Behavior data processing method, device, equipment and storage medium

Abstract

The present invention discloses a behavior data processing method, apparatus, equipment and storage medium. When a data processing request is detected, user behavior data and current webpage product association data are obtained according to the data processing request; the user behavior data and the current webpage product association data are spliced ​​to generate behavior features; the corresponding current interest vector is determined according to the behavior features; the current interest vector is input into a hidden layer in a preset deep neural network to obtain the current product recommendation degree of different products, and the hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree; the target product is selected based on the current product recommendation degree and the target product is displayed, which can fully mine the user's real-time behavior data, improve the ability to capture the user's real-time interest, and realize personalized product recommendation.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a behavior data processing method, device, equipment and storage medium. Background Art

[0002] At the recall level, recall currently uses multiple recall methods such as search-based recall and target group index (TGI) indicator recall. For specific scenarios, data is retrieved from the search server (Elastic Search, ES) database through some specified filtering conditions (such as: primary category, secondary category, item tag). Through recall filtering, the candidate set is compressed from millions or hundreds of thousands to hundreds of levels, thereby reducing the pressure on the sorting service.

[0003] At the ranking level, the main models used for sorting include wide&deep and deepfm. The wide&deep model combines the memory capacity of the logistic regression (LR) model and the generalization ability of the deep neural network (DNN) model, but still requires a large amount of manual feature combination at the wide layer. The DeepFm model uses the factorization machine (FM) model at the wide layer, which can achieve automatic feature cross-pollination, but it takes more time than the wide&deep model in online services.

[0004] During the recall process, only the user's historical preferences are taken into account, and no preference mining and analysis of the user's real-time behavior is conducted, resulting in the recall lacking real-time characteristics. During sorting, the granularity of user behavior mining is relatively coarse, with clicks, downloads, and other behaviors accumulated over multiple days. This lacks the ability to capture the user's real-time interests, and the ranking of items cannot be adjusted based on real-time user feedback. Furthermore, during recall and sorting, there is a lack of capturing the real-time performance of items. For example, during the Double Eleven shopping festival, shopping products will generate a large amount of demand in the short term or in the evening and at noon of a single day. It is difficult to capture such changes in data distribution at the model level, resulting in the model being unable to meet the real-time needs of users. Summary of the Invention

[0005] The main purpose of the present invention is to provide a behavioral data processing method, device, equipment and storage medium, aiming to solve the technical problems in the existing technology that recall lacks real-time performance, sorting lacks the capture of users' real-time interests, and data cannot meet users' real-time needs.

[0006] To achieve the above object, the present invention provides a behavior data processing method, which includes the following steps:

[0007] When a data processing request is detected, obtaining user behavior data and current webpage product association data according to the data processing request;

[0008] Combine the user behavior data and the current webpage product association data to generate a behavior feature;

[0009] Determining a corresponding current interest vector according to the behavior characteristics;

[0010] Inputting the current interest vector into a hidden layer in a preset deep neural network to obtain the current product recommendation degree of different products, wherein the hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree;

[0011] A target product is selected based on the current product recommendation degree, and the target product is displayed.

[0012] Optionally, determining the corresponding current interest vector according to the behavior feature includes:

[0013] Obtaining user click data, search data, and download data from the behavioral characteristics;

[0014] A current interest vector is determined according to the click data, the search data, and the download data.

[0015] Optionally, determining a current interest vector according to the click data, the search data, and the download data includes:

[0016] Determining product data that the user is interested in based on the click data, the search data, and the download data;

[0017] The feature vector corresponding to the product data that the user is interested in is used as the current interest vector.

[0018] Optionally, determining the product data of interest to the user based on the click data, the search data, and the download data includes:

[0019] Obtaining the number of clicks on each product according to the click data, and comparing each click number with a preset number of clicks to obtain a first comparison result;

[0020] Obtaining a keyword search count for each product based on the search data, and comparing the keyword search count for each product with a preset search count to obtain a second comparison result;

[0021] Obtaining a total data download amount for each product based on the downloaded data, and comparing the total data download amount with a preset total data amount to obtain a third comparison result;

[0022] The commodity data that the user is interested in is determined according to the first comparison result, the second comparison result, and the third comparison result.

[0023] Optionally, selecting a target product based on the current product recommendation degree and displaying the target product includes:

[0024] Determining target products that the user is interested in from the current product recommendation degree, and obtaining an interest value for each target product based on the current product recommendation degree;

[0025] Sort the target products according to their interest values ​​and generate a sorting result;

[0026] The target products are displayed according to the ranking results.

[0027] Optionally, sorting the target products according to their interest values ​​and generating sorting results includes:

[0028] Get the purchase quantity and click quantity of each target product;

[0029] Obtaining a conversion rate of each target product based on the product purchase quantity and the click quantity;

[0030] Determine the predicted score of each target product according to the interest value of each target product;

[0031] Calculating a final interest value of each target product based on the predicted score and the conversion rate;

[0032] The target products are sorted according to the final interest value, and a sorting result is generated.

[0033] Optionally, the final interest value of each target product is calculated according to the predicted score and the conversion rate using the following formula, including:

[0034] Iv=α*A+β*B

[0035] Among them, Iv is the final value of interest, A is the predicted score, B is the conversion rate, α is the preset score weight ratio, and β is the preset conversion rate weight ratio.

