An explainable online short-term rental recommendation method integrating trust and explicit preference

Abstract

The present invention provides an explainable online short-term rental recommendation method integrating trust and explicit preference. The method constructs a heterogeneous preference recognition model, and constructs a training set by acquiring a historical user set, a historical housing source set and a transaction record set of an online short-term rental platform to train the heterogeneous preference recognition model. Compared with the prior art, the present invention effectively realizes the separation of explicit heterogeneous preferences at the feature level by disentangling the entangled representations of landlords and housing sources in the heterogeneous preference recognition model, and can better explore the user's preferences for the selection of houses and landlords. At the same time, by obtaining the importance explanation of each item-level representation and each preference representation, a more practical and feasible housing source evaluation reference can be provided to the landlord; at the same time, loss calculation is set for each module in the heterogeneous preference recognition model, and while identifying heterogeneous preferences, the recommendation effect is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of house representation learning, and more specifically, to an explainable online short-term rental recommendation method integrating trust and explicit preference. Background Art

[0002] The sharing economy provides new and flexible employment opportunities where individuals can provide and acquire basic skills and services. Service or item providers in the sharing economy are often referred to as "micro-entrepreneurs" whose livelihoods depend on how they interact with algorithms. Shared accommodation is an indispensable subset of the sharing economy. The explicit order of different listings in the recommendation list may affect their likelihood of being booked. However, hosts know little about the factors that affect the recommendation list. Hosts can only improve their work intuitively, which leads to uncertainty, frustration and anxiety in the hosts' working environment. Therefore, while making recommendations, an explainable shared accommodation recommendation method that can show the basis for the recommendation and the factors considered by the algorithm is crucial.

[0003] Most landlords do not understand the algorithm and believe that the algorithm will evaluate them based on factors that they have little control over, such as location and noise. Therefore, some landlords have proposed that "super houses" and "super landlords" should be two separate things. If the landlord is very good, but the location, noise or other aspects of the apartment or building are problematic, the landlord should also be able to become a "super landlord". Therefore, explaining the user's preference for houses and the preference for landlords is crucial to alleviating landlord anxiety.

[0004] Existing explainable recommendation methods focus on explaining the importance of certain features or aspects. However, most explainable recommendations are currently considered from the perspective of buyers, and are not sufficiently explainable to sellers. Due to the existence of heterogeneous preferences of users, some features contain information about heterogeneous preference entanglement. Explaining these preference entanglement features cannot provide specific guidance for the work of landlords. Therefore, it is crucial to identify feature-level heterogeneous preferences from the features of heterogeneous preference entanglement and explain the actual impact of house features in recommendations under specific preferences. Summary of the invention

[0005] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide an explainable online short-term rental recommendation method that integrates trust and explicit preference, so as to effectively separate feature-level explicit heterogeneous preferences in online short-term rentals and realize the interpretation of feature importance.

[0006] The technical solution adopted by the present invention is:

[0007] The present invention provides an explainable online short-term rental recommendation method integrating trust and explicit preference, the method comprising:

[0008] Obtain a historical user set, a historical property set and a transaction record set of online short-term rentals, construct a training set based on the historical user set, the historical property set and the transaction record set, and obtain a historical transaction property sequence for each user based on the transaction record set; the historical transaction property sequence represents the properties that the user has historically stayed in and their order;

[0009] For all the houses in the historical house collection, obtain the house feature representation vector and the landlord feature representation vector of each house according to the house information and the landlord information of the house, and extract the entangled feature representation vector of each house according to the relevant comments of the user on the house, and construct the house feature representation set, landlord feature representation set and entangled feature representation set of all houses;

[0010] Constructing a heterogeneous preference recognition model, training the heterogeneous preference recognition model according to the house feature representation set, the landlord feature representation set and the entanglement feature representation set of the training set, and obtaining the trained heterogeneous preference recognition model;

[0011] The heterogeneous preference identification model includes a representation interpretation module, an entangled representation decomposition module, a representation fusion module, a user preference module and a score calculation module;

[0012] The representation interpretation module is used to obtain an interpretable weight of each representation vector in the representation set according to the input representation set, and obtain an item-level representation set corresponding to the input representation set according to the interpretable weight;

[0013] The entangled representation decomposition is used to disentangle the input representation set and decompose the input representation set into a plurality of corresponding item-level representation sets;

[0014] The representation fusion module is used to splice and fuse several input representation sets;

