A method and device for matchmaking and friend-making recommendations based on image content analysis

Abstract

The present invention provides a method for recommending marriage and dating based on image content analysis. The method for recommending marriage and dating based on image content analysis includes the following steps: S1. Use a publicly available dataset to train a face recognition model and a face detection model as extractors of image features, and offline extract the image features of user photos; S2. According to user behavior data, offline train ALS as a recall model to generate a list of users to be recommended for each user. The method for recommending marriage and dating based on image content analysis provided by the present invention adds image features, can effectively capture users' preferences for image content, increases the dimension of user portraits, ensures the accuracy of the model for recommended content, and thus efficiently and accurately recommends a matching partner for users, improving the efficiency of users' partner selection.

Description

Technical Field

[0001] The present invention relates to the technical field of matchmaking and dating recommendation, and particularly to a matchmaking and dating recommendation method and device based on image content analysis. Background Art

[0002] With the continuous development of human life, marriage, as an essential life event that people must experience, is generated by the mutual love between men and women. Due to work needs, people's opportunities to travel have decreased, and it is difficult for the opposite-sex friends they meet to be accepted by both parties, resulting in a sharp decline in the number of love affairs and marriages. Therefore, using online dating has become the future development trend of new dating methods.

[0003] Matchmaking and dating websites provide a platform for single men and women to meet and get to know each other. However, they also make single men and women overwhelmed by the vast amount of information and various choices. In order to accurately find suitable potential partners for users, a recommendation system has emerged. The matchmaking and dating recommendation system is based on the mate selection needs of Internet users and uses recommendation algorithms for filtering and screening, so as to push potential partners that users may be interested in to them.

[0004] Traditional matchmaking and dating recommendations mainly match users based on information such as user attributes and user behavior, and fail to effectively integrate multimodal data such as user pictures and texts for association and recommendation. In the scenario of matchmaking and dating, visual elements such as appearance and looks often have a higher priority in the mate selection process. Therefore, adding image content analysis to matchmaking and dating recommendations and judging the current needs of users based on the similarity of image content is an important way to improve the matching degree of matchmaking users.

[0005] Therefore, it is necessary to provide a matchmaking and dating recommendation method and device based on image content analysis to solve the above technical problems. Summary of the Invention

[0006] The present invention provides a matchmaking and dating recommendation method based on image content analysis, which solves the problem that the integration degree of user information in online matchmaking and dating recommendations needs to be improved.

[0007] To solve the above technical problems, the matchmaking and dating recommendation method based on image content analysis provided by the present invention includes the following steps:

[0008] S1. Use a publicly available dataset to train a face recognition model and a face detection model as image feature extractors, and offline extract the image features of user photos;

[0009] S2. According to user behavior data, offline train ALS as a recall model to generate a list of users to be recommended for each user;

[0010] S3. Based on the photos of the opposite sex that the user likes and the facial image features of the photos, recommend similar-looking users to each user to expand the list of users to be recommended, and re-rank the recalled users by comprehensively considering the facial similarity and the ALS model score;

[0011] S4, obtain user basic attributes, mate selection conditions, image feature data, user click data, and train Wide & Deep model for online sorting;

[0012] S5. With the help of Web service technology, we provide user recommendation service, which can realize the process of ALS plus image feature recall, W&D model sorting, and business rule filtering to achieve personalized and accurate recommendations for each user.

[0013] Preferably, the image feature extractor in step S1 includes a face detection model and a face recognition model, wherein the face detection model uses the scrfd model, which obtains the optimal network structure through a network search method under the premise of a fixed amount of calculation, and ensures that the backbone, neck, and head parts are structurally balanced, and improves the accuracy of model detection through the strategies of training sample balance, ATSS, and DIOU to obtain a training model.

[0014] Preferably, the face recognition model adopts the insightface structure. The insightface network improves the inter-class separability of the spatial distribution of facial features by adding margin to the loss function, while strengthening the intra-class compactness and inter-class differences, thereby improving the accuracy of face recognition. In order to ensure that the model has better real-time performance, the backbone network uses a lightweight mobilenetv2 structure.

[0015] Preferably, the extractor is used to extract facial features of photos uploaded by users, and the results are stored in a hive table for easy reading during recall.

[0016] Preferably, in step S3, the recall of users adopts a similar face recall strategy, recommends photos similar to the user's favorite user to the user, and performs PQ encoding on the facial features.

