A data processing method, device, medium and equipment for a recommendation system

Abstract

The data processing method, device, medium and equipment of a recommendation system provided by an embodiment of the present invention have the following beneficial effects: obtaining sample data of multiple periodic statistics, calculating the ratio of the actual click rate to the estimated click rate average; using the ratio as the regression label, batch gradient descent solves to obtain the click rate influence factor model; selecting reference products from the recall list product set according to the preset target ratio; obtaining the click rate influence factor of each reference product from the click rate influence factor model; calculating the corresponding reference score; presetting the screening set; cyclically calculating the feature score of each reference product according to the characteristics of the product; and forming a recommendation set with products with high feature scores. The present invention performs weight reduction control on the exposed product data, adopts a multi-way recall fusion method, fully considers the frequency and marginal sensitivity, realizes the data processing optimization of the recommendation system, effectively improves the data precision and thus improves the accuracy of the recommendation system.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic data processing, and in particular to a data processing method, device, medium and equipment for a recommendation system. Background Art

[0002] Recommendation system technology usually adopts a layered solution, which includes a recall layer, a fusion layer, a sorting layer, and a re-ranking layer. In the recall layer, certain rules, algorithms, or models are used to find products that users may be interested in. This layer usually uses multiple methods or uses multiple methods at the same time to find a large number of candidate products. In the fusion layer, the product lists recalled by each method are fused into a product list according to certain algorithm rules, and then the first hundreds are taken for sorting. The sorting layer uses a unified model to score the products selected by the fusion layer and sort them from high to low according to the scores. The re-ranking layer will intervene in the sorted product list, such as placing certain products at the top, breaking up similar products, and weighting / de-weighting products with certain attributes. After the intervention is finally completed, the first few or dozens of products in the list are taken for output. However, in such a system framework, it corresponds to the processing of a large amount of product and user data, and each layer is progressive and iterative. The accuracy of data calculation is the core factor that determines the accuracy of the recommendation system.

[0003] Therefore, how to improve the accuracy of the recommendation system is a technical problem that technical personnel in this field need to solve urgently. Summary of the invention

[0004] In order to overcome the shortcomings of poor accuracy of existing recommendation systems, the present invention proposes a data processing method, device, medium and equipment for a recommendation system.

[0005] In order to achieve the above objective, according to a first aspect of the present invention, an embodiment of the present invention provides a data processing method for a recommendation system, the method comprising the following steps:

[0006] Acquire multiple periodic statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click rates, and click flags corresponding to the product and the user;

[0007] Divide the sample data into multiple data groups according to the historical exposure times and the historical click times. In the data groups where the sample data volume is greater than or equal to the threshold, count the actual click rate by the click mark, and calculate the ratio of the actual click rate to the average value of the estimated click rate.

[0008] Using the ratio as a regression label, batch gradient descent is performed to obtain a click rate influence factor model corresponding to the product and the user, wherein the click rate influence factor model obtains a click rate influence factor in response to historical exposure times and historical click times;

[0009] Select reference products from the recalled list products according to the preset target ratio;

[0010] The click rate influence factor of each reference product is obtained from the click rate influence factor model;

[0011] Calculate the corresponding reference score based on the price, estimated click-through rate, estimated conversion rate and click-through rate impact factor of each reference product;

[0012] A preset filter set, wherein the filter set includes products that are recently recommended to the user, or reference products with the highest reference scores;

[0013] According to the characteristics of the products, the feature score of each reference product relative to the products in the screening set is calculated cyclically, and the reference products with high scores are added to the screening set. The feature score is the weighted difference between the reference score and the similarity of the reference product and the products in the screening set.

[0014] Prioritize products with high feature scores to form a recommendation set.

[0015] Optionally, the ratio is used as a regression label, and batch gradient descent is used to solve the click rate influencing factor model corresponding to the product and the user, including calculating through the formula:

[0016] +

[0017] ,

[0018] ,

[0019]

[0020] in, is the click impact factor, a, b, is the model parameter obtained by batch gradient descent, N is the total data volume of historical exposure times or historical click times recorded in a period, C is the click impact, E is the exposure impact, is the i-th data in the historical exposure times recorded in a period. is the i-th data in the historical click count recorded in a period, Y is the ratio, is the actual click-through rate, Estimate the average click rate; use C and E as features and Y as regression labels to fit C, E and The mapping relationship.

[0021] Optionally, selecting reference products from the recalled list products according to a preset target ratio includes:

[0022] Calculate the Euclidean distance of each recall in the recall list product set. The formula is as follows:

[0023]

[0024] The estimated recall ratios of each route after recalling one product from route i are as follows:

[0025]

[0026] in, is the recall ratio of the current i-th path, is the number of recalled products on the current i-th route, is the total number of current recalls; is the recall ratio after the i-th route recalls one product, is the recall ratio of the jth route after the ith route recalls one product, is the number of recalled products in route j except route i; is the Euclidean distance of the ith path, is the target ratio of the jth road;

[0027] Recall one product from the route with the minimum Euclidean distance in a loop until the preset total number of recalls is reached, including: when the number of products in the jth route is 0, update the target ratio of other routes according to the following formula, and recall one product from other routes with the minimum Euclidean distance in a loop:

[0028]

[0029] in, is the target ratio of the i-th route, and n is the total number of routes whose current number of products is not 0.

[0030] Optionally, the calculation of the corresponding reference score by the price, estimated click rate, estimated conversion rate and click rate impact factor of each reference product includes:

[0031]

[0032] Among them, Score is the reference score, is the click impact factor, is the estimated click-through rate for the corresponding product. is the estimated conversion rate of the corresponding product, and Price is the price of the corresponding product.

