Commodity recommendation method and device based on multi-channel recall and fusion strategy

Abstract

The embodiment of the present application provides a product recommendation method and device based on multi-channel recall and fusion strategy. By acquiring user behavior data of the entire domain; using preset multi-channel statistical recall rules to perform portrait mining on user behavior data, a first project list corresponding to each statistical recall rule is obtained; the user portrait and policy project portrait are input into a preset deep structure semantic model to obtain a user coding vector and a project coding vector; responding to the user's viewing request; obtaining each first project list from the database according to the user ID, obtaining the user coding vector from the database according to the user ID, and retrieving a preset number of second project lists from the search library according to the user coding vector; fusing and sorting each first project list and the second project list to obtain a recommended product list, and recommending it to the user. The embodiment of the present application can make precise and differentiated recommendations and realize rapid conversion of product value.

Description

Technical Field

[0001] The present application relates to the field of computer application technology, and in particular to a product recommendation method and device based on multi-channel recall and fusion strategy. Background Art

[0002] The application platform has added a new mall business section, which can recommend various products of interest to users (such as policy-related, patent-related, industry-university-research, and legal-related products). However, in the early stages of building the mall platform, there is little or no direct user behavior data on products, and these products have very strong business attributes. If recommendations are only based on simple business rules, the recommended products for all users will not be very differentiated, and it will be impossible to accurately recommend policy-related products that are suitable for users, resulting in a low conversion rate of product value. Summary of the Invention

[0003] This application provides a product recommendation method, device, electronic device and readable storage medium based on multi-channel recall and fusion strategy, which can recommend policy project products of interest to each user, make precise and differentiated recommendations, and realize rapid conversion of product value.

[0004] The technical solution of this application is as follows:

[0005] In a first aspect, the present application provides a product recommendation method based on a multi-channel recall and fusion strategy, the method comprising:

[0006] Acquire user behavior data across the entire domain, the user behavior data including user profiles and policy project profiles corresponding to the user profiles;

[0007] Performing profile mining on the user behavior data using preset multi-channel statistical recall rules to obtain a first item list corresponding to each channel of the statistical recall rules, and storing each channel of the first item list in a preset database;

[0008] Inputting the user profile and the policy project profile into a preset deep structure semantic model to obtain a user coding vector and a project coding vector, and storing the user coding vector in the database and the project coding vector in a preset search library;

[0009] Responding to a user's viewing request, the viewing request including a user ID, the user ID corresponding to the user profile;

[0010] Obtaining each of the first item lists from the database according to the user ID, obtaining a user code vector from the database according to the user ID, and retrieving a preset number of second item lists from the search library according to the user code vector;

[0011] The first item lists and the second item lists of each path are merged and sorted to obtain a recommended product list, and the recommended product list is recommended to the user.

[0012] In the above scheme, firstly, the user behavior data of the entire domain is obtained. The user behavior data includes user portraits and policy project portraits corresponding to the user portraits. By obtaining the user behavior data of the entire domain, more data can be obtained to avoid inaccurate recommendation results due to insufficient data; the user behavior data is subjected to portrait mining using the preset multi-channel statistical recall rules to obtain the first item list corresponding to each statistical recall rule, and each first item list is stored in a preset database, and the first item list of the recalled data is stored for subsequent query; the user portrait and policy project portrait are input into the preset deep structure semantic model to obtain the user encoding vector and the item encoding vector. The deep structure semantic model has a fast training speed and can fully extract the correlation information between the user portrait and the policy project portrait, making the recommendation results more accurate. The user coding vector is stored in the database, and the project coding vector is stored in the preset search library for easy subsequent query; in response to the user's viewing request, the viewing request includes the user ID, and the user ID corresponds to the user portrait; according to the user ID, each first project list is obtained from the database, and the user coding vector is obtained from the database according to the user ID, and a preset number of second project lists are retrieved from the search library according to the user coding vector. The obtained first project list and second project list are related to the user, so as to recommend products of interest to the user; the first project list and the second project list of each path are fused and sorted to obtain a recommended product list, and the information that better meets the user's needs is selected through fusion to recommend the recommended product list to the user, which can make accurate and differentiated recommendations and realize rapid conversion of product value. Compared with the prior art that cannot accurately recommend products suitable for users and reduce the product value conversion rate, the embodiment of the present application can recommend policy project products of interest to each user, make accurate and differentiated recommendations, and realize rapid conversion of product value.

[0013] In some embodiments of the present application, before inputting the user profile and the policy project profile into a preset deep structure semantic model to obtain a user coding vector and a project coding vector, the method further includes:

[0014] Obtaining positive samples and target negative samples based on the user behavior data collection;

[0015] The user profile and the policy project profile are constructed based on the positive sample and the target negative sample.

[0016] In the above technical solution, by collecting positive and negative samples and constructing user portraits and policy project portraits, data samples that meet the requirements of model training are obtained, which can make the model recall data more accurate and provide users with accurate and differentiated recommendations.

