A customized recommendation method and recommendation system based on multi-platform user data

By building a time-measurement user behavior evaluation model and multi-platform recall strategy in the recommendation system, the problem of user behavior scoring in the existing technology is solved, and the real-time and diversity of recommendation results are improved. sex.

CN113850649BActive Publication Date: 2024-11-22AEROSPACE INFORMATION RES INST CAS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111147286.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-29
Publication Date
2024-11-22
Estimated Expiration
2041-09-29

AI Technical Summary

Technical Problem

When the existing recommendation system is building a user behavior scoring system, time factors are not considered, resulting in a lack of real-time recommendation results; at the same time, the information between multiple portal systems cannot be correlated, resulting in the loss and duplication of recommendation information, reducing the recommendation 结果的准确性和多样性。

Method used

Using a customized recommendation method based on multi-platform user data, a user behavior evaluation model containing time measurement is constructed, combined with multi-modal entity mapping and multi-platform behavior fusion recall strategy, establish associations through multi-platform data to improve the timeliness of recommendation results and accuracy.

Benefits of technology

It realizes the time sensitivity of the user behavior scoring system, reduces the loss and duplication of recommendation information, and improves the diversity and accuracy of recommendation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113850649B_ABST
    Figure CN113850649B_ABST
Patent Text Reader

Abstract

The present invention discloses a customized recommendation method based on multi-platform user data, including: collecting user data from multiple platforms; constructing a multi-platform user portrait based on the user data, wherein the multi-platform user portrait includes user behavior information and user attribute information; constructing a user behavior evaluation model including time measurement based on the multi-platform user portrait, wherein the user behavior evaluation model is used to calculate the user's preference value for an entity; constructing a multi-platform recall strategy based on the calculation result of the user behavior evaluation model, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method; obtaining a recommendation result using a ranking model based on the output result of the multi-platform recall strategy; feeding back the recommendation result to the user; and outputting a customized recommendation result based on the user's feedback result. The present invention also discloses a customized recommendation system based on multi-platform user data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of multi-platform recommendation strategy customization in a recommendation system, and in particular to a customized recommendation method and a recommendation system based on multi-platform user data. Background Art

[0002] A recommendation system is an information filtering system that predicts a user's rating or preference for an item. With the development of information technology and the Internet, people have gradually moved from an era of information scarcity to an era of information overload. In this case, consumers want to find information of interest from a large amount of information (such as items in an online shopping mall), while information producers want the information they produce to stand out. One way to solve the above problem is to use a recommendation system to connect users and information and build a bridge between the two. At the same time, another problem that the recommendation system needs to solve is to discover user behavior, find users' personalized needs, and help users find information that interests them.

[0003] The process of generating recommendation results by existing recommendation systems mainly includes two stages: the first stage is the recall stage, which uses a variety of fast-running recall strategies to perform initial screening of items in the database and generate a candidate set of user-favorite items; the second stage is the sorting stage, which uses a sorting algorithm to score and sort the candidate set generated in the first stage to produce ordered recommendation results.

[0004] In the recall stage, existing recommendation systems usually adopt algorithms based on user collaborative filtering, item collaborative filtering, content recommendation, etc.; in the sorting stage, commonly used sorting algorithms can be roughly divided into two categories: the first category is the strategy-based sorting method, that is, the system performs mixed sorting and recommendation on candidate data sets according to weighted mixing and other methods; the second category is the model-based sorting method, and commonly used models include logistic regression, gradient boosting tree, linear and deep models Wide And Deep models, etc.

[0005] At present, there have been many studies on recommendation systems and recommendation strategies. In the existing technologies, the focus of research is on increasing the number of recall sources and using deep learning models for scoring, so as to increase the richness and accuracy of recommendation effects. However, in the actual application of this type of method, there are still two key problems that need to be solved: First, when building a user behavior scoring system, the scoring of user behavior is not affected by time. For example, a user's browsing behavior a week ago and the browsing behavior that occurred recently received the same score, which makes the system believe that the importance of the two behaviors is similar. In this case, the real-time nature of the recommendation results cannot be reflected. Second, there are few recommendation systems or strategies that generate unified recommendation results for multiple portal systems. This is because the information in different platforms may not be able to be correlated with each other, resulting in the loss and duplication of recommendation information, and reducing the accuracy and diversity of recommendation results. Summary of the invention

[0006] In view of this, the main purpose of the present invention is to provide a customized recommendation method and recommendation system based on multi-platform user data, in order to at least partially solve at least one of the above-mentioned technical problems.

[0007] According to one aspect of the present invention, a customized recommendation method based on multi-platform user data is provided, comprising:

[0008] A customized recommendation method based on multi-platform user data, comprising:

[0009] Collect user data from multiple platforms;

[0010] Build a multi-platform user portrait based on user data, where the multi-platform user portrait includes user behavior information and user attribute information;

[0011] Based on multi-platform user portraits, a user behavior evaluation model including time measurement is constructed, where the user behavior evaluation model is used to calculate the user's preference value for the entity;

[0012] According to the calculation results of the user behavior evaluation model, a multi-platform recall strategy is constructed, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method;

[0013] According to the output results of the multi-platform recall strategy, the ranking model is used to obtain the recommendation results;

[0014] Feedback the recommendation results to the user;

[0015] Output customized recommendation results based on user feedback.

[0016] According to an embodiment of the present invention, constructing a multi-platform user portrait based on user data includes:

[0017] Processing the user data to obtain processed user data;

[0018] Conduct statistical analysis on the processed user data to generate multi-platform user portraits;

[0019] Receive user data from multiple platforms in real time and dynamically update user portraits on multiple platforms;

[0020] Among them, building a user behavior evaluation model including time measurement includes:

[0021] Preset different weights for different user behaviors on entities;

[0022] Calculate the score of the user's first behavior on the entity based on the weight;

[0023] According to the time sequence of the user's behavior on the entity, the first behavior score is decayed using the decay model to obtain the second behavior score;

[0024] The decayed behavior scores are stored according to the platform and entity type.

[0025] According to an embodiment of the present invention, the above attenuation model is expressed by equation (1) and equation (2):

[0026]

[0027] score desc =w*score act (2)

[0028] Wherein, formula (1) is used to calculate the attenuation coefficient, and formula (2) is used to calculate the second behavior score; where n is the number of times the user has performed the behavior, w is the attenuation coefficient, t is the position of the entity in the behavior sequence, and the behavior sequence is used to indicate the time sequence of the behavior. act Indicates that the first line is the score, score dxsc Indicates that the second line is the score.