[0015] The user preference module is used to obtain the preference representation vector of the corresponding user according to the relationship between the representation vectors in the input representation set, and construct the preference representation set of each user;

[0016] The score calculation module is used to calculate the matching score between the user's preference representation set and each item-level representation vector in the corresponding item-level representation set;

[0017] The preference fusion interpretation module is used to splice the input representation set and obtain the interpretable preference weight of the spliced ​​representation set;

[0018] Obtain a target user set and a target housing set, input the target user set and the target housing set into the trained heterogeneous preference recognition model, obtain the matching score output by the score calculation module and recommend it to the user;

[0019] The interpretable weights output by the representation interpretation module and the interpretable preference weights output by the preference fusion interpretation module are respectively obtained, and the interpretable weights and interpretable preference weights are respectively converted into corresponding evaluation scores and recommended to the landlord.

[0020] Furthermore, the training of the heterogeneous preference recognition model according to the house feature representation set, the landlord feature representation set and the entanglement feature representation set of the training set specifically includes:

[0021] Obtaining, through the representation interpretation module, item-level house item representation sets, landlord item representation sets, and entangled item representation sets corresponding to the house feature representation sets, landlord feature representation sets, and entangled feature representation sets;

[0022] Decomposing the entangled item representation set into an item-level first item representation set related to the house, a second item representation set related to the landlord, and a shared item representation set related to both the house and the landlord by the entangled representation decomposition module;

[0023] The house item representation set, the first item representation set and the shared item representation set are fused by the representation fusion module to obtain a house fused item representation set, and the landlord item representation set, the second item representation set and the shared item representation set are fused to obtain a landlord fused item representation set;

[0024] Based on the house item representation set, the landlord item representation set, the first item representation set, the second item representation set, the house fusion item representation set and the landlord fusion item representation set, respectively extract the item-level house representation set, the landlord representation set, the first representation set, the second representation set, the house fusion representation set and the landlord fusion representation set corresponding to each user through the historical transaction house source sequence;

[0025] Input the house representation set, landlord representation set, first representation set, second representation set, house fusion representation set and landlord fusion representation set corresponding to each user into the user preference module, and obtain the house preference representation set, landlord preference representation set, house feature-related preference representation set, landlord feature-related preference representation set, house fusion preference representation set and landlord fusion preference representation set of each user respectively;

[0026] The score calculation module calculates the matching scores of the house preference representation set, the landlord preference representation set, the house feature-related preference representation set and the landlord feature-related preference representation set of each user and each item-level representation vector in the corresponding item-level representation set respectively;

[0027] Input the house fusion preference representation set and the landlord fusion preference representation set, as well as the house fusion representation set and the landlord fusion representation set into the preference fusion interpretation module for splicing, and obtain the user's interpretable preference weights for the house and the landlord respectively through the spliced ​​representation sets;

[0028] Respectively calculating the disentanglement regularization loss between the first item representation set and the house item representation set and the landlord item representation set, and between the second item representation set and the house item representation set and the landlord item representation set, and obtaining the first loss of the heterogeneous preference recognition model according to the disentanglement regularization loss;

[0029] Respectively calculating the orthogonal constraint losses of the shared item representation set and the house item representation set, the landlord item representation set, the first item representation set, and the second item representation set, and obtaining the second loss of the heterogeneous preference recognition model according to the orthogonal constraint losses;

[0030] Calculate a first BPR loss according to the matching scores of the house preference representation set, the landlord preference representation set, the house feature-related preference representation set and the landlord feature-related preference representation set of the corresponding user and each item-level representation vector in the corresponding item-level representation set, and obtain a third loss of the heterogeneous preference recognition model according to the first BPR loss;

[0031] Calculate a second BPR loss according to the user's interpretable preference weights for the house and the landlord, and the matching scores of each item-level representation vector in the corresponding item-level representation set, and obtain a fourth loss of the heterogeneous preference recognition model according to the second BPR loss;

[0032] Weighting the first loss, the second loss, the third loss and the fourth loss to obtain the final loss of the heterogeneous preference identification model;

[0033] Reversely updating the heterogeneous preference recognition model according to the final loss, and finally obtaining the trained heterogeneous preference recognition model;

[0034] Obtain a target user set and a target housing set, input the target user set and the target housing set into the trained heterogeneous preference recognition model, obtain the matching score output by the score calculation module and recommend it to the user;

[0035] The interpretable weights output by the representation interpretation module and the interpretable preference weights output by the preference fusion interpretation module are respectively obtained, and the interpretable weights and interpretable preference weights are respectively converted into corresponding evaluation scores and recommended to the landlord.