[0036] In addition, to achieve the above-mentioned purpose, the present invention further proposes a behavior data processing device, the behavior data processing device comprising:

[0037] A data acquisition module is used to acquire user behavior data and current webpage product association data according to the data processing request when a data processing request is detected;

[0038] A feature generation module, configured to combine the user behavior data with the current webpage product association data to generate a behavior feature;

[0039] A vector acquisition module, configured to determine a corresponding current interest vector according to the behavior characteristics;

[0040] An input module, configured to input the current interest vector into a hidden layer in a preset deep neural network to obtain current product recommendation degrees for different products, wherein the hidden layer in the preset deep neural network reflects a mapping relationship between the interest vector and the product recommendation degree;

[0041] The recommendation module is used to select a target product based on the current product recommendation degree and display the target product.

[0042] In addition, to achieve the above-mentioned purpose, the present invention also proposes a behavior data processing device, which includes: a memory, a processor, and a behavior data processing program stored on the memory and executable on the processor, wherein the behavior data processing program is configured to implement the steps of the behavior data processing method described above.

[0043] In addition, to achieve the above-mentioned purpose, the present invention further proposes a storage medium, on which a behavior data processing program is stored. When the behavior data processing program is executed by a processor, the steps of the behavior data processing method described above are implemented.

[0044] The behavioral data processing method proposed in the present invention obtains user behavior data and current webpage product association data according to the data processing request when a data processing request is detected; splices the user behavior data and the current webpage product association data to generate behavioral features; determines the corresponding current interest vector according to the behavioral features; inputs the current interest vector into the hidden layer of a preset deep neural network to obtain the current product recommendation degree of different products, and the hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree; selects the target product based on the current product recommendation degree and displays the target product, which can fully mine the user's real-time behavior data, and through the collection of the user's real-time behavior and product real-time data, mines the user's category preference and applies it to the recall and sorting levels; improves the ability to capture the user's real-time interest and realizes personalized product recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present invention;

[0046] Figure 2 This is a flow chart of a first embodiment of a behavior data processing method according to the present invention;

[0047] Figure 3 This is a flow chart of a second embodiment of the behavior data processing method of the present invention;

[0048] Figure 4 This is a flow chart of a third embodiment of the behavior data processing method of the present invention;

[0049] Figure 5 This is a flow chart of a fourth embodiment of the behavior data processing method of the present invention;

[0050] Figure 6 This is a flow chart of a fifth embodiment of the behavior data processing method of the present invention;

[0051] Figure 7 This is a functional module diagram of the first embodiment of the behavior data processing device of the present invention.

[0052] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0053] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present invention.

[0055] like Figure 1 As shown, the device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a button, and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), such as a disk storage. The memory 1005 may optionally be a storage device independent of the aforementioned processor 1001.

[0056] Those skilled in the art will understand that Figure 1 The device structure shown in the figure does not constitute a limitation of the device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0057] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and a behavior data processing program.

[0058] exist Figure 1 In the device shown, the network interface 1004 is mainly used to connect to the external network and communicate data with other network devices; the user interface 1003 is mainly used to connect to the user device and communicate data with the device; the device of the present invention calls the behavior data processing program stored in the memory 1005 through the processor 1001, and executes the implementation method of behavior data processing provided by the embodiment of the present invention.

[0059] Based on the above hardware structure, an embodiment of the behavior data processing method of the present invention is proposed.

[0060] Reference Figure 2 , Figure 2 Schematic diagram of the flow of the first embodiment of the behavior data processing method of the present invention.

[0061] In a first embodiment, the behavior data processing method includes the following steps:

[0062] Step S10: When a data processing request is detected, user behavior data and current webpage product association data are obtained according to the data processing request.

[0063] It should be noted that the data processing request is a data processing request generated by the user through actions such as clicking or touching. Through the data processing request, user behavior data and current web page product association data can be obtained. The user behavior data is the real-time data of the user browsing products; the current web page product association data is the product-related data corresponding to the web page currently browsed by the user.

[0064] Step S20: Combine the user behavior data and the current webpage product association data to generate behavior features.

[0065] It is understandable that the user behavior data and the current webpage product association data can generally be spliced ​​together to form behavioral features for use in model prediction.

[0066] Step S30: Determine the corresponding current interest vector according to the behavior characteristics.

[0067] It should be understood that different behavioral characteristics correspond to different interest vectors. Determining the current interest vector of the current user based on the behavioral characteristics helps to improve the pertinence of subsequent product recommendations.

[0068] Step S40: Input the current interest vector into a hidden layer in a preset deep neural network to obtain the current product recommendation degree of different products. The hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree.

[0069] It can be understood that the preset deep neural network is a pre-set deep neural network for obtaining product recommendation degrees. The hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree. It can be a relationship obtained through a large amount of data training, or a mapping relationship set by technical personnel based on daily operating experience, or a mapping relationship determined by other means. This embodiment does not limit this.

[0070] Step S50: Select a target product based on the current product recommendation degree, and display the target product.

