Social e-commerce resource matching method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of human-computer interaction, and provides a social e-commerce resource matching method and device, electronic equipment and a storage medium. The method comprises the following steps: performing multi-source heterogeneous data preprocessing on an original user behavior data set to obtain a standardized time-space behavior sequence, performing label extraction on the standardized time-space behavior sequence to obtain a user interest label system, and performing dynamic attenuation weight distribution and semantic enhancement processing on the user interest label system to obtain a time-space weighted user portrait vector; meanwhile, performing scene knowledge graph embedding and functional entity extraction processing on the commodity description text to obtain a commodity scene feature matrix, and performing multi-target matching degree calculation according to the space-time weighted user portrait vector and the commodity scene feature matrix to obtain a user-commodity matching probability tensor; and performing conflict resolution and global optimal allocation processing on the user-commodity matching probability tensor to obtain a resource allocation table. Through data processing and machine learning, the resource matching precision of the social e-commerce platform is improved.

Description

Technical Field

[0001] This application relates to the field of human-computer interaction technology, and in particular, to a social e-commerce resource matching method, device, electronic device, and storage medium. Background Art

[0002] With the continuous expansion of the functions of social platforms, a typical integration function of social platform + shopping represented by sending gifts on WeChat has emerged. In traditional social interactions, giving gifts has always been an important means to enhance feelings and express intentions. Combining gift-giving with online shopping greatly improves the convenience and immediacy of gift-giving. Social platforms also provide users with a fast experience of gift selection, purchase, sending, and receiving, lowering the threshold of gift-giving and making gift delivery no longer restricted by geography and time.

[0003] The existing resource matching method for social platform group chats (i.e., the gift-giving gameplay) adopts a random allocation method. However, as a functional carrier, gifts can only meet the needs of users within a certain functional range, and there are significant differences in the interests and tastes of users participating in gift lotteries. This makes the gifts unable to meet the needs or preferences of the lottery users, resulting in gifts being transferred or rejected and failing to fulfill the wishes of the gift-giving users. Summary of the Invention

[0004] In view of this, this application provides a social e-commerce resource matching method, device, electronic device, and storage medium to solve the problem of inaccurate social resource matching.

[0005] The first aspect of this application provides a social e-commerce resource matching method, and the method includes: Performing target data analysis on the first target device and the second target device according to a preset data collection method to obtain a commodity description text and an original user behavior data set; Performing multi-source heterogeneous data preprocessing on the original user behavior data set to obtain a standardized spatio-temporal behavior sequence; Performing user interest tag extraction on the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and performing dynamic decay weight assignment and semantic enhancement processing on the user interest tag system to obtain a spatio-temporal weighted user portrait vector; Performing scene knowledge graph embedding and functional entity extraction processing on the commodity description text to obtain a commodity scene feature matrix; Calculating a multi-objective matching degree according to the spatio-temporal weighted user portrait vector and the commodity scene feature matrix to obtain a user-commodity matching probability tensor; Performing conflict resolution and global optimal allocation processing on the user-commodity matching probability tensor to obtain a resource allocation table.

[0006] In an alternative embodiment, the preprocessing of the multi-source heterogeneous data for the original user behavior dataset to obtain a standardized spatio-temporal behavior sequence includes: Perform UTC unified conversion on the timestamps in the original user behavior dataset to obtain spatio-temporally aligned data; Perform Geohash encoding processing on the geographical location information in the original user behavior dataset to obtain a spatio-temporal joint index; Filter the abnormal behavior data from the original user behavior dataset according to the spatio-temporal joint index and the spatio-temporally aligned data to obtain an abnormal behavior filtered dataset; Perform Z-score standardization processing on the numerical fields in the abnormal behavior filtered dataset to obtain standardized numerical data; Perform One-Hot encoding processing on the discrete fields in the abnormal behavior filtered dataset to obtain sparse vectors; Stitch and fuse the standardized numerical data and the sparse vectors according to a preset stitching method to obtain the standardized spatio-temporal behavior sequence.

[0007] In an alternative embodiment, the extraction of user interest tags from the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and the dynamic decay weight assignment and semantic enhancement processing of the user interest tag system to obtain a spatio-temporally weighted user portrait vector includes: Perform semantic parsing processing on the target words in the standardized spatio-temporal behavior sequence through a preset BERT-Base model to construct an initial tag system; Perform dynamic decay processing on the initial tag system according to the time difference between the timestamp in the standardized spatio-temporal behavior sequence and the current time to obtain a decay-optimized tag system; Determine the time decay factor of each tag in the decay-optimized tag system according to the time difference, and perform weighted processing on the decay-optimized tag system according to the time decay factor to obtain a time-weighted tag system; Perform geographical clustering processing on each tag in the time-weighted tag system according to the spatio-temporal joint index to obtain the regional tag corresponding to each tag in the time-weighted tag system, and merge the regional tag with the time-weighted tag system to generate the user interest tag system; Perform weight stitching processing on each tag in the user interest tag system in a preset dimension to obtain the spatio-temporally weighted user portrait vector.

[0008] In an alternative embodiment, the embedding of the scene knowledge graph and the extraction of functional entities from the commodity description text to obtain a commodity scene feature matrix includes: Tokenize the commodity description text to extract functional words and scenario keywords; Perform named entity recognition on the functional words through a pre-set BiLSTM-CRF model to obtain a list of functional entities; Match the list of functional entities and the scenario keywords according to a pre-set domain dictionary to obtain a commodity scenario knowledge graph; Perform graph embedding on the commodity scenario knowledge graph to obtain entity vectors for each scenario entity; Calculate the degree of association between the scenario entities and the scenario keywords in the commodity scenario knowledge graph to obtain the relationship strength for each scenario entity; Weight the entity vectors according to the relationship strength to obtain scenario entity vectors; Construct a matrix for the scenario entity vectors to obtain the commodity scenario feature matrix.

[0009] In an alternative embodiment, the multi-objective matching degree calculation based on the spatio-temporal weighted user profile vector and the commodity scenario feature matrix to obtain a user-commodity matching probability tensor includes: Perform matrix multiplication on the spatio-temporal weighted user profile vector and the commodity scenario feature matrix to obtain a matching degree score between the user profile and the commodity scenario; Calculate the cosine similarity between each user and each commodity scenario according to the matching degree score, the weights of each label in the spatio-temporal weighted user profile vector, and the scenario entity vectors in the commodity scenario feature matrix to obtain an initial matching degree matrix; Normalize each matching degree score in the initial matching degree matrix according to the quantities of each commodity recorded in the commodity description text to obtain a normalized matching degree score; Determine the priority of each user according to the historical behavior data in the spatio-temporal weighted user profile vector, and weight the normalized matching degree scores according to the priority to obtain an optimized matching degree matrix; Perform Softmax probability mapping on the optimized matching degree matrix to obtain an initial user-commodity matching probability tensor; Normalize the initial user-commodity matching probability tensor to obtain the user-commodity matching probability tensor.

[0010] In an alternative embodiment, the conflict resolution and global optimal allocation processing of the user-commodity matching probability tensor to obtain a resource allocation table includes: Sort the matching probabilities between each user and each product in the user-product matching probability tensor to determine the preliminary candidate product set for each user; Sort the matching candidate users for each product according to the preliminary candidate product set and the number of products to generate the preliminary candidate user set for each product; Determine a preset table conversion method according to the preliminary candidate user set, and perform format conversion on the preliminary candidate user set according to the table conversion method to obtain the resource allocation table.

