Social user matching method and device based on large model, medium and electronic equipment

Abstract

The invention relates to the technical field of computers, and discloses a social user matching method and device based on a large model, a medium and electronic equipment. And collecting behavior data of the main state user and attribute data of the customer state user, respectively inputting the behavior data and the attribute data into the fine-tuned large model to obtain main state sequence features corresponding to the main state user and customer state attribute features corresponding to the customer state user, and recommending the customer state user to the main state user by adopting a preset recommendation strategy based on the main state sequence features and the customer state attribute features. According to the method, characterization is carried out based on a large model of fine tuning, long and short term dependency relationships and hidden modes in a user behavior sequence can be fully, deeply and comprehensively mined, trapping into convergence is avoided, the recall rate and the click rate are effectively improved, and better generalization ability is achieved; and meanwhile, the fine tuning training cost of the large model is lower, and low-cost and high-efficiency iterative updating can be realized under data driving, so that the dynamic change of the social interest point of the user can be tracked in time, and lag is avoided.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology. More specifically, the embodiments of the present disclosure relate to a social user matching method based on a large model, a social user matching device based on a large model, a computer-readable storage medium, and an electronic device. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims, and no statement herein is admitted to be prior art by inclusion in this section.

[0003] In user matching in social scenarios, collaborative filtering algorithms are commonly used to perform similarity analysis based on the behavioral data of recommended users; or, based on social networks, a graph neural network structure with users as nodes and social behaviors as edges is constructed to perform similarity analysis and make recommendations based on similarity.

[0004] The application of collaborative filtering algorithms requires consideration of data sparsity. Users' social interaction behavior data is often sparse, or there may be no interaction at all, resulting in inaccurate similarity calculations and impacting recommendation accuracy. Furthermore, similar user groups tend to have similar preferences, which can lead to single recommendations and hinder the discovery of new points of interest and the expansion of diversity. The application of graph neural networks requires consideration of the cost of training complex structures and long-term dynamic maintenance. Graph neural networks are complex in structure and require high training resources and labeled data. Furthermore, as the user base iterates, graph neural networks also need to be updated in a timely manner to avoid lags. This results in high implementation costs and difficulty in rapid deployment and efficient maintenance. Summary of the Invention

[0005] In this context, embodiments of the present disclosure are intended to provide a social user matching method based on a large model, a social user matching device based on a large model, a computer-readable storage medium, and an electronic device.

[0006] According to a first aspect of an embodiment of the present disclosure, a social user matching method based on a big model is provided. The method may include: collecting behavioral data of a primary user and attribute data of a guest user; inputting the behavioral data into a fine-tuned big model to obtain a primary sequence feature corresponding to the primary user; inputting the attribute data into the fine-tuned big model to obtain a guest attribute feature corresponding to the guest user; and recommending at least one guest user to the primary user using a preset recommendation strategy based on the primary sequence feature and the guest attribute feature.

[0007] In an exemplary embodiment, the primary state sequence feature corresponds to the user identifier of the primary state user. Based on the primary state sequence feature and the guest state attribute feature, a preset recommendation strategy is adopted to recommend at least one guest state user to the primary state user, including: in response to a recommendation request for a target user, obtaining the corresponding target primary state sequence feature using the user identifier of the target user; performing similarity matching on the target primary state sequence feature and the guest state attribute feature, and determining at least one guest state user to be recommended to the target user based on the matching result.

[0008] In an exemplary embodiment, the primary state sequence feature corresponds to the user identifier of the primary state user. Based on the primary state sequence feature and the guest state attribute feature, a preset recommendation strategy is adopted to recommend at least one guest state user to the primary state user, including: in response to a recommendation request for a target user, obtaining the corresponding target primary state sequence feature with the user identifier of the target user; inputting the target primary state sequence feature and the guest state attribute feature into a recommendation model, and determining at least one guest state user to be recommended to the target user based on the recall result of the recommendation model.

[0009] In an example embodiment, after recommending at least one guest user to the primary user based on the primary sequence features and the guest attribute features, the method further includes: continuously obtaining multi-dimensional feedback data of the primary user; and updating at least one of the large model and the preset recommendation strategy with the feedback data.

[0010] In an exemplary embodiment, continuously obtaining feedback data of the primary user includes: periodically obtaining multi-dimensional feedback data of the primary user at a dynamic time interval, and the dynamic time interval is adjusted based on at least one of data update frequency, data accumulation speed, and data processing efficiency.

[0011] In an exemplary embodiment, the large model includes an input layer and a base, and the steps of fine-tuning the large model are as follows: collect historical behavior data of the main user, and the historical behavior data includes historical behavior sequence and historical next behavior; input the historical behavior sequence into the large model, perform vector encoding through the input layer, and extract and predict it by the base based on the self-attention mechanism to obtain the predicted next behavior output by the large model for the main user; calculate the prediction loss based on the predicted next behavior and the historical next behavior; and fine-tune the large model with the prediction loss.

[0012] In an exemplary embodiment, the large model also includes a vector conversion layer, the weight parameters of the vector conversion layer are two-dimensional matrices, the weight parameters represent the number of samples and the hidden layer dimension, and the vector conversion layer is updated after each generation of fine-tuning training.

[0013] In an exemplary embodiment, the vector encoding includes word unit encoding, word unit position encoding, word unit type encoding, and sequence position encoding.

[0014] In an exemplary embodiment, the self-attention mechanism includes self-attention calculation based on a sliding window mechanism, a hole sliding window mechanism, and a global information fusion sliding window mechanism.

[0015] In an exemplary embodiment, the prediction loss is a global cross entropy loss.

[0016] In an exemplary embodiment, the behavioral data includes social interaction behavior data of the primary user and the guest user; the attribute data includes at least one of identity identification data, identity feature data, social relationship data, physiological and life feature data, and personality and expression feature data.

[0017] In an exemplary embodiment, after collecting the behavioral data of the primary user and the attribute data of the guest user, the method further includes: preprocessing the behavioral data and the attribute data, wherein the preprocessing includes at least one of data cleaning, time window screening, random sampling, and data integration.

[0018] According to a second aspect of an embodiment of the present disclosure, a social user matching device based on a large model is provided, which may include: a data collection module for collecting behavioral data of a primary user and attribute data of a guest user; a first characterization module for inputting the behavioral data into a fine-tuned large model to obtain primary sequence features corresponding to the primary user; a second characterization module for inputting the attribute data into the fine-tuned large model to obtain guest attribute features corresponding to the guest user; and a user recommendation module for recommending at least one guest user to the primary user using a preset recommendation strategy based on the primary sequence features and the guest attribute features.

