A method and system for friend recommendation based on user information

Abstract

This invention discloses a method and system for recommending friends based on user information, belonging to the field of social recommendation technology. It collects behavioral data of target users within a preset time period and scenario, constructing a user behavior vector group containing multiple weak behavioral feature sub-vectors. The sub-vectors are input into a feature weight analysis model to calculate their relevance weight coefficients under different dimensions of friend-making motivation. Based on the weight coefficients, the social need intensity score of the target user in the current context is calculated. The behavioral vector groups and need scores of other users are compared to construct a user matching similarity matrix. Based on this matrix, the top N users with the highest similarity are selected, and a personalized recommendation list is generated by combining the activity score. This invention can dynamically mine users' potential social intentions, achieving high-precision and highly adaptable friend-making recommendations.

Description

Technical Field

[0001] This invention relates to the field of social recommendation technology, specifically to a method and system for making friends based on user information. Background Technology

[0002] Currently, most social media platforms primarily rely on user interest tags, geographical location, or simple behavioral data such as likes and browsing history for friend recommendations. However, this type of recommendation often suffers from problems such as overly generalized matching results, a lack of in-depth understanding, and poor user engagement.

[0003] Especially among urban commuters, many users experience high stress and low willingness to communicate during specific time periods, making their behavioral data unreliable in reflecting their true social needs. For example, a user frequently browsing short videos during their subway commute might be identified by the recommendation system as having a preference for solitude, while ignoring their high level of interaction during lunch breaks or at night, leading to a misjudgment of their social tendencies.

[0004] Furthermore, different users exhibit highly heterogeneous behavioral patterns in specific scenarios, and simple, uniform recommendation models cannot effectively extract these scenario-specific features. For example, the same level of browsing interest similarity has completely different recommendation value in a coffee shop scenario and a scenario of being alone at night. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for recommending friends based on user information, in order to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for recommending friends based on user information, comprising: S1. Collect behavioral data of target users in preset time periods and scenarios; S2. Construct a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; S3. Input the weak behavioral feature sub-vectors into the preset feature weight analysis model, and calculate the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation; S4. Calculate the social demand intensity score R of the target user in the current context based on the relevance weighting coefficient T. S5. Compare the behavioral vector groups F′ of other users with their corresponding social demand intensity scores R′, and construct the user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social demand dimension. S6. Based on the user matching similarity matrix M, filter out the top N users who are closest to the target user in the current social needs dimension, and generate a personalized dating recommendation list.

[0007] Preferably, the behavioral data includes, but is not limited to: interaction frequency, content dwell time, input behavior intensity, and interface jump frequency.

[0008] Preferably, the calculation of the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation includes: For each weak behavior feature subvector in the user behavior vector group F, input it according to the time period-scene combination label, and then perform feature extraction; A deep neural network model is used to perform multi-layer feature fusion processing on weak behavioral feature sub-vectors to extract the response distribution of each feature dimension under the dimension of friendship motivation. Based on the preset set of semantic tags for dating motivations, the fused feature vectors are mapped to the relevance scoring space of each motivation dimension to obtain the corresponding relevance weight coefficient T.

[0009] Preferably, the calculation of the target user's social need intensity score R in the current context includes: Perform a one-to-one feature mapping between the weak behavioral feature subvectors of the target user under the current time period and scenario combination and the corresponding relevance weight coefficients; The weighted behavioral response vector is calculated based on the product of each feature dimension and its corresponding weight. The weighted behavioral response value vector is input into a first-order linear combination function to obtain the original social demand score; The original social need score is normalized using a preset dynamic adjustment function, and the social need intensity score R of the target user in the current context is output.

[0010] Preferably, constructing the user matching similarity matrix M includes: Obtain the social demand intensity score and weak behavioral feature sub-vectors of the target users, and select several comparison users with the same time period and scenario combination; The user's behavior vector set and social need intensity score are compared and standardized. Based on the defined social needs dimension, the weighted behavioral response value vectors of the target user and each comparison user are input into the cosine similarity function to calculate the similarity score; The calculation results are filled into the similarity matrix according to the correspondence between user pairs to generate the user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social needs dimension.