[0036] The present invention constructs the heterogeneous preference recognition model, and constructs a training set by obtaining a historical user set, a historical house source set and a transaction record set of an online short-term rental platform. In the process of constructing the training set, the entangled feature representation vector of each house source is obtained by analyzing the relevant comments of the user on the house source. In the training process, the entangled representation decomposition module set in the heterogeneous preference recognition model is used to represent the features only related to the house, the features only related to the landlord and the features related to both the house and the landlord. By disentangling the feature representations, the separation of explicit heterogeneous preferences at the feature level is effectively achieved, and the user's preference for the selection of houses and landlords can be better explored.

[0037] At the same time, by setting the representation interpretation module in the heterogeneous preference recognition model, in the process of converting the input representation set into the corresponding item-level representation set, the interpretable weight of each representation in the input item-level representation set is obtained, and the interpretable weight represents the importance of the corresponding feature to the user's housing selection, and by setting the preference fusion interpretation module in the heterogeneous preference recognition model, the user's preference representation set is interpreted for importance, and the interpretable preference weight of the preference representation set is obtained. The interpretable preference weight represents the user's preference for the representation of the corresponding housing source. By obtaining the interpretable weight and the interpretable preference weight and recommending them to the landlord, the landlord can have a clearer understanding of the situation of his own housing source among users, which can effectively alleviate the landlord's anxiety and thereby improve his own housing source situation; conversely, by setting the score calculation module in the heterogeneous preference recognition model to calculate the matching scores of each preference representation set of the user and each item-level representation vector, the recommendation of each feature representation of the house and the landlord is reflected, providing a reference for the user's choice of housing.

[0038] In addition, it is equivalent to setting a corresponding loss calculation for each module in the heterogeneous preference recognition model, which can better improve the accuracy of the recognition calculation of the corresponding module.

[0039] Furthermore, the represented interpretation module includes a plurality of first attention pooling layers arranged in parallel;

[0040] The first attention pooling layer is used to parse the representation vectors in the input representation set according to the input representation set and using the attention pooling mechanism, obtain the interpretable weight of each representation vector, and perform weighted concatenation on each representation vector according to the interpretable weight to obtain the item-level representation set corresponding to the input representation set;

[0041] It manifests as:

[0042]

[0043]

[0044]

[0045] In the formula, Representation set representing the input The lth characterization vector in Represents the representation set The interpretable weight calculated from the lth representation vector in, represents the concatenation function, Represents the representation set The vector after weighted concatenation of all representation vectors of The output corresponds to the representation set The item-level representation set, is the corresponding input representation set The parameter vector to be learned of the first attention pooling layer represents the overall representation set; is the corresponding input representation set The parameter vector to be learned of the first attention pooling layer, , is the weight matrix to be learned, R is the matrix identifier, and Indicates that the dimensions of the matrix are dx1 and dxd respectively.

[0046] The representation interpretation module captures the importance of each representation vector and the house in the input representation set through the attention mechanism, and interprets the house through the corresponding representation vector through the importance, which can provide an intuitive reference for the landlord. The number of settings of the first attention pooling layer corresponds to the number of representation sets that need to be input.

[0047] Furthermore, the user preference module includes a plurality of preference networks arranged in parallel; the preference network includes a multi-head self-attention layer and a second attention pooling layer;

[0048] The multi-head self-attention layer is used to obtain the relationship between each representation vector in the input representation set by using the self-attention mechanism, and enhance the corresponding representation vector through the relationship;

[0049] The second attention pooling layer is used to parse each enhanced representation vector by using the attention mechanism, obtain the preference weight of each enhanced representation vector, and perform weighted concatenation on each enhanced representation vector according to the preference weight to obtain a preference representation set of the corresponding representation set;

[0050] It manifests as:

[0051]

[0052]

[0053]

[0054]

[0055] In the formula, and Respectively represent the input item-level representation set for user u The i-th characterization vector and the j-th characterization vector in is the corresponding input representation set The weight of the multi-head self-attention layer in the preference network represents the relationship between the i-th representation vector and the j-th representation vector, The input representation set The enhanced representation vector of the i-th representation vector, is the corresponding input representation set The weight matrix to be learned of the multi-head attention layer in the preference network, R is the matrix identifier, Indicates that the dimension of the matrix is ​​dxd;

