A home decoration designer intelligent recommendation method and system

Abstract

The application relates to a home decoration designer intelligent recommendation method and system, which comprises the following steps: obtaining user demand information and historical behavior information, and extracting demand features and historical behavior features; obtaining a designer database and classifying the same to form a multi-dimensional feature vector; inputting the demand features, the historical behavior features and the multi-dimensional feature vector into a matching model to generate a matching degree matrix; generating a demand solution of the matching degree matrix based on a loss function and a solving algorithm; and generating a recommendation list based on the generated demand solution; the application extracts demand features and historical behavior features, inputs the multi-dimensional feature vector into a matching model to generate a matching degree matrix, generates a demand solution of the matching degree matrix based on a loss function and a solving algorithm, and generates and outputs a recommendation list, so that dynamic preference mining of home decoration design and designer recommendation combined with multi-dimensional features are realized, and the practicability of a recommendation system and user satisfaction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent analysis, in particular to a home decoration designer intelligent recommendation method and system. BACKGROUND

[0002] As an important part of residential environment transformation, the personalized service demand of home decoration design is increasing. With the popularization of Internet home decoration platforms, how to efficiently match users and designers has become a key problem to improve the service efficiency of the industry.

[0003] The existing home decoration designer recommendation system usually adopts keyword matching, collaborative filtering or content similarity calculation and other ways for recommendation, such as directly comparing the designer tags according to the static demand filled by the user, such as budget, style preference, etc.

[0004] However, the above method has obvious limitations: first, it often ignores the dynamic preferences implied in the user's historical behavior, resulting in a lack of personalization in the recommendation results; second, most systems only consider single-dimensional matching (such as style or price), and fail to comprehensively evaluate the multi-dimensional features of designers (such as service response speed, regional coverage, and quality of past cases), making the recommendation results one-sided and lacking in practicality. SUMMARY

[0005] In order to solve the above-mentioned defects, the present application provides a home decoration designer intelligent recommendation method and system.

[0006] The above invention purpose of the present application is realized by the following technical scheme:

[0007] A home decoration designer intelligent recommendation method, comprising the steps of:

[0008] Obtaining user demand information and historical behavior information, and extracting demand features and historical behavior features respectively;

[0009] Obtaining a designer database, and classifying the designer information in the designer database based on a pre-set classification algorithm to form a multi-dimensional feature vector corresponding to the designer information association;

[0010] Inputting the demand features, historical behavior features and multi-dimensional feature vector into the same pre-trained matching model to generate a matching degree matrix;

[0011] When receiving the matching degree matrix output by the matching model, generating a plurality of demand solutions based on a pre-defined loss function and a solving algorithm associated with the loss function;

[0012] Generating a recommendation list based on the generated plurality of demand solutions and sending it to the user end that sent the user demand information.

[0013] By adopting the technical scheme, the user demand information and the historical behavior information are acquired, the corresponding demand features and historical behavior features are extracted, the demand features, the historical behavior features and the multi-dimensional feature vector of the designer formed by the classification algorithm in the designer database are input into the pre-trained matching model to generate a matching degree matrix, and a demand solution of the matching degree matrix is generated based on a loss function and a solving algorithm to generate and output a recommendation list, so that dynamic preference mining of home decoration design and accurate recommendation of the designer combined with multi-dimensional features are realized: the dynamic preference of the user is captured through the historical behavior feature extraction, and the problem of insufficient personalization caused by neglecting the user behavior data in the prior art is solved; meanwhile, the limitation of single-dimensional recommendation is solved through the collaborative matching of the multi-dimensional feature vector of the designer and the demand features, so that the recommendation result not only comprehensively considers the comprehensive ability of the designer, but also meets the actual demand of the user, and has the effect of improving the practicability of the recommendation system and the user satisfaction.

[0014] In a preferred example, the application can be further configured to: the steps of acquiring user demand information and historical behavior information, and extracting demand features and historical behavior features respectively, include steps of:

[0015] identifying identity information corresponding to the user demand information, and acquiring historical behavior information of the identity information;

[0016] inputting the historical behavior information into the pre-trained feature extraction model, so that the feature extraction model extracts and outputs the historical behavior features based on the historical behavior information.

[0017] By adopting the technical scheme, the application realizes mining and conversion of user behavior data by identifying the identity information corresponding to the user demand information and acquiring the historical behavior data, and inputting the historical behavior information into the pre-trained feature extraction model for historical behavior feature extraction: the pre-trained feature extraction model is used to process the historical behavior information, automatically identify the rules in unstructured data such as user browsing records, collection preferences and interactive feedback, reduce subjectivity and inefficiency, not only significantly improve the accuracy and consistency of feature extraction, but also dynamically adapt to the differences in behavior patterns of different users, and provide behavior features with representation for the subsequent matching model.

[0018] In a preferred example, the application can be further configured to: the feature extraction model includes a quantization layer, an association layer, an identification layer and an output layer, the historical behavior information includes browsing behavior, collection behavior, consultation behavior and negative feedback, and the step of inputting the historical behavior information into the pre-trained feature extraction model so that the feature extraction model extracts and outputs the historical behavior features based on the historical behavior information includes steps of:

[0019] the quantization layer performs feature quantization on the browsing behavior and the collection behavior, and outputs a basic feature vector;

[0020] The association layer captures the association behavior between the browsing behavior, the collection behavior, the consultation behavior and the negative feedback, and optimizes the basic feature vector based on the association behavior, and outputs an optimized feature vector;

[0021] The recognition layer generates an implicit behavior vector and an abnormal behavior vector based on the browsing behavior, the collection behavior, the consultation behavior and the negative feedback;

[0022] The output layer generates a historical behavior feature based on the optimized feature vector, the implicit behavior vector and the abnormal behavior vector.

