House renting recommendation method and system based on multi-modal data fusion analysis

Abstract

The application discloses a house renting recommendation method and system based on multi-modal data fusion analysis, and relates to the technical field of big data analysis.The method comprises the following steps: in the case that a user trigger operation for putting materials on a house renting platform is detected, obtaining the put channel information, user geographic position and put material attribute data corresponding to the new user flow; inputting the put channel information and put material attribute data into a house renting demand prediction model to construct a house renting demand matrix corresponding to the new user flow; screening a candidate house source set from a house source database according to the user geographic position; for each candidate house source, calculating a matching degree relative to the house renting demand matrix according to the house source description data of the candidate house source and the corresponding travel distance; and generating a house source recommendation list for the new user flow according to a preset number of candidate house sources ranked in the front according to the matching degree. Thus, through the multi-modal data fusion mode, accurate house renting house source recommendation in the cold start case is realized.

Description

Technical Field

[0001] The present application relates to the field of big data analysis technology, and in particular to a rental recommendation method and system based on multimodal data fusion analysis. Background Art

[0002] With the continuous development of Internet technology and big data applications, rental recommendation systems have gradually become an indispensable tool in modern urban life, improving the rental experience and efficiency by providing users with housing information that meets their needs.

[0003] The current rental recommendation system mainly adopts a recommendation method based on user portraits. By collecting users' personal information (such as requirements for houses, such as house type, environment preferences, rental preferences, etc.) and users' historical behavior data, it builds a user "portrait" model and makes personalized recommendations based on the portrait.

[0004] However, the cold start problem is a significant drawback of these rental recommendation systems. Specifically, when the system lacks sufficient user behavior data, it cannot effectively make personalized recommendations. In particular, the rental preferences of newly registered users are not yet fully understood, resulting in the recommendation system being unable to quickly adapt to the needs of new users. Similarly, when new listings are added to the system, the lack of historical user interaction data often prevents the system from accurately predicting the new listing's potential user base, leading to poor recommendation results.

[0005] To address the above issues, the industry has not yet proposed a better technical solution. Summary of the Invention

[0006] The present application provides a rental recommendation method, system, storage medium, computer program product and electronic device based on multimodal data fusion analysis, which is used to at least solve the cold start problem of traditional rental recommendation systems when attracting new users, resulting in low relevance of recommendation results due to the lack of user historical behavior data.

[0007] In a first aspect, an embodiment of the present application provides a rental recommendation method based on multimodal data fusion analysis, comprising: upon detecting a user triggering operation to place materials on a rental platform, obtaining the placement channel information, user geographic location, and placement material attribute data of the corresponding new user; the placement material attribute data includes a material preference group type and a material property description information; inputting the placement channel information and the placement material attribute data into a rental demand prediction model to output multiple rental demand types and corresponding prediction confidences, thereby constructing a rental demand matrix corresponding to the new user; each row of the rental demand matrix represents a rental demand type, and each column represents a corresponding prediction confidence; screening a set of candidate properties from a property library based on the user's geographic location, wherein the travel distance between the property location indicated by the candidate property and the user's geographic location is less than a preset threshold; for each candidate property, calculating the matching degree relative to the rental demand matrix based on the property description data and the corresponding travel distance of the candidate property; and generating a property recommendation list for the new user based on a preset number of candidate properties ranked top by the corresponding matching degree.

[0008] In a second aspect, an embodiment of the present application provides a house rental recommendation system based on multimodal data fusion analysis, comprising: a traffic data acquisition unit for acquiring, upon detecting a user triggering operation for placing materials on a house rental platform, the delivery channel information, user geographic location, and delivery material attribute data corresponding to the new user being attracted; the delivery material attribute data includes material preference group type and material house description information; a house rental demand prediction unit for inputting the delivery channel information and the delivery material attribute data into a house rental demand prediction model to output a plurality of house rental demand types and corresponding prediction confidences, thereby constructing a house rental demand matrix corresponding to the new user being attracted; the Each row of the rental demand matrix represents a type of rental demand, and each column represents a corresponding prediction confidence; a candidate housing screening unit is used to screen a set of candidate housing sources from a housing library according to the user's geographic location, wherein the travel distance between the geographic location of the candidate housing source and the geographic location of the user is less than a preset threshold; a housing matching calculation unit is used to calculate the matching degree relative to the rental demand matrix for each candidate housing source based on the housing description data of the candidate housing source and the corresponding travel distance; a drainage housing recommendation unit is used to generate a housing recommendation list for the drainage new user based on a preset number of candidate housing sources ranked high in corresponding matching degrees.

[0009] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the steps of the rental recommendation method based on multimodal data fusion analysis of any embodiment of the present application.

[0010] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the rental recommendation method based on multimodal data fusion analysis of any embodiment of the present application are implemented.

[0011] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the rental recommendation method based on multimodal data fusion analysis of any embodiment of the present application.

[0012] The housing rental recommendation method and system based on multimodal data fusion analysis provided by this application can produce at least the following technical effects:

[0013] (1) By introducing multimodal data such as delivery channel information and lead material attributes (including material preference group type and house description information), an initial rental demand matrix is ​​constructed for newly registered users. By combining the attribute information of delivery materials with the delivery channel information of lead users and applying model prediction technology, the system can preliminarily predict the type of rental demand and its confidence level of new users based on the preferences of the target group. Even without the user's historical interaction data, a recommendation list that meets their potential preferences can be generated, thereby effectively alleviating the cold start problem of new users.

[0014] (2) The system filters the candidate housing set based on the user's geographic location to ensure that the recommended results meet the user's travel convenience needs. In addition, the system matches the property description data of each candidate property with the user's demand matrix and sorts the candidate properties based on the matching calculation results, thereby ensuring that the recommended properties are highly consistent with the user's needs, which can significantly improve the relevance and personalization of the recommended properties.

[0015] This technical solution analyzes new user traffic information, predicts their rental needs based on the distribution channels and properties of the corresponding traffic materials, and quickly matches them with candidate properties within a short travel distance. This generates highly accurate recommended properties without requiring historical user interaction data. This multimodal data fusion enables accurate rental recommendations even in cold-start scenarios, effectively improving the recommendation system's cold-start performance and ensuring that recommendations closely align with new users' needs, optimizing the user experience on the rental platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0017] Figure 1 A flowchart of an example of a rental recommendation method based on multimodal data fusion analysis according to an embodiment of the present application is shown;

[0018] Figure 2 A schematic diagram showing the structural connection of an example of a generator in the CGAN architecture of the rental demand prediction model according to an embodiment of the present application is shown;

[0019] Figure 3 An operational flow chart of an example of matching degree calculation according to an embodiment of the present application is shown;

[0020] Figure 4 A structural block diagram of an example of a rental recommendation system based on multimodal data fusion analysis according to an embodiment of the present application is shown;

[0021] Figure 5 This is a schematic structural diagram of an embodiment of an electronic device of the present application. DETAILED DESCRIPTION

[0022] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0023] In the technical solutions of this application, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.

[0024] Figure 1 A flowchart of an example of a rental recommendation method based on multimodal data fusion analysis according to an embodiment of the present application is shown.