[0056] is the corresponding input representation set The weight of the second attention pooling layer in the preference network represents the importance of the i-th and j-th enhanced representation vectors to the user u; is the corresponding input representation set The preference representation set output by the preference network; is the parameter vector to be learned in the preference network, representing the entire historical transaction house sequence; represents the concatenation function; represents the natural exponential function; Represents the transpose of a matrix; and Respectively represent the corresponding input representation set The parameter vector and weight matrix to be learned of the second attention pooling layer in the preference network, R is the matrix identifier, and Indicates that the dimensions of the matrix are dx1 and dxd respectively.

[0057] Furthermore, the house fusion preference representation set and the landlord fusion preference representation set, as well as the house fusion representation set and the landlord fusion representation set are input into the preference fusion interpretation module for splicing, and the user's interpretable preference weights for the house and the landlord are obtained through the spliced ​​representation sets, wherein the user's interpretable preference weights for the house are obtained as follows:

[0058]

[0059] The user's interpretable preference weights for the landlord are obtained as follows:

[0060]

[0061] In the formula, They represent the house fusion preference representation set and the landlord fusion preference representation set respectively, They represent the house fusion representation set and the landlord fusion representation set respectively, They represent the parameter vectors to be learned, represents the weight matrix to be learned, R is the matrix identifier, and Respectively indicate that the dimensions of the matrix are dx1 and dxd; Indicates concatenation of input parameters; Represents the transpose of a matrix.

[0062] The interpretable preference weights of the input preference representation set are obtained through the preference fusion interpretation module, and the user's preference for each property feature is represented by the interpretable preference weights, which can further provide a reference for the landlord.

[0063] Furthermore, the score calculation module is used to calculate the matching score between the user's preference representation set and each item-level representation in the corresponding item-level representation set. The specific formula is as follows:

[0064]

[0065] In the formula, represents the i-th representation vector in the item-level representation set output by the corresponding representation interpretation module or representation fusion module; represents the preference representation set of user u output by the corresponding user preference module, Represents the transpose of a matrix.

[0066] That is, the score calculation module is used to obtain the match between the characteristics of each house and the user's preferences, providing explainable information for the user's choice of house and landlord.

[0067] Furthermore, the disentanglement regularization loss between the first item representation set and the house item representation set and the landlord item representation set, and between the second item representation set and the house item representation set and the landlord item representation set are calculated respectively, and the corresponding disentanglement regularization loss is calculated respectively by the following formula:

[0068]

[0069]

[0070]

[0071]

[0072] In the formula, is the first item representation set, is the second item representation set, is a collection of house item representations, A collection of listings for said landlord; represents the distance covariance between two matrices, represents the distance variance of the matrix;

[0073] The first loss of the heterogeneous preference identification model is obtained according to the regularization loss, and is specifically calculated by the following formula:

[0074]

[0075] In the formula, Represents the first loss.

[0076] By calculating the correlation between the first item representation set and the second item representation set obtained by de-entanglement and the house item representation set and the landlord item representation set respectively to construct the first loss, the de-entanglement effect of the entanglement representation decomposition module can be well improved according to the first loss.

[0077] Furthermore, the orthogonal constraint losses of the shared item representation set and the house item representation set, the landlord item representation set, the first item representation set, and the second item representation set are calculated respectively, and the corresponding orthogonal constraint losses are calculated respectively by the following formulas:

[0078]

[0079]

[0080]

[0081]

[0082] In the formula, represents the square of the F norm, is the first item representation set, is the second item representation set, is a collection of house item representations, is a collection of landlord listings, is the shared item representation set, Represents the transpose of a matrix;

[0083] The second loss of the heterogeneous preference identification model is obtained according to the orthogonal constraint loss, and is specifically calculated by the following formula:

[0084]

[0085] In the formula, Represents the second loss.

[0086] In order to better decompose the entangled item representation set into a first item representation set related to house characteristics, a second item representation set related to landlord characteristics, and a shared item representation set related to both the house and the landlord, orthogonal constraints are added between the shared item representation set and the house item representation set, the landlord item representation set, the first item representation set, and the second item representation set, respectively, to further improve the de-entanglement effect of the entangled representation decomposition module.