[0023] By adopting the above technical solutions, the application generates a basic feature vector by quantifying the browsing behavior and the collection behavior through the quantification layer, outputs an optimized feature vector by analyzing the association behavior between the browsing behavior, the collection behavior, the consultation behavior and the negative feedback through the association layer, combines the implicit behavior vector and the abnormal behavior vector generated by the recognition layer, and finally fuses and outputs a historical behavior feature by the output layer, so as to realize multi-dimensional modeling of user behavior data; the layered feature extraction architecture is adopted, the quantification layer is used to ensure the numerical expression of the basic behavior feature, the association layer is used to mine the potential relationship between the multiple types of user behaviors to improve the feature representation ability, the recognition layer is used to capture the implicit pattern and abnormal feedback in the user preference, and finally the output layer is used to realize the comprehensive optimization of the features, so that the generated historical behavior feature can not only reflect the explicit behavior preference of the user, but also identify the implicit demand and sensitive problems, and provide more comprehensive and more accurate user portrait data for the subsequent matching model.

[0024] In a preferred example, the application can be further configured to: the step of generating, by the recognition layer, an implicit behavior vector and an abnormal behavior vector based on the browsing behavior, the collection behavior, the consultation behavior and the negative feedback, comprises the steps of:

[0025] The recognition layer performs text analysis on the browsing behavior and the collection behavior, and extracts preference style labels and urgency labels;

[0026] The preference intensity of the extracted preference style labels and the urgency intensity of the urgency labels are calculated;

[0027] The implicit behavior vector is generated based on the preference intensity and the urgency intensity.

[0028] By adopting the technical scheme, the application extracts the preference style label and the urgency label through text analysis of the browsing behavior and the collection behavior by the recognition layer, and calculates the preference intensity and the urgency intensity to generate the implicit behavior vector, so as to realize the mining of the potential design demand and the decision psychology of the user; the semantic information in the user behavior data is analyzed by using the text analysis technology, the preference intensity of the preference style label is quantified, the degree of inclination of the user to a specific design style is accurately captured, the time sensitivity of the user decision is analyzed through the urgency intensity of the urgency label, the generated implicit behavior vector can reflect the aesthetic preference and the project progress demand of the user at the same time, a double reference basis containing the explicit demand and the implicit demand is provided for the designer matching, the understanding ability of the recommendation system to the complex intention of the user is effectively improved, so as to meet the multi-dimensional demand of the user.

[0029] In a preferred example, the application can be further configured to: the step of generating the implicit behavior vector and the abnormal behavior vector based on the browsing behavior, the collection behavior, the consultation behavior and the negative feedback by the recognition layer comprises the steps of:

[0030] extracting the contradictory behavior features in the browsing behavior, the collection behavior, the consultation behavior and the negative feedback based on the pre-set contradiction detection rule;

[0031] extracting the abnormal frequency features in the browsing behavior, the collection behavior, the consultation behavior and the negative feedback based on the pre-set frequency detection rule;

[0032] generating the abnormal behavior vector based on the contradictory behavior features and the abnormal frequency features.

[0033] By adopting the technical scheme, the application extracts the contradictory behavior features in the browsing behavior, the collection behavior, the consultation behavior and the negative feedback based on the pre-set contradiction detection rule, and simultaneously extracts the abnormal frequency features in the browsing behavior, the collection behavior, the consultation behavior and the negative feedback by using the frequency detection rule, finally generates the abnormal behavior vector by fusion, so as to realize the identification of the potential abnormal pattern in the user behavior data; by adopting the double detection mechanism, the inconsistency in the user behavior logic is identified by the contradiction detection rule, and the abnormal operation mode is captured by the frequency detection rule, so that the generated abnormal behavior vector can accurately reflect the hesitation, the contradiction and the special demand in the decision-making process of the user, and provide an important behavior abnormality early warning index for the recommendation system, thereby improving the reliability of the recommendation result and the user satisfaction.

[0034] In a preferred example, the application can be further configured to: the matching model comprises an embedding layer, an interaction layer and an output layer, and the step of inputting the demand features, the historical behavior features and the multi-dimensional feature vector into the same pre-trained matching model to make the matching model generate a matching degree matrix comprises the steps of:

[0035] The embedding layer maps the demand features, the historical behavior features and the multi-dimensional feature vectors to the same embedding space to generate a user embedding vector and a designer embedding vector;

[0036] The interaction layer identifies the interaction correlation between the user embedding vector and the designer embedding vector to generate an interaction feature vector;

[0037] The output layer generates a matching degree matrix based on the interaction feature vector.