[0025] Regarding the execution subject of the method of the embodiment of the present application, it can be any controller or processor with computing or processing capabilities. Specifically, it can be implemented by a housing recommendation analysis system or a rental platform server. By performing data mining analysis on the drainage information of new users, integrating and analyzing multiple modal data sources such as delivery channel information, material attribute data, geographic location and housing description data, the recommendation system no longer relies solely on the user's personal portrait and historical behavior data, but can infer the user's potential needs based on multi-level information. Through the fusion analysis of multimodal data, the user's rental needs and preferences can be more comprehensively portrayed, thereby improving the accuracy and applicability of the recommendation.

[0026] In some examples, it can be integrated into an electronic device or terminal through software, hardware, or a combination of software and hardware, and the type of terminal or electronic device can be diverse, such as a mobile phone, tablet computer, or desktop computer, etc.

[0027] like Figure 1 As shown, in step S110, when a user triggering operation to release materials on the rental platform is detected, the release channel information, user geographic location and release material attribute data of the corresponding new users are obtained. The lead material attribute data includes the material preference group type and material house description information.

[0028] In some embodiments, the material property description information includes at least one of the following: the type of the material property, the decoration style, the price range, and the highlights of the property.

[0029] For example, a rental platform promotes itself on various advertising platforms (for example, short video platforms, search engine platforms, etc.) by placing traffic-generating materials. When a user is interested in a certain traffic-generating material, such as clicking, following, liking, collecting or downloading, and entering the rental platform system, the relevant information of the traffic-generating material can be analyzed, including the delivery channel information (for example, which advertising channel is used for the traffic, such as social media, search engines, etc.), the user's geographic location information, and the attribute data of the traffic-generating material.

[0030] It's important to note that each lead-generating creative on rental platforms is typically carefully designed, often featuring distinct listing characteristics. These features can simultaneously capture multiple dimensions of information through listing tags. These include listing type (e.g., studio, apartment, entire rental), decor style (e.g., modern, minimalist, retro), price range (e.g., 1,000-1,500 yuan / month), and listing highlights (e.g., "direct subway access" or "nearby supermarkets"). These attribute data provide an initial basis for predicting user demand.

[0031] In step S120, the delivery channel information and the delivery material attribute data are input into the rental demand prediction model to output a plurality of rental demand types and corresponding prediction confidence, thereby constructing a rental demand matrix corresponding to the new user.

[0032] Specifically, the system inputs the collected delivery channel information and delivery material attribute data as input features into the rental demand prediction model. The model can be a multi-classification model based on machine learning or deep learning, which can identify the possible rental demand types of the user after training.

[0033] It should be understood that the type of rental demand prediction model can be diverse, which is not limited in the present embodiment. In one example of the present embodiment, the rental demand prediction model can adopt a rule-based model to infer the user's demand through a series of pre-defined rules, which can be set according to platform historical data and expert experience, for example, if the user comes from a certain delivery channel (such as the "fashion home" page of a social platform), he or she may prefer a house with a fashionable decoration style. In another example of the present embodiment, the rental demand prediction model can also adopt a deep learning model (such as convolutional neural network CNN, long short-term memory network LSTM, etc.) to process a large amount of heterogeneous data and automatically extract high-dimensional features.

[0034] In some embodiments, the rental demand type includes at least one of the following: high cost performance, fine decoration, convenient transportation, complete life supporting, single living or family living. Further, the model will predict different rental demand types (such as "high cost performance", "fine decoration", "convenient transportation", etc.), output the prediction confidence of different demand types, and form a rental demand matrix. Each row of the rental demand matrix represents a rental demand type, and each column represents a corresponding prediction confidence, for example, for a user who prefers "high cost performance" house, the model may assign a higher confidence to this demand type.

[0035] Here, the rental demand matrix can be regarded as a preliminary portrait of user demand, so that the system quickly captures the new user preference in the cold start stage. Using the information of the lead material contacted by the user, the system can infer the rental demand type of the user. This demand prediction method based on delivery channel and material attribute breaks through the dependence on historical user behavior data and realizes the identification of user demand in the cold start state. By constructing the demand matrix, the system can accurately express the user's multiple potential demands in a quantitative way, providing a basis for personalized and accurate recommendation of house resources.

[0036] In step S130, candidate house resources are filtered from the house resource library according to the user's geographical location, wherein the travel distance between the geographical location indicated by the candidate house resource and the user's geographical location is less than a preset threshold.

[0037] In some embodiments, based on the user's geographic location, houses that meet the user's travel distance requirements are screened out from the house library to form a candidate house set. The travel distance can be calculated in real time through a map API (such as Baidu Maps, Amap, etc.) to screen out houses with a travel distance less than a threshold. The travel distance threshold can be adjusted according to the platform's usage scenario. For example, the city center area can be set to 3 kilometers, and the suburbs can be set to 5 kilometers or more, so as to flexibly adapt to the needs of different users.

[0038] This way, distance becomes a key factor for users in rental recommendations, and the screening mechanism for candidate listings significantly improves the relevance of recommended listings. By initially screening candidate listings based on user location information synchronized with the traffic generation platform, we ensure that recommended listings are located within the user's living radius and reduce system resource consumption for comparing and analyzing massive listings.

[0039] In step S140 , for each candidate housing source, the matching degree relative to the housing rental demand matrix is ​​calculated based on the housing source description data and the corresponding travel distance of the candidate housing source.

[0040] It should be noted that the matching degree can be calculated in a variety of ways. For example, based on the cosine similarity calculation results or the dot product calculation results, the feature vector of the property description and the demand type characteristics in the user's demand matrix are calculated to evaluate the correlation between the property and the user's needs. No restrictions are imposed here for the time being.

[0041] In some embodiments, the matching degree of the selected candidate listings relative to the user's needs is calculated based on the listing's descriptive data (e.g., type of listing, decor, price, etc.) and the travel distance from the user's geographic location, combined with the confidence values ​​of each requirement type in the rental demand matrix. This matching degree calculation can be implemented using a weighted scoring approach, multiplying each listing's attributes by the corresponding requirement confidence values ​​in the demand matrix and summing them to produce an overall matching degree score. For example, a user who prefers "high cost-effectiveness" and "convenient transportation" will prioritize listings that meet these criteria and assign them a higher matching degree.

[0042] In step S150, a list of recommended properties for attracting new users is generated based on a preset number of candidate properties ranked high in terms of corresponding matching scores.

[0043] In some implementations, the candidate listings are sorted in descending order of match scores, and a preset number of top-ranked listings are selected to generate a final list of recommended listings. Furthermore, the order of presentation in this list can be determined based on the match scores, ensuring that users prioritize the listings that best meet their needs when browsing the recommended list.

[0044] It should be understood that the recommendation list can be presented in various formats, such as lists, cards, and maps, to help users intuitively understand the listing information. Furthermore, each listing can be annotated in the recommendation list to highlight its compatibility with user needs (e.g., "close proximity" or "high value"), to enhance user understanding and trust in the recommendations.