[0071] It should be understood that, based on the current product recommendation degree, several products can be selected as target products and the target products can be displayed. The display method can be to display them in a certain order according to the size of the product recommendation degree, or to display some target products in an excellent and concentrated manner according to the product recommendation degree, or to display them in other ways, which is not limited in this embodiment.

[0072] This embodiment adopts the above scheme, by acquiring user behavior data and current webpage commodity association data according to the data processing request when a data processing request is detected; splicing the user behavior data and the current webpage commodity association data to generate behavior features; determining the corresponding current interest vector according to the behavior features; inputting the current interest vector into the hidden layer of the preset deep neural network to obtain the current commodity recommendation degree of different commodities, and the hidden layer of the preset deep neural network reflects the mapping relationship between the interest vector and the commodity recommendation degree; selecting the target commodity based on the current commodity recommendation degree, and displaying the target commodity, which can fully mine the user's real-time behavior data, and through the collection of the user's real-time behavior and product real-time data, mine the user's category preference and apply it to the recall and sorting levels; improving the ability to capture the user's real-time interest and realizing personalized commodity recommendation.

[0073] Furthermore, Figure 3 This is a flow chart of the second embodiment of the behavior data processing method of the present invention. Figure 3 As shown, a second embodiment of the behavior data processing method of the present invention is proposed based on the first embodiment. In this embodiment, step S30 includes the following steps:

[0074] Step S31: Obtain the user's click data, search data, and download data from the behavior characteristics.

[0075] It should be noted that the behavioral characteristics include the user's click data, search data and download data; the click data is real-time click data generated by the user's click, the search data is real-time click data generated by the user's search operation, and the download data is data downloaded in response to the user's download request.

[0076] Step S32: Determine a current interest vector based on the click data, the search data, and the download data.

[0077] It can be understood that the current interest vector can be determined through the click data, the search data and the download data. Generally, these data are analyzed to obtain the interest vector corresponding to these data, that is, the user's real-time interest can be reflected through the click data, the search data and the download data.

[0078] Furthermore, the step S32 further includes the following steps:

[0079] Determining product data that the user is interested in based on the click data, the search data, and the download data;

[0080] The feature vector corresponding to the product data that the user is interested in is used as the current interest vector.

[0081] It should be understood that the click data, the search data and the download data can be analyzed to determine the products that the user is interested in, and then determine the product data that the user is interested in; and then obtain the feature vector corresponding to the product data that the user is interested in, and use the feature vector as the current interest vector.

[0082] Furthermore, the step of determining the product data that the user is interested in based on the click data, the search data, and the download data specifically includes the following steps:

[0083] Obtaining the number of clicks on each product according to the click data, and comparing each click number with a preset number of clicks to obtain a first comparison result;

[0084] Obtaining a keyword search count for each product based on the search data, and comparing the keyword search count for each product with a preset search count to obtain a second comparison result;

[0085] Obtaining a total data download amount for each product based on the downloaded data, and comparing the total data download amount with a preset total data amount to obtain a third comparison result;

[0086] The commodity data that the user is interested in is determined according to the first comparison result, the second comparison result, and the third comparison result.

[0087] It should be noted that the preset number of clicks is a preset click threshold, the preset number of searches is a preset search threshold, and the preset total amount of data is a preset total amount of data threshold. The preset number of clicks, the preset number of searches and the preset total amount of data can be thresholds obtained in advance through a large amount of experimental training, or thresholds determined by technical personnel based on daily operating experience, or thresholds determined by other means. This embodiment does not impose any restrictions on this; the number of keyword searches is the number of times a keyword appears in the search process obtained by analyzing the search data, and the total amount of data loaded is the cumulative total amount of data downloaded for each product that performs related download operations. The first comparison result, the second comparison result and the third comparison result obtained after comparison can determine the product data corresponding to the product that the user is more interested in.

[0088] Furthermore, the step of determining the product data that the user is interested in based on the first comparison result, the second comparison result, and the third comparison result includes:

[0089] The product with a click count greater than the preset click count in the first comparison result is used as the first product of interest;

[0090] The product with the keyword search times greater than the preset search times in the second comparison result is used as the second product of interest;

[0091] The product with the total amount of data downloaded being greater than the preset total amount of data in the third comparison result is used as a third product of interest;

[0092] The data corresponding to the first interest product, the second interest product, and the third interest product are respectively used as product data of interest to the user.

[0093] It should be understood that after comparison, commodities with values ​​greater than various preset thresholds may be regarded as commodities of interest, and the data corresponding to the corresponding commodities may be regarded as commodity data of interest to the user.

[0094] Through the above scheme, this embodiment obtains the user's click data, search data and download data from the behavioral characteristics; determines the current interest vector based on the click data, search data and download data, and can specifically determine the products that the user is interested in. By analyzing the user's real-time behavioral data of clicks, searches and downloads, it can fully mine the user's real-time behavioral data. By collecting the user's real-time behavior and real-time product data, the user's category preferences are mined and applied to the recall and sorting levels. The ability to capture the user's real-time interests is improved, and personalized product recommendations are achieved.

[0095] Furthermore, Figure 4 FIG. 1 is a flow chart of a third embodiment of the behavior data processing method of the present invention. Figure 4 As shown, a third embodiment of the behavior data processing method of the present invention is proposed based on the first embodiment. In this embodiment, before step S40, the data processing method further includes the following steps:

[0096] Step S401: Acquire the user's historical behavior data and historical behavior results, and use the historical behavior data and historical behavior results as training data.