[0017] In some embodiments of the present application, the user behavior data includes user click behavior data, which is data obtained by collecting user click operations;

[0018] The obtaining of positive samples and target negative samples based on the user behavior data collection includes:

[0019] In chronological order, the user click behavior data is used as a positive sample, and the untrusted data in the user click behavior data is used as the first negative sample. In the case of historical embedding, the second negative sample is extracted according to the business scenario of the user click behavior data using preset business rules. If the number of the target negative samples is insufficient, the third negative sample is obtained by sampling from the user behavior data of the entire domain.

[0020] The first negative sample, the second negative sample, and the third negative sample constitute the target negative sample, and the positive sample is balanced with the target negative sample.

[0021] In the above technical solution, negative samples are sampled in different ways to balance positive and negative samples, making the recommendation results more accurate.

[0022] In some embodiments of the present application, the steps of merging and sorting the first item lists and the second item lists to obtain a recommended product list include:

[0023] Setting weights for the items in the first item list and the items in the second item list on each path, to obtain a first item weight table and a second item weight table on each path, wherein the first item weight table and the second item weight table each include a plurality of weights corresponding to each item;

[0024] According to the first item weight table and the second item weight table of each path, the items in the first item list and the items in the second item list of each path are integrated and scored to obtain multiple item scores;

[0025] The item scores are sorted to obtain the recommended product list.

[0026] In the above technical solution, by integrating and selecting information that better meets user needs, accurate and differentiated recommendations can be made, achieving rapid conversion of product value.

[0027] In some embodiments of the present application, setting weights for the items in the first item list and the items in the second item list on each path to obtain the first item weight table and the second item weight table on each path includes:

[0028] Grouping the items in the first item list and the items in the second item list of each path to obtain a grouping table;

[0029] Obtaining a user click table, associating the group table with the user click table to obtain an association table, and querying from the association table to obtain a first recall path corresponding to each item in the first item list and a second recall path corresponding to each item in the second item list;

[0030] Splitting the plurality of first recall paths and the second recall paths and counting the number of clicks to obtain the number of clicks on the items in each recall path, wherein the items in each recall path are items in the first item list or the second item list;

[0031] The first item weight table and the second item weight table of each path are obtained according to the number of clicks.

[0032] In the above technical solution, the items in each first item list and the items in the second item list are grouped, and then the grouping table is associated with the user click table. Subsequently, the multiple first recall paths and second recall paths obtained by querying the association table are split and the number of times is counted, and the weight is set according to the number of clicks so that subsequent fusion processing can be performed according to the weight.

[0033] In some embodiments of the present application, before obtaining global user behavior data, the method further includes:

[0034] Collect log information offline through the preset log collection system to obtain multiple offline user behavior logs;

[0035] Collect data online through the preset data collection interface to obtain multiple online user behavior logs;

[0036] The user behavior data is obtained according to each of the offline user behavior logs and each of the online user behavior logs.

[0037] In the above technical solution, by collecting data offline and online, user behavior data can be enriched, thereby achieving personalized recommendations and recommending products of interest to users.

[0038] In some embodiments of the present application, obtaining the user behavior data according to each of the offline user behavior logs and each of the online user behavior logs includes:

[0039] Performing data de-dirt processing on each of the offline user behavior logs and each of the online user behavior logs to obtain a plurality of first offline behavior data and a plurality of first online behavior data;

[0040] Deduplication is performed on the plurality of first offline behavior data and the plurality of first online behavior data to obtain a plurality of second offline behavior data and a plurality of second online behavior data;

[0041] performing desensitization processing on the plurality of second offline behavior data and the plurality of second online behavior data respectively to obtain a plurality of third offline behavior data and a plurality of third online behavior data;

[0042] The plurality of the third offline behavior data and the plurality of the third online behavior data are combined into the user behavior data.

[0043] In the above technical solution, data cleaning is performed by removing dirty data, deduplicating data, and desensitizing the collected data, which not only improves the accuracy of recommending policy project products, but also protects user privacy.

[0044] In a second aspect, an embodiment of the present application provides a product recommendation device based on a multi-channel recall and fusion strategy, the device comprising:

[0045] A first data acquisition module is used to acquire user behavior data of the entire domain, wherein the user behavior data includes user profiles and policy project profiles corresponding to the user profiles;

[0046] a first data mining module, configured to perform profile mining on the user behavior data using preset multi-channel statistical recall rules, obtain a first item list corresponding to each channel of the statistical recall rules, and store each first item list in a preset database;

[0047] A second data mining module is configured to input the user profile and the policy project profile into a preset deep structure semantic model to obtain a user coding vector and a project coding vector, and store the user coding vector in the database and the project coding vector in a preset search library;

[0048] A data response module, configured to respond to a user's viewing request, wherein the viewing request includes a user ID corresponding to the user portrait;

[0049] a second data acquisition module, configured to acquire the first item lists of each path from the database according to the user ID, acquire a user coding vector from the database according to the user ID, and retrieve a preset number of second item lists from the search library according to the user coding vector;

[0050] The fusion recommendation module is used to fuse the first item list and the second item list of each channel, and sort them to obtain a recommended product list, and recommend the recommended product list to the user.