[0019] In an exemplary embodiment, the primary state sequence feature corresponds to the user identifier of the primary state user; the user recommendation module is specifically used to respond to a recommendation request for a target user, obtain the target primary state sequence feature with the user identifier of the target user; perform similarity matching on the target primary state sequence feature and the guest state attribute feature, and determine at least one guest state user recommended to the target user based on the matching result.

[0020] In an exemplary embodiment, the primary state sequence feature corresponds to the user identifier of the primary state user; the user recommendation module is specifically used to respond to a recommendation request for a target user and obtain the target primary state sequence feature with the user identifier of the target user; the target primary state sequence feature and the guest state attribute feature are input into the recommendation model, and at least one guest state user recommended to the target user is determined based on the recall result of the recommendation model.

[0021] In an exemplary embodiment, the device further includes a continuous updating module for continuously acquiring multi-dimensional feedback data of primary users; and updating at least one of the large model and the preset recommendation strategy with the feedback data.

[0022] In an exemplary embodiment, the continuous update module is specifically used to periodically obtain multi-dimensional feedback data of the primary user at a dynamic time interval, and the dynamic time interval is adjusted based on at least one of the data update frequency, data accumulation speed, and data processing efficiency.

[0023] In an exemplary embodiment, the large model includes an input layer and a base, and the device also includes a model training module for collecting historical behavior data of the main user, the historical behavior data including historical behavior sequence and historical next behavior; the historical behavior sequence is input into the large model, vector encoded through the input layer, and extracted and predicted by the base based on the self-attention mechanism to obtain the predicted next behavior output by the large model for the main user; the prediction loss is calculated based on the predicted next behavior and the historical next behavior; and the large model is fine-tuned with the prediction loss.

[0024] In an exemplary embodiment, the large model also includes a vector conversion layer, the weight parameters of the vector conversion layer are two-dimensional matrices, and the weight parameters represent the number of samples and the hidden layer dimension; the model training module is also used to update the vector conversion layer after each generation of fine-tuning training is completed.

[0025] In an exemplary embodiment, the vector encoding includes word unit encoding, word unit position encoding, word unit type encoding, and sequence position encoding.

[0026] In an exemplary embodiment, the self-attention mechanism includes self-attention calculation based on a sliding window mechanism, a hole sliding window mechanism, and a global information fusion sliding window mechanism.

[0027] In an exemplary embodiment, the prediction loss is a global cross entropy loss.

[0028] In an exemplary embodiment, the behavioral data includes social interaction behavior data of the primary user and the guest user; the attribute data includes at least one of identity identification data, identity feature data, social relationship data, physiological and life feature data, and personality and expression feature data.

[0029] In an exemplary embodiment, the data collection module is further used to pre-process the behavior data and attribute data, and the pre-processing includes at least one of data cleaning, time window screening, random sampling, and data integration.

[0030] According to a third aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, any of the above-mentioned social user matching methods based on a large model is implemented.

[0031] According to a fourth aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute any one of the above-mentioned large model-based social user matching methods by executing the executable instructions.

[0032] According to the social user matching method based on the big model in the embodiment of the present disclosure, the social user matching device, computer-readable storage medium and electronic device based on the big model can collect the behavioral data of the main user and the attribute data of the guest user, and input them into the fine-tuned big model to obtain the main sequence features corresponding to the main user and the guest attribute features corresponding to the guest user; on this basis, the guest user can be recommended to the main user using a preset recommendation strategy based on the main sequence features and the guest attribute features. In the embodiment of the present disclosure, the behavioral data of the main user and the attribute data of the guest user are characterized separately based on the fine-tuned big model, which can fully, deeply and comprehensively mine the long-term and short-term dependencies and hidden patterns in the user behavior sequence, avoid falling into convergence, effectively improve the recall rate and click-through rate, and have better generalization ability; at the same time, the fine-tuning training cost of the big model is lower, and it can also be iteratively updated at low cost and high efficiency under data-driven, so as to timely track the dynamic changes of users' social interests and avoid lagging. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings, in which several embodiments of the present disclosure are shown by way of example and not limitation, wherein:

[0034] Figure 1 One of the flow charts of a social user matching method based on a large model according to an embodiment of the present disclosure is shown;

[0035] Figure 2 A flow chart showing the fine-tuning steps of a large model according to an embodiment of the present disclosure is shown;

[0036] Figure 3 One of the flow charts of a social user matching method based on a large model according to an embodiment of the present disclosure is shown;

[0037] Figure 4 The second flowchart of the social user matching method based on the big model according to the embodiment of the present disclosure is shown;

[0038] Figure 5 A schematic diagram of a social user matching device based on a large model according to an embodiment of the present disclosure is shown;

[0039] Figure 6A schematic diagram of a storage medium according to an embodiment of the present disclosure is shown;

[0040] Figure 7 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0041] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. Rather, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present disclosure may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0043] According to an embodiment of the present disclosure, a social user matching method based on a large model, a social user matching device based on a large model, a computer-readable storage medium, and an electronic device are provided.

[0044] In this document, any number of elements in the drawings is for illustration and not for limitation, and any naming is for distinction only and does not have any limiting meaning.

[0045] The principles and spirit of the present disclosure are described in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0047] Collaborative filtering algorithms can be used for user matching in social scenarios. For example, user-based collaborative filtering can calculate similar users based on the interaction behavior data of the recommended user and then recommend similar users or items they interact with. Alternatively, item-based collaborative filtering can calculate similar items based on their co-occurrence in user behavior data and then recommend similar items or users who interact with similar items. However, this method suffers from data sparsity, which affects the accuracy of similarity calculations. It is also subject to biases based on interest groups, leading to biased and simplistic recommendation results.

[0048] Alternatively, a graph neural network can be used, with nodes constructed from users and edges constructed from social behaviors. User attribute information such as age and gender can be converted into a vector representation. This vector representation is then combined with the node embedding vector learned by the algorithm within the graph neural network structure to obtain a user representation that combines network structural features with attribute information. Similarity calculations can then be performed based on this user representation for subsequent recommendations. This approach results in a complex graph neural network structure, high training resource requirements, high requirements for the quality of the labeled data, and difficulty in long-term dynamic maintenance.

[0049] In the embodiments of the present invention, the behavioral data of the main user and the attribute data of the guest user are characterized separately based on the fine-tuning of the large model, which can fully, deeply and comprehensively mine the long-term and short-term dependencies and hidden patterns in the user behavior sequence, avoid falling into convergence, effectively improve the recall rate and click-through rate, and have better generalization ability; at the same time, the fine-tuning training cost of the large model is lower, and it can be iteratively updated at low cost and high efficiency under data-driven, so as to timely track the dynamic changes of users' social interests and avoid lagging behind.

[0050] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention are described in detail below.

[0051] Example application scenarios

[0052] It should be noted that the following application scenarios are only provided to facilitate understanding of the spirit and principles of the present invention, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0053] The large model-based social user matching method of the embodiments of the present disclosure can be applied to various application scenarios involving social user matching.