[0011] Preferably, generating a personalized dating recommendation list includes: Extract the data of the row containing the target user from the user matching similarity matrix to form a set of similarity scores with all compared users; Based on the defined social needs dimension, user pairs with similarity scores higher than a preset threshold are selected from the set of similarity scores; The selected user pairs are sorted from high to low according to their similarity scores, and the top N users are selected as the recommendation targets. By combining the activity ratings of each recommended user within the current time period and scenario, the final recommendation order is adjusted to generate a personalized dating recommendation list.

[0012] This invention also provides a dating recommendation system based on user information, comprising: The behavior data collection module collects behavior data of target users within a preset time period and scenario. The behavior feature construction module constructs a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; The feature weight analysis module inputs the weak behavior feature sub-vectors into a preset feature weight analysis model and calculates the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation. The social needs modeling module calculates the social needs intensity score R of the target user in the current context based on the relevance weight coefficient T. The similarity calculation module compares the behavior vector groups F′ of other users and their corresponding social need intensity scores R′ to construct a user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social need dimension. The recommendation generation module, based on the user matching similarity matrix M, filters out the top N users who are closest to the target user in the current social needs dimension, and generates a personalized dating recommendation list.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention introduces multi-dimensional behavioral feature modeling and contextual scene recognition mechanisms to accurately mine weak behavioral features of users in different time periods and scenarios. It also combines these with semantic tags representing friendship motivations to construct a feature weight analysis model, thereby achieving refined modeling of users' social needs. Compared to existing recommendation methods that rely solely on static tags or single behavioral data, this invention can dynamically perceive users' potential social intentions and contextual changes, significantly improving the relevance and matching degree of recommendation results.

[0014] 2. This invention constructs a user matching similarity matrix based on social needs and combines it with user activity factors for comprehensive ranking, thereby improving the real-time performance and interactivity of recommendation responses while ensuring matching accuracy. The overall technical solution has the advantages of strong scalability, high matching quality, and excellent user acceptance, and is suitable for various types of social platform scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] Example 1, please refer to Figure 1 As shown in this embodiment, a friend recommendation method based on user information includes: S1. Collect behavioral data of target users in preset time periods and scenarios; S2. Construct a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; S3. Input the weak behavioral feature sub-vectors into the preset feature weight analysis model, and calculate the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation; S4. Calculate the social demand intensity score R of the target user in the current context based on the relevance weighting coefficient T. S5. Compare the behavioral vector groups F′ of other users with their corresponding social demand intensity scores R′, and construct the user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social demand dimension. S6. Based on the user matching similarity matrix M, filter out the top N users who are closest to the target user in the current social needs dimension, and generate a personalized dating recommendation list.

[0020] In this invention, step S1 is used to collect behavioral data of the target user in a preset time period and scenario.

[0021] Specifically, the activity cycle of the target users is first divided into multiple preset time periods, such as: morning peak (7:00–9:00), work period (9:00–12:00), lunch break (12:00–13:30), afternoon work period (13:30–18:00), evening leisure period (18:00–23:00), and nighttime period (23:00–7:00 the next day). At the same time, combined with the user's application usage scenario, such as "home network environment", "commuting environment (based on base station / accelerometer sensing)", "office Wi-Fi environment", etc., a "time period-scenario" combination index is formed.

[0022] For each time period and scenario combination, the system collects the following behavioral data: Interaction frequency: Records the number of times a user takes initiative on social media platforms, including liking, commenting, sending private messages, adding friends, and viewing other people's profiles; Content dwell time: Monitors the average browsing time of users on a certain type of content (such as short videos, graphic and text updates) to determine their content engagement. Input behavior intensity: Collect user behavior characteristics in the input box, such as typing speed, pause duration, deletion frequency, etc., to infer their expression intention and emotional fluctuations; Interface jump frequency: Monitors the number and order in which users jump between different interfaces (such as homepage, recommendation page, chat page, friend list page) to determine their browsing path and exploration intent.

[0023] The above data will be anonymized after collection and stored in a structured manner according to time period and scenario dimension, providing a data foundation for the subsequent construction of user weak behavior feature vectors.

[0024] S2. Construct a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; After completing the behavioral data collection described in step S1, the method described in this embodiment is further executed to construct a user behavior vector group F based on the behavioral data, so as to realize the structured expression of weak behavioral features of users in multi-temporal and spatial scenarios.