[0033] Optionally, according to the characteristics of the products, the feature score of each reference product relative to the products in the screening set is cyclically calculated, and the reference products with high scores are added to the screening set, wherein the feature score is a weighted difference between the reference score and the similarity between the reference product and the products in the screening set, including:

[0034]

[0035]

[0036] in, is the feature score of the i-th reference product, is the relevance weight, represents the click impact factor of the i-th reference product, S is the current filter set, is the similarity between the i-th reference product and the j-th product in the screening set, The interval between the i-th reference product and the j-th product in the screening set, ; is the title edit distance between the i-th reference product and the j-th product in the filter set, is the third-level catalog matching value between the i-th reference product and the j-th product in the filter set, is the secondary catalog matching value between the i-th reference product and the j-th product in the filter set, is the brand matching value between the i-th reference product and the j-th product in the filter set, is the code matching value between the i-th reference product and the j-th product in the filter set.

[0037] According to a second aspect of the present invention, an embodiment of the present invention further provides a data processing device for a recommendation system, comprising:

[0038] An acquisition module, used to acquire multiple periodic statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click rates, and click flags corresponding to products and users;

[0039] A statistical module is used to divide the sample data into multiple data groups according to the historical exposure times and the historical click times. In the data groups where the sample data volume is greater than or equal to the threshold, the actual click rate is counted by the click mark, and the ratio of the actual click rate to the average value of the estimated click rate is calculated;

[0040] A training module, configured to use the ratio as a regression label and perform batch gradient descent to obtain a click rate influence factor model corresponding to the product and the user, wherein the click rate influence factor model obtains the click rate influence factor in response to the historical exposure times and the historical click times;

[0041] The recall module is used to select reference products from the recall list products according to a preset target ratio;

[0042] A factor calculation module is used to obtain the click rate influence factor of each reference product from the click rate influence factor model;

[0043] A rating calculation module, used to calculate the corresponding reference rating through the price, estimated click rate, estimated conversion rate and click rate influencing factor of each reference product;

[0044] A configuration module, used to preset a filter set, wherein the filter set includes products recently recommended to the user, or reference products with the highest reference scores;

[0045] A screening module, for cyclically calculating the feature score of each reference product relative to the products in the screening set according to the features of the products, and adding the reference products with high scores to the screening set, wherein the feature score is the weighted difference between the reference score and the similarity between the reference product and the products in the screening set;

[0046] The result module is used to prioritize the products with high feature scores into the recommendation set.

[0047] Optionally, the training module is further used to obtain a click rate influencing factor model corresponding to the product and the user by solving the following formula:

[0048] +

[0049] ,

[0050] ,

[0051]

[0052] in, is the click impact factor, a, b, is the model parameter obtained by batch gradient descent, N is the total data volume of historical exposure times or historical click times recorded in a period, C is the click impact, E is the exposure impact, is the i-th data in the historical exposure times recorded in a period. is the i-th data in the historical click count recorded in a period, Y is the ratio, is the actual click-through rate, Estimate the average click rate; use C and E as features and Y as regression labels to fit C, E and The mapping relationship.

[0053] Optionally, the screening module calculates the feature score by the following formula:

[0054]

[0055]

[0056] in, is the feature score of the i-th reference product, is the relevance weight, represents the click impact factor of the i-th reference product, S is the current filter set, is the similarity between the i-th reference product and the j-th product in the screening set, The interval between the i-th reference product and the j-th product in the screening set, ; is the title edit distance between the i-th reference product and the j-th product in the filter set, is the third-level catalog matching value between the i-th reference product and the j-th product in the filter set, is the secondary catalog matching value between the i-th reference product and the j-th product in the filter set, is the brand matching value between the i-th reference product and the j-th product in the filter set, is the code matching value between the i-th reference product and the j-th product in the filter set.

[0057] According to the third aspect of the present invention, an embodiment of the present invention further provides a storage medium, wherein at least one instruction, at least one program, a code set or an instruction set is stored in the storage medium, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the steps of the data processing method of the recommendation system in any of the above embodiments.

[0058] According to a fourth aspect of the present invention, an embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the data processing method of the recommendation system in any of the above embodiments are implemented.

[0059] As described above, a data processing method, device, medium and equipment for a recommendation system provided by an embodiment of the present invention has the following beneficial effects: obtaining a plurality of periodically statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click-through rates and click marks corresponding to products and users; dividing the sample data into a plurality of data groups according to the historical exposure times and historical click times, and in the data groups where the sample data volume is greater than or equal to a threshold, counting the actual click-through rate by the click mark, and calculating the ratio of the actual click-through rate to the average of the estimated click-through rate; using the ratio as a regression label, batch gradient descent is performed to obtain a click-through rate influencing factor model corresponding to products and users, wherein the click-through rate influencing factor model A click-through rate influence factor is obtained in response to the number of historical exposures and the number of historical clicks; reference products are selected from the recall list product set according to a preset target ratio; the click-through rate influence factor of each reference product is obtained from the click-through rate influence factor model; the corresponding reference score is calculated by the price, estimated click-through rate, estimated conversion rate and click-through rate influence factor of each reference product; a preset screening set is set, and the screening set includes products that have been recently recommended to users, or reference products with the highest reference scores; based on the characteristics of the products, the feature score of each reference product relative to the products in the screening set is cyclically calculated, and the feature score is the weighted difference between the reference score and the similarity of the reference product and the product in the screening set; products with high feature scores are preferentially formed into the recommendation set. The present invention performs downweighting control on the exposed product data, adopts a multi-way recall fusion method, fully considers frequency and marginal sensitivity, and optimizes the processing of data in the recommendation system, effectively improving data accuracy and thus improving the accuracy of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a flowchart of a data processing method of a recommendation system provided by an embodiment of the present invention;

[0061] Figure 2 It is a flowchart of a multi-channel recall method provided by an embodiment of the present invention;

[0062] Figure 3 is a structural schematic diagram of a data processing device of a recommendation system provided by an embodiment of the present invention;

[0063] Figure 4 It is a schematic diagram of the hardware structure of an electronic device for executing the data processing method of the recommendation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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 should fall within the scope of protection of the present invention.

