS (Limited-memory BFGS) algorithm, we can get The x corresponding to 0 ij , and then get the derivatives corresponding to α and γ:

[0077]

[0078]

[0079] in, is the derivative corresponding to α; is the derivative corresponding to γ; t is the number of iteration steps.

[0080] Then, α and γ can be solved by gradient boosting.

[0081] Optionally, in the case of online solution, the flow control system uses a log estimation algorithm to approximate the Lagrangian dual function of the Lagrangian dual algorithm to perform online solution.

[0082] It can be understood here that α and γ can be obtained offline based on historical log data, but cannot be obtained online by Set to 0 to solve for x ij .

[0083] Therefore, the present invention adopts log estimation Approximating L(x, α, γ), we get:

[0084]

[0085] Among them, β i is the Lagrange multiplier; E jk is the total number of exposures of merchant j on page k; is the historical exposure times of merchant j on page k.

[0086] Will Set to 0, you can get x ij The parsed form is:

[0087]

[0088] Among them, Z ij is the intermediate term,

[0089] Because β i It does not affect the user's request x ij The sorting needs to satisfy the constraints ∑ j x ij =n i , we can get x ij for:

[0090]

[0091]

[0092] Among them, J is all candidate merchants, J' is a subset of J; A ij is the middle term.

[0093] The meaning of this formula is that in A ij Take the subscripts corresponding to the n largest numbers and make the x corresponding to these subscripts ij is 1, based on the final solution x ij , the merchant j that needs to be exposed in the i-th request can be determined. Thus, the traffic control system can determine the set of all target merchants that need to be exposed.

[0094] In step S500, the traffic control system models and solves the target merchant set based on the second preset constraint condition and the ranking value to obtain the display positions of the merchants in the target merchant set.

[0095] The second preset constraint condition includes at least an efficiency loss constraint and a maximum flow control flow rate.

[0096] For example, in step S400, the traffic control system determines the merchants that need to be exposed, but the specific display location of each merchant in each page display scene of the e-commerce platform has not yet been determined. The traffic control system obtains the specific display location of each merchant that needs to be exposed through modeling and solving.

[0097] Exemplarily, the performance loss constraint may be that the click-through rate of the scenario cannot be lower than 95% compared to a case without flow control; maximizing the flow control traffic may be to maximize the traffic obtained by the flow control merchant.

[0098] Alternatively, if Figure 4 As shown, in step S500, steps S510-S520 may also be included.

[0099] In step S510, the flow control system determines a modeling formula of a position rearrangement algorithm based on an efficiency loss constraint and a maximized flow control flow and a ranking value.

[0100] For example, the modeling formula of the position rearrangement algorithm is:

[0101]

[0102] Among them, y ij Whether merchant j is displayed at location i, y ij ∈{0,1};v ij is the exposure value of merchant j at position i; w i is the attenuation coefficient at position i; s j is the ranking score of merchant j; δ1 is the weight that can tolerate the loss of scene performance; G is the ideal value of DCG (Discounted Cumulative Gain).

[0103] y ij The value of is relaxed to between 0 and 1, and the constraints of the above modeling formula are relaxed using Lagrange, and we can get:

[0104]

[0105] In step S520, the traffic control system solves the modeling formula online through the Lagrange dual algorithm and the gradient boosting algorithm to obtain the display positions of the merchants in the target merchant set.

[0106] For example, the modeling formula can be solved online based on the Lagrange dual algorithm and the gradient boosting algorithm to obtain the required y ij , so as to determine the specific display location of the merchant.

[0107] Optionally, the flow control system can also regard the inner loop as an optimal transmission problem and use the Sinkhorn algorithm (an algorithm for solving the optimal transmission problem) to calculate y in each iteration. ij Accelerated solving can speed up the online solving speed.

[0108] The present invention responds to user requests, determines a merchant set based on a search and push general link, and determines the ranking values ​​of all merchants in the merchant set, models and solves all merchants based on a first preset constraint condition and the ranking value, and can obtain a target merchant set that needs to be displayed, and models and solves the target merchant set based on a second preset constraint condition and the ranking value, and can obtain the display positions of the merchants in the target merchant set.