[0087] Further, the first BPR loss is calculated based on the matching scores of the house preference representation set, the landlord preference representation set, the house feature-related preference representation set and the landlord feature-related preference representation set of the corresponding user and each item-level representation vector in the corresponding item-level representation set. Specifically, the corresponding first BPR losses are calculated respectively by the following formulas:

[0088]

[0089]

[0090]

[0091]

[0092] In the formula, represents the training set, represents the positive sample set of user u in the historical user set in the historical house set based on the transaction record set, represents a negative sample set of user u in the historical house set based on the transaction record set;

[0093] The third loss of the heterogeneous preference identification model is obtained according to the first BPR loss, and is specifically calculated by the following formula:

[0094]

[0095] In the formula, represents the third loss.

[0096] By obtaining the corresponding first BPR loss by matching the corresponding user preferences with the corresponding item-level representation vectors in the item-level representation set, and constructing the third loss, the representation interpretation capability of the representation interpretation module for each item-level representation set can be improved, as well as the recognition capability of the user preference module for the user preference representation.

[0097] Furthermore, the second BPR loss is calculated according to the user's interpretable preference weights for the house and the landlord, and the matching scores of each item-level representation in the corresponding item-level representation set, specifically including:

[0098] The matching scores of the house fusion preference representation set of each user and the representation vectors of each item level in the house fusion representation set, as well as the matching scores between the landlord fusion preference representation set of each user and the landlord fusion representation set are obtained respectively. The specific formula is:

[0099]

[0100]

[0101] In the formula, and denote the i-th representation vector in the house fusion representation set and the landlord fusion representation set respectively, and They respectively represent the house fusion preference representation set and the user-landlord fusion preference representation set obtained based on user u;

[0102] Get the user's interpretable preference weights for houses and interpretable preference weights for landlords ;

[0103] The user preference matching score is calculated using the following formula:

[0104]

[0105] The corresponding second BPR loss is calculated by the following formulas:

[0106]

[0107]

[0108]

[0109] In the formula, represents the training set, represents the positive sample set of user u in the historical user set in the historical house set based on the transaction record set, represents a negative sample set of user u in the historical house set based on the transaction record set;

[0110] The fourth loss of the heterogeneous preference identification model is obtained according to the second BPR loss, and is specifically calculated by the following formula:

[0111]

[0112] In the formula, represents the fourth loss.

[0113] By obtaining the matching scores of the house fusion preference representation set of each user and the representation vectors of each item level in the house fusion representation set, and correspondingly calculating the user's explainable preference weight for the house and the user's explainable preference weight for the landlord, the representation interpretation module can better obtain the preference relationship between the user's various representations of the house and the various representations of the landlord.

[0114] Compared with the prior art, the present invention has the following beneficial effects:

[0115] 1. The present invention identifies and recommends the matching between housing resources and users by constructing a heterogeneous preference recognition model. In the process of analyzing the representation of housing resources, the entangled representation decomposition module set in the heterogeneous preference recognition model is used to disentangle the entangled representations related to landlords and housing resources, effectively realizing the separation of explicit heterogeneous preferences at the feature level, and better mining the user's preferences for housing and landlords;

[0116] 2. The present invention converts the feature-level representation of the house into the corresponding item-level representation through the representation interpretation module set in the heterogeneous preference recognition model, so as to facilitate the heterogeneous preference model to recognize the item-level representation of the user for the house, and obtains the interpretable weights that express the importance of each item-level representation in the conversion process, and realizes the importance interpretation of each representation of the house through the interpretable weights; at the same time, through the preference fusion interpretation module set in the heterogeneous preference recognition model, the importance interpretation of the preference representation set is performed, and the interpretable preference weights that express the user's preference for each house representation are obtained, so as to provide a reference for the landlord for the value and rectification of his own house through the interpretable weights and interpretable preference weights;

[0117] 3. In the training process of the heterogeneous preference recognition model, the present invention is equivalent to setting a loss calculation for each module therein, and in the training process, the parameters of the corresponding module can be updated through the loss calculation, thereby improving the accuracy of the recognition calculation of the corresponding module. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1It is a diagram of the method steps of the present invention.

[0119] Figure 2 It is a structural diagram of the heterogeneous preference model of the present invention.