[0065] See also Figures 1 to 4 It should be noted that the illustrations provided in this embodiment are only used to illustrate the basic concept of the present invention in a schematic manner, and the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0066] See also Figure 1 , is a flow chart of a data processing method of a recommendation system provided by an embodiment of the present invention, such as Figure 1 As shown, the embodiment of the present invention shows the flow of the data processing method of the recommendation system.

[0067] Step S101: Acquire multiple periodic statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click rates, and click flags corresponding to products and users.

[0068] The present invention adopts a method for controlling and devaluing products that are adaptively exposed based on user feedback. On the basis of recommended filtering or exposed filtering, based on a certain time window, products that have been exposed or recommended within a period of time are directly filtered out. In addition, it includes some optimizations, such as adding exposure frequency constraints, filtering out products after being exposed for more than a few times, etc. The following four factors are further considered: 1. The user's tolerance for repeated exposure of products gradually decreases with the frequency. 2. The user's tolerance for exposed products increases with time, that is, the user has forgetfulness, and the longer the exposure, the smaller the impact. 3. The user's tolerance for repeated exposure of products that have been clicked recently is relatively high, but it will gradually decrease with the click time. 4. The user's behavior pattern is slowly changing, and it needs to be dynamically learned and adjusted from user feedback.

[0069] In this application, the behavior logs of user recommendation scene exposure and product clicks are recorded, and the historical exposure times and historical click times of a certain product by the user are counted according to time intervals (the time interval in one implementation case of the present invention is taken as one day / 24 hours as an example).

[0070] The number of historical exposures and clicks on the product in the user's most recent N time intervals (7 days, i.e., one cycle, is selected in one implementation case of the present invention), together with the estimated click-through rate ctr (English: Click Through Rate, Chinese: click-through rate) at that time, are recorded. Here, each of the N time intervals has the number of historical exposures and clicks of the user on a certain product in that time interval. If not, it is recorded as 0. It should be noted that the estimated click-through rate is a deep neural network model trained by historical data in the recommendation system, which calculates the click-through rate when the product is displayed to the user based on user characteristics and product characteristics, in order to estimate the click probability of the product after it is recommended to the user. The specific process will not be repeated in the embodiments of this application. In addition, the sample data further records whether the user actually clicks after the estimated click probability, i.e., the click mark.

[0071] Through the above description, a complete sample data contains the combination of the historical exposure times, historical click times, estimated click rate and click flag corresponding to the product and user. An exemplary sample data format and its recording process are as follows:

[0072] Record the number of times the user has been exposed to the product every day for the past 7 days (including the current day) and the number of clicks, and store them in redis (English: Remote Dictionary Server, Chinese: Remote Dictionary Service). When recommending products to users, send the user's previous exposure times and historical click times of the product, along with the estimated click-through rate ctr, to the log system for snapshot recording. Obtain user record information from the log and organize it into sample data as shown in the following table:

[0073]

[0074] Taking the first sample data as an example, the historical exposure count [3,0,2,1,1,0,1] represents the number of times the product was recommended to users every day in the past 7 days, for example, 3 times on the first day, 0 times on the second day, and 2 times on the third day. The historical click count [1,0,0,0,1,0,0] represents the number of times users clicked given the number of recommendations in the past 7 days, for example, 1 click on the first day, 0 clicks on the second day, etc. The estimated click rate is the click rate of the next recommendation estimated by the deep neural network model in the recommendation system. The estimated result here is 0.3, which means that there is a 30% probability that the user will click on the product. The click mark is whether the user actually clicks on the product after the estimated click rate is pushed to the user. The click mark "1" indicates that a click has occurred, and the click mark "0" indicates that no click has occurred.

[0075] Step S102: Divide the sample data into multiple data groups according to the historical exposure times and the historical click times. In the data groups where the sample data volume is greater than or equal to the threshold, count the actual click rate by clicking on the mark, and calculate the ratio of the actual click rate to the average estimated click rate.

[0076] According to the structure of the sample data described in the above embodiment, statistics are collected from the behavior log according to the user's exposure and click situations on the products within N time intervals. That is, sample data with the same historical exposure times and historical click times are organized into a data group. The actual click rate is counted by the click mark, and the average of the estimated click rate is counted. The two are divided to obtain the corresponding ratio, and the amount of sample data (that is, the number of samples) is also counted.

[0077] Based on the above example, the sample data after this step is processed is as follows:

[0078]

[0079] Taking the first data group as an example, based on the sample data recorded in the previous step, there are 4 sample data corresponding to the historical exposure times [3,0,2,1,1,0,1] and the historical click times [1,0,0,0,1,0,0], which are organized into a data group, and the sample data volume of this data group is 4. The estimated click-through rates of the 4 sample data are averaged, and the average estimated click-through rate is 0.32. Among these 4 sample data, there is only 1 click mark "1", so the actual click-through rate is the number of click marks "1" 1 divided by the sample data volume of this data group 4, so the actual click-through rate is 1 / 4=0.25.

