Household service platform matching method and system

Abstract

The invention provides a household service platform matching method and system, and relates to the technical field of household service recommendation, and the method comprises the steps: obtaining a real-time demand text of a target user, extracting a keyword, and generating a demand vector, obtaining skill data, historical service data and user evaluation of a server at the same time, and generating a capability vector based on the skill data, the historical service data and the user evaluation; inputting the two into a pre-training matching model to obtain a matching degree score, and recommending a server according to the matching degree score; and after the order is completed, adjusting parameters of the matching model by using an incremental learning algorithm according to evaluation data of the target user to the server. The reliability of the recommendation result can be improved.

Description

Technical Field

[0001] This application relates to the technical field of domestic service recommendation, and in particular to a matching method and system for domestic service platforms. Background Technology

[0002] With the rapid development of the social economy and the ever-accelerating pace of life, the domestic service industry has ushered in unprecedented development opportunities. Modern families have increasingly diversified demands for domestic services, covering multiple areas such as cleaning, nannies, maternity nurses, nursing care, and appliance repair. At the same time, consumers are placing higher demands on the quality and professionalism of domestic services, expecting not only service personnel to possess the necessary skills but also services that precisely meet their individual needs.

[0003] Driven by internet technology, domestic service platforms have emerged, breaking down the information asymmetry inherent in traditional domestic services and building a convenient communication bridge between service providers and consumers. Through online platforms, consumers can quickly post their domestic service needs, and service providers can promptly access job opportunities, greatly improving the efficiency of supply and demand matching in the domestic service sector. However, with the continuous expansion of the domestic service market and the increasing complexity of user needs, how to achieve more accurate and efficient matching on domestic service platforms has become a key issue facing the industry.

[0004] The relevant technologies utilize machine learning or deep learning algorithms to predict users' preferences for service providers based on their historical behavior and the behavior of similar users, thereby enabling personalized recommendations.

[0005] However, most machine learning or deep learning models in related technologies are trained on static data. Once the model is trained, it is difficult to adjust it according to real-time user feedback and dynamic changes in service providers, thus affecting the accuracy and reliability of the recommendation results. Summary of the Invention

[0006] To improve the accuracy and reliability of recommendation results, this application provides a matching method and system for a domestic service platform.

[0007] Firstly, this application provides a matching method for a domestic service platform, employing the following technical solution: A matching method for domestic service platforms includes the following steps: Obtain the real-time demand text of the target user, extract the keywords from the real-time demand text, generate a demand vector based on the keywords in the real-time demand text, obtain the service provider's skill data, historical service data and user evaluations, and generate a capability vector based on the skill data, historical service data and user evaluations; Input the demand vector and capability vector into the pre-trained matching model to obtain the matching score, and recommend service providers to the target user based on the matching score; After an order is completed, an incremental learning algorithm is used to adjust the parameters of the matching model using the target users' evaluation data of the service provider.

[0008] This application improves the accuracy of matching results by acquiring real-time demand text from target users, which contains key information such as the service content expected by the target users. This text extracts keywords and converts them into vector form, making it more likely that target users will obtain services that meet their current urgent needs. Furthermore, this application generates capability vectors by acquiring service provider data from multiple dimensions, enabling the generated capability vectors to more accurately represent the service provider's competitiveness in the market.

[0009] Subsequently, this application inputs the demand vector and capability vector into a pre-trained matching model to calculate a matching score. The pre-trained matching model can capture the complex underlying relationships between the demand vector and capability vector, thus providing a relatively accurate matching score. Recommending service providers to target users based on these scores can improve the accuracy and effectiveness of recommendations to a certain extent, making it easier for users to find suitable service providers, while also increasing the chances of service providers obtaining orders.

[0010] After an order is completed, this application employs an incremental learning algorithm to adjust the parameters of the matching model using target users' evaluation data of service providers. This algorithm allows for local adjustments and updates to the matching model using user evaluation data without retraining the entire model. This enables the matching model to adapt to changes in the market and user needs. As more orders are completed and user evaluation data accumulates, the model can be continuously optimized. By continuously optimizing the matching model, it can provide target users with service provider recommendations that better meet their needs, thereby improving the overall user experience. Users are more likely to obtain satisfactory services, thus increasing their trust and loyalty to the platform. Simultaneously, for service providers, accurate recommendations enable them to obtain more suitable orders, improving job satisfaction and income levels. This also helps incentivize service providers to continuously improve their service quality, fostering a healthy market ecosystem.