[0038] By adopting the above technical solution, the user demand and the designer feature are deeply semantically matched by mapping the demand features, the historical behavior features and the multi-dimensional feature vectors of the designer to the same embedding space through the embedding layer to generate the user embedding vector and the designer embedding vector, analyzing the interaction correlation between the two through the interaction layer to generate the interaction feature vector, and finally outputting the matching degree matrix by the output layer. The embedding space mapping technology is adopted to solve the problem that the multi-source heterogeneous features are difficult to be directly compared. By projecting the features on the user side and the designer side to the unified vector space, the features in different dimensions such as style preference and service ability are comparable. The interaction layer can identify the deep matching relationship that is difficult to be found by the traditional method, such as the matching degree between the implicit demand of the user and the characteristic specialty of the designer, by capturing the potential correlation pattern between the user and the designer features. The matching degree matrix finally outputted not only contains the similarity of the surface features, but also contains the potential synergistic effect of the user and the designer combination, so that the recommendation result can not only meet the explicit demand of the user, but also mine the potential matching possibility that has not been explicitly expressed, thereby significantly improving the intelligent level and the precision of the recommendation system.

[0039] In a preferred example, the application can be further configured to, when the matching degree matrix outputted by the matching model is received, generate a plurality of demand solutions based on a pre-defined loss function and a solving algorithm associated with the loss function, including the steps of:

[0040] matching the matching error loss function and the diversity constraint loss function corresponding to the pre-set based on the matching degree matrix;

[0041] constructing a multi-objective loss function based on the matched matching error loss function and the diversity constraint loss function, and matching the associated solving algorithm;

[0042] generating a set of recommendation solutions based on the multi-objective loss function;

[0043] generating a plurality of demand solutions based on the matched solving algorithm for the set of recommendation solutions.

[0044] By adopting the technical scheme, the application triggers joint optimization of the matching error loss function and the diversity constraint loss function through the matching degree matrix, constructs a multi-objective loss function and generates a recommended solution set, and finally outputs several demand solutions through a solving algorithm to realize intelligent balancing between the accuracy and diversity of the recommended results; the multi-objective optimization framework is adopted, the matching error loss function is used to ensure the matching degree of the recommended results and the user demand, and the diversity constraint loss function is used to avoid the homogenization of the recommended list, so that the generated recommended solution set can accurately match the core demand of the user and cover different types of designer resources; based on the further optimization processing of the solving algorithm on the recommended solution set, the recommendation strategy can be dynamically adjusted according to the actual business demand, the reasonable distribution of the recommended results is realized on the premise of ensuring the recommendation quality, the problem that the traditional recommendation system is easy to fall into local optimum and lead to the singleization of the recommended list is effectively solved, the selection space and satisfaction of the user are improved, and the balanced exposure of the platform designer resources is promoted.

[0045] The second application object is realized by the following technical scheme:

[0046] The application object is realized by the following technical scheme:

[0047] The feature extraction module is used to obtain user demand information and historical behavior information, and extract demand features and historical behavior features, respectively;

[0048] The classification module is used to obtain a designer database, and classify the designer information in the designer database based on a pre-set classification algorithm, so as to form a multi-dimensional feature vector associated with the designer information;

[0049] The input module is used to input the demand features, the historical behavior features and the multi-dimensional feature vector into the same pre-trained matching model, so that the matching degree matrix is generated;

[0050] The demand solution generation module is used to generate several demand solutions based on a pre-defined loss function and a solving algorithm associated with the loss function when receiving the matching degree matrix output by the matching model;

[0051] The recommended list generation module is used to generate a recommended list based on the generated several demand solutions and send the recommended list to the user terminal of the user who sends the user demand information.

[0052] By adopting the technical scheme, the feature extraction module is used to acquire user demand information and historical behavior information, and extract demand features and historical behavior features respectively; the classification module is used to acquire a designer database, and classify designer information in the designer database based on a pre-set classification algorithm, so as to correspondingly form a multi-dimensional feature vector associated with the designer information; the input module is used to input the demand features, the historical behavior features and the multi-dimensional feature vector into a same pre-trained matching model, so that the matching model generates a matching degree matrix; the demand solution generation module is used to generate a plurality of demand solutions based on a pre-defined loss function and a solving algorithm associated with the loss function when receiving the matching degree matrix output by the matching model; and the recommendation list generation module is used to generate a recommendation list based on the plurality of generated demand solutions and send the recommendation list to a user terminal of a user who sends the user demand information.

[0053] In summary, the present application includes at least one of the following beneficial technical effects:

[0054] 1. The present application acquires user demand information and historical behavior information, extracts corresponding demand features and historical behavior features, combines a multi-dimensional feature vector formed by a classification algorithm in a designer database, inputs the demand features, the historical behavior features and the designer multi-dimensional feature vector into a pre-trained matching model to generate a matching degree matrix, generates demand solutions of the matching degree matrix based on a loss function and a solving algorithm, and generates and outputs a recommendation list, so as to realize dynamic preference mining of home decoration design and accurate recommendation of designers combined with multi-dimensional features: the dynamic preference of the user is captured by extracting historical behavior features, the problem of insufficient personalization caused by neglecting user behavior data in the prior art is solved, and at the same time, the limitation of single-dimensional recommendation is solved by the collaborative matching of the designer multi-dimensional feature vector and the demand features, so that the recommendation result not only comprehensively considers the comprehensive ability of the designer, but also meets the actual demand of the user, and has the effect of improving the practicability of the recommendation system and the user satisfaction;