[0011] In an alternative embodiment, the method further includes: When obtaining the order data sent by the first target device, compare the number of devices of the second target device with the number of text in the product description text; When the number of devices is greater than the number of text, perform data partitioning on the order data according to the number of devices and a preset partitioning method to obtain order sub-data fragments; Update the preset data fragment allocation table according to a preset allocation method, the order sub-data fragments, and the second target device; According to a preset order data integrity monitoring method, perform order sub-data fragment classification and data integrity judgment on the order sub-data fragments corresponding to each second target device in the data fragment allocation table; When the order sub-data fragment passes the data integrity judgment, perform target data extraction processing on the data fragment allocation table according to the order sub-data fragment to obtain the resource allocation table.

[0012] The second aspect of the present application provides a social e-commerce resource matching device, and the device includes: A data acquisition module, configured to perform target data analysis on the first target device and the second target device according to a preset data acquisition method to obtain a product description text and an original user behavior data set; A spatio-temporal standard module, configured to perform multi-source heterogeneous data preprocessing on the original user behavior data set to obtain a standardized spatio-temporal behavior sequence; A portrait vector module, configured to extract user interest tags from the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and perform dynamic decay weight allocation and semantic enhancement processing on the user interest tag system to obtain a spatio-temporally weighted user portrait vector; A feature matrix module, configured to perform scene knowledge graph embedding and functional entity extraction processing on the product description text to obtain a product scene feature matrix; A probability tensor module for calculating multi-object matching degrees based on the spatio-temporal weighted user profile vector and the commodity scenario feature matrix to obtain a user-commodity matching probability tensor; A resource allocation module for performing conflict resolution and global optimal allocation processing on the user-commodity matching probability tensor to obtain a resource allocation table.

[0013] A third aspect of the present application provides an electronic device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the social e-commerce resource matching method described above are implemented.

[0014] A fourth aspect of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the social e-commerce resource matching method described above are implemented.

[0015] In summary, the present application at least includes the following beneficial technical effects: 1. By combining the spatio-temporal behavior data of users and the scenario features of commodities to calculate multi-object matching degrees, the interest preferences and behavior characteristics of users are deeply mined, and combined with the function and scenario information of commodities, the gift allocation is made more in line with the needs of users, improving the accuracy of gift allocation and the satisfaction of users.

[0016] 2. By dynamically decaying and semantically enhancing user interest tags and combining the semantic parsing of the spatio-temporal behavior sequence by the BERT-Base model, the user profile can reflect their latest behavior characteristics and interest changes in real time.

[0017] 3. Through the embedding of the scenario knowledge graph and the extraction of functional entities, the scenario features of commodities are deeply understood, and an accurate commodity scenario feature matrix is constructed, enabling the system to better identify the matching degree between commodities and users, and thus making more personalized allocations according to the scenarios and usage environments where the commodities are located. Description of the Drawings

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a flowchart of a social e-commerce resource matching method provided by an embodiment of the present application; Figure 2It is a functional module diagram of a social e-commerce resource matching device provided by an embodiment of the present application; Figure 3 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Specific embodiments

[0020] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0021] It should be understood that the social e-commerce resource matching method is used to match the winning users and the gift-giving products during the process of "giving gifts" in the social groups of the social platform, so as to improve the satisfaction of the users who receive the gifts. The social e-commerce resource matching method provided by the present application is executed by the server. Correspondingly, the social e-commerce resource matching device runs in the server. The server is grafted onto the social platform (such as WeChat, QQ, etc.) through a secondary domain name, resolves and binds the secondary domain name to the social platform system, obtains the social platform account sales interface through the secondary domain name grafted onto the social platform account, enables micro-payment, and establishes its own ecosystem. Specifically, the gift-giving user purchases products through the online shopping system of the online mall. After successful payment and order placement, the order is grouped and sent to the corresponding social group through a link. The users in the social group participate in the gift-snatching activity by clicking on the link. The server selects the corresponding number of winning users from the users participating in the gift-snatching activity according to the number of products. The server transfers and binds the preset disposal rights of the products to the winning users. Therefore, the winning users obtain operation permissions including receiving, transferring, returning, and exchanging products.

[0022] Next, the social e-commerce resource matching method provided by the embodiments of the present application will be described from the perspective of the server in combination with the process of giving gifts for multiple products in the same social group.

[0023] As Figure 1 shown, it is a flowchart of the social e-commerce resource matching method provided by the embodiments of the present application. The social e-commerce resource matching method provided by the embodiments of the present application includes the following steps.

[0024] Step S1: Perform target data analysis on the first target device and the second target device according to a preset data collection method to obtain product description text and an original user behavior data set.

[0025] Among them, the first target device is the mobile device of the gift-giving user. The gift-giving user can perform related operations such as selecting gifts in the online shopping mall system through the first target device, sending gift requests to the social platform, and setting lottery parameters (such as the number of gift-receiving users). The second target device is the mobile device of the gift-receiving user, and the gift-receiving user is randomly selected from the users participating in the lottery in the social group through a random function preset by the server.

[0026] The product description text is used to record the detailed information of each product, including the product's functions, features, applicable scenarios, etc. The server collects product description data (i.e., product description text) from the corresponding online shopping mall system through the API interface according to the order information of the products selected by the gift-giving user. Each product description text is usually stored in a structured format, and the product description text includes, but is not limited to, text fields, picture or video descriptions, prices, applicable scenarios, and other contents. The server can perform natural language processing and analysis on the product description to extract the main features of the product (such as "smartphone", "sports shoes", "desk", etc.).

[0027] The original user behavior dataset refers to the user data collected by the server from the second target device. User data includes, but is not limited to, product browsing records (such as browsing time, browsing frequency, categories of browsed products on the shopping platform, etc.), search records, interaction records (i.e., user interaction behaviors in the social platform, such as comments, likes, shares, etc.), shopping cart records, and purchase history records, etc. The original user behavior dataset is collected in real time through the event tracking mechanism of the application or web page on the second target device (such as user session tracking, clickstream analysis). When the gift-receiving user interacts with the platform (such as browsing products, searching for products, etc.), the second target device will record these behaviors and upload them to the server for storage and analysis. It should be understood that when obtaining the original user behavior dataset, it is necessary to ensure the protection of user privacy. The server needs to obtain the authorization of the user (such as through an explicit consent data privacy policy) before collecting their behavior data. The collection of user data should comply with relevant laws and regulations to ensure the user's informed consent to data collection and use.

[0028] After the server obtains the data sent by the online shopping mall system or the second target device, it cleans and analyzes the collected product description text through natural language processing technology. The cleaning and analysis process includes removing irrelevant characters, part-of-speech tagging, keyword extraction, etc. Finally, the content of the product description text will be converted into structured data (such as product functions, applicable scenarios, product types, etc.). The original user behavior dataset is cleaned and standardized to convert the user behavior dataset into a structured dataset containing information such as user ID, behavior type, product ID, timestamp, etc.

[0029] Step S2: Perform multi-source heterogeneous data preprocessing on the original user behavior dataset to obtain a standardized spatio-temporal behavior sequence.

[0030] The timestamps in the original user behavior dataset may be recorded in different time zone formats, resulting in inconsistent time information among the data. To ensure time alignment in subsequent analyses, it is first necessary to uniformly convert the timestamps of all data in the original user behavior dataset to UTC. UTC is a standard time format widely used for unified coordination among global time zones. Specifically, traverse each data record in the user behavior dataset to obtain the timestamp information, and convert the time zone information to the UTC time format according to the time zone information of each data. The time of each data in the converted spatio-temporal aligned data will be used as a time feature to ensure the consistency of time information in subsequent analyses.