[0011] Optionally, the matching model adopts a dual-tower model, which includes a user demand tower, a service provider capability tower, and a similarity calculation layer. The user demand tower performs deep semantic encoding on the demand vector through a multi-layer Transformer structure and outputs demand features. The service provider capability tower performs dynamic weighted fusion processing on the capability vector through an attention layer and outputs capability features. The similarity calculation layer calculates the cosine similarity between the demand features and the capability features, and converts the cosine similarity into a matching score through a sigmoid activation function.

[0012] The user demand pyramid employs a multi-layered Transformer structure to perform deep semantic encoding on demand vectors. The Transformer structure has powerful sequence modeling capabilities, enabling it to capture long-distance dependencies and complex semantic information between words in the demand text. Through multi-layered processing, it can delve deeper into the underlying intentions and semantic levels behind user demands, outputting more accurate and comprehensive demand features.

[0013] The service provider capability pyramid uses an attention layer to dynamically weight and fuse capability vectors. This attention mechanism automatically assigns different weights based on the importance of different capability dimensions in the matching process. Service provider capabilities encompass multiple aspects, such as professional skills, work experience, and communication skills. The attention layer highlights capability dimensions that are more critical to user needs in the current matching scenario, while weakening relatively less important dimensions. This ensures that the output capability features better reflect the service provider's comprehensive advantages in the actual matching context, improving the alignment between the service provider's actual capabilities and user needs.

[0014] The similarity calculation layer first calculates the cosine similarity between demand features and ability features. Cosine similarity measures the degree of similarity between two vectors in a direction, intuitively reflecting the close relationship between user demand features and service provider ability features. Then, the cosine similarity is converted into a matching score through the sigmoid activation function. The sigmoid function maps the output value between 0 and 1, making the matching score more intuitive to interpret and facilitating service provider recommendations based on the score, thus improving the operability and practicality of the matching results.

[0015] The dual-tower model separates user demand processing and service provider capability processing into two independent towers. This allows for simultaneous feature extraction of multiple user demands and multiple service provider capabilities when calculating large-scale user-service provider matching, improving computational efficiency and enabling rapid generation of matching scores. This meets the needs of real-time recommendations, reduces user waiting time, and enhances user experience.

[0016] Optionally, the matching model includes a gradient inversion layer, and the method further includes: Obtain the target user's historical demand text, extract keywords from the historical demand text, compare the keywords in the real-time demand text with those in the historical demand text, and calculate the semantic difference of the conflicting demand pairs through a preset semantic similarity model. If the semantic difference exceeds the preset difference threshold, the keyword pair is marked as a conflicting demand pair, the historical demand text in the conflicting demand pair is recorded as the target demand text, the keywords of the target demand text are used to generate a historical vector, adversarial features are generated based on the historical vector, the gradient inversion layer uses the adversarial features to perform adversarial training, and the matching model after adversarial training is used to replace the pre-trained matching model to calculate the matching score.

[0017] This application obtains historical and real-time request texts from target users, extracts and compares keywords, and uses a pre-defined semantic similarity model to identify conflicting request pairs. This allows for precise identification of significantly changed parts of user requests, facilitating a deeper understanding of the user's current true intent and minimizing matching biases caused by relying solely on real-time or historical requests, thereby improving the accuracy of matching results. Furthermore, by calculating the semantic difference between conflicting request pairs and comparing it to a pre-defined difference threshold, this application comprehensively considers changes in user needs, generating matching scores based on more accurate information and enhancing the alignment between the recommendation service provider and the user's current needs.

[0018] Subsequently, this application generates historical vectors using keywords from the target demand text, thereby generating adversarial features. The gradient inversion layer uses these adversarial features for adversarial training. The introduction of adversarial training enables the matching model to learn more robust feature representations, reducing its over-reliance on specific data distributions. In practical applications, user needs and market environments are constantly changing. Through adversarial training, the matching model can better adapt to these changes, improving its generalization ability in different scenarios, thus providing relatively accurate matching scores under various complex conditions. This application achieves dynamic model updates by using an adversarially trained matching model instead of a pre-trained matching model to calculate matching scores. As user historical data accumulates and real-time demands change, the matching model can adjust its parameters and feature representations in a timely manner, always remaining synchronized with user needs and market conditions.

[0019] Optionally, the matching model further includes an orthogonal neural network layer, and the step of generating adversarial features based on historical vectors includes the following steps: The historical vector is decomposed into core capability components and non-capability-related components through orthogonal neural network layers. Feature-level attention weighting is applied to the core capability components to output adversarial features.

