A method, device, computer equipment and storage medium for intelligent product recommendation

Abstract

The present invention provides a method, device, computer equipment and storage medium for intelligent recommendation of commodities, and relates to the field of computer software technology. The method comprises: for each commodity, obtaining the historical exchange activity record corresponding to the commodity, and extracting the net income and the crowd interception ratio of the commodity from the historical exchange activity record; grouping the commodities based on the activities to which they belong, and dividing each group into multiple sub-buckets based on the crowd interception ratio, and then calculating the cumulative net income in each sub-bucket; constructing a net income distribution map through all the cumulative net income, and dynamically adjusting the crowd interception ratio based on the net income distribution map, and using the dynamic adjustment result as the candidate crowd interception ratio; obtaining the integral segment of the commodity, calculating the average value of the candidate crowd interception ratio in the integral segment, and using the default crowd interception ratio as the default crowd interception ratio of the integral segment; and intercepting the marketing crowd of the commodity using the default crowd interception ratio. The present invention can improve the marketing crowd interception accuracy and recommendation effect.

Description

Technical Field

[0001] The present invention relates to the field of computer software technology, and in particular to a commodity intelligent recommendation method, device, computer equipment and storage medium. Background Art

[0002] The existing SMS product intelligent recommendation system is a recommendation system based on deep learning models. For the selected products, the system will score the user's willingness to buy. These users' purchase intention scores will be used as screening indicators, aiming to cover as many potential purchasing users as possible with as few marketing SMS messages as possible. Reduce user fatigue and marketing costs, and maximize profits.

[0003] The logic of selecting users in existing projects is: for a selected product, users are sorted based on their purchase intention scores, and then the top N% of the population are selected to send marketing text messages. The specific interception ratio for each product is obtained based on the percentile of the highest point in the historical reverse purchase intention score-net benefit curve. The problem with this population interception method is that it cannot dynamically adjust the interception ratio based on the latest purchase records, and cannot guarantee real-time performance. Summary of the invention

[0004] The embodiments of the present invention provide a method, apparatus, computer device and storage medium for intelligent product recommendation, aiming to improve the accuracy of marketing crowd capture and recommendation effect.

[0005] In a first aspect, an embodiment of the present invention provides a method for intelligently recommending products, which includes:

[0006] For each commodity, obtain the historical redemption activity record corresponding to the commodity, and extract the net income and population interception ratio of the commodity from the historical redemption activity record;

[0007] The products are grouped based on the activities they belong to, and each group is divided into multiple sub-buckets based on the population interception ratio, and then the cumulative net income in each sub-bucket is calculated;

[0008] Constructing a net income distribution graph through all accumulated net incomes, dynamically adjusting the population interception ratio based on the net income distribution graph, and using the dynamic adjustment result as the candidate population interception ratio;

[0009] Obtain the score segment corresponding to the product, calculate the average value of the candidate population interception ratio within the score segment, and use the calculation result as the default population interception ratio of the corresponding score segment;

[0010] The marketing population of the product is intercepted using the default population interception ratio, and the population interception ratio is updated in real time based on the intercepted redemption activity records.

[0011] In a second aspect, an embodiment of the present invention provides a commodity intelligent recommendation device, comprising:

[0012] A record acquisition unit, for acquiring, for each commodity, a historical redemption activity record corresponding to the commodity, and extracting the net income and the population interception ratio of the commodity from the historical redemption activity record;

[0013] A partitioning unit is used to group products based on the activities they belong to, and to divide each group into multiple sub-buckets based on the population interception ratio, and then calculate the cumulative net income in each sub-bucket;

[0014] A dynamic adjustment unit, configured to construct a net income distribution diagram through all accumulated net income, dynamically adjust the population interception ratio based on the net income distribution diagram, and use the dynamic adjustment result as the candidate population interception ratio;

[0015] an average value calculation unit, used to obtain the score segment corresponding to the product, calculate the average value of the candidate population interception ratio within the score segment, and use the calculation result as the default population interception ratio of the corresponding score segment;

[0016] The real-time updating unit is used to intercept the marketing population of the product by using the default population interception ratio, and to update the population interception ratio in real time based on the intercepted redemption activity record.

[0017] In a third aspect, an embodiment of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for intelligent product recommendation as described in the first aspect is implemented.

[0018] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for intelligent product recommendation as described in the first aspect is implemented.

[0019] The embodiment of the present invention provides a method, apparatus, computer equipment and storage medium for intelligent recommendation of commodities, the method comprising: for each commodity, obtaining the historical redemption activity record corresponding to the commodity, and extracting the net income and population interception ratio of the commodity from the historical redemption activity record; grouping the commodities based on the activities to which they belong, and dividing each group into multiple sub-buckets based on the population interception ratio, and then calculating the cumulative net income in each sub-bucket; constructing a net income distribution graph through all the cumulative net income, and dynamically adjusting the population interception ratio based on the net income distribution graph, and using the dynamic adjustment result as the candidate population interception ratio; obtaining the integral segment corresponding to the commodity, calculating the average value of the candidate population interception ratio within the integral segment, and using the calculated result as the default population interception ratio of the corresponding integral segment; intercepting the marketing population of the commodity using the default population interception ratio, and updating the population interception ratio in real time based on the intercepted redemption activity record. The embodiment of the present invention realizes dynamic adjustment of the population interception ratio by grouping the current population interception ratio and net profit of the product, constructing a net profit distribution map, etc., so that the dynamically adjusted population interception ratio is more in line with the marketing recommendation of the product, thereby improving the marketing population interception accuracy and recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0021] Figure 1 A flowchart of a method for intelligently recommending products provided by an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of a sub-process of a commodity intelligent recommendation method provided by an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of another sub-process of a commodity intelligent recommendation method provided by an embodiment of the present invention;

