A method and system for matching based on user portraits

Abstract

The application discloses a kind of based on user portrait's love match method and system, it is related to user portrait matching technical field, including: the static portrait of the registration user of love platform is collected, dynamic behavior, three kinds of data of value text, and generate standardization preference time series dataset;Bidirectional ESN preference reserve pool embedding preference mutation detection gate is constructed and trained, long-term preference dependence features are extracted by forward reserve pool, short-term preference mutation features are extracted by reverse reserve pool, implicit preference feature vectors are output by fusing and mining target user implicit preference, and initial implicit preference candidate set is generated by screening;Bidirectional preference double-tower deep network embedding asymmetric bidirectional mutual attention layer is constructed, and the bidirectional matching comprehensive score of target user and user in effective candidate set is calculated;According to bidirectional matching comprehensive score, generate matching recommendation list and return to target user, avoid one-way invalid recommendation, greatly improve matching rationality and success rate.

Description

Technical Field

[0001] This invention relates to the field of user profile matching technology, specifically to a matchmaking method and system based on user profiles. Background Technology

[0002] With the rapid popularization of internet technology and the digital transformation of the marriage and dating service industry, online dating platforms have become the core channel for people of marriageable age to complete their dating and matchmaking. The number of registered users on these platforms continues to expand, and users' demands for matchmaking services are constantly upgrading from basic screening to more precise, personalized, and highly compatible services. Traditional offline dating models have inherent limitations such as information asymmetry, low matching efficiency, and limited reach. Online dating platforms, through digital means, break down geographical and social barriers, providing a massive number of users with mate selection information exchange and matching recommendation services. However, at the same time, the growth in the platform's user base has also brought about the problem of information overload. Users need to spend a lot of time sifting through massive amounts of candidate information to find suitable matches. Whether the platform can achieve accurate and efficient two-way matching directly determines the user experience, retention rate, and the final success rate of marriage and dating. This has become a key component of the core competitiveness of dating platforms.

[0003] Current online dating platforms primarily employ rule-based matching technologies based on static user profile tags. These technologies rely on fixed, user-provided tags such as age, region, education, and income to filter candidates and rank them. A few solutions incorporate collaborative filtering algorithms and basic user profiling techniques to optimize matching. However, these approaches suffer from several unavoidable flaws in practical applications: First, inherent biases exist at the feature input level. Dynamic behavioral data fails to distinguish between genuine preferences and noise, allowing users' accidental swiping or curious browsing to directly interfere with preference modeling. Second, user-submitted texts about values ​​and views on marriage often exhibit "persona packaging," with significant discrepancies between the textual descriptions and actual behavioral preferences. Existing technologies lack cross-validation and correction for these two types of data, resulting in insufficient reliability of feature inputs and directly limiting matching accuracy. First, the accuracy of implicit preference temporal modeling is insufficient. Conventional temporal networks can only capture long-term stable preferences and cannot promptly identify sudden changes in users' mate selection preferences, such as preference shifts caused by a single blind date experience or changes in life status. Moreover, the incremental updates of the model mostly adopt a fixed-period full-scale fine-tuning method, which has both update lag and is prone to overall preference drift due to local preference changes, making it impossible to balance update efficiency and model stability. Second, the two-way matching mechanism does not conform to the characteristics of blind date scenarios. Existing two-way matching mostly adopts a symmetrical mutual attention architecture, assuming that the mate selection focus dimensions and weights of the matched parties are completely consistent. However, in blind date scenarios, the mate selection decision logic of different individuals has significant asymmetry. For example, some users take value compatibility as the core decision-making basis, while some users take lifestyle habit matching as the priority condition. Symmetrical attention cannot adapt to this differentiated focus dimension, resulting in the deviation between the two-way fit calculation and the actual decision logic. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a matchmaking method and system based on user profiles. It improves the quality of input features through behavioral credibility calibration and value consistency correction, achieves accurate temporal mining of implicit preferences through bidirectional ESN with mutation detection, adapts differentiated mate selection logic for matchmaking scenarios through asymmetric bidirectional mutual attention, and comprehensively improves matching accuracy and scenario adaptability by combining elastic hard constraints and confidence-weighted incremental closed loops. This solves the technical problems of large feature deviations, coarse preference modeling, and distorted bidirectional matching logic in existing technologies.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a matchmaking method based on user profiles, comprising: Collect three types of data from registered users on dating platforms: static profiles, dynamic behaviors, and value texts. Generate corresponding feature vectors, concatenate the feature vectors according to a preset time window, and generate a standardized preference time series dataset. A bidirectional ESN preference reserve pool with embedded preference mutation detection gating is constructed and trained. After receiving the matchmaking request of the target user, the standardized preference time series data is extracted and input into the trained preference reserve pool. Long-term preference dependency features are extracted through the forward reserve pool, and short-term preference mutation features are extracted through the reverse reserve pool. The implicit preferences of the target user are fused and mined, and the implicit preference feature vector is output. The initial implicit preference candidate set is then generated. Based on the hard constraints preset by the target user, a rigid filter is performed on the initial implicit preference candidate set to generate an effective candidate set. A bidirectional preference dual-tower deep network with an embedded asymmetric bidirectional mutual attention layer is constructed. The preference vectors of the target user and the candidates are mapped to dimensions through their respective independent feature projection matrices. The asymmetric bidirectional attention weights are calculated to generate bidirectional fusion feature vectors. The bidirectional matching comprehensive score between the target user and the users in the effective candidate set is calculated. A matching recommendation list is generated based on the comprehensive score of two-way matching and returned to the target user. Two-way interaction feedback data is collected and the user preference time series data is updated based on the feedback data. The parameters of the two-way preference dual-tower deep network are adjusted by a confidence-weighted incremental learning method.