[0020] Orthogonal neural network layers can decompose historical vectors into core competency components and non-competency-related components. Core competency components focus on the truly critical and essential capabilities or needs of service providers or users in their historical behavior, while non-competency-related components contain information that is less correlated with core competencies or may interfere with them. For example, in service provider competency assessment, core competency components can accurately extract core elements such as the service provider's professional skills and key project experience, while separating some incidental and non-critical behaviors or information into non-competency-related components, enabling the matching model to more accurately grasp key features.

[0021] Subsequently, this application applies feature-level attention weighting to the core capability components, further strengthening the role of core features in the matching model. The attention mechanism can automatically assign different weights based on the importance of different core capability features in the matching process. Through attention weighting, the matching model can highlight the features in the core capability components, making the generated adversarial features more reflective of the core advantages and needs of service providers or users in the matching process, thereby improving the accuracy and effectiveness of feature representation.

[0022] The adversarial features generated after orthogonal decomposition and attention weighting have stronger discriminative power. After this processing, the core capabilities of different service providers can more clearly show their respective advantages and differences, and the user's needs can be captured and represented more accurately. This makes it easier for the matching model to distinguish different situations when faced with many service providers and complex needs, recommend the most suitable service provider for each user, and improve the accuracy and personalization of the matching.

[0023] Optionally, the method further includes: Based on the attention weight of the core competency components, the score of each feature in the core competency components is set, the emotional polarity of the core competency components in the evaluation data is analyzed, the score of the corresponding feature is adjusted according to the emotional polarity, and the feature whose adjusted score is lower than the preset value is adjusted to a non-competency-related component.

[0024] This application sets the score of each feature based on the attention weight of the core competency components, which can more accurately reflect the relative importance of different core competency features in matching. By assigning corresponding scores to different skill features through attention weights, this difference in importance can be accurately reflected, enabling the matching model to more accurately consider the role of each feature in the subsequent matching process, thereby improving the accuracy of matching.

[0025] This application analyzes the sentiment polarity of core competency components in the evaluation data and adjusts the scores of corresponding features based on the sentiment polarity. This introduces actual feedback information from users or relevant parties regarding the service provider's core competencies into the model, enabling the matching model to be closer to the actual situation, taking into account users' subjective feelings and evaluations, and further improving the accuracy of matching.

[0026] This application modifies features with scores below a preset value into non-capability-related components. This helps filter out core capability features that perform poorly in user reviews or have little impact on matching. These low-scoring features may be relatively weak capabilities of the service provider or capabilities that are not very relevant in the current matching scenario. By modifying them into non-capability-related components, the model can focus more on core capability features when calculating the matching degree, reducing the influence of interfering information and thus improving the quality of matching results.

[0027] Optionally, when the matching score is lower than a preset threshold, the method further includes: The keywords in the real-time requirement text are compared with the preset capability requirement dimension library to obtain the capability requirement dimension to which each keyword in the real-time requirement text belongs. The priority of each capability requirement dimension is set according to the number of keywords under each capability requirement dimension. Target learning resources are selected from the preset learning resource library according to the priority of each capability requirement dimension, and the target learning resources are sent to the server.

[0028] When the matching score is below a preset threshold, it indicates a gap between the service provider's current capabilities and the user's needs. This application compares keywords in the real-time requirement text with a preset capability requirement dimension library to accurately pinpoint the specific capability dimensions involved in the user's needs. Then, by prioritizing keywords based on the number of keywords under each capability requirement dimension, it clearly identifies which capability dimensions are the most urgent and critical to the user's needs. Based on this priority, this application selects target learning resources from a preset learning resource library, providing service providers with targeted learning content to help them quickly improve their capabilities in that area. After improving their capabilities through learning the target resources, service providers will see a significant increase in the fit between their capabilities and user needs during subsequent matching processes. When encountering similar needs again, the matching score is more likely to reach or exceed the preset threshold, thus increasing the chance of a successful match.

[0029] Optionally, before sending the target learning resources to the server, the method further includes: The specific service score is calculated based on the capability requirement dimension of the historical service data. It is then determined whether the specific service score is greater than the preset score threshold. If it is, the target learning resources are not sent to the service provider; otherwise, the target learning resources are sent to the service provider.

[0030] This application calculates specialized service scores based on the capability requirement dimensions of historical service data. This allows for a deeper understanding of the service provider's past performance across various capability dimensions. If a service provider's specialized service score in a certain dimension is high and exceeds a preset score threshold, it indicates that they possess strong capabilities in that area, suggesting that the service provider specializes in that field and is unlikely to work in other areas. In this case, target learning resources will no longer be sent to the service provider, thus minimizing the need to blindly send large amounts of target learning resources to the service provider and reducing costs related to data transmission, storage, and management. Simultaneously, it also minimizes interference with the service provider.

