Commodity personalized recommendation method based on intelligent property platform

Abstract

The invention discloses a commodity personalized recommendation method based on an intelligent property platform, and relates to the technical field of intelligent property management, and the method comprises the following steps: obtaining a relation recommendation score in advance, obtaining a depth recommendation score, obtaining a scene adaptation score, obtaining a comprehensive recommendation score, and generating a personalized recommendation list according to the comprehensive recommendation score, and the recommendation result is displayed at the applet end. According to the invention, property specific data such as enterprise welfare charge rules, community activity types, user repair records and the like are incorporated into a recommendation model, a welfare scene adaptation score calculation module, a community activity scene adaptation score calculation module and a life service scene adaptation score calculation module are constructed, and deep coupling of recommendation strategies and community life scenes is realized.

Description

Technical Field

[0001] The present invention relates to the field of smart property management technology, and in particular to a method for personalized product recommendation based on a smart property platform. Background Art

[0002] With the development of smart property platforms, online shopping mall services have become an important component. Currently, smart property platforms have transformed from traditional property management models to integrated, intelligent service platforms, covering diverse scenarios such as community security, repair services, neighborhood social interaction, and online shopping. When owners use smart property platforms, their consumption behavior is deeply tied to community life scenarios. For example, they are more inclined to purchase products related to home repairs, community activities, and property welfare benefits, and their consumption decisions are often influenced by factors such as neighbor recommendations and property activity rules. However, traditional recommendation algorithms such as collaborative filtering and content-based recommendations are mostly designed based on general e-commerce scenarios. They cannot effectively capture the unique attributes of owners in smart property scenarios, such as life service preferences, community welfare consumption habits, and neighborhood interactions. This leads to deviations between recommendation results and actual user needs.

[0003] In existing technical solutions, some recommendation systems attempt to match basic user attributes with product categories, but lack consideration of the dynamic characteristics of community scenarios, such as temporary property activities, seasonal changes in service demand, and adjustments to welfare fee usage rules. At the same time, traditional algorithms are prone to reduced recommendation accuracy when processing small sample data unique to the community. For example, the behavioral data of new owners is sparse, or there are insufficient interaction records for niche community services and products, which makes it difficult for the recommendation model to accurately portray user preferences. In addition, user behavior data in smart property platforms has multi-source characteristics, including online shopping records, offline service appointments, and participation in community activities. Traditional algorithms find it difficult to effectively integrate multi-source heterogeneous data and cannot fully capture the potential needs of users in community life, which leads to insufficient recommendation accuracy and cannot effectively meet the personalized recommendation needs of smart property platforms. Therefore, there is an urgent need for a personalized product recommendation method that can adapt to smart property scenarios to solve the adaptability problem of traditional algorithms in community scenarios.

[0004] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0005] In response to the problems in the related art, the present invention proposes a personalized product recommendation method based on a smart property platform to overcome the above-mentioned technical problems existing in the existing related art.

[0006] The technical solution of the present invention is achieved as follows:

[0007] A method for personalized product recommendation based on a smart property platform, comprising:

[0008] Pre-obtain relationship recommendation score CF (u,i), at least including: obtaining the user-based recommendation score score YH (u,i) and get the item-based recommendation score score WP (u,i), and calculate the relationship recommendation score score CF (u,i), expressed as:

[0009] score CF (u,i)=α·score YH (u,i)+(1-α)·score WP (u,i);

[0010] Get the deep recommendation score DL (u,i), including: building a multi-layer neural network model, taking user characteristics and product characteristics as input, and predicting the user's preference score for the product;

[0011] Get the scene adaptation score CJ (u,i), at least including: calculating the adaptation score of product i in the welfare scenario based on the enterprise welfare fee quota and usage rules, calculating the adaptation score of product i in the community activity scenario based on the activity type and time, and calculating the adaptation score of product i in the life service scenario based on the user's repair record and housekeeping needs, and calculating the scenario adaptation score CJ (u,i), expressed as:

[0012] score CJ (u,i)=score FL (i)+score HD (i)+score FW (i);

[0013] Get the comprehensive recommendation score score(u,i), expressed as:

[0014] score(u,i)=ω1·score CF (u,i)+ω2·score DL (u,i)+ω3·score CJ (u,i);

[0015] Among them, score CF (u,i) is the relationship recommendation score, score DL (u,i) is the depth recommendation score, score CJ (u,i) is the scene adaptation score, ω1, ω2, ω3 are weight coefficients;

[0016] Generate a personalized recommendation list based on the comprehensive recommendation score score(u,i) and display the recommendation results on the mini program.

[0017] The following steps are also included:

[0018] According to the user's real-time behavior, the real-time recommendation model is used to update the recommendation results, which can be expressed as:

[0019] score 实时 (u,i)=(1-θ)·score(u,i)+θ·score 实时行为 (u,i);

[0020] Among them, score 实时 (u,i) is the recommendation score after real-time update, score 实时行为 (u,i) is the recommendation score based on the user’s real-time behavior, and θ is the real-time update coefficient.

[0021] Wherein, the recommendation score based on the user is obtained YH (u,i), including the following steps:

[0022] Constructing the user similarity matrix includes calculating the similarity between users, which is expressed as:

[0023]

[0024] Among them, X u and X v are the behavioral feature vectors of users u and v respectively; ||·|| represents the L2 norm of the vector;

[0025] Screening the neighbor user set includes: sorting by similarity, selecting K users most similar to the target user u, and forming a neighbor set S(u,K);

[0026] Calculate the recommendation score, including: for item i, if the neighboring user v∈S(u,K) has interacted with i, then calculate the recommendation score by weighted average, expressed as:

[0027]

[0028] Where N(i) represents the set of users who have interacted with item i; r v,i is the rating of user v for item i.

[0029] Wherein, the recommendation score based on the item is obtained WP (u,i), including the following steps:

[0030] Constructing the item similarity matrix, including calculating the similarity between users, is expressed as:

[0031]

[0032] Among them, N(i) and N(j) are the sets of users who have interacted with items i and j respectively;

[0033] Screening the similar item set includes: selecting M items that are most similar to the target item i to form a similarity set S(i,M);

[0034] Calculating the recommendation score includes: weighted calculation of the preference for item i based on the historical interaction item set B(u) of user u, expressed as:

[0035]

[0036] Among them, r u,j is the rating of user u on historical item j.

[0037] Wherein, the depth recommendation score is obtained DL (u,i), including the following steps:

[0038] Pre-calibrate the embedding representation of users and items, including mapping the original features of user u and item i into low-dimensional dense vectors, represented as:

[0039] User embed:p u =Embedding u (u)∈R k ;

[0040] Item embed: q i =Embeddingi(i)∈R k ;

[0041] Among them, Embedding u (u) and Embeddingi(i) are trainable embedding matrices, and k is the embedding dimension;

[0042] A multi-layer perceptron (MLP) is used to capture nonlinear interactions, including concatenating the embedding vectors of users and items and inputting them into the MLP, which can be expressed as:

[0043] z l =[p u ;q i ]∈R 2k ;

[0044] Z l+1 =σ(W l Z l +b l ),l=1,2,...,L;

[0045] Among them, W l and b l is the weight matrix and bias vector of the lth layer; σ is the activation function; L is the number of layers of the MLP;

[0046] Performing attention mechanism weighted historical behavior includes the following steps:

[0047] Assume that the historical interaction item set of user u is

[0048]

[0049] Calculate for each historical item i t The association weight with the current item i is expressed as:

[0050]

[0051] Among them, w a is the attention weight matrix;

[0052] Generating attention-enhanced representations of users

[0053]

[0054] The prediction score is made and the output of MLP is combined with the attention-enhanced user representation to obtain the deep recommendation score, which is expressed as:

[0055] score DL (u,i)=h T [z L ;p u ′]+b;

[0056] Among them, h is the output layer weight vector; b is the bias term; [z L ;p u ′] represents the concatenation of the final output of MLP and the attention enhancement vector.