[0006] Furthermore, differentiated standardization preprocessing was performed on the three types of collected data. Discrete tags in the static profile data were processed using one-hot encoding, and continuous tags in the static profile data were processed using Min-Max normalization to generate static feature vectors. Dynamic behavior data was sorted in ascending order by timestamp, and behavior types were numerically encoded according to fixed rules. Behavior duration and page dwell time were processed using Min-Max normalization to generate initial dynamic behavior feature vectors. Semantic features were extracted from the value text data using a BERT-based semantic coding model pre-trained in Chinese. The output of the semantic coding model's CLS token was used to generate initial value semantic feature vectors.

[0007] Furthermore, for a single dynamic behavior data point, if there is a continuous interaction link after the behavior, the behavior credibility weight is assigned to 1.2; if the behavior is only a single browsing and the dwell time is less than the preset time threshold, it is judged as a novelty noise behavior and the credibility weight is assigned to 0.3; the credibility weight of other behaviors is assigned to 1.0; the credibility weight is multiplied by the corresponding dimension of the initial dynamic behavior feature vector to complete the weighted calibration and generate the final dynamic behavior feature vector; The distribution of candidate tags corresponding to the dynamic behavior of the target user in the past 30 days is statistically analyzed to generate a behavior preference distribution vector. The cosine similarity between the behavior preference distribution vector and the initial value semantic feature vector is calculated. If the similarity is less than 0.5, the value semantic feature vector is linearly interpolated and corrected according to the behavior preference distribution vector with a correction ratio of 0.4 to generate the final value semantic feature vector.

[0008] Furthermore, a single time window is set to a calendar day. The static feature vectors, weighted and calibrated dynamic behavioral feature vectors, and consistency-corrected value semantic feature vectors generated within the same time window are concatenated dimensionally in a fixed order to generate the initial feature vector for a single time window. For the differences in the temporal change frequency of the three types of features (static, dynamic, and value), a dynamic time warping algorithm is used to perform cross-modal temporal alignment. All time window aligned feature vectors are arranged in chronological order to form the preference time-series data for a single platform registered user. For the same-dimensional features of the preference time-series data for all platform registered users, a Z-score normalization method is used to perform global normalization processing, completing the construction of the standardized preference time-series dataset.

[0009] Furthermore, the constructed bidirectional ESN preference reserve pool with embedded preference mutation detection gating includes an input layer, a bidirectional reserve pool layer, a preference mutation detection gating, and an output layer. The bidirectional reserve pool layer simultaneously receives the forward and reverse sequences of standardized preference time-series data. Standardized preference time-series data of registered users on all platforms are extracted and used as training input data for the bidirectional ESN preference reserve pool. The training objective is to memorize the preference association patterns of registered users on different platforms. The output layer weights of the bidirectional ESN preference reserve pool are trained using the ridge regression method. The training termination condition is set: when the recognition accuracy of the bidirectional ESN preference reserve pool for user preference association patterns on the validation set reaches or exceeds the accuracy threshold, the training is stopped. The preference mutation detection gating monitors the feature distribution of newly added time-series data in real time, calculates the KL divergence between the new data features and the mean of historical features, and if the KL divergence exceeds the preset mutation threshold, it triggers a local incremental update of the weights of the reserve pool output layer, only adjusting the weight parameters corresponding to the mutation feature dimension, while the other weights remain fixed.

[0010] Furthermore, the system receives matchmaking requests initiated by target users, extracts standardized preference time-series data of the target user, inputs this data into a bidirectional ESN preference reserve pool, outputs long-term preference features from the forward reserve pool and short-term preference mutation features from the reverse reserve pool, and concatenates and merges the two types of features with a weight of 7:3. The output layer then outputs the latent preference feature vector of the target user. Simultaneously, the system selects latent preference feature vectors of other platform registered users with a similarity ≥ 0.7 to the target user's latent preference feature vector from the bidirectional ESN preference reserve pool to form an initial latent preference candidate set.

[0011] Furthermore, the hard constraints of the target users who initiated the matchmaking request are extracted. According to the fixed rule that each hard constraint corresponds to a dimension, a hard constraint feature vector with a fixed dimension is generated. The dimension that satisfies the corresponding hard constraint is assigned a value of 1, and the dimension that does not satisfy the hard constraint is assigned a value of 0. The hard constraint gating mask layer is called to compare the hard constraint feature vector of the target user with the hard constraint feature vectors of all platform registered users in the initial latent preference candidate set bit by bit, and calculate the sum of the element-wise products of the two vectors. For platform registered users whose product sum is less than the dimension of the hard constraint feature vector, their matching score is directly set to 0 and they are excluded, generating a rigidly filtered candidate subset. If the number of users in the candidate subset after rigid filtering is lower than the preset minimum threshold, the constraints are relaxed in order of preset constraint priority from low to high. The priorities from low to high are geographical scope, age range, income requirements, and marital history requirements. The number of candidate subsets is recalculated for each level of constraint relaxation until the number of candidate subsets reaches the minimum threshold or all non-core constraints have been relaxed, and finally a valid candidate set is generated.