[0045] In this embodiment, the delivery channel and material attribute information triggered by the user is input into the rental demand prediction model, which outputs multiple demand types and their prediction confidence levels, and constructs a demand matrix to reflect the diverse demand trends of new users. This multi-type demand prediction not only increases the flexibility of recommendations but also improves the reliability of recommendations, allowing the system to provide more comprehensive recommendation options when users first contact it.

[0046] Based on the user's geographic location, the system then filters candidate listings within a preset travel distance threshold, ensuring that recommended listings meet the user's commuting or living area requirements. Furthermore, by calculating and ranking the degree of match between the candidate listings' descriptions and the requirements matrix, it prioritizes listings that best meet the user's needs. This not only improves the relevance of recommendation results but also significantly enhances the personalization of the recommendation list, improving the user experience.

[0047] Through the embodiments of the present application, multimodal data is used to construct a preliminary demand profile of a new user, and the calculation of the matching degree between geographic location and demand is introduced into the housing recommendation analysis system, which can effectively solve the cold start problem, improve the accuracy of the recommendation results and user satisfaction, and optimize the housing recommendation effect of new users in the cold start case of the rental platform.

[0048] In some examples of the embodiments of the present application, the rental demand prediction model adopts a conditional generative adversarial network (CGAN), which includes a generator and a discriminator, and uses the adversarial training mechanism of the generator and the discriminator to generate high-quality demand prediction results.

[0049] Specifically, the discriminator distinguishes between real demand (e.g., historical user demand data or expert-annotated demand forecasts) and generated demand types. It uses feedback to help the generator adjust its generation process, making its generated demand forecasts more realistic and accurate. The discriminator generally only participates in optimizing the generator during the model training phase and does not operate during the inference phase. The structure of the discriminator in current related technologies can be referenced or borrowed, so we will not elaborate on it here.

[0050] As the core part of the rental demand prediction model CGAN, the generator is used to generate the conditional output of the rental demand prediction. Specifically, the generator generates the rental demand prediction matched with the given condition input (such as the attribute of the material put, the information of the channel put, etc.) according to the given condition input. The generator generates a possible demand distribution of a user, that is, a rental demand matrix, according to the condition information, for example, the rental demand type (such as "high cost performance", "refurbished" and the like) of the user and the prediction confidence of each demand type.

[0051] In the training process of the CGAN, the generator and the discriminator are trained in an antagonistic manner and optimize each other. The training target of the generator is to make the demand prediction generated by the generator as close as possible to the real demand data, so as to deceive the discriminator and avoid being identified as false data by the discriminator. The generator adjusts its parameters in the training process until the difference between the demand prediction generated by the generator and the real demand data is minimized. The training target of the discriminator is to accurately distinguish the generated false data from the real data, and to help the generator gradually improve its generation strategy.

[0052] Through the embodiments of the present application, the CGAN architecture is used to construct the rental demand prediction model. When the user lacks historical behavior data, the conditional generative adversarial network can generate preliminary demand prediction relying on the existing external information (such as the channel put, the material attribute, etc.), thereby alleviating the cold start problem.

[0053] Figure 2 A structure connection diagram of an example of the generator in the CGAN architecture of the rental demand prediction model according to an embodiment of the present application is shown.

[0054] As shown in Figure 2 , the generator 200 includes an input layer 210, a fusion layer 220, a feature decomposition layer 230 and an output layer 240.

[0055] The input layer 210 is used to embed the input channel put information and the input material attribute data into a unified vector space through multi-modal feature encoding to generate the corresponding encoding features of each modality.

[0056] Here, the main function of the input layer 210 is to convert various input data (such as delivery channel information, delivery material attribute data, etc.) into a unified vector space representation through multimodal feature encoding. For example, discretization encoding (such as one-hot encoding) can be used to convert channel information (such as social media, search engines, advertising platforms, etc.) into vector representation. In addition, for delivery material attribute data, such as material preference group type (such as "students", "office workers") and material property description information (such as "decoration style"), it can be vectorized through the embedding layer. For numerical data (such as price range), it is standardized; if it is categorical data (such as property type), it is represented by one-hot encoding or embedding, and so on.

[0057] As a result, the encoded features of each modality will be embedded in a unified vector space, which can effectively eliminate the dimensionality differences between different features and enable subsequent processing to be performed based on a unified representation.

[0058] The fusion layer 220 is used to fuse the coding features of each modality to determine the corresponding multimodal fusion features.

[0059] Here, the fusion layer is used to fuse features from different modalities (delivery channel information, material attributes, etc.) to generate multimodal fusion features. It should be understood that the fusion method can be diverse, such as feature splicing, weighted fusion, or other methods, which are not limited here.

[0060] The feature decomposition layer 230 is used to perform hierarchical processing on the multimodal fusion features according to the condition information to determine the personalized demand matrix of attracting new users at the coarse-grained and refined demand levels.

[0061] Conditional information is defined based on delivery channel information, delivery material attribute data, and a pre-set set of rental demand types. Delivery channel information reflects the user's initial interest or needs. The delivery material attribute data, including information such as the type of property and price range, and the rental demand type, including high cost-effectiveness, well-decorated, and convenient transportation, help the model understand the context and hierarchy of user needs.

[0062] In some embodiments, first, in the coarse-grained demand hierarchy analysis, a coarse-grained prediction of user demand is generated based on the user's delivery channel information and material attributes. For example, based on the user's click behavior from the "cost-effective housing" advertisement, it is predicted that the user may be more interested in "budget-friendly housing". Then, in the refined demand hierarchy analysis, the detailed attribute information of the delivery material (such as the decoration style of the housing, the specific price range, etc.) is combined for refined decomposition to obtain user demand predictions on certain details. For example, the user's preference for "convenient transportation" may be further refined into the demand for "close to the subway station". Afterwards, the feature decomposition layer outputs a personalized demand matrix, in which each element represents the user's interest level or demand confidence in a certain demand type.

[0063] Through the feature decomposition layer and hierarchical demand decomposition, we can capture the different dimensions of user needs and avoid over-simplification of user needs. Hierarchical processing can provide more fine-grained personalized demand predictions, achieve accurate modeling of user needs, and enhance the depth of understanding of user needs.

[0064] The output layer 240 is used to output multiple rental demand types and corresponding prediction confidence levels according to the personalized demand matrix.

[0065] In some implementations, the output layer outputs multiple rental demand types and their corresponding prediction confidence scores based on the decomposed personalized demand matrix, representing the user's interest in each demand type. Furthermore, the output layer can calculate the probability of each demand type through forward propagation and use this probability as the corresponding confidence score, or calibrate the personalized demand matrix to output a highly accurate match for each demand type. Demand types with high confidence scores are more relevant to the user.

[0066] It should be noted that in cGAN, the output of the generator is not necessarily the true demand type distribution, but through adversarial training with the discriminator, it can generate demand prediction results close to the true distribution.

[0067] Through the generator structure provided in the embodiment of the present application, relying on multimodal data fusion, conditional constraints and information hierarchical processing, the demand prediction distribution for attracting new users is effectively generated, and the authenticity and diversity of the generated effect are improved by utilizing the adversarial training in the CGAN architecture, ultimately achieving higher demand prediction accuracy when user historical interaction data is scarce.