[0057] The adaptability score of product i in the welfare scenario is calculated based on the enterprise welfare fund quota and usage rules, and is expressed as:

[0058] score FL (i) = β1·f KY (i)+β2·f ZK (i)+β3·f YE (i)

[0059] Among them, score FL (i) is the welfare scenario adaptation score, f KY (i) indicates whether commodity i can be purchased with welfare funds (0 or 1), f ZK (i) represents the discount strength of product i in the welfare scenario, fYE (i) represents the coverage of user u’s welfare balance to commodity i, and β1, β2, and β3 are weight coefficients.

[0060] The adaptability score of product i in the community activity scenario is calculated based on the activity type and time, and is expressed as:

[0061] score HD (i) = γ1·a LX (i)+γ2·a SJ (i);

[0062] Among them, score HD (i) is the community activity scene adaptation score, a LX (i) represents the matching degree between product i and activity type, a SJ (i) represents the compatibility between product i and activity time, and γ1 and γ2 are weight coefficients.

[0063] The adaptability score of product i in the life service scenario is calculated based on the user's repair record and housekeeping needs, and is expressed as:

[0064] score FW (i) = δ1·s LX (i)+δ2·s PL (u,i);

[0065] Among them, score FW (i) is the adaptation score for life service scenarios, s LX (i) represents the matching degree between product i and service type, s PL (u,i) represents the frequency of user u’s demand for related services, and δ1 and δ2 are weight coefficients.

[0066] Beneficial effects of the present invention:

[0067] 1. The present invention implements multi-dimensional scenario modeling to significantly improve recommendation accuracy and user experience. By incorporating property-specific data such as corporate welfare fee rules, community activity types, and user repair records into the recommendation model, a calculation module for welfare scenario adaptation scores, community activity scenario adaptation scores, and life service scenario adaptation scores is constructed to achieve deep coupling of recommendation strategies with community life scenarios. The priority of product recommendations can be dynamically adjusted according to the user's welfare balance, which greatly improves the conversion rate of products in welfare scenarios and effectively solves the "scenario disconnection" problem of traditional algorithms. At the same time, the weight of users' recent behavior is enhanced through the attention mechanism. Even if the data of new owners is sparse, suitable products can be quickly located based on repair records, which greatly improves the accuracy of recommendations for new users and significantly improves the recommendation effect in cold start scenarios.

[0068] 2. The hybrid recommendation architecture and real-time dynamic update mechanism of the present invention break through the bottleneck of traditional technology. By constructing a three-dimensional association model of "user-item-scenario", integrating collaborative filtering, deep neural networks and scenario adaptation scores, a multi-level recommendation system is formed. In the scenario of recommending niche products in the community, the accuracy is greatly improved compared with the single collaborative filtering algorithm, and the recall rate of long-tail product recommendations is significantly optimized. The real-time recommendation model can respond to user browsing, adding to purchases and other behaviors in milliseconds. In addition, the model captures neighborhood consumption preferences through the user similarity matrix, and combines personal welfare balance, activity participation history and other characteristics to form a recommendation logic of "community commonality + personal characteristics", so that the recommended content is in line with the community living habits and meets the needs of individual differences, and user satisfaction is greatly improved. This solution not only increases the average monthly transaction volume for the property platform online mall, but also promotes the digital transformation of smart properties from basic management to personalized life services. It has significant commercial value and industry demonstration significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0070] Figure 1 The present invention is a flowchart of a method for personalized product recommendation based on a smart property platform according to an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention are within the scope of protection of the present invention.

[0072] According to an embodiment of the present invention, a method for personalized product recommendation based on a smart property platform is provided.