[0012] Furthermore, a bidirectional preference dual-tower deep network is constructed, comprising a user preference encoding tower and a candidate preference encoding tower. The static feature vector, value semantic feature vector, and latent preference feature vector of the target user are concatenated in a fixed order and input into the user preference encoding tower, outputting the initial preference vector of the target user. Similarly, the static feature vector, value semantic feature vector, and latent preference feature vector of each registered user on the platform in the effective candidate set are concatenated in a fixed order and input into the candidate preference encoding tower, outputting the initial preference vector of the candidate. An asymmetric bidirectional mutual attention layer is constructed, with independent trainable projection matrices Wu and Wc configured for the target user side and the candidate side, respectively. The initial preference vector U of the target user is projected onto Wu to obtain U', and the initial preference vector matrix C of the candidate is projected onto Wc to obtain C'. The mutual attention weight matrix is ​​calculated. The mutual attention weight matrix A is Softmax normalized row-wise to obtain the attention weight α of the target user for each candidate, and the mutual attention weight matrix A is Softmax normalized column-wise to obtain the attention weight β of each candidate for the target user. Based on the attention weights α and β, a bidirectional fused feature vector is generated. The bidirectional fused feature vector is input into two consecutive fully connected layers, where the first layer uses the ReLU function and the second layer uses the Sigmoid function. Based on the preset joint loss function, the bidirectional matching comprehensive score between the target user and the corresponding candidate is obtained. The joint loss function is composed of a weighted sum of bidirectional acceptance probability loss and latent preference similarity loss.

[0013] Furthermore, all registered users on the platform in the effective candidate set are sorted from high to low according to the comprehensive score of two-way matching. The top 50 registered users are used to generate an initial matching list. Registered users on the platform who have already viewed candidate information, liked, rejected, or chatted with the target user are filtered out. The final matching recommendation list is generated and returned to the target user who initiated the matchmaking request. Collect bidirectional interactive feedback data and perform confidence grading according to interaction depth. Pull the graded feedback data once per hour and use it as new dynamic behavior data to update the user preference time series data. Use the graded and weighted feedback samples as new training data. Use mini-batch gradient descent method to adjust the parameters of the third fully connected layer of the bidirectional preference dual-tower deep network according to the sample weight.

[0014] A matchmaking system based on user profiles includes: The data acquisition module collects three types of data from registered users on the dating platform: static profiles, dynamic behaviors, and value texts. It generates corresponding feature vectors, concatenates the feature vectors according to a preset time window, and generates a standardized preference time series dataset. The latent preference mining module constructs and trains a bidirectional ESN preference reserve pool with an embedded preference mutation detection gate. After receiving a matchmaking request from a target user, it extracts the standardized preference time series data of the target user and inputs it into the trained preference reserve pool. It extracts long-term preference dependency features through the forward reserve pool and short-term preference mutation features through the reverse reserve pool. It then merges and mines the latent preferences of the target user and outputs the latent preference feature vector. Finally, it selects and generates an initial latent preference candidate set. The bidirectional matching scoring module performs rigid filtering on the initial implicit preference candidate set based on the hard constraints preset by the target user, generates an effective candidate set, constructs a bidirectional preference dual-tower deep network with an embedded asymmetric bidirectional mutual attention layer, performs dimensional mapping on the preference vectors of the target user and the candidates through their respective independent feature projection matrices, calculates the asymmetric bidirectional attention weights, generates a bidirectional fusion feature vector, and calculates the bidirectional matching comprehensive score between the target user and users in the effective candidate set. The recommendation generation module generates a matching recommendation list based on the comprehensive score of two-way matching and returns it to the target user. It collects two-way interaction feedback data and updates the user preference time series data based on the feedback data. It also uses a confidence-weighted incremental learning method to adjust the parameters of the two-way preference dual-tower deep network.

[0015] (III) Beneficial Effects This invention provides a matchmaking method and system based on user profiles, which has the following beneficial effects: (1) Improve data credibility from the source of features to ensure the basic quality of modeling. Dynamic behavior noise filtering is achieved through dual verification of behavior links and dwell time, which weakens the interference of misoperation and curiosity browsing on preference modeling. At the same time, the value text features are corrected in reverse through the distribution of behavior preferences, which eliminates the text bias caused by the packaging of user persona, and makes the input features more in line with the user's real mate selection preferences, thereby improving the upper limit of matching accuracy from the data level.

[0016] (2) Achieve high-precision and high-response implicit preference time series modeling. By capturing long-term stable preferences and short-term abrupt preferences through bidirectional ESN, and combining preference abrupt detection gating to achieve triggered local incremental updates, it not only retains the high efficiency of ESN time series modeling, but also solves the problems of fixed-period update lag and full-scale fine-tuning that easily cause preference drift. It can accurately adapt to the dynamic changes of users' mate selection preferences.

[0017] (3) Construct an asymmetric bidirectional matching mechanism adapted to the matchmaking scenario. Asymmetric calculation of bidirectional attention is achieved through an independent projection matrix. This fits the real characteristics of the differences in the mate selection focus dimensions of both parties in the matchmaking scenario, breaks the default assumption of traditional symmetric attention, and makes the bidirectional compatibility calculation more in line with the real mate selection decision logic, thus significantly improving the bidirectional matching success rate.