[0029] According to an embodiment of the present invention, the above-mentioned construction of a multi-platform recall strategy includes:

[0030] Acquire attribute information of an entity of the first platform, wherein the attribute information includes text information and image information;

[0031] Processing the text information using a feature extraction model to obtain a semantic feature vector, wherein the dimension of the semantic feature vector is equal to a first preset value;

[0032] Processing the image information using the image feature extraction model to obtain an image feature vector, wherein the image feature vector and the semantic feature vector have the same vector dimension;

[0033] When the attribute information of the entity of the first platform is missing, the value of the feature vector corresponding to the missing attribute information is initialized to zero, wherein the dimension of the feature vector corresponding to the missing attribute information is the same as the first preset value;

[0034] The semantic feature vector and the image feature vector are concatenated to obtain a concatenated result, and the concatenated result is compressed by a fully connected network to obtain a multimodal semantic feature of the entity of the first platform;

[0035] Acquire multimodal semantic features of multiple entities under each of the multiple second platforms;

[0036] Calculating the cosine similarity scores of the multimodal semantic features of the entity of the first platform and the multimodal semantic features of each entity of the second platform, and sorting the cosine similarity scores in descending order to obtain a sorting result;

[0037] According to the ranking results, the entity of the second platform ranked first is set as a completely similar associated entity of the entity of the first platform; multiple entities of the second platform ranked within a range from the second to the second preset value are set as approximate associated entities of the entity of the first platform;

[0038] Completely similar associated entities and approximately associated entities of an entity of a first platform on other second platforms are obtained.

[0039] According to an embodiment of the present invention, the multimodal entity mapping recall method includes a recall strategy based on heat and a recall strategy based on semantic similarity; the multi-platform behavior fusion recall method includes a user-based cross-platform collaborative filtering recall strategy, an item-based cross-platform collaborative filtering recall strategy, a matrix decomposition-based cross-platform recall strategy, a tag-based cross-platform recall strategy, and a space-based cross-platform recall strategy;

[0040] When performing a recall strategy based on popularity, the popularity of the entity on the first platform is calculated; the popularity of each completely similar associated entity on the first platform is calculated on its second platform; the popularity of the entity on the first platform and its completely similar associated entity are added together to obtain the total popularity of the entity on the first platform; using the total popularity, the entities on the first platform and multiple second platforms are sorted in descending order of popularity, and the entities on the first platform and multiple second platforms whose rankings are within the third preset value range are used as the recall results of the recall strategy based on popularity;

[0041] When a recall strategy based on semantic similarity is performed, the approximate associated items of the entity of the first platform are mapped using the completely similar associated entities of the entity of the first platform to obtain multiple completely associated entities of the entity of the first platform on multiple second platforms; the multiple completely associated entities are sorted in descending order of similarity and duplicates are removed; and the completely associated entities of the multiple entities of the first platform whose rankings are within a fourth preset value range are obtained as the recall results based on the semantic similarity recall strategy;

[0042] When performing a user-based cross-platform collaborative filtering recall strategy, the user evaluation model is used to obtain the scores of the target user and multiple other users on the same entity; the improved Pearson similarity between the target user and each other user is calculated using the scores; multiple similar users of the target user are obtained based on the improved Pearson similarity; entities browsed by similar users are obtained based on the attribute information of similar users; entities browsed by the target user are filtered out, and the filtered entities are sorted in descending order using a heat algorithm, and entities ranked within the fifth preset value range are obtained as the recall results of the user-based cross-platform collaborative filtering recall strategy;

[0043] When performing a cross-platform collaborative filtering recall strategy based on items, the user evaluation model is used to obtain the target user's rating of the entity; the rating is used to obtain other entities browsed by the target user; the improved co-occurrence similarity is used to calculate the co-occurrence similarity values ​​of the entity and other entities on multiple platforms; other entities are sorted in descending order according to the co-occurrence similarity values, other entities browsed by the target user are filtered out, and multiple entities ranked within the sixth preset value range are used as the recall results of the cross-platform collaborative filtering recall strategy based on items;

[0044] When a cross-platform recall strategy based on matrix decomposition is performed, the user's behavior matrices on multiple platforms are concatenated, and the concatenated matrices are decomposed using an improved loss function. The decomposed matrices are used to score entities that the user has not browsed and the scores are sorted in descending order. According to the sorting results, entities ranked within the seventh preset value range are selected as the recall results of the cross-platform recall strategy based on matrix decomposition;

[0045] When implementing a tag-based cross-platform recall strategy, the user evaluation model is used to obtain the user's rating of the entity and obtain the entity with the highest score; the tag of the entity with the highest score is used to obtain other entities with the same tag as the entity with the highest score; the other entities are sorted in descending order of total popularity, the entities browsed by the user are filtered out, and the entities ranked within the eighth preset value range are used as the recall results of the tag-based cross-platform recall strategy;

[0046] When a cross-platform recall strategy is based on space, the user evaluation model is used to obtain users' ratings of entities and obtain the entity with the highest score. Based on the geographic location of the entity with the highest score, multiple entities around the geographic location of the entity with the highest score are screened out. Entities that have been browsed by users are filtered out, and multiple entities are sorted in descending order of total popularity. Entities ranked within the ninth preset value range are used as the recall results of the cross-platform recall strategy based on space.

[0047] According to an embodiment of the present invention, the output result of the above-mentioned multi-platform recall strategy is the union of a popularity-based recall strategy, a semantic similarity-based recall strategy, a user-based cross-platform collaborative filtering recall strategy, an item-based cross-platform collaborative filtering recall strategy, a matrix decomposition-based cross-platform recall strategy, a tag-based cross-platform recall strategy, and a space-based cross-platform recall strategy.

[0048] According to an embodiment of the present invention, the improved Pearson similarity is expressed by formula (3):

[0049]

[0050] Among them, r is the similarity; T is the number of platforms; t represents each platform; X t Y is the score given by A to the items viewed by users A and B; t is the score given by B to the items viewed by users A and B; N is the number of items viewed by users A and B; ∑X t Y t The product of the ratings of items viewed by users A and B is summed up;

[0051] The improved co-occurrence similarity is expressed by formula (4):

[0052]

[0053] Where T is the total number of platforms; |N t (i)∩N t (j)| represents the number of times item i and item j appear together in the user's browsing list on platform t; |N t (i)| represents the number of times item i appears in the user's browsing list on platform t; |N t (j)| represents the number of times item j appears in the user's browsing list on platform t;

[0054] The improved loss function is expressed by formula (5):

[0055]

[0056] Where R ij is the rating of item j by user i on a single platform, is the vector of user i in the user matrix after matrix decomposition, S j is the vector corresponding to item j in the item matrix after matrix decomposition; n is the number of users; m is the number of items.

[0057] According to an embodiment of the present invention, obtaining a recommendation result using a ranking model based on the output results of the multi-platform recall strategy includes:

[0058] Obtain user behavior characteristics and user attribute characteristics to form user portrait characteristics;

[0059] Construct training samples, where the samples include positive samples and negative samples;

[0060] Using training samples to train the sorting model;

[0061] The trained sorting model is used to sort the output results of the multi-platform recall strategy to obtain the recommendation results.

[0062] According to an embodiment of the present invention, outputting customized recommendation results according to the user's feedback results includes:

[0063] When the user is not satisfied with the recommendation result, the results generated by the multimodal entity mapping recall method and the multi-platform behavior fusion recall method are displayed to the user;

[0064] Generate customized recommendation results based on the data source, recall source and sorting algorithm selected by the user.