[0025] Specifically, all behavioral data of the target user are categorized and organized according to "time period-scenario combination", constructing the following vector set: F={f1,f2,...,fn}; where each sub-vector fi corresponds to an independent time period-scenario combination Ci, i.e.: fi=[xi1,xi2,xi3,xi4]; the dimensions of the above vectors represent, in order: xi1: Normalized value of interaction frequency; xi2: Average content dwell time (average number of seconds per unit time); xi3: Input behavior intensity index (such as the behavior energy value weighted by typing speed and pause rate); xi4: Interface jump frequency index (reflects the activity level of user exploration behavior in the current scenario).

[0026] To enhance the contrast and adaptability of features, the original collected data is normalized, and a weighted moving average algorithm is used to smooth out outliers in order to eliminate the interference of occasional behavior on the overall feature representation.

[0027] For example, for a user in a "weekday morning" scenario, their behavioral data is as follows: interaction frequency 12 times, average content dwell time 32 seconds, input behavior intensity score 0.78, and interface jump frequency 4 times per minute. After standardization, the system constructs it into the following weak behavioral feature sub-vector: fwork_am=[0.45,0.60,0.78,0.52]. Repeating the above process, behavioral sub-vectors for the target user under all preset time period-scenario combinations are generated sequentially, ultimately forming a complete behavioral vector set F. This vector set serves as input for subsequent social motivation modeling and recommendation algorithms, providing multi-dimensional and multi-contextual feature support for judging user social needs.

[0028] S3. Input the weak behavioral feature sub-vectors into the preset feature weight analysis model, and calculate the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation.

[0029] First, each weak behavioral feature sub-vector in the behavioral vector group is bound to its corresponding time period label and scene label to form a triplet data structure, represented as: [time period label, scene label, behavioral feature sub-vector]. For example, the label of a user's sub-vector under the combination of "evening time period - home scene" is [E5,S2,f5], where f5 is a four-dimensional behavioral feature vector.

[0030] Subsequently, the triplet data is input into the feature extraction structure. The feature extraction process employs standard feature normalization and dimensionality normalization methods to limit the value range of all behavioral feature sub-vectors to the [0,1] interval, avoiding interference caused by scale inconsistencies between feature dimensions.

[0031] The normalized weak behavioral feature vectors are input into a multi-layer feedforward neural network structure. The network as a whole adopts a four-layer structure consisting of an input layer, two hidden layers, and an output layer. Each hidden layer has 64 neurons, and the rectified linear unit function (ReLU) is used as the activation function.

[0032] This network structure introduces an attention weight mechanism to dynamically adjust the contribution of each feature dimension to the output representation. Specifically, assuming the input feature vector is f=[x1,x2,x3,x4], the network calculates the attention scores a1 to a4 corresponding to each feature dimension and generates a fusion vector v using a weighted summation method. The fusion result v is equal to the sum of the products of the feature values ​​and their corresponding attention scores; that is, the fusion vector is equal to the sum of the products of the feature vector's dimension values ​​and their attention scores. Through the above fusion, a dense feature vector v′ for motivational response modeling is generated, with a dimension of 1×32.

[0033] At this stage, a set of semantic tags for friendship motivation is set to represent common social motivation dimensions, including but not limited to: interest-based social interaction (labeled M1), career-based connection (M2), acquaintance based on geographical location (M3), need for emotional companionship (M4), and knowledge-sharing interaction (M5), for a total of 5 friendship motivation dimensions.

[0034] A five-dimensional motivation rating space is constructed, with each dimension corresponding to a dating motivation label. The aforementioned fused feature vector v′ is input into a motivation mapping model, which is a set of five parallel fully connected output layers, each used to output a response rating for each dating motivation dimension.

[0035] The motivation mapping model is trained using a normalized cross-entropy loss function, enabling it to predict the response intensity of the current behavior vector under each motivation dimension based on the user's past behavior data. The output is a set of weight coefficients T=[t1,t2,t3,t4,t5], where t1 to t5 correspond to the relevance scores of the five dating motivation dimensions mentioned above.

[0036] The output relevance weight coefficient T will be associated with and stored with the original weak behavior feature sub-vectors, and used in subsequent steps to quantitatively model the strength of users' social needs.

[0037] S4. Calculate the social demand intensity score R of the target user in the current context based on the relevance weight coefficient T.