[0031] Optionally, after sending the target learning resources to the server, the method further includes: Service providers upload video data related to the target learning resources. The pre-set action recognition model quantifies the service provider's operational standardization. If the operational standardization meets expectations, the skill description of the capability requirement dimension of the target learning resources is updated in the service provider's skill data.

[0032] This application utilizes a pre-defined action recognition model to quantitatively analyze video data uploaded by service providers regarding target learning resources, enabling an objective and accurate assessment of the service provider's operational standardization. When the service provider's operational standardization meets expectations, the skill description of the capability requirement dimension of the target learning resource is updated in the service provider's skill data, ensuring that the service provider's skill data always reflects their latest capability level. As service providers continuously learn and practice, their skills will gradually improve and expand. By updating the skill data in a timely manner, more comprehensive and accurate information can be provided for subsequent service matching.

[0033] Optionally, after recommending service providers to the target user based on the matching score, the method further includes: the service provider and the target user completing order matching through a two-way confirmation mechanism.

[0034] Secondly, this application provides a domestic service platform matching system, which adopts the following technical solution: A domestic service platform matching system includes: a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in the first aspect.

[0035] In summary, this application includes at least one of the following beneficial technical effects: 1. This application improves the accuracy of matching results by acquiring real-time demand text from target users, which contains key information such as the service content expected by the target users. This text extracts keywords and converts them into vector form, making it more likely that target users will obtain services that meet their current urgent needs. This application also generates capability vectors by acquiring service provider data from multiple dimensions, enabling the generated capability vectors to more accurately represent the service provider's competitiveness in the market.

[0036] 2. This application inputs the demand vector and capability vector into a pre-trained matching model to calculate a matching score. The pre-trained matching model can capture the complex underlying relationships between the demand vector and capability vector, thus providing a relatively accurate matching score. Recommending service providers to target users based on these scores can improve the accuracy and effectiveness of recommendations to a certain extent, making it easier for users to find suitable service providers, while also increasing the chances of service providers obtaining orders.

[0037] 3. After an order is completed, this application employs an incremental learning algorithm to adjust the parameters of the matching model using target users' evaluation data of service providers. This algorithm allows for local adjustments and updates to the matching model based on user feedback data without retraining the entire model. This enables the matching model to adapt to market changes and evolving user needs. As more orders are completed and user feedback data accumulates, the model can be continuously optimized. By continuously optimizing the matching model, it can provide target users with service provider recommendations that better meet their needs, thereby improving the overall user experience. Users are more likely to obtain satisfactory services, thus increasing their trust and loyalty to the platform. Simultaneously, for service providers, accurate recommendations enable them to obtain more suitable orders, improving job satisfaction and income levels. This also helps incentivize service providers to continuously improve their service quality, fostering a healthy market ecosystem. Attached Figure Description

[0038] Figure 1 This is a flowchart of Embodiment 1 of this application; Figure 2 This is a flowchart of Embodiment 2 of this application; Figure 3 This is a flowchart of Embodiment 3 of this application. Detailed Implementation

[0039] The following combination Figures 1 to 3 This application will be described in further detail.

[0040] Example 1: This example discloses a matching method for a domestic service platform, referring to... Figure 1 The method includes: S11 data acquisition and processing, S12 recommendation, and S13 feedback. First, the real-time demand text of the target user is acquired and keywords are extracted to generate a demand vector. At the same time, service provider skill data, historical service data, and user evaluations are acquired, and a capability vector is generated based on the skill data, historical service data, and user evaluations. Then, the two are input into a pre-trained matching model to obtain a matching score, and service providers are recommended accordingly. After the order is completed, the parameters of the matching model are adjusted using the target user's evaluation data of the service provider and an incremental learning algorithm. The execution process of each step in this embodiment is described in detail below.

[0041] S11 data acquisition and processing acquires the real-time demand text of the target user and uses natural language processing technology to extract keywords from the real-time demand text, including using the bag-of-words model to extract keywords from the real-time demand text and using the TF-IDF algorithm to extract keywords from the real-time demand text.

[0042] The bag-of-words model treats real-time demand text as a set of words, with each word serving as a feature. The text is represented by counting the number of times each word appears in the real-time demand text.

[0043] The TF-IDF algorithm takes into account the importance of words in a document and determines keywords by calculating term frequency (TF) and inverse document frequency (IDF).