[0024] Figure 4 A schematic block diagram of a commodity intelligent recommendation device provided by an embodiment of the present invention;

[0025] Figure 5 A schematic block diagram of a device for intelligently recommending products according to an embodiment of the present invention;

[0026] Figure 6 Another schematic block diagram of a commodity intelligent recommendation device provided by an embodiment of the present invention;

[0027] Figure 7 A schematic diagram of a net profit distribution diagram in a commodity intelligent recommendation method provided by an embodiment of the present invention;

[0028] Figure 8 Another schematic diagram of a net profit distribution diagram in a commodity intelligent recommendation method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0031] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0032] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0033] See below Figure 1 , Figure 1 A flowchart of a method for intelligently recommending products provided by an embodiment of the present invention specifically includes steps S101 to S105.

[0034] S101. For each commodity, obtain the historical redemption activity record corresponding to the commodity, and extract the net income and the population interception ratio of the commodity from the historical redemption activity record;

[0035] S102, grouping the products based on the activities they belong to, and dividing each group into multiple sub-buckets based on the population interception ratio, and then calculating the cumulative net income in each sub-bucket;

[0036] S103, constructing a net income distribution graph through all accumulated net income, dynamically adjusting the population interception ratio based on the net income distribution graph, and using the dynamic adjustment result as the candidate population interception ratio;

[0037] S104, obtaining the score segment corresponding to the product, calculating the average value of the candidate population interception ratio within the score segment, and using the calculation result as the default population interception ratio of the corresponding score segment;

[0038] S105: intercepting the marketing population of the product by using the default population interception ratio, and updating the population interception ratio in real time based on the intercepted redemption activity records.

[0039] In this embodiment, for each commodity, the corresponding population interception ratio and net income are captured from its historical exchange activity records, wherein the population interception ratio refers to sorting based on the user's purchase intention score, and then selecting the top N% of the population to send marketing text messages; the net income is the actual net income in a certain exchange activity, and the net income = commodity profit - channel cost. For example, the population interception ratio corresponding to each commodity is captured from the SMS exchange activity records within a month. Then the commodities are grouped and bucketed, and the net income distribution chart is constructed based on the net income of the commodity and the current population interception ratio. Then, the population interception ratio of the commodity is dynamically adjusted according to the cumulative net income trend in the net income distribution chart, and based on the dynamically adjusted population interception ratio (i.e., the candidate population interception ratio), the default population interception ratio of the commodity under different score segments is calculated, and finally the marketing population of the commodity is intercepted and marketed (such as SMS marketing or email marketing, etc.) using the default population interception ratio. At the same time, in order to ensure the accuracy and real-time performance of the default population interception ratio, a dynamic adjustment can be made every preset time interval (such as 12 hours or 24 hours, etc.).

[0040] This embodiment realizes dynamic adjustment of the population interception ratio by grouping the current population interception ratio and net profit of the product, constructing a net profit distribution map, etc., so that the dynamically adjusted population interception ratio is more in line with the marketing recommendation of the product, thereby improving the marketing population interception accuracy and recommendation effect.

[0041] This embodiment realizes the dynamic adjustment of the interception ratio of the marketing population corresponding to each product based on the latest marketing data and model prediction results. In the prior art, a fixed ratio is used to intercept the marketing population each time, which cannot fully utilize the accumulated marketing data, and may cause inaccurate interception of the population due to inconsistent distribution of new data and historical data. This embodiment can adjust the ratio of the marketing population in a timely manner based on the latest data, which can minimize excessive marketing to people with no intention, and can also reduce ineffective marketing costs, improve the overall efficiency of SMS marketing activities, and realize a data-driven intelligent product recommendation strategy.

[0042] In one embodiment, the intelligent commodity recommendation method further includes:

[0043] For each of the products, a matching score is given to each of the products and each user in the population cut-off ratio by using a scoring model;

[0044] A benchmark table is constructed by combining the activity ID, product ID, user ID, net revenue, population interception ratio and matching score to which the product belongs.

[0045] In this embodiment, for each product, the scoring model is used to score the matching degree of the product and each user in the population interception ratio, so that a benchmark table can be constructed by combining the activity ID, product ID, user ID, net income, interception ratio and model score to which the product belongs. The benchmark table can be shown in Table 1:

[0046] Event ID Product ID User ID Net income Cut ratio Model Scoring 1 1 1 2 30 0.0672 1 1 2 -0.03 30 0.0004 2 5 3 5 50 0.0832 3 3 4 20 5 0.0735 3 3 5 10 5 0.0634 3 3 6 5 5 0.0613 3 3 7 8 5 0.0421 3 3 8 -10 5 0.0321

[0047] Table 1

[0048] In one embodiment, for each of the commodities, a matching score is scored between each of the commodities and each user in the population cut-off ratio by using a scoring model, including:

[0049] Extracting a product feature vector and a user feature vector for the product and the user respectively;

[0050] The product feature vector and the user feature vector are input into a Wide&Deep model, and the Wide&Deep model is used to output the matching degree between the product and the user.

[0051] In this embodiment, the Wide&Deep model is composed of a shallow (or single-layer) Wide neural network and a deep Deep multi-layer neural network. The output layer uses softmax or logistics regression to synthesize the outputs of the Wide and Deep parts. Among them, the Wide part is conducive to enhancing the "memory ability" of the model, and the Deep part is conducive to enhancing the "generalization ability" of the model. The Wide part is a basic linear model, expressed as y=WX+b. The X feature part includes basic features and cross features. The cross features are very important in the Wide part. They can capture the interaction between features and play a role in adding nonlinearity; the Deep part is a feedforward network model. The product feature vector and the user feature vector are first converted into low-dimensional dense vectors, and the dimensions are usually O(10)-O(100). Then the vectors are randomly initialized and the model is trained after minimizing the function at any time. The activation function uses Relu. In the joint model, the outputs of the Wide and Deep parts are merged together in a weighted manner, and the final output is obtained through the logistic loss function.