[0051] In a third aspect, an embodiment of the present application provides an electronic device comprising a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes any one of the methods provided in the first aspect above.

[0052] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed, any one of the methods provided in the first aspect is executed.

[0053] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0054] 1. By acquiring global user behavior data, mining the user behavior data using multi-channel statistical recall rules and deep semantic structure models, and writing the data into a database and search library, and then fusing the data to recommend a list of recommended products to the user, the present invention effectively solves the problem in related technologies of being unable to recommend policy-related products of interest to users, resulting in a low conversion rate of product value. The embodiments of the present application can acquire more behavioral data, recommend policy-related products of interest to users, make precise and differentiated recommendations, and achieve rapid conversion of product value.

[0055] 2. Based on the acquisition of global user behavior data, sampling and processing the data can alleviate sample imbalance and make the recommendation results more accurate.

[0056] 3. By collecting data offline and online, we can enrich user behavior data and then implement personalized recommendations to recommend products that users are interested in.

[0057] 4. Clean the collected data to improve the accuracy of recommended policy project products. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flowchart of a product recommendation method based on multi-channel recall and fusion strategy provided by an embodiment of the present application;

[0059] Figure 2 This is a flowchart of a product recommendation method based on multi-channel recall and fusion strategy provided by another embodiment of the present application;

[0060] Figure 3 yes Figure 2 A schematic flow chart of a sub-step of step S900;

[0061] Figure 4 yes Figure 1 A schematic flow chart of a sub-step of step S600;

[0062] Figure 5 yes Figure 4 A schematic flow chart of a sub-step of step S610;

[0063] Figure 6 This is a structural diagram of a product recommendation device based on a multi-channel recall and fusion strategy provided by an embodiment of the present application;

[0064] Figure 7 This is a schematic structural diagram of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0066] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0067] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0068] In related technologies, Qizhidao's product is a big data science and innovation service platform. The mall platform is the primary business module on the product homepage, providing immediate recommendations for various products of interest to users (including those related to policies, patents, industry-university-research cooperation, and legal issues). During the initial launch of the mall platform, there was little data on user behavior, resulting in an imbalanced sample size. Furthermore, due to the multiple recall channels, the scores for recalled items varied across different dimensions, and the integration of items also affected the recommendations of products of interest to users.

[0069] Based on this, the embodiment of the present application provides a product recommendation method and device based on multi-channel recall and fusion strategy, the product recommendation method based on multi-channel recall and fusion strategy obtains global user behavior data, the user behavior data includes user portraits and policy project portraits corresponding to the user portraits, by obtaining global user behavior data, more data can be obtained, avoiding inaccurate recommendation results due to insufficient data; the user behavior data is subjected to portrait mining using preset multi-channel statistical recall rules, and the first item list corresponding to each statistical recall rule is obtained, and each first item list is stored in a preset database, and the first item list of the recalled data is stored for convenience of subsequent query; the user portrait and the policy project portrait are input into a preset deep structure semantic model to obtain a user encoding vector and an item encoding vector, and the deep structure semantic model has a fast training speed and can fully extract user The association information between the portrait and the policy project portrait makes the recommendation results more accurate, and stores the user coding vector in the database, and stores the project coding vector in a preset search library to facilitate subsequent queries; responding to the user's viewing request, the viewing request includes the user ID, and the user ID corresponds to the user portrait; obtaining each first project list from the database according to the user ID, obtaining the user coding vector from the database according to the user ID, and retrieving a preset number of second project lists from the search library according to the user coding vector. The obtained first project list and second project list are related to the user, so as to recommend products of interest to the user; fusing the first project list and the second project list of each path and sorting them to obtain a recommended product list. By fusing and selecting information that better meets the user's needs, the recommended product list is recommended to the user, which can make accurate and differentiated recommendations and achieve rapid conversion of product value. Compared with the prior art that cannot accurately recommend products suitable for users and reduce the product value conversion rate, the embodiment of the present application can recommend policy project products of interest to each user, make accurate and differentiated recommendations, and achieve rapid conversion of product value.

[0070] It should be noted that the product recommendation method based on multi-channel recall and fusion strategies can be applied to recommending policy project products, patent products, and digital transformation standards. This product recommendation method based on multi-channel recall and fusion strategies can recommend policy project products of interest to each user, making precise and differentiated recommendations, achieving rapid conversion of product value, and has wide application in recommending industry-university-research projects.

[0071] The technical solutions provided in the embodiments of the present application are further described below in conjunction with the accompanying drawings.