[0017] Preferably, the PQ coding method is to divide the 512-dimensional facial features into N-dimensional subspaces, and each subspace is clustered into M categories using the Kmeans algorithm. The clustered category ID is used as the PQ code. Since the clustering centers are distributed in the feature space, the feature codes near the PQ codes are obtained, and matching is performed based on the adjacent feature code values ​​to improve the speed of global search.

[0018] Preferably, during the offline search process, the PQ codes of the user photos and the adjacent PQ codes are respectively written into the hive table. The adjacent PQ codes of the photos liked by the user and the PQ codes of the photos to be searched are used as the matching rules, and with the computing power of Spark, a fast offline global search function is realized. Combining the scoring of ALS and the face similarity in step S2, a rough ranking of the recall results is performed.

[0019] Preferably, in step S4, the user's basic attributes are obtained. The basic attributes include but are not limited to gender, age, height, constellation, hobbies, income, the province, county, and city where the user lives and works for a long time, mate selection conditions, and face value score data are integrated. The user can swipe left and right on the card recommendation page, and the Wide&Deep model is trained. The optimal offline model under the current conditions is obtained by comprehensively considering indicators such as Accuracy, Recall, F1-score, AUC indicator, and gAUC indicator. In order to improve the response speed of the recommendation server, the model parameters trained offline and the encoding results of user features are cached separately. The online inference part uses a self-built wide&deep framework based on pytorh to achieve ms-level inference.

[0020] Preferably, some conditional filtering operations are performed on the ALS recall pool in the early stage, and the results are directly used. The data structure is relatively simple and cached using Redis; for new users, this group has not generated any behavior, so there are no corresponding recall results in the ALS recall pool. Here, we use Mongodb to cache some active users and attach some key matching information. During the recall process, the matching information of the user is used as a condition to search for target users that meet the conditions in the cache pool. In step S4, the trained W&D model uses the obtained recall results as candidates to enter the online W&D inference logic, and the best candidates are recommended according to the scoring results finally output by the W&D model. Considering that the provided recommendation service project is not complex, the Python lightweight web asynchronous framework Sanic is used to encapsulate the recommendation service, and Gunicorn is used as the WSGI to handle asynchronous requests.

[0021] The present invention also provides a marriage and dating recommendation device based on image content analysis, including:

[0022] A user face image feature database, a user attribute and behavior data database, a user recall list database, and a user recommendation web server;

[0023] The user face image feature database includes user id, picture id, face feature encoding. In the hive table, the data of all user photos are stored in the hive table after offline calculation;

[0024] The user attribute and behavior data database includes basic information such as user ID, gender, age, height, and mate selection conditions, as well as behavior data such as user clicks and browsing, which are stored in a Hive table;

[0025] The user recall list database includes user ID, to-be-recommended user ID, and a Redis database, and the data is stored in the Redis database;

[0026] The user recommendation web server includes modules for obtaining to-be-recommended users, obtaining user basic attributes and mate selection condition features, obtaining image features, Wide&Deep model inference, and business rule filtering.

[0027] Compared with related technologies, the marriage and dating recommendation method based on image content analysis provided by the present invention has the following beneficial effects:

[0028] The present invention provides a marriage and dating recommendation method based on image content analysis. This method incorporates image features, can effectively capture users' preferences for image content, increases the dimension of user portraits, ensures the accuracy of the model for recommended content, and thus efficiently and accurately recommends suitable mate selection objects for users, improving the users' mate selection efficiency. Brief Description of the Drawings

[0029] Figure 1 It is a flowchart of the method for the marriage and dating recommendation method based on image content analysis provided by the present invention;

[0030] Figure 2 It is a flowchart of image search for the marriage and dating recommendation method based on image content analysis provided by the present invention;

[0031] Figure 3 It is a block diagram of the web interface part of the marriage and dating recommendation method based on image content analysis provided by the present invention;

[0032] Figure 4 It is a schematic structural diagram of the marriage and dating recommendation device based on image content analysis provided by the present invention;

[0033] Figure 5 It is a schematic structural diagram of the server installation device for the marriage and dating recommendation method based on image content analysis provided by the present invention;

[0034] Figure 6 For Figure 5 It is a schematic structural diagram of the heat exchange mechanism part shown;

[0035] Figure 7 For Figure 5 It is a schematic structural diagram of the mounting bracket part shown.