[0052] Furthermore, in one embodiment, the step of inputting the product feature vector and the user feature vector into a Wide&Deep model and using the Wide&Deep model to output the degree of matching between the product and the user includes:

[0053] According to the following formula, the Wide network in the Wide&Deep model is used to perform cross-feature construction on the product feature vector and the user feature vector:

[0054]

[0055] y=w T [x,φ(x)]+b

[0056] Where y represents the predicted value output by the Wide network, φ k (x) represents the cross product transfer function, c ki is a Boolean variable, x=[x1, x2, ..., xd] is the product feature vector and user feature vector, w=[w1w2, ..., wd] is the model parameter, and b is the deviation coefficient;

[0057] According to the following formula, the Deep network in the Wide & Deep model is used to concatenate the embedding vector and the dense vector:

[0058] a (l+1) =f(W (l) a (l) +b (l) )

[0059] In the formula, 1 represents the number of layers of the Deep network, f represents the activation function, and a (l) , b (l) and W (l) are the activation function, bias coefficient and model weight of the lth layer respectively;

[0060] Based on the output results of the Wide network and the Deep network, a joint prediction is performed according to the following formula to obtain the final matching degree P between the product and the user:

[0061]

[0062] Where Y is the binary classification label, σ(.) is the sigmoid function, φ(x) is the cross-product transformation of the original feature x, b is the bias term, and W Wide is the weight vector of the Wide network, W Deep For the final activation function The weight of .

[0063] In this embodiment, the Wide & Deep model connects the Wide part of the single input layer with the Deep part consisting of the Embedding layer and multiple hidden layers, and inputs them together into the final output layer. The single-layer Wide part is good at processing a large number of sparse ID-type features; the Deep part uses the strong expressiveness of the neural network to perform deep feature cross-talk and mine the data patterns hidden behind the features. Finally, using the logistic regression model, the output layer combines the Wide part and the Deep part to form a unified model.

[0064] In one embodiment, the S102 includes:

[0065] Obtaining the product ID and the activity ID to which it belongs in the reference table;

[0066] The historical redemption activity records with the same product ID and the corresponding activity ID are divided into the same group, and the historical redemption activity records with different product IDs and / or corresponding activity IDs are divided into different groups;

[0067] In each of the groups, the group is divided into N sub-buckets according to the population interception ratio N% corresponding to the commodity;

[0068] All net revenues in each of the sub-buckets are summed up to obtain the cumulative net revenue of each of the sub-buckets.

[0069] In this embodiment, the product ID and its corresponding activity ID are obtained, and then the products in the historical redemption activity records are grouped according to (activity ID, product ID), that is, the same product ID and the corresponding activity ID are divided into the same group, and the product ID and / or the corresponding activity ID are divided into different groups. Further, in each group, the group is further divided into multiple sub-buckets according to the crowd interception ratio corresponding to the product, that is, if the crowd interception ratio is N%, then the group is divided into N sub-buckets. And all net income in each sub-bucket is summed up to obtain the cumulative net income of each sub-bucket. For example, if a group contains 100 data and the crowd interception ratio is 50%, then the group is divided into 50 sub-buckets, and each sub-bucket contains two data. In a specific application scenario, combined with the matching score calculated in the step, in each group, the data are sorted according to the matching score, for example, sorted from high to low (of course, it can also be sorted from low to high), and then starting from the data with the highest score, every two data are divided into a sub-bucket.

[0070] Table 2 is an example of the calculation results for a group corresponding to (activity ID, product ID). For each group, a distribution D of the intercept ratio bucket ID vs. cumulative net revenue can be obtained.

[0071] Event ID Product ID Cut ratio Intercept proportion bucket ID Net income in the barrel Cumulative net income 3 3 5 1 20 20 3 3 5 2 10 30 3 3 5 3 5 35 3 3 5 4 8 43 3 3 5 5 -10 33

[0072] Table 2

[0073] In one embodiment, if Figure 2 Said S103 includes: steps S201 to S204.

[0074] S201, setting the horizontal axis interval [0, N] according to the population interception ratio N%, and taking the cumulative net income as the vertical axis, to construct the net income distribution graph;

[0075] S202, determining whether the cumulative net income trend in the net income distribution graph is a trend of first increasing and then decreasing or a trend of always increasing;

[0076] S203, if the cumulative net income shows a trend of first increasing and then decreasing, it is determined that the cumulative net income has a maximum value in the interval [0, N], and the population interception ratio corresponding to the maximum value is set as the candidate population interception ratio;

[0077] S204. If the cumulative net income shows a rising trend, it is determined that the population interception ratio N% is not the optimal solution, and a polynomial regression model is fitted based on the cumulative net income in the interval [0, N], and the candidate population interception ratio is calculated using the polynomial regression model.