[0055] 2.The application triggers the joint optimization of the matching error loss function and the diversity constraint loss function through the matching degree matrix, constructs a multi-objective loss function and generates a recommended solution set, and finally outputs several demand solutions through a solving algorithm to realize the intelligent balance between the accuracy and diversity of the recommended results; a multi-objective optimization framework is adopted, the matching error loss function is used to ensure the matching degree of the recommended results and the user demand, and the diversity constraint loss function is used to avoid the homogenization of the recommended list, so that the recommended solution set can accurately match the core demand of the user and cover different types of designer resources; based on the further optimization processing of the solving algorithm on the recommended solution set, the recommendation strategy can be dynamically adjusted according to the actual business demand, the reasonable distribution of the recommended results is realized on the premise of ensuring the recommendation quality, the problem that the traditional recommendation system is easy to fall into local optimum and lead to the singleization of the recommended list is effectively solved, the selection space and satisfaction of the user are improved, and the balanced exposure of the platform designer resources is promoted. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a flowchart of an embodiment of the intelligent recommendation method for home decoration designers of the application;

[0057] Figure 2 is an implementation flowchart of step S12 in an embodiment of the intelligent recommendation method for home decoration designers of the application;

[0058] Figure 3 is an implementation flowchart of step S123 in an embodiment of the intelligent recommendation method for home decoration designers of the application;

[0059] Figure 4 is another implementation flowchart of step S123 in an embodiment of the intelligent recommendation method for home decoration designers of the application;

[0060] Figure 5 is an implementation flowchart of step S30 in an embodiment of the intelligent recommendation method for home decoration designers of the application;

[0061] Figure 6 is an implementation flowchart of step S40 in an embodiment of the intelligent recommendation method for home decoration designers of the application. DETAILED DESCRIPTION

[0062] The following will be described in detail in combination with the accompanying drawings Figures 1-6 The application will be further described in detail.

[0063] In an embodiment, as shown in Figure 1 The application discloses an intelligent recommendation method for home decoration designers, which specifically comprises the following steps:

[0064] S10: Obtain user demand information and historical behavior information, and extract demand features and historical behavior features, respectively;

[0065] In this embodiment, the user demand information is the user's expressed demand based on home decoration design, such as including budget, style preference (such as modern, Nordic, Chinese style, etc.), house area, functional demand (such as storage, light optimization, etc.), etc.; the historical behavior information is the operation record behavior information of the user on the platform; the demand feature is the structured data extracted from the user demand information, such as budget range, style label, function label, etc.; the historical behavior feature is the implicit preference mined from the user historical behavior, such as style tendency, decision urgency, sensitivity to designer service quality, etc.

[0066] Specifically, the home decoration demand information (such as budget, style preference, etc.) and the historical behavior information of the user are obtained, and the demand features (explicit demand) and the historical behavior features (implicit preference) are extracted, respectively.

[0067] S20: Obtain a designer database, and classify the designer information in the designer database based on a pre-set classification algorithm, to correspondingly form a multi-dimensional feature vector associated with the designer information;

[0068] In this embodiment, the designer database is a structured data set storing designer information, including the personal profile, case works, service evaluation, response speed, regional coverage, etc. of the designer; the multi-dimensional feature vector is a numerical representation formed after classifying and quantifying the designer information, covering professional skills, service ability, user evaluation, etc.

[0069] Specifically, the designer information (such as case works, service score, response speed, etc.) is read from the designer database, and a classification algorithm (such as clustering, label classification) is used to classify the designers, and the designer information is converted into a multi-dimensional feature vector (such as design style score, service score, response speed, etc.).

[0070] S30: Input the demand feature, the historical behavior feature, and the multi-dimensional feature vector into the same pre-trained matching model, so as to generate a matching degree matrix;

[0071] In this embodiment, the matching model is a pre-trained deep learning model, which is used to calculate the matching degree of the user features and the designer features; the matching degree matrix is the matching score matrix between the user and the designer output by the matching model, reflecting the matching degree of different combinations;

[0072] Specifically, the demand feature, the historical behavior feature of the user and the multi-dimensional feature vector of the designer are input into the matching model, so that the matching model calculates the matching degree of the user and each designer, and outputs the corresponding matching degree matrix, i.e. the user-designer score table.

[0073] S40: When the matching degree matrix output by the matching model is received, a plurality of solutions are generated based on a predefined loss function and a solving algorithm associated with the loss function;

[0074] In this embodiment, the loss function is a predefined optimization objective function for evaluating matching errors and recommendation diversity; the solving algorithm is a mathematical method (such as gradient descent, genetic algorithm, etc.) applied to optimize the loss function and generate a recommended solution; and the solution is a candidate recommended solution generated after optimization;

[0075] Specifically, the matching degree matrix is adjusted by the predefined loss function and the solving algorithm associated with the loss function to generate a plurality of solutions, i.e., candidate recommended solutions.

[0076] S50: Generating a recommendation list based on the generated plurality of solutions and sending it to the user terminal that sent the user demand information;

[0077] In this embodiment, the recommendation list is a designer recommendation ranking list finally presented to the user terminal;

[0078] Specifically, the solutions are sorted or fused to generate a final recommendation list, and the recommendation list is sent to the user terminal for selection or further consultation.

[0079] In an embodiment, step S10 includes the following steps:

[0080] S11: Identifying the identity information corresponding to the user demand information and obtaining the historical behavior information of the identity information;

[0081] S12: Inputting the historical behavior information into the pre-trained feature extraction model to enable the feature extraction model to extract and output historical behavior features based on the historical behavior information.