[0054] In one application scenario, a social platform may be involved. Typically, in this application scenario, users can search, browse, and publish posts in various forms, and can also interact with other users, such as following other users, or liking, commenting on, forwarding, collecting, and sharing other users' posts to other platforms. On a social platform, posts published by other users can be recommended to users, and other users can also be recommended to users. In this application scenario, the large-scale model-based social user matching method of the embodiment of the present disclosure can be used to perform vector representations of the user and other users respectively through a fine-tuned large-scale model, and then make recommendations based on the vector representations using a preset recommendation strategy.

[0055] In another application scenario, an information platform may be used to recommend information publishers to users; an e-commerce platform may be used to recommend merchants or other buyers to users, etc. Those skilled in the art may implement the large model-based social user matching method provided in the embodiments of the present disclosure in different application scenarios according to actual needs.

[0056] Exemplary Methods

[0057] In combination with the above application scenarios, Figure 1 The following describes a social user matching method based on a large model according to an exemplary embodiment of the present disclosure.

[0058] like Figure 1 One of the processes of the large model-based social user matching method according to the exemplary embodiment of the present disclosure shown may include the following steps 101 to 104 .

[0059] In step 101, the behavior data of the primary user and the attribute data of the guest user are collected.

[0060] In the disclosed embodiments, among multiple users, the primary user is any user who is recommended, and the guest users can be some or all users other than the primary user. The multiple users can be active users in different scenarios, such as social platforms, information platforms, communication platforms, and e-commerce platforms. They can be multiple users on a single platform or multiple users across multiple platforms. Based on the determination of the primary user and its corresponding guest user, data collection and screening can be performed on the corresponding platforms where the multiple users are active.

[0061] Among them, the main user can collect his behavioral data, which may include the sequence of interactive behaviors implemented by the main user. The behavioral data describes the main user's actions and the order in which the actions occur; the guest user can collect his attribute data, which may include static attributes of the guest user that are not related to the interactive behavior, and may also include the guest user's interactive status, which may describe the guest user's active tendencies.

[0062] In an optional method embodiment of the present disclosure, the behavior data includes social interaction behavior data between the primary user and the guest user.

[0063] In the disclosed embodiments, behavioral data may include social interaction behavior data of the primary user with the guest user. Social interaction behavior can include proactive social interaction behavior of the primary user with the guest user, such as the primary user clicking on or sharing the guest user's personal homepage, sending a private message to the guest user, following the guest user, and so on. For proactive social interaction behavior, the time of occurrence, the object of the interactive guest user, and the type of behavior can be collected to characterize the interactive relationship and behavioral patterns between the primary user and different guest users. Social interaction behavior can also include the primary user's consumer interaction behavior with published content, such as the browsing time of text, images, videos, and other content posted by the guest user, as well as consumer interaction behaviors such as likes, comments, favorites, and forwarding. For consumer interaction behavior, the time of occurrence, interactive content, and behavior type can be collected. Comprehensive collection of social interaction behavior data allows for a more thorough and in-depth analysis of the primary user's behavioral habits and content tendencies, better supporting subsequent reasoning and analysis.

[0064] In an optional method embodiment of the present disclosure, the attribute data includes at least one of identity identification data, identity feature data, social relationship data, physiological and life feature data, and personality and expression feature data.

[0065] In the disclosed embodiments, attribute data may include multi-dimensional, comprehensive descriptive information about guest users. Identity data may include information such as the guest user's username and nickname, which reflect the guest user's personalized identity. Identity feature data may include data such as gender and province / city, which reflect the guest user's basic identity features. Social relationship data may include data such as occupation, education level, and marital status, which reflect the guest user's socioeconomic characteristics. Physiological and lifestyle data may include data such as age, height, weight, and zodiac sign, which reflect the guest user's personal physiological and lifestyle characteristics. Personality and expression feature data may include data such as the guest user's published personalized content or interaction status, which reflect the guest user's personality traits and self-expression tendencies. This data may include the guest user's signature, published text, images, videos, audio, and other content, active topics or subject areas, frequency of activity, and number of interactions. Interactions may include the number of likes, comments, favorites, and shares. By collecting extensive and in-depth attribute data on the guest user side, as well as on the guest user's publishing and interaction with content, a more complete user profile of the guest user can be constructed to support subsequent analysis and reasoning.

[0066] In step 102, the behavior data is input into the fine-tuned large model to obtain the main state sequence features corresponding to the main state user.

[0067] In the disclosed embodiments, the fine-tuned large model can be obtained by fine-tuning and training a large model that vectorizes input representations through methods such as generation and encoding. Based on this, the behavioral data of the primary user can be input into the fine-tuned large model to obtain the primary state sequence features output by the fine-tuned large model through inference. The primary state sequence features are used to characterize the type and sequence information of the primary user's interactive behavior.

[0068] In step 103, the attribute data is input into the fine-tuned macro model to obtain the guest attribute features corresponding to the guest user.

[0069] In the disclosed embodiment, the attribute data of the guest user can also be input into the fine-tuned large model to obtain the guest attribute features output by the fine-tuned large model after inference. The guest attribute features are used to represent information such as the static attributes and interaction status of the guest user.

[0070] It should be noted that the fine-tuned large model involved in step 102 and step 103 can be a single deployed fine-tuned large model, and the fine-tuned large model can be executed in the order of step 102 first and then step 103, or step 103 first and then step 102 (see Appendix). Figure 2 It is also possible to deploy two identical fine-tuned large models, and then perform steps 102 and 103 on the two identical fine-tuned large models (see FIG. Figure 2 Alternatively, the large model can also be fine-tuned based on the different types of data sources to adapt to the representation needs of different types of users, so that two different fine-tuned large models can be deployed respectively, and steps 102 and 103 can be performed on the two different fine-tuned large models respectively (as shown in the attached figure). Figure 2 Examples), the above examples are for illustration only, and those skilled in the art may make selections based on actual deployment conditions, characterization requirements, etc., and the embodiments of the present disclosure do not impose specific limitations on this.

[0071] In step 104, based on the primary state sequence characteristics and the guest state attribute characteristics, a preset recommendation strategy is adopted to recommend at least one guest state user to the primary state user.

[0072] In the embodiments of the present disclosure, based on characterizing the primary user based on primary state sequence features and the guest user based on guest state attribute features, the primary state can be matched with the guest user using the primary state sequence features and the guest state attribute features, thereby recommending at least one guest user to the primary user based on whether they match. The preset recommendation strategy can be a pre-selected and configured method for matching the primary state with the guest user, as well as a strategy for selecting recommended items based on the matching results. Those skilled in the art can select and configure this strategy based on actual needs, and the embodiments of the present disclosure do not impose specific limitations on this.