[0078] In this embodiment, a net income distribution diagram is first constructed based on the cumulative net income and the population interception ratio. The net income distribution diagram may include: Figure 7 and Figure 8 The two forms shown, namely Figure 7 In the , the cumulative net income shows a trend of first increasing and then decreasing, and there is a maximum value in the population interception ratio interval [0, N], which means that the population interception ratio N is too large, including people who have no intention to redeem, resulting in a downward trend in the cumulative net income when the population interception ratio is close to N. Figure 1 In the case of , the population interception ratio corresponding to the maximum cumulative net income can be selected as the dynamic adjustment standard for this time, that is, the population interception ratio corresponding to the maximum cumulative net income is used as the candidate population interception ratio, such as Figure 1 The population interception ratio corresponding to the maximum cumulative net income is 4, so the population interception ratio after this dynamic adjustment is set to 4. Another form of combination Figure 8 It can be seen that in the net income distribution diagram, the cumulative net income shows an upward trend in the population interception ratio interval [0, N], that is, the cumulative net income has been rising in the interval [0, N] without reaching the highest point, which means that the current population interception ratio N is not the optimal solution and does not cover all people with redemption intentions. Therefore, the population interception ratio can be further expanded. In this case, a polynomial regression model can be fitted based on the cumulative net income data in the interval [0, N].

[0079] In one embodiment, if Figure 3 Said S204 includes: steps S301 to S306.

[0080] S301, set array X[1,2,3,4] to represent the first interception ratio, and array Y[20,30,35,25] to represent the second interception ratio;

[0081] S302: construct a polynomial feature [1, x, x] for each element x in the array X. 2 ,x 3 ,…x 7 ], get the matrix X;

[0082] S303, fitting the matrix X and the array Y to obtain a polynomial function: ax+bx 2 +cx 3 +…+gx 7 +h, where a, b, c…h are all parameters to be optimized in the polynomial function;

[0083] S304, obtaining the prediction result y_model corresponding to each eigenvector in the matrix X through the polynomial function, thereby constructing the polynomial regression model;

[0084] S305, calculating the derivative of the polynomial regression model corresponding to each population interception ratio within the interval [N-4, N+1];

[0085] S306, calculating the candidate population interception ratio n according to the following formula:

[0086] n=N+(arctan(A)*180 / pi / 90)*5

[0087] Where A is the average value of all derivatives and pi is the circumference of a circle.

[0088] In this embodiment, fitting a polynomial regression model according to the cumulative net income data in the interval [0, N] may specifically include:

[0089] Assume that array X[1,2,3,4] represents the first interception ratio, and array Y[20,30,35,25] represents the second interception ratio;

[0090] For each element x in array X, construct the polynomial feature [1, x, x 2 ,x 3 ,…x 7 ]. After feature construction, the array X is transformed into a (4, 8) matrix;

[0091] Fit the matrix X and array Y to get the polynomial function: ax+bx 2 +cx 3 +…+gx 7 +h, where a, b, c…h are all parameters to be optimized in the polynomial function, which can be obtained by random initialization;

[0092] Thus, each eigenvector in the matrix X can be used to obtain a prediction result y_model through the polynomial function. Furthermore, through the objective function f(a,b)=sqrt(abs(a 2 -b 2 ))Measure the error between the predicted result y_model and the actual net profit y, and reduce the error of the objective function through the gradient descent method, thereby optimizing the polynomial parameters so that it can fit the real curve.

[0093] After obtaining the polynomial regression model, calculate the derivative corresponding to each population interception ratio of the polynomial regression model in the interval [N-4, N+1], and then average all derivatives to get A. The significance of this step is to combine historical data and model predicted data to infer the trend of cumulative net income after point N. The reason why the derivative is calculated in the interval [N-4, N+1] is that the derivative of the curve represents the upward and downward trend of the image. Since the population interception ratio N needs to be adjusted, it is hoped that the interception ratio will be fine-tuned according to the upward and downward trend near point N. Specifically, the first-order derivative of the polynomial function f(x) is calculated to obtain a new function g(x). Substituting the interception ratio x into it, the corresponding derivative value can be obtained.

[0094] Then calculate the interception ratio of the candidate population according to the following formula: N+(arctan(A)*180 / pi / 90)*5, and use the calculated result as the interception ratio of the population after this dynamic adjustment, that is, the interception ratio of the candidate population. Here, if A is larger, it means that the cumulative net income after point N grows faster with N, so N will be increased by a larger amount as the new interception ratio; otherwise, N will be increased by a smaller amount.

[0095] In one embodiment, in step S104, the interception ratio of each product ID in each activity ID is averaged, and the interception ratio of each product ID in the crowd corresponding to the activity ID can be calculated. Further, the grouping is performed according to the score segment corresponding to each product, and the average interception ratio of the crowd in the score segment group is calculated to serve as the default interception ratio of the score segment. Of course, for products without activity records, their interception ratio can be set according to the default interception ratio of the corresponding score segment.

[0096] Here, the point segment is an inherent attribute of a product and is related to its corresponding business logic. For example, a product with a selling price of 10 points is in the point segment of 10-15.

[0097] In one embodiment, in step S105, the default crowd interception ratio is used as the latest marketing data to update the interception ratio of each commodity, thereby ensuring the real-time nature of the data. In order to improve the updating efficiency and real-time nature, the crowd interception ratio can be dynamically adjusted once a day or every few hours.

[0098] Figure 4 A schematic block diagram of a commodity intelligent recommendation device 400 provided in an embodiment of the present invention, the device 400 includes:

[0099] The record acquisition unit 401 is used to acquire, for each commodity, a historical redemption activity record corresponding to the commodity, and extract the net income and the population interception ratio of the commodity from the historical redemption activity record;

[0100] A division unit 402 is used to group the products based on the activities they belong to, and divide each group into multiple sub-buckets based on the population interception ratio, and then calculate the cumulative net income in each sub-bucket;

[0101] A dynamic adjustment unit 403 is used to construct a net income distribution diagram through all accumulated net income, and dynamically adjust the population interception ratio based on the net income distribution diagram, and use the dynamic adjustment result as the candidate population interception ratio;

[0102] The average value calculation unit 404 is used to obtain the score segment corresponding to the product, calculate the average value of the candidate population interception ratio within the score segment, and use the calculation result as the default population interception ratio of the corresponding score segment;

[0103] The real-time updating unit 405 is used to intercept the marketing population of the product by using the default population interception ratio, and to update the population interception ratio in real time based on the intercepted redemption activity records.