[0072] Reference Figure 1 , Figure 1 1 is a flow chart of a product recommendation method based on a multi-channel recall and fusion strategy provided by an embodiment of the present application. The product recommendation method based on a multi-channel recall and fusion strategy includes steps S100, S200, S300, S400, S500, and S600.

[0073] Step S100: Acquire user behavior data for the entire domain, where the user behavior data includes user profiles and policy project profiles corresponding to the user profiles.

[0074] In one embodiment, since there is almost no user behavior data in the initial launch phase of the platform, global user behavior data is obtained through a preset data reading interface. The user behavior data includes user profiles and policy project profiles corresponding to the user profiles. By obtaining global user behavior data, more user behavior data can be obtained, avoiding inaccurate recommendation results due to insufficient data. Among them, the user profile includes user click behavior data, which is data obtained by collecting user click operations, and also includes user information, exemplarily, user ID, gender, age, user type, whether internal user, access platform preference, access time preference, whether to apply for a project, region, device model, device type, operating system, network type and other contextual information. The policy project profile is based on the user's viewing and clicking of policies or patents, as well as the project product data exposed to the user. For example, the policy project profile includes project ID, project type, project level, funding amount, project province, city and district code, application start time, application deadline and the number of clicks and exposures of the project in the last 7 / 14 / 30 days. The preset data reading interface can be an interface for reading a database, or an interface for reading a saved log file, etc. It can read data log files and obtain global user behavior data, which will not be described in detail here.

[0075] In one embodiment, before obtaining global user behavior data, Figure 2 As shown, the product recommendation method based on the multi-channel recall and fusion strategy also includes but is not limited to the following steps:

[0076] Step S700: Collect log information offline through a preset log collection system to obtain multiple offline user behavior logs.

[0077] In one possible embodiment of the present application, the preset log collection system may be Flume, a distributed, reliable, and highly available system for collecting, aggregating, and transmitting massive amounts of logs. Flume-agent can be used to collect log information offline, generating multiple offline user behavior logs. Offline data collection can generate historical user data, allowing for a sufficient amount of user behavior data.

[0078] Step S800: Collect data online through a preset data collection interface to obtain multiple online user behavior logs.

[0079] In one possible embodiment of the present application, the preset data collection interface can be provided by Kafka, which can process all consumer action stream data on the website. By connecting Kafka to Flink for real-time consumption and online data collection, multiple online user behavior logs can be obtained. Online data collection can obtain historical and current user data, thereby obtaining a sufficient amount of user behavior data.

[0080] Step S900 : obtaining user behavior data according to each offline user behavior log and each online user behavior log.

[0081] In a possible embodiment of the present application, user behavior data is obtained based on the offline user behavior log obtained in step S700 and the online user behavior log obtained in step S800. By collecting data offline and online, user behavior data can be enriched, thereby achieving personalized recommendations and providing data support for recommending products of interest to users.

[0082] like Figure 3 As shown, obtaining user behavior data based on each offline user behavior log and each online user behavior log includes but is not limited to the following steps:

[0083] Step S910 : performing data de-dirt processing on each offline user behavior log and each online user behavior log to obtain a plurality of first offline behavior data and a plurality of first online behavior data.

[0084] In one possible embodiment of the present application, since the obtained offline user behavior logs and online user behavior logs contain dirty data and duplicate data, each offline user behavior log and each online user behavior log are first subjected to data de-dirt processing to obtain multiple first offline behavior data and multiple first online behavior data. Pre-processing the data can improve the accuracy of recommendations.

[0085] Step S920 : Deduplication processing is performed on the plurality of first offline behavior data and the plurality of first online behavior data to obtain a plurality of second offline behavior data and a plurality of second online behavior data.

[0086] In a possible embodiment of the present application, the first offline behavior data and the first online behavior data obtained in step S910 are deduplicated respectively to obtain a plurality of second offline behavior data and a plurality of second online behavior data. Through the deduplication process, data redundancy can be reduced and the storage space occupied by the data can be reduced.

[0087] Step S930 : Desensitizing the plurality of second offline behavior data and the plurality of second online behavior data to obtain a plurality of third offline behavior data and a plurality of third online behavior data.

[0088] In a possible embodiment of the present application, since the user's behavioral operations are involved, the user's private data may exist. Therefore, the second offline behavior data and the second online behavior data obtained according to step S920 are desensitized respectively to obtain multiple third offline behavior data and multiple third online behavior data, which can protect the user's privacy from being leaked.

[0089] Step S940: Combining the plurality of third offline behavior data and the plurality of third online behavior data into user behavior data.

[0090] In one possible embodiment of the present application, the plurality of third offline behavior data and the plurality of third online behavior data obtained in step S930 are combined into user behavior data, and the data is stored in a Hive database. This data cleaning not only protects privacy and reduces data redundancy, but also helps improve the accuracy and diversity of recommendations.