[0038] After calculating the relevance weighting coefficients, the modeling process for the social demand intensity score begins, which measures the target user's degree of social willingness within the current time period and scenario combination. This process includes the following four steps: Perform a one-to-one feature mapping between the weak behavioral feature vectors of the target user under the current time period and scenario combination and the corresponding relevance weight coefficients: First, weak behavioral feature sub-vectors of the target user in the current time period and scenario combination are extracted from the behavioral vector group, denoted as: fcurrent=[x1,x2,x3,x4], which represent the normalized interaction frequency, content dwell time, input behavior intensity and interface jump frequency, respectively.

[0039] Simultaneously, the relevance weight coefficient vector Tcurrent = [t1, t2, t3, t4] corresponding to this sub-vector is extracted. This vector is output by the aforementioned relevance analysis model of dating motivation, representing the response importance of each feature dimension in the current context. Next, a one-to-one feature mapping is performed, matching each feature dimension in fcurrent with the weight of the corresponding dimension in Tcurrent.

[0040] The weighted behavioral response vector Ra is calculated by multiplying each dimension sequentially, with the formula: Ra = [x1×t1, x2×t2, x3×t3, x4×t4]. This operation can be seen as assigning different behavioral features their actual influence within the context of current social motivation, strengthening the contribution of high-weight features, and weakening low-weight or interfering features.

[0041] To synthesize the four weighted behavioral response values, a first-order linear combination function S is constructed to output the original social need score, denoted as Roriginal. This function is defined as follows: Roriginal = w1×(x1×t1) + w2×(x2×t2) + w3×(x3×t3) + w4×(x4×t4) + b, where w1 to w4 are fixed coefficients obtained by the system training based on historical data, representing the global importance of each behavior type in the final score, and b is a bias term constant. It is suggested that the initial weights be set to w1=0.3, w2=0.25, w3=0.25, w4=0.2, and the bias term b be 0.05. These can be fine-tuned through training after actual deployment.

[0042] To prevent calculation bias caused by differences in the distribution of original scores across different scenarios, a dynamic adjustment function f-adjustment is introduced to normalize the original scores, yielding the final social demand intensity score R: R = f-adjustment(Roriginal, μcurrent, σcurrent), where μcurrent is the mean of the original scores for the user group in the current scenario, and σcurrent is the standard deviation. The adjustment function employs a piecewise normalization method based on a normal distribution, with the formula: R = 1 ÷ (1 + exp[-(Roriginal - μcurrent) ÷ σcurrent]). This function maps the original social demand scores to the (0,1) interval, adapting to the distribution characteristics of social behavior in different scenarios and enhancing the comparability and stability of the results.

[0043] The final output score R, representing the intensity of social needs, can serve as a foundational indicator for subsequent user matching and friend recommendations, enabling more targeted and timely social push notifications.

[0044] S5. Compare the behavioral vector groups F′ of other users with their corresponding social need intensity scores R′, and construct a user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social need dimension.

[0045] First, extract the weak behavioral feature sub-vector f and the social demand intensity score R from the target user's historical behavior data, which correspond to the current time period and scenario combination.

[0046] Next, other users with the same time period and scenario combination are retrieved from the user database to form a comparison user set, denoted as U′. Each comparison user j contains its corresponding weak behavioral feature sub-vector fj and social need intensity score Rj.

[0047] This constraint method ensures that all users participating in similarity calculations are in a consistent behavioral environment, which helps improve the accuracy and contextual relevance of matching results.

[0048] To eliminate the impact of differences in individual user behavior on subsequent similarity calculations, all feature vectors and score data involved in the calculation are standardized.

[0049] The specific method employs the Z-score standardization algorithm, which is defined as follows: for each feature dimension x, the mean μ and standard deviation σ of this feature in the user set are used to transform the result, and the standardized value x′ is calculated. .

[0050] Similarly, the social need intensity score R is processed in the same way to obtain the standardized score R′. The standardization operation ensures that all vectors are within the same numerical range, avoiding the disproportionate influence of a high-value dimension on the cosine calculation.

[0051] After standardization, for each comparison user j, extract its weighted behavioral response value vector fj′ and the target user's weighted vector ftarget′, and perform pairwise calculations based on the set social needs dimension.