[0044] Keywords are mapped to a vector space to generate a demand vector. In this embodiment, one-hot encoding is used, representing each keyword as a vector with the same dimension as the vocabulary size, with the keyword position set to 1 and the rest to 0. Subsequently, word embedding techniques such as Word2Vec and GloVe are used to map the keywords to a low-dimensional vector space, where each vector captures the semantic information of the word. Finally, the demand vector is obtained by performing weighted summation or other aggregation operations on the keyword vectors.

[0045] For example, a real-time demand text for a housekeeping service provider who can clean a three-bedroom apartment and cook a simple dinner can be extracted using natural language processing technology. Keywords extracted include: cleaning, three-bedroom apartment, cooking, and dinner.

[0046] The pre-built vocabulary includes 10 common household-related terms such as cleaning, three-bedroom apartment, cooking, dinner, window cleaning, and childcare. For the keywords cleaning, three-bedroom apartment, cooking, and dinner, the corresponding one-hot vectors are set to 1 at the appropriate positions and 0 elsewhere, resulting in four high-dimensional vectors, as shown below: The one-hot vector corresponding to cleaning is: [1, 0, 0, 0, 0, 0, 0, 0, 0, 0] The one-hot vector corresponding to the three-bedroom apartment is: [0, 1, 0, 0, 0, 0, 0, 0, 0, 0] The one-hot vector corresponding to cooking is: [0, 0, 1, 0, 0, 0, 0, 0, 0, 0] The one-hot vector corresponding to dinner is: [0, 0, 0, 1, 0, 0, 0, 0, 0, 0] The four one-hot vectors are mapped to a low-dimensional space using a pre-trained Word2Vec model, as follows: The cleaned-up vector is: [0.2, 0.1, 0.3] The vector mapped from the three-bedroom apartment is: [0.1, 0.4, 0.2]. The vector after cooking mapping is: [0.3, 0.2, 0.1] The vector mapped to dinner is: [0.2, 0.3, 0.2] The weighted sum is calculated based on the importance of the keywords. In this embodiment, the weight of each of the four keywords is 0.25, and the final demand vector is [0.2, 0.25, 0.2].

[0047] Obtain service provider skill data, historical service data, and user reviews; generate capability vectors based on skill data, historical service data, and user reviews, including the following steps: For skills data, the same method used to generate the demand vector is employed to convert it into a vector representation. Historical service data is quantified using metrics such as service frequency, service duration, and service satisfaction, and these metrics are then combined into a vector. For user reviews, NLP techniques are used to perform sentiment analysis and feature extraction on the text, converting it into numerical features, which are then added to the capability vector. Finally, by fusing these data vectors from different sources, the service provider's capability vector is obtained.

[0048] Taking a service provider's skill description as being proficient in cleaning various types of houses and able to cook home-style dishes as an example, the process of obtaining the service provider's ability vector is as follows: Natural language processing (NLP) techniques were used to extract keywords. From a pre-constructed vocabulary (containing 10 common household-related terms such as cleaning, three-bedroom apartment, cooking, dinner, window cleaning, and childcare), the keywords "cleaning" and "home-style cooking" were selected. Since "home-style cooking" is semantically similar to "cooking" in the vocabulary, it was chosen to represent "home-style cooking." The one-hot vectors corresponding to these two keywords are as follows: The one-hot vector corresponding to cleaning is: [1, 0, 0, 0, 0, 0, 0, 0, 0, 0] The one-hot vector corresponding to cooking is: [0, 0, 1, 0, 0, 0, 0, 0, 0, 0] Using the same pre-trained Word2Vec model as the demand vector, these two one-hot vectors are mapped to a low-dimensional space, as follows: The cleaned-up vector is: [0.2, 0.1, 0.3] The vector after cooking mapping is: [0.3, 0.2, 0.1] Assuming that the keywords "cleaning" and "cooking" each have a weight of 0.5, the skill data vector is obtained by weighted summation [0.25, 0.15, 0.2].

[0049] Service provider A provided services 50 times in the past year, with an average service duration of 3 hours and a service satisfaction score of 4.5 out of 5. This is integrated into a vector [50, 3, 4.5]. For ease of calculation, this vector is standardized to obtain a historical service data vector [0.5, 0.6, 0.9].

[0050] Collect user reviews of service provider A, such as "the cleaning was very clean" and "the food was delicious".