[0073] In an optional method embodiment of the present disclosure, the primary state sequence feature corresponds to the user identifier of the primary state user.

[0074] In the disclosed embodiment, different primary-state users can correspond to different primary-state sequence features, and each primary-state user can also have a unique corresponding user identifier. Based on this, multiple users can be pre-selected as recommended users and subjected to the aforementioned primary-state user-specific processing. The obtained primary-state sequence features can then be stored separately. During storage, the primary-state user's user identifier is used to distinguish between different primary-state sequence features. This allows recommendations to be made to the primary-state user directly based on the user identifier, eliminating the need for on-site characterization and enabling rapid subsequent reasoning and analysis.

[0075] For example, the pre-acquired primary state sequence features and guest state attribute features can be exported and stored in a distributed file system (HDFS). HDFS is highly fault-tolerant and suitable for processing large amounts of data, enabling secure and stable storage of the aforementioned vector representations. Furthermore, the primary state sequence features and guest state attribute features stored in HDFS can be sequentially imported into a search engine for subsequent feature retrieval based on recommendation requests, and then matching and recommendation based on the search results.

[0076] On this basis, in an optional method embodiment of the present disclosure, step 104 may include the following steps A1 to A2.

[0077] In step A1, in response to a recommendation request for a target user, a corresponding target main state sequence feature is obtained using the user identifier of the target user.

[0078] In the disclosed embodiments, in a recommendation process, the target user is the recommended user. A recommendation request can be triggered automatically when a new page is opened, when returning to an old page, or in other situations. It can also be triggered based on instructions such as updating or supplementing page content. Based on this, a user ID of the target user can be requested in response to a recommendation request. Furthermore, the primary state feature sequence can be retrieved based on the user ID to obtain the target primary state sequence features obtained by processing the behavioral data of the fine-tuned large model when the target user is the primary state user.

[0079] In step A2, similarity matching is performed between the target main state sequence features and the guest state attribute features, and at least one guest state user recommended to the target user is determined based on the matching result.

[0080] In an embodiment of the present disclosure, based on the determination of the target main state sequence features, the target main state sequence features can be matched with the guest state attribute features for similarity. For example, cosine similarity can be used to calculate the cosine value of the vector angle between the target main state feature sequence and the guest state attribute features, and the similarity is evaluated using this cosine value. The closer the cosine value is to 1, the more similar the social behavior pattern and content interest preferences of the target user and the corresponding guest state user are; conversely, the closer the cosine value is to 0, the greater the difference in social behavior pattern and content interest preferences between the target user and the corresponding guest state user.

[0081] On this basis, the similarity matching results can include the cosine similarity between the target main state sequence features and different guest state attribute features. Based on the size of the cosine similarity, guest state users close to the target user are selected as candidate recommendation items and recommended to the target user.

[0082] Those skilled in the art may also choose other similarity calculation methods according to actual needs, such as Pearson correlation coefficient, distance similarity, etc., and the embodiments of the present disclosure do not impose specific limitations on this.

[0083] Alternatively, in an optional method embodiment of the present disclosure, step 104 may include the following steps B1 to B2.

[0084] In step B1, in response to a recommendation request for a target user, a corresponding target main state sequence feature is obtained using the user identifier of the target user.

[0085] In the embodiment of the present disclosure, step B1 may correspond to the aforementioned step A1, and will not be described again here to avoid repetition.

[0086] In step B2, the target main state sequence features and the guest state attribute features are input into the recommendation model, and at least one guest state user recommended to the target user is determined based on the recall result of the recommendation model.

[0087] In the disclosed embodiments, a recommendation model can be used to perform inference based on the target's primary state sequence features and guest state attribute features to obtain recommendation scores for recommending different guest state users to the target user. Through training, the recommendation model can calculate similarity scores between the target's primary state sequence features and guest state attribute features, thereby comprehensively evaluating the recommendation scores of guest state users corresponding to the guest state attribute features. These users are then ranked based on the recommendation scores to obtain a recall result from the recommendation model, which is then used to recommend corresponding guest state users to the target user.

[0088] It should be noted that steps A1 to A2 and steps B1 to B2 can be different preset recommendation strategies. Those skilled in the art can also combine steps A1 to A2 and steps B1 to B2 according to their needs. For example, after similarity matching of the target main state sequence features and the guest state attribute features, the guest state users whose similarity meets the candidate needs are screened out as recall sources based on the similarity matching results, and then the similarity matching results are used as one of the reference dimensions of the recommendation score of the recommendation model. The recommendation model comprehensively evaluates the recommendation score based on the target main state sequence features, the guest state attribute features and the similarity matching results, and then sorts based on the recommendation score to obtain the recall result of the recommendation model. Those skilled in the art can also choose other preset recommendation strategies for recommendation, such as using the main state sequence features and the guest state attribute features as the feature vectors of the refined ranking model, etc. The embodiments of the present disclosure do not impose specific restrictions on this.

[0089] According to the large-scale model-based social user matching method of the disclosed embodiment, the behavioral data of the primary user and the attribute data of the guest user can be collected and respectively input into the fine-tuned large-scale model to obtain the primary sequence features corresponding to the primary user and the guest attribute features corresponding to the guest user; on this basis, a preset recommendation strategy can be adopted based on the primary sequence features and the guest attribute features to recommend the guest user to the primary user. In the disclosed embodiment, the fine-tuned large-scale model is used to characterize the behavioral data of the primary user and the attribute data of the guest user, respectively, which can fully, deeply, and comprehensively mine the long-term and short-term dependencies and hidden patterns in the user behavior sequence, avoid falling into convergence, effectively improve the recall rate and click-through rate, and have better generalization ability; at the same time, the fine-tuning training cost of the large-scale model is lower, and it can also be iteratively updated at low cost and high efficiency under data-driven conditions, thereby timely tracking the dynamic changes of users' social interests and avoiding lags.

[0090] like Figure 2 The flowchart of the fine-tuning steps of the large model according to the exemplary embodiment of the present disclosure may include steps 201 to 204 .

[0091] In step 201, historical behavior data of the primary user is collected, and the historical behavior data includes a historical behavior sequence and a historical next behavior.

[0092] In the disclosed embodiments, the large model can be fine-tuned using the primary user's historical behavior data so that the large model can predict the primary user's next possible behavior based on their behavior sequence. Based on this, the primary user's historical behavior data can include a historical behavior sequence and a historical next behavior, where the historical next behavior is the next historical behavior performed by the primary user in the time series of the historical behavior sequence.

[0093] For example, if the primary user clicked on posts A, B, and C in the past 10 minutes, the historical behavior sequence may be "click on post A - click on post B", and the next historical behavior may be "click on post C".