[0065] According to another aspect of the present invention, a customized recommendation system based on multi-platform user data is provided, comprising:

[0066] The collection module is used to collect user data from multiple platforms;

[0067] A first construction module is used to construct a multi-platform user portrait based on user data, wherein the multi-platform user portrait includes user behavior information and user attribute information;

[0068] The second construction module is used to construct a user behavior evaluation model including time measurement based on multi-platform user portraits, wherein the user behavior evaluation model is used to calculate the user's preference value for the item;

[0069] The third building module is used to build a multi-platform recall strategy according to the calculation results of the user behavior evaluation model, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method;

[0070] The recommendation module is used to obtain recommendation results using the ranking model based on the output results of the multi-platform recall strategy;

[0071] Feedback module, used to feed back recommendation results to users;

[0072] The customization module is used to output customized recommendation results based on user feedback.

[0073] The customized recommendation method based on multi-platform user data provided by the present invention can establish associations for related information in multiple portal platforms, solving the problem of loss and duplication of recommended information when making recommendations across multiple portal platforms; at the same time, the above recommendation method can combine the data advantages of multiple platforms to improve the diversity and accuracy of recommendation results. The customized recommendation system based on multi-platform user data provided by the present invention can improve the timeliness of recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flow chart of a customized recommendation method based on multi-platform user data according to an embodiment of the present invention;

[0075] Figure 2 is a flowchart of constructing a multi-platform user portrait according to an embodiment of the present invention;

[0076] Figure 3 is a flow chart of constructing a user behavior evaluation model including time measurement according to an embodiment of the present invention;

[0077] Figure 4 is a flow chart of building a multi-platform recall strategy according to an embodiment of the present invention;

[0078] Figure 5 is a flow chart of outputting customized recommendation results according to an embodiment of the present invention;

[0079] Figure 6 is a structural diagram of a customized recommendation system based on multi-platform user data according to an embodiment of the present invention;

[0080] Figure 7 is a workflow diagram of a recommendation system according to an embodiment of the present invention;

[0081] Figure 8 is an architectural diagram of a recommendation system according to an embodiment of the present invention;

[0082] Fig. 9 is a result display diagram of a recommendation system according to an embodiment of the present invention;

[0083] Fig.10 is a configuration diagram of a recommendation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0084] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0085] According to one aspect of the present invention, a customized recommendation method based on multi-platform user data is provided. Figure 1 The process shown includes steps S110 to S170.

[0086] In step S110, user data from multiple platforms is collected.

[0087] In step S120, a multi-platform user portrait is constructed based on the user data, wherein the multi-platform user portrait includes user behavior information and user attribute information.

[0088] Among them, taking three portal platforms (such as the knowledge graph platform, the map platform, and the mall platform) as examples, in the knowledge graph platform and the map platform, there are four types of user behaviors: click, browse, collect, and download; in the mall platform, there are six types of user behaviors: click, browse, collect, place an order, add to shopping cart, and evaluate.

[0089] In step S130, a user behavior evaluation model including time measurement is constructed based on the multi-platform user portraits, wherein the user behavior evaluation model is used to calculate the user's preference value for the entity.

[0090] The above entities include items of multiple categories, such as concepts on the knowledge graph platform, geographical areas on the map platform, and commodities on the mall platform.

[0091] The above user portrait can be constructed by counting the total number of behaviors of users on each platform towards entities or products, such as the total number of clicks, views and other data.

[0092] In step S140, a multi-platform recall strategy is constructed according to the calculation results of the user behavior evaluation model, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method.

[0093] In step S150, a ranking model is used to obtain a recommendation result according to the output result of the multi-platform recall strategy.

[0094] In step S160, the recommendation result is fed back to the user.

[0095] In step S170, customized recommendation results are output according to the user's feedback results.

[0096] The recommendation method disclosed in the above-mentioned embodiment of the present invention innovatively constructs a user behavior scoring system that includes time measurement, which solves the problem that the time impact cannot be measured in the scoring system of the recommendation system; it innovatively constructs a multi-platform recall strategy to solve the problem of loss and duplication of recommendation information when making recommendations across multiple portal platforms. At the same time, the multi-platform behavior fusion recall method can combine the data advantages of multiple platforms to improve the diversity and accuracy of recommendation results.

[0097] Figure 2 It is a flowchart of constructing a multi-platform user portrait according to an embodiment of the present invention, including steps S210 to S230.

[0098] According to an embodiment of the present invention, constructing a multi-platform user profile based on user data includes:

[0099] In step S210, the user data is processed to obtain processed user data.

[0100] In step S220, statistical analysis is performed on the processed user data to generate a multi-platform user portrait.

[0101] In step S230, user data from multiple platforms is received in real time, and user portraits from multiple platforms are dynamically updated. For example, user portraits may be updated regularly every day.

[0102] Figure 3 It is a flowchart of constructing a user behavior evaluation model including time measurement according to an embodiment of the present invention, including steps S310 to S340.

[0103] like Figure 3 As shown in Figure 2, constructing a user behavior evaluation model that includes time measurement includes:

[0104] In step S310, different weights are preset for different behaviors of the user on the entity.

[0105] In step S320, a score of the user's first behavior on the entity is calculated according to the weight.

[0106] In step S330, the first behavior score is decayed using a decay model according to the time sequence of the user's behavior on the entity to obtain a second behavior score.

[0107] In step S340, the attenuated behavior scores are stored according to the platform and entity types.

[0108] Constructing a user behavior evaluation model that includes time measurement can effectively improve the real-time performance of recommendation results. For example, in a shopping mall, if a user browsed T-shirts a week ago and hoodies an hour ago, the behavior scores of the two behaviors are the same, so the system believes that the user likes the two products to the same degree. However, the user should like the product that he viewed an hour ago more than the product that he viewed a week ago. Therefore, introducing a scoring system that includes time measurement makes the scoring more biased towards the user's recent behavior records, so that the recommendation results are real-time.

[0109] According to an embodiment of the present invention, the above attenuation model is expressed by equation (1) and equation (2):

[0110]

[0111] score desc =w*sCore act (2)

[0112] Wherein, formula (1) is used to calculate the attenuation coefficient, and formula (2) is used to calculate the second behavior score; where n is the number of times the user has performed the behavior, w is the attenuation coefficient, t is the position of the entity in the behavior sequence, and the behavior sequence is used to indicate the time sequence of the behavior. act Indicates that the first line is the score, score desc Indicates that the second line is the score.

[0113] The above decay model fully considers the time factor of user behavior and recalculates the weight of user behavior at different time points, which can better reflect the user's preference for the entity. For example, for a certain entity, assuming that the behaviors related to it in the platform are browsing, collecting, and downloading, after determining the importance of each behavior through human experience, the platform can obtain the behavior score of the behavior, as shown in formula (6):

[0114] score act =0.2*n brow +0.4*n fav +0.4*n down (6)

[0115] Among them, n brow is the number of times the user browses the entity or product, n fav Whether the user has collected the entity or product, n downThe number of times the user has downloaded the entity or product; in order to distinguish the impact of each behavior on the platform, the number of behaviors is multiplied by the importance weight coefficient. For example, in this platform, 0.2 is the weight coefficient for browsing behavior, and 0.4 is the weight coefficient for collection and download, which indicates that the importance of collection and download behavior on this platform should be higher than browsing behavior. In order to generate recommendation results for the behavior data of a single platform and for the behavior data of multiple platforms, the score is divided into two categories according to the platform and the type of item. desc Save the data in the format of "id-type-platform-score", where id is the data storage number; type is the data type; platform is the data source platform; score is the score. desc ,This storage format facilitates fast query when subsequently calculating the recall results.