[0068] In some examples of the present application, fusion layer 220 employs an adaptive attention-based fusion mechanism. Specifically, the fusion layer applies adaptive attention to the features of each modality to identify the different impacts of each modality on the user's rental demand. Based on adaptive attention, the model can highlight the features that are more critical to demand prediction within multimodal information.

[0069] Specifically, the fusion layer 220 is used to perform the following operations:

[0070] Calculate the attention weights corresponding to the encoding features of each modality:

[0071]

[0072]

[0073] Where, α i represents the normalized attention weight of modality i, is the embedding vector of the i-th modality, N represents the number of modalities, represents the attention weight of modality i, σ att is the activation function used to calculate the attention weight; is the projection bias term, is the projection weight matrix, which transforms the feature dimension d of mode i into i Convert to uniform dimension d att ; is the sum of attention weights of all modalities.

[0074] Based on the attention weight, the encoding features of each modality are weighted and fused to obtain weighted multimodal features:

[0075]

[0076] Where Z fusion represents weighted multimodal features.

[0077] Through the adaptive attention mechanism, the fusion layer can dynamically assign weights to multimodal information based on the specific characteristics of different users, thereby ensuring that during the fusion process, modal features that have a more significant impact on demand forecasting are given higher weights.

[0078] In addition, in order to further strengthen the interactive relationship between the features of each modality, a modal interaction mechanism is introduced in the fusion layer so that information from different modalities can influence each other, thereby improving the accuracy of demand forecasting.

[0079] Specifically, a bilinear interaction method is used to generate modal interaction terms to capture the correlation characteristics between different modalities:

[0080]

[0081] wherein I ij denotes the modal interaction term between modal i and modal j, is the transpose vector of , σ int is the activation function of the modal interaction; is the weight matrix of the modal interaction, denoting the correlation strength between modal i and j; d int denotes the feature dimension of the modal interaction, d fusion denotes the feature dimension of Z fusion ; is the bias term of the modal interaction, used to adjust the linear combination of the inter-modal interaction features.

[0082] The modal interaction mechanism is used to calculate the interaction term between modalities, capture the correlation information between different modal features, effectively utilize the complementarity of multi-modal features, so that each modal feature not only is simply weighted and superimposed, but also interacts with each other, thereby providing a more comprehensive expression of user demand. For example, the user's demand may be influenced by both the attributes and the geographic location of the material to be put in, and the modal interaction mechanism can capture these correlations to generate a more refined demand prediction for the user.

[0083] The modal interaction term and the weighted multi-modal feature are spliced to form the final fusion feature representation containing interaction information:

[0084] Z final = [Z fusion ; I ij ], equation (5)

[0085] wherein Z final denotes the multi-modal fusion feature.

[0086] Through the weighted fusion of multi-modal features and the generation of interaction terms, the multi-modal fusion feature Z final output by the fusion layer contains personalized information of user preferences, so that the model can more accurately predict the user's rental demand type and the corresponding preference strength.

[0087] Through the embodiments of the present application, the adaptive attention weight distribution and the modal interaction mechanism are used in the fusion layer, which improves the expression ability of the rental demand prediction model in multi-modal feature processing, enhances the adaptability of the model to personalized demand and the generalization ability of the model to diversified user demand, and reduces information redundancy, which helps to improve the prediction accuracy of rental demand.

[0088] In some examples of the embodiments of the present application, the feature decomposition layer 230 is used to perform multi-level calculation processing on the multi-modal fusion feature from coarse granularity to fine granularity. Specifically, it performs the following operations to perform hierarchical processing on the multi-modal fusion feature:

[0089] Calculate the demand category weight vector for multimodal fusion features:

[0090] γ=softmax(W γ ·Z final +b γ ), Formula (6)

[0091] Where γ represents the demand category weight vector, which indicates the user's preference for each broad demand category. The softmax function is used to normalize each weight to ensure that the sum of the weights of each category is 1; W γ and b γ They represent the weight matrix and bias term used to generate the demand category weight vector respectively.

[0092] Based on the user's multimodal fusion features, a demand category weight vector γ is generated. This adaptively assigns weights to different demand categories, enabling the model to dynamically adjust its focus on different demand categories based on the user's performance in delivery channel information and material attribute data. This allows the model to more accurately capture users' broad rental needs (such as high cost-effectiveness and convenient transportation), laying the foundation for personalized recommendations.

[0093] The multimodal fusion features are weighted and aggregated according to the demand category weight vector to generate a coarse-grained demand matrix:

[0094] D coarse =γ⊙(W coarse ·Z final +b coarse ), Formula (7)

[0095] Where ⊙ represents element-wise multiplication, D coarse represents the coarse-grained demand matrix, W coarse and b coarse They represent the weight matrix and bias term used to generate coarse-grained features respectively.

[0096] Here, we leverage users' multimodal features to generate a personalized demand weight vector γ, representing their preference for each broad demand category. This dynamic aggregation of multimodal features, rather than a fixed weight matrix, generates a coarse-grained demand matrix. This dynamic aggregation of demand categories allows us to extract a preliminary match to demand from delivery channels and creative attributes, even when user information is limited. This improves the model's adaptability to cold-start users.

[0097] Calculate the preliminary matching degree of the delivery channel information and delivery material attribute data with respect to each rental demand type in the rental demand type set, and construct condition information based on each preliminary matching degree:

[0098]

[0099] Where, E C is the embedding representation of the delivery channel information, T q represents the embedding representation of the qth requirement type, ||E C || is E C The L2 norm of ||T q || is T q L2 norm of M C,q is the delivery channel information and the qth demand type T q The initial matching degree indicates the similarity between the delivery channel information and the qth demand type; M M,q To put the material attributes and T q The initial matching degree indicates the similarity between the attributes of the delivered material and the qth demand type; E M is the embedded representation of the delivery material attributes, ||E M || is the L2 norm of the delivery material attribute; C cond Embedding the conditional information to represent the user's personalized needs and preferences; β q represents the weighting coefficient of the qth rental demand type, and Q represents the total number of rental demand types in the rental demand type set.

[0100] Here, the cosine similarity is used to calculate the initial matching degree between the delivery channel information and the delivery material attributes and each demand type, thereby generating the conditional information embedding C that reflects the user preference. cond , ensuring that the model can extract the user's real demand preferences from the delivery information and material attributes, and convert them into conditional information vectors.

[0101] The high-dimensional relationship between coarse-grained features and conditional information is captured through multi-dimensional convolutional layers to generate a personalized demand matrix:

[0102] D input =[D coarse ; C cond ] , Formula (11)

[0103]

[0104] Where, represents the personalized demand matrix, which is used to represent the fine-grained demand preferences of attracting new users. h′ and w′ are the height and width of the personalized demand matrix, d fine The characteristic dimension that represents the refined demand characteristics; D input Represents the combined input features of the multi-dimensional convolutional layer; Represents a multidimensional convolution kernel, which is used to capture the complex relationship between coarse-grained features and conditional information in the spatial dimension and feature dimension, Wconv The size in the spatial dimension is k h ×k w ,* represents convolution operation, b conv represents the convolution bias term, σ conv represents the activation function of the convolution layer, d input Represents the feature dimension of the combined input features; u and v are the convolution kernels W conv The spatial position index of represents the index of the convolution kernel in the height direction and width direction respectively; Represents the weight matrix of the convolution kernel at position (u,v).