[0080] Furthermore, we calculate the ratio of the actual click-through rate to the average estimated click-through rate. Taking Group 1 as an example, the ratio is 0.25 actual click-through rate / 0.32 estimated click-through rate = 0.78125. The specific results are as follows:

[0081]

[0082] Of course, the above embodiments are only exemplary. In specific implementations, data groups with smaller sample data volumes will also be removed. For example, taking the threshold of 3000 as an example, data groups with sample data volumes greater than or equal to the threshold of 3000 will be selected to increase statistical confidence.

[0083] Step S103: using the ratio as a regression label, batch gradient descent is performed to obtain a click rate influence factor model corresponding to the product and the user, wherein the click rate influence factor model obtains a click rate influence factor in response to historical exposure times and historical click times.

[0084] The data processed by the above embodiment is input into the model, and the model parameters are solved by the gradient descent method to obtain the click rate influencing factor model. The specific data processing process is as follows:

[0085] +

[0086] ,

[0087] ,

[0088]

[0089] in, is the click impact factor, a, b, is the model parameter obtained by batch gradient descent, N is the total data volume of historical exposure times or historical click times recorded in a period, C is the click impact, E is the exposure impact, is the i-th data in the historical exposure times recorded in a period. is the i-th data in the historical click count recorded in a period, Y is the ratio, is the actual click-through rate, Estimate the average click rate; use C and E as features and Y as regression labels to fit C, E and The mapping relationship.

[0090] In this application, it is assumed that the impact of user exposure and click events on the future is geometrically attenuated. The number of clicks of a certain user on a certain product in N time intervals (i.e., 7 days in the above embodiment, with one time interval per day) is recorded from far to near as The attenuation ratio of each time interval is a, and the exposure is recorded as , the influence decay ratio of each time interval is b, then the click influence , Exposure Effect E , actual ctr impact .

[0091] In this way, C and E are used as features, Y is used as the regression label, and an M-order polynomial is selected to fit the mapping relationship between C, E and Y to predict the impact of ctr .

[0092] Further, choose MSE as the loss, that is , using batch gradient descent method to solve, get the model parameters a, b, ~ Therefore, based on the historical exposure times and historical click times of a product recommended to a user, the click rate impact factor corresponding to the product and the user can be calculated.

[0093] Step S104: Select reference products from the recalled list products according to a preset target ratio.

[0094] After obtaining the click rate influencing factor model, further, according to the configuration, the recall list product set is obtained. The recall list set is a collection of a series of products, filtering out the recommended products, out-of-stock products, and off-shelf products on all pages before the current page, and the products that cannot be recommended due to special reasons. Then, the recall list product set can further include multiple subsets, represented by paths, and multiple paths select reference products according to a preset target ratio. The target ratio is the ratio of the selected products in each path. Moreover, the total number of selected reference products (i.e., the total recall number) is also pre-configured, such as 500, etc., which is not limited in the embodiments of the present invention.

[0095] In specific implementation, the embodiment of the present application first calculates, assuming that one recall is made this time, the recall ratio of each route after the recall, and then selects the route whose ratio after the recall is closest to the target ratio for recall. Then repeat until the recall quantity is finally reached. If a recall is empty in the middle, then the target ratio of this recall is proportionally distributed to the remaining recall routes, and the cycle continues. The method provided by the present application is concise in concept, and the result guarantees the desired ratio of each route to the greatest extent, and the ratio can be arbitrarily controlled, which is convenient for parameter adjustment experiments. See Figure 2 , is a flowchart of a multi-channel recall method provided by an embodiment of the present invention. As shown in the figure, the specific steps of the method are as follows:

[0096] Step S1041: Calculate the Euclidean distance of each recall in the recall list product set.

[0097] For example, the recall list product set has n recalls, and the target proportions of each recall are The current number of recalled products is , the current total recall quantity is T.

[0098] Define the current recall ratio of each channel as:

[0099] in, is the recall ratio of the current i-th path, is the number of recalled products on the current i-th route, is the current total recall quantity.

[0100] Assume that a product is recalled from the i-th route, the proportion of recalls from each route after the recall becomes:

[0101]

[0102]

[0103] in, is the recall ratio after the i-th route recalls one product, is the recall ratio of the jth route after the ith route recalls one product, is the number of recalled products in the jth route except the i-th route.

[0104] Furthermore, after recalling one product from each route, the difference between the recall ratio of each route and the target ratio is calculated. The difference can be represented by Euclidean distance or other distance measurement methods. The embodiment of the present invention uses Euclidean distance:

[0105]

[0106] in, is the Euclidean distance of the ith path, is the target ratio of the j-th road.

[0107] Step S1042: recall one product from the path with the minimum Euclidean distance.

[0108] use The next time a recall is made, after the recall, the current number of recalled products on each channel and the total recall amount will be updated, and the cycle will continue until the recall quantity is met.

[0109] Step S1043: When the number of commodities in the j-th route is 0, the target ratios of other routes are updated.

[0110] When a recall is completed but the overall recall quantity is not met, the ratio is redistributed as follows. According to the description of the above embodiment, there are n recalls in total, and the target ratio of each recall is .