[0116] Figure 4 It is a flowchart of constructing a multi-platform recall strategy according to an embodiment of the present invention, including steps S410 to S490.

[0117] In step S410, attribute information of an entity of the first platform is obtained, wherein the attribute information includes text information and image information.

[0118] In step S420, the text information is processed using a feature extraction model to obtain a semantic feature vector, wherein the dimension of the semantic feature vector is equal to a first preset value.

[0119] The text information can be processed using the BERT feature extraction model, where the dimension of the semantic feature vector is 1*768.

[0120] In step S430, the picture information is processed using the picture feature extraction model to obtain a picture feature vector, wherein the picture feature vector and the semantic feature vector have the same vector dimension.

[0121] The pre-trained VGG-19 image feature extraction model is used to process the image information to obtain an initial image feature vector, where the dimension of the initial image feature vector is 196*512. The first dimension of the initial image feature vector is averaged by columns and the second dimension is mapped using an existing fully connected network to obtain an image feature vector with the same dimension as the semantic feature vector, that is, the dimension of the processed image feature vector is 1*768.

[0122] In step S440, when the attribute information of the entity of the first platform is missing, the value of the feature vector corresponding to the missing attribute information is initialized to zero, wherein the dimension of the feature vector corresponding to the missing attribute information is the same as the first preset value.

[0123] For example, if an entity has no text information or picture information, the semantic feature vector or picture feature vector corresponding to the entity is set to a 1*768-dimensional zero-value vector by default, where the first preset value is 1*768 dimensions.

[0124] In step S450, the semantic feature vector and the image feature vector are concatenated to obtain a concatenation result, and the concatenation result is compressed using a fully connected network to obtain a multimodal semantic feature of the entity of the first platform.

[0125] The dimension of the above multimodal semantic features is 1*768, which is consistent with the dimension of the semantic feature vector and the image feature vector.

[0126] In step S460, multimodal semantic features of multiple entities under each of the multiple second platforms are obtained.

[0127] There are multiple second platforms, and each second platform has multiple entities. The multimodal semantic features of the multiple entities on each second platform are calculated respectively, so as to obtain the multimodal semantic features of the multiple entities on the multiple platforms.

[0128] In step S470, the cosine similarity scores of the multimodal semantic features of the entity of the first platform and the multimodal semantic features of each entity of the second platform are calculated, and the cosine similarity scores are sorted in descending order to obtain a sorting result.

[0129] In step S480, according to the sorting results, the entity of the second platform ranked first is set as a completely similar associated entity of the entity of the first platform; and multiple entities of the second platform ranked within a range from the second to the second preset value are set as approximate associated entities of the entity of the first platform.

[0130] Calculate the cosine similarity score between the entity of the first platform and multiple entities of one of the second platforms, wherein, for example, the entity or item with the highest similarity score is set as the complete similarity association X of the entity or item of the first platform most , set the top 10 items with the highest similarity as approximate association X related , the above second preset value is 10.

[0131] In step S490, completely similar associated entities and approximately associated entities of the entity of the first platform on other second platforms are obtained.

[0132] By repeating the above steps, completely similar associated entities and multiple approximately associated entities of the entity of the above first platform on each second platform can be obtained.

[0133] According to an embodiment of the present invention, the multimodal entity mapping recall method includes a recall strategy based on heat and a recall strategy based on semantic similarity; the multi-platform behavior fusion recall method includes a user-based cross-platform collaborative filtering recall strategy, an item-based cross-platform collaborative filtering recall strategy, a matrix decomposition-based cross-platform recall strategy, a tag-based cross-platform recall strategy, and a space-based cross-platform recall strategy;

[0134] When implementing a recall strategy based on popularity, calculate the popularity of entities on the first platform on the first platform; calculate the popularity of each completely similar associated entity of the entity on the first platform on its second platform; add the popularity values ​​of the entity on the first platform and its completely similar associated entities to obtain the total popularity value of the entities on the first platform; use the total popularity value to sort the entities on the first platform and multiple second platforms in descending order of popularity, and use the entities on the first platform and multiple second platforms whose rankings are within the third preset value range as the recall results of the recall strategy based on popularity.

[0135] For example, when performing a recall strategy based on popularity, the existing recall method based on popularity usually calculates the popularity value of the entity on a certain platform and obtains its recall result. However, the embodiment of the present invention provides a recall strategy based on popularity, using the completely similar association X obtained in the above embodiment. most Calculate the total popularity of entities or items in multiple portal platforms, so as to make related recommendations for entities or items on multiple platforms. The total popularity is expressed by formula (7):

[0136] heat all (X)=heat1+heat2+heat3 (7),

[0137] Among them, heat1, heat2, and heat3 are the popularity of a certain item in three different portal platforms; heat all (X) is the total heat of entities in the three portal platforms. all Sort the entities in each platform by heat. all The 20 entities with the highest values ​​(i.e. within the third preset value range) are taken as the final result D of this recall source. heat .

[0138] When a recall strategy based on semantic similarity is performed, the approximate associated items of the entity of the first platform are mapped using the completely similar associated entities of the entity of the first platform to obtain multiple completely associated entities of the entity of the first platform on multiple second platforms; the multiple completely associated entities are sorted in descending order of similarity and duplicates are removed; and the completely associated entities of multiple entities of the first platform whose rankings are within a fourth preset value range are obtained as the recall results based on the semantic similarity recall strategy.

[0139] The existing recall strategy based on semantic similarity directly uses the deep learning model to obtain the entity or item with the highest text similarity on the same platform as the recall result. However, the recall strategy based on semantic similarity disclosed in the embodiment of the present invention uses the approximate association X obtained in the above embodiment. related Calculate the result D of this recall source semantic , so as to make recommendations based on the multimodal semantic similarity between entities or items in each system. For example, when a user browses an entity or product on a certain platform, the approximate association X of the currently browsed entity on multiple other platforms can be obtained. related Assuming the total number of platforms is N, the approximate total number of associations obtained for the currently browsed entity or product is X related_all is 10*(N-1); 10*(N-1) approximate associations are used as completely similar associations X most By mapping, we can get 10*(N-1) fully related entities in the user's platform. These entities may be repeated, and after deduplication, they can be used as the result of this recall source D semantic If D semantic If there are too many entities, they can be deleted by similarity sorting, and the n entities with the highest similarity (i.e., entities within the fourth preset value range) are selected as the recommendation results of this recall source.