[0105] Through multi-dimensional conditional convolution, the feature decomposition layer can capture the high-dimensional relationship between coarse-grained demand features and condition information, and can convolve coarse-grained features and condition information in spatial and feature dimensions, extract their local patterns, generate more refined user demand expressions, and realize the refined demand matrix D. fine This enables the model to capture users’ preferences for detailed needs such as rental range, decoration style, and transportation convenience at a fine-grained level, effectively reflecting users’ personalized needs in specific scenarios.

[0106] Therefore, the feature decomposition layer realizes the fusion expression of broad demand and fine-grained demand in the rental demand prediction model, which not only improves the accuracy of the model for user needs, but also enhances the adaptability of the model.

[0107] In some examples of the embodiments of the present application, the output layer 240 is used to perform the following operations:

[0108] Based on the global weighted pooling mechanism, the features of each spatial position in the personalized demand matrix are weighted pooled to obtain the weighted pooled demand features:

[0109]

[0110] Where D pool is the weighted pooling demand feature; α u,v is the pooling weight, which represents the feature weight at the spatial position (u, v); is the feature representation of the personalized demand matrix at the spatial position (u, v), W α and b α Respectively represent the weight matrix and bias term used to calculate the pooling weight, and exp represents the exponential function; is a normalization term to ensure that all α u,v The sum of is 1.

[0111] Through the global weighted pooling mechanism, the output layer can refine the demand matrix D fineAssign different weights to the features of each spatial position and generate adaptive pooling weights α u,v , adaptively performing weighted summation on features at different spatial positions, suppressing irrelevant information and strengthening the influence of key feature positions.

[0112] The weighted pooled demand features are processed based on the softmax function to obtain the preliminary predicted probability distribution of the demand type:

[0113]

[0114] Where, Represents the preliminary predicted probability distribution of the demand type, where the predicted probability value is used to quantify the confidence of the corresponding rental demand type; and Represent the classification weight matrix and classification bias term respectively.

[0115] Here, the predicted probability of the model is directly used as the confidence, and the predicted probability output by softmax naturally reflects the confidence of the model.

[0116] The preliminary predicted probability distribution of the demand type is calibrated based on the conditional information to output multiple rental demand types and corresponding prediction confidence levels:

[0117]

[0118] Where, represents the corrected probability distribution of demand types, and Represent the conditional correction weight matrix and conditional correction bias term respectively; σ sig Represents the sigmoid activation function to ensure that the correction coefficient of the conditional information is in the [0,1] interval.

[0119] Here, after obtaining the prediction probability, a dynamic correction mechanism based on conditional information is introduced to further enhance the personalized ability of prediction. cond Dynamically adjust the predicted probability of the demand type, and supervise and adjust the predicted probability based on the user's delivery channel information and material attributes, so that the final output of the model is more in line with the user's actual needs.

[0120] It should be noted that in cold start scenarios, historical user behavior data may be limited, making it difficult for the model to accurately grasp user preferences. Through a global weighted pooling mechanism and conditional information correction, existing distribution channel information and creative attributes can be reused for supervision. The pooling mechanism ensures the effective aggregation of key features, while conditional correction infers potential preferences based on information such as user registration path and the ad creatives they see.

[0121] This combined global weighted pooling mechanism and a supervised correction mechanism based on conditional information ensures that the model's predictions are more tailored to the user's individual needs. Whether in cold-start scenarios or diverse user preference scenarios, the model can analyze new users' needs and preferences based on traffic flow information.

[0122] It should be noted that the generator and discriminator of the cGAN-based rental demand prediction model are trained alternately based on the data sample set. Specifically, during training, the generator is first fixed and the discriminator is trained to better distinguish between real and generated data, reducing the discriminator's loss. Then, the discriminator is fixed and the generator is trained to generate data that is more capable of "deceiving" the discriminator, reducing the generator's loss.

[0123] In the adversarial network of the embodiment of the present application, the task of the discriminator is to judge the difference between the generated data (demand forecast distribution generated based on conditional information) and the real data. To achieve this goal, the discriminator can receive both the generated data and the conditional information C at the input. cond , so as to more effectively identify whether the generated data is reasonable.

[0124] Specifically, the loss function of the discriminator is expressed as follows:

[0125]

[0126] Where S is the number of samples in the data sample set, s is the sample index; x (s) represents the input feature of the sth sample, Is the generator based on x (s) and condition information The generated demand forecast distribution, represents the conditional information of the sth sample, Represents the one-hot encoding of the real rental demand type label of the s-th sample; is the discriminator's response to the true demand label y (s) and condition information The output probability of , with an expected value of 1, indicates that the demand type is true, so that the discriminator can learn to establish the correct association between real data and conditional information; is the output of the discriminator to the generator and condition information The output probability of , with an expected value of 0, indicates that the demand type is generated data, so that the discriminator can learn to distinguish generated data (under the same conditional information) from real data.

[0127] Here, by introducing the condition information C cond,The discriminator can detect whether the generated demand distribution meets the requirements of the conditional information, and then provide effective feedback to the generator, helping the generator to better utilize the conditional information to generate personalized solutions that meet user needs in subsequent iterations.

[0128] In cGAN, the input of the generator G contains not only the user’s input data x (i.e., the user’s feature data), but also the conditional information C cond In this way, the generator controls the generated demand type prediction distribution through conditional information, making the generated results more in line with the actual needs of users.

[0129] The loss function of the generator is expressed as follows:

[0130]

[0131] Where, represents the loss function of the generator, Represents the loss term of the generator's classification of the rental demand type for the s-th sample in the data sample set, which is used to measure the difference between the generated rental demand type prediction and the true label; represents the adversarial loss term for the sth sample in the data sample set, indicating the likelihood that the generated data is judged as real by the discriminator; λ1 and λ2 represent the corresponding loss term hyperparameters respectively; Represents a generator based on x (s) and condition information The generated predicted probability of rental demand type l.

[0132] Here, the generator loss jointly optimizes the classification loss and the adversarial loss to ensure that the demand distribution generated by the generator not only conforms to the real distribution of user demand, but also confuses the discriminator, making it difficult to distinguish the generated data. In this framework, the conditional information C cond Embedded in the generator's input, this directly influences the category distribution and feature representation of the generated results, ensuring that the generated demand distribution meets the user's specific conditional requirements. Furthermore, the generator generates a personalized demand prediction distribution guided by conditional information through classification and adversarial losses.

[0133] Specifically, in each round, the generator is fixed and the discriminator loss is optimized using the real data and the data generated by the generator. Maximizing this loss improves the discriminator’s ability to distinguish between real and generated data. In addition, in each round, the discriminator is fixed and the generator loss is minimized. The data generated by the generator can "fool" the discriminator, while making the demand type prediction as close to the true label as possible.