[0018] (4) Construct a robust feedback loop iteration system, divide sample confidence by interaction depth, correct negative sample bias by browsing time, update model parameters by weight in incremental learning, effectively avoid low confidence samples misleading model iteration, ensure the stability of matching accuracy in long-term operation, and form a matching loop that can be continuously optimized. Attached Figure Description

[0019] Figure 1This is a schematic diagram illustrating the steps of the matchmaking method based on user profiles according to the present invention; Figure 2 This is a schematic diagram of the matchmaking method based on user profiles according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-2 This invention provides a matchmaking method based on user profiles, including the following: Step 1: Collect three types of data from registered users on the dating platform: static profiles, dynamic behaviors, and value texts. Generate corresponding feature vectors, concatenate the feature vectors according to a preset time window, and generate a standardized preference time series dataset. Step one includes the following: Step 101: Collect three types of core raw data from registered users on the dating platform. The first type is static profile data, including fixed tags such as age, gender, region, education, monthly income, marital history, height, and weight verified through real-name authentication. This data is collected through registration page entries combined with real-name authentication interface verification. The second type is dynamic behavioral data, including candidate profile browsing history, like / rejection operation history, chat duration, interaction frequency, and page dwell time. This data is collected at the minute level through front-end tracking and back-end logs, and basic cleaning is performed to remove duplicates and invalid null values. The third type is value text data, including self-statements of views on marriage and love, texts of mate selection requirements, and texts of answers to marriage and love questionnaires. This data is collected through personal profile pages and personalized questionnaire pages to construct the raw user dataset. Step 102: Perform standardization processing adapted to the data characteristics of the three types of raw data. For static profile data, discrete labels such as gender, region, education level, and marital history are processed using one-hot encoding, while continuous labels such as age, monthly income, height, and weight are processed using Min-Max normalization to generate a static feature vector of fixed dimensions. For dynamic behavior data, sort the data in ascending order by timestamp, perform fixed-rule numerical encoding on behavior types, and perform Min-Max normalization on behavior duration and page dwell time. At the same time, extract the candidate static labels pointing to the behavior and complete one-hot encoding to generate an initial dynamic behavior feature vector. For value text data, use a Chinese pre-trained BERT-base model to perform fine-grained semantic feature extraction, take the CLS token output results, and generate an initial value semantic feature vector of fixed dimensions. Step 103: To address the behavioral noise and text packaging issues in the matchmaking scenario, feature correction is performed, including dynamic behavior weighted calibration and value semantic correction. Dynamic behavior weighted calibration: For a single behavioral data point, if there is a subsequent continuous interaction chain, such as clicking "like" or initiating a chat after browsing, it is determined to be a strong genuine preference behavior, with a credibility weight of 1.2; if it is only a single browsing session and the dwell time is less than the 2-second duration threshold, it is determined to be a novelty-related accidental noise behavior, with a credibility weight of 0.3; other regular behaviors are weighted at 1.0. The credibility weights are then compared with the initial dynamic behavior features. The feature vectors are multiplied element-wise along their corresponding dimensions to generate the final dynamic behavior feature vector. Value semantic correction: The distribution of candidate tags corresponding to the user's dynamic behavior over the past 30 days is statistically analyzed to generate a behavior preference distribution vector. The cosine similarity between this vector and the initial value semantic feature vector is calculated. If the similarity is less than 0.5, it is determined that there is a significant deviation between the text description and the behavior preference. Linear interpolation correction is performed according to the behavior preference distribution vector, with a correction weight of 0.4. That is, the corrected vector = 0.6 × initial value vector + 0.4 × behavior preference distribution vector, generating the final value semantic feature vector. Step 104: Set a single time window to one calendar day. Concatenate the static feature vector, calibrated dynamic behavior feature vector, and corrected value semantic feature vector within the same window in a fixed order dimension to generate the initial feature vector for a single time window. Addressing the cross-modal temporal differences caused by slow changes in static features, rapid changes in dynamic features, and low update frequency of text features, the Dynamic Time Warping (DTW) algorithm is used to perform cross-modal temporal alignment, correcting temporal offsets of different modal features and ensuring semantic correspondence of features within the same time window in the time dimension. Arrange the aligned feature vectors of all time windows in chronological order to form single-user preference time-series data. The initial time series length is fixed at 30 consecutive time windows. Newly registered users are supplemented using the average feature vectors of users in the same age group and region. Perform Z-score global standardization on the same-dimensional features of all user preference time-series data to unify the numerical distribution range and eliminate dimensional differences. Store the data in the time-series database to complete the construction of the standardized preference time-series dataset.

[0022] When using this method, refer to steps 101 to 104: By improving data credibility from the source of features and ensuring the basic quality of modeling, dynamic behavioral noise filtering is achieved through dual verification of behavioral links and dwell time, reducing the interference of misoperation and novelty browsing on preference modeling. At the same time, the value text features are corrected in reverse through behavioral preference distribution, eliminating the text bias caused by user persona packaging, making the input features more in line with the user's real mate selection preferences, and improving the upper limit of matching accuracy from the data level.