[0104] In this embodiment, for each commodity, the corresponding population interception ratio and net income are captured from its historical exchange activity records, wherein the population interception ratio refers to sorting based on the user's purchase intention score, and then selecting the top N% of the population to send marketing text messages; the net income is the actual net income in a certain exchange activity, and the net income = commodity profit - channel cost. For example, the population interception ratio corresponding to each commodity is captured from the SMS exchange activity records within a month. Then the commodities are grouped and bucketed, and the net income distribution chart is constructed based on the net income of the commodity and the current population interception ratio. Then, the population interception ratio of the commodity is dynamically adjusted according to the cumulative net income trend in the net income distribution chart, and based on the dynamically adjusted population interception ratio (i.e., the candidate population interception ratio), the default population interception ratio of the commodity under different score segments is calculated, and finally the marketing population of the commodity is intercepted and marketed (such as SMS marketing or email marketing, etc.) using the default population interception ratio. At the same time, in order to ensure the accuracy and real-time performance of the default population interception ratio, a dynamic adjustment can be made every preset time interval (such as 12 hours or 24 hours, etc.).

[0105] This embodiment realizes dynamic adjustment of the population interception ratio by grouping the current population interception ratio and net profit of the product, constructing a net profit distribution map, etc., so that the dynamically adjusted population interception ratio is more in line with the marketing recommendation of the product, thereby improving the marketing population interception accuracy and recommendation effect.

[0106] This embodiment realizes the dynamic adjustment of the interception ratio of the marketing population corresponding to each product based on the latest marketing data and model prediction results. In the prior art, a fixed ratio is used to intercept the marketing population each time, which cannot fully utilize the accumulated marketing data, and may cause inaccurate interception of the population due to inconsistent distribution of new data and historical data. This embodiment can adjust the ratio of the marketing population in a timely manner based on the latest data, which can minimize excessive marketing to people with no intention, and can also reduce ineffective marketing costs, improve the overall efficiency of SMS marketing activities, and realize a data-driven intelligent product recommendation strategy.

[0107] In one embodiment, the intelligent commodity recommendation device 400 further includes:

[0108] A scoring unit, for scoring the matching degree of each of the commodities and each user in the population cut-off ratio by using a scoring model;

[0109] The table construction unit is used to construct a benchmark table by combining the activity ID, product ID, user ID, net income, population interception ratio and matching score to which the product belongs.

[0110] In this embodiment, for each product, the scoring model is used to score the matching degree of the product and each user in the population interception ratio, so that a benchmark table can be constructed by combining the activity ID, product ID, user ID, net income, interception ratio and model score to which the product belongs. The benchmark table can be shown in Table 1:

[0111] Event ID Product ID User ID Net income Cut ratio Model Scoring 1 1 1 2 30 0.0672 1 1 2 -0.03 30 0.0004 2 5 3 5 50 0.0832 3 3 4 20 5 0.0735 3 3 5 10 5 0.0634 3 3 6 5 5 0.0613 3 3 7 8 5 0.0421 3 3 8 -10 5 0.0321

[0112] Table 1

[0113] In one embodiment, the scoring sheet includes:

[0114] A vector extraction unit, used to extract a product feature vector and a user feature vector for the product and the user respectively;

[0115] A vector input unit is used to input the product feature vector and the user feature vector into the Wide&Deep model, and use the Wide&Deep model to output the matching degree between the product and the user.

[0116] In this embodiment, the Wide&Deep model is composed of a shallow (or single-layer) Wide neural network and a deep Deep multi-layer neural network. The output layer uses softmax or logistics regression to integrate the outputs of the Wide and Deep parts. Among them, the Wide part is conducive to enhancing the "memory ability" of the model, and the Deep part is conducive to enhancing the "generalization ability" of the model. The Wide part is the basic linear model, expressed as y=WX+b. The X feature part includes basic features and cross features. The cross feature is very important in the Wide part, which can capture the interaction between features and play a role in adding nonlinearity; the Deep part is a feedforward network model. The features are first converted into low-dimensional dense vectors, and the dimensions are usually O(10)-O(100). The vector is randomly initialized and the model is trained by minimizing the random function. The activation function uses Relu. In the joint model, the outputs of the Wide and Deep parts are merged together in a weighted manner and finally output through the logistic loss function.

[0117] In one embodiment, the vector input unit comprises:

[0118] The Wide network unit is used to construct cross-features of the product feature vector and the user feature vector using the Wide network in the Wide&Deep model according to the following formula:

[0119]

[0120] y=w T [x,φ(x)]+b

[0121] Where y represents the predicted value output by the Wide network, φ k (x) represents the cross product transfer function, c ki is a Boolean variable, x=[x1, x2, ..., xd] is the product feature vector and user feature vector, w=[w1w2, ..., wd] is the model parameter, and b is the deviation coefficient;

[0122] The Deep network unit is used to concatenate the embedding vector and the dense vector using the Deep network in the Wide & Deep model according to the following formula:

[0123] a (l+1) =f(W (l) a (l) +b (l) )

[0124] In the formula, 1 represents the number of layers of the Deep network, f represents the activation function, and a (l) , b(l) and W (l) are the activation function, bias coefficient and model weight of the lth layer respectively;

[0125] The joint prediction unit is used to perform joint prediction based on the output results of the wide network and the deep network according to the following formula to obtain the final matching degree P between the product and the user:

[0126]

[0127] Where Y is the binary classification label, σ(.) is the sigmoid function, φ(x) is the cross-product transformation of the original feature x, b is the bias term, and W Wide is the weight vector of the Wide network, W Deep For the final activation function The weight of .