[0097] It should be noted that the historical behavior data and historical behavior results are the behavior data and corresponding behavior results generated by the user during the previous web browsing process. By obtaining the historical behavior data and historical behavior results of the user, the historical behavior data and historical behavior results can be used as data for ready training.

[0098] Step S402: Train the initial deep neural network according to the data to be trained to obtain a preset deep neural network.

[0099] It can be understood that the initial deep neural network is a pre-set deep neural network for training user historical behavior data. The initial deep neural network can be trained and learned through the data to be trained to obtain a preset deep neural network.

[0100] Furthermore, step S402 includes the following steps:

[0101] Obtaining candidate feature data of preset candidate products;

[0102] Associating the candidate feature data with the data to be trained;

[0103] Performing weighted summation on the associated candidate feature data and the data to be trained to generate data to be embedded;

[0104] An initial deep neural network is trained according to the data to be trained and the data to be embedded to obtain a preset deep neural network.

[0105] It should be noted that the preset candidate products are pre-set candidate products selected from the database corresponding to the search server (Elastic Search, ES). Of course, they can also be products customized by technicians, or they can be other types of candidate products. This embodiment does not limit this. By obtaining the candidate feature data of the preset candidate products and associating the candidate feature data with the data to be trained, generally performing embedding management and performing weighted summation on the associated data, the corresponding data to be embedded can be generated, thereby training the initial deep neural network according to the selected data and the data to be embedded to obtain the preset deep neural network.

[0106] Furthermore, the step of training the initial deep neural network according to the data to be trained and the data to be embedded to obtain a preset deep neural network includes the following steps:

[0107] Training the initial deep neural network with the data to be trained to obtain an initial training result;

[0108] Inputting the data to be embedded into the hidden layer of the initial deep neural network to obtain an embedding training result;

[0109] Inputting the data to be embedded into a preset fusion shallow model to obtain a fusion training result;

[0110] The initial deep neural network is trained based on the initial training result, the embedding training result and the fusion training result to obtain a preset deep neural network.

[0111] It can be understood that after obtaining the data to be trained, a preset deep neural network can be obtained based on the initial training results obtained from the training and combined with the embedding training results obtained by training the hidden layer of the initial deep neural network based on the data to be embedded. Generally, the data to be trained and the embedded data are connected and input into the hidden layer, and jointly trained with the preset fusion shallow model, that is, the initial deep neural network is trained based on the initial training results, the embedding training results and the fusion training results, and then the preset deep neural network is obtained.

[0112] Furthermore, the step of obtaining candidate feature data of preset candidate commodities includes the following steps:

[0113] The category feature data and context feature data of the preset candidate commodity are obtained, and the category feature data and the context feature data are used as candidate feature data.

[0114] It should be understood that the preset candidate products correspond to category feature data and context feature data. The category feature data are some specified screening conditions, such as primary categories, secondary categories, item labels, etc. Different category feature data correspond to users' category preferences at different levels; the context feature data are the context information corresponding to the preset candidate products. By using the category feature data and the context feature data as candidate feature data, the initial deep neural network can be trained more comprehensively to ensure the data validity and accuracy of the preset deep neural network.

[0115] Furthermore, the step of performing weighted summation on the associated candidate feature data and the data to be trained to generate data to be embedded includes the following steps:

[0116] Performing weighted summation on the associated candidate feature data and the to-be-trained data according to a preset weight coefficient to obtain a summation result, and writing the summation result into a preset distributed file system;

[0117] Obtain user download behavior information recorded in a preset database warehouse tool table, and obtain log information in the preset distributed file system;

[0118] The user download behavior information and the log information are combined to generate data to be embedded.

[0119] It should be noted that the preset weight coefficient is a preset weight ratio, and the preset weight coefficient can be used to perform weighted summation on the associated candidate feature data and the data to be trained to obtain a summation result; the preset distributed file system is a preset system for storing distributed files such as various types of log information, and the preset database warehouse tool table is a preset data table for the storage warehouse tool. User download behavior information can be obtained from the preset database warehouse tool table, and log information can be obtained from the preset distributed file system; by merging the log information and the user download behavior information, the data to be embedded for neural network model training can be obtained.

[0120] Through the above scheme, this embodiment obtains the user's historical behavior data and historical behavior results, and uses the historical behavior data and historical behavior results as the data to be trained; the initial deep neural network is trained according to the data to be trained to obtain a preset deep neural network. The initial deep neural network can be trained based on the historical data to explore the user's category preferences and apply them to the recall and sorting levels; the ability to capture the user's real-time interests is improved, and personalized product recommendations are achieved.

[0121] Furthermore, Figure 5This is a flow chart of the fourth embodiment of the behavior data processing method of the present invention. Figure 5 As shown, a fourth embodiment of the behavior data processing method of the present invention is proposed based on the first embodiment. In this embodiment, step S50 includes the following steps:

[0122] Step S51: determining target products that the user is interested in based on the current product recommendation degree, and obtaining an interest value of each target product based on the current product recommendation degree.