[0023] Step 2: Construct and train a bidirectional ESN preference reserve pool with embedded preference mutation detection gating. After receiving the matchmaking request of the target user, extract its standardized preference time series data and input it into the trained preference reserve pool. Extract long-term preference dependency features through the forward reserve pool and extract short-term preference mutation features through the reverse reserve pool. Fuse and mine the latent preferences of the target user and output the latent preference feature vector. Filter and generate the initial latent preference candidate set. Step two includes the following: Step 201: Construct a bidirectional ESN (Echo State Network) preference reservoir with an embedded preference mutation detection gate. The architecture includes an input layer, a bidirectional reservoir layer, a preference mutation detection gate, and an output layer. The input layer dimension is completely consistent with the dimension of the standardized preference time series data. The bidirectional reservoir layer adopts a sparse connection structure with a fixed connection rate of 1%, a leakage rate α=0.8, and a spectral radius ρ=0.95, satisfying the echo state characteristics. The bidirectional reservoir simultaneously receives forward and reverse time series sequences. The forward sequence is from early to late time and is used to extract long-term preference dependency patterns. The reverse sequence is from late to early time and is used to extract short-term preference mutation features. The output layer dimension is fixed at 256 dimensions and is used to output latent preference feature vectors. The reservoir connection weights are initialized to random values ​​in the range [-0.1, 0.1] and remain fixed after initialization. The output layer weights are set to a trainable state. Step 202: Extract the full set of standardized user preference time-series data from the time-series database, divide it into training and validation sets in an 8:2 ratio, and train the output layer weights using the Ridge regression method with the goal of memorizing user preference association patterns. This enables the ESN model to discover implicit association rules such as "users with fitness preferences implicitly prefer regular lifestyles" and "users who are interested in cooking implicitly prefer family-oriented partners." When the accuracy of preference association pattern recognition on the validation set reaches the threshold of 88.5%, training is stopped, model parameters are saved, and the model is deployed. Configure a preference mutation detection gating system to monitor the feature distribution of newly added user time-series data in real time, calculate the KL divergence between the features of the newly added data and the average features of the user's historical 30 days. If the KL divergence exceeds the preset mutation threshold of 0.3, it is determined that a preference mutation has occurred, triggering a local incremental update of the reserve pool output layer. Only the output layer weights corresponding to the mutation feature dimension are adjusted, while the other weights remain fixed. The update learning rate is 1e-5 to avoid preference drift caused by full updates. Step 203: After receiving the matchmaking request from the target user, extract their latest standardized preference time series data and input it into the trained bidirectional ESN preference reserve pool. The forward reserve pool outputs 256-dimensional long-term preference features, and the reverse reserve pool outputs 256-dimensional short-term mutation preference features. The two types of features are concatenated and fused with a weight of 7:3. The final 256-dimensional latent preference feature vector is generated through the output layer. The cosine similarity between the latent preference vectors of all other users and the target user vector is calculated. Users with a similarity ≥ 0.7 are selected to form an initial latent preference candidate set. The size of the candidate set is controlled within the range of 1000~5000 people.

[0024] When using it, combine steps 201 to 203: Achieving high-precision and high-response implicit preference time series modeling, it captures long-term stable preferences and short-term abrupt preferences through bidirectional ESN, and achieves triggered local incremental updates by combining preference abrupt detection gating. This not only retains the efficiency of ESN time series modeling, but also solves the problems of fixed-period update lag and full-scale fine-tuning that easily cause preference drift, and can accurately adapt to the dynamic changes of users' mate selection preferences.

[0025] Step 3: Based on the hard constraints preset by the target user, perform rigid filtering on the initial implicit preference candidate set to generate an effective candidate set. Construct a bidirectional preference dual-tower deep network with an embedded asymmetric bidirectional mutual attention layer. Perform dimensional mapping on the preference vectors of the target user and the candidates through their respective independent feature projection matrices, calculate the asymmetric bidirectional attention weights, generate bidirectional fusion feature vectors, and calculate the bidirectional matching comprehensive score between the target user and users in the effective candidate set. Step three includes the following: Step 301: Extract the veto-type hard constraints actively set by the target user. By default, these include region, marital history, age range, and fertility intention. Generate a hard constraint feature vector for each constraint corresponding to one dimension. Assign a value of 1 if the condition is met, and 0 if it is not met. Call the hard constraint gating mask layer to multiply the hard constraint vectors of the target user and the candidate user bit by bit and calculate the element-wise sum. If the element-wise sum is less than the vector dimension, it is determined that there is a hard constraint that is not met. Set the matching score to 0 and exclude the user. This completes the rigid filtering and generates a rigid candidate subset. Count the number of users in the rigid candidate subset. If the number is less than the preset minimum threshold of 200, start the elastic relaxation mechanism: relax the restrictions of the corresponding constraints in order of priority from low to high, such as relaxing the region from the city to surrounding cities and relaxing the upper and lower limits of the age range by 2 years. Recount the number of candidates after each relaxation level until the number reaches the minimum threshold or all non-core constraints have been relaxed. Finally, generate an effective candidate set. Step 302: Construct a bidirectional preference dual-tower deep network, which includes a user preference encoding tower and a candidate preference encoding tower with consistent structure and non-shared parameters. Both towers adopt a 3-layer fully connected structure and the activation function is uniformly ReLU. The static feature vector, value semantic feature vector, and latent preference feature vector of the target user are concatenated in a fixed order and input into the user preference encoding tower. The output is a 128-dimensional initial preference vector U of the target user. The three types of feature vectors of each user in the effective candidate set are concatenated in the same order and input into the candidate preference encoding tower. The output is a 128-dimensional initial preference vector of the candidate. All candidate vectors form a matrix C. Step 303: Construct an asymmetric bidirectional mutual attention layer. Configure independent trainable projection matrices Wu and Wc for the target user side and candidate side respectively, to map the preference vectors of both sides to their respective attention feature spaces. Project the initial preference vector U of the target user to obtain U' = U × Wu, and project the candidate preference matrix C to obtain C' = C × Wc. Calculate the mutual attention weight matrix A = U' × (C')^T, where T is the transpose operation. Perform Softmax normalization on matrix A by rows to obtain the attention weight α of the target user for each candidate. Perform Softmax normalization on matrix A by columns to obtain the attention weight β of each candidate for the target user. Based on the bidirectional attention weights, perform... Feature weighting and fusion are used to generate a bidirectional fused feature vector F = α・U + β・C. The fused feature vector is then input into two consecutive fully connected layers. The first layer uses ReLU as the activation function, and the second layer uses Sigmoid. A bidirectional matching comprehensive score is calculated based on a joint loss function L = 0.6 × L1 + 0.4 × L2. Here, L1 is the bidirectional acceptance probability cross-entropy loss, calculated using cross-entropy loss based on bidirectional like and rejection tags from the historical interaction data of registered users on the platform. L2 is the latent preference cosine similarity loss, calculated based on the cosine similarity between the latent preference feature vectors of the target user and the corresponding candidate. The final score ranges from 0 to 1, with a higher score indicating a higher bidirectional fit.