[0052] Similarity calculation uses the cosine similarity function, which is mathematically defined as: similarity Where, "•" represents the vector dot product, "" represents the Euclidean norm of the vector (i.e., the length of the vector).

[0053] The calculation result is a real number, ranging from -1 to 1. The closer the value is to 1, the more similar the behavioral patterns of the two users are in the current social needs dimension.

[0054] To further improve differentiation, the system can set a similarity score threshold δ. It is recommended to use the default value of δ=0.6 and only retain user pairs with similarity greater than this threshold for subsequent recommendations.

[0055] Construct a two-dimensional matrix M with dimensions n×n, where n is the total number of users participating in the matching calculation. Fill the corresponding position Mi,j in the matrix with the similarity score of each user pair (user i, user j).

[0056] Each element Mi,j in the matrix represents the matching similarity between user i and user j under the current defined social needs dimension. To ensure matrix symmetry, the similarity results must satisfy Mi,j = Mj,i.

[0057] The completed user matching similarity matrix M will serve as the basic input data for subsequent user recommendation filtering and ranking, providing target users with high-precision, scenario-matched personalized social recommendation services.

[0058] S6. Based on the user matching similarity matrix M, filter out the top N users who are closest to the target user in the current social needs dimension, and generate a personalized dating recommendation list.

[0059] First, in the generated user matching similarity matrix, locate the row containing the target user, denoted as row i. This row contains the similarity scores between the target user and all other users being compared.

[0060] By extracting all elements from the row except itself, a one-dimensional array structure is formed, denoted as Si=[s1,s2,...,sn], where sn represents the similarity score between the target user and user n under the defined social needs dimension. This set serves as the basic input for subsequent recommendation object filtering and ranking.

[0061] To ensure that recommended users have sufficient similarity in behavioral patterns, a similarity score screening threshold δ is introduced. This threshold is set based on historical user matching experience, with an initial recommendation value of 0.65, which can be dynamically adjusted according to actual platform data.

[0062] In the similarity score set Si mentioned above, all user pairs that satisfy sj≥δ are selected to form a candidate user set U.

[0063] This step can filter out user pairs with insufficient similarity, improving the matching quality of the final recommendation results and user satisfaction.

[0064] The user pairs in candidate U are sorted in descending order according to their corresponding similarity scores to obtain the sorted user sequence U.

[0065] Set the upper limit parameter N for the recommended quantity. The value of N is set according to the product design scenario, and common values ​​are 5, 10 or 20.

[0066] The top N users from the ranking results are selected to form the initial recommendation set U, which serves as the core dating candidate group for the current recommendation period.

[0067] To improve the real-time nature and interactivity of recommended content, an activity scoring factor Aj is introduced to reflect the online activity of each candidate user j in the current time period and scenario combination.

[0068] The activity score Aj is calculated by weighting the following behavioral indicators: Aj = q1 × login frequency weight + q2 × page dwell time weight + q3 × recent interaction frequency weight, where q1, q2, and q3 are system-preset weight coefficients that can be set according to actual business scenarios, with initial recommended values ​​of 0.3, 0.4, and 0.3 respectively. The similarity score sj and the activity score Aj are then fused using a first-order linear combination to obtain the final comprehensive recommendation score: Comprehensive Score = w1 × sj + w2 × Aj. The initial values ​​of both w1 and w2 are set to 0.5, indicating that the recommendation system maintains a balance between recommendation quality and real-time responsiveness.

[0069] Users are re-ranked based on their overall scores, and a final personalized dating recommendation list is generated for front-end display.

[0070] Example 2, please refer to Figure 2 As shown in this embodiment, a dating recommendation system based on user information includes: The behavior data collection module collects behavior data of target users within a preset time period and scenario. The behavior feature construction module constructs a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; The feature weight analysis module inputs the weak behavior feature sub-vectors into a preset feature weight analysis model and calculates the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation. The social needs modeling module calculates the social needs intensity score R of the target user in the current context based on the relevance weight coefficient T. The similarity calculation module compares the behavior vector groups F′ of other users and their corresponding social need intensity scores R′ to construct a user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social need dimension. The recommendation generation module, based on the user matching similarity matrix M, filters out the top N users who are closest to the target user in the current social needs dimension, and generates a personalized dating recommendation list.