[0094] In step 202, the historical behavior sequence is input into the large model, vector-encoded through the input layer, and extracted and predicted by the base based on the self-attention mechanism to obtain the predicted next behavior output by the large model for the main user.

[0095] In the embodiment of the present disclosure, the large model may include an input layer and a base. The input layer is used to encode the input data to support the subsequent reasoning and analysis of the base. When fine-tuning the large model using historical behavior data, the input layer converts the specific behavior (item) in the historical behavior sequence into an attribute representation that reflects its characteristics, and then encodes it according to the encoding rules set for the attribute representation, and the different encodings can be further vectorized (embedding) to obtain a vector representation with certain semantic information. On this basis, different vector representations are spliced in a specific order to form a feature vector that comprehensively represents the historical behavior sequence. The feature vector is used as the input of the subsequent base to support its subsequent further reasoning and analysis. In the embodiment of the present disclosure, the aforementioned input layer encoding and processing method is used to more accurately and comprehensively convert the historical behavior sequence of the main user into a representation form that is understandable and processable by the base.

[0096] In the disclosed embodiment, the base can be an open source large model that captures key information such as the long-short dependencies in the historical behavior sequence and the importance of each element in the sequence based on the self-attention mechanism. Across long sequence intervals, deep connections are mined in the continuous behavior of the main user. During the operation of the large model, the base gradually refines and integrates the input feature vector through forward calculation and back propagation at different layers. The features of the basic behavior are first extracted and preliminarily processed, and abstractions are gradually integrated during the transmission process to form more abstract and general high-level features. On this basis, the next behavior (nextitem) of the main user in the time series is predicted.

[0097] In an optional method embodiment of the present disclosure, the vector encoding includes word unit encoding, word unit position encoding, word unit type encoding and sequence position encoding.

[0098] Token encoding. After the historical behavior sequence is converted into text, it can be segmented to obtain different tokens. Each token is assigned a unique encoding based on encoding rules. This encoding can accurately identify the corresponding token information, facilitating subsequent recognition and understanding of the token.

[0099] Token position encoding. In the historical behavior sequence, the position of each token holds key semantic and logical significance. For each token, a position-specific encoding can be set based on encoding rules to record the token's position in the historical behavior sequence. Position encoding facilitates comprehensive consideration of token position information during subsequent reasoning and analysis, better capturing the order-related semantic information in the historical behavior sequence.

[0100] Token type encoding. In historical behavior sequences, items may have different attribute types after being represented by attributes. To distinguish tokens of different attribute types, we can set the corresponding attribute type encoding for each token based on encoding rules. This allows for targeted analysis and processing based on the attribute type corresponding to the token in subsequent reasoning, while also avoiding interference caused by confusion between different attribute types.

[0101] Sequence position encoding. In a historical behavior sequence, the position of an item in the historical behavior sequence is associated with the behavior pattern of the primary user. Each item can also be encoded based on its position in the historical behavior sequence based on encoding rules to record the item's position in the historical behavior sequence. This facilitates subsequent reasoning to understand the relative positional relationships between items in the historical behavior sequence and the meaning reflected by the overall structure.

[0102] In the embodiment of the present disclosure, when performing vector encoding at the input layer, the aforementioned encoding method can be used for encoding and then vectorization, so that the comprehensive feature vector can include various information such as position and type, and more comprehensively reflect the behavior pattern of the main user in the historical behavior sequence.

[0103] In an optional method embodiment of the present disclosure, the self-attention mechanism includes self-attention calculation based on a sliding window mechanism, a hole sliding window mechanism, and a global information fusion sliding window mechanism.

[0104] In the disclosed embodiment, the base can use the open-source Longformer large model. Based on the multi-layer self-attention mechanism structure of the Transformer, a self-attention mechanism is implemented whose spatiotemporal complexity is linearly related to the length of the text sequence, enabling deep mining on long sequences. The self-attention mechanism can include self-attention calculations based on a sliding window mechanism, a hole sliding window mechanism, and a global information fusion sliding window mechanism.

[0105] Figure 3 A schematic diagram of a self-attention mechanism provided in an embodiment of the present disclosure is shown in FIG. Figure 3As shown, compared with the current attention mechanism 301, the sliding window mechanism 302 can limit the scope of attention calculation, thereby improving calculation efficiency and capturing local information in the historical behavior sequence in a targeted manner; the hollow sliding window mechanism 303 expands the coverage of attention calculation on this basis, introduces intervals in the sliding window, and can mine effective information within a larger coverage range to avoid excessive concentration of local information; the global information fusion sliding window mechanism 304 integrates global information into local calculations, so that while processing local information, it can also take into account global contextual information.

[0106] The above-mentioned Longformer large model is only used as an example. Those skilled in the art can select a suitable open source large model according to actual needs, such as GPT, etc., and the embodiments of the present disclosure do not impose specific limitations on this.

[0107] In step 203 , the prediction loss is calculated based on the predicted next behavior and the historical next behaviors.

[0108] In the disclosed embodiments, a prediction loss can be calculated based on the output predicted next behavior and the actual collected historical next behavior. This prediction loss represents the deviation between the large model's prediction of the next item and the expected situation. Those skilled in the art can select an appropriate loss function based on actual needs and computing conditions, and the disclosed embodiments do not impose specific limitations on this.

[0109] For example, the large model predicts the next behavior based on the output of "clicking post A - clicking post B" as "the probability of clicking post C is 40%, and the probability of clicking post D is 60%", then calculates the prediction loss based on the predicted next behavior and the historical next behavior of "clicking post C".

[0110] In an optional method embodiment of the present disclosure, the prediction loss is a global cross entropy loss.

[0111] In the disclosed embodiments, prediction loss can be implemented using global cross-entropy loss to measure the degree of difference between the probability distribution of the predicted next behavior and the true distribution of the historical next behavior. After the historical behavior sequence is input into the large model, the large model processes it based on its own structure and parameters, and makes probabilistic predictions for different situations of the next item. On this basis, the global cross-entropy loss is used to calculate the degree of difference between the predicted next behavior and the historical next behavior, indicating the training direction for the large model.

[0112] In step 204, the large model is fine-tuned using the prediction loss.

[0113] In the disclosed embodiment, based on the prediction loss, the backpropagation algorithm can be used to adjust the parameters of the large model. Through iterative optimization, the predicted next behavior output by the large model is made as close as possible to the historical next behavior. Based on this, the large model can predict the next item based on the historical behavior sequence of the primary user, supporting matching and recommendation of objects that the primary user is likely to interact with.

[0114] Specifically, after obtaining the fine-tuned large model, inference analysis can be performed based on the fine-tuned large model. During the inference analysis process, the fine-tuned large model is mainly used to process and analyze data of large-scale and widely representative primary and guest users.