[0091] In step S200 , the user behavior data is profiled and mined using the preset multi-channel statistical recall rules to obtain a first item list corresponding to each channel of statistical recall rules, and each first item list is stored in a preset database.

[0092] In one embodiment, the preset multi-channel statistical recall rules include popular regional policies, user recent behavior sequences, user policy category preferences, user high-value behaviors (submitting a reservation or project consultation), user policy search term predictions, etc. After the data is cleaned according to the above step 900, the data is extracted based on hive sql to recommend online services. The user behavior data is profiled and mined using dimensions such as popular regional policies, user recent behavior sequences, user policy category preferences, user high-value behaviors (submitting a reservation or project consultation), and user policy search term predictions to obtain a first item list corresponding to each statistical recall rule. Through multi-channel information recall, user behavior data can be obtained, making recommendations more accurate and differentiated, thereby achieving rapid conversion of commodity value. Multiple first item lists are obtained through various statistical recall rules, and each first item list is stored in a preset database to facilitate subsequent acquisition of the first item list for recommendation. The database is a redis library.

[0093] Step S300: Input the user portrait and policy project portrait into the preset deep structure semantic model to obtain the user coding vector and the project coding vector, and store the user coding vector in the database and the project coding vector in the preset search library.

[0094] In one embodiment, the preset Deep Structured Semantic Model (DSSM) is a fast-training model that eliminates the need to save model files and directly stores the data generated by model training. User profiles and policy project profiles are input into the preset deep structured semantic model to generate user and project encoding vectors. The user encoding vectors are stored in a database, and the project encoding vectors are stored in a preset search library. This allows for subsequent retrieval of the user and project encoding vectors, data fusion, and recommendations to users. The database is a Redis database, and the search library is a Faiss database.

[0095] In one embodiment, because the DSSM requires a large number of samples, sample sampling is necessary. Before inputting user and policy project profiles into a preset deep structure semantic model to obtain user and project encoding vectors, the product recommendation method based on a multi-channel recall and fusion strategy also includes collecting positive and target negative samples based on user behavior data to obtain a large amount of data, facilitate training of the DSSM, and improve recommendation accuracy. The positive and target negative samples are then vectorized, and user and policy project profiles are constructed using preset feature engineering to rapidly train the DSSM.

[0096] In another embodiment, the user behavior data includes user click behavior data, and the user click behavior data is data obtained by collecting the user's click operations. Since the ratio of exposure to click samples in the recommendation system may be extremely unbalanced, it is necessary to collect a certain ratio of negative samples, and obtain positive samples and target negative samples based on the user behavior data in the following manner. The sampling effect is evaluated using a positive and negative sample ratio of (1:3, 1:7, 1:9, 1:11, 1:15 and 1:17). Since the effect of direct random sampling in the initial stage is very unsatisfactory, the effect and performance of the sampling ratio of 1:11 are relatively good, so sampling is performed based on certain rules. Based on the above sampling ratio, one sampling method is to use the user click behavior data as the positive sample according to the exposure event in chronological order, and the untrusted data in the user click behavior data as the first negative sample, which is conducive to obtaining a sufficient number of negative samples later. Among them, untrusted data refers to data that the user did not click in the exposure event. According to the above sampling ratio rule, for example, the sampling ratio is 1:5, event 5 is clicked, which is the user's click behavior data and serves as a positive sample; events 1-4 are not clicked, which are untrusted data and serve as the first negative sample.

[0097] Another sampling method is to extract a second negative sample based on the business scenario of the user's click behavior data using preset business rules when there are historical tracking points. This helps to obtain a sufficient number of negative samples later. For example, when viewing the company details page, the positive sample is when the user clicks on a policy item in the company's list of applicable projects. The negative sample has no tracking point, so the policy interface needs to be used to extract the current company's list of applicable projects for sampling.

[0098] Another sampling method is to sample from global user behavior data to obtain a third negative sample when the target negative sample number is insufficient. This helps to obtain a sufficient number of negative samples later. The first, second, and third negative samples obtained from the above sampling constitute the target negative sample. The positive samples are balanced with the target negative samples. By reading data and training the DSSM model in batches, the model is exposed to a wider range of negative samples, which to a certain extent enhances the model's generalization ability.

[0099] Step S400: responding to a viewing request from a user, the viewing request includes a user ID, and the user ID corresponds to a user portrait.

[0100] In one embodiment, a user views shopping mall information on the platform in response to a user's viewing request, which includes a user ID, and the user ID corresponds to a user portrait, so that the user ID can subsequently search for the first item list and the user encoding vector.

[0101] Step S500: Obtain each first item list from the database according to the user ID, obtain the user coding vector from the database according to the user ID, and retrieve a preset number of second item lists from the search library according to the user coding vector.