[0082] In this embodiment, the identity information is a unique identifier of the user, such as a user ID, a mobile phone number, an account, etc., used to associate the user's historical data;

[0083] Specifically, the identity information corresponding to the user demand information is identified and its historical behavior data is obtained, the historical behavior information is input into the pre-trained feature extraction model for historical behavior feature extraction, and the user behavior data is mined and converted.

[0084] In an embodiment, the feature extraction model includes a quantization layer, an association layer, an identification layer, and an output layer, and the historical behavior information includes browsing behavior, collection behavior, consultation behavior, and negative feedback, such as Figure 2 As shown in FIG. 1 1, step S12 includes the following steps:

[0085] S121: The quantization layer quantizes the browsing behavior and the collection behavior to output a basic feature vector;

[0086] S122: The association layer captures the association behavior between the browsing behavior, the collection behavior, the consultation behavior, and the negative feedback, and optimizes the basic feature vector based on the association behavior, and outputs an optimized feature vector;

[0087] S123: The recognition layer generates an implicit behavior vector and an abnormal behavior vector based on the browsing behavior, the collection behavior, the consultation behavior, and the negative feedback;

[0088] S124: The output layer generates a historical behavior feature based on the optimized feature vector, the implicit behavior vector, and the abnormal behavior vector.

[0089] In the embodiment, the quantization layer is a first layer module in the feature extraction model, which is used for basic feature extraction of the browsing behavior and the collection behavior; the basic feature vector is a preliminary feature representation output by the quantization layer, which contains the quantitative statistical results of the user's basic behavior (such as browsing and collection); the association layer is a second layer module in the feature extraction model, which is used for analyzing the potential association relationship between different behavior types; the browsing behavior refers to the operation record of the user viewing the design case or the designer homepage on the platform, including the browsing object, the number of times, the time length, and other dimensional data; the collection behavior refers to the operation record of the user actively saving a specific case or a designer, which reflects the explicit preference selection of the user; the consultation behavior refers to the direct interaction record between the user and the designer, including online consultation, appointment communication, and other interactive behaviors; the negative feedback refers to the behavior record of the user expressing dissatisfaction, including canceling the collection, complaint evaluation, and other negative operations; the optimized feature vector is an enhanced feature representation obtained after the association layer processing, which not only contains the basic behavior statistics, but also integrates the association mode information between behaviors; the recognition layer is a third layer module in the feature extraction model, which is used for identifying deep patterns and abnormal situations from user behaviors; the implicit behavior vector is a feature representation reflecting the potential preferences of the user generated by the recognition layer, which captures the secondary preferences and potential needs that are not explicitly expressed but exist continuously in the user behavior; the abnormal behavior vector is a feature representation reflecting the abnormal behavior of the user generated by the recognition layer, which is used to identify the sudden or contradictory behavior features that do not conform to the user's regular behavior patterns;

[0090] For example, the historical behavior information of a certain user on the platform includes:

[0091] Browsing behavior: viewing 8 Nordic style cases and 3 industrial style cases, and the viewing time is greater than the preset time; Collection behavior: saving the homepage of 2 Nordic style designers; Consultation behavior: contacting 1 designer (Nordic style); Negative feedback: complaining about the consultation once due to "high price".

[0092] Specifically, the browsing behavior and the collection behavior are quantified to generate a basic feature vector by a quantification layer, an optimized feature vector is output by analyzing the association between the browsing behavior, the collection behavior, the consultation behavior and the negative feedback by an association layer, the historical behavior features are finally fused and output by an output layer in combination with the implicit behavior vector and the abnormal behavior vector generated by a recognition layer, so as to realize multi-dimensional modeling of the user behavior data.

[0093] In one embodiment, as shown in FIG. 12, step S123 includes steps of: Figure 3

[0094] S1231A: The recognition layer performs text analysis on the browsing behavior and the collection behavior, and extracts preference style labels and urgency labels;

[0095] S1232A: The preference intensity of the extracted preference style label and the urgency intensity of the urgency label are calculated;

[0096] S1233A: An implicit behavior vector is generated based on the preference intensity and the urgency intensity.

[0097] In the embodiment, the text analysis is a process of natural language processing on the text information (such as case title, designer label, description copy, etc.) in the user browsing and collection content, and extracting key semantic features; the preference style label is a design style category (such as “modern and simple”, “new Chinese style”, “industrial style”, etc.) identified by text analysis, used to quantify the aesthetic preference direction of the user; the urgency label is a time-sensitive feature (such as “urgent”, “as soon as possible”, “before the end of the month”, etc.) extracted from the behavior text, reflecting the timeliness requirement of the user demand; the preference intensity is a quantitative value of the preference degree for a specific style label, calculated based on the frequency, duration, etc. of the style-related behavior; the urgency intensity is a quantitative evaluation value of the time-sensitive demand, generated in combination with the appearance frequency, text emphasis degree, etc. of the urgency label; the implicit behavior vector is a two-dimensional feature representation composed of the preference intensity and the urgency intensity, reflecting the potential demand features of the user not explicitly expressed;

[0098] For example, the browsing behavior of a user on the platform is: “Modern and Simple Three-bedroom” (staying for 8 minutes) and “Urgent! Two-week completion of minimalist apartment” (staying for 11 minutes); the collection behavior is: collecting the homepage of designer Wang (the homepage includes labels: “modern style” and “efficient construction”) and collecting the homepage of designer Li (the homepage includes cases: “48-hour fast installation plan”);

[0099] ​Text analysis process: extract style tags: modern minimalist (2 occurrences) and minimalism (1 occurrence), extract urgency tags: "Hurry! Two weeks to complete" and "48 hours fast installation"; intensity calculation process: preference intensity: modern minimalist: 0.8 (based on 2 views + 1 collection); minimalism: 0.6; urgency intensity: 0.9 (high frequency of time limit words); implicit behavior vector generation: {style preference: {“modern minimalist”: 0.8, “minimalism”: 0.6}, time urgency: 0.9}, which indicates that the user has a strong preference for modern minimalist style and requires high construction efficiency.