[0120] Figure 3 A diagram showing the training steps of the heterogeneous preference model of the present invention. DETAILED DESCRIPTION

[0121] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0122] Example 1

[0123] like Figure 1 As shown, this embodiment provides an explainable online short-term rental recommendation method integrating trust and explicit preference, and the method includes:

[0124] S1: Get the historical user set of online short-term rentals 、Historical Listings Collection and a transaction record set, construct a training set based on the historical user set, historical property set and transaction record set, and obtain the historical transaction property sequence of each user u based on the transaction record set ;

[0125] in, represents the number of users in the historical user set, Indicates the number of listings in the historical listings set. is the number of houses in the historical transaction house source sequence of the corresponding user. In this embodiment, the house source includes house information and information about the landlord to whom the house belongs. The historical transaction house source sequence is a sequence of rental situations of each user in the historical user set in the historical house source set, such as represents the first house v rented by user u, Indicates the second house v rented by user u, and so on.

[0126] S2: For the historical housing collection For all the houses in the list, we obtain the house feature representation vector and the landlord feature representation vector of each house according to the house information and the landlord information of each house, and extract the entangled feature representation vector of each house according to the user's comments on the house, and construct the house feature representation set of all houses. , Landlord Feature Set and entanglement feature representation set ;

[0127] In this step, for all the houses in the historical house collection, feature extraction is first performed based on the house information and landlord information of each house, wherein the extracted features include the trust features of the user in the house and the landlord. Specifically, for the feature extraction of the house, the number of people who can live in the house, the rental price of the house, the type of house, the supporting facilities in the house, the area of ​​the house and the number of rooms, etc., can be considered as features; for the feature extraction of the landlord, the gender and expression of the landlord's avatar, as well as the landlord's reputation, platform honors and landlord verification information, etc. can be considered as features; the specific settings will be based on the corresponding public information of the online short-term rental platform.

[0128] Since users' evaluation of listings usually includes evaluation of the house and the landlord, the users' comments on the listings are used as entangled features of the house and the landlord, including users' ratings of the house and the landlord, the number of users' comments, the sentiment polarity scores of different granularity sentences in user comments, the relevance of comments to the listing, and the complete text of user comments.

[0129] After obtaining the house features, landlord features, and entanglement features, they need to be represented and vectorized. Specifically, the corresponding features can be converted into a set of embedded vectors of the same dimension through the embedding layer, and finally the house feature representation set is obtained. , Landlord Feature Set and entanglement feature representation set ,in, represents the bth feature representation vector in the house feature representation set, B is the number of feature representation vectors in the house feature representation set, represents the mth feature representation vector in the landlord feature representation set, M is the number of feature representation vectors in the landlord feature representation set, represents the nth feature characterization vector in the entangled feature characterization set, and N is the number of feature characterization vectors in the entangled feature characterization set.

[0130] S3: constructing a heterogeneous preference recognition model, training the heterogeneous preference recognition model, and obtaining the trained heterogeneous preference recognition model;

[0131] like Figure 2 As shown, the heterogeneous preference identification model includes a representation interpretation module, an entangled representation decomposition module, a representation fusion module, a user preference module, a score calculation module and a preference fusion interpretation module;

[0132] The representation interpretation module is used to obtain an interpretable weight of each representation vector in the representation set according to the input representation set, and obtain an item-level representation set corresponding to the input representation set according to the interpretable weight;

[0133] Specifically, the represented interpretation module includes several first attention pooling layers set in parallel;

[0134] The first attention pooling layer is used to parse the representation vectors in the input representation set according to the input representation set and using the attention pooling mechanism, obtain the interpretable weight of each representation vector, and perform weighted concatenation on each representation vector according to the interpretable weight to obtain the item-level representation set corresponding to the input representation set;

[0135] The number of the first attention pooling layers is set as required by inputting the representation interpretation module to obtain the corresponding interpretable weights or outputting the corresponding item-level representation sets, and each of the first attention pooling layers corresponds to an input representation set;

[0136] It manifests as:

[0137]

[0138]

[0139]

[0140] In the formula, Representation set representing the input The lth characterization vector in Represents the representation set The interpretable weight calculated from the lth representation vector in, represents the concatenation function, Represents the representation set The vector after weighted concatenation of all representation vectors of The output corresponds to the representation set The item-level representation set, is the corresponding input representation set The parameter vector to be learned of the first attention pooling layer represents the overall representation set; is the corresponding input representation set The parameter vector to be learned of the first attention pooling layer, , is the weight matrix to be learned, R is the matrix identifier, and Indicates that the dimensions of the matrix are dx1 and dxd respectively.