[0073] like Figure 1 As shown, the personalized product recommendation method based on the smart property platform according to an embodiment of the present invention includes the following steps:

[0074] Step S1: Pre-acquire the relationship recommendation score CF (u,i), including the following steps:

[0075] Step S101, pre-acquiring a recommendation score based on the user, includes the following steps:

[0076] Construct a user similarity matrix, including using cosine similarity to calculate the similarity between users, expressed as:

[0077]

[0078] Among them, X u and X v are the behavioral feature vectors of users u and v respectively (such as one-hot encoding or embedding vectors of browsing and purchase history); ||·|| represents the L2 norm of the vector.

[0079] Screening the neighbor user set includes: sorting by similarity, selecting K users most similar to the target user u, and forming a neighbor set S(u,K);

[0080] Calculate the recommendation score, including: for item i, if the neighboring user v∈S(u,K) has interacted with i, then calculate the recommendation score by weighted average, expressed as:

[0081]

[0082] Where N(i) represents the set of users who have interacted with item i; r v,i is the rating of user v for item i.

[0083] Step S102, obtaining a recommendation score based on the item, includes the following steps:

[0084] Construct the item similarity matrix, including the improved cosine similarity, which is expressed as:

[0085]

[0086] Among them, N(i) and N(j) are the sets of users who have interacted with items i and j respectively; the logarithmic term in the denominator is used to penalize popular items to avoid "popular preferences" dominating recommendations.

[0087] Screening the similar item set includes: selecting M items that are most similar to the target item i to form a similarity set S(i,M);

[0088] Calculating the recommendation score includes: weighted calculation of the preference for item i based on the historical interaction item set B(u) of user u, expressed as:

[0089]

[0090] Among them, r u,j is the rating of user u on historical item j.

[0091] Step S103: Calculate the relationship recommendation score CF (u,i), expressed as:

[0092] score CF (u,i)=α·score YH (u,i)+(1-α)·score WP (u,i);

[0093] Among them, α∈[0,1] is the weight coefficient, and the optimal value is determined by cross-validation.

[0094] Step S2: Get the depth recommendation score DL (u,i), including: building a multi-layer neural network model, taking user features and product features as input, and mapping them into a low-dimensional dense vector p through the embedding layer u and q i , then capture the nonlinear relationship through the multi-layer perceptron (MLP), and finally predict the user's preference score for the product through the output layer as the deep recommendation score DL (u,i). Specifically includes the following steps:

[0095] Pre-calibrate the embedding representation of users and items, including mapping the original features of user u and item i into low-dimensional dense vectors, represented as:

[0096] User embed:p u =Embedding u (u)∈R k ;

[0097] Item embed: q i =Embeddingi(i)∈R k ;

[0098] Among them, Embedding u (u) and Embeddingi(i) are trainable embedding matrices, and k is the embedding dimension.

[0099] A multi-layer perceptron (MLP) is used to capture nonlinear interactions, which involves concatenating the embedding vectors of users and items and inputting them into the MLP, expressed as:

[0100] z l =[p u ;q i ]∈R 2k ;

[0101] Z l+1 =σ(W l Z l +b l ),l=1,2,…,L;

[0102] Among them, W l and bl are the weight matrix and bias vector of layer l; σ is the activation function; and L is the number of layers in the MLP.

[0103] Perform an attention mechanism to weight historical behaviors. This introduces an attention mechanism to weight the user's historical interaction items and enhance the impact of recent behaviors. This includes the following steps:

[0104] Assume that the historical interaction item set of user u is

[0105]

[0106] Calculate for each historical item i t The association weight with the current item i is expressed as:

[0107]

[0108] Among them, w a is the attention weight matrix.