[0115] Corresponding reference Figure 1 As shown in the figure, the dominant user's behavioral data is input into the fine-tuned large model. After tokenization at the input layer, a forward propagation operation is performed within the large model. Through multi-layer feature extraction, transformation, and integration, a feature representation corresponding to the behavioral data is generated. The dominant sequence features are selected from the output of the large model as the representation of the dominant user. This effectively characterizes the dominant user's behavior patterns and interests, thus supporting further data analysis and recommendation decisions.

[0116] For guest users, attribute data is input into the fine-tuned master model. After tokenization at the input layer, forward propagation is performed within the master model. Through multiple layers of feature extraction, transformation, and integration, a feature representation corresponding to the attribute data is generated. Guest attribute features are selected from the master model's output as a representation of the guest user. This effectively characterizes the guest user's attribute preferences and behavior patterns, supporting subsequent data analysis and recommendation decisions. Recommendations are then made based on both primary sequence features and guest attribute features.

[0117] In an optional method embodiment of the present disclosure, the large model also includes a vector conversion layer, the weight parameters of the vector conversion layer are two-dimensional matrices, the weight parameters represent the number of samples and the hidden layer dimension, and the vector conversion layer is updated after each generation of fine-tuning training is completed.

[0118] In the embodiment of the present disclosure, the vector conversion layer is used to sample the input data to reduce the hardware requirements for data processing. The weight parameters of the vector conversion (item embedding) layer can be a two-dimensional matrix, which represents the number of samples and the hidden layer dimension. In the behavioral data of the main user, the behavior of the main user can correspond to the guest user with whom the interaction occurs. In this case, the number of samples can be expressed as the number of guest users. During the fine-tuning training process of the large model, the weight parameters of the vector conversion layer are updated after each epoch of fine-tuning training. That is, the vector conversion layer is updated when a pre-training data set has completely passed through the large model for one forward propagation and one backward propagation. When there are still some training samples in a pre-training data set that have not passed through the large model, the vector conversion layer is not updated.

[0119] Specifically, when the large model also includes a vector conversion layer, the fine-tuning training process of the large model may include first loading the pre-training parameters of the open source large model as the base, and then inputting the historical behavior data of the main user as the fine-tuning training set. After vector conversion layer sampling, input layer vector encoding and base inference output, the prediction loss is calculated and backpropagation is performed to update the parameters.

[0120] It should be noted that the vector conversion layer is updated after each generation of fine-tuning training is completed, and is not updated during the fine-tuning training of one generation.

[0121] Furthermore, the large model can be verified and tested, and its performance on the validation and test datasets can be evaluated to select a fine-tuned large model whose performance meets the application requirements.

[0122] like Figure 4 The second process of the large model-based social user matching method according to the exemplary embodiment of the present disclosure may include the following steps 401 to 404 .

[0123] In step 401, the behavior data of the primary user and the attribute data of the guest user are collected.

[0124] In the embodiment of the present disclosure, step 401 may correspond to the relevant description of the aforementioned step 101, and will not be described again here to avoid repetition.

[0125] In an optional method embodiment of the present disclosure, step C may be further included after step 401 .

[0126] In step C, the behavior data and attribute data are preprocessed, and the preprocessing includes at least one of data cleaning, time window screening, random sampling, and data integration.

[0127] In the disclosed embodiments, to improve the quality and usability of the collected data, the collected behavioral data and attribute data can be preprocessed before being processed by the fine-tuned large model. The preprocessing can include data cleaning, time window screening, random sampling, data integration, etc.

[0128] Data cleaning can be done by removing duplicate, erroneous, or abnormal collected data through methods such as data duplication checking and logical verification, to avoid interference with fine-tuning training of large models or task reasoning.

[0129] Time window filtering can be used to filter behavioral data within a certain time window in the collected data according to the actual application scenario and data characteristics, so as to focus on some behavioral data based on timeliness and relevance.

[0130] Random sampling can be performed randomly in the collected data to reduce the complexity and computational cost of data processing.

[0131] Data integration can be the aggregation and organization of data collected from different data sources for the same user. For example, the behavioral data of the primary user on different guest users and their corresponding published content can be integrated based on the behavior type and time sequence, and the attribute data of the guest users across platforms can be integrated to build a more complete and coherent behavioral sequence for the primary user, or the attribute data of the guest users can be integrated from multiple data sources and platforms to more accurately discover and mine the user's behavioral patterns and potential characteristics.

[0132] It should be noted that during the preprocessing process, data cleaning, time window screening, random sampling, and data integration can be performed individually or in combination. For example, random sampling can be performed on top of time window screening to ensure data representativeness while reducing data processing complexity and computational cost.

[0133] In step 402, the behavior data is input into the fine-tuned large model to obtain the main state sequence features corresponding to the main state user.

[0134] In the embodiment of the present disclosure, step 402 may correspond to the relevant description of the aforementioned step 102, and will not be described again here to avoid repetition.

[0135] In step 403, the attribute data is input into the fine-tuned macro model to obtain the guest attribute features corresponding to the guest user.

[0136] In the embodiment of the present disclosure, step 403 may correspond to the relevant description of the aforementioned step 103, and will not be described again here to avoid repetition.

[0137] In step 404, based on the primary state sequence characteristics and the guest state attribute characteristics, a preset recommendation strategy is adopted to recommend at least one guest state user to the primary state user.

[0138] In the embodiment of the present disclosure, step 404 may correspond to the relevant description of the aforementioned step 104, and will not be described again here to avoid repetition.

[0139] In step 405, multi-dimensional feedback data of the primary user is continuously obtained.

[0140] In the disclosed embodiments, after executing the recommendation task, feedback data from the primary user on the recommendation results can be continuously obtained. This feedback data can be multi-dimensional dynamic data, such as the primary user's click history, browsing history, favorite history, comments, likes or dislikes on the recommendation results, etc., to comprehensively, deeply, and fully reflect the primary user's acceptance of the recommendation results and potential interests and preferences.

[0141] In step 406, at least one of the large model and the preset recommendation strategy is updated using the feedback data.

[0142] In the disclosed embodiments, the large model, preset recommendation strategies, and the like can be updated based on feedback data. The large model can be updated by supplementing the feedback data with the primary user's behavioral data so that the large model can dynamically adjust its representation of the primary user. The large model can also be retrained, and its structure and parameters dynamically adjusted to ensure that it can promptly adapt to changes in the primary user's behavioral patterns and interests in social scenarios. The preset recommendation strategies can also be updated. For example, similarity calculation methods, data screening and sorting methods, recall source selection rules, and recommendation model selection and iteration can be collaboratively optimized based on the dynamically supplemented behavioral data, though this is not specifically limited in the disclosed embodiments. By continuously and dynamically supplementing feedback data, the primary user's behavioral data can be dynamically expanded and updated in real time, keeping pace with changes in the primary user's behavioral patterns and interests. Based on this, the large model and preset recommendation strategies can accurately and real-time capture the dynamic evolution trends of the primary user's behavioral patterns and interests based on the dynamically expanded behavioral data, providing more targeted and accurate data support for subsequent reasoning and analysis.