[0134] Through the adversarial loss of the generator and discriminator, the model can continuously optimize the generated results during the adversarial training process, gradually making the generated data distribution closer to the real data distribution. Ultimately, the generator is prompted to learn the characteristics of the real demand distribution, thereby reflecting higher authenticity in the generated rental demand forecast. By embedding conditional information into the input of the generator and discriminator and using this conditional information to guide the adversarial training of the generator and discriminator, the model can dynamically adjust the generated demand forecast based on the user's specific context (such as delivery channel, material attributes, etc.), generating a more personalized demand forecast distribution.

[0135] Figure 3 An operational flow chart of an example of matching degree calculation according to an embodiment of the present application is shown.

[0136] like Figure 3 As shown, in step S310, the property description data and the corresponding travel distance are concatenated to generate a corresponding property feature vector.

[0137] Specifically, when constructing the property feature vector, travel distance is considered a specific feature dimension and integrated into the same vector along with the property description features. Travel distance is a numerical value that can be normalized or other data standardization methods to convert it to a similar scale for easy integration with other features.

[0138] F prop =[f1,f2,…,f G-1 ,d distance ] , Formula (21)

[0139] Where f1, f2, …, f G-1 Represents the information characteristics of each dimension in the property description data, d distance Represents the normalized characteristic of travel distance, F prop represents the property feature vector;

[0140] In step S320, a bilinear pooling matching operation is performed on the housing resource feature vector and the rental demand matrix to obtain a matching result matrix through interactive feature analysis.

[0141] Bilinear pooling establishes multi-layered, second-order interactions between rental demand characteristics and listing characteristics, capturing the complex relationships between different demand types and listing characteristics. Compared to simple linear weighting, this approach more accurately reflects users' focus on specific needs. Furthermore, by assigning different weights to each feature interaction, it more intelligently handles the correlation between different demand types and listing characteristics, enhancing personalized matching.

[0142] Here, travel distance is an important component of the property feature vector and can be given a higher weight to highlight its influence in the matching calculation. This can be achieved by adjusting the weights in the bilinear pooling. For example, the last dimension of the feature vector containing travel distance (i.e., d distance ) are given higher weights.

[0143]

[0144] Where, F represents the prediction confidence of the qth rental demand type in the rental demand matrix; prop is the property feature vector, containing G-dimensional features; represents the g-th eigenvalue in the property feature vector, where the eigenvalue of the last dimension represents the travel distance feature; W (q,g) represents the weight parameter of bilinear pooling, which is used to control the interaction between rental demand characteristics and housing supply feature vectors; d distance represents the normalized travel distance feature; W (q,G) Represents the weight parameter of the travel distance feature, which is used to control the importance of the travel distance feature in the matching calculation; μ is the additional weight coefficient of the travel distance feature, and η is the hyperparameter for adjusting the rate of change of the travel distance weight; P match Represents the matching result matrix; C qg Score the correlation between demand type q and the g-th eigenvalue in the property feature vector; is a normalized term, which represents the sum of the correlation scores of all demand and housing feature vectors in the rental demand matrix.

[0145] In the above formula, by normalizing the correlation scores between demand characteristics and property characteristics, we can better balance the contributions of different demand types and property characteristics, making the matching calculation more stable and avoiding the excessive influence of individual characteristics on the results.

[0146] It should be noted that by weighting the distance between the user's location and the listing's location as an independent feature, the weight of the travel distance is dynamically adjusted based on the distance, so that listings with shorter distances receive a higher degree of matching. This allows for a sensitive response to the user's travel needs, ensuring that the system prioritizes the user's geographic location when recommending listings.

[0147] In addition, the two types of hyperparameters for travel distance mentioned above can, on the one hand, be customized according to the needs of the scenario, and on the other hand, the weight of the travel distance feature can be automatically optimized during model training, so that the model can adaptively adjust the importance of travel distance in the matching calculation according to the needs and preferences of different users, thereby achieving personalized recommendations.

[0148] In step S330 , the matching result matrix is ​​projected by using feature compression and linear projection to obtain the matching degree.

[0149] S match =σ mat (W proj ·vec(P match )+b proj ), Formula (26)

[0150] Where S match represents the matching degree of candidate housing, σ mat represents the nonlinear activation function, W proj is the projection matrix, which is used to map the high-dimensional matching matrix to the low-dimensional space; vec(·) represents the operation of expanding the matrix into a vector; b proj Represents the bias vector.

[0151] Here, by adopting feature compression and linear projection, the high-dimensional matching matrix can be mapped into a low-dimensional space, thereby greatly reducing the amount of calculation, effectively retaining the core information between multi-dimensional features, and at the same time improving the speed of matching calculation, ensuring real-time responsiveness in large-scale recommendation scenarios and meeting the real-time requirements of high-concurrency recommendation systems.

[0152] In the embodiment of the present application, travel distance is explicitly included in the property feature vector. By assigning a higher weight, the travel distance feature is highlighted in the matching calculation, making the recommendation results more in line with the user's geographical needs. Furthermore, bilinear pooling is used to capture all second-order interactions between user needs and property features. As part of the property features, the travel distance feature interacts with the demand matrix in multiple ways, which can more carefully reflect the correlation between the user's travel preferences and other needs. Projecting on the high-dimensional matching matrix generated by bilinear pooling maps the high-dimensional information to a low-dimensional matching score, which not only retains rich interaction information but also maintains computational efficiency.

[0153] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of combined actions, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0154] Figure 4A structural block diagram of an example of a rental recommendation system based on multimodal data fusion analysis according to an embodiment of the present application is shown.

[0155] like Figure 4 As shown, the rental recommendation system 400 based on multimodal data fusion analysis includes a traffic data acquisition unit 410, a rental demand prediction unit 420, a candidate housing source screening unit 430, a housing source matching calculation unit 440 and a traffic-generating housing source recommendation unit 450.

[0156] The traffic data acquisition unit 410 is used to obtain the delivery channel information, user geographic location and delivery material attribute data of the corresponding new user when a user triggering operation to deliver materials to the rental platform is detected; the traffic material attribute data includes the material preference group type and material housing description information.

[0157] The rental demand prediction unit 420 is used to input the delivery channel information and the delivery material attribute data into the rental demand prediction model to output multiple rental demand types and corresponding prediction confidence levels, thereby constructing a rental demand matrix corresponding to the new users attracted; each row of the rental demand matrix represents a rental demand type, and each column represents a corresponding prediction confidence level.

[0158] The candidate listing screening unit 430 is configured to screen a set of candidate listings from a listing database based on the user's geographic location, wherein the travel distance between the geographic location indicated by the candidate listing and the user's geographic location is less than a preset threshold.

[0159] The housing matching calculation unit 440 is used to calculate the matching degree of each candidate housing source relative to the housing rental demand matrix based on the housing description data and the corresponding travel distance of the candidate housing source.

[0160] The house recommendation unit 450 is used to generate a house recommendation list for the new user to be attracted based on a preset number of candidate houses ranked high in corresponding matching degrees.

[0161] In some embodiments, an embodiment of the present application provides a non-volatile computer-readable storage medium, which stores one or more programs including execution instructions, and the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute any of the steps of the above-mentioned rental recommendation method based on multimodal data fusion analysis in the present application.