[0123] It should be noted that the target product that the user is interested in can be determined based on the current product recommendation degree, and the interest value corresponding to each target product can be obtained based on the current product recommendation degree.

[0124] Step S52: sort the target products according to their interest values ​​and generate a sorting result.

[0125] It should be understood that by sorting the target products according to their interest values, the generated sorting results can be used to intuitively understand the user's interest level in each product, which helps to display products to users in a targeted manner.

[0126] Furthermore, the step S52 includes the following steps:

[0127] Get the purchase quantity and click quantity of each target product;

[0128] Obtaining a conversion rate of each target product based on the product purchase quantity and the click quantity;

[0129] Determine the predicted score of each target product according to the interest value of each target product;

[0130] Calculating a final interest value of each target product based on the predicted score and the conversion rate;

[0131] The target products are sorted according to the final interest value, and a sorting result is generated.

[0132] It should be noted that the purchase quantity and click quantity of the goods are generally determined by retrieving the background purchase records and click records corresponding to each target product. Of course, they can also be obtained by tracing back to the source of specific information, or by other methods. This embodiment does not limit this. The conversion rate of the corresponding goods can be determined by the purchase quantity and click quantity of the goods. Generally, the ratio of the purchase quantity of goods to the click quantity is the conversion rate of the corresponding goods. The predicted score of each target product can be determined by the interest value of each target product, and then the final interest value of each target product is calculated based on the predicted score and the conversion rate, and the target products are sorted according to the final interest value to generate a sorting result.

[0133] Furthermore, the final interest value of each target product is calculated according to the predicted score and the conversion rate using the following formula, including:

[0134] Iv=α*A+β*B

[0135] Among them, Iv is the final value of interest, A is the predicted score, B is the conversion rate, α is the preset score weight ratio, and β is the preset conversion rate weight ratio.

[0136] It is understandable that in the process of online services, the final interest value of an item is determined by combining the estimated score, that is, the predicted score and the conversion rate.

[0137] Step S53: Display the target products according to the ranking results.

[0138] It is understandable that the target products are displayed based on the ranking results, generally on a display page of a terminal device currently used by the user, or through projection or other means, which is not limited in this embodiment.

[0139] Through the above scheme, this embodiment determines the target products that the user is interested in based on the current product recommendation degree, and obtains the interest value of each target product based on the current product recommendation degree; sorts each target product according to the interest value of each target product, and generates a sorting result; displays the target products according to the sorting result, and can sort and display the target products according to the interest value of the target products, thereby improving the ability to capture the user's real-time interests and realizing personalized product recommendations.

[0140] Furthermore, Figure 6 FIG. 5 is a flow chart of a fifth embodiment of a behavior data processing method according to the present invention. Figure 6 As shown, a fifth embodiment of the behavior data processing method of the present invention is proposed based on the first embodiment. In this embodiment, after step S50, the behavior data processing method further includes the following steps:

[0141] Step S60: After receiving the initial address clicked and downloaded by the user, generate a secondary category address of the product according to the initial address, and generate a recall sorting condition according to the secondary category address.

[0142] It should be noted that the initial address is the product address corresponding to the user's real-time click and download behavior. The secondary category address can be generated through the initial address, that is, after the secondary category of the product is sought, the secondary category that the user may be interested in can be obtained, and then the address corresponding to the secondary category can be determined, thereby generating recall sorting conditions based on the secondary category address.

[0143] Step S70: query the search server for recall data according to the recall sorting condition.

[0144] It is understandable that the full amount of data of the candidate products is generally stored in the search server, and the corresponding recall data can be found in the search server through the recall sorting condition.

[0145] Step S80: Adjust the currently displayed products according to the recall data.

[0146] It should be understood that the order of currently displayed commodities may be adjusted according to the recall data.

[0147] Furthermore, the step S80 includes the following steps:

[0148] Obtaining the exposure of the currently displayed product in the recall data;

[0149] Sort the products according to the exposure, and obtain the product with the lowest exposure ranking;

[0150] Increase the number of recommendations for the lowest-ranked product based on the preset number of exposures.

[0151] It should be noted that the recall data corresponds to the exposure of different products. The products can be sorted according to the exposure, thereby increasing the number of recommendations for the product with the lowest exposure, making the exposure data diversified and balanced.

[0152] Furthermore, the step of querying the recall data in the search server according to the recall sorting condition includes the following steps:

[0153] When a new recommended product is detected, the recall sorting condition is updated according to the recommended product and the recall data;

[0154] The corresponding recall data is re-queried in the search server according to the updated recall sorting condition.

[0155] It is understandable that when newly added recommended products are detected, the recall sorting conditions can be updated according to the recommended products and the recall data, thereby obtaining new recall data.

[0156] In a specific implementation, the data queried from the search server will correspond to a query score. When the online model prediction times out, the recall score can be used as a degradation strategy to sort by the recall score to reduce the risk of sorting failure caused by timeout. The real-time data includes the exposure and conversion data of each item. At the recall layer, frequency control can be performed based on the exposure of the item to increase the number of recommendations for low-exposure items, making the exposure data diverse and balanced.