[0143] In the disclosed embodiments, updates to the large model and preset recommendation strategies based on feedback data can be performed regularly at set time intervals. The length of the time intervals can be flexibly adjusted based on recommendation needs and computing and operating costs, taking into full consideration the frequency of changes in the primary user's behavior patterns and interest preferences, the speed of data accumulation, and the efficiency of data processing. This avoids frequent updates due to too short time intervals, which would be costly and disruptive to normal business operations, and avoids untimely updates due to too long time intervals, which would affect the effectiveness and accuracy of recommendations. At the expiration of each time interval, the large model, preset recommendation strategies, and the like can be adjusted and updated based on dynamically supplemented behavioral data.

[0144] According to the large-scale model-based social user matching method of the disclosed embodiment, the behavioral data of the primary user and the attribute data of the guest user can be collected and respectively input into the fine-tuned large-scale model to obtain the primary sequence features corresponding to the primary user and the guest attribute features corresponding to the guest user; on this basis, a preset recommendation strategy can be adopted based on the primary sequence features and the guest attribute features to recommend the guest user to the primary user. In the disclosed embodiment, the fine-tuned large-scale model is used to characterize the behavioral data of the primary user and the attribute data of the guest user, respectively, which can fully, deeply, and comprehensively mine the long-term and short-term dependencies and hidden patterns in the user behavior sequence, avoid falling into convergence, effectively improve the recall rate and click-through rate, and have better generalization ability; at the same time, the fine-tuning training cost of the large-scale model is lower, and it can also be iteratively updated at low cost and high efficiency under data-driven conditions, thereby timely tracking the dynamic changes of users' social interests and avoiding lags.

[0145] Furthermore, the disclosed embodiments can also continuously obtain multi-dimensional feedback data from primary users, and use this feedback data to update the macro model, preset recommendation strategies, and the like. Through continuous dynamic updates and maintenance, the macro model-based social user matching method provided by the disclosed embodiments can continuously and effectively adapt to the dynamic changes in primary users' behavior patterns and interests in complex and ever-changing social scenarios, thereby maintaining and improving recommendation accuracy and providing more personalized, precise, and dynamically changing recommendation services.

[0146] Exemplary devices

[0147] After introducing the social user matching method based on the big model according to the exemplary embodiment of the present disclosure, Figure 5 A social user matching device based on a large model according to an exemplary embodiment of the present disclosure is described.

[0148] It should be noted that other specific details of the various functional modules of the social user matching device based on a large model in the embodiment of the present disclosure have been described in detail in the embodiment of the social user matching method based on a large model above, and will not be repeated here.

[0149] Figure 5 A social user matching device 500 based on a big model according to an exemplary embodiment of the present disclosure is shown. The device 500 may include: a data collection module 501 for collecting behavioral data of a primary user and attribute data of a guest user; a first characterization module 502 for inputting the behavioral data into a fine-tuned big model to obtain a primary sequence feature corresponding to the primary user; a second characterization module 503 for inputting the attribute data into the fine-tuned big model to obtain a guest attribute feature corresponding to the guest user; and a user recommendation module 504 for recommending at least one guest user to the primary user using a preset recommendation strategy based on the primary sequence feature and the guest attribute feature.

[0150] In an exemplary embodiment, the primary state sequence feature corresponds to the user identifier of the primary state user; the user recommendation module 504 is specifically used to respond to a recommendation request for a target user, obtain the target primary state sequence feature with the user identifier of the target user; perform similarity matching on the target primary state sequence feature and the guest state attribute feature, and determine at least one guest state user recommended to the target user based on the matching result.

[0151] In an exemplary embodiment, the primary state sequence feature corresponds to the user identifier of the primary state user; the user recommendation module 504 is specifically used to respond to a recommendation request for a target user and obtain the target primary state sequence feature with the user identifier of the target user; input the target primary state sequence feature and the guest state attribute feature into the recommendation model, and determine at least one guest state user recommended to the target user based on the recall result of the recommendation model.

[0152] In an exemplary embodiment, the device further includes a continuous updating module for continuously acquiring multi-dimensional feedback data of primary users; and updating at least one of the large model and the preset recommendation strategy with the feedback data.

[0153] In an exemplary embodiment, the continuous update module is specifically used to periodically obtain multi-dimensional feedback data of the primary user at a dynamic time interval, and the dynamic time interval is adjusted based on at least one of the data update frequency, data accumulation speed, and data processing efficiency.

[0154] In an exemplary embodiment, the large model includes an input layer and a base, and the device also includes a model training module for collecting historical behavior data of the main user, the historical behavior data including historical behavior sequence and historical next behavior; the historical behavior sequence is input into the large model, vector encoded through the input layer, and extracted and predicted by the base based on the self-attention mechanism to obtain the predicted next behavior output by the large model for the main user; the prediction loss is calculated based on the predicted next behavior and the historical next behavior; and the large model is fine-tuned with the prediction loss.

[0155] In an exemplary embodiment, the large model also includes a vector conversion layer, the weight parameters of the vector conversion layer are two-dimensional matrices, and the weight parameters represent the number of samples and the hidden layer dimension; the model training module is also used to update the vector conversion layer after each generation of fine-tuning training is completed.

[0156] In an exemplary embodiment, the vector encoding includes word unit encoding, word unit position encoding, word unit type encoding, and sequence position encoding.

[0157] In an exemplary embodiment, the self-attention mechanism includes self-attention calculation based on a sliding window mechanism, a hole sliding window mechanism, and a global information fusion sliding window mechanism.

[0158] In an exemplary embodiment, the prediction loss is a global cross entropy loss.

[0159] In an exemplary embodiment, the behavioral data includes social interaction behavior data of the primary user and the guest user; the attribute data includes at least one of identity identification data, identity feature data, social relationship data, physiological and life feature data, and personality and expression feature data.

[0160] In an exemplary embodiment, the data collection module 501 is further used to pre-process the behavior data and attribute data, and the pre-processing includes at least one of data cleaning, time window screening, random sampling, and data integration.