[0109] The present invention can determine the merchants that need to be exposed and displayed in response to user needs each time in the first modeling and solving stage, and determine the specific display locations of the merchants that need to be exposed and displayed in the second modeling and solving stage. The present invention can reduce the complexity of the solution through two-stage modeling and solving (i.e., two-stage flow control method).

[0110] The present invention sets exposure distribution difference constraints, exposure total amount constraints and maximized scene performance in the first modeling and solving stage, which can make the difference between merchants supported by traffic before and after flow control smaller, making the merchant traffic distribution more reasonable and stable, thereby improving the user's browsing experience. And the present invention sets performance loss constraints and maximizes flow control flow in the second modeling and solving stage for online solving, which can improve scene performance and reduce online pressure.

[0111] Figure 5 The experimental data comparison of the total ranking scores and uncompleted flow gaps based on several different flow control methods such as no flow control, fixed weighting method (existing), one-stage flow control method and two-stage flow control method (the present invention) is shown. Figure 5 As shown, the two-stage flow control method provided by the present invention has only a small loss in the total ranking score compared to the method without flow control, but the unfinished flow gap is significantly reduced.

[0112] Figure 6 The experimental data comparison of the distribution difference of the two-stage flow control method without exposure distribution difference constraint and with exposure distribution difference constraint in the present invention is shown. Figure 6 As shown, by setting the exposure distribution difference constraint, the total KL divergence of the distribution difference and the number of merchants with large KL distribution difference (such as distribution difference KL greater than 0.5) can be significantly reduced, that is, before and after flow control, the present invention can effectively keep the distribution of merchants consistent before and after flow control.

[0113] Figure 7 The figure shows the comparison of supported merchant traffic, scenario UV_CTR (i.e., user-based click-through rate), and supported merchant enquiry volume based on several different flow control methods, such as no flow control, fixed weighting method (existing) and two-stage flow control method (the present invention). Figure 7 As shown, the two-stage flow control method provided by the present invention can significantly increase the traffic and inquiry volume of supported merchants, and can also significantly improve the scene UV_CTR.

[0114] According to another aspect of the present invention, the present invention provides a traffic control system based on an e-commerce platform. Figure 8 As shown, the traffic control system 1 includes a merchant processing module 10 , a first-stage modeling and solving module 20 , and a second-stage modeling and solving module 30 .

[0115] According to an exemplary embodiment, the merchant processing module 10 determines the natural merchants and the flow-controlled merchants based on the search and push universal link in response to the user request.

[0116] For example, a user request may be a search request from a user based on an e-commerce platform, and a search and push general link may be a general link used for data transmission and processing in a search engine or recommendation system. A natural merchant may be a merchant directly associated with a user's search request, and a traffic control merchant may be a traffic control merchant that needs traffic support.

[0117] The merchant processing module 10 responds to the user's search request and recalls the merchants corresponding to the search request through the search and push general link. Based on the user's search request, the present invention adds an additional recall of flow-controlled merchants on the basis of the original recall of natural merchants, so that the recall results including natural merchants and flow-controlled merchants can be obtained.

[0118] According to an exemplary embodiment, the merchant processing module 10 determines a merchant set according to natural merchants and flow-controlled merchants. The merchant processing module 10 determines the ranking values ​​of all merchants in the merchant set.

[0119] For example, natural merchants and flow control merchants are combined to form a merchant set. The merchant processing module 10 inputs the merchant set into a deep model sorting module based on a deep model sorting algorithm, sorts all merchants in the merchant set, and thus obtains the ranking values ​​(ranking scores) of all merchants.

[0120] According to an exemplary embodiment, the first-stage modeling and solving module 20 performs modeling and solving on all merchants based on the first preset constraint condition and the ranking value, so as to solve and obtain a target merchant set that needs to be displayed in the merchant set.

[0121] The first preset constraint condition includes at least an exposure distribution difference constraint, an exposure total amount constraint and maximizing scene efficiency.

[0122] For example, the target merchant set may be a set of merchants that need to be exposed and displayed in the merchant set obtained by the first-stage modeling and solving module 20 through modeling and solving.

[0123] The exposure distribution difference constraint may be an exposure distribution difference constraint of each page of the merchant before and after traffic control, for example, it may be a KL divergence difference (Kullback-Leibler divergence, an asymmetric measure of the difference between two probability distributions) of the exposure distribution of each page of the merchant before and after traffic control.