[0031] Meanwhile, the geographical location information in the original user behavior dataset is usually recorded in the form of longitude and latitude, but different records may cover different geographical regions. To enable unified processing of the geographical location information in all data in subsequent analyses, it is necessary to use the Geohash coding technique to convert the longitude and latitude information into a Geohash code, which can represent different precisions of geographical locations. Specifically, obtain the longitude and latitude information in the user behavior data, and use the Geohash algorithm to encode the longitude and latitude information to obtain a Geohash string of a fixed length. The precision of the Geohash coding can be adjusted according to requirements. Usually, the coding precision is set to 8 digits or higher to ensure a relatively fine-grained representation of geographical information. Further, associate the UTC time after timestamp conversion and the obtained Geohash code to obtain a spatio-temporal joint index.

[0032] The original user behavior dataset may contain some abnormal behavior records (i.e., abnormal behavior data), such as abnormal clicks and high-frequency browsing of users. Abnormal behavior data will affect subsequent data analyses and produce misleading results. Therefore, it is necessary to filter the abnormal behavior data. Specifically, according to the spatio-temporal joint index, by comparing the normal behavior patterns of users (such as browsing frequency, click pattern, etc.), use the Z-score algorithm or model-based anomaly detection methods (e.g., Isolation Forest) to identify and filter out the data that does not conform to the normal behavior. In the implementation of this application, the Z-score algorithm is used to calculate the Z-score value of each behavior record in the original user behavior dataset, and filter out the behavior records whose Z-score value exceeds the preset anomaly threshold, thereby obtaining an abnormal behavior filtered dataset. The Z-score algorithm can be expressed as the following specific formula: where, Represents the Z-score value after standardization, represents the original user behavior data (e.g., number of clicks, dwell time, etc.), represents the mean of the behavior data, represents the standard deviation of the behavior data. Filter abnormal behavior data through Z-score to ensure the quality and accuracy of the dataset.

[0033] The numerical fields in the original user behavior dataset (e.g., dwell time, number of purchases, etc.) may have different dimensions or orders of magnitude, so it is necessary to perform Z-score standardization on these numerical values. After standardization, all numerical data will have the same scale, avoiding unfair impacts on the model caused by different scales of different features. The Z-score algorithm screening method and principle for Z-score standardization of numerical fields are similar to those for filtering abnormal behavior data, and the details will not be elaborated here. For details, please refer to the operation of filtering abnormal behavior data. After performing Z-score standardization on the numerical fields, the scales of different numerical fields in the obtained standardized numerical data are unified, avoiding over-reliance of subsequent models on certain fields.

[0034] The discrete fields in the user behavior dataset (e.g., behavior type, product category, etc.) need to be converted into a format that can be processed by machine learning algorithms. One-Hot encoding is a commonly used conversion method that converts discrete categorical variables into a sparse vector, with only one position in each vector being 1 and the other positions being 0. Specifically, for each discrete field (e.g., the "behavior type" field, product category, etc.), it is first necessary to determine all its possible values. These values form the value space or category space of the field. Allocate these possible values to different positions in the vector in a certain order (usually alphabetical order or insertion order). For each discrete value, generate a vector with the same size as the value space, and mark the index position corresponding to this discrete value in the vector as 1, and the other positions as 0.

[0035] Exemplarily, there is a field "behavior type" with three set values: "click", "purchase", and "browse". At the same time, the values of the "behavior type" field are allocated indexes in the shown order: "click" corresponds to index 0; "purchase" corresponds to index 1; "browse" corresponds to index 2. When the user's "behavior type" is "click", the corresponding One-Hot encoded vector is [1, 0, 0]; when the user's "behavior type" is "purchase", the corresponding One-Hot encoded vector is [0, 1, 0]; when the user's "behavior type" is "browse", the corresponding One-Hot encoded vector is [0, 0, 1].

[0036] Finally, the numerical data and the sparse vectors are fused according to a preset splicing method to form a standardized spatio-temporal behavior sequence. Among them, the standardized spatio-temporal behavior sequence is a feature vector, which is convenient for subsequent analysis and model use.

[0037] Exemplarily, assume that the numerical data is , and the sparse vector is . Then the standardized spatio-temporal behavior sequence after splicing can be expressed by the following formula: Wherein, represents the final spliced spatio-temporal behavior sequence, represents the standardized numerical data, represents the sparse vector after One-Hot encoding.

[0038] Step S3: Extract user interest tags from the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and perform dynamic decay weight assignment and semantic enhancement processing on the user interest tag system to obtain a spatio-temporally weighted user portrait vector.

[0039] Among them, the spatio-temporally weighted user portrait vector is a comprehensive feature vector generated based on multi-dimensional information such as the user's behavior data, time information, geographical location, and user interests. It synthesizes the timeliness of behavior and geographical location factors, and can reflect the interests and needs of the recipient users. By extracting the user's interest tags from the standardized spatio-temporal behavior sequence, and performing dynamic decay, semantic enhancement, and spatio-temporal weighting processing on the interest tags, the spatio-temporally weighted user portrait vector is obtained.

[0040] First, extract the target words in the standardized spatio-temporal behavior sequence. The target words are usually keywords or important items in the user behavior data, such as "smartphone", "festival promotion", or "office equipment", etc. The target words can reflect the user's interest direction. Thus, the Bidirectional Encoder Representations from Transformers (BERT) model is used to comprehensively capture the semantic information of the words based on the context of the target words. The BERT-Base model can effectively understand the meaning of the target words in the context through its bidirectional encoding characteristics, so as to accurately extract the user's interest tags.

[0041] Specifically, target words in the standardized spatio-temporal behavior sequence (such as product categories, user behavior types, etc.) are input into a pre-set BERT-Base model for encoding. The BERT-Base model learns the semantic representation of the target words through context relationships and generates semantic vectors of the target words. Among them, the semantic vectors of the target words contain the meanings of the target words in specific situations and the interest characteristics of users. Through the output of the BERT-Base model, an initial user interest label system is constructed, and each label represents the user's interest in a certain field or product category.

[0042] Since the user's demand for products is affected by time and location, generally, the longer the time, the greater the attenuation of the user's demand. The demand for the same product is also different in different geographical locations (for example, the demand for humidifiers among users in the south is relatively small, while the demand for humidifiers among users in the north is relatively large). In order to make the user portrait more timely and dynamic, it is necessary to dynamically assign attenuation weights to each label in the initial label system according to time and geographical information.

[0043] First, calculate the time difference of each label according to the difference between the current time and the timestamp of the behavior corresponding to the label, and then determine the time decay factor according to the time difference and the pre-set mapping table. The time decay factor usually uses the exponential decay formula, and the decay factor gradually decreases over time. After obtaining the time decay factor of each label, calculate the weight impact of the time decay factor on each label according to the following dynamic decay formula, and then perform weighted processing on the weight of each label according to the time decay factor to obtain the time-weighted label system.

[0044] Among them, represents the weight of the label in the initial interest label system. represents the time decay rate, which determines the speed of decay. The time difference is the gap between the current time and the behavior time corresponding to the label. represents the weight of the label after attenuation. By adjusting the weight of each label, it is ensured that the weight of the label after attenuation is more in line with the current needs and interests of the user.