[0036] Reference numerals in the figures:

[0037] 1. Mounting frame

[0038] 2. Heat exchange box, 201. Adjusting hole, 210. Circulation cavity, 220. Heat exchange cavity

[0039] 3. Telescopic member

[0040] 4. Lifting frame, 41. Mounting table

[0041] 5. Isolation net

[0042] 6. Adjusting plate

[0043] 7. Mounting frame, 71. Ice pack

[0044] 8. Heat exchange mechanism, 81. Water pump, 82. Heat exchange pipe, 83. Heat exchange rod Detailed implementation mode

[0045] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0046] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 , wherein, Figure 1 is the method flow chart of the marriage and dating recommendation method based on image content analysis provided by the present invention; Figure 2 is the image search flow chart of the marriage and dating recommendation method based on image content analysis provided by the present invention; Figure 3 is the block diagram of the web interface part of the marriage and dating recommendation method based on image content analysis provided by the present invention; Figure 4 is the structural schematic diagram of the marriage and dating recommendation device based on image content analysis provided by the present invention; Figure 5 is the structural schematic diagram of the server installation device of the marriage and dating recommendation method based on image content analysis provided by the present invention; Figure 6 is Figure 5 the structural schematic diagram of the heat exchange mechanism part shown in; Figure 7 is Figure 5 the structural schematic diagram of the mounting frame part shown in.

[0047] A marriage and dating recommendation method based on image content analysis includes the following steps:

[0048] S1. Use a publicly available data set to train a face recognition model and a face detection model as image feature extractors, and extract the image features of user photos offline;

[0049] The image feature extractor in step S1 includes a face detection model and a face recognition model, wherein the face detection model uses the scrfd model, which obtains the best network structure through a network search method under the premise of a fixed amount of calculation, and ensures that the backbone, neck, and head parts are balanced. The accuracy of model detection is improved through the strategies of training sample balance, ATSS, and DIOU to obtain a training model;

[0050] The face recognition model adopts the insightface structure. The insightface network adds margin to the loss function to improve the inter-class separability of face features in spatial distribution, while strengthening the intra-class compactness and inter-class differences, thereby improving the accuracy of face recognition. In order to ensure the model has good real-time performance, the backbone network uses the lightweight mobilenetv2 structure.

[0051] Use the extractor to extract facial features from photos uploaded by users, and store the results in a hive table for easy reading during recall;

[0052] S2. Based on user behavior data, offline training of ALS as a recall model is performed to generate a list of users to be recommended for each user.

[0053] Collect user behavior data. Here we select some large and representative behavior data generated by users on the APP, such as: left and right swipe and click data of users on the home card recommendation page, data of users sending and reading letters, data of users browsing other people's homepages, data of users browsing, liking, and commenting on dynamic publishing pages, etc. According to the degree of fit between the user's behavior category and the final goal, different behavior scores are manually set for each type of behavior. At the same time, in order to reflect the difference of the same behavior, a time decay factor is constructed, and the time decay factor is used to weight the behavior score when scoring the behavior. Finally, using spark's excellent processing performance for big data, all behavior data are merged to obtain the final User-Iterm Rating Matrix, and then the ALS model is used for explicit feedback training. The trained model is used to make personalized recommendations for each user to obtain the user's personalized recall pool;

[0054] S3. Based on the photos of the opposite sex that the user likes and the facial image features of the photos, recommend similar-looking users to each user to expand the list of users to be recommended, and re-rank the recalled users by comprehensively considering the facial similarity and the ALS model score;

[0055] In step S3, the user is recalled using a similar face recall strategy, and photos similar to the user's favorite user are recommended to the user, and the facial features are PQ-encoded;

[0056] The described PQ coding method divides the 512-dimensional face features into N-dimensional subspaces. Each subspace uses the Kmeans algorithm for clustering into M categories, and the category ID after clustering is used as the PQ coding. Since the cluster centers are distributed in the feature space, the feature codes near the PQ coding are obtained and matched according to the values of the adjacent feature codes, greatly improving the speed of global search;

[0057] During the offline search process, the PQ coding of the user's photo and the adjacent PQ coding are respectively written into the hive table. The adjacent PQ coding of the user's favorite photo and the PQ coding of the photo to be searched are used as the matching rules, and with the computing power of spark, a fast offline global search function is realized. Combining the scoring of ALS and the face similarity in step S2, a rough ranking of the recall results is performed;