[0111] When the number of recalled products on route j is 0, the recall ratios of other routes are updated according to the following formula:

[0112]

[0113] in, is the target ratio of the i-th route, n is the total number of routes whose current number of products is not 0, so continue to take recalled products from the remaining n-1 recall routes according to the previous method until the total number of recalled products reaches the preset total number of reference products.

[0114] The following is an exemplary embodiment, describing a method for processing data during a specific recall process:

[0115] Assume there are 3 recalls, and the target ratios are

[0116] The current number of recalled products is: , the total recall number is T=9.

[0117] First, calculate the proportion of products recalled from a certain route:

[0118] Assume recall from route 1:

[0119]

[0120]

[0121]

[0122] = 0

[0123] Assume recall from route 2:

[0124]

[0125]

[0126]

[0127]

[0128] Assume recall from route 3:

[0129]

[0130]

[0131]

[0132]

[0133] Secondly, by comparing the three-way distance, it is found that after the first recall, the gap with the target ratio is the smallest, so the first recall is selected.

[0134] Then, after the recall, update the recall quantity of each channel:

[0135] If a recall route is used up, update the recall ratios of the remaining routes.

[0136] Repeat the above steps until the recall quantity is full and the total quantity of the reference product is reached.

[0137] It should be noted that in actual use, the present application can evaluate the most reasonable proportion of each recall through offline simulation and online testing.

[0138] Step S105: Obtaining the click rate influence factor of each reference product from the click rate influence factor model.

[0139] The click rate influence factor model obtained in step S103 calculates the click rate influence factor of each reference product by inputting the historical exposure times and historical click times of each product in the most recent period (7 days) for the reference products obtained in step S104.

[0140] Step S106: Calculate the corresponding reference score based on the price, estimated click-through rate, estimated conversion rate and click-through rate influencing factor of each product.

[0141] Get the price of the product; get the features of the user and all reference products, and request the CTR model (click-through rate model) and CVR model (conversion rate model) trained in the recommendation system in batches, and the estimated click-through rate and estimated conversion rate corresponding to each reference product; finally, combine the click-through rate influence factor calculated in the above embodiment steps, and calculate the corresponding reference score according to the following formula:

[0142]

[0143] Among them, Score is the reference score, is the click impact factor, is the estimated click-through rate for the corresponding product. is the estimated conversion rate of the corresponding product, and Price is the price of the corresponding product

[0144] In addition, if the product user has not been exposed, The click impact factor defaults to 1, and then the reference scores are sorted from large to small.

[0145] S107: Preset a filter set, where the filter set includes products that are recently recommended to the user, or reference products with the highest reference scores.

[0146] The filter set is a collection of products that have been recently recommended to users. For example, when recommending on a computer or mobile terminal web page, you can obtain the list of products that have been recommended on this web page before and take the last 20 as the filter set. Or, if this web page is page 1, take the first one in the above embodiment, that is, the reference product with the highest reference score, and add it to the filter set. This filter set is used as the initial set for data processing in the subsequent steps.

[0147] Step S108: Based on the characteristics of the products, the feature score of each product relative to the products in the screening set is calculated cyclically, and reference products with high scores are added to the screening set. The feature score is the weighted difference between the reference score and the similarity of the reference product and the products in the screening set.

[0148] In the data processing of recommended products, the diversity of products is further dispersed. The rule-based dispersion takes frequency control into account but does not take user interest into account. The more interested a user is in a product, the higher the tolerance for repeated similar products. On the contrary, if a user is not very interested in a product, the user may be disgusted by the appearance of two copies. The dispersion based on MMR (MaximalMarginal Relevance) takes user interest into account but does not take frequency control into account, and only considers the maximum marginal similarity. The embodiment of the present invention comprehensively considers user interest and takes into account the frequency of occurrence of similar products. The specific method is as follows:

[0149]

[0150]

[0151] in, is the feature score of the i-th reference product, is the relevance weight, represents the click impact factor of the i-th reference product, S is the current filter set, is the similarity between the i-th reference product and the j-th product in the screening set, The time interval between the push of the i-th reference product and the j-th product in the filter set, ; is the title edit distance between the i-th reference product and the j-th product in the filter set, is the third-level catalog matching value between the i-th reference product and the j-th product in the filter set, is the secondary catalog matching value between the i-th reference product and the j-th product in the filter set, is the brand matching value between the i-th reference product and the j-th product in the filter set, is the code matching value between the i-th reference product and the j-th product in the filter set.

[0152] It should be noted that Indicates reference product The correlation between it and user interest U.

[0153] Indicates reference product Diversity with the selected filter set items. Display products Items with current filter set The similarity between .

[0154] Inner Layer is a softmax function used to convert similarity into weight distribution and introduce a decay factor. Is for all selected filter set products The attenuation factor Reference products Filter set products The interval between them, where the interval can be understood as the push time interval, or one or more combinations of the push product category difference or the web page interval. The larger the interval, the greater the attenuation. .

[0155] exist In the data calculation process, the title edit distance is calculated for the title names of two products and is used to measure the difference between two strings. It represents the minimum number of single-character edit operations required to convert from one string to another. These edit operations include:

[0156] 1. Insertion: Insert a character into a string.

[0157] 2. Deletion: Delete a character from a string.

[0158] 3. Substitution: Replace one character in a string with another character.

[0159] Title edit distance is often used in natural language processing, spell checking, DNA sequence analysis and other fields to help determine the similarity between two strings. In this application, title edit distance It is the normalized value of the product title edit distance divided by the maximum title length.