[0124] It can be understood here that the current modeling method for the flow control algorithm is to constrain the flow of the flow control merchant to be greater than a given value, and to maximize the ranking points of all merchants. However, the inventors found that when this modeling method is adopted, some merchants with larger flow target setting values ​​will occupy the user's homepage exposure for a long time, resulting in a poor user experience. Therefore, in the modeling process, the present invention adds the exposure distribution difference of each page of the merchant before and after flow control (such as KL divergence difference constraint), which can make the exposure distribution of the merchant before and after flow control remain unchanged as much as possible, thereby making the merchant flow distribution more stable.

[0125] The total exposure constraint can be used to constrain the exposure of all merchants during the modeling and solving process. The total exposure constraint is not a fixed value (such as a fixed value of 1), but the number of merchants that need to be returned on each page.

[0126] The inventors also found that in real scenarios, when users make requests, several merchants are generally returned per page. However, the current modeling methods for flow control algorithms all assume that only one merchant is returned each time, which deviates from the real scenario. Therefore, in the modeling process, the present invention constrains the exposure of all merchants and the number of merchants that need to be returned per page. This setting can align offline modeling with users' online requests.

[0127] Maximizing scene effectiveness can be the click-through rate of the scene.

[0128] Optionally, the first-stage modeling and solving module 20 determines a modeling formula of a merchant selection algorithm based on exposure distribution difference constraints, exposure total amount constraints, maximization of scene effectiveness and ranking values.

[0129] For example, the modeling formula for the merchant selection algorithm can be:

[0130]

[0131] Among them, x ij is the exposure probability of merchant j in the i-th request; r ij is the ranking value of merchant j in the i-th request; T j is the traffic value required by merchant j; is the KL divergence difference of the exposure distribution of each page of merchant j before and after flow control; δ is the manually set flow control coefficient, which is used to control the distribution difference before and after flow control; is the exposure ratio of merchant j on page k; I k For all user requests for page k; n i is the number of merchants that need to be returned in the i-th request.

[0132] The first-stage modeling solution module 20 relaxes the modeling formula, introduces regularization terms and simplifies the exposure distribution difference constraints.

[0133] For example, to speed up the solution, we can ij Relax and introduce the regular term x into the modeling formula ij lnx ij , and at the same time simplify the KL divergence difference and convert it into standard form:

[0134]

[0135] Among them, U is the objective function; λ is the regularization term coefficient; p ik Indicates whether the i-th request requests the k-th page; D is the intermediate item.

[0136] The first-stage modeling solution module 20 solves the modeling formula online through the Lagrange dual algorithm and the gradient boosting algorithm to obtain the target merchants that need to be displayed in the merchant set.

[0137] The first-stage modeling and solving module 20 determines a target merchant set based on all target merchants.

[0138] The Lagrange dual algorithm is a method for solving constrained optimization problems. By introducing Lagrange multipliers, the original optimization problem is converted into a dual problem, thereby simplifying the solution process.

[0139] For example, the first-stage modeling solution module 20 uses the Lagrangian dual algorithm for solution, where the Lagrangian dual function is:

[0140]

[0141] Among them, α and γ are Lagrange multipliers.

[0142] Using the L-BFGS (Limited-memory BFGS) algorithm, we can get The x corresponding to 0 ij , and then get the derivatives corresponding to α and γ:

[0143]

[0144]

[0145] in, is the derivative corresponding to α; is the derivative corresponding to γ; t is the number of iteration steps.

[0146] Then, α and γ can be solved by gradient boosting.

[0147] Optionally, in the case of online solution, the one-stage modeling solution module 20 uses a log estimation algorithm to approximate the Lagrangian dual function of the Lagrangian dual algorithm to perform online solution.

[0148] It can be understood here that α and γ can be obtained offline based on historical log data, but cannot be obtained online by Set to 0 to solve for x ij .

[0149] Therefore, the present invention adopts log estimation Approximating L(x, α, γ), we get:

[0150]

[0151] Among them, β i is the Lagrange multiplier; E jk is the total number of exposures of merchant j on page k; is the historical exposure times of merchant j on page k.