[0058] S4. Obtain the user's basic attributes, mate selection conditions, and image feature data. The user clicks on the data to train the Wide&Deep model for online ranking;

[0059] In step S4, the user's basic attributes are obtained. The basic attributes include, but are not limited to, gender, age, height, constellation, hobbies, income, the province, county, and city of the long-term residence and work location, mate selection conditions, and appearance score data. The user can swipe left and right on the card recommendation page, and the Wide&Deep model is trained. The optimal offline model under the current conditions is obtained by integrating the Accuracy, Recall, F1-score, AUC index, and gAUC index. To improve the response speed of the recommendation server, the model parameters of the offline training and the encoding results of the user features are respectively cached, and the online inference part uses a self-built wide&deep framework based on pytorh to achieve ms-level inference;

[0060] S5. With the help of web service technology, provide user recommendation services. This service can implement the process of ALS plus image feature recall, W&D model fine ranking, and business rule filtering to achieve personalized and accurate recommendations for each user;

[0061] Some conditional filtering operations were performed on the ALS recall pool in the early stage, and the results were directly used. The data structure is relatively simple, and Redis is used for caching. For new users, this group has not generated any behaviors, so there are no corresponding recall results in the ALS recall pool. Here, we use Mongodb to cache some active users and attach some key matching information. During the recall process, the matching information of the user is used as a condition to search for target users that meet the conditions in the cache pool. In step S4, the trained W&D model uses the obtained recall results as candidates to enter the online W&D inference logic, and the recommended results are selected based on the scoring results finally output by the W&D model. Considering that the provided recommendation service project is not complex, the lightweight web asynchronous framework Sanic of Python is used to encapsulate the recommendation service, and Gunicorn is used as the WSGI to handle asynchronous requests.

[0062] The present invention also provides a marriage and dating recommendation device based on image content analysis, including:

[0063] A user face image feature database, a user attribute and behavior data database, a user recall list database, and a user recommendation web server;

[0064] The user face image feature database includes user ID, picture ID, face feature encoding. In the hive table, all user photo data is processed offline and stored in the hive table;

[0065] The user attribute and behavior data database includes basic information such as user ID, gender, age, height, and mate selection conditions, as well as user click and browse behavior data, which is stored in the hive table;

[0066] The user recall list database includes user ID, user ID to be recommended, and redis database, and the data is stored in the redis database;

[0067] The user recommendation web server includes modules for obtaining users to be recommended, obtaining user basic attributes and mate selection condition features, obtaining image features, Wide&Deep model inference, and business rule filtering.

[0068] By training the image model to analyze the face similarity and fully considering the personalized information of users, the matching degree between users can be effectively improved, the problem of lacking image content information in traditional recommendation algorithms can be solved, the efficiency of users to find the best partner can be helped, and the user experience and user activity can be improved.

[0069] Compared with the related technologies, the marriage and dating recommendation method based on image content analysis provided by the present invention has the following beneficial effects:

[0070] This method incorporates image features, which can effectively capture users' preferences for image content, increase the dimensions of user portraits, ensure the accuracy of the model for recommended content, and thus efficiently and accurately recommend matching potential partners to users, improving the efficiency of users' partner selection.

[0071] During the installation and use of the marriage and dating recommendation device based on image content analysis provided by the present invention, a server installation device needs to be used. The server installation device includes:

[0072] An installation frame 1 and a heat exchange box 2. The installation frame 1 is fixedly connected to the heat exchange box 2, and an adjustment hole 201 is provided on the heat exchange box 2;

[0073] A telescopic member 3, which is fixedly installed on the installation frame 1;

[0074] A lifting frame 4, which is fixedly installed at the shaft end of the telescopic member 3. The lifting frame 4 is of an L-shaped structure, and an installation table 41 is installed on the lifting frame 4;

[0075] An isolation net 5, which is fixedly installed inside the heat exchange box 2. The isolation net 5 divides the heat exchange box 2 into a circulation chamber 210 and a heat exchange chamber 220;

[0076] An adjustment plate 6, which is movably installed on the heat exchange box 2;