[0071] 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 method for recommending friends based on user information, characterized in that: include: S1. Collect behavioral data of target users in preset time periods and scenarios; S2. Construct a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; S3. Input the weak behavioral feature sub-vectors into the preset feature weight analysis model, and calculate the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation; S4. Calculate the social demand intensity score R of the target user in the current context based on the relevance weighting coefficient T. S5. Compare the behavioral vector groups F′ of other users with their corresponding social demand intensity scores R′, and construct the user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social demand dimension. S6. Based on the user matching similarity matrix M, filter out the top N users who are closest to the target user in the current social needs dimension, and generate a personalized dating recommendation list.

2. The method for recommending friends based on user information according to claim 1, characterized in that: The behavioral data includes interaction frequency, content dwell time, input behavior intensity, and interface jump frequency.

3. The method for recommending friends based on user information according to claim 1, characterized in that: The calculation of the relevance weight coefficient T for each sub-vector under different dimensions of friendship motivation includes: For each weak behavior feature subvector in the user behavior vector group F, input it according to the time period-scene combination label, and then perform feature extraction; A deep neural network model is used to perform multi-layer feature fusion processing on weak behavioral feature sub-vectors to extract the response distribution of each feature dimension under the dimension of friendship motivation. Based on the preset set of semantic tags for dating motivations, the fused feature vectors are mapped to the relevance scoring space of each motivation dimension to obtain the corresponding relevance weight coefficient T.

4. The method for recommending friends based on user information according to claim 1, characterized in that: The calculation of the target user's social need intensity score R in the current context includes: Perform a one-to-one feature mapping between the weak behavioral feature subvectors of the target user under the current time period and scenario combination and the corresponding relevance weight coefficients; The weighted behavioral response vector is calculated based on the product of each feature dimension and its corresponding weight. The weighted behavioral response value vector is input into a first-order linear combination function to obtain the original social demand score; The original social need score is normalized using a preset dynamic adjustment function, and the social need intensity score R of the target user in the current context is output.

5. The method for recommending friends based on user information according to claim 1, characterized in that: The construction of the user matching similarity matrix M includes: Obtain the social demand intensity score and weak behavioral feature sub-vectors of the target users, and select several comparison users with the same time period and scenario combination; The user's behavior vector set and social need intensity score are compared and standardized. Based on the defined social needs dimension, the weighted behavioral response value vectors of the target user and each comparison user are input into the cosine similarity function to calculate the similarity score; The calculation results are filled into the similarity matrix according to the correspondence between user pairs to generate the user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social needs dimension.

6. The method for recommending friends based on user information according to claim 1, characterized in that: The generation of a personalized dating recommendation list includes: Extract the data of the row containing the target user from the user matching similarity matrix to form a set of similarity scores with all compared users; Based on the defined social needs dimension, user pairs with similarity scores higher than a preset threshold are selected from the set of similarity scores; The selected user pairs are sorted from high to low according to their similarity scores, and the top N users are selected as the recommendation targets. By combining the activity ratings of each recommended user within the current time period and scenario, the final recommendation order is adjusted to generate a personalized dating recommendation list.

7. A friend recommendation system based on user information, used to implement the friend recommendation method based on user information as described in any one of claims 1-6, characterized in that, include: The system includes a behavior data acquisition module, a behavior feature construction module, a feature weight analysis module, a social demand modeling module, a similarity calculation module, and a recommendation generation module; among them, The behavior data collection module collects behavior data of target users within a preset time period and scenario; The behavior feature construction module constructs a user behavior vector group F based on the behavior data, wherein F contains weak behavior feature sub-vectors of the target user under multiple time period-scenario combinations; The feature weight analysis module inputs the weak behavior feature sub-vectors into a preset feature weight analysis model and calculates the relevance weight coefficient T of each sub-vector under different dimensions of friendship motivation. The social demand modeling module calculates the social demand intensity score R of the target user in the current context based on the relevance weight coefficient T. The similarity calculation module compares the behavior vector groups F′ of other users with their corresponding social need intensity scores R′ and constructs a user matching similarity matrix M, where each element Mi,j of M represents the similarity between user i and user j under the set social need dimension. The recommendation generation module, based on the user matching similarity matrix M, filters out the top N users who are closest to the target user in the current social needs dimension, and generates a personalized dating recommendation list.

Images