[0045] Furthermore, based on the spatio-temporal joint index, geographical clustering is performed on each tag, so as to group user behavior tags with similar geographical locations, thereby improving the geographical relevance of the user profile. Specifically, according to the geographical information of each tag in the time-weighted tag system, common clustering algorithms (such as K-Means) are used to group the tags and assign the tags to different geographical region tags. The geographical location of each tag is associated with the spatio-temporal joint index in the user behavior data. Finally, the tag system after geographical clustering is merged with the time-weighted tag system to generate a user interest tag system. The user interest tag system contains tag information optimized by time decay and geographical clustering, and can display the preferences of users within a specific time and geographical region. Thus, it can better reflect the interests and needs of users.

[0046] Perform weight splicing processing on each tag in the user interest tag system. Splice the weight information of each tag with its corresponding semantic features, time information, and geographical information into a high-dimensional feature vector, and finally form a spatio-temporal weighted user profile vector. The splicing information of each tag includes the weight of the tag, the time decay factor, the geographical clustering information, etc. The spliced vector will be used as the spatio-temporal weighted user profile vector for subsequent matching and recommendation. Among them, the splicing process can be expressed by the following formula: Among them, The finally spliced spatio-temporal weighted user profile vector, is the weight of the tag , is the time information of the tag (i.e., the decay factor), is the geographical information of the tag (i.e., the region tag). The spatio-temporal weighted user profile vector provides an accurate representation of user needs for subsequent personalized recommendation and resource matching.

[0047] Step S4: Perform scene knowledge graph embedding and functional entity extraction processing on the commodity description text to obtain a commodity scene feature matrix.

[0048] Among them, the commodity scene feature matrix integrates multiple dimensions of information of the commodity (such as function, scene, use, etc.), and can understand the potential use value of the commodity. For example, whether a certain commodity is suitable for a specific festival, whether it is suitable for use in an office environment, etc., so that the scene information of the commodity can be used as an input in machine learning and recommendation models to help the algorithm perform more accurate resource allocation.

[0049] First, by performing word segmentation on the product description text, the text is decomposed into individual words (or sub-words). Word segmentation is a basic task in natural language processing. Through word segmentation, the text can be split into basic language units. When performing word segmentation, it is necessary to preprocess the text according to the characteristics of the product description, remove irrelevant characters (such as punctuation marks), and standardize special symbols and units. From the results of word segmentation, keywords related to the product function and scenario are extracted. For example, "waterproof", "high performance", "suitable for outdoors", etc. in the product description can be used as function words; "festival gift giving", "office needs" can be used as scenario keywords.

[0050] Furthermore, the BiLSTM-CRF model is used to perform named entity recognition on the function words after word segmentation. Both the BiLSTM-CRF model and the conditional random field model are deep learning models widely used in sequence labeling tasks and are suitable for processing named entities in text (such as product functions, brands, types, etc.). The BiLSTM-CRF model will identify and label the function entities in the product description through context information. For example, "smartphone", "high-definition TV" in the product description will be labeled as function entities. Thus, a preset domain dictionary is used to match the function entities and scenario keywords extracted from the product description. Among them, the domain dictionary contains predefined keywords related to the product (such as "smartphone", "office equipment", "home theater", etc.) and is used to match the entity words and scenario keywords extracted from the product description with the actual defined categories. Through the matching process, it can be ensured that each entity word and scenario keyword can be associated with a well-defined category, thereby constructing a product scenario knowledge graph. Among them, the product scenario knowledge graph contains the relationships between the function entities (such as "waterproof", "high performance") and scenario entities (such as "festival gift", "home entertainment") of the product.

[0051] To facilitate subsequent data processing, it is necessary to perform graph embedding on the product scenario knowledge graph, thereby converting the complex graph structure into a low-dimensional numerical vector (i.e., entity vector). Graph embedding converts the entities in the knowledge graph (such as "waterproof", "festival gift giving", etc.) into vector form according to the connection relationships of the nodes (i.e., entities) in the product scenario knowledge graph.

[0052] Furthermore, a product may involve multiple scenarios, and different products have different applicability in different scenarios. By calculating the degree of association between the scenario entities and scenario keywords in the product scenario knowledge graph, these complex multi-dimensional relationships can be quantified, enabling the server to better understand and express the potential of the product in different scenarios. The relationship strength between each scenario entity and other scenario keywords is obtained through the following degree of association calculation formula to determine their closeness.

[0053] Among them, and are two scene entities in the commodity scene knowledge graph respectively. Embed is the embedding vector of the scene entity. Sim is the similarity between scene entities.

[0054] Furthermore, according to the relationship strength between entities, adjust the weight of each scene entity in the feature vector. Entities with stronger relationship strength will occupy a more important position in the generated scene entity vector. Through the following formula, weight the vector of each scene entity according to the relationship strength.

[0055] Among them, The weighted scene entity vector. Sim is the relationship strength between scene entities. Embed is the embedding vector of the scene entity . Through the weighting process, it can be ensured that the scene entity vector can better reflect the relationship between entities and its impact on commodity functions.

[0056] Finally, by concatenating all the weighted scene entity vectors, construct a commodity scene feature matrix. Each row in the commodity scene feature matrix represents the scene features of a commodity, including the functions, applicable scenes of the commodity, and the relationships between them. The commodity scene feature matrix provides detailed scene information for each commodity, which is convenient for subsequent resource allocation and recommendation algorithms to use.

[0057] Step S5, calculate the multi-objective matching degree according to the spatio-temporal weighted user portrait vector and the commodity scene feature matrix to obtain a user-commodity matching probability tensor.

[0058] Among them, the user-commodity matching probability tensor is a multi-dimensional data structure that records the matching probability between each user and each commodity. Quantify the matching degree between each user and each commodity through the user-commodity matching probability tensor, which helps the server to more accurately understand the relationship between user needs and commodity characteristics, so as to optimize the recommendation and resource allocation process.

[0059] First, through matrix multiplication processing, multiply the spatio-temporal weighted user portrait vector by the commodity scene feature matrix to obtain the matching degree score between the user and the commodity. Among them, the spatio-temporal weighted user portrait vector is obtained by spatio-temporally weighting the user behavior data, which contains the interests and needs of the user. The commodity scene feature matrix represents the features such as functions and scenes of each commodity. The matrix multiplication can be expressed as the following calculation formula: ​​ Among them, is the spatio-temporal weighted portrait vector representing the user . is the scene feature vector representing the product . is the matching degree score between the user and the product , representing the interest intensity of the user in the product scene.

[0060] At the same time, the similarity between the spatio-temporal weighted user portrait vector and the product scene feature vector is calculated through cosine similarity, further refining the matching degree score. The cosine similarity is calculated through the following formula: Among them, represents element-wise multiplication, represents the label weight vector of the user , represents the scene feature vector of the product , represents the label interest vector of the user . The contribution degree of the user interest vector is adjusted through the label weight, enhancing the influence of high-weight labels on the matching result.

[0061] Furthermore, the initial matching degree matrix is further calculated using the matching degree score and cosine similarity. The initial matching degree matrix is a two-dimensional matrix representing the matching intensity between all users and all product scenes. The initial matching degree matrix is obtained through the following calculation formula: Among them, is the initial matching degree matrix, representing the comprehensive matching degree between the user and the product scene . is the matching degree score between the user and the product scene . CosineSim is the cosine similarity between the user and the product scene . The initial matching degree matrix combines the matching degree score and cosine similarity, reflecting the comprehensive interest degree of the user in the product scene.