[0152] Will Set to 0, you can get x ij The parsed form is:

[0153]

[0154] Among them, Z ij is the intermediate term,

[0155] Because β i It does not affect the user's request x ij The sorting needs to satisfy the constraints ∑ j x ij =n i , we can get x ij for:

[0156]

[0157]

[0158] Among them, J is all candidate merchants, J' is a subset of J; A ij is the middle term.

[0159] The meaning of this formula is that in A ij Take the subscripts corresponding to the n largest numbers and make the x corresponding to these subscripts ij is 1, based on the final solution x ij, the merchant j that needs to be exposed in the i-th request can be determined. Thus, the first-stage modeling and solving module 20 can determine the set of all target merchants that need to be exposed.

[0160] According to an exemplary embodiment, the two-stage modeling and solving module 30 performs modeling and solving on the target merchant set based on the second preset constraint condition and the ranking value to solve and obtain the display positions of the merchants in the target merchant set.

[0161] The second preset constraint condition includes at least an efficiency loss constraint and a maximum flow control flow rate.

[0162] For example, the first-stage modeling and solving module 20 determines the merchants that need to be exposed, but the specific display position of each merchant in each page display scene of the e-commerce platform has not yet been determined. The second-stage modeling and solving module 30 obtains the specific display position of each merchant that needs to be exposed through modeling and solving.

[0163] Exemplarily, the performance loss constraint may be that the click-through rate of the scenario cannot be lower than 95% compared to a case without flow control; maximizing the flow control traffic may be to maximize the traffic obtained by the flow control merchant.

[0164] Optionally, the two-stage modeling and solving module 30 determines a modeling formula of a position rearrangement algorithm based on an efficiency loss constraint and a maximized flow control flow and a ranking value.

[0165] For example, the modeling formula of the position rearrangement algorithm is:

[0166]

[0167]

[0168]

[0169]

[0170]

[0171]

[0172] Among them, y ij Whether merchant j is displayed at location i, y ij ∈{0,1};v ij is the exposure value of merchant j at position i; w i is the attenuation coefficient at position i; s j is the ranking score of merchant j; δ1 is the weight that can tolerate the loss of scene performance; G is the ideal value of DCG (Discounted Cumulative Gain).

[0173] y ij The value of is relaxed to between 0 and 1, and the constraints of the above modeling formula are relaxed using Lagrange, and we can get:

[0174]

[0175]

[0176]

[0177] The two-stage modeling and solving module 30 solves the modeling formula online through the Lagrange dual algorithm and the gradient boosting algorithm to obtain the display positions of the merchants in the target merchant set.

[0178] For example, the modeling formula can be solved online based on the Lagrange dual algorithm and the gradient boosting algorithm to obtain the required y ij , so as to determine the specific display location of the merchant.

[0179] Optionally, the two-stage modeling and solving module 30 may also regard the inner loop as an optimal transmission problem and use the Sinkhorn algorithm (an algorithm for solving the optimal transmission problem) to solve y in each iteration. ij Accelerated solving can speed up the online solving speed.

[0180] The present invention responds to user requests, determines a merchant set based on a search and push general link, and determines the ranking values ​​of all merchants in the merchant set, models and solves all merchants based on a first preset constraint condition and the ranking value, and can obtain a target merchant set that needs to be displayed, and models and solves the target merchant set based on a second preset constraint condition and the ranking value, and can obtain the display positions of the merchants in the target merchant set.

[0181] The present invention can determine the merchants that need to be exposed and displayed in response to user needs each time in the first modeling and solving stage, and determine the specific display locations of the merchants that need to be exposed and displayed in the second modeling and solving stage. The present invention can reduce the complexity of the solution through two-stage modeling and solving (i.e., two-stage flow control method).

[0182] The present invention sets exposure distribution difference constraints, exposure total amount constraints and maximized scene performance in the first modeling and solving stage, which can make the difference between merchants supported by traffic before and after flow control smaller, making the merchant traffic distribution more reasonable and stable, thereby improving the user's browsing experience. And the present invention sets performance loss constraints and maximizes flow control flow in the second modeling and solving stage for online solving, which can improve scene performance and reduce online pressure.

[0183] According to another aspect of the present invention, an electronic device is provided, which includes: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors can implement the flow control method described above.

[0184] According to another aspect of the present invention, a non-volatile computer-readable storage medium is provided, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the flow control method described above can be implemented.