[0141] The entangled representation decomposition module is used to disentangle the input representation set and decompose the input representation set into a plurality of corresponding item-level representation sets.

[0142] Specifically, the entangled representation set includes several decomposition network layers, each of which receives an input representation set and inputs the input representation set into a decomposition network layer for processing, which is shown as follows:

[0143]

[0144] In the formula, represents the input representation set, represents the representation set corresponding to the output of the decomposition network layer, and They represent the parameter vector and weight matrix to be learned of the corresponding i-th decomposition network layer, R is the matrix identifier, and Respectively indicate that the dimensions of the matrix are dx1 and dxd;.

[0145] In this embodiment, since it is necessary to disentangle the entangled features of the house and the landlord and decompose the entangled features into features related only to the landlord, features related only to the house source, and features related to both the house source and the landlord, the decomposition network layer in the entanglement representation decomposition module in this embodiment is set to at least 3, and each decomposition network layer corresponds to obtaining a decomposed feature.

[0146] The representation fusion module is used to splice and fuse several input representation sets; specifically, the representation fusion module includes at least one feature connection network, and when the feature connection networks are set to be no less than two, each of the feature connection networks is set in parallel; each of the feature connection networks receives a group of representation sets that need to be spliced ​​and fused, and splices and fuses all the representation sets in the group;

[0147] The characteristic connection network is shown as follows:

[0148]

[0149] In the formula, represents the fused item-level representation set, Represents the input representation set To splice, A learnable weight matrix that connects the network to the corresponding features.

[0150] In this embodiment, in order to better obtain the user's preference for houses and landlords, it is necessary to fuse the house-related representation set and the landlord-related representation set respectively, so at least two feature connection networks are set in the representation fusion module in this implementation.

[0151] The user preference module is used to obtain the preference representation vector of the corresponding user according to the relationship between the representation vectors in the input representation set, and construct the preference representation set of each user;

[0152] Specifically, the user preference module includes several preference networks arranged in parallel; the preference network includes a multi-head self-attention layer and a second attention pooling layer;

[0153] The multi-head self-attention layer is used to obtain the relationship between each representation vector in the input representation set by using the self-attention mechanism, and enhance the corresponding representation vector through the relationship;

[0154] The second attention pooling layer is used to parse each enhanced representation vector by using the attention mechanism, obtain the preference weight of each enhanced representation vector, and perform weighted concatenation on each enhanced representation vector according to the preference weight to obtain a preference representation set of the corresponding representation set;

[0155] The preference network is specifically manifested as follows:

[0156]

[0157]

[0158]

[0159]

[0160] In the formula, and Respectively represent the input item-level representation set for user u The i-th characterization vector and the j-th characterization vector in is the corresponding input representation set The weight of the multi-head self-attention layer in the preference network represents the relationship between the i-th representation vector and the j-th representation vector, The input representation set The enhanced representation vector of the i-th representation vector, is the corresponding input representation set The weight matrix to be learned of the multi-head attention layer in the preference network, R is the matrix identifier, Indicates that the dimension of the matrix is ​​dxd;

[0161] is the corresponding input representation set The weight of the second attention pooling layer in the preference network represents the importance of the i-th and j-th enhanced representation vectors to the user u; is the corresponding input representation set The preference representation set output by the preference network; is the parameter vector to be learned in the preference network, representing the entire historical transaction house sequence; represents the concatenation function; represents the natural exponential function; Represents the transpose of a matrix; and Respectively represent the corresponding input representation set The parameter vector and weight matrix to be learned of the second attention pooling layer in the preference network, R is the matrix identifier, and Indicates that the dimensions of the matrix are dx1 and dxd respectively.

[0162] In order to more comprehensively obtain the user's preference for each item-level representation set, a plurality of preference networks are provided to match the number of representation sets that need to be analyzed.

[0163] The preference fusion interpretation module is used to concatenate the input representation set and obtain the interpretable preference weight of the concatenated representation set. ;

[0164] The specific manifestations are:

[0165]

[0166] In the formula, Represents the input representation set To splice, They represent the parameter vectors to be learned, represents the weight matrix to be learned, R is the matrix identifier, and Respectively indicate that the dimensions of the matrix are dx1 and dxd;

[0167] In this embodiment, the preference fusion interpretation module is used to obtain the user's preference for the housing source representation, so the representation set input into the preference fusion interpretation module needs to include a preference representation set.