[0051] Using NLP techniques, sentiment analysis was performed on the review text to determine that the review was positive. The features "clean" and "appetizing" were then extracted and assigned corresponding numerical values. Clauses containing "clean" and "appetizing" were segmented from the review text, and sentiment analysis models (such as SnowNLP, TextBlob, or BERT) were used to calculate the sentiment polarity score of the clauses (0~1, with closer to 1 indicating a more positive sentiment). "Clean" corresponded to a positive sentiment score of 0.8, and "appetizing" corresponded to a positive sentiment score of 0.7.

[0052] The extracted feature values ​​are combined into a user rating vector [0.8, 0.7].

[0053] The skill data vector, historical service data vector, and user evaluation vector are fused using a weighted summation method. Assuming the skill data vector has a weight of 0.4, the historical service data vector has a weight of 0.3, and the user evaluation vector has a weight of 0.3, the resulting capability vector is [0.49, 0.45, 0.35].

[0054] S12 recommendation inputs the demand vector and capability vector into a pre-trained matching model to obtain a matching score, and recommends service providers to the target user based on the matching score.

[0055] The pre-trained matching model can be a machine learning or deep learning-based model, such as logistic regression, support vector machine, neural network, etc. During the training process, the matching model uses a large number of historical demand vectors and capability vectors, as well as their corresponding matching results (such as whether the user selected the service provider), as training data to learn the mapping relationship between demand vectors and capability vectors.

[0056] When new demand vectors and capability vectors are input, the pre-trained matching model calculates their matching score based on the learned mapping relationship. This score represents the degree of matching between the service provider and the target user's needs.

[0057] S13 feedback indicates that after an order is completed, the housekeeping service platform will collect the target user's evaluation data of the service provider, and this evaluation data can be used as new training data.

[0058] Subsequently, the evaluation data, along with the corresponding demand and capability vectors, is input into an incremental learning algorithm. This algorithm adjusts the parameters of the pre-trained matching model based on the evaluation data, enabling the model to better adapt to new data and user needs. For example, if a user rates a service provider highly, it indicates a relatively accurate match between the service provider's capability vector and the user's demand vector. The incremental learning algorithm will then adjust the parameters of the matching model accordingly, ensuring that similar capability and demand vectors achieve higher matching scores in future matches.

[0059] By adopting the above solution, this embodiment can achieve accurate matching from demand to service provider, and continuously optimize the recommendation effect through feedback loop.

[0060] In other embodiments, the method further includes: the service provider and the target user completing order matching through a two-way confirmation mechanism.

[0061] Example 2: Refer to Figure 2 The difference between this embodiment and embodiment 1 is that in this embodiment, the matching model adopts a dual-tower model, which includes a user demand tower, a service provider capability tower, a similarity calculation layer, a gradient inversion layer, and an orthogonal neural network layer.

[0062] S21 calculates the semantic difference degree by obtaining the target user's historical demand text, extracting keywords from the historical demand text, comparing the keywords in the real-time demand text with those in the historical demand text, and calculating the semantic difference degree of the conflicting demand pair using a preset semantic similarity model (such as Sentence-BERT). The calculation model for the semantic difference degree 'a' is as follows: ; Where 'a' represents the semantic difference between keyword A in the real-time requirement text and keyword B in the historical requirement text. It is a model specifically optimized for sentence-level semantic similarity calculation.

[0063] S22 Semantic difference determination: Determine whether the semantic difference exceeds a preset difference threshold. If so, mark the keyword pair as a conflicting demand pair and perform S23 adversarial training.

[0064] If not, then the S12 recommendation is executed, which includes the following steps: The user demand pyramid performs deep semantic encoding on the demand vector through a multi-layer Transformer structure, outputting demand features. The process is as follows: The demand vector is projected into a high-dimensional space through a linear transformation. In this embodiment, the demand vector is projected into a 768-dimensional vector to match the input requirements of the Transformer encoder.

[0065] After passing through a multi-layer Transformer structure, the demand vector is encoded into 768-dimensional demand features. The multi-layer Transformer structure consists of stacked Transformer encoder layers (e.g., 6 layers), each layer containing: Multi-head self-attention is used to calculate the association weights of each dimension in the demand vector and capture long-distance dependencies.

[0066] Feedforward neural networks (FFNs) are used to perform nonlinear transformations on self-attention outputs to enhance their expressive power.

[0067] Residual connections and layer normalization are used to alleviate the vanishing gradient problem and stabilize the training process.

[0068] The service provider capability tower calculates the correlation weights of each dimension in the capability vector through the attention layer, highlighting the capabilities relevant to the current needs and outputting 768-dimensional capability features.

[0069] The similarity calculation layer uses the cosine similarity algorithm to calculate the similarity between demand features and ability features, and converts the similarity into a matching score through the sigmoid activation function.