[0109] Generating attention-enhanced representations of users

[0110]

[0111] The prediction score is made and the output of MLP is combined with the attention-enhanced user representation to obtain the deep recommendation score, which is expressed as:

[0112] score DL (u,i)=h T [z L ;p u ′]+b;

[0113] Among them, h is the output layer weight vector; b is the bias term; [z L ;p u ′] represents the concatenation of the final output of MLP and the attention enhancement vector.

[0114] Step S3: Get the scene adaptation score CJ (u,i), including the following steps:

[0115] Step S301: Calculate the suitability score of product i in the welfare scenario based on the enterprise welfare fund quota and usage rules, expressed as:

[0116] score FL (i) = β1·f KY (i)+β2·f ZK (i)+β3·f YE (i)

[0117] Among them, score FL(i) is the welfare scenario adaptation score, f KY (i) indicates whether commodity i can be purchased with welfare funds (0 or 1), f ZK (i) represents the discount strength of product i in the welfare scenario, f YE (i) represents the coverage of user u’s welfare balance to commodity i, and β1, β2, and β3 are weight coefficients.

[0118] Step S302: Calculate the suitability score of product i in the community activity scenario based on the activity type and time, expressed as:

[0119] score HD (i) = γ1·a LX (i)+γ2·a SJ (i);

[0120] Among them, score HD (i) is the community activity scene adaptation score, a LX (i) represents the matching degree between product i and activity type, a SJ (i) represents the compatibility between product i and activity time, and γ1 and γ2 are weight coefficients.

[0121] Step S303: Calculate the suitability score of product i in the life service scenario based on the user's repair record and housekeeping needs, expressed as:

[0122] score FW (i) = δ1·s LX (i)+δ2·s PL (u,i);

[0123] Among them, score FW (i) is the adaptation score for life service scenarios, s LX (i) represents the matching degree between product i and service type, s PL (u,i) represents the frequency of user u’s demand for related services, and δ1 and δ2 are weight coefficients.

[0124] Step S304: Obtain scene adaptation score CJ (u,i), expressed as:

[0125] score CJ (u,i)=score FL (i)+score HD (i)+score FW (i).

[0126] Step S4, obtain the comprehensive recommendation score score(u,i), expressed as:

[0127] score(u,i)=ω1·score CF (u,i)+ω2·score DL (u,i)+ω3·score CJ (u,i);

[0128] Among them, score CF (u,i) is the relationship recommendation score, score DL (u,i) is the depth recommendation score, score CJ (u,i) is the scene adaptation score, ω1, ω2, ω3 are weight coefficients, satisfying ω1+ω2+ω3=1,

[0129] Step S5: Generate a personalized recommendation list based on the comprehensive recommendation score score(u,i) and display the recommendation results on the mini program side.

[0130] Step S6: Based on the user's real-time behavior, the real-time recommendation model is used to update the recommendation results. The update formula of the model is expressed as:

[0131] score 实时 (u,i)=(1-θ)·score(u,i)+θ·score 实时行为 (u,i);

[0132] Among them, score 实时 (u,i) is the recommendation score after real-time update, score 实时行为 (u,i) is the recommendation score based on the user’s real-time behavior, and θ is the real-time update coefficient.

[0133] In summary, with the help of the above technical solution of the present invention, the following effects can be achieved:

[0134] 1. The present invention implements multi-dimensional scenario modeling to significantly improve recommendation accuracy and user experience. By incorporating property-specific data such as corporate welfare fee rules, community activity types, and user repair records into the recommendation model, a calculation module for welfare scenario adaptation scores, community activity scenario adaptation scores, and life service scenario adaptation scores is constructed to achieve deep coupling of recommendation strategies with community life scenarios. The priority of product recommendations can be dynamically adjusted according to the user's welfare balance, which greatly improves the conversion rate of products in welfare scenarios and effectively solves the "scenario disconnection" problem of traditional algorithms. At the same time, the weight of users' recent behavior is enhanced through the attention mechanism. Even if the data of new owners is sparse, suitable products can be quickly located based on repair records, which greatly improves the accuracy of recommendations for new users and significantly improves the recommendation effect in cold start scenarios.