[0062] After calculating the initial matching degree scores, it is necessary to consider the quantity of various commodities in the link. Therefore, it is necessary to count the quantity according to the types of commodities recorded in the commodity description text, so as to normalize each matching degree score in the initial matching degree matrix to ensure that the matching degree score of each commodity can reflect its actual supply quantity. The normalization process adjusts the matching scores of different commodities to the same scale, so that the quantity of commodities will not affect the final resource allocation. The normalized matching degree score can be obtained through the following formula: Where, is the matching degree score between the normalized user and the commodity . is the matching degree score between the user and the commodity . is a set of commodities, representing all possible commodities. is the sum of the matching degree scores of the user and all commodities .

[0063] Perform a Softmax probability mapping on the optimized matching degree scores to convert them into probability values. The Softmax function converts the scores into probability values between 0 and 1, representing the matching probability of each commodity for each user. The Softmax probability mapping can be performed using the following formula: Where, is the matching probability between the user and the commodity . is the optimized matching degree score. is the exponential operation on the optimized matching degree score. is the sum of the exponentialized matching degree scores of all commodities. The Softmax mapping ensures that the sum of the matching probabilities of all commodities is 1.

[0064] Finally, normalize the obtained initial user-commodity matching probability tensor to ensure that the sum of the matching probabilities of each recipient user is 1, so that the resource allocation is fair among all recipient users. The normalized user-commodity matching probability tensor can be obtained through the following formula: Where, is the finally normalized matching probability. Initial matching probability. is the sum of the matching probabilities of all commodities for the user .

[0065] Step S6: Perform conflict resolution and global optimal allocation processing on the user-item matching probability tensor to obtain a resource allocation table.

[0066] In the user-item matching probability tensor, there may be a problem of competition for item resources among multiple recipient users. By performing conflict resolution and global optimal allocation on the user-item matching probability tensor, the resource competition problem between users and items is solved, and fair and optimal resource allocation is carried out according to the matching probability and user priority.

[0067] After obtaining the user-item matching probability tensor, sort each user and each item according to the matching probability score from high to low, so as to determine the matching priority between the user and the item according to the matching degree score, and generate a preliminary candidate item set for the user. The preliminary candidate item set contains the items that the user is most likely to be interested in, sorted according to the matching probability from high to low.

[0068] After determining the preliminary candidate item set, further screen the preliminary candidate item set for each item's recipient users by combining the quantity of each type of item in the items. Since the quantity of items is limited, it is necessary to sort each item to determine which users will receive the item from the candidate user set. For each item, sort the candidate users according to the matching degree score. Specifically, calculate the matching degree score between each user and the item, and sort all candidate users from high to low according to the score. After obtaining the sorted candidate users, determine the number of users that can be finally selected from the preliminary candidate user set according to the quantity of each item corresponding to the item. For example, assume the quantity of the item is 10, then select the first 10 users from the sorted candidate user set as the candidate user set for this item.

[0069] It should be understood that through the preliminary candidate item set, the candidate item sorting for each recipient user can be obtained. Combining the preliminary candidate item set and the quantity of each item, the preliminary candidate user set for each item that meets the item quantity limit can be screened. In this process, there may be a situation where multiple users compete for the same item during the screening of recipient users, resulting in a situation where the number of candidate users in the finally obtained preliminary candidate user set is greater than the quantity of items in the link.

[0070] Exemplarily, assume that the quantity of Product A is 1 piece. After sorting by the matching probability score, the preliminary candidate product sets of User 1 and User 2 are obtained. Product A ranks first in the preliminary candidate product sets of both User 1 and User 2. During the process of screening the receiving users, User 1 and User 2, for whom Product A ranks first in the preliminary candidate product set, are selected as candidate users according to the preliminary candidate product set, and the matching degree scores of candidate User 1 and candidate User 2 for Product A are obtained. However, the matching degree scores of User 1 and User 2 for Product A are the same and are the highest scores among all the pending users. Therefore, when sorting the pending users according to the matching degree scores, User 1 and User 2 are tied for the first place. When generating the preliminary candidate user set of Product A according to the user ranking, all the candidate users ranked second and after are deleted according to the product quantity, and only User 1 and User 2 recorded in the first place are retained, and the preliminary candidate user set of Product A is generated, which results in the number of candidate users recorded in the preliminary candidate user set exceeding the product quantity.

[0071] After obtaining the preliminary candidate user set, the number of candidate users in the preliminary candidate user set is compared with the product quantity. When the number of candidate users in the preliminary candidate user set is less than or equal to the product quantity, each product in the preliminary candidate user set and its corresponding allocated user information are formatted through a preset table conversion method (for example, two-dimensional table or JSON format) to generate a resource allocation table. The resource allocation table lists which users each product is allocated to and which products each user obtains.

[0072] When the number of candidate users in the preliminary candidate user set is greater than or equal to the product quantity, a maximum matching based on the Hungarian algorithm is performed according to the preliminary candidate product set and the preliminary candidate user set to obtain an optimized candidate user set, and each product in the optimized candidate user set and its corresponding allocated user information are formatted through a preset table conversion method (for example, two-dimensional table or JSON format) to generate a resource allocation table. It should be understood that the process of matching products with receiving users is similar to a bipartite graph matching problem. Among them, one set of nodes represents users, and the other set of nodes represents products. Through the Hungarian algorithm, the optimal matching can be found between products and users. The Hungarian algorithm will perform maximum matching based on the matching degree scores of the preliminary candidate user set to ensure the maximum matching degree between each product and user, while satisfying the finiteness of resources (i.e., the product quantity limit). The objective function of the Hungarian algorithm can be expressed as the following formula: where, is the user and the product the matching degree score between them. is a binary variable indicating whether the user is allocated to the product The sum represents a summation operation, which means summing up the matching scores for all users and products.

[0073] In an optional implementation, to provide more accurate resource matching in the future social group gift-giving process, thereby improving user satisfaction with resource matching. By monitoring the behavior of the recipient user after receiving the gift (such as operations like receiving, forwarding, returning, or exchanging), collecting the satisfaction scores of the users for the gift, and updating the user profile of the recipient user based on the feedback of the satisfaction scores.

[0074] After the second target device receives the winning product information sent by the server, the recipient user can perform the following subsequent operation instructions on the product: Receiving: The user confirms to accept the product, indicating the user's most satisfied attitude towards the product.

[0075] Forwarding: The user forwards the product to others.

[0076] Returning: The user is not satisfied with the product and requests a return.

[0077] Exchanging: The user has an opinion on the product and hopes to exchange it for another product.

[0078] The server obtains the above operation instructions of the user through the second target device, and converts each operation instruction into a different satisfaction score according to a preset scoring standard. Among them, the receiving operation is regarded as the most satisfied score (for example, 5 points). Thus, a user feedback data set is generated based on the user's operation instructions and satisfaction scores. The user feedback data set includes user identification information, each user's operation instructions and the corresponding satisfaction scores, etc.

[0079] Based on the satisfaction scores in the user's feedback data set, perform exponential moving average processing on the weight of each user's interest label. Among them, the portrait vector of each user contains the user's interest label and its corresponding weight. The size of the weight represents the degree of preference of the user for a certain interest label. The formula for exponential moving average is as follows: Among them, is the feedback score of the user for the product . is the updated label weight. is the smoothing factor, which controls the decay coefficient of the historical weight, and its value range is [0, 1], usually set to 0.7 or 0.8. is the indicator function, indicating the label whether it appears in the product . $n$ represents the number of product categories, indicating how many products the user has received. Through the exponential moving average method, the weights of the user interest tags can be smoothed, making the impact of recent feedback on the user profile greater, so as to better reflect the user's current interests and needs.