[0100] Specifically, by recognizing the layer, the text analysis of the browsing behavior and the collection behavior extracts the preferred style tags and the urgency tags, and calculates the preference intensity and the urgency intensity to generate the implicit behavior vector, thereby realizing the mining of the user's potential design demand and decision-making psychology.

[0101] In an embodiment, as shown in FIG. 12B, step S123 includes steps of: Figure 4

[0102] S1231B: Extracting contradictory behavior features in the browsing behavior, the collection behavior, the consultation behavior and the negative feedback based on the pre-set contradiction detection rule;

[0103] S1232B: Extracting abnormal frequency features in the browsing behavior, the collection behavior, the consultation behavior and the negative feedback based on the pre-set frequency detection rule;

[0104] S1233B: Generating an abnormal behavior vector based on the contradictory behavior features and the abnormal frequency features.

[0105] In the embodiment, the contradiction detection rule is a pre-defined behavior logic conflict judgment standard for identifying self-contradictory patterns in the user behavior, and the contradiction detection rule includes: style preference conflict (such as frequently browsing modern style but collecting classical style), inconsistent behavior (such as canceling immediately after consultation), evaluation contradiction (such as giving a good evaluation but canceling collection), etc.; the contradictory behavior features are quantitative indexes extracted by the contradiction detection rule, reflecting the degree of logical inconsistency in the user behavior, including: style conflict intensity, behavior reversal frequency, evaluation contradiction index, etc.; the frequency detection rule is a pre-set behavior frequency anomaly judgment standard for identifying behavior mutations deviating from the user's regular patterns, and the frequency detection rule detects behavior content including: sudden high-frequency behavior, abnormal behavior interval, abnormal time period distribution, etc.; the abnormal frequency features are quantitative indexes extracted by the frequency detection rule, reflecting the degree of irregularity of the behavior frequency, including: frequency deviation value, behavior density index, time period abnormality degree, etc.; the abnormal behavior vector is a comprehensive abnormal index composed of the contradictory behavior features and the abnormal frequency features, used to represent the abnormal and atypical patterns of the user behavior;

[0106] ​Exemplarily, a user's browsing behavior on the platform: 15 modern minimalist cases, the browsing time is greater than the preset time; collection behavior: collect 3 industrial wind designers' homepages; consulting behavior: consult 2 industrial wind designers; negative feedback: cancel 1 industrial wind designer consultation halfway; historical behavior pattern (compared with the benchmark): past three months average: browse 10 modern minimalist cases per week, collect 1-2 modern minimalist designers, and rarely contact industrial wind; contradiction detection process: mainly browse modern minimalist (15 times) but collect / consult industrial wind (3 times); contradiction characteristics: style conflict value: 0.8 (full score 1), behavior reversal rate: 0.6; frequency detection process finds: industrial wind contact times increase by 300%, and consult in non-active period (midnight); abnormal characteristics: frequency deviation: 0.9, time period abnormal value: 0.7; abnormal behavior vector generation: {contradiction index: 0.8, frequency abnormality: 0.9, comprehensive abnormal value: 0.85}.

[0107] Specifically, the contradiction detection rules are pre-set to extract the contradiction behavior characteristics in the browsing behavior, the collection behavior, the consulting behavior and the negative feedback, and the frequency detection rules are used to extract the abnormal frequency characteristics in the browsing behavior, the collection behavior, the consulting behavior and the negative feedback, and finally the abnormal behavior vector is generated by fusion to realize the identification of the potential abnormal pattern in the user behavior data.

[0108] In an embodiment, the matching model includes an embedding layer, an interaction layer and an output layer, as shown in Figure 5 The step S30 includes the following steps:

[0109] S31: The embedding layer maps the demand characteristics, the historical behavior characteristics and the multi-dimensional feature vector to the same embedding space to generate the user embedding vector and the designer embedding vector;

[0110] S32: The interaction layer identifies the interaction correlation of the user embedding vector and the designer embedding vector to generate the interaction feature vector;

[0111] S33: The output layer generates the matching degree matrix based on the interaction feature vector.

[0112] In the embodiment, the embedding layer is the first layer module of the matching model, used to map heterogeneous features to a unified low-dimensional vector space, and the embedding layer is realized by a nonlinear transformation: dimension unification and semantic alignment; the user embedding vector is a user feature representation output by the embedding layer, generated by fusing demand features and historical behavior features, including explicit demand (such as budget, style preference) and implicit features (such as behavior pattern, potential preference); the designer embedding vector is a designer feature representation output by the embedding layer, converted from a multi-dimensional feature vector, including professional dimensions (style proficiency, case quality) and service dimensions (response speed, price interval) and the like; the interaction layer is the second layer module of the matching model, which mines vector dot product (calculates basic matching degree), cross network (captures nonlinear interaction relationship) and attention mechanism (identifies key matching dimensions) through a deep network; the interaction feature vector is a high-order feature representation output by the interaction layer, including explicit interaction features (such as style matching degree) and implicit association features (such as behavior-case potential association); the output layer is the third layer module of the matching model, which calculates the final matching score through a fully connected network and generates an N x M-dimensional matching degree matrix (N users x M designers); the matching degree matrix is a structured result output by the matching model;

[0113] Specifically, the demand features, historical behavior features and designer multi-dimensional feature vectors are uniformly mapped to the same embedding space through the embedding layer to generate the user embedding vector and the designer embedding vector, the interaction association between the two is analyzed through the interaction layer to generate the interaction feature vector, and finally the matching degree matrix is output by the output layer, realizing the deep semantic matching of user demand and designer features.