[0026] When using this method, refer to the content of steps 301 to 303: An asymmetric bidirectional matching mechanism adapted to matchmaking scenarios is constructed. Asymmetric calculation of bidirectional attention is achieved through an independent projection matrix, which fits the real characteristics of the differences in the mate selection focus dimensions of both parties in matchmaking scenarios. It breaks the default assumption of traditional symmetric attention, making the bidirectional fit calculation more in line with the real mate selection decision logic and significantly improving the bidirectional matching success rate.

[0027] Step 4: Generate a matching recommendation list based on the comprehensive score of two-way matching and return it to the target user. Collect two-way interaction feedback data and update the user preference time series data based on the feedback data. Adjust the parameters of the two-way preference dual-tower deep network using a confidence-weighted incremental learning method.

[0028] Step four includes the following: Step 401: Sort the valid candidate users in descending order of their bidirectional matching comprehensive score, take the top 50 to generate an initial matching list, filter out users who have been viewed, liked, rejected, or chatted with by the target user, generate the final matching recommendation list and return it. Collect two types of bidirectional interactive feedback through front-end tracking and back-end logs: one is the target user's interaction data with the recommended candidates, such as likes, rejections, browsing time, and chat data; the other is the candidate's reverse interaction data with the target user. All feedback data is synchronously written to the Kafka message queue. Step 402: Pull feedback data from Kafka every hour, perform confidence level classification and assign sample weights according to interaction depth and behavioral motivation. Positive sample classification: mutual selection of likes and successful exchange of contact information are high-confidence positive samples with a weight of 1.5; one-sided liking and receiving a reply are medium-confidence positive samples with a weight of 1.0. Negative sample correction and classification: active rejection and browsing time of less than 3 seconds are judged as first impression rejection with exposure bias and are low-confidence negative samples with a weight of 0.5; active rejection and browsing time of more than 10 seconds are judged as rejection after in-depth understanding and are high-confidence negative samples with a weight of 1.2; read without reply for more than 72 hours are medium-confidence negative samples with a weight of 0.8. Step 403: The graded feedback data is used as new dynamic behavior data. Feature vectors are generated according to the preprocessing rules in Step 1. The single time window features and standardized preference time series data of the corresponding users are updated and stored synchronously in the time series database. The model is updated using incremental learning. The updated time series data is input into the bidirectional ESN preference reserve pool. Local updates are triggered according to the mutation gating rules. The graded weighted feedback samples are used as new training data. The mini-batch gradient descent method is used with a batch size of 32 and a learning rate of 1e-4. The parameters of the third fully connected layer of the bidirectional preference dual-tower network are fine-tuned. The parameters of the first two layers are frozen during the training process. The parameters are updated every 4 hours to avoid degradation of the original matching ability of the model. Step 404: Dynamically adjust the static tag weights based on continuous interactive behavior. If a user likes a candidate with a certain static tag three times in a row, the weight of that tag in the user profile is increased by 10%. If a user rejects a candidate with a certain tag three times in a row, the weight of that tag is decreased by 10%. All weights are limited to the range of 0.5 to 2.0. After the weight adjustment, the static feature vector is updated synchronously to provide more accurate input for the next matching.

[0029] When using this method, please refer to the content of steps 401 to 404: A robust feedback loop iterative system is constructed. By dividing the confidence of samples through interactive depth and correcting negative sample bias by combining browsing time, the model parameters are updated according to weights in incremental learning. This effectively avoids low-confidence samples from misleading model iteration, ensures the stability of matching accuracy in long-term operation, and forms a continuously optimized matching loop.

[0030] This invention also provides a matchmaking system based on user profiles, comprising: a data acquisition module, a latent preference mining module, a two-way matching scoring module, and a recommendation generation module, wherein: The data acquisition module collects three types of data from registered users on the dating platform: static profiles, dynamic behaviors, and value texts. It generates corresponding feature vectors, concatenates the feature vectors according to a preset time window, and generates a standardized preference time series dataset. The latent preference mining module constructs and trains a bidirectional ESN preference reserve pool with an embedded preference mutation detection gate. After receiving a matchmaking request from a target user, it extracts the standardized preference time series data of the target user and inputs it into the trained preference reserve pool. It extracts long-term preference dependency features through the forward reserve pool and short-term preference mutation features through the reverse reserve pool. It then merges and mines the latent preferences of the target user and outputs the latent preference feature vector. Finally, it selects and generates an initial latent preference candidate set. The bidirectional matching scoring module performs rigid filtering on the initial implicit preference candidate set based on the hard constraints preset by the target user, generates an effective candidate set, constructs a bidirectional preference dual-tower deep network with an embedded asymmetric bidirectional mutual attention layer, performs dimensional mapping on the preference vectors of the target user and the candidates through their respective independent feature projection matrices, calculates the asymmetric bidirectional attention weights, generates a bidirectional fusion feature vector, and calculates the bidirectional matching comprehensive score between the target user and users in the effective candidate set. The recommendation generation module generates a matching recommendation list based on the comprehensive score of two-way matching and returns it to the target user. It collects two-way interaction feedback data and updates the user preference time series data based on the feedback data. It also uses a confidence-weighted incremental learning method to adjust the parameters of the two-way preference dual-tower deep network.