[0070] In S23 adversarial training, the historical demand text in the conflict demand pair is recorded as the target demand text, and historical vectors are generated using the keywords of the target demand text.

[0071] The historical vector is decomposed into core capability components and non-capability-related components through orthogonal neural network layers. The decomposition model is as follows: ; Where h is the history vector; and Let be the projection matrix, where , ; c is the core capability component, and , Let represent the set of all vectors consisting of k real numbers; n is a non-capability-related component, and , Represents the set of all vectors consisting of dk real numbers; Through orthogonal constraints It can be determined that the core capability component c is independent of the non-capability-related component n.

[0072] The core capability component c is weighted by feature-level attention using a self-attention mechanism to obtain the weighted core capability component c'. The calculation model of the weighted core capability component c' is as follows: ; ; ; in, Let i be the attention weight of the i-th element in the core capability component c. Let be the i-th element in the core capability component c; Q and K are learnable parameters.

[0073] The weighted core capability vector c′ is concatenated with the non-capability-related component n and input into the adversarial generator to output adversarial features.

[0074] During backpropagation, the gradient inversion layer inverts the gradient sign, forcing the matching model to maximize the difficulty of distinguishing conflicting requirement pairs when optimizing the adversarial loss. The calculation model is as follows: ; in, For historical vectors; This is the demand vector.

[0075] Simultaneously optimize the original matching loss (such as cosine similarity loss) and adversarial loss to obtain the matching model after adversarial training. The total loss function of the matching model at this point is... as follows: ; in, These are hyperparameters used to control the intensity of adversarial training. This represents the original matching loss.

[0076] By using a gradient reversal layer, the matching model is forced to learn feature representations that are insensitive to conflict requirements while reducing the matching loss, thereby improving robustness. Then, S12 recommendation is performed. During the execution of S12 recommendation, the matching model trained adversarially is used instead of the pre-trained matching model to calculate the matching score.

[0077] S24 Adjust the components. After the order is completed, set the score of each feature in the core capability component based on the attention weight of the core capability component. In this embodiment, the product of the attention weight and the upper limit of the score is used as the score of the corresponding feature.

[0078] Sentiment scores are calculated using sentiment analysis models (such as SnowNLP or BERT-finetuned). If the sentiment score of a feature is lower than a preset sentiment score threshold, its score is reduced by a preset step size. Features with adjusted scores lower than the preset value are then adjusted to non-ability-related components.

[0079] Example 3: Reference Figure 3 The difference between this embodiment and embodiment 1 is that the preset capability requirement dimension library in this embodiment contains predefined capability dimensions, such as cleaning, cooking, childcare, and elderly care. Each dimension is associated with corresponding words. For example, the capability dimension of elderly care is associated with smart elderly care services, elderly meal assistance services, and accompanying the elderly. The learning resource library stores various learning resources according to capability dimensions.

[0080] When the matching score is lower than a preset threshold, the method further includes: S31 filters resources by comparing the keywords in the real-time requirement text with the preset capability requirement dimension library to obtain the capability requirement dimension to which each keyword in the real-time requirement text belongs.

[0081] Count the number of keywords matched under each capability requirement dimension to establish a correspondence between capability requirement dimension and keyword count. Sort the dimensions in descending order of keyword count, with higher-count capability requirement dimensions having higher priority. When the number of keywords is the same, the two capability requirement dimensions have the same priority.

[0082] The system retrieves a pre-defined learning resource library, where resources are pre-labeled with their corresponding competency requirements. Based on the priority of these competency requirements, the system prioritizes and filters learning resources corresponding to higher-priority competency requirements; these are the target learning resources.

[0083] The S32 resource sending judgment maps historical service data to the corresponding capability requirement dimension, and calculates the special service score based on the capability requirement dimension to which the historical service data belongs. The calculation model of the special service score is as follows: ; in, The score for the specific service in the i-th capability requirement dimension; The number of times the i-th capability requirement dimension is served in the historical service data; The total number of services in the historical service data.

[0084] If the score for a specific service exceeds a preset score threshold, the target learning resource is not sent to the service provider; otherwise, the target learning resource is sent to the service provider.

[0085] S33 skill update: After learning the skill, service providers upload video data related to the target learning resources. The preset action recognition model outputs a quantitative score of operational standardization through frame analysis, action comparison, etc., which may include: Completion rate of key actions (whether all necessary steps are covered); Movement accuracy (the degree of matching with the demonstrated movements in the learning resources); Process compliance (whether the steps are in the correct order, and whether there are any omissions or errors).