[0080] After obtaining the updated weight of each tag through exponential moving average processing, the corresponding tag weights in the user interest tag system are updated according to the updated weights. Then, the updated weights of each tag are concatenated, and these weights are combined into the user interest tag system. The concatenated tag system will be used as the new user profile vector. Through the concatenation process, the optimized user profile vector provides more accurate user needs and interest characteristics of the user for the subsequent social group gift-giving process.

[0081] In an alternative embodiment, the gift-giving user can choose to divide the product into several data fragments to be given to multiple users in the social group. First, order data from a first target device is received. The order data contains the product information selected by the gift-giving user, which is transmitted in a structured format (e.g., JSON or XML) and includes one or more product description texts. At the same time, the device quantity information of a second target device is obtained from the social group, and the device quantity represents the number of pre-set recipient users. The number of pre-set recipient users is specified by the first target device in the order data. According to the above, it can be known that the product description text and the product are in one-to-one correspondence, and the number of product description texts in the order data is the number of products. In the resource matching process, comparing the device quantity with the text quantity is a key step, and its purpose is to determine whether the order data needs to be further split. Only when the device quantity > text quantity, is it necessary to split the order data into multiple data fragments so that each recipient user can obtain the corresponding order data fragment.

[0082] When the device quantity > text quantity, the order data is split according to a pre-set division method. The order data contains the detailed description texts of one or more products selected by the gift-giving user, and the description texts record the functions, prices, descriptions, and other relevant information of the products. The goal of data splitting is to split these order data into several data fragments, and each data fragment represents part of the order data for subsequent random distribution of the data fragments to multiple recipient users. The pre-set division method can adopt a random splitting algorithm to ensure the randomness and a certain degree of balance of the splitting result, or adopt a fixed ratio division algorithm to make the data volume of each fragment basically the same after splitting.

[0083] After the order data is split, the order sub-data fragments are allocated to the second target device according to a preset allocation method. The allocation method can be set to random allocation or polling allocation to ensure that all participating winning users have the opportunity to obtain order data fragments. The allocation method can be random allocation, polling allocation, etc. After the allocation is completed, the preset data fragment allocation table is updated according to the result of this allocation. The data fragment allocation table is a preset classification table used to record each second target device and its corresponding allocated data fragment number.

[0084] Meanwhile, classify the order sub-data fragments obtained by each winning user (i.e., the second target device) in the data fragment allocation table, and determine whether these fragments can be pieced together into complete order data. First, classify the order sub-data fragments according to the second target device recorded in the data fragment allocation table. Merge the order data fragments of the same commodity into a set. Thus, perform integrity detection on the data fragment set of each second target device in the data fragment allocation table to determine whether the order data fragments of the same commodity are sufficient to be pieced together into complete order data. The detection method can adopt the statistical method or the data consistency detection method. For example, calculate the coverage rate or similarity score of the spliced data and compare the result with a preset threshold. When the order data fragments of the same commodity obtained by the second target device meet or exceed the threshold, it is considered that the order data of this commodity obtained by the second target device is complete.

[0085] When complete order data is detected, for each second target device that meets the data integrity condition, splice the order sub-data in its data fragment set to extract the complete target order data. This splicing process must be carried out according to the preset data splicing rules to ensure that all order data fragments can be merged into a complete data record in order. Finally, summarize all second target devices that meet the data integrity condition and their target order data, and thus generate the final resource allocation table according to the preset table conversion format.

[0086] This application is applied to the field of human-computer interaction technology, and provides a social e-commerce resource matching method, device, electronic device, and storage medium. By performing multi-source heterogeneous data preprocessing on the original user behavior data set to obtain a standardized spatio-temporal behavior sequence, extracting tags from the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and performing dynamic decay weight assignment and semantic enhancement processing on the user interest tag system to obtain a spatio-temporal weighted user portrait vector. At the same time, perform scene knowledge graph embedding and functional entity extraction processing on the product description text to obtain a product scene feature matrix, calculate the multi-objective matching degree according to the spatio-temporal weighted user portrait vector and the product scene feature matrix to obtain a user-product matching probability tensor, and perform conflict resolution and global optimal allocation processing on the user-product matching probability tensor to obtain a resource allocation table. This application improves the resource matching accuracy of the social e-commerce platform through data processing and machine learning.

[0087] As Figure 2 shown, it is a functional module diagram of a social e-commerce resource matching device provided by an embodiment of this application.

[0088] In some embodiments, the social e-commerce resource matching device 2 may include multiple functional modules composed of computer program segments. The computer programs of each program segment in the social e-commerce resource matching device 2 can be stored in the memory of the server and executed by at least one processor to execute (see details in Figure 1 the description) the functions of the social e-commerce resource matching method.

[0089] In this embodiment, the social e-commerce resource matching device 2 can be divided into multiple functional modules according to the functions it executes. The functional modules may include: a data collection module 21, a spatio-temporal standard module 22, a portrait vector module 23, a feature matrix module 24, a probability tensor module 25, a resource allocation module 26, and a data fragmentation module 27. The module referred to in the present invention means a series of computer program segments that can be executed by at least one processor and can complete fixed functions, and are stored in the memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0090] The data collection module 21 is used to perform target data analysis on the first target device and the second target device according to a preset data collection method to obtain a product description text and an original user behavior data set.

[0091] The spatio-temporal standard module 22 is used to perform multi-source heterogeneous data preprocessing on the original user behavior data set to obtain a standardized spatio-temporal behavior sequence.

[0092] In an optional implementation manner, the spatio-temporal standard module 22 is specifically used for: Perform UTC unified conversion on the timestamps in the original user behavior dataset to obtain spatio-temporally aligned data; Perform Geohash encoding processing on the geographical location information in the original user behavior dataset to obtain a spatio-temporal joint index; Filter the abnormal behavior data from the original user behavior dataset according to the spatio-temporal joint index and the spatio-temporally aligned data to obtain an abnormal behavior filtered dataset; Perform Z-score standardization processing on the numerical fields in the abnormal behavior filtered dataset to obtain standardized numerical data; Perform One-Hot encoding processing on the discrete fields in the abnormal behavior filtered dataset to obtain sparse vectors; Concatenate and fuse the standardized numerical data and the sparse vectors according to a preset concatenation method to obtain the standardized spatio-temporal behavior sequence.

[0093] The portrait vector module 23 is used to extract user interest tags from the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and perform dynamic decay weight assignment and semantic enhancement processing on the user interest tag system to obtain a spatio-temporally weighted user portrait vector.

[0094] In an alternative embodiment, the portrait vector module 23 is specifically configured to: Perform semantic parsing processing on the target words in the standardized spatio-temporal behavior sequence through a preset BERT-Base model to construct an initial tag system; Perform dynamic decay processing on the initial tag system according to the time difference between the timestamp in the standardized spatio-temporal behavior sequence and the current time to obtain a decay-optimized tag system; Determine the time decay factor of each tag in the decay-optimized tag system according to the time difference, and perform weighted processing on the decay-optimized tag system according to the time decay factor to obtain a time-weighted tag system; Perform geographical clustering processing on each tag in the time-weighted tag system according to the spatio-temporal joint index to obtain the regional tag corresponding to each tag in the time-weighted tag system, and merge the regional tag with the time-weighted tag system to generate the user interest tag system; Perform weight concatenation processing on each tag in the user interest tag system in a preset dimension to obtain the spatio-temporally weighted user portrait vector.