[0160] In addition, the third-level directory matching value , Secondary directory matching value , brand matching value , encoding matching value The value is calculated based on whether the two products match. For example, and filter focused products All belong to a third-level directory, then The value is 1, otherwise it is 0. Similarly, the secondary catalog matching value determines whether the two products belong to the same secondary catalog, the brand matching value determines whether the two products belong to the same brand, and the code matching value determines whether the two products belong to the same product code. If so, the corresponding value is 1, otherwise it is 0. I will not go into details here.

[0161] The data processing method provided in this application considers the similarity between the current reference product and all selected products in the screening set when calculating diversity, and introduces an attenuation factor to give products that are farther away from the current position a lower weight. Overall, the parameters balance diversity and user interests, solve the problems of existing mainstream methods, have higher data accuracy, and achieve better user experience in practical applications.

[0162] Step S109: prioritize products with high feature scores to form a recommendation set.

[0163] In the present application, a certain number of screening sets are obtained through a cyclic calculation of the above steps. When the preset number is met (for example, 10), the set of products with high feature scores can be further combined into a recommendation set (including 10 products in the screening set with high feature scores) for the recommendation system to call.

[0164] It can be seen from the description of the above embodiments that an embodiment of the present invention provides a data processing method for a recommendation system, including obtaining sample data of multiple periodic statistics, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click-through rates, and click marks corresponding to products and users; dividing the sample data into multiple data groups according to the historical exposure times and historical click times, and in the data groups where the sample data volume is greater than or equal to a threshold, counting the actual click-through rate through the click mark, and calculating the ratio of the actual click-through rate to the average value of the estimated click-through rate; using the ratio as a regression label, batch gradient descent is used to solve a click-through rate influencing factor model corresponding to the products and users, wherein the click-through rate influencing factor model responds to the historical exposure times and historical click times. The click-through rate influence factor is obtained by the number of clicks; reference products are selected from the recall list product set according to the preset target ratio; the click-through rate influence factor of each reference product is obtained by the click-through rate influence factor model; the corresponding reference score is calculated by the price, estimated click-through rate, estimated conversion rate and click-through rate influence factor of each reference product; a preset screening set is set, and the screening set includes the products that have been recommended to the user recently, or the reference products with the highest reference scores; according to the characteristics of the products, the feature score of each reference product relative to the products in the screening set is cyclically calculated, and the reference products with high scores are added to the screening set, and the feature score is the weighted difference between the reference score and the similarity of the reference product and the products in the screening set; products with high feature scores are preferentially formed into the recommendation set. The present invention performs weight reduction control on the exposed product data, comprehensively considers the influence of various data dynamic factors, and effectively corrects the data in the sorting layer; at the recall layer, a multi-way recall fusion method is adopted, which not only ensures the orderliness after fusion, but also takes into account the fusion ratio; moreover, in the process of diversity dispersion control, frequency and marginal sensitivity are fully considered. In summary, the data processing method of the recommendation system of the present application optimizes the data of the recommendation system, effectively improves the data accuracy and thus improves the accuracy of the recommendation system.

[0165] Through the description of the above method embodiments, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk.

[0166] An embodiment of the present invention provides a non-volatile computer storage medium, wherein the computer storage medium stores computer executable instructions, and the computer executable instructions can execute the data processing method of the recommendation system in any of the above method embodiments.

[0167] Corresponding to the data processing method embodiment of the recommendation system provided by the present invention, the present invention also provides a data processing device of the recommendation system.

[0168] See also Figure 3 , is a schematic diagram of the structure of a data processing device of a recommendation system provided by an embodiment of the present invention. As shown in the figure, the device includes:

[0169] An acquisition module 11 is used to acquire a plurality of periodic statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click rates, and click flags corresponding to the product and the user;

[0170] A statistics module 12 is used to divide the sample data into multiple data groups according to the historical exposure times and the historical click times, and in the data groups where the sample data volume is greater than or equal to the threshold, count the actual click rate by the click mark, and calculate the ratio of the actual click rate to the average value of the estimated click rate;

[0171] A training module 13 is used to use the ratio as a regression label and perform batch gradient descent to obtain a click rate influence factor model corresponding to the product and the user, wherein the click rate influence factor model obtains the click rate influence factor in response to the historical exposure times and the historical click times;

[0172] The recall module 14 is used to select reference products from the recall list products according to a preset target ratio;

[0173] A factor calculation module 15, used to obtain a click rate influence factor of each reference product from a click rate influence factor model;

[0174] The score calculation module 16 is used to calculate the corresponding reference score according to the price, estimated click rate, estimated conversion rate and click rate influencing factor of each reference product;

[0175] Configuration module 17, used for presetting a filter set, wherein the filter set includes the products recommended to the user recently, or the reference products with the highest reference scores;

[0176] A screening module 18 is used to cyclically calculate the feature score of each reference product relative to the products in the screening set according to the features of the products, and add the reference products with high scores to the screening set, wherein the feature score is the weighted difference between the reference score and the similarity between the reference product and the products in the screening set;

[0177] The result module 19 is used to prioritize the products with high feature scores into a recommendation set.

[0178] In an exemplary embodiment, the training module 13 is further used to obtain a click rate influencing factor model corresponding to the product and the user by solving the following formula:

[0179] +

[0180] ,

[0181] ,

[0182]

[0183] in, is the click impact factor, a, b, is the model parameter obtained by batch gradient descent, N is the total data volume of historical exposure times or historical click times recorded in a period, C is the click impact, E is the exposure impact, is the i-th data in the historical exposure times recorded in a period. is the i-th data in the historical click count recorded in a period, Y is the ratio, is the actual click-through rate, Estimate the average click rate; use C and E as features and Y as regression labels to fit C, E and The mapping relationship.