[0114] In an embodiment, as shown in Figure 6 Step S40 includes the following steps:

[0115] S41: matching the matching error loss function and the diversity constraint loss function corresponding to the preset based on the matching degree matrix;

[0116] S42: constructing a multi-objective loss function based on the matched matching error loss function and diversity constraint loss function, and matching the associated solving algorithm;

[0117] S43: generating a recommended solution set based on the multi-objective loss function;

[0118] S44: generating a number of demand solutions based on the recommended solution set matched by the solving algorithm.

[0119] In the embodiment, the matching error loss function is a function for evaluating the difference between the predicted matching degree and the ideal matching degree, and is usually a mean square error (measuring the accuracy of score prediction), a cross-entropy loss (processing a binary classification matching problem), or a contrastive loss (enhancing the distinction between positive and negative samples); the diversity constraint loss function is a penalty term for ensuring the diversity of the recommended results, and common types include a category distribution entropy (promoting uniform coverage of dimensions such as style and price), a similarity penalty (suppressing the similarity of designers in the recommended list), and a long-tail compensation (improving the exposure opportunity of unpopular designers); the multi-objective loss function is a composite objective function composed of multiple sub-losses; the solving algorithm is a mathematical method for minimizing the loss function, including gradient descent, evolutionary algorithms, and heuristic algorithms (simulated annealing); the recommended solution set is a set of candidate recommended schemes generated in the solving process, including recommended lists under different weight configurations and schemes with different diversity levels; and the demand solution is a recommended scheme that has passed the final screening.

[0120] Specifically, the matching error loss function and the diversity constraint loss function are jointly optimized through the matching degree matrix, a multi-objective loss function is constructed, and a recommended solution set is generated, and finally a plurality of demand solutions are output through the solving algorithm, so as to realize intelligent balancing between the accuracy and diversity of the recommended results.

[0121] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0122] In an embodiment, a home decoration designer intelligent recommendation system is provided, which corresponds one-to-one to the home decoration designer intelligent recommendation method in the above embodiment. The home decoration designer intelligent recommendation system comprises:

[0123] A feature extraction module is configured to obtain user demand information and historical behavior information, and extract demand features and historical behavior features, respectively;

[0124] A classification module is configured to obtain a designer database, and classify the designer information in the designer database based on a pre-set classification algorithm, so as to form a multi-dimensional feature vector associated with the designer information;

[0125] An input module is configured to input the demand features, the historical behavior features, and the multi-dimensional feature vector into the same pre-trained matching model, so that the matching model generates a matching degree matrix;

[0126] A demand solution generation module is configured to, when receiving the matching degree matrix output by the matching model, generate a plurality of demand solutions based on a pre-defined loss function and a solving algorithm associated with the loss function;

[0127] The recommendation list generation module is configured to generate a recommendation list based on the generated requirements and send the recommendation list to the user terminal of the user who sends the user requirement information.

[0128] The specific limitation of the intelligent recommendation system for a home designer can refer to the limitation of the intelligent recommendation method for a home designer, which will not be described here. Each module in the intelligent recommendation system for a home designer can be realized by software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0129] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not change the essence of the corresponding technical solutions, and should be included in the protection scope of the present application.