[0095] The feature matrix module 24 is used to perform scene knowledge graph embedding and functional entity extraction processing on the product description text to obtain a product scene feature matrix.

[0096] In an optional implementation manner, the feature matrix module 24 is specifically configured to: Perform word segmentation on the product description text to extract function words and scenario keywords; Perform named entity recognition processing on the function words through a preset BiLSTM-CRF model to obtain a function entity list; Perform matching processing on the function entity list and the scenario keywords according to a preset domain dictionary to obtain a product scenario knowledge graph; Perform graph embedding processing on the product scenario knowledge graph to obtain entity vectors of each scenario entity; Calculate the association degree between the scenario entities and the scenario keywords in the product scenario knowledge graph to obtain the relationship strength of each scenario entity; Perform weighted processing on the entity vectors according to the relationship strength to obtain scenario entity vectors; Perform matrix construction on the scenario entity vectors to obtain the product scenario feature matrix.

[0097] The probability tensor module 25 is configured to perform multi-object matching degree calculation according to the spatio-temporal weighted user profile vector and the product scenario feature matrix to obtain a user-product matching probability tensor.

[0098] In an optional implementation manner, the probability tensor module 25 is specifically configured to: Perform matrix multiplication on the spatio-temporal weighted user profile vector and the product scenario feature matrix to obtain a matching degree score between the user profile and the product scenario; Calculate the cosine similarity between each user and each product scenario according to the matching degree score, the weight of each label in the spatio-temporal weighted user profile vector, and the scenario entity vector in the product scenario feature matrix to obtain an initial matching degree matrix; Perform normalization processing on each matching degree score in the initial matching degree matrix according to the quantities of each product recorded in the product description text to obtain a normalized matching degree score; Determine the priority of each user according to the historical behavior data in the spatio-temporal weighted user profile vector, and perform weighted processing on the normalized matching degree score according to the priority to obtain an optimized matching degree matrix; Perform Softmax probability mapping processing on the optimized matching degree matrix to obtain an initial user-product matching probability tensor; Perform normalization processing on the initial user-product matching probability tensor to obtain the user-product matching probability tensor.

[0099] The resource allocation module 26 is configured to perform conflict resolution and global optimal allocation processing on the user-item matching probability tensor to obtain a resource allocation table.

[0100] In an alternative embodiment, the resource allocation module 26 is specifically configured to: Sort the matching probabilities between each user and each item in the user-item matching probability tensor to determine a preliminary candidate item set for each user; Sort the candidate matching users for each item according to the preliminary candidate item set and the number of items to generate a preliminary candidate user set for each item; Determine a preset table conversion method according to the preliminary candidate user set, and perform format conversion on the preliminary candidate user set according to the table conversion method to obtain the resource allocation table.

[0101] In an alternative embodiment, the social e-commerce resource matching device 2 further includes a data fragmentation module 27, and the data fragmentation module 27 is specifically configured to: When obtaining the order data sent by the first target device, compare the number of devices of the second target device with the number of texts of the item description text; When the number of devices is greater than the number of texts, perform data partitioning on the order data according to the number of devices and a preset partitioning method to obtain order sub-data fragments; Update a preset data fragmentation allocation table according to a preset allocation method, the order sub-data fragments, and the second target device; Perform order sub-data fragment classification and data integrity judgment on the order sub-data fragments corresponding to each second target device in the data fragmentation allocation table according to a preset order data integrity monitoring method; When the order sub-data fragments pass the data integrity judgment, perform target data extraction processing on the data fragmentation allocation table according to the order sub-data fragments to obtain the resource allocation table.

[0102] It should be understood that the various change modes and specific embodiments in the methods provided in the above embodiments are equally applicable to the social e-commerce resource matching device in this embodiment. Through the foregoing detailed description of the social e-commerce resource matching method, those skilled in the art can clearly know the implementation method of the social e-commerce resource matching device in this embodiment. For the sake of brevity of the specification, it will not be described in detail here.

[0103] As Figure 3 shown, it is a schematic structural diagram of an electronic device provided by an embodiment of the present application.

[0104] In a preferred embodiment of the present invention, the electronic device 3 may include, but is not limited to: a memory 31, at least one processor 32, and at least one communication bus 33.

[0105] Those skilled in the art should understand that Figure 3 the structure of the illustrated electronic device 3 does not constitute a limitation on the embodiments of the present invention. The electronic device 3 may also include more or fewer other hardware or software than shown, or different component arrangements.

[0106] In some embodiments, the electronic device 3 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits, programmable gate arrays, digital signal processors, and embedded devices, etc.

[0107] It should be noted that the electronic device 3 is only an example. Other existing or future possible electronic products that can be adapted to this application should also be included within the protection scope of this application and are hereby incorporated by reference.

[0108] In some embodiments, a computer program is stored in the memory 31. When the computer program is executed by the at least one processor 32, all or part of the steps in the social e-commerce resource matching method as described above are implemented. The memory 31 includes a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc memories, magnetic disk memories, tape memories, or any other computer-readable medium capable of carrying or storing data. Further, the computer-readable storage medium mainly includes a storage program area and a storage data area. Among them, the storage program area may store an operating system, application programs required for at least one function, etc.

[0109] In some embodiments, the at least one processor 32 is the control core (Control Unit) of the electronic device 3, connecting various components of the entire electronic device 3 through various interfaces and circuits. By running or executing programs or modules stored in the memory 31, and by invoking data stored in the memory 31, it performs various functions of the electronic device 3 and processes data. For example, when the at least one processor 32 executes the computer program stored in the memory 31, it implements all or part of the steps of the social e-commerce resource matching method described in the embodiments of the present application; or implements all or part of the functions of the social e-commerce resource matching device. The at least one processor 32 may be composed of integrated circuits. For example, it may be composed of a single packaged integrated circuit, or may be composed of multiple integrated circuits with the same or different functions packaged, including a combination of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips, etc.

[0110] In some embodiments, the at least one communication bus 33 is configured to enable connection communication between the memory 31 and the at least one processor 32, etc. Although not shown, the electronic device 3 may further include a power source (such as a battery) for powering each component. Preferably, the power source can be logically connected to the at least one processor 32 through a power management device, so as to implement functions such as management of charging, discharging, and power consumption management through the power management device. The power source may also include any components such as one or more DC or AC power sources, a recharge device, a power failure detection circuit, a power converter or inverter, and a power status indicator. The electronic device 3 may also include various sensors, a Bluetooth module, a Wi-Fi module, etc., which will not be elaborated here.

[0111] The above-mentioned integrated unit implemented in the form of software function modules can be stored in a computer-readable storage medium. The above-mentioned software function modules are stored in a storage medium, including several instructions for causing an electronic device (which may be a personal computer, an electronic device, or a network device, etc.) or a processor to execute part of the methods described in the various embodiments of the present application.

[0112] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0113] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical unit, and it may be located in one place or distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0114] The above are all preferred embodiments of this application, and the protection scope of this application is not limited thereby. Therefore, all equivalent changes made according to the structure, shape, and principle of this application shall be covered within the protection scope of this application.