[0102] In one embodiment, since the user portrait corresponds to the user ID, and the first item list is obtained based on the user behavior data, the user behavior data includes the user ID. Therefore, each first item list is obtained from the database according to the user ID, the user encoding vector is obtained from the database according to the user ID, and a preset number of second item lists are retrieved from the search library according to the user encoding vector. This is conducive to the subsequent fusion processing of each first item list and second item list to recommend a list of products to the user. Exemplarily, there are 6 recall channels, which are the channels corresponding to recall strategies such as regional popular policies, user recent behavior sequences, user policy category preferences, user high-value behaviors, user policy search term predictions, and deep structure semantic models. Among them, the first 5 correspond to the first item list, and the last one corresponds to the second item list.

[0103] Step S600: Merge and sort the first item lists and the second item lists to obtain a recommended product list, and recommend the recommended product list to the user.

[0104] like Figure 4 As shown, the first item list and the second item list of each path are merged and sorted to obtain a recommended product list, including but not limited to the following steps:

[0105] Step S610 , weights are set for the items in each first item list and the items in the second item list to obtain each first item weight table and second item weight table, wherein the first item weight table and the second item weight table both include multiple weights corresponding to each item.

[0106] In one possible embodiment of the present application, the first item weight table includes multiple weights corresponding to each item, and the second item weight table includes multiple weights corresponding to each item. Weights are assigned to the items in the first item list and the items in the second item list for each path to obtain the first item weight table and the second item weight table for each path. The weights can be dynamically updated on a T+1 basis to facilitate subsequent fusion scoring.

[0107] like Figure 5 As shown, weights are set for the items in the first item list and the items in the second item list of each path to obtain the first item weight table and the second item weight table of each path, including but not limited to the following steps:

[0108] Step S611 , grouping the items in the first item list and the items in the second item list of each path to obtain a grouping table.

[0109] In a possible embodiment of the present application, the first item list includes multiple items, and the second item list includes multiple items. Since different recall rules will recall overlapping items, the items in each of the first item list and the second item list are grouped to obtain a grouping table, which facilitates the subsequent acquisition of an association table based on the grouping table.

[0110] Step S612: Obtain the user click table, associate the group table with the user click table to obtain an association table, and query the association table to obtain the first recall path corresponding to the item in each first item list and the second recall path corresponding to the item in the second item list.

[0111] In one possible embodiment of the present application, a user click table can be constructed based on user click operations and stored in a database. The user click table is then retrieved and the grouping table is associated with the user click table to obtain an association table. Since items are retrieved through user behavior data and are associated with user click behaviors, the association table records the relationship between users, items, and the recall paths corresponding to the items. The first recall path corresponding to each item in the first item list and the second recall path corresponding to each item in the second item list can be obtained from the association table, facilitating subsequent statistics on the number of clicks on the items.

[0112] Step S613 , splitting the plurality of first recall paths and second recall paths and counting the number of clicks to obtain the number of clicks on the items in each recall path, wherein the items in each recall path are items in the first item list or the second item list.

[0113] In one possible embodiment of the present application, one recall path corresponds to one recall strategy, and each item in the recall path is either an item in the first item list or an item in the second item list, i.e., one recall path corresponds to one item list. Based on the obtained multiple first and second recall paths, the paths are split according to the recall paths, and the number of clicks on the items in each recall path is counted to obtain the click count, which facilitates the subsequent generation of the first and second item weight tables.

[0114] Step S614: Obtain the first item weight table and the second item weight table for each path according to the number of clicks.

[0115] In one possible embodiment of the present application, the number of clicks obtained in step S613 is sorted. If the sorting is in descending order, the weight of the items set earlier is greater, and vice versa. Based on the weight of each item, a first item weight table and a second item weight table are obtained for each path, so that the various strategies can be subsequently integrated to provide personalized recommendations to users and accurately recommend products.

[0116] Step S620 , based on the first item weight table and the second item weight table of each path, the items in the first item list and the items in the second item list of each path are integrated and scored to obtain a plurality of item scores.

[0117] In one possible embodiment of the present application, to make product recommendations more accurate, the item weights in each first-item weight table and second-item weight table are first normalized. After normalization, the items in each first-item list and second-item list are fused and scored based on the weights to obtain multiple item scores to create a list of recommended products. For example, the fusion metric is recall@50 (representing the proportion of actual user clicks on the top 50 policy items in each recall).

[0118] Step S630: sort the item scores to obtain a list of recommended products.

[0119] In one possible embodiment of the present application, the project scores obtained in step S620 are sorted. The sorting may be in descending order, with the top n items recommended as a recommended product list. Alternatively, the sorting may be in ascending order, with the bottom n items recommended as a recommended product list. Where n can be 3 or 5, and is set as required, which will not be described in detail here. The recommended product list is displayed to the user. The recommended products in the recommended product list are products that the user is interested in, which can improve the conversion of product value.