[0077] An installation frame 7, which is fixedly installed on the adjustment plate 6, and an ice pack 71 is installed on the installation frame 7. The installation frame 7 is inserted into the heat exchange chamber 220 through the adjustment hole 201;

[0078] A heat exchange mechanism 8, which includes a water pump 81, a heat exchange pipe 82, and a heat exchange rod 83. The water pump 81 is installed in the circulation chamber 210. The output end of the water pump 81 is installed with a heat exchange pipe 82. The output end of the heat exchange pipe 82 penetrates through the heat exchange box 2 and the installation frame 1. A heat exchange rod 83 is fixedly installed on the heat exchange pipe 82. The heat exchange rod 83 is installed inside the installation frame 1, and the heat exchange rod 83 is misaligned with the telescopic member 3 and the installation table 41. The output end of the heat exchange pipe 82 penetrates through the heat exchange box 2 and extends into the heat exchange chamber 220.

[0079] The installation table 41 on the lifting frame 4 facilitates the installation of server equipment. The lifting frame 4 can be conveniently lifted and adjusted through the telescopic member 3, so as to contract the installed server on the installation frame 1, keep the server in a closed state, reduce the interference of the external environment, and be safer, hidden, more stable and reliable. It is convenient to carry out normal heat exchange and cooling inside the installation frame 1 in cooperation with the heat exchange mechanism 8, and ensure the normal heat exchange during the operation of the equipment.

[0080] The telescopic member 3 is an electric telescopic rod, which is connected to an external power supply during use to provide a stable power source for the lifting adjustment of the lifting frame 4.

[0081] The heat exchange box 2 is filled with a heat exchange water source. A water pump 81 is installed inside the heat exchange box 2. The water pump 81 is connected to an external power supply during use to drain the water source after heat exchange inside the circulation cavity 210. The heat exchange water source exchanges heat with the air inside the mounting frame 1 through the heat exchange pipe 82 and the heat exchange rod 83, providing a stable and reliable cooling source for the space where the server is installed, and maintaining stable heat exchange cooling and operation of the server in a closed state.

[0082] The isolation net 5 is made of stainless steel mesh, which provides a limit for the installation of the mounting frame 7. The mounting frame 7 facilitates the installation of the ice bag 71. After the ice bag 71 is installed, it can cool and exchange heat with the water source inside the heat exchange cavity 220. The output end of the heat exchange pipe 82 can re-transmit the heat-exchanged water source back into the range of the ice bag 71, realizing the circulating cooling of the heat-exchanged water source.

[0083] The adjusting plate 6 drives the ice bag 71 to move up and down through the mounting frame 7, facilitating the installation and disassembly of the ice bag 71.

[0084] The beneficial effects of the server installation device provided by the present invention:

[0085] The installation platform 41 on the lifting frame 4 facilitates the installation of server devices. The lifting frame 4 can be easily lifted and adjusted through the telescopic member 3, so as to contract the installed server on the mounting frame 1, keep the server in a closed state, reduce the interference of the external environment, and be safer, hidden, more stable and reliable. Cooperating with the heat exchange mechanism 8, it is convenient to carry out normal heat exchange cooling inside the mounting frame 1 and ensure normal heat exchange during the operation of the equipment.

[0086] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied to other related technical fields, shall be included in the patent protection scope of the present invention by the same token.

Claims

1. A method for recommending marriage and dating based on image content analysis, characterized in that, it includes the following steps: S1. Use a publicly available dataset to train a face recognition model and a face detection model as extractors of image features, and offline extract the image features of user photos; S2. According to user behavior data, offline train ALS as a recall model to generate a list of users to be recommended for each user; S3. According to the photos of the opposite sex liked by the user, and combined with the face image features of the photos, recommend similar face users for each user to expand the list of users to be recommended, and comprehensively consider the face similarity and the scoring of the ALS model to re-rank the recall; S4. Obtain user basic attributes, mate selection conditions, image feature data, and user click data, and train the Wide&Deep model for online ranking; S5. With the help of web service technology, provide a user recommendation service. This service can implement the process of ALS plus image feature recall, Wide&Deep model fine ranking, and business rule filtering to achieve personalized and accurate recommendations for each user.

2. The method for recommending marriage and dating based on image content analysis according to claim 1, characterized in that, the extractor of image features in step S1 includes a face detection model and a face recognition model. Among them, the face detection model uses the scrfd model. This model uses the method of network search to obtain the best network structure on the premise of fixed computational complexity, and ensures the balance of the backbone, neck, and head parts. Through the strategies of training sample balance, ATSS, and DIOU, the detection accuracy of the model is improved to obtain a trained model.