[0031] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0032] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0033] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0034] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A matchmaking method based on user profiles, characterized in that: include: Collect three types of data from registered users on dating platforms: static profiles, dynamic behaviors, and value texts. Generate corresponding feature vectors, concatenate the feature vectors according to a preset time window, and generate a standardized preference time series dataset. A bidirectional ESN preference reserve pool with embedded preference mutation detection gating is constructed and trained. After receiving the matchmaking request of the target user, the standardized preference time series data is extracted and input into the trained preference reserve pool. Long-term preference dependency features are extracted through the forward reserve pool, and short-term preference mutation features are extracted through the reverse reserve pool. The implicit preferences of the target user are fused and mined, and the implicit preference feature vector is output. The initial implicit preference candidate set is then generated. Based on the hard constraints preset by the target user, a rigid filter is performed on the initial implicit preference candidate set to generate an effective candidate set. A bidirectional preference dual-tower deep network with an embedded asymmetric bidirectional mutual attention layer is constructed. The preference vectors of the target user and the candidates are mapped to dimensions through their respective independent feature projection matrices. The asymmetric bidirectional attention weights are calculated to generate bidirectional fusion feature vectors. The bidirectional matching comprehensive score between the target user and the users in the effective candidate set is calculated. A matching recommendation list is generated based on the comprehensive score of two-way matching and returned to the target user. Two-way interaction feedback data is collected and the user preference time series data is updated based on the feedback data. The parameters of the two-way preference dual-tower deep network are adjusted by a confidence-weighted incremental learning method.

2. The matchmaking method based on user profiles according to claim 1, characterized in that: Differential standardization preprocessing was performed on the three types of collected data. Discrete tags in the static profile data were processed using one-hot encoding, and continuous tags in the static profile data were processed using Min-Max normalization to generate static feature vectors. Dynamic behavior data was sorted in ascending order by timestamp, and behavior types were numerically encoded according to fixed rules. Behavior duration and page dwell time were processed using Min-Max normalization to generate initial dynamic behavior feature vectors. Semantic features were extracted from the value text data using a BERT-based semantic coding model pre-trained in Chinese. The output of the semantic coding model's CLS token was used to generate initial value semantic feature vectors.

3. The matchmaking method based on user profiles according to claim 2, characterized in that: For a single dynamic behavior data point, if there is a continuous interaction link after the behavior, the behavior credibility weight is assigned to 1.

2. If the behavior is only a single browsing and the dwell time is less than the preset time threshold, it is judged as a novelty noise behavior and the credibility weight is assigned to 0.

3. The credibility weight of other behaviors is assigned to 1.

0. The credibility weight is multiplied by the corresponding dimension of the initial dynamic behavior feature vector to complete the weighted calibration and generate the final dynamic behavior feature vector. The distribution of candidate tags corresponding to the dynamic behavior of the target user in the past 30 days is statistically analyzed to generate a behavior preference distribution vector. The cosine similarity between the behavior preference distribution vector and the initial value semantic feature vector is calculated. If the similarity is less than 0.5, the value semantic feature vector is linearly interpolated and corrected according to the behavior preference distribution vector with a correction ratio of 0.4 to generate the final value semantic feature vector.

4. The matchmaking method based on user profiles according to claim 3, characterized in that: The unit of a single time window is set to one calendar day. The static feature vectors, weighted and calibrated dynamic behavioral feature vectors, and consistency-corrected value semantic feature vectors generated within the same time window are concatenated dimensionally in a fixed order to generate the initial feature vector for a single time window. For the differences in the temporal change frequency of the three types of features (static, dynamic, and value), a dynamic time warping algorithm is used to perform cross-modal temporal alignment. All time-window aligned feature vectors are arranged in chronological order to form the preference time-series data for a single platform registered user. For the same-dimensional features of the preference time-series data for all platform registered users, the Z-score normalization method is used to perform global normalization processing, completing the construction of the standardized preference time-series dataset.

5. The matchmaking method based on user profiles according to claim 1, characterized in that: The constructed bidirectional ESN preference reservoir with embedded preference mutation detection gating includes an input layer, a bidirectional reservoir layer, a preference mutation detection gating, and an output layer. The bidirectional reservoir layer simultaneously receives the forward and reverse sequences of standardized preference time-series data. Standardized preference time-series data of registered users on all platforms are extracted and used as training input data for the bidirectional ESN preference reserve pool. The training objective is to memorize the preference association patterns of registered users on different platforms. The output layer weights of the bidirectional ESN preference reserve pool are trained using the ridge regression method. The training termination condition is set: when the recognition accuracy of the bidirectional ESN preference reserve pool for user preference association patterns on the validation set reaches or exceeds the accuracy threshold, the training is stopped. The preference mutation detection gating monitors the feature distribution of newly added time-series data in real time, calculates the KL divergence between the new data features and the mean of historical features, and if the KL divergence exceeds the preset mutation threshold, it triggers a local incremental update of the weights of the reserve pool output layer, only adjusting the weight parameters corresponding to the mutation feature dimension, while the other weights remain fixed.