Claims

1. A social e-commerce resource matching method, characterized in that, The method includes: Performing target data analysis on a first target device and a second target device according to a preset data collection method to obtain a product description text and an original user behavior dataset; Performing multi-source heterogeneous data preprocessing on the original user behavior dataset to obtain a standardized spatio-temporal behavior sequence; Performing user interest label extraction on the standardized spatio-temporal behavior sequence to obtain a user interest label system, and performing dynamic decay weight assignment and semantic enhancement processing on the user interest label system to obtain a spatio-temporal weighted user portrait vector; Performing scene knowledge graph embedding and functional entity extraction processing on the product description text to obtain a product scene feature matrix; Calculating a multi-object matching degree according to the spatio-temporal weighted user portrait vector and the product scene feature matrix to obtain a user-product matching probability tensor; Performing conflict resolution and global optimal allocation processing on the user-product matching probability tensor to obtain a resource allocation table.

2. The social e-commerce resource matching method according to claim 1, wherein The performing multi-source heterogeneous data preprocessing on the original user behavior dataset to obtain a standardized spatio-temporal behavior sequence includes: Performing UTC unified conversion on the timestamps in the original user behavior dataset to obtain spatio-temporal aligned data; Performing Geohash coding processing on the geographical location information in the original user behavior dataset to obtain a spatio-temporal joint index; Filtering abnormal behavior data from the original user behavior dataset according to the spatio-temporal joint index and the spatio-temporal aligned data to obtain an abnormal behavior filtered dataset; Performing Z-score standardization processing on the numerical fields in the abnormal behavior filtered dataset to obtain standardized numerical data; Performing One-Hot coding processing on the discrete fields in the abnormal behavior filtered dataset to obtain sparse vectors; Performing splicing and fusion on the standardized numerical data and the sparse vectors according to a preset splicing method to obtain the standardized spatio-temporal behavior sequence.

3. The social e-commerce resource matching method according to claim 2, wherein The performing user interest label extraction on the standardized spatio-temporal behavior sequence to obtain a user interest label system, and performing dynamic decay weight assignment and semantic enhancement processing on the user interest label system to obtain a spatio-temporal weighted user portrait vector includes: Performing semantic parsing processing on the target words in the standardized spatio-temporal behavior sequence through a preset BERT-Base model to construct an initial label system; Performing dynamic decay processing on the initial label system according to the time difference between the timestamp in the standardized spatio-temporal behavior sequence and the current time to obtain a decay optimized label system; Determining the time decay factor of each label in the decay optimized label system according to the time difference, and performing weighted processing on the decay optimized label system according to the time decay factor to obtain a time weighted label system; Performing geographical clustering processing on each label in the time weighted label system according to the spatio-temporal joint index to obtain the regional label corresponding to each label in the time weighted label system, and merging the regional label with the time weighted label system to generate the user interest label system; Perform weighted splicing processing on each tag in the user interest tag system in a preset dimension to obtain the spatio-temporal weighted user portrait vector.

4. The social e-commerce resource matching method according to claim 1, wherein The processing of performing scene knowledge graph embedding and functional entity extraction on the commodity description text to obtain the commodity scene feature matrix includes: Perform word segmentation processing on the commodity description text to extract functional words and scene keywords; Perform named entity recognition processing on the functional words through a preset BiLSTM-CRF model to obtain a list of functional entities; Perform matching processing on the list of functional entities and the scene keywords according to a preset domain dictionary to obtain a commodity scene knowledge graph; Perform graph embedding processing on the commodity scene knowledge graph to obtain entity vectors of each scene entity; Calculate the correlation degree between the scene entities and the scene keywords in the commodity scene knowledge graph to obtain the relationship strength of each scene entity; Perform weighted processing on the entity vectors according to the relationship strength to obtain scene entity vectors; Perform matrix construction on the scene entity vectors to obtain the commodity scene feature matrix.

5. The social e-commerce resource matching method according to claim 1, wherein The calculation of the multi-object matching degree according to the spatio-temporal weighted user portrait vector and the commodity scene feature matrix to obtain the user-commodity matching probability tensor includes: Perform matrix multiplication processing on the spatio-temporal weighted user portrait vector and the commodity scene feature matrix to obtain the matching degree score between the user portrait and the commodity scene; Calculate the cosine similarity between each user and each commodity scene according to the matching degree score, the weight of each tag in the spatio-temporal weighted user portrait vector, and the scene entity vector in the commodity scene feature matrix to obtain an initial matching degree matrix; Perform normalization processing on each matching degree score in the initial matching degree matrix according to the quantity of each commodity recorded in the commodity description text to obtain a normalized matching degree score; Determine the priority of each user according to the historical behavior data in the spatio-temporal weighted user portrait vector, and perform weighted processing on the normalized matching degree score according to the priority to obtain an optimized matching degree matrix; Perform Softmax probability mapping processing on the optimized matching degree matrix to obtain an initial user-commodity matching probability tensor; Perform normalization processing on the initial user-commodity matching probability tensor to obtain the user-commodity matching probability tensor.

6. The social e-commerce resource matching method according to claim 5, wherein The conflict resolution and global optimal allocation processing of the user-commodity matching probability tensor to obtain the resource allocation table includes: Sort the matching probabilities between each user and each commodity in the user-commodity matching probability tensor to determine the preliminary candidate commodity set of each user; Sort the matching candidate users of each commodity according to the preliminary candidate commodity set and the commodity quantity to generate the preliminary candidate user set of each commodity; Determine a preset table conversion method according to the preliminary candidate user set, and perform format conversion on the preliminary candidate user set according to the table conversion method to obtain the resource allocation table.

7. The social e-commerce resource matching method according to claim 1, wherein The method further includes: When obtaining the order data sent by the first target device, compare the number of devices of the second target device with the number of text in the commodity description text; When the number of devices is greater than the number of text, divide the order data according to the number of devices and a preset division method to obtain order sub-data fragments; Update a preset data fragment allocation table according to a preset allocation method, the order sub-data fragments, and the second target device; According to a preset order data integrity monitoring method, classify the order sub-data fragments corresponding to each second target device in the data fragment allocation table and judge data integrity; When the order sub-data fragments pass the data integrity judgment, perform target data extraction processing on the data fragment allocation table according to the order sub-data fragments to obtain the resource allocation table.

8. A social e-commerce resource matching device, characterized in that, The device includes: A data acquisition module, configured to perform target data analysis on the first target device and the second target device according to a preset data acquisition method to obtain a commodity description text and an original user behavior data set; A spatio-temporal standard module, configured to perform multi-source heterogeneous data preprocessing on the original user behavior data set to obtain a standardized spatio-temporal behavior sequence; A portrait vector module, configured to extract user interest tags from the standardized spatio-temporal behavior sequence to obtain a user interest tag system, and perform dynamic decay weight allocation and semantic enhancement processing on the user interest tag system to obtain a spatio-temporal weighted user portrait vector; A feature matrix module, configured to perform scene knowledge graph embedding and functional entity extraction processing on the commodity description text to obtain a commodity scene feature matrix; A probability tensor module, configured to calculate a multi-objective matching degree according to the spatio-temporal weighted user portrait vector and the commodity scene feature matrix to obtain a user-commodity matching probability tensor; A resource allocation module, configured to perform conflict resolution and global optimal allocation processing on the user-commodity matching probability tensor to obtain a resource allocation table.

9. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the social e-commerce resource matching method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the social e-commerce resource matching method according to any one of claims 1 to 7 are implemented.

Images