[0128] In this embodiment, the Wide & Deep model connects the Wide part of the single input layer with the Deep part consisting of the Embedding layer and multiple hidden layers, and inputs them together into the final output layer. The single-layer Wide part is good at processing a large number of sparse ID-type features; the Deep part uses the strong expressiveness of the neural network to perform deep feature cross-talk and mine the data patterns hidden behind the features. Finally, using the logistic regression model, the output layer combines the Wide part and the Deep part to form a unified model.

[0129] In one embodiment, the dividing unit 402 includes:

[0130] An ID acquisition unit, used to acquire a product ID and its corresponding activity ID from the reference table;

[0131] A grouping unit, configured to group historical redemption activity records with the same product ID and activity ID into the same group, and group historical redemption activity records with different product IDs and / or activity IDs into different groups;

[0132] A sub-bucket division unit is used to divide each of the groups into N sub-buckets according to a population interception ratio N% corresponding to the commodity;

[0133] The summing calculation unit is used to sum all the net income in each of the sub-buckets to obtain the accumulated net income of each of the sub-buckets.

[0134] In this embodiment, the product ID and its corresponding activity ID are obtained, and then the products in the historical redemption activity records are grouped according to (activity ID, product ID), that is, the same product ID and the corresponding activity ID are divided into the same group, and the product ID and / or the corresponding activity ID are divided into different groups. Further, in each group, the group is further divided into multiple sub-buckets according to the crowd interception ratio corresponding to the product, that is, if the crowd interception ratio is N%, then the group is divided into N sub-buckets. And all net income in each sub-bucket is summed up to obtain the cumulative net income of each sub-bucket. For example, if a group contains 100 data and the crowd interception ratio is 50%, then the group is divided into 50 sub-buckets, and each sub-bucket contains two data. In a specific application scenario, combined with the matching score calculated in the step, in each group, the data are sorted according to the matching score, for example, sorted from high to low (of course, it can also be sorted from low to high), and then starting from the data with the highest score, every two data are divided into a sub-bucket.

[0135] Table 2 is an example of the calculation results for a group corresponding to (activity ID, product ID). For each group, a distribution D of the intercept ratio bucket ID vs. cumulative net revenue can be obtained.

[0136] Event ID Product ID Cut ratio Intercept proportion bucket ID Net income in the barrel Cumulative net income 3 3 5 1 20 20 3 3 5 2 10 30 3 3 5 3 5 35 3 3 5 4 8 43 3 3 5 5 -10 33

[0137] Table 2

[0138] In one embodiment, if Figure 5 As shown, the dynamic adjustment unit 403 includes:

[0139] A coordinate setting unit 501 is used to set the horizontal coordinate interval [0, N] according to the population interception ratio N%, and to construct the net income distribution diagram with the cumulative net income as the vertical coordinate;

[0140] A trend determination unit 502 is used to determine whether the cumulative net income trend in the net income distribution diagram is a trend of first increasing and then decreasing or a trend of continuously increasing;

[0141] The first determination unit 503 is configured to determine that the cumulative net income has a maximum value in the interval [0, N] if the cumulative net income shows a trend of first increasing and then decreasing, and set the population interception ratio corresponding to the maximum value as the candidate population interception ratio;

[0142] The second determination unit 504 is used to determine that the population interception ratio N% is not an optimal solution if the cumulative net income shows a rising trend, and to obtain a polynomial regression model based on the cumulative net income in the interval [0, N], and to calculate the candidate population interception ratio using the polynomial regression model.

[0143] In this embodiment, a net income distribution diagram is first constructed based on the cumulative net income and the population interception ratio. The net income distribution diagram may include: Figure 7 and Figure 8 The two forms shown, namely Figure 7 In the , the cumulative net income shows a trend of first increasing and then decreasing, and there is a maximum value in the population interception ratio interval [0, N], which means that the population interception ratio N is too large, including people who have no intention to redeem, resulting in a downward trend in the cumulative net income when the population interception ratio is close to N. Figure 1 In the case of , the population interception ratio corresponding to the maximum cumulative net income can be selected as the dynamic adjustment standard for this time, that is, the population interception ratio corresponding to the maximum cumulative net income is used as the candidate population interception ratio, such as Figure 1 The population interception ratio corresponding to the maximum cumulative net income is 4, so the population interception ratio after this dynamic adjustment is set to 4. Another form of combination Figure 8 It can be seen that in the net income distribution diagram, the cumulative net income shows an upward trend in the population interception ratio interval [0, N], that is, the cumulative net income has been rising in the interval [0, N] without reaching the highest point, which means that the current population interception ratio N is not the optimal solution and does not cover all people with redemption intentions. Therefore, the population interception ratio can be further expanded. In this case, a polynomial regression model can be fitted based on the cumulative net income data in the interval [0, N].