[0157] Furthermore, when a new recommended product is detected, the step updates the recall sorting conditions according to the recommended product and the recall data, including:

[0158] When a new recommended product is detected, the new recommended product is added to the product recommendation list, thereby updating the recall sorting condition;

[0159] or,

[0160] When newly added recommended products are detected, the newly added recommended products are added to the product recommendation list, and the products with the lowest user interest are deleted from the product recommendation list, thereby updating the recall sorting conditions.

[0161] It is understandable that by updating the recall sorting conditions or directly deleting the products with the lowest user interest from the product recommendation list, the real-time nature of the recall sorting conditions can be ensured, the real-time interests of users can be effectively followed up, and the real-time interests of users can be captured in a timely manner.

[0162] Through the above-mentioned scheme, this embodiment generates a secondary category address of the product based on the initial address after receiving the initial address clicked and downloaded by the user, and generates a recall sorting condition based on the secondary category address; queries recall data in the search server according to the recall sorting condition; and adjusts the currently displayed products according to the recall data, thereby ensuring the real-time nature of the recall sorting condition, effectively following up the real-time interests of the user, and realizing timely capture of the real-time interests of the user.

[0163] Based on the above behavior data processing method, the present invention further provides a behavior data processing device and proposes an embodiment of the behavior data processing device of the present invention.

[0164] Reference Figure 7 , Figure 7 This is a functional module diagram of the first embodiment of the behavior data processing device of the present invention.

[0165] In a first embodiment of the behavior data processing device of the present invention, the behavior data processing device includes:

[0166] The data acquisition module 10 is used to acquire user behavior data and current webpage product association data according to the data processing request when a data processing request is detected.

[0167] The feature generation module 20 is used to combine the user behavior data and the current webpage product association data to generate behavior features.

[0168] The vector acquisition module 30 is configured to determine the corresponding current interest vector according to the behavior characteristics.

[0169] Furthermore, the vector acquisition module 30 includes:

[0170] The feature data extraction module is used to obtain the user's click data, search data and download data from the behavior features.

[0171] The vector determination module is used to determine the current interest vector according to the click data, the search data and the download data.

[0172] Furthermore, the vector determination module includes:

[0173] The product data determination module is used to determine the product data that the user is interested in based on the click data, the search data and the download data.

[0174] The feature vector determination module is used to use the feature vector corresponding to the product data that the user is interested in as the current interest vector.

[0175] Furthermore, the commodity data determination module includes:

[0176] The first comparison module is used to obtain the number of clicks of each product according to the click data, and compare each click number with a preset click number to obtain a first comparison result.

[0177] The second comparison module is configured to obtain the number of keyword searches for each product based on the search data, and compare the number of keyword searches with a preset number of searches to obtain a second comparison result.

[0178] The third comparison module is configured to obtain a total data download amount for each product based on the downloaded data, and compare the total data download amount with a preset data total amount to obtain a third comparison result.

[0179] The comparison result determination module is used to determine the product data that the user is interested in based on the first comparison result, the second comparison result and the third comparison result.

[0180] Furthermore, the comparison result determination module includes:

[0181] The first interest module is configured to take the product with a click count greater than the preset click count in the first comparison result as a first interest product.

[0182] The second interest module is configured to take the product in the second comparison result whose keyword search times are greater than the preset search times as the second interest product.

[0183] The third interest module is configured to take the commodity whose total data download amount is greater than the preset total data amount in the third comparison result as a third interest commodity.

[0184] The product data corresponding module is used to use the data corresponding to the first interest product, the second interest product and the third interest product respectively as the product data of interest to the user.

[0185] The input module 40 is used to input the current interest vector into the hidden layer of the preset deep neural network to obtain the current product recommendation degree of different products. The hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree.

[0186] Accordingly, the behavior data processing device further includes:

[0187] The historical data acquisition module is used to obtain the user's historical behavior data and historical behavior results, and use the historical behavior data and historical behavior results as training data.

[0188] The historical data training module is used to train the initial deep neural network according to the data to be trained to obtain a preset deep neural network.

[0189] Furthermore, the historical data training module includes:

[0190] The candidate feature acquisition module is used to obtain candidate feature data of preset candidate products.

[0191] An associating module is used to associate the candidate feature data with the data to be trained.

[0192] The weighted summation module is used to perform weighted summation on the associated candidate feature data and the data to be trained to generate data to be embedded.

[0193] A joint training module is used to train the initial deep neural network according to the data to be trained and the data to be embedded to obtain a preset deep neural network.

[0194] Furthermore, the joint training module includes:

[0195] The initial training module is used to train the initial deep neural network with the data to be trained to obtain initial training results.

[0196] The embedding training module is used to input the data to be embedded into the hidden layer of the initial deep neural network to obtain an embedding training result.

[0197] The fusion module is used to input the data to be embedded into a preset fusion shallow model to obtain a fusion training result.

[0198] A deep training module is used to train the initial deep neural network based on the initial training result, the embedding training result and the fusion training result to obtain a preset deep neural network.

[0199] Furthermore, the candidate feature acquisition module includes:

[0200] The context acquisition module is used to acquire category feature data and context feature data of preset candidate commodities, and use the category feature data and the context feature data as candidate feature data.