[0140] When implementing a user-based cross-platform collaborative filtering recall strategy, a user evaluation model is used to obtain the ratings of a target user and multiple other users on the same entity; the ratings are used to calculate the improved Pearson similarity between the target user and each other user; based on the improved Pearson similarity, multiple similar users of the target user are obtained; based on the attribute information of the similar users, entities browsed by the similar users are obtained; entities browsed by the target user are filtered out, and the filtered entities are sorted in descending order using a heat algorithm, and entities ranked within the fifth preset value range are obtained as the recall results of the user-based cross-platform collaborative filtering recall strategy.

[0141] In the existing recommendation methods, the collaborative filtering algorithm only uses the user's data on one platform to find similar users, and then recalls the entities or products that the similar users are interested in and that the user has not browsed. This method cannot combine the advantages of multi-platform data, resulting in reduced diversity and accuracy of recommendations. The user-based cross-platform collaborative filtering recall strategy provided in the embodiment of the present invention can use multi-platform data to more effectively capture user behavior, thereby finding the true nearest neighbor user and giving accurate recommendations. Through the above-mentioned user-based cross-platform collaborative filtering recall strategy, the improved Pearson similarity based on multi-platform user behavior is calculated, and then the 40 entities and products with the highest similarity are taken, and then the items browsed by the user are filtered, and the 20 entities or products with the highest popularity (i.e., entities within the fifth preset value range) are selected as the result of this recall source D user_rec

[0142] When implementing an item-based cross-platform collaborative filtering recall strategy, a user evaluation model is used to obtain the target user's rating of an entity; the rating is used to obtain other entities browsed by the target user; the improved co-occurrence similarity is used to calculate the co-occurrence similarity values ​​of entities and other entities on multiple platforms; other entities are sorted in descending order according to the co-occurrence similarity values, other entities browsed by the target user are filtered out, and multiple entities ranked within the sixth preset value range are used as the recall results of the item-based cross-platform collaborative filtering recall strategy.

[0143] The existing cross-platform collaborative filtering recall strategy based on items only uses single-platform data to calculate item similarity, which reduces the diversity and accuracy of recommendation results. The multi-platform behavior fusion recall strategy based on item collaborative filtering provided by the embodiment of the present invention can better find other entities similar to a certain entity through improved co-occurrence similarity, and at the same time filter the items browsed by the user and then intercept the 20 most popular entities or commodities (i.e., multiple entities within the sixth preset value range) as the result of this recall source D item_rec .

[0144] When implementing a cross-platform recall strategy based on matrix decomposition, the user's behavior matrices on multiple platforms are spliced, and the spliced ​​matrices are decomposed using an improved loss function. The decomposed matrices are used to score entities that the user has not browsed and the scores are sorted in descending order. According to the sorting results, entities ranked within the seventh preset value range are selected as the recall results of the cross-platform recall strategy based on matrix decomposition.

[0145] Assume that R1 is the user's behavior matrix on the knowledge graph platform; R2 is the user's behavior matrix on the mall platform; and R3 is the user's behavior matrix on the map platform. In order to obtain the user's behavior characteristics on multiple platforms, the corresponding user's behavior records on multiple platforms are spliced, and the data without behavior is filled with 0. R′ is the matrix after the splicing of multiple platform behaviors, which is expressed as (8):

[0146]

[0147] The loss function in the prior art is expressed by formula (9):

[0148]

[0149] Where R ij is the rating of item j by user i on a single platform, is the vector of user i in the user matrix after matrix decomposition, S j is the vector corresponding to item j in the item matrix after matrix decomposition; n is the number of users; m is the number of items; Different from the existing matrix decomposition strategy, the cross-platform recall strategy based on matrix decomposition provided by the embodiment of the present invention can calculate matrix decomposition for multiple platforms at the same time. After matrix decomposition, the system will predict and score the items that the user has not seen, and intercept the first 20 entities or commodities (i.e., entities within the seventh preset value range) as the result D of this recall source matrix .

[0150] When implementing a tag-based cross-platform recall strategy, the user evaluation model is used to obtain users' ratings of entities and obtain the entity with the highest score; the label of the entity with the highest score is used to obtain other entities with the same label as the entity with the highest score; other entities are sorted in descending order of total popularity, and entities that have been browsed by users are filtered out, and entities ranked within the eighth preset value range are used as the recall results of the tag-based cross-platform recall strategy.

[0151] In the sorting results, after filtering the items browsed by the user, the top 20 items with the highest popularity (i.e., entities within the eighth preset value range) are obtained as the result D of this recall source. tag .

[0152] When a cross-platform recall strategy is based on space, the user evaluation model is used to obtain users' ratings of entities and obtain the entity with the highest score. Based on the geographic location of the entity with the highest score, multiple entities around the geographic location of the entity with the highest score are screened out. Entities that have been browsed by users are filtered out, and multiple entities are sorted in descending order of total popularity. Entities ranked within the ninth preset value range are used as the recall results of the cross-platform recall strategy based on space.

[0153] First, obtain an entity or item that is most popular among users on all portal platforms; if the item or entity has a geographical location stored in the system, 50 candidate entities or commodities around its location are selected based on the location; secondly, filter the items that the user has browsed, and use the popularity score to intercept the 20 most popular entities (i.e., within the ninth preset value range) as the result of the space-based cross-platform recall strategy D space .

[0154] Through the above embodiments, by utilizing the above multiple recall strategies and combining user data from multiple platforms, more accurate and comprehensive recommendation results can be obtained.

[0155] According to an embodiment of the present invention, the output result of the above-mentioned multi-platform recall strategy is the union of a popularity-based recall strategy, a semantic similarity-based recall strategy, a user-based cross-platform collaborative filtering recall strategy, an item-based cross-platform collaborative filtering recall strategy, a matrix decomposition-based cross-platform recall strategy, a tag-based cross-platform recall strategy, and a space-based cross-platform recall strategy.

[0156] Using formula (10), we can get the total candidate set D in the generation and recall phase: cand :

[0157] D cand =D heat ∪D semantic ∪D user_rec ∪D item_rec ∪D matrix ∪D tag ∪D space (10).

[0158] The total candidate set D cand That is the output result of the above multi-platform recall strategy.

[0159] According to an embodiment of the present invention, the improved Pearson similarity is expressed by formula (3):

[0160]

[0161] Among them, r is the similarity; T is the number of platforms; t represents each platform; X t Y is the score given by A to the items viewed by users A and B; t is the score given by B to the items viewed by users A and B; N is the number of items viewed by users A and B; ∑X t Y t The product of the ratings of items viewed by users A and B is summed up;

[0162] The improved co-occurrence similarity is expressed by formula (4):

[0163]

[0164] Where T is the total number of platforms; |N t (i)∩N t (j)| represents the number of times item i and item j appear together in the user's browsing list on platform t; |N t (i)| represents the number of times item i appears in the user's browsing list on platform t; |N t (j)| represents the number of times item j appears in the user's browsing list on platform t;

[0165] The improved loss function is expressed by formula (5):

[0166]

[0167] Where R′ ij is the rating of item j by user i on a single platform, is the vector of user i in the user matrix after matrix decomposition, S j is the vector corresponding to item j in the item matrix after matrix decomposition; n is the number of users; m is the number of items.