6. The matchmaking method based on user profiles according to claim 5, characterized in that: The system receives matchmaking requests initiated by target users, extracts their standardized preference time-series data, and inputs this data into a bidirectional ESN preference reserve pool. The system outputs long-term preference features from the forward reserve pool and short-term preference mutation features from the reverse reserve pool. These two types of features are then concatenated and fused with a 7:3 weighting, and the output layer outputs the target user's latent preference feature vector. Simultaneously, the system selects latent preference feature vectors from other platform registered users whose latent preference feature vectors have a similarity ≥ 0.7 with the target user's latent preference feature vector from the bidirectional ESN preference reserve pool, forming an initial latent preference candidate set.

7. The matchmaking method based on user profiles according to claim 6, characterized in that: Extract the hard constraints of the target users who initiated the matchmaking request. Generate a fixed-dimensional hard constraint feature vector according to the fixed rule that each hard constraint corresponds to a dimension. The dimension that satisfies the corresponding hard constraint is assigned a value of 1, and the dimension that does not satisfy the hard constraint is assigned a value of 0. Call the hard constraint gating mask layer to compare the hard constraint feature vector of the target user with the hard constraint feature vectors of all platform registered users in the initial implicit preference candidate set bit by bit, and calculate the sum of the element-wise products of the two vectors. For platform registered users whose product sum is less than the dimension of the hard constraint feature vector, their matching score is directly set to 0 and they are excluded, generating a rigidly filtered candidate subset. If the number of users in the candidate subset after rigid filtering is lower than the preset minimum threshold, the constraints are relaxed in order of preset constraint priority from low to high. The priorities from low to high are geographical scope, age range, income requirements, and marital history requirements. The number of candidate subsets is recalculated for each level of constraint relaxation until the number of candidate subsets reaches the minimum threshold or all non-core constraints have been relaxed, and finally a valid candidate set is generated.

8. The matchmaking method based on user profiles according to claim 7, characterized in that: A bidirectional preference dual-tower deep network is constructed, comprising a user preference encoding tower and a candidate preference encoding tower. The static feature vector, value semantic feature vector, and latent preference feature vector of the target user are concatenated in a fixed order and input into the user preference encoding tower, outputting the initial preference vector of the target user. The static feature vector, value semantic feature vector, and latent preference feature vector of each registered user on the platform in the effective candidate set are concatenated in a fixed order and input into the candidate preference encoding tower, outputting the initial preference vector of the candidate. Construct an asymmetric bidirectional mutual attention layer, and configure independent trainable projection matrices Wu and Wc for the target user side and the candidate side, respectively. The initial preference vector U of the target user is projected through Wu to obtain U', and the initial preference vector matrix C of the candidates is projected through Wc to obtain C'. The mutual attention weight matrix is ​​calculated. The mutual attention weight matrix A is Softmax normalized row by row to obtain the attention weight α of the target user for each candidate. The mutual attention weight matrix A is Softmax normalized column by column to obtain the attention weight β of each candidate for the target user. Based on the attention weights α and β, a bidirectional fused feature vector is generated. The bidirectional fused feature vector is input into two consecutive fully connected layers, where the first layer uses the ReLU function and the second layer uses the Sigmoid function. Based on the preset joint loss function, the bidirectional matching comprehensive score between the target user and the corresponding candidate is obtained. The joint loss function is composed of a weighted sum of bidirectional acceptance probability loss and latent preference similarity loss.

9. The matchmaking method based on user profiles according to claim 1, characterized in that: All registered users on the platform in the valid candidate set are sorted from high to low according to the comprehensive score of two-way matching. The top 50 registered users are used to generate an initial matching list. Registered users on the platform who have already viewed candidate information, liked, rejected, or chatted with the target user are filtered out. The final matching recommendation list is generated and returned to the target user who initiated the matchmaking request. Collect bidirectional interactive feedback data and perform confidence grading according to interaction depth. Pull the graded feedback data once per hour and use it as new dynamic behavior data to update the user preference time series data. Use the graded and weighted feedback samples as new training data. Use mini-batch gradient descent method to adjust the parameters of the third fully connected layer of the bidirectional preference dual-tower deep network according to the sample weight.

10. A user profile-based matchmaking system for implementing the user profile-based matchmaking method according to any one of claims 1 to 9, characterized in that: include: The data acquisition module collects three types of data from registered users on the dating platform: static profiles, dynamic behaviors, and value texts. It generates corresponding feature vectors, concatenates the feature vectors according to a preset time window, and generates a standardized preference time series dataset. The latent preference mining module constructs and trains a bidirectional ESN preference reserve pool with an embedded preference mutation detection gate. After receiving a matchmaking request from a target user, it extracts the standardized preference time series data of the target user and inputs it into the trained preference reserve pool. It extracts long-term preference dependency features through the forward reserve pool and short-term preference mutation features through the reverse reserve pool. It then merges and mines the latent preferences of the target user and outputs the latent preference feature vector. Finally, it selects and generates an initial latent preference candidate set. The bidirectional matching scoring module performs rigid filtering on the initial implicit preference candidate set based on the hard constraints preset by the target user, generates an effective candidate set, constructs a bidirectional preference dual-tower deep network with an embedded asymmetric bidirectional mutual attention layer, performs dimensional mapping on the preference vectors of the target user and the candidates through their respective independent feature projection matrices, calculates the asymmetric bidirectional attention weights, generates a bidirectional fusion feature vector, and calculates the bidirectional matching comprehensive score between the target user and users in the effective candidate set. The recommendation generation module generates a matching recommendation list based on the comprehensive score of two-way matching and returns it to the target user. It collects two-way interaction feedback data and updates the user preference time series data based on the feedback data. It also uses a confidence-weighted incremental learning method to adjust the parameters of the two-way preference dual-tower deep network.

Images