[0135] 2. The hybrid recommendation architecture and real-time dynamic update mechanism of the present invention break through the bottleneck of traditional technology. By constructing a three-dimensional association model of "user-item-scenario", integrating collaborative filtering, deep neural networks and scenario adaptation scores, a multi-level recommendation system is formed. In the scenario of recommending niche products in the community, the accuracy is greatly improved compared with the single collaborative filtering algorithm, and the recall rate of long-tail product recommendations is significantly optimized. The real-time recommendation model can respond to user browsing, adding to purchases and other behaviors in milliseconds. In addition, the model captures neighborhood consumption preferences through the user similarity matrix, and combines personal welfare balance, activity participation history and other characteristics to form a recommendation logic of "community commonality + personal characteristics", so that the recommended content is in line with the community living habits and meets the needs of individual differences, and user satisfaction is greatly improved. This solution not only increases the average monthly transaction volume for the property platform online mall, but also promotes the digital transformation of smart properties from basic management to personalized life services. It has significant commercial value and industry demonstration significance.

[0136] The foregoing is merely a preferred embodiment of the present invention and is not intended to limit the present invention. A person skilled in the art will readily appreciate other embodiments of the present invention after considering the disclosure in the specification and examples. This application is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered merely exemplary, and the true scope and spirit of the present invention are indicated by the claims.

[0137] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A personalized product recommendation method based on a smart property platform, characterized in that: The following steps are involved: Pre-obtain relationship recommendation score CF (u,i), at least including: obtaining the user-based recommendation score score YH (u,i) and get the item-based recommendation score score WP (u,i), and calculate the relationship recommendation score score CF (u,i), expressed as: score CF (u,i)=α·score YH (u,i)+(1-α)·score WP (u,i); Get the deep recommendation score DL (u,i), including: building a multi-layer neural network model, taking user characteristics and product characteristics as input, and predicting the user's preference score for the product; Get the scene adaptation score CJ (u,i), at least including: calculating the adaptation score of product i in the welfare scenario based on the enterprise welfare fee quota and usage rules, calculating the adaptation score of product i in the community activity scenario based on the activity type and time, and calculating the adaptation score of product i in the life service scenario based on the user's repair record and housekeeping needs, and calculating the scenario adaptation score CJ (u,i), expressed as: score CJ (u,i)=score FL (i)+score HD (i)+score FW (i); Get the comprehensive recommendation score score(u,i), expressed as: score(u,i)=ω1·score CF (u,i)+ω2·score DL (u,i)+ω3·score CJ (u,i); Among them, score CF (u,i) is the relationship recommendation score, score DL (u,i) is the depth recommendation score, score CJ (u,i) is the scene adaptation score, ω1, ω2, ω3 are weight coefficients; Generate a personalized recommendation list based on the comprehensive recommendation score score(u,i) and display the recommendation results on the mini program.

2. The method for personalized product recommendation based on the smart property platform according to claim 1, characterized in that: The following steps are also included: According to the user's real-time behavior, the real-time recommendation model is used to update the recommendation results, which can be expressed as: score 实时 (u,i)=(1-θ)·score(u,i)+θ·score 实时行为 (u,i); Among them, score 实时 (u,i) is the recommendation score after real-time update, score 实时行为 (u,i) is the recommendation score based on the user’s real-time behavior, and θ is the real-time update coefficient.

3. The method for personalized product recommendation based on the smart property platform according to claim 1, characterized in that: The obtained recommendation score based on the user YH (u,i), including the following steps: Constructing the user similarity matrix includes calculating the similarity between users, which is expressed as: Among them, X u and X v are the behavioral feature vectors of users u and v respectively; ||·|| represents the L2 norm of the vector; Screening the neighbor user set includes: sorting by similarity, selecting K users most similar to the target user u, and forming a neighbor set S(u,K); Calculate the recommendation score, including: for item i, if the neighboring user v∈S(u,K) has interacted with i, then calculate the recommendation score by weighted average, expressed as: Where N(i) represents the set of users who have interacted with item i; r v,i is the rating of user v for item i.