[0185] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions of the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A traffic control method based on an e-commerce platform, characterized in that: include: In response to user requests, determine natural merchants and flow-controlled merchants based on the general search and push links; Determine a merchant set according to the natural merchants and the flow control merchants; Determine the ranking values ​​of all merchants in the merchant set; Modeling and solving all the merchants based on the first preset constraint condition and the ranking value to obtain a target merchant set that needs to be displayed in the merchant set; Modeling and solving the target merchant set based on the second preset constraint condition and the ranking value to obtain the display positions of merchants in the target merchant set; Among them, the first preset constraint condition includes at least exposure distribution difference constraint, total exposure constraint and maximizing scene efficiency; the second preset constraint condition includes at least efficiency loss constraint and maximizing flow control flow.

2. The flow control method according to claim 1, characterized in that: The modeling and solving of all the merchants based on the first preset constraint condition and the ranking value to obtain a target merchant set to be displayed in the merchant set includes: Determine a modeling formula for a merchant selection algorithm based on the exposure distribution difference constraint, the exposure total amount constraint, the maximized scene effectiveness, and the ranking value; Relaxing the modeling formula, introducing regularization terms and simplifying the exposure distribution difference constraint; The modeling formula is solved online by using a Lagrange dual algorithm and a gradient boosting algorithm to obtain a target merchant to be displayed in the merchant set; The target merchant set is determined according to all the target merchants.

3. The flow control method according to claim 2, characterized in that: In the case of online solution, a log estimation algorithm is used to approximate the Lagrangian dual function of the Lagrangian dual algorithm to perform the online solution.

4. The flow control method according to claim 1, characterized in that: The modeling and solving the target merchant set based on the second preset constraint condition and the ranking value to obtain the display positions of merchants in the target merchant set includes: Determine a modeling formula of a position rearrangement algorithm based on the efficiency loss constraint, the maximized flow control flow rate and the ranking value; The modeling formula is solved online by Lagrange dual algorithm and gradient boosting algorithm to obtain the display positions of merchants in the target merchant set.

5. A traffic control system based on an e-commerce platform, characterized in that: include: A merchant processing module, in response to a user request, determines a natural merchant and a flow-controlled merchant based on a search and push general link, determines a merchant set according to the natural merchant and the flow-controlled merchant, and determines a ranking value of all merchants in the merchant set; A first-stage modeling and solving module performs modeling and solving on all the merchants based on the first preset constraint condition and the ranking value, so as to solve and obtain a target merchant set that needs to be displayed in the merchant set; A second-stage modeling and solving module is used to model and solve the target merchant set based on the second preset constraint condition and the ranking value, so as to solve and obtain the display positions of merchants in the target merchant set; Among them, the first preset constraint condition includes at least exposure distribution difference constraint, total exposure constraint and maximizing scene efficiency; the second preset constraint condition includes at least efficiency loss constraint and maximizing flow control flow.

6. The flow control system according to claim 5, characterized in that: The first-stage modeling and solving module determines a modeling formula of a merchant selection algorithm based on the exposure distribution difference constraint, the exposure total amount constraint, the maximized scene effectiveness and the ranking value; The first-stage modeling and solving module relaxes the modeling formula, introduces regularization terms, and simplifies the exposure distribution difference constraint; The first-stage modeling solution module solves the modeling formula online through the Lagrange dual algorithm and the gradient boosting algorithm to obtain the target merchant to be displayed in the merchant set; The first-stage modeling and solving module determines the target merchant set based on all target merchants.

7. The flow control system according to claim 6, characterized in that: In the case of online solution, the one-stage modeling and solving module uses a log estimation algorithm to approximate the Lagrangian dual function of the Lagrangian dual algorithm to perform the online solution.

8. The flow control system according to claim 5, characterized in that: The two-stage modeling and solving module determines a modeling formula of a position rearrangement algorithm based on the efficiency loss constraint, the maximized flow control flow and the ranking value; The two-stage modeling and solving module solves the modeling formula online through the Lagrange dual algorithm and the gradient boosting algorithm to obtain the display positions of the merchants in the target merchant set.

9. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the flow control method as described in any one of claims 1-4.

10. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the flow control method according to any one of claims 1 to 4 is implemented.

Images