[0168] Figure 4 4 is a flowchart of a recommendation result according to an embodiment of the present invention, including steps S410 to S440.

[0169] like Figure 4 As shown in the figure, based on the output of the multi-platform recall strategy, the recommendation results obtained by using the ranking model include:

[0170] In step S410, user behavior characteristics and user attribute characteristics are obtained to form user portrait characteristics.

[0171] First, based on the feature construction method, one-hot features are used to construct user behavior features and user attribute features and then concatenated; secondly, sparse features and numerical features are constructed using the feature construction method to obtain user attribute features. act , where sparse features are used to train the logistic regression sorting model, and numerical features are used to train the XGBoost sorting model; again, the user's age, department, birthplace, living area, and favorite product type features are selected to obtain the user attribute feature feature prop ; act and feature prop After splicing, we can get the user portrait feature.

[0172] In step S420, training samples are constructed, where the samples include positive samples and negative samples.

[0173] In the above steps, the positive and negative examples of the samples are distinguished by counting the user's browsing time. For example, if a user browses an item for more than 3 seconds, it is considered a positive example, and if the browsing time is less than 3 seconds, it is considered a negative example.

[0174] In step S430, the sorting model is trained using the training samples;

[0175] In step S440, the trained sorting model is used to sort the output results of the multi-platform recall strategy to obtain a recommendation result.

[0176] In the above step S440, the trained sorting model is used to sort the total candidate set D of the recall source cand , score and sort, and obtain the recommended results.

[0177] The above recommendation method utilizes a ranking model, fully utilizes the user portrait provided by the embodiment of the present invention, and through sample training, enables the generated recommendation results to better meet the needs of users.

[0178] Figure 5 It is a flowchart of outputting customized recommendation results according to an embodiment of the present invention, including steps S510 to S520.

[0179] like Figure 5 As shown in the figure, according to the user's feedback, the output customized recommendation results include:

[0180] In step S510, when the user is not satisfied with the recommendation result, the results generated by the multimodal entity mapping recall method and the multi-platform behavior fusion recall method are displayed to the user;

[0181] In step S520, customized recommendation results are generated according to the data source, recall source and sorting algorithm selected by the user.

[0182] By customizing the recommendation results generated by the above recommendation method, the generated recommendation results can be closer to the user's needs, giving full play to the role of the user in the above recommendation method, making the recommendation results more accurate.

[0183] According to another aspect of the present invention, a customized recommendation system based on multi-platform user data is provided. Figure 6 As shown, the recommendation system 600 includes a collection module 610, a first construction module 620, a second construction module 630, a third construction module 640, a recommendation module 650, a feedback module 660 and a customization module 670; wherein:

[0184] A collection module 610 is used to collect user data from multiple platforms;

[0185] A first construction module 620 is used to construct a multi-platform user portrait based on user data, wherein the multi-platform user portrait includes user behavior information and user attribute information;

[0186] The second construction module 630 is used to construct a user behavior evaluation model including time measurement based on multi-platform user portraits, wherein the user behavior evaluation model is used to calculate the user's preference value for the item;

[0187] The third construction module 640 is used to construct a multi-platform recall strategy according to the calculation results of the user behavior evaluation model, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method;

[0188] A recommendation module 650 is used to obtain a recommendation result using a ranking model according to the output result of the multi-platform recall strategy;

[0189] Feedback module 660, used to feed back the recommendation results to the user;

[0190] The customization module 670 is used to output customized recommendation results based on the user's feedback results.

[0191] The above recommendation system combines user data from multiple platforms, builds portraits and user evaluation models, adopts multiple recommendation strategies to process user data and fully considers user personalized operations, generating rich and reliable recommendation results for users, greatly improving the user experience.

[0192] Figure 7 is a workflow diagram of a recommendation system according to an embodiment of the present invention, Figure 8 is an architecture diagram of a recommendation system according to an embodiment of the present invention; Figure 7 and Figure 8 The above recommendation system is further described in detail.

[0193] like Figure 7 As shown, the above system first collects user information from multiple platforms, and constructs multi-platform user portraits and user evaluation models including time measurements, and then constructs a multi-platform recall strategy, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method. The sorting model is used to sort and filter the recommendation results generated by the multi-platform recall strategy, and the relevant recommendation results are fed back to the user. When the user is satisfied with the current result, the relevant recommendation result is output. If the user is not satisfied with the current result, the user customization process is entered to generate a result that satisfies the user. Figure 8 The detailed architecture diagram of the above recommendation system is further shown; Figure 8As shown, user data includes user behavior information and user attribute information, which are used to extract user features; entities or commodities include text information and image information of entities or commodities, which are used to extract multimodal entity features; multi-platform behavior fusion recall methods include user-based cross-platform collaborative filtering recall strategies, item-based cross-platform collaborative filtering recall strategies, matrix decomposition-based cross-platform recall strategies, label-based cross-platform recall strategies, and space-based cross-platform recall strategies, and multi-modal entity mapping recall methods include semantic similarity-based recall strategies and heat-based recall strategies; the above-mentioned multi-modal entity mapping recall methods and multi-platform behavior fusion recall methods are used to generate recall results of multiple recall sources; the recall results are selected using a ranking model to generate the final recommendation results, wherein the ranking model includes a gradient boosting tree model and a logistic regression model.

[0194] Combine the following Fig. 9 and Fig.10 The process of the above recommendation system outputting customized recommendation results is described in detail.

[0195] Fig. 9 is a result display diagram of the recommendation system according to an embodiment of the present invention. Fig. 9 As shown, the recommendation system supports users to select recall strategies and sorting algorithms by clicking; when users are not satisfied with the recommendation results, the recommendation system will display the reasons for the recall results, so that users can understand the reasons for the recall results and delete or add recommendation strategies supported by the recommendation method.

[0196] Fig.10 is a configuration diagram of a recommendation system according to an embodiment of the present invention. Fig.10 As shown in Figure 1, users can customize the recommendation results through the relevant configuration of the above recommendation system, where the customizable parts include the source of user data, the selection of recall sources, and the selection of sorting algorithms. When users customize the data source, the system uses the user data P of all platforms by default. all Generate recommendation results. Take three user platforms as an example. The knowledge graph platform data is P1, the mall platform data is P2, and the map platform data is P3. The total data source of the system is: P all =P1∪P2∪P3 Assuming that the user does not want to receive the recommendation results from the mall platform, the data of the mall platform can be cancelled. At this time, the data source is: P all =P1∪P3. When the user customizes the recall source, the system analyzes Figure 6 The reason for the recall result shown in the figure is that the user can delete or add the recall sources supported by the system to change the total candidate set data and thus change the final recommendation result. For example, after the user cancels the recall source based on heat and the recall source based on matrix decomposition, the total candidate set D cand Change to D cand =Dheat ∪Ds emantc ∪D user_rec ∪D item_rec ∪D tag ∪D space When users customize the sorting algorithm, the system supports the selection of two different sorting algorithms, including the logistic regression model and the gradient boosting tree algorithm. The logistic regression algorithm has a faster training and parameter update speed, and the gradient boosting tree algorithm has a more accurate sorting result. After the above customization operation, the system outputs the recommendation results customized by the user.