4. The method for personalized product recommendation based on the smart property platform according to claim 3, characterized in that: The obtained recommendation score based on the item WP (u,i), including the following steps: Constructing the item similarity matrix, including calculating the similarity between users, is expressed as: Among them, N(i) and N(j) are the sets of users who have interacted with items i and j respectively; Screening the similar item set includes: selecting M items that are most similar to the target item i to form a similarity set S(i,M); Calculating the recommendation score includes: weighted calculation of the preference for item i based on the historical interaction item set B(u) of user u, expressed as: Among them, r u,j is the rating of user u on historical item j.

5. The method for personalized product recommendation based on a smart property platform according to claim 1, characterized in that: The depth recommendation score is obtained DL (u,i), including the following steps: Pre-calibrate the embedding representation of users and items, including mapping the original features of user u and item i into low-dimensional dense vectors, represented as: User embed:p u =Embedding u (u)∈R k ; Item embed: q i =Embeddingi(i)∈R k ; Among them, Embedding u (u) and Embeddingi(i) are trainable embedding matrices, and k is the embedding dimension; A multi-layer perceptron (MLP) is used to capture nonlinear interactions, including concatenating the embedding vectors of users and items and inputting them into the MLP, which can be expressed as: z l =[p u ;q i ]∈R 2k ; WITH l+1 =σ(W l WITH l +b l ),l=1,2,…,L; Among them, W l and b l is the weight matrix and bias vector of the lth layer; σ is the activation function; L is the number of layers of the MLP; Performing attention mechanism weighted historical behavior includes the following steps: Assume that the historical interaction item set of user u is Calculate for each historical item i t The association weight with the current item i is expressed as: Among them, w a is the attention weight matrix; Generating attention-enhanced representations of users The prediction score is made and the output of MLP is combined with the attention-enhanced user representation to obtain the deep recommendation score, which is expressed as: score DL (u,i)=h T [z L ;p u ′]+b; Among them, h is the output layer weight vector; b is the bias term; [z L ;p u ′] represents the concatenation of the final output of MLP and the attention enhancement vector.

6. The method for personalized product recommendation based on a smart property platform according to claim 1, characterized in that: According to the enterprise welfare fund quota and usage rules, the adaptation score of product i in the welfare scenario is calculated and expressed as: score FL (i)=β1·f KY (i)+β2·f ZK (i)+β3·f YE (i) Among them, score FL (i) is the welfare scenario adaptation score, f KY (i) indicates whether commodity i can be purchased with welfare funds (0 or 1), f ZK (i) represents the discount strength of product i in the welfare scenario, f YE (i) represents the coverage of user u’s welfare balance to commodity i, and β1, β2, and β3 are weight coefficients.

7. The method for personalized product recommendation based on a smart property platform according to claim 6, characterized in that: The adaptation score of product i in the community activity scenario is calculated based on the activity type and time, which is expressed as: score HD (i)=γ1·a LX (i)+γ2·a SJ (i); Among them, score HD (i) is the community activity scene adaptation score, a LX (i) represents the matching degree between product i and activity type, a SJ (i) represents the compatibility between product i and activity time, and γ1 and γ2 are weight coefficients.

8. The method for personalized product recommendation based on a smart property platform according to claim 7, characterized in that: According to the user's repair record and housekeeping needs, the adaptation score of product i in the life service scenario is calculated and expressed as: score FW (i)=δ1·s LX (i)+δ2·s PL (u,i); Among them, score FW (i) is the adaptation score for life service scenarios, s LX (i) represents the matching degree between product i and service type, s PL (u,i) represents the frequency of user u’s demand for related services, and δ1 and δ2 are weight coefficients.

Images