[0120] like Figure 6As shown, an embodiment of the present application provides a product recommendation device 100 based on a multi-channel recall and fusion strategy, which obtains global user behavior data through a first data acquisition module 110. The user behavior data includes a user portrait and a policy project portrait corresponding to the user portrait. By obtaining global user behavior data, more data can be obtained to avoid inaccurate recommendation results due to insufficient data; then, a first data mining module 120 is used to perform portrait mining on the user behavior data using a preset multi-channel statistical recall rule to obtain a first item list corresponding to each statistical recall rule, and each first item list is stored in a preset database, and the first item list of the recalled data is stored for convenience of subsequent query; a second data mining module 130 is used to input the user portrait and the policy project portrait into a preset deep structure semantic model to obtain a user coding vector and a project coding vector. The deep structure semantic model has a fast training speed and can fully extract the correlation information between the user portrait and the policy project portrait. , making the recommendation results more accurate, and storing the user coding vector in the database, and storing the item coding vector in the preset search library to facilitate subsequent queries; using the data response module 140 to respond to the user's viewing request, the viewing request includes the user ID, and the user ID corresponds to the user portrait; then using the second data acquisition module 150 to obtain each first item list from the database according to the user ID, obtain the user coding vector from the database according to the user ID, and retrieve a preset number of second item lists from the search library according to the user coding vector. The obtained first item list and second item list are related to the user, so as to recommend products of interest to the user; finally, using the fusion recommendation module 160 to fuse the first item list and the second item list of each path, and sort them to obtain a recommended product list, and recommend the recommended product list to the user. By fusing and selecting information that better meets user needs, the recommended product list is recommended to the user, which can make accurate and differentiated recommendations and realize rapid conversion of product value.

[0121] It should be noted that the first data acquisition module 110 is connected to the first data mining module 120, the first data mining module 120 is connected to the second data mining module 130, the second data mining module 130 is connected to the data response module 140, the data response module 140 is connected to the second data acquisition module 150, and the second data acquisition module 150 is connected to the fusion recommendation module 160. The above-mentioned product recommendation method based on the multi-way recall and fusion strategy is applied to the product recommendation device 100 based on the multi-way recall and fusion strategy. The product recommendation device 100 based on the multi-way recall and fusion strategy obtains user behavior data from the entire domain, and uses multi-way statistical recall rules and deep semantic structure models to perform data mining on the user behavior data, and writes the data into the database and search library respectively. After fusion processing, it recommends a list of recommended products to the user, can obtain more behavior data, and recommend policy project products of interest to the user, make accurate and differentiated recommendations, and realize rapid conversion of product value.

[0122] It should also be noted that the above embodiments provide devices that implement their functions using only the division of the above functional modules as examples. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0123] This application also discloses an electronic device. Figure 7 , Figure 7 Schematic diagram of the structure of an electronic device disclosed in an embodiment of the present application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0124] The communication bus 502 is used to implement the connection and communication between these components.

[0125] The user interface 503 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0126] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0127] The processor 501 may include one or more processing cores. Using various interfaces and circuits, the processor 501 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 505, as well as accesses data stored in the memory 505, to perform various server functions and process data. Optionally, the processor 501 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 501 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may also be implemented as a separate chip, rather than integrated into the processor 501.

[0128] Among them, the memory 505 may include a random access memory (RAM) or a read-only memory (Read-Only Memory). Optionally, the memory 505 includes a non-transitory computer-readable storage medium. The memory 505 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 505 may also be optionally at least one storage device located away from the aforementioned processor 501. Reference Figure 7 , the memory 505 as a computer storage medium may include an operating system, a network communication module, a user interface module and an application program of a product recommendation method based on multi-channel recall and fusion strategy.

[0129] exist Figure 7In the electronic device 500 shown, the user interface 503 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 501 can be used to call the application program stored in the memory 505 for a product recommendation method based on multi-way recall and fusion strategy. When executed by one or more processors 501, the electronic device 500 executes one or more methods in the above-mentioned embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should know that this application is not limited to the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0130] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0131] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0132] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0133] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0134] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory includes various media that can store program code, such as USB flash drives, mobile hard drives, magnetic disks, or optical disks.

[0135] The above are merely exemplary embodiments of the present disclosure and are not intended to limit the scope of the present disclosure. In other words, any equivalent variations and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the disclosure and the practical implications thereof.

[0136] This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not described herein. The description and examples are to be considered as exemplary only, and the scope and spirit of the present disclosure are to be defined by the claims.