[0162] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to perform any step of the above-mentioned rental recommendation method based on multimodal data fusion analysis.

[0163] In some embodiments, an embodiment of the present application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the steps of the rental recommendation method based on multimodal data fusion analysis.

[0164] Figure 5 This is a hardware structure diagram of an electronic device for performing a rental recommendation method based on multimodal data fusion analysis provided by another embodiment of the present application, such as Figure 5 As shown, the device includes:

[0165] One or more processors 510 and memory 520, Figure 5 A processor 510 is taken as an example.

[0166] The device for executing the housing rental recommendation method based on multimodal data fusion analysis may further include: an input device 530 and an output device 540 .

[0167] The processor 510, the memory 520, the input device 530 and the output device 540 may be connected via a bus or other means. Figure 5 The bus connection is taken as an example.

[0168] Memory 520, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the rental recommendation method based on multimodal data fusion analysis in the embodiments of the present application. Processor 510 executes the non-volatile software programs, instructions, and modules stored in memory 520 to execute various server functional applications and data processing, thereby implementing the rental recommendation method based on multimodal data fusion analysis in the aforementioned method embodiment.

[0169] The memory 520 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device, etc. In addition, the memory 520 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 520 may optionally include a memory remotely located relative to the processor 510, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0170] The input device 530 can receive input digital or character information and generate signals related to user settings and function control of the electronic device. The output device 540 can include a display device such as a display screen.

[0171] The one or more modules are stored in the memory 520 , and when executed by the one or more processors 510 , perform the housing recommendation method based on multimodal data fusion analysis in any of the above method embodiments.

[0172] The above-mentioned product can execute the method provided in the embodiment of this application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of this application.

[0173] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to:

[0174] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and their primary purpose is to provide voice and data communications. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0175] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers and have computing and processing capabilities, and generally also have mobile Internet access. These terminals include PDAs, MIDs, and UMPCs.

[0176] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0177] (4) Other onboard electronic devices with data interaction functions, such as onboard computer devices installed in vehicles.

[0178] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0179] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, or of course, by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the relevant technology, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiment.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A rental recommendation method based on multimodal data fusion analysis, comprising: When a user triggering operation to place a material on the rental platform is detected, the distribution channel information, user geographic location, and placement material attribute data of the corresponding new user are obtained; the placement material attribute data includes the material preference group type and material housing description information; Inputting the delivery channel information and the delivery material attribute data into a rental demand prediction model to output multiple rental demand types and corresponding prediction confidences, thereby constructing a rental demand matrix corresponding to the new users attracted; each row of the rental demand matrix represents a rental demand type, and each column represents a corresponding prediction confidence; Filtering a set of candidate properties from a property library based on the user's geographic location, wherein the travel distance between the property's geographic location indicated by the candidate properties and the user's geographic location is less than a preset threshold; For each candidate housing source, calculating a matching degree relative to the housing rental demand matrix based on the housing description data of the candidate housing source and the corresponding travel distance; Generate a list of recommended properties for the new user based on a preset number of candidate properties ranked by matching scores; The rental demand prediction model adopts a conditional generative adversarial network, which includes a generator and a discriminator; wherein the generator includes an input layer, a fusion layer, a feature decomposition layer and an output layer: The input layer is used to embed the input delivery channel information and delivery material attribute data into a unified vector space through multimodal feature encoding to generate corresponding coding features of each modality; The fusion layer is used to fuse the coding features of each modality to determine the corresponding multimodal fusion features; The feature decomposition layer is used to perform hierarchical processing on the multimodal fusion features according to condition information to determine a personalized demand matrix for attracting new users at the coarse-grained and refined demand levels; the condition information is defined based on the delivery channel information, the delivery material attribute data, and a preset set of rental demand types; The output layer is used to output multiple rental demand types and corresponding prediction confidence levels according to the personalized demand matrix.

2. The method according to claim 1, wherein The material property description information includes at least one of the following: the type, decoration style, price range and highlights of the material property; the rental demand type includes at least one of the following: high cost performance, fine decoration, convenient transportation, complete living facilities, single residence or family residence.

3. The method according to claim 1, wherein The fusion layer adopts a fusion layer based on an adaptive attention mechanism and is used to perform the following operations: Calculate the attention weights corresponding to the encoding features of each modality: Where, α i represents the normalized attention weight of modality i, is the embedding vector of the i-th modality, N represents the number of modalities, represents the attention weight of modality i, σ att is the activation function used to calculate the attention weight; is the projection bias term, is the projection weight matrix, which transforms the feature dimension d of mode i into i Convert to uniform dimension d att ; is the sum of attention weights of all modalities; Based on the attention weight, the encoding features of each modality are weighted and fused to obtain weighted multimodal features: Where Z fusion represents weighted multimodal features; A bilinear interaction method is used to generate modal interaction terms to capture the correlation characteristics between different modalities: Where, I ij represents the modal interaction term between modality i and modality j, for The transposed vector, σ int is the activation function of modal interaction; is the weight matrix of modal interaction, which represents the association strength between modalities i and j; d int Represents the characteristic dimension of modal interaction, d fusion Represents Z fusion characteristic dimensions; is the bias term of modal interaction, which is used to adjust the linear combination of inter-modal interaction features; The modal interaction terms and weighted multimodal features are concatenated to form a fused feature representation containing interaction information: WITH final =[Z fusion ;AND ij ], Where Z final Represents multimodal fusion features.

4. The method according to claim 3, wherein: The feature decomposition layer performs hierarchical processing on the multimodal fusion features by performing the following operations: Calculate the demand category weight vector for the multimodal fusion feature: γ=softmax(W γ ·WITH final +b γ ), Where γ represents the demand category weight vector, which indicates the user's preference for each broad demand category. The softmax function is used to normalize each weight to ensure that the sum of the weights of each category is 1; W γ and b γ denote the weight matrix and bias term used to generate the demand category weight vector respectively; The multimodal fusion features are weighted and aggregated according to the demand category weight vector to generate a coarse-grained demand matrix: D coarse =γ⊙(W coarse ·Z final +b coarse ), Where ⊙ represents element-wise multiplication, D coarse represents the coarse-grained demand matrix, W coarse and b coarse Represent the weight matrix and bias term used to generate coarse-grained features respectively; Calculate the preliminary matching degree of the delivery channel information and the delivery material attribute data with respect to each rental demand type in the rental demand type set, and construct condition information based on each preliminary matching degree: Where, E C is the embedded representation of the delivery channel information, T q represents the embedding representation of the qth requirement type, ‖E C ‖ is E C The L2 norm of , ‖T q ‖ is T q L2 norm of M C,q is the delivery channel information and the qth demand type T q The initial matching degree indicates the similarity between the delivery channel information and the qth demand type; M M,q To put material attributes and T q The initial matching degree indicates the similarity between the attributes of the delivered material and the qth demand type; E M is the embedded representation of the delivery material attributes, ‖E M ‖ is the L2 norm of the delivery material attributes; C cond Embedding conditional information to represent the user's personalized needs and preferences; β q represents the weighting coefficient of the qth rental demand type, and Q represents the total number of rental demand types in the rental demand type set; The high-dimensional relationship between coarse-grained features and conditional information is captured through multi-dimensional convolutional layers to generate a personalized demand matrix: D input =[D coarse ;C cond ], Where, represents the personalized demand matrix, which is used to represent the fine-grained demand preferences of attracting new users. h′ and w′ are the height and width of the personalized demand matrix, d fine Characteristic dimensions that represent refined demand characteristics; D input Represents the combined input features of the multi-dimensional convolutional layer; Represents a multidimensional convolution kernel, which is used to capture the complex relationship between coarse-grained features and conditional information in the spatial dimension and feature dimension, W conv The size in the spatial dimension is k h ×k w ,* represents convolution operation, b conv represents the convolution bias term, σ conv represents the activation function of the convolution layer, d input Represents the feature dimension of the combined input features; u and v are the convolution kernels W conv The spatial position index of represents the index of the convolution kernel in the height direction and width direction respectively; Represents the weight matrix of the convolution kernel at position (u,v).