[0197] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A customized recommendation method based on multi-platform user data, include: Collecting user data from multiple platforms, including a knowledge graph platform, a map platform, and a shopping mall platform; Based on the user data, construct a multi-platform user portrait, wherein the multi-platform user portrait includes user behavior information and user attribute information, and the user behavior includes click, browse, collect and download behaviors on the knowledge graph platform and the map platform, and click, browse, collect, place an order, add to shopping cart and evaluate behaviors on the mall platform; Based on the multi-platform user portraits, a user behavior evaluation model including time measurement is constructed, wherein the user behavior evaluation model is used to calculate the user's preference value for an entity, wherein the entity includes a multi-category item consisting of concepts on a knowledge graph platform, geographic areas on a map platform, and commodities on a shopping mall platform; According to the calculation results of the user behavior evaluation model, a multi-platform recall strategy is constructed, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method; According to the output results of the multi-platform recall strategy, a ranking model is used to obtain a recommendation result; Feedback the recommendation result to the user; Outputting customized recommendation results according to the user's feedback results; Among them, building a multi-platform recall strategy includes: obtaining attribute information of an entity of a first platform, wherein the attribute information includes text information and image information; using a feature extraction model to process the text information to obtain a semantic feature vector, wherein the dimension of the semantic feature vector is equal to a first preset value; using an image feature extraction model to process the image information to obtain an image feature vector, wherein the image feature vector and the semantic feature vector have the same vector dimension; when the attribute information of an entity of the first platform is missing, initializing the value of the feature vector corresponding to the missing attribute information to zero, wherein the dimension of the feature vector corresponding to the missing attribute information is the same as the first preset value; splicing the semantic feature vector and the image feature vector to obtain a splicing result, and performing a multi-platform recall strategy on the entity of the first platform. The splicing result is compressed by a fully connected network to obtain the multimodal semantic features of the entity of the first platform; the multimodal semantic features of multiple entities under each of the multiple second platforms are obtained; the cosine similarity scores of the multimodal semantic features of the entity of the first platform and the multimodal semantic features of each entity of the second platform are calculated, and the cosine similarity scores are sorted in descending order to obtain the sorting results; according to the sorting results, the entity of the second platform ranked first is set as the completely similar associated entity of the entity of the first platform; multiple entities of the second platform ranked within the range from the second to the second preset value are set as approximate associated entities of the entity of the first platform; the completely similar associated entities and the approximate associated entities of the entity of the first platform on other second platforms are obtained; Wherein, the multimodal entity mapping recall method includes a recall strategy based on heat and a recall strategy based on semantic similarity; When the recall strategy based on popularity is performed, the popularity value of the entity of the first platform on the first platform is calculated; the popularity value of each completely similar associated entity of the entity of the first platform on its second platform is calculated; the popularity value of the entity of the first platform and its completely similar associated entity are added together to obtain the total popularity value of the entity of the first platform; using the total popularity value, the entities on the first platform and multiple second platforms are sorted in descending order of popularity, and the entities on the first platform and multiple second platforms whose rankings are within the third preset value range are used as the recall results of the recall strategy based on popularity; When the recall strategy based on semantic similarity is performed, the approximate associated items of the entity of the first platform are mapped using the completely similar associated entities of the entity of the first platform to obtain multiple completely associated entities of the entity of the first platform on multiple second platforms; the multiple completely associated entities are sorted in descending order of similarity and duplicates are removed; and the completely associated entities of multiple entities of the first platform whose rankings are within a fourth preset value range are obtained as the recall result of the recall strategy based on semantic similarity; Wherein, constructing a multi-platform user portrait according to the user data includes: Processing the user data to obtain processed user data; Performing statistical analysis on the processed user data to generate the multi-platform user portrait; Receive the user data from multiple platforms in real time and dynamically update the user portraits of the multiple platforms; The step of constructing a user behavior evaluation model including time measurement includes: Presetting different weights for different behaviors of the user on the entity; Calculate the first behavior score of the user on the entity according to the weight; According to the time sequence of the user's behavior on the entity, the first behavior score is decayed by using a decay model to obtain a second behavior score; storing the attenuated behavior score according to the platform and the type of the entity; The attenuation model is expressed by equation (1) and equation (2): (1), (2), Wherein, formula (1) is used to calculate the attenuation coefficient, and formula (2) is used to calculate the second behavior score; where n is the number of times the user has performed the behavior, w is the attenuation coefficient, and t is the position of the entity in the behavior sequence. The behavior sequence is used to represent the time sequence of the behavior. represents the score of the first behavior, Indicates the second behavior is a score; Wherein, outputting customized recommendation results according to the user's feedback results includes: When the user is not satisfied with the recommendation result, the results generated by the multimodal entity mapping recall method and the multi-platform behavior fusion recall method are displayed to the user; Generate customized recommendation results based on the data source, recall source and sorting algorithm selected by the user.

2. The method according to claim 1, in, The multi-platform behavior fusion recall method includes a user-based cross-platform collaborative filtering recall strategy, an item-based cross-platform collaborative filtering recall strategy, a matrix decomposition-based cross-platform recall strategy, a tag-based cross-platform recall strategy, and a space-based cross-platform recall strategy; When performing a user-based cross-platform collaborative filtering recall strategy, the user evaluation model is used to obtain the scores of the target user and multiple other users on the same entity; the scores are used to calculate the improved Pearson similarity between the target user and each of the other users; According to the improved Pearson similarity, obtaining multiple similar users of the target user; According to the attribute information of the similar users, obtaining entities browsed by the similar users; Filter out the entities browsed by the target user, sort the filtered entities in descending order using a heat algorithm, and obtain entities whose rankings are within a fifth preset value range as recall results of a user-based cross-platform collaborative filtering recall strategy; When performing a cross-platform collaborative filtering recall strategy based on items, the user evaluation model is used to obtain the target user's rating of the entity; the rating is used to obtain other entities browsed by the target user; the improved co-occurrence similarity is used to calculate the co-occurrence similarity values ​​of the entity and the other entities on multiple platforms; the other entities are sorted in descending order according to the co-occurrence similarity values, the other entities browsed by the target user are filtered out, and the multiple entities ranked within the sixth preset value range are used as the recall results of the cross-platform collaborative filtering recall strategy based on items; When a cross-platform recall strategy based on matrix decomposition is performed, the behavior matrices of the user on multiple platforms are spliced, and the spliced ​​matrices are decomposed using an improved loss function, and the decomposed matrices are used to score entities that the user has not browsed and the scores are sorted in descending order, and according to the sorting result, entities whose rankings are within the seventh preset value range are selected as the recall result of the cross-platform recall strategy based on matrix decomposition; When performing a tag-based cross-platform recall strategy, the user evaluation model is used to obtain the user's score for the entity, and the entity with the highest score is obtained; the tag of the entity with the highest score is used to obtain other entities with the same tag as the entity with the highest score; the other entities are sorted in descending order of total popularity, and the entities browsed by the user are filtered out, and the entities ranked within the eighth preset value range are used as the recall results of the tag-based cross-platform recall strategy; When the cross-platform recall strategy is based on space, the user evaluation model is used to obtain the user's score on the entity, and the entity with the highest score is obtained; according to the geographical location of the entity with the highest score, multiple entities around the geographical location of the entity with the highest score are screened out; Entities browsed by the user are filtered out, the multiple entities are sorted in descending order of total popularity, and entities ranked within a ninth preset value range are used as recall results of the space-based cross-platform recall strategy.