Claims

1. A product recommendation method based on multi-channel recall and fusion strategy, characterized in that: The method comprises: Acquire user behavior data across the entire domain, the user behavior data including user profiles and policy project profiles corresponding to the user profiles; Performing profile mining on the user behavior data using preset multi-channel statistical recall rules to obtain a first item list corresponding to each channel of the statistical recall rules, and storing each channel of the first item list in a preset database; Inputting the user profile and the policy project profile into a preset deep structure semantic model to obtain a user coding vector and a project coding vector, and storing the user coding vector in the database and the project coding vector in a preset search library; Responding to a user's viewing request, the viewing request including a user ID, the user ID corresponding to the user profile; Obtaining each of the first item lists from the database according to the user ID, obtaining a user code vector from the database according to the user ID, and retrieving a preset number of second item lists from the search library according to the user code vector; The first item lists and the second item lists of each path are merged and sorted to obtain a recommended product list, and the recommended product list is recommended to the user.

2. The method according to claim 1, characterized in that Before inputting the user profile and the policy project profile into a preset deep structure semantic model to obtain a user coding vector and a project coding vector, the method further includes: Obtaining positive samples and target negative samples based on the user behavior data collection; The user profile and the policy project profile are constructed based on the positive sample and the target negative sample.

3. The method according to claim 2, characterized in that The user behavior data includes user click behavior data, which is data obtained by collecting user click operations; The obtaining of positive samples and target negative samples based on the user behavior data collection includes: In chronological order, the user click behavior data is used as a positive sample, and the untrusted data in the user click behavior data is used as the first negative sample. In the case of historical embedding, the second negative sample is extracted according to the business scenario of the user click behavior data using preset business rules. If the number of the target negative samples is insufficient, the third negative sample is obtained by sampling from the user behavior data of the entire domain. The first negative sample, the second negative sample, and the third negative sample constitute the target negative sample, and the positive sample is balanced with the target negative sample.

4. The method according to claim 1, wherein The step of merging and sorting the first item lists and the second item lists to obtain a recommended product list includes: Setting weights for the items in the first item list and the items in the second item list on each path, to obtain a first item weight table and a second item weight table on each path, wherein the first item weight table and the second item weight table each include a plurality of weights corresponding to each item; According to the first item weight table and the second item weight table of each path, the items in the first item list and the items in the second item list of each path are integrated and scored to obtain multiple item scores; The item scores are sorted to obtain the recommended product list.

5. The method according to claim 4, characterized in that The step of setting weights for the items in the first item list and the items in the second item list of each path to obtain a first item weight table and a second item weight table of each path includes: Grouping the items in the first item list and the items in the second item list of each path to obtain a grouping table; Obtaining a user click table, associating the group table with the user click table to obtain an association table, and querying from the association table to obtain a first recall path corresponding to each item in the first item list and a second recall path corresponding to each item in the second item list; Splitting the plurality of first recall paths and the second recall paths and counting the number of clicks to obtain the number of clicks on the items in each recall path, wherein the items in each recall path are items in the first item list or the second item list; The first item weight table and the second item weight table of each path are obtained according to the number of clicks.

6. The method according to claim 1, characterized in that Before obtaining the global user behavior data, the method further includes: Collect log information offline through the preset log collection system to obtain multiple offline user behavior logs; Collect data online through the preset data collection interface to obtain multiple online user behavior logs; The user behavior data is obtained according to each of the offline user behavior logs and each of the online user behavior logs.

7. The method according to claim 6, characterized in that The obtaining of the user behavior data according to each of the offline user behavior logs and each of the online user behavior logs includes: Performing data de-dirt processing on each of the offline user behavior logs and each of the online user behavior logs to obtain a plurality of first offline behavior data and a plurality of first online behavior data; Deduplication is performed on the plurality of first offline behavior data and the plurality of first online behavior data to obtain a plurality of second offline behavior data and a plurality of second online behavior data; performing desensitization processing on the plurality of second offline behavior data and the plurality of second online behavior data respectively to obtain a plurality of third offline behavior data and a plurality of third online behavior data; The plurality of the third offline behavior data and the plurality of the third online behavior data are combined into the user behavior data.

8. A product recommendation device based on multi-channel recall and fusion strategy, characterized in that: The device comprises: A first data acquisition module is used to acquire user behavior data of the entire domain, wherein the user behavior data includes user profiles and policy project profiles corresponding to the user profiles; a first data mining module, configured to perform profile mining on the user behavior data using preset multi-channel statistical recall rules, obtain a first item list corresponding to each channel of the statistical recall rules, and store each first item list in a preset database; A second data mining module is configured to input the user profile and the policy project profile into a preset deep structure semantic model to obtain a user coding vector and a project coding vector, and store the user coding vector in the database and the project coding vector in a preset search library; A data response module, configured to respond to a user's viewing request, wherein the viewing request includes a user ID corresponding to the user portrait; a second data acquisition module, configured to acquire the first item lists of each path from the database according to the user ID, acquire a user coding vector from the database according to the user ID, and retrieve a preset number of second item lists from the search library according to the user coding vector; The fusion recommendation module is used to fuse the first item list and the second item list of each channel, and sort them to obtain a recommended product list, and recommend the recommended product list to the user.

9. An electronic device, characterized in that: The electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 7 is executed.

Images