[0201] Furthermore, the weighted summation module includes:

[0202] A summing and writing module is used to perform weighted summation on the associated candidate feature data and the data to be trained according to a preset weight coefficient to obtain a summation result, and write the summation result into a preset distributed file system;

[0203] The log acquisition module is used to obtain user download behavior information recorded in the preset database warehouse tool table and obtain log information in the preset distributed file system.

[0204] The embedding and merging module is used to merge the user download behavior information and the log information to generate data to be embedded.

[0205] The recommendation module 50 is configured to select a target product based on the current product recommendation degree and display the target product.

[0206] Furthermore, the recommendation module 50 includes:

[0207] The interest value acquisition module is used to determine the target products that the user is interested in from the current product recommendation degree, and obtain the interest value of each target product according to the current product recommendation degree.

[0208] The interest ranking module is used to rank the target products according to their interest values ​​and generate ranking results.

[0209] A display module is used to display the target product according to the sorting result.

[0210] Furthermore, the interest sorting module includes:

[0211] The purchase click module is used to obtain the purchase quantity and click quantity of each target product.

[0212] The conversion rate determination module is used to obtain the conversion rate of each target product based on the purchase quantity of the product and the click quantity.

[0213] The prediction score module is used to determine the prediction score of each target product according to the interest value of each target product.

[0214] The interest value calculation module is used to calculate the final interest value of each target product based on the predicted score and the conversion rate.

[0215] The final sorting module is used to sort the target products according to the final interest value and generate a sorting result.

[0216] Furthermore, the behavior data processing device further includes:

[0217] The secondary address generation module is used to generate a secondary category address of the product according to the initial address after receiving the initial address clicked and downloaded by the user, and generate a recall sorting condition according to the secondary category address.

[0218] The sorting query module is used to query the recall data in the search server according to the recall sorting condition.

[0219] The sorting adjustment module is used to adjust the currently displayed products according to the recall data.

[0220] Furthermore, the sorting adjustment module includes:

[0221] An exposure acquisition module, configured to acquire the exposure of the currently displayed product in the recall data;

[0222] An exposure ranking module is used to sort the products according to the exposure level and obtain the product with the lowest exposure ranking;

[0223] Add a recommendation module to increase the number of recommendations for the products with the lowest exposure ranking based on the preset number of exposures.

[0224] Furthermore, the ranking query module includes:

[0225] An update sorting module, configured to update the recall sorting conditions according to the recommended products and the recall data when a new recommended product is detected;

[0226] The re-query module is used to re-query the corresponding recall data in the search server according to the updated recall sorting condition.

[0227] Furthermore, the update sorting module includes:

[0228] The adding and updating module is used to add the newly added recommended products to the product recommendation list when a newly added recommended product is detected, thereby updating the recall sorting condition.

[0229] The deletion and update module is used to add the newly recommended products to the product recommendation list when new recommended products are detected, and delete the products with the lowest user interest from the product recommendation list, thereby updating the recall sorting conditions.

[0230] The steps implemented by the functional modules of the behavior data processing device may refer to the various embodiments of the behavior data processing method of the present invention, and will not be described in detail here.

[0231] In addition, an embodiment of the present invention further provides a storage medium on which a behavior data processing program is stored. When the behavior data processing program is executed by a processor, the implementation method of behavior data processing provided by the above embodiment of the present invention is realized.

Claims

1. A behavior data processing method, characterized in that: The behavior data processing method includes: When a data processing request is detected, obtaining user behavior data and current webpage product association data according to the data processing request; Combine the user behavior data and the current webpage product association data to generate a behavior feature; Determining a corresponding current interest vector according to the behavior characteristics; Inputting the current interest vector into a hidden layer in a preset deep neural network to obtain the current product recommendation degree of different products, wherein the hidden layer in the preset deep neural network reflects the mapping relationship between the interest vector and the product recommendation degree; Selecting a target product based on the current product recommendation level and displaying the target product; The determining of the corresponding current interest vector according to the behavior feature includes: Obtaining user click data, search data, and download data from the behavioral characteristics; Obtaining the number of clicks on each product according to the click data, and comparing each click number with a preset number of clicks to obtain a first comparison result; Obtaining a keyword search count for each product based on the search data, and comparing the keyword search count for each product with a preset search count to obtain a second comparison result; Obtaining a total data download amount for each product based on the downloaded data, and comparing the total data download amount with a preset total data amount to obtain a third comparison result; Determine the product data that the user is interested in based on the first comparison result, the second comparison result, and the third comparison result, where the feature vector corresponding to the product data is the current interest vector; After selecting a target product based on the current product recommendation degree and displaying the target product, the behavior data processing method includes: After receiving the initial address clicked and downloaded by the user, generating a secondary category address of the product based on the initial address, and generating a recall sorting condition based on the secondary category address; Querying the recall data in the search server according to the recall sorting condition; The currently displayed products are adjusted according to the recall data.

2. The behavioral data processing method according to claim 1, wherein: The determining of the product data of interest to the user based on the first comparison result, the second comparison result, and the third comparison result includes: The product with a click count greater than the preset click count in the first comparison result is used as the first product of interest; The product with the keyword search times greater than the preset search times in the second comparison result is used as the second product of interest; The product with the total amount of data downloaded being greater than the preset total amount of data in the third comparison result is used as a third product of interest; The data corresponding to the first interest product, the second interest product, and the third interest product are respectively used as product data of interest to the user.