3. The method for recommending marriage and dating based on image content analysis according to claim 2, characterized in that, the face recognition model adopts the insightface structure. The insightface network improves the inter-class separability of face features in space by adding margin to the loss function, and at the same time strengthens the intra-class compactness and inter-class differences, thereby improving the accuracy of face recognition. In order for the model to have better real-time performance, the backbone network uses the lightweight mobilenetv2 structure.

4. The method for recommending marriage and dating based on image content analysis according to claim 3, characterized in that, Use the extractor to extract the face features of the user-uploaded photos and store the results in a hive table for easy reading during recall.

5. The method for recommending marriage and dating based on image content analysis according to claim 4, characterized in that, In step S3, the recall users adopt a similar face recall strategy to recommend photos similar to the users liked by themselves, and perform PQ coding on the face features.

6. The method for recommending marriage and dating based on image content analysis according to claim 5, characterized in that, The method of PQ coding divides the 512-dimensional face features into N-dimensional subspaces. Each subspace uses the Kmeans algorithm for M-category clustering, and the category ID after clustering is used as the PQ coding. Since the cluster centers are distributed in the feature space, the feature coding near the PQ coding is obtained, and matching is performed according to the values of the adjacent feature codings to improve the speed of global search.

7. The method for matchmaking and dating recommendation based on image content analysis according to claim 6, wherein, During the offline search process, the PQ coding of the user's photo and the adjacent PQ codings are respectively written into the hive table. The adjacent PQ coding of the photo liked by the user and the PQ coding of the photo to be searched are used as the matching rules, and with the computing power of spark, a fast offline global search function is realized. Combining the scoring of ALS and the face similarity in step S2, a rough ranking of the recall results is performed.

8. The method for matchmaking and dating recommendation based on image content analysis according to claim 7, wherein, In step S4, the user's basic attributes are obtained. The basic attributes include, but are not limited to, the integration of data such as gender, age, height, constellation, hobbies, income, the province, county, and city where the user lives and works, mate selection conditions, and face value scores. The user can swipe left and right on the card recommendation page, and the Wide&Deep model is trained. The optimal offline model under the current conditions is obtained by comprehensively considering indicators such as Accuracy, Recall, F1-score, AUC index, and gAUC index. In order to improve the response speed of the recommendation server, the model parameters trained offline and the encoding results of user features are cached respectively. The online inference part uses a self-built Wide&Deep model based on pytorh to achieve ms-level inference.

9. The method for matchmaking and dating recommendation based on image content analysis according to claim 8, wherein, The ALS recall pool has performed conditional filtering operations in the early stage, and the results are directly used. The data structure is simple and cached using Redis; for new users, this group has not generated behaviors, so there are no corresponding recall results in the ALS recall pool. Here, Mongodb is used to cache some active users, and some key matching information is attached. During the recall process, the matching information of the user is used as a condition to search for target users that meet the conditions in the cache pool. In step S4, the trained Wide&Deep model uses the obtained recall results as candidates to enter the online Wide&Deep model inference logic, and the best recommendation is made according to the scoring results finally output by the Wide&Deep model. Considering that the provided recommendation service project is not complex, the lightweight web asynchronous framework Sanic of Python is used to encapsulate the recommendation service, and Gunicorn is used as the WSGI to handle asynchronous requests.

10. A device for matchmaking and dating recommendation based on image content analysis, including the method for matchmaking and dating recommendation based on image content analysis according to any one of claims 1-9, wherein, including: User face image feature database, user attribute and behavior data database, user recall list database, user recommendation web server; The user face image feature database includes user ID, picture ID, face feature encoding. In the Hive table, all user photo data is processed offline and stored in the Hive table; The user attribute and behavior data database includes basic information such as user ID, gender, age, height, as well as mate selection conditions and user click and browsing behavior data, which are stored in the Hive table; The user recall list database includes user ID and to-be-recommended user ID, and the data is stored in the Redis database; The user recommendation web server includes to-be-recommended user acquisition, user basic attribute and mate selection condition feature acquisition, image feature acquisition, Wide&Deep model inference, and business rule filtering modules.

Images