[0144] In one embodiment, if Figure 6 As shown, the second determination unit 504 includes:

[0145] An array setting unit 601 is used to set an array X[1,2,3,4] to represent a first interception ratio, and an array Y[20,30,35,25] to represent a second interception ratio;

[0146] The matrix construction unit 602 is used to construct a polynomial feature [1, x, x] for each element x in the array X. 2 ,x 3 ,…x 7 ], get the matrix X;

[0147] The fitting unit 603 is used to fit the matrix X and the array Y to obtain a polynomial function: ax+bx 2 +cx 3 +…+gx 7 +h, where a, b, c…h are all parameters to be optimized in the polynomial function;

[0148] A model building unit 604 is used to obtain the prediction result y_model corresponding to each feature vector in the matrix X through the polynomial function, so as to build the polynomial regression model;

[0149] The derivative calculation unit 605 is used to calculate the derivative of the polynomial regression model corresponding to each population interception ratio within the interval [N-4, N+1];

[0150] The ratio calculation unit 606 is used to calculate the candidate group interception ratio n according to the following formula:

[0151] n=N+(arctan(A)*180 / pi / 90)*5

[0152] Where A is the average value of all derivatives and pi is the circumference of a circle.

[0153] In this embodiment, fitting a polynomial regression model according to the cumulative net income data in the interval [0, N] may specifically include:

[0154] Assume that array X[1,2,3,4] represents the first interception ratio, and array Y[20,30,35,25] represents the second interception ratio;

[0155] For each element x in array X, construct the polynomial feature [1, x, x 2 ,x 3 ,…x 7 ]. After feature construction, the array X is transformed into a (4, 8) matrix;

[0156] Fit the matrix X and array Y to get the polynomial function: ax+bx 2 +cx 3 +…+gx 7 +h, where a, b, c…h are all parameters to be optimized in the polynomial function, which can be obtained by random initialization;

[0157] Thus, each eigenvector in the matrix X can be used to obtain a prediction result y_model through the polynomial function. Furthermore, through the objective function f(a,b)=sqrt(abs(a 2 -b 2 ))Measure the error between the predicted result y_model and the actual net profit y, and reduce the error of the objective function through the gradient descent method, thereby optimizing the polynomial parameters so that it can fit the real curve.

[0158] After obtaining the polynomial regression model, calculate the derivative corresponding to each population interception ratio of the polynomial regression model in the interval [N-4, N+1], and then average all derivatives to get A. The significance of this step is to combine historical data and model predicted data to infer the trend of cumulative net income after point N. The reason why the derivative is calculated in the interval [N-4, N+1] is that the derivative of the curve represents the upward and downward trend of the image. Since the population interception ratio N needs to be adjusted, it is hoped that the interception ratio will be fine-tuned according to the upward and downward trend near point N. Specifically, the first-order derivative of the polynomial function f(x) is calculated to obtain a new function g(x). Substituting the interception ratio x into it, the corresponding derivative value can be obtained.

[0159] Then calculate the interception ratio of the candidate population according to the following formula: N+(arctan(A)*180 / pi / 90)*5, and use the calculated result as the interception ratio of the population after this dynamic adjustment, that is, the interception ratio of the candidate population. Here, if A is larger, it means that the cumulative net income after point N grows faster with N, so N will be increased by a larger amount as the new interception ratio; otherwise, N will be increased by a smaller amount.

[0160] In one embodiment, in the average value calculation unit 404, the interception ratio of each product ID in each activity ID is averaged, and the interception ratio of the crowd of each product ID in the corresponding activity ID can be calculated. Further, the grouping is performed according to the score segment corresponding to each product, and the average interception ratio of the crowd in the score segment group is calculated to serve as the default interception ratio of the score segment. Of course, for products without activity records, their interception ratio can be set according to the default interception ratio of the corresponding score segment.

[0161] Here, the point segment is an inherent attribute of a product and is related to its corresponding business logic. For example, a product with a selling price of 10 points is in the point segment of 10-15.

[0162] In one embodiment, the real-time updating unit 405 uses the default crowd interception ratio as the latest marketing data to update the interception ratio of each commodity, thereby ensuring the real-time nature of the data. In order to improve the updating efficiency and real-time nature, the crowd interception ratio can be dynamically adjusted once a day or every few hours.

[0163] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed, the steps provided in the above embodiment can be implemented. The storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0164] The embodiment of the present invention also provides a computer device, which may include a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps provided in the above embodiment may be implemented. Of course, the computer device may also include various network interfaces, power supplies and other components.

[0165] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of this application.

[0166] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A method for intelligent product recommendation, characterized in that: include: For each commodity, obtain the historical redemption activity record corresponding to the commodity, and extract the net income and population interception ratio of the commodity from the historical redemption activity record; The products are grouped based on the activities they belong to, and each group is divided into multiple sub-buckets based on the population interception ratio, and then the cumulative net income in each sub-bucket is calculated; Constructing a net income distribution graph through all accumulated net incomes, dynamically adjusting the population interception ratio based on the net income distribution graph, and using the dynamic adjustment result as the candidate population interception ratio; Obtain the score segment corresponding to the product, calculate the average value of the candidate population interception ratio within the score segment, and use the calculation result as the default population interception ratio of the corresponding score segment; The marketing population of the product is intercepted by using the default population interception ratio, and the population interception ratio is updated in real time based on the intercepted redemption activity records; Also includes: For each of the products, a matching score is given to each of the products and each user in the population cut-off ratio by using a scoring model; A benchmark table is constructed by combining the activity ID, product ID, user ID, net income, population interception ratio and matching score to which the product belongs; For each of the commodities, a matching score is scored between each of the commodities and each user in the population cut-off ratio by using a scoring model, including: Extracting a product feature vector and a user feature vector for the product and the user respectively; Inputting the product feature vector and the user feature vector into a Wide&Deep model, and using the Wide&Deep model to output a matching degree between the product and the user; The step of inputting the product feature vector and the user feature vector into a Wide&Deep model and using the Wide&Deep model to output a matching degree between the product and the user includes: According to the following formula, the Wide network in the Wide&Deep model is used to perform cross-feature construction on the product feature vector and the user feature vector: y=w T [x,φ(x)]+b Where y represents the predicted value output by the Wide network, Ф(x) represents the cross product conversion function, and c ki is a Boolean variable, x=[x1,x2,…x i , …x d ] is the product feature vector or user feature vector, w is the model parameter, and b is the deviation coefficient; According to the following formula, the Deep network in the Wide & Deep model is used to concatenate the embedding vector and the dense vector: a (l+1) =f(W (l) a (l) +b (l) ) In the formula, a (l+1) represents the activation function of the l+1th layer, l represents the number of layers of the Deep network, f(.) represents the activation function, a (l) , b (l) and W (l) are the activation function, bias coefficient and model weight of the lth layer respectively; Based on the output results of the Wide network and the Deep network, a joint prediction is performed according to the following formula to obtain the final matching degree between the product and the user: Where P(Y=1|x) represents the matching degree, Y is the binary classification label, σ(.) is the sigmoid function, and W wide is the weight vector of the Wide network, W deep For the final activation function The weight of .