3. The method according to claim 2, in, The output result of the multi-platform recall strategy is the union of a heat-based recall strategy, a semantic similarity-based recall strategy, a user-based cross-platform collaborative filtering recall strategy, an item-based cross-platform collaborative filtering recall strategy, a matrix decomposition-based cross-platform recall strategy, a tag-based cross-platform recall strategy, and a space-based cross-platform recall strategy.

4. The method according to claim 2, in, The improved Pearson similarity is expressed by formula (3): (3), Among them, r is the similarity; T is the number of platforms; t represents each platform; Give A the scores for the items that users A and B have viewed together; is the score given by B to the items viewed by users A and B together; N is the number of items viewed by users A and B together; is the sum of the product of the scores of the items viewed by users A and B together; The improved co-occurrence similarity is expressed by formula (4): (4), Where T is the total number of platforms; represents the number of times item i and item j appear together in the user's browsing list on platform t; Represents the number of times item i appears in the user's browsing list on platform t; represents the number of times item j appears in the user's browsing list on platform t; The improved loss function is expressed by formula (5): (5), in is the rating of item j by user i on a single platform, is the vector of user i in the user matrix after matrix decomposition, is the vector corresponding to item j in the item matrix after matrix decomposition; n is the number of users; m is the number of items.

5. The method according to claim 1, wherein the ranking model is used to obtain the recommendation result according to the output result of the multi-platform recall strategy include: Acquire the user behavior characteristics and the user attribute characteristics to form the user portrait characteristics; Constructing training samples, wherein the samples include positive samples and negative samples; Using the training samples to train the sorting model; The trained sorting model is used to sort the output results of the multi-platform recall strategy to obtain the recommendation result.

6. A customized recommendation system based on multi-platform user data, include: A collection module, used to collect user data from multiple platforms, wherein the platforms include a knowledge graph platform, a map platform, and a mall platform; A first construction module is used to construct a multi-platform user portrait based on the user data, wherein the multi-platform user portrait includes user behavior information and user attribute information, and the user behavior includes click, browse, collect and download behaviors on the knowledge graph platform and the map platform, and click, browse, collect, place an order, add to shopping cart and evaluate behaviors on the mall platform; A second construction module is used to construct a user behavior evaluation model including time measurement based on the multi-platform user portraits, wherein the user behavior evaluation model is used to calculate the user's preference value for items, and the entities include multi-category items consisting of concepts on the knowledge graph platform, geographical areas on the map platform, and commodities on the mall platform; A third construction module is used to construct a multi-platform recall strategy according to the calculation results of the user behavior evaluation model, wherein the multi-platform recall strategy includes a multi-modal entity mapping recall method and a multi-platform behavior fusion recall method; A recommendation module, used to obtain a recommendation result using a ranking model according to the output result of the multi-platform recall strategy; A feedback module, used for feeding back the recommendation result to the user; A customization module, used to output customized recommendation results according to the feedback results of the user; Among them, building a multi-platform recall strategy includes: Acquire attribute information of an entity of the first platform, wherein the attribute information includes text information and picture information; Processing the text information using a feature extraction model to obtain a semantic feature vector, wherein a dimension of the semantic feature vector is equal to a first preset value; Processing the image information using an image feature extraction model to obtain an image feature vector, wherein the image feature vector and the semantic feature vector have the same vector dimension; When the attribute information of the entity of the first platform is missing, initializing the value of the feature vector corresponding to the missing attribute information to zero, wherein the dimension of the feature vector corresponding to the missing attribute information is the same as the first preset value; splicing the semantic feature vector and the image feature vector to obtain a splicing result, and performing full-connection network compression on the splicing result to obtain multimodal semantic features of the entity of the first platform; Acquire multimodal semantic features of multiple entities under each of the multiple second platforms; Calculating the cosine similarity scores of the multimodal semantic features of the entity of the first platform and the multimodal semantic features of each entity of the second platform, and sorting the cosine similarity scores in descending order to obtain a sorting result; According to the ranking result, the entity of the second platform ranked first is set as a completely similar associated entity of the entity of the first platform; and multiple entities of the second platform ranked within a range from the second to a second preset value are set as approximate associated entities of the entity of the first platform; Acquire the completely similar associated entities and the approximately associated entities of the entity of the first platform on other second platforms; Wherein, the multimodal entity mapping recall method includes a recall strategy based on heat and a recall strategy based on semantic similarity; When the recall strategy based on popularity is performed, the popularity value of the entity of the first platform on the first platform is calculated; the popularity value of each completely similar associated entity of the entity of the first platform on its second platform is calculated; the popularity value of the entity of the first platform and its completely similar associated entity are added together to obtain the total popularity value of the entity of the first platform; using the total popularity value, the entities on the first platform and multiple second platforms are sorted in descending order of popularity, and the entities on the first platform and multiple second platforms whose rankings are within the third preset value range are used as the recall results of the recall strategy based on popularity; When the recall strategy based on semantic similarity is performed, the approximate associated items of the entity of the first platform are mapped using the completely similar associated entities of the entity of the first platform to obtain multiple completely associated entities of the entity of the first platform on multiple second platforms; the multiple completely associated entities are sorted in descending order of similarity and duplicates are removed; and the completely associated entities of multiple entities of the first platform whose rankings are within a fourth preset value range are obtained as the recall result of the recall strategy based on semantic similarity; Wherein, constructing a multi-platform user portrait according to the user data includes: Processing the user data to obtain processed user data; Performing statistical analysis on the processed user data to generate the multi-platform user portrait; Receive the user data from multiple platforms in real time and dynamically update the user portraits of the multiple platforms; The step of constructing a user behavior evaluation model including time measurement includes: Presetting different weights for different behaviors of the user on the entity; Calculate the first behavior score of the user on the entity according to the weight; According to the time sequence of the user's behavior on the entity, the first behavior score is decayed by using a decay model to obtain a second behavior score; storing the attenuated behavior score according to the platform and the type of the entity; The attenuation model is expressed by equation (1) and equation (2): (1), (2), Wherein, formula (1) is used to calculate the attenuation coefficient, and formula (2) is used to calculate the second behavior score; where n is the number of times the user has performed the behavior, w is the attenuation coefficient, and t is the position of the entity in the behavior sequence. The behavior sequence is used to represent the time sequence of the behavior. represents the score of the first behavior, Indicates the second behavior is a score; Wherein, outputting customized recommendation results according to the user's feedback results includes: When the user is not satisfied with the recommendation result, the results generated by the multimodal entity mapping recall method and the multi-platform behavior fusion recall method are displayed to the user; Generate customized recommendation results based on the data source, recall source and sorting algorithm selected by the user.