[0184] In an exemplary embodiment, the recall module 14 is further configured to:

[0185] Calculate the Euclidean distance of each recall in the recall list product set. The formula is as follows:

[0186]

[0187] The estimated recall ratios of each route after recalling one product from route i are as follows:

[0188]

[0189]

[0190]

[0191] in, is the recall ratio of the current i-th path, is the number of recalled products on the current i-th route, is the total number of current recalls; is the recall ratio after the i-th route recalls one product, is the recall ratio of the jth route after the ith route recalls one product, is the number of recalled products in route j except route i; is the Euclidean distance of the ith path, is the target ratio of the jth road;

[0192] Recall one product from the route with the minimum Euclidean distance in a loop until the preset total number of recalls is reached, including: when the number of products in the jth route is 0, update the target ratio of other routes according to the following formula, and recall one product from other routes with the minimum Euclidean distance in a loop:

[0193]

[0194] in, is the target ratio of the i-th route, and n is the total number of routes whose current number of products is not 0.

[0195] In an exemplary embodiment, the score calculation module 16 calculates the reference score by the following formula:

[0196]

[0197] Among them, Score is the reference score, is the click impact factor, is the estimated click-through rate for the corresponding product. is the estimated conversion rate of the corresponding product, and Price is the price of the corresponding product.

[0198] In an exemplary embodiment, the screening module 18 calculates the above feature score by the following formula:

[0199]

[0200]

[0201] in, is the feature score of the i-th reference product, is the relevance weight, represents the click impact factor of the i-th reference product, S is the current filter set, is the similarity between the i-th reference product and the j-th product in the screening set, The interval between the i-th reference product and the j-th product in the screening set, ; is the title edit distance between the i-th reference product and the j-th product in the filter set, is the third-level catalog matching value between the i-th reference product and the j-th product in the filter set, is the secondary catalog matching value between the i-th reference product and the j-th product in the filter set, is the brand matching value between the i-th reference product and the j-th product in the filter set, is the code matching value between the i-th reference product and the j-th product in the filter set.

[0202] Figure 4 is a schematic diagram of the hardware structure of an electronic device for executing the data processing method of the recommendation system provided by an embodiment of the present invention, such as Figure 4 As shown, the device includes:

[0203] One or more processors 410 and memory 420, Figure 4 A processor 410 is taken as an example.

[0204] The device for executing the data processing method of the recommendation system may further include: an input device 430 and an output device 440 .

[0205] The processor 410, the memory 420, the input device 430 and the output device 440 may be connected via a bus or other means. Figure 4 The example of connecting through bus is taken in the following.

[0206] The memory 420 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the data processing method of the recommendation system in the embodiment of the present invention (for example, the attached Figure 3The processor 410 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 420, that is, the data processing method of the recommendation system of the above method embodiment is implemented.

[0207] The memory 420 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the data processing device of the recommendation system, etc. In addition, the memory 420 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 420 may optionally include a memory remotely arranged relative to the processor 410, and these remote memories may be connected to the data processing device of the recommendation system via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0208] The input device 430 can receive input digital or character information and generate key signal input related to user settings and function control of the data processing device of the recommendation system. The output device 440 can include display devices such as a display screen.

[0209] The one or more modules are stored in the memory 420 , and when executed by the one or more processors 410 , the data processing method of the recommendation system in any of the above method embodiments is executed.

[0210] The above product can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0211] The electronic device of the embodiment of the present invention exists in various forms, including but not limited to:

[0212] (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (such as iPhone), multimedia phones, functional phones, and low-end phones.

[0213] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing capabilities, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0214] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0215] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0216] (5) Other electronic devices with data interaction functions.

[0217] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place or distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0218] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0219] It should be noted that, in this article, relational terms such as "first" and "second" 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 "include", "comprise" 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, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0220] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data processing method for a recommendation system, characterized in that: include: Acquire multiple periodic statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click rates, and click flags corresponding to the product and the user; Divide the sample data into multiple data groups according to the historical exposure times and the historical click times. In the data groups where the sample data volume is greater than or equal to the threshold, count the actual click rate by the click mark, and calculate the ratio of the actual click rate to the average value of the estimated click rate. Using the ratio as a regression label, batch gradient descent is performed to obtain a click rate influence factor model corresponding to the product and the user, wherein the click rate influence factor model obtains a click rate influence factor in response to historical exposure times and historical click times; Select reference products from the recalled list products according to the preset target ratio; The click rate influence factor of each reference product is obtained from the click rate influence factor model; Calculate the corresponding reference score based on the price, estimated click-through rate, estimated conversion rate and click-through rate impact factor of each reference product; A preset filter set, wherein the filter set includes products that are recently recommended to the user, or reference products with the highest reference scores; According to the characteristics of the products, the feature score of each reference product relative to the products in the screening set is calculated cyclically, and the reference products with high scores are added to the screening set. The feature score is the weighted difference between the reference score and the similarity of the reference product and the products in the screening set. Prioritize products with high feature scores to form a recommendation set; The step of selecting reference products from the recalled list products according to a preset target ratio includes: Calculate the Euclidean distance of each recall in the recall list product set. The formula is as follows: ， The estimated recall ratios of each route after recalling one product from route i are as follows: ， ， in, is the recall ratio of the current i-th path, is the number of products currently recalled on the ith route, and T is the total number of products currently recalled; is the recall ratio after the i-th route recalls one product, is the recall ratio of the jth route after the ith route recalls one product, is the number of recalled products in route j except route i; is the Euclidean distance of the ith path, is the target ratio of the jth road; Recall one product from the route with the minimum Euclidean distance in a loop until the preset total number of recalls is reached, including: when the number of products in the jth route is 0, update the target ratio of other routes according to the following formula, and recall one product from other routes with the minimum Euclidean distance in a loop: ， in, is the target ratio of the i-th route, and n is the total number of routes whose current number of products is not 0.