[0086] The quantitative score is compared with the preset expected threshold for standardization. If the level of standardization meets the expectation (i.e., the quantitative score is greater than the expected threshold), it is determined that the service provider has mastered the skill, and the skill description of the ability requirement dimension of the target learning resource is updated to the service provider's skill data. If it does not meet the standard, the specific aspects that need improvement are fed back, and the service provider is guided to relearn and submit the video.

[0087] Example 4: This example discloses a domestic service platform matching system, a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes the computer program stored on the computer-readable storage medium, it implements the aforementioned method for matching domestic service platforms.

[0088] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A matching method for a domestic service platform, characterized in that, include: Obtain the real-time demand text of the target user, extract the keywords from the real-time demand text, generate a demand vector based on the keywords in the real-time demand text, obtain the service provider's skill data, historical service data and user evaluations, and generate a capability vector based on the skill data, historical service data and user evaluations; Input the demand vector and capability vector into the pre-trained matching model to obtain the matching score, and recommend service providers to the target user based on the matching score; After an order is completed, an incremental learning algorithm is used to adjust the parameters of the matching model using the target users' evaluation data of the service provider.

2. The matching method for a domestic service platform according to claim 1, characterized in that, The matching model adopts a dual-tower model, which includes a user demand tower, a service provider capability tower, and a similarity calculation layer. The user demand tower performs deep semantic encoding on the demand vector through a multi-layer Transformer structure and outputs demand features. The service provider capability tower performs dynamic weighted fusion processing on the capability vector through an attention layer and outputs capability features. The similarity calculation layer calculates the cosine similarity between the demand features and the capability features, and converts the cosine similarity into a matching score through a sigmoid activation function.

3. The matching method for a domestic service platform according to claim 2, characterized in that, The matching model includes a gradient inversion layer, and the method further includes: Obtain the target user's historical demand text, extract keywords from the historical demand text, compare the keywords in the real-time demand text with those in the historical demand text, and calculate the semantic difference of the conflicting demand pairs through a preset semantic similarity model. If the semantic difference exceeds the preset difference threshold, the keyword pair is marked as a conflicting demand pair, the historical demand text in the conflicting demand pair is recorded as the target demand text, the keywords of the target demand text are used to generate a historical vector, adversarial features are generated based on the historical vector, the gradient inversion layer uses the adversarial features to perform adversarial training, and the matching model after adversarial training is used to replace the pre-trained matching model to calculate the matching score.

4. The matching method for a domestic service platform according to claim 3, characterized in that, The matching model further includes an orthogonal neural network layer, and the generation of adversarial features based on historical vectors includes the following steps: The historical vector is decomposed into core capability components and non-capability-related components through orthogonal neural network layers. Feature-level attention weighting is applied to the core capability components to output adversarial features.

5. The matching method for a domestic service platform according to claim 4, characterized in that, The method further includes: Based on the attention weight of the core competency components, the score of each feature in the core competency components is set, the emotional polarity of the core competency components in the evaluation data is analyzed, the score of the corresponding feature is adjusted according to the emotional polarity, and the feature whose adjusted score is lower than the preset value is adjusted to a non-competency-related component.

6. The matching method for a domestic service platform according to any one of claims 1-5, characterized in that, When the matching score is lower than a preset threshold, the method further includes: The keywords in the real-time requirement text are compared with the preset capability requirement dimension library to obtain the capability requirement dimension to which each keyword in the real-time requirement text belongs. The priority of each capability requirement dimension is set according to the number of keywords under each capability requirement dimension. Target learning resources are selected from the preset learning resource library according to the priority of each capability requirement dimension, and the target learning resources are sent to the server.

7. The matching method for a domestic service platform according to claim 6, characterized in that, Before sending the target learning resources to the server, the method further includes: The specific service score is calculated based on the capability requirement dimension of the historical service data. It is then determined whether the specific service score is greater than the preset score threshold. If it is, the target learning resources are not sent to the service provider; otherwise, the target learning resources are sent to the service provider.

8. The matching method for a domestic service platform according to claim 7, characterized in that, After sending the target learning resources to the server, the method further includes: Service providers upload video data related to the target learning resources. The pre-set action recognition model quantifies the service provider's operational standardization. If the operational standardization meets expectations, the skill description of the capability requirement dimension of the target learning resources is updated in the service provider's skill data.

9. The matching method for a domestic service platform according to claim 1, characterized in that, After recommending service providers to the target user based on the matching score, the method further includes: the service provider and the target user completing the order matching through a two-way confirmation mechanism.

10. A domestic service platform matching system, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in any one of claims 1-9.

Images