2. The intelligent commodity recommendation method according to claim 1, characterized in that: The grouping is based on the activities to which the products belong, and each group is divided into multiple sub-buckets based on the population interception ratio, and then the cumulative net income in each sub-bucket is calculated, including: Obtaining the product ID and the activity ID to which it belongs in the reference table; The historical redemption activity records with the same product ID and the corresponding activity ID are divided into the same group, and the historical redemption activity records with different product IDs and / or corresponding activity IDs are divided into different groups; In each of the groups, the group is divided into N sub-buckets according to the population interception ratio N% corresponding to the commodity; All net revenues in each of the sub-buckets are summed up to obtain the cumulative net revenue of each of the sub-buckets.

3. The intelligent commodity recommendation method according to claim 1, characterized in that: The method of constructing a net income distribution diagram through all accumulated net income, dynamically adjusting the population interception ratio based on the net income distribution diagram, and using the dynamic adjustment result as the candidate population interception ratio includes: The horizontal axis interval [0, N] is set according to the population interception ratio N%, and the cumulative net income is used as the vertical axis to construct the net income distribution diagram; Determine whether the cumulative net income trend in the net income distribution graph is a trend of first increasing and then decreasing or a trend of continuously increasing; If the cumulative net income shows a trend of first increasing and then decreasing, it is determined that the cumulative net income has a maximum value in the interval [0, N], and the population interception ratio corresponding to the maximum value is set as the candidate population interception ratio; If the cumulative net income shows a rising trend, it is determined that the population interception ratio N% is not the optimal solution, and a polynomial regression model is fitted based on the cumulative net income in the interval [0, N], and the candidate population interception ratio is calculated using the polynomial regression model.

4. A commodity intelligent recommendation device, characterized in that: include: A record acquisition unit, for acquiring, for each commodity, a historical redemption activity record corresponding to the commodity, and extracting the net income and the population interception ratio of the commodity from the historical redemption activity record; A partitioning unit is used to group products based on the activities they belong to, and to divide each group into multiple sub-buckets based on the population interception ratio, and then calculate the cumulative net income in each sub-bucket; A dynamic adjustment unit, configured to construct a net income distribution diagram through all accumulated net income, dynamically adjust the population interception ratio based on the net income distribution diagram, and use the dynamic adjustment result as the candidate population interception ratio; an average value calculation unit, used to obtain the score segment corresponding to the product, calculate the average value of the candidate population interception ratio within the score segment, and use the calculation result as the default population interception ratio of the corresponding score segment; A real-time updating unit, configured to intercept the marketing population of the product using the default population interception ratio, and to update the population interception ratio in real time based on the intercepted redemption activity record; The intelligent commodity recommendation device further includes: A scoring unit, for scoring the matching degree of each of the commodities and each user in the population cut-off ratio by using a scoring model; A table construction unit, used to construct a benchmark table by combining the activity ID, product ID, user ID, net income, population interception ratio and matching score to which the product belongs; The scoring unit comprises: A vector extraction unit, used to extract a product feature vector and a user feature vector for the product and the user respectively; A vector input unit, used to input the product feature vector and the user feature vector into a Wide&Deep model, and output the matching degree between the product and the user using the Wide&Deep model; The vector input unit comprises: The Wide network unit is used to construct cross-features of the product feature vector and the user feature vector using the Wide network in the Wide&Deep model according to the following formula: y=w T [x,φ(x)]+b Where y represents the predicted value output by the Wide network, Ф(x) represents the cross product conversion function, and c ki is a Boolean variable, x=[x1,x2,…x i , …x d ] is the product feature vector or user feature vector, w is the model parameter, and b is the deviation coefficient; The Deep network unit is used to concatenate the embedding vector and the dense vector using the Deep network in the Wide & Deep model according to the following formula: a (l+1) =f(W (l) a (l) +b (l) ) In the formula, a (l+1) represents the activation function of the l+1th layer, l represents the number of layers of the Deep network, f(.) represents the activation function, a (l) , b (l) and W (l) are the activation function, bias coefficient and model weight of the lth layer respectively; The joint prediction unit is used to perform joint prediction based on the output results of the wide network and the deep network according to the following formula to obtain the final matching degree between the product and the user: Where P(Y=1|x) represents the matching degree, Y is the binary classification label, σ(.) is the sigmoid function, and W wide is the weight vector of the Wide network, W deep For the final activation function The weight of .

5. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the commodity intelligent recommendation method as claimed in any one of claims 1 to 3 when executing the computer program.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for intelligently recommending commodities according to any one of claims 1 to 3 is implemented.

Images