2. The data processing method of the recommendation system according to claim 1, characterized in that: The ratio is used as a regression label, and batch gradient descent is used to solve the click rate influencing factor model corresponding to the product and the user, including calculating through the formula: ， in, is the click impact factor, a, b, is the model parameter obtained by batch gradient descent, N is the total data volume of historical exposure times or historical click times recorded in a period, C is the click impact, E is the exposure impact, is the i-th data in the historical exposure times recorded in a period, is the i-th data in the historical click count recorded in a period, Y is the ratio, is the actual click-through rate, Estimate the average click rate; use C and E as features and Y as regression labels to fit C, E and The mapping relationship.

3. The data processing method of the recommendation system according to claim 1, characterized in that: The corresponding reference score is calculated by the price, estimated click rate, estimated conversion rate and click rate influencing factor of each reference product, including: ， Among them, Score is the reference score, is the click impact factor, is the estimated click-through rate for the corresponding product. is the estimated conversion rate of the corresponding product, and Price is the price of the corresponding product.

4. The data processing method of the recommendation system according to claim 1, characterized in that: According to the characteristics of the products, the characteristic score of each reference product relative to the products in the screening set is cyclically calculated, and the reference products with high scores are added to the screening set. The characteristic score is the weighted difference between the reference score and the similarity between the reference product and the products in the screening set, including: ， in, is the feature score of the i-th reference product, is the relevance weight, represents the click impact factor of the i-th reference product, S is the current filter set, is the similarity between the i-th reference product and the j-th product in the screening set, Characterize the attenuation factor, represents the interval between the i-th reference product and the j-th product in the screening set, is the interval attenuation factor; is the title edit distance between the i-th reference product and the j-th product in the filter set, is the third-level catalog matching value between the i-th reference product and the j-th product in the filter set, is the secondary catalog matching value between the i-th reference product and the j-th product in the filter set, is the brand matching value between the i-th reference product and the j-th product in the filter set, is the code matching value between the i-th reference product and the j-th product in the filter set.

5. A data processing device for a recommendation system, used to implement the method according to any one of claims 1 or 3, characterized in that: include: An acquisition module, used to acquire multiple periodic statistical sample data, wherein the sample data includes a combination of historical exposure times, historical click times, estimated click rates, and click flags corresponding to products and users; A statistical module is used to divide the sample data into multiple data groups according to the historical exposure times and the historical click times. In the data groups where the sample data volume is greater than or equal to the threshold, the actual click rate is counted by the click mark, and the ratio of the actual click rate to the average value of the estimated click rate is calculated; A training module, configured to use the ratio as a regression label and perform batch gradient descent to obtain a click rate influence factor model corresponding to the product and the user, wherein the click rate influence factor model obtains the click rate influence factor in response to the historical exposure times and the historical click times; The recall module is used to select reference products from the recall list products according to a preset target ratio; A factor calculation module is used to obtain the click rate influence factor of each reference product from the click rate influence factor model; A rating calculation module, used to calculate the corresponding reference rating through the price, estimated click rate, estimated conversion rate and click rate influencing factor of each reference product; A configuration module, used to preset a filter set, wherein the filter set includes products recently recommended to the user, or reference products with the highest reference scores; A screening module, for cyclically calculating the feature score of each reference product relative to the products in the screening set according to the features of the products, and adding the reference products with high scores to the screening set, wherein the feature score is the weighted difference between the reference score and the similarity between the reference product and the products in the screening set; The result module is used to prioritize the products with high feature scores into the recommendation set.

6. The data processing device of the recommendation system according to claim 5, characterized in that: The training module is also used to obtain a click rate influencing factor model corresponding to the product and the user by solving the following formula: ， in, is the click impact factor, a, b, is the model parameter obtained by batch gradient descent, N is the total data volume of historical exposure times or historical click times recorded in a period, C is the click impact, E is the exposure impact, is the i-th data in the historical exposure times recorded in a period, is the i-th data in the historical click count recorded in a period, Y is the ratio, is the actual click-through rate, Estimate the average click rate; use C and E as features and Y as regression labels to fit C, E and The mapping relationship.

7. The data processing device of the recommendation system according to claim 5, characterized in that: The screening module calculates the feature score by the following formula: ， in, is the feature score of the i-th reference product, is the relevance weight, represents the click impact factor of the i-th reference product, S is the current filter set, is the similarity between the i-th reference product and the j-th product in the screening set, Characterize the attenuation factor, represents the interval between the i-th reference product and the j-th product in the screening set, is the interval attenuation factor; is the title edit distance between the i-th reference product and the j-th product in the filter set, is the third-level catalog matching value between the i-th reference product and the j-th product in the filter set, is the secondary catalog matching value between the i-th reference product and the j-th product in the filter set, is the brand matching value between the i-th reference product and the j-th product in the filter set, is the code matching value between the i-th reference product and the j-th product in the filter set.

8. A computer-readable storage medium, characterized in that: The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the steps of the data processing method of the recommendation system according to any one of claims 1 to 4.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the data processing method of the recommendation system according to any one of claims 1 to 4 are implemented.

Images