Claims

1. A method for intelligent recommendation by home decoration designers, characterized in that: Including the following steps: Obtain user demand information and historical behavior information, and extract demand features and historical behavior features respectively; Identify the identity information corresponding to the user's demand information, and obtain the historical behavior information of that identity information; Historical behavior information is input into a pre-trained feature extraction model, which then extracts and outputs historical behavior features based on the historical behavior information. The feature extraction model includes a quantization layer, a correlation layer, a recognition layer, and an output layer; the historical behavior information includes browsing behavior, collection behavior, consultation behavior, and negative feedback. The quantization layer quantizes the features of browsing and bookmarking behaviors and outputs basic feature vectors. The association layer captures the association behaviors between browsing behavior, collection behavior, consultation behavior and negative feedback, and optimizes the basic feature vector based on the association behaviors to output the optimized feature vector; The recognition layer generates implicit behavior vectors and abnormal behavior vectors based on browsing behavior, collection behavior, consultation behavior, and negative feedback; The recognition layer performs text analysis on browsing and bookmarking behaviors, and extracts preference style tags and urgency tags; Urgency tags are time-sensitive features extracted from behavioral text; Calculate the preference intensity of the extracted preference style tags and the urgency intensity of the urgency tags; Preference intensity is a quantitative value representing the degree of preference for a specific style tag; Urgency intensity is a quantitative assessment value for time-sensitive needs; Generate implicit behavior vectors based on preference intensity and urgency intensity; Based on pre-set contradiction detection rules, contradictory behavioral features are extracted from browsing behavior, collection behavior, inquiry behavior, and negative feedback. Abnormal frequency features are extracted from browsing behavior, collection behavior, inquiry behavior, and negative feedback based on pre-set frequency detection rules; Generate abnormal behavior vectors based on contradictory behavior characteristics and abnormal frequency characteristics; The output layer generates historical behavior features based on optimized feature vectors, latent behavior vectors, and abnormal behavior vectors; Obtain the designer database and classify the designer information in the database based on a pre-set classification algorithm to form a multi-dimensional feature vector that associates the designer information. Input demand features, historical behavior features, and multi-dimensional feature vectors into the same pre-trained matching model to generate a matching degree matrix. The matching model includes an embedding layer, an interaction layer, and an output layer; The embedding layer maps demand features, historical behavior features, and multi-dimensional feature vectors to the same embedding space, generating user embedding vectors and designer embedding vectors. The interaction layer identifies the interaction relationships between user-embedded vectors and designer-embedded vectors, and generates interaction feature vectors. The output layer generates a matching degree matrix based on the interaction feature vectors; When the matching degree matrix output by the matching model is received, several required solutions are generated based on the predefined loss function and the solution algorithm associated with the loss function. Based on the matching degree matrix, a pre-defined matching error loss function and a diversity constraint loss function are used for matching; A multi-objective loss function is constructed based on the matching error loss function and the diversity constraint loss function, and a matching and associated solution algorithm is then applied. Generate a recommended solution set based on a multi-objective loss function; The matching-based solution algorithm generates several required solutions from the recommended solution set; Based on the generated demand solutions, a recommendation list is generated and sent to the user who submitted the demand information.

2. A smart recommendation system for home decoration designers, characterized in that: include: The feature extraction module is used to obtain user demand information and historical behavior information, and extract demand features and historical behavior features respectively. The first module is used to identify the identity information corresponding to the user's demand information and obtain the historical behavior information of that identity information; The second module is used to input historical behavior information into a pre-trained feature extraction model, so that the feature extraction model can extract and output historical behavior features based on the historical behavior information. The third module, the feature extraction model includes a quantization layer, a correlation layer, a recognition layer and an output layer, and the historical behavior information includes browsing behavior, collection behavior, consultation behavior and negative feedback; The fourth module, the quantization layer, performs feature quantization on browsing and bookmarking behaviors, and outputs basic feature vectors. The fifth module, the association layer, captures the association behaviors between browsing behavior, collection behavior, consultation behavior, and negative feedback, and optimizes the basic feature vector based on the association behaviors, outputting an optimized feature vector; The sixth module, the recognition layer, generates implicit behavior vectors and abnormal behavior vectors based on browsing behavior, collection behavior, consultation behavior, and negative feedback. The seventh module, the recognition layer, performs text analysis on browsing and bookmarking behaviors, and extracts preference style tags and urgency tags; The eighth module, urgency labels, are time-sensitive features extracted from behavioral text; The ninth module is used to calculate the preference intensity of the extracted preference style tags and the urgency intensity of the urgency tags; Module 10, Preference Intensity, is a quantitative value representing the degree of preference for a specific style tag; The eleventh module, with urgency intensity defined as a quantitative assessment of time-sensitive requirements; Twelve modules are used to generate implicit behavior vectors based on preference intensity and urgency intensity; Thirteen modules are used to extract contradictory behavioral features from browsing behavior, collection behavior, inquiry behavior, and negative feedback based on pre-set contradiction detection rules; Fourteen modules are used to extract abnormal frequency features from browsing behavior, collection behavior, consultation behavior, and negative feedback based on pre-set frequency detection rules; Fifteen modules are used to generate abnormal behavior vectors based on contradictory behavior characteristics and abnormal frequency characteristics; The sixteen modules, with the output layer generating historical behavior features based on optimized feature vectors, latent behavior vectors, and abnormal behavior vectors; The classification module is used to obtain the designer database and classify the designer information in the database based on a pre-set classification algorithm to form a multi-dimensional feature vector that associates the designer information. The input module is used to input demand features, historical behavior features, and multi-dimensional feature vectors into the same pre-trained matching model to generate a matching degree matrix. The matching model comprises seventeen modules, including an embedding layer, an interaction layer, and an output layer. The eighteen modules, the embedding layer maps demand features, historical behavior features and multi-dimensional feature vectors to the same embedding space, generating user embedding vectors and designer embedding vectors; The nineteenth module, the interaction layer, identifies the interaction relationships between user-embedded vectors and designer-embedded vectors, and generates interaction feature vectors. The twenty modules, with the output layer generating a matching degree matrix based on the interaction feature vectors; The requirement solution generation module is used to generate several requirement solutions based on a predefined loss function and the solution algorithm associated with the loss function when the matching degree matrix output by the matching model is received. Module 21 is used to match a preset matching error loss function and a diversity constraint loss function based on the matching degree matrix. The 22 module is used to construct a multi-objective loss function based on the matched matching error loss function and the diversity constraint loss function, and to match the associated solution algorithm. Modules two and three are used to generate a recommended solution set based on a multi-objective loss function; Module 24 is used to generate several required solutions from the recommended solution set using a matching-based solution algorithm; The recommendation list generation module is used to generate a recommendation list based on several generated demand solutions and send it to the user terminal that sent the user demand information.

Images