3. The behavioral data processing method according to claim 1, wherein: Before inputting the current interest vector into a hidden layer in a preset deep neural network to obtain the current product recommendation degrees of different products, the method includes: Obtaining the user's historical behavior data and historical behavior results, and using the historical behavior data and historical behavior results as training data; The initial deep neural network is trained according to the data to be trained to obtain a preset deep neural network.

4. The behavioral data processing method according to claim 3, wherein: The step of training the initial deep neural network according to the data to be trained to obtain a preset deep neural network includes: Obtaining candidate feature data of preset candidate products; Associating the candidate feature data with the data to be trained; Performing weighted summation on the associated candidate feature data and the data to be trained to generate data to be embedded; An initial deep neural network is trained according to the data to be trained and the data to be embedded to obtain a preset deep neural network.

5. The behavior data processing method according to claim 4, characterized in that: The initial deep neural network is trained according to the data to be trained and the data to be embedded to obtain a preset deep neural network, including: Training the initial deep neural network with the data to be trained to obtain an initial training result; Inputting the data to be embedded into the hidden layer of the initial deep neural network to obtain an embedding training result; Inputting the data to be embedded into a preset fusion shallow model to obtain a fusion training result; The initial deep neural network is trained based on the initial training result, the embedding training result and the fusion training result to obtain a preset deep neural network.

6. The behavioral data processing method according to claim 5, characterized in that: The step of obtaining candidate feature data of a preset candidate commodity includes: The category feature data and context feature data of the preset candidate commodity are obtained, and the category feature data and the context feature data are used as candidate feature data.

7. The behavioral data processing method according to claim 6, characterized in that: The step of performing weighted summation on the associated candidate feature data and the data to be trained to generate data to be embedded includes: Performing weighted summation on the associated candidate feature data and the to-be-trained data according to a preset weight coefficient to obtain a summation result, and writing the summation result into a preset distributed file system; Obtain user download behavior information recorded in a preset database warehouse tool table, and obtain log information in the preset distributed file system; The user download behavior information and the log information are combined to generate data to be embedded.

8. The behavioral data processing method according to claim 1, wherein: The adjusting of the currently displayed products according to the recall data includes: Obtaining the exposure of the currently displayed product in the recall data; Sort the products according to the exposure, and obtain the product with the lowest exposure ranking; Increase the number of recommendations for the lowest-ranked product based on the preset number of exposures.

9. The behavioral data processing method according to claim 1, wherein: The querying of the recall data in the search server according to the recall sorting condition includes: When a new recommended product is detected, the recall sorting condition is updated according to the recommended product and the recall data; The corresponding recall data is re-queried in the search server according to the updated recall sorting condition.

10. The behavior data processing method according to claim 9, characterized in that: When a new recommended product is detected, updating the recall sorting condition according to the recommended product and the recall data includes: When a new recommended product is detected, the new recommended product is added to the product recommendation list, thereby updating the recall sorting condition; or, When newly added recommended products are detected, the newly added recommended products are added to the product recommendation list, and the products with the lowest user interest are deleted from the product recommendation list, thereby updating the recall sorting conditions.

11. A behavior data processing device, characterized in that: The behavior data processing device includes: A data acquisition module is used to acquire user behavior data and current webpage product association data according to the data processing request when a data processing request is detected; A feature generation module, configured to combine the user behavior data with the current webpage product association data to generate a behavior feature; A vector acquisition module, configured to determine a corresponding current interest vector according to the behavior characteristics; An input module, configured to input the current interest vector into a hidden layer in a preset deep neural network to obtain current product recommendation degrees for different products, wherein the hidden layer in the preset deep neural network reflects a mapping relationship between the interest vector and the product recommendation degree; A recommendation module, configured to select a target product based on the current product recommendation degree and display the target product; The vector acquisition module is further used to obtain the user's click data, search data and download data from the behavior characteristics; Obtaining the number of clicks on each product according to the click data, and comparing each click number with a preset number of clicks to obtain a first comparison result; Obtaining a keyword search count for each product based on the search data, and comparing the keyword search count for each product with a preset search count to obtain a second comparison result; Obtaining a total data download amount for each product based on the downloaded data, and comparing the total data download amount with a preset total data amount to obtain a third comparison result; Determine the product data that the user is interested in based on the first comparison result, the second comparison result, and the third comparison result, where the feature vector corresponding to the product data is the current interest vector; The recommendation module is further configured to, after receiving the initial address clicked and downloaded by the user, generate a secondary category address of the product based on the initial address, and generate a recall sorting condition based on the secondary category address; Querying the recall data in the search server according to the recall sorting condition; The currently displayed products are adjusted according to the recall data.

12. A behavioral data processing device, characterized in that: The behavior data processing device includes: a memory, a processor, and a behavior data processing program stored in the memory and executable on the processor, wherein the behavior data processing program is configured to implement the steps of the behavior data processing method according to any one of claims 1 to 10.

13. A storage medium, characterized in that: The storage medium stores a behavior data processing program, which, when executed by a processor, implements the steps of the behavior data processing method according to any one of claims 1 to 10.

Images