[0161] According to the large-scale model-based social user matching device of the disclosed embodiment, the behavioral data of the primary user and the attribute data of the guest user can be collected, and the data can be input into the fine-tuned large-scale model to obtain the primary sequence features corresponding to the primary user and the guest attribute features corresponding to the guest user; on this basis, a preset recommendation strategy can be adopted based on the primary sequence features and the guest attribute features to recommend the guest user to the primary user. In the disclosed embodiment, the fine-tuned large-scale model is used to characterize the behavioral data of the primary user and the attribute data of the guest user, respectively, which can fully, deeply, and comprehensively mine the long-term and short-term dependencies and hidden patterns in the user behavior sequence, avoid falling into convergence, effectively improve the recall rate and click-through rate, and have better generalization ability; at the same time, the fine-tuning training cost of the large-scale model is lower, and it can also be iteratively updated at low cost and high efficiency under data-driven conditions, thereby timely tracking the dynamic changes of users' social interests and avoiding lags.

[0162] It should be noted that although the detailed description above mentions several modules or units of the large-scale model-based social user matching device, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in a single module or unit. Conversely, the features and functions of a single module or unit described above can be further divided and embodied by multiple modules or units.

[0163] Exemplary Storage Media

[0164] The storage medium according to the exemplary embodiment of the present disclosure will be described below.

[0165] In this exemplary embodiment, referring to Figure 6 As shown, a program product 600 for implementing the above method according to an exemplary embodiment of the present disclosure is described. For example, a portable compact disk read-only memory (CD-ROM) may be used and includes program code, and can be run on a device such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0166] The program product 600 can be implemented in any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0167] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0168] The program code contained on the readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, RE, etc., or any suitable combination of the foregoing.

[0169] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, and the like, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device via any type of network, including a local area network (FAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0170] Exemplary electronic devices

[0171] refer to Figure 7 An electronic device according to an exemplary embodiment of the present disclosure will be described.

[0172] Figure 7 The electronic device 700 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0173] like Figure 7 As shown, electronic device 700 is implemented as a general-purpose computing device. Components of electronic device 700 may include, but are not limited to, at least one processing unit 710, at least one storage unit 720, a bus 730 connecting various system components (including storage unit 720 and processing unit 710), and a display unit 740.

[0174] The storage unit stores program codes, which can be executed by the processing unit 710, so that the processing unit 710 performs the steps according to various exemplary embodiments of the present disclosure described in the "Exemplary Method" section above. For example, the processing unit 710 can perform the following steps: Figure 1 、 2 Or the method steps shown in 4, etc.

[0175] The storage unit 720 may include a volatile storage unit, such as a random access memory unit (RAM) 721 and / or a cache memory unit 722 , and may further include a read-only memory unit (ROM) 723 .

[0176] The storage unit 720 may also include a program / utility 724 having a set (at least one) of program modules 725, such program modules 725 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0177] The bus 730 may include a data bus, an address bus, and a control bus.

[0178] The electronic device 700 can also communicate with one or more external devices 800 (e.g., a keyboard, a pointing device, a Bluetooth device, etc.), and such communication can be performed via an input / output (I / O) interface 750. The electronic device 700 also includes a display unit 740, which is connected to the input / output (I / O) interface 750 for display. In addition, the electronic device 700 can also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 760. As shown, the network adapter 760 communicates with other modules of the electronic device 700 via a bus 730. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0179] It should be noted that although several modules or submodules of the device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in a single unit / module. Conversely, the features and functions of a single unit / module described above can be further divided and embodied by multiple units / modules.

[0180] Furthermore, although the operations of the disclosed method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0181] Although the spirit and principles of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features in these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the appended claims.

Claims

1. A social user matching method based on a large model, characterized in that: The method comprises: Collect behavioral data of primary users and attribute data of guest users: Inputting the behavior data into the fine-tuned large model to obtain the main state sequence features corresponding to the main state user; Inputting the attribute data into the fine-tuned large model to obtain the guest attribute features corresponding to the guest user; Based on the main state sequence feature and the guest state attribute feature, a preset recommendation strategy is adopted to recommend at least one guest state user to the main state user.

2. The method according to claim 1, characterized in that The primary state sequence feature corresponds to the user identifier of the primary state user, and recommending at least one guest state user to the primary state user using a preset recommendation strategy based on the primary state sequence feature and the guest state attribute feature includes: In response to a recommendation request for a target user, obtaining a corresponding target main state sequence feature using the user identifier of the target user; Similarity matching is performed on the target main state sequence feature and the guest state attribute feature, and at least one guest state user recommended to the target user is determined according to the matching result.

3. The method according to claim 1, characterized in that The primary state sequence feature corresponds to the user identifier of the primary state user, and recommending at least one guest state user to the primary state user using a preset recommendation strategy based on the primary state sequence feature and the guest state attribute feature includes: In response to a recommendation request for a target user, obtaining a corresponding target main state sequence feature using the user identifier of the target user; The target main state sequence feature and the guest state attribute feature are input into a recommendation model, and at least one guest state user recommended to the target user is determined according to a recall result of the recommendation model.

4. The method according to claim 1, wherein After recommending at least one guest state user to the primary state user based on the primary state sequence feature and the guest state attribute feature, the method further includes: Continuously obtain multi-dimensional feedback data of the primary user; At least one of the large model and the preset recommendation strategy is updated using the feedback data.

5. The method according to claim 4, characterized in that The continuously obtaining feedback data of the primary user includes: Multi-dimensional feedback data of the primary user is periodically obtained at a dynamic time interval, and the dynamic time interval is adjusted based on at least one of data update frequency, data accumulation speed, and data processing efficiency.

6. The method according to claim 1, characterized in that The large model includes an input layer and a base. The fine-tuning steps of the large model are as follows: Collecting historical behavior data of the primary user, wherein the historical behavior data includes a historical behavior sequence and a historical next behavior; The historical behavior sequence is input into the large model, vectorized and encoded through the input layer, and extracted and predicted by the base based on the self-attention mechanism to obtain the predicted next behavior output by the large model for the primary user; Calculating a prediction loss based on the predicted next behavior and the historical next behavior; The large model is fine-tuned with the prediction loss.

7. The method according to claim 6, characterized in that The large model also includes a vector conversion layer, the weight parameters of the vector conversion layer are two-dimensional matrices, and the weight parameters represent the number of samples and the hidden layer dimension. The vector conversion layer is updated after each generation of fine-tuning training.

8. A social user matching device based on a large model, characterized in that: The device comprises: The data collection module is used to collect the behavioral data of primary users and the attribute data of guest users: A first characterization module is configured to input the behavior data into a fine-tuned large model to obtain a main state sequence feature corresponding to the main state user; A second characterization module is configured to input the attribute data into the fine-tuned large model to obtain guest attribute features corresponding to the guest user; The user recommendation module is configured to recommend at least one guest state user to the primary state user by adopting a preset recommendation strategy based on the primary state sequence feature and the guest state attribute feature.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the social user matching method based on a large model according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: processor; as well as a memory for storing executable instructions of the processor; The processor is configured to execute the large model-based social user matching method according to any one of claims 1 to 7 by executing the executable instructions.

Images