5. The method according to claim 4, wherein The output layer is used to perform the following operations: Based on the global weighted pooling mechanism, the features of each spatial position in the personalized demand matrix are weighted pooled to obtain the weighted pooled demand features: Where D pool is the weighted pooling demand feature; α u,v is the pooling weight, which represents the feature weight at the spatial position (u, v); is the feature representation of the personalized demand matrix at the spatial position (u, v), W α and b α Respectively represent the weight matrix and bias term used to calculate the pooling weight, and exp represents the exponential function; is a normalization term to ensure that all α u,v The sum of is 1; The weighted pooled demand features are processed based on the softmax function to obtain a preliminary predicted probability distribution of the demand type: Where, Represents the preliminary predicted probability distribution of the demand type, where the predicted probability value is used to quantify the confidence of the corresponding rental demand type; and Represent the classification weight matrix and classification bias term respectively; The preliminary predicted probability distribution of the demand type is calibrated based on the condition information to output multiple rental demand types and corresponding prediction confidence levels: Where, represents the corrected probability distribution of demand types, and Represent the conditional correction weight matrix and conditional correction bias term respectively; σ sig Represents the sigmoid activation function to ensure that the correction coefficient of the conditional information is in the [0,1] interval.

6. The method according to claim 5, wherein: The generator and discriminator of the rental demand prediction model are trained alternately based on the data sample set; The loss function of the discriminator is expressed as follows: Where S is the number of samples in the data sample set, s is the sample index; x (s) represents the input feature of the sth sample, Is the generator based on x (s) and condition information The generated demand forecast distribution, represents the conditional information of the sth sample, Represents the one-hot encoding of the real rental demand type label of the s-th sample; is the discriminator's response to the true demand label y (s) and condition information The output probability of , the expected value is 1, indicating that the demand type is true; is the output of the discriminator to the generator and condition information The output probability of , the expected value is 0, indicating that the demand type is generated data; The loss function of the generator is expressed as follows: Where, represents the loss function of the generator, Represents the loss term of the generator's classification of the rental demand type for the s-th sample in the data sample set, which is used to measure the difference between the generated rental demand type prediction and the true label; represents the adversarial loss term for the sth sample in the data sample set, indicating the possibility that the generated data is judged as real by the discriminator; λ1 and λ2 represent the corresponding loss term hyperparameters; Represents a generator based on x (s) and condition information The generated predicted probability of rental demand type l.

7. The method according to any one of claims 3 to 6, wherein Calculating a matching degree relative to the housing demand matrix based on the housing description data of the candidate housing and the corresponding travel distance includes: Concatenate the property description data and the corresponding travel distance to generate the corresponding property feature vector: F prop =[f1,f2,…,f G-1 ,d distance ], Where f1, f2, …, f G-1 Represents the information characteristics of each dimension in the property description data, d distance Represents the normalized characteristic of travel distance, F prop represents the property feature vector; A bilinear pooling matching operation is performed on the housing source feature vector and the rental demand matrix to obtain a matching result matrix through interactive feature analysis: Where, F represents the prediction confidence of the qth rental demand type in the rental demand matrix; prop is the property feature vector, containing G-dimensional features; represents the g-th eigenvalue in the property feature vector, where the eigenvalue of the last dimension represents the travel distance feature; W (q,g) represents the weight parameter of bilinear pooling, which is used to control the interaction between rental demand characteristics and housing supply feature vectors; d distance represents the normalized travel distance feature; W (q,G) Represents the weight parameter of the travel distance feature, which is used to control the importance of the travel distance feature in the matching calculation; μ is the additional weight coefficient of the travel distance feature, and η is the hyperparameter for adjusting the rate of change of the travel distance weight; P match Represents the matching result matrix; C qg Score the correlation between demand type q and the g-th eigenvalue in the property feature vector; is a normalized term, representing the sum of the correlation scores of all demand and housing feature vectors in the rental demand matrix; The matching result matrix is ​​projected using feature compression and linear projection to obtain the matching degree: S match =s mat (W proj ·vec(P match )+b proj ), Where S match represents the matching degree of candidate housing, σ mat represents the nonlinear activation function, W proj is the projection matrix, which is used to map the high-dimensional matching matrix to the low-dimensional space; vec(·) represents the operation of expanding the matrix into a vector; b proj Represents the bias vector.

8. A rental recommendation system based on multimodal data fusion analysis, comprising: The traffic data acquisition unit is used to obtain the distribution channel information, user geographic location and distribution material attribute data of the corresponding new user when a user triggers the distribution of materials on the rental platform; the distribution material attribute data includes the type of material preference group and material housing description information; a rental demand prediction unit, configured to input the delivery channel information and the delivery material attribute data into a rental demand prediction model to output a plurality of rental demand types and corresponding prediction confidences, thereby constructing a rental demand matrix corresponding to the new users attracted; wherein each row of the rental demand matrix represents a rental demand type, and each column represents a corresponding prediction confidence; a candidate listing screening unit, configured to screen a set of candidate listings from a listing database based on the user's geographic location, wherein the travel distance between the geographic location indicated by the candidate listing and the user's geographic location is less than a preset threshold; a housing matching calculation unit, configured to calculate, for each candidate housing source, a matching degree relative to the housing rental demand matrix based on the housing description data of the candidate housing source and the corresponding travel distance; A house recommendation unit for attracting new users, configured to generate a house recommendation list for the new users according to a preset number of candidate houses ranked by corresponding matching degrees; The rental demand prediction model adopts a conditional generative adversarial network, which includes a generator and a discriminator; wherein the generator includes an input layer, a fusion layer, a feature decomposition layer and an output layer: The input layer is used to embed the input delivery channel information and delivery material attribute data into a unified vector space through multimodal feature encoding to generate corresponding coding features of each modality; The fusion layer is used to fuse the coding features of each modality to determine the corresponding multimodal fusion features; The feature decomposition layer is used to perform hierarchical processing on the multimodal fusion features according to condition information to determine a personalized demand matrix for attracting new users at the coarse-grained and refined demand levels; the condition information is defined based on the delivery channel information, the delivery material attribute data, and a preset set of rental demand types; The output layer is used to output multiple rental demand types and corresponding prediction confidence levels according to the personalized demand matrix.

Images