[0168] The score calculation module is used to calculate the matching score between the user's preference representation set and each item-level representation vector in the corresponding item-level representation set;

[0169] The specific manifestations are:

[0170]

[0171] In the formula, represents the i-th representation vector in the item-level representation set output by the corresponding representation interpretation module or representation fusion module; represents the preference representation set of user u output by the corresponding user preference module, Represents the transpose of the matrix. That is, the score calculation module is used to obtain the matching between the characteristics of each house and the user's preferences, providing explainable information for the user's choice of house and landlord.

[0172] like Figure 3 As shown, the training of the heterogeneous preference recognition model is specifically the training of each module in the above-mentioned heterogeneous preference recognition model, and the training steps specifically include:

[0173] A1: Obtaining the house feature representation set through the representation interpretation module , Landlord Feature Set and entanglement feature representation set Corresponding item-level housing item representation set , Landlord Item Listing Collection and the entangled object representation set ;

[0174] Specifically, the house feature representation set , Landlord Feature Set and entanglement feature representation set Input to the corresponding first attention pooling layer in the representation interpretation module. In this embodiment, since the representation interpretation module needs to process the house feature representation set , Landlord Feature Set and entanglement feature representation set Therefore, in this embodiment, at least three first attention pooling layers are set, and each of the first attention pooling layers processes a representation set respectively, which are respectively expressed as follows:

[0175] For the house feature representation set :

[0176]

[0177]

[0178]

[0179] Represents the set of house features The interpretable weight obtained from the b-th feature representation vector;

[0180] For the landlord feature representation set :

[0181]

[0182]

[0183]

[0184] Represents the landlord feature set The interpretable weight obtained from the m-th feature representation vector;

[0185] For the entanglement feature representation set :

[0186]

[0187]

[0188]

[0189] Denotes the set represented by the entanglement feature The interpretable weight obtained from the nth feature representation vector;

[0190] A2: The entangled object representation set is decomposed by the entangled representation decomposition module. Decomposed into the first set of house-related item representations at the item level , the second set of item representations related to the landlord and a set of shared item representations related to both the house and the landlord ;

[0191] Specifically, the entangled item representation set The first item representation set is respectively input into the three decomposition network layers in the entanglement representation decomposition module, and is calculated to obtain the first item representation set. , the second item representation set and shared item representation set , specifically:

[0192]

[0193]

[0194]

[0195] In the formula, and are the parameter vector and weight matrix to be learned for the corresponding decomposition network layer, R is the matrix identifier, and The dimensions of the matrices are dx1 and dxd respectively. The entangled item representation set includes representations affected by the house and the landlord. In order to better identify the representations of the house and the landlord, the entangled item representation set needs to be disentangled.

[0196] A3: The house object representation set is integrated by the representation fusion module , the first item representation set and shared item representation set Fusion to obtain the house fusion item representation set , and the landlord item representation is aggregated , the second item representation set and shared item representation set Fusion to obtain the landlord fusion item representation set ;

[0197] Specifically, the house item representation is collected , the first item representation set and shared item representation set As a set of inputs into one of the feature connection networks in the representation fusion module, the house fusion item representation set is calculated to obtain , collect the landlord item representation , the second item representation set and shared item representation set As another set of inputs, another feature connection network in the representation fusion module is calculated to obtain the landlord fusion item representation set , specifically:

[0198]

[0199]

[0200] In the formula, and They are respectively the learnable weight matrices of the corresponding feature connection network.

[0201] The house fusion item representation set Represents the set of all item-level representations related to the house. Correspondingly, the house fusion item representation set It represents the set of all item-level representations related to the landlord.

[0202] A4: Based on the house item representation set , Landlord Item Listing Collection , the first item representation set , the second item representation set , House Fusion Item Representation Set Merge item representation sets with landlords , through the historical transaction listing sequence Extract the item-level house representation set corresponding to each user separately , Landlord List Collection , the first representation set , the second representation set , House Fusion Representation Collection and the Landlord Fusion Representation Collection ;

[0203] Specifically, in this step, for any user u in the historical user set U, according to the historical transaction house sequence Get all the houses that user u has rented, and match the houses to the house item representation set , Landlord Item Listing Collection , the first item representation set , the second item representation set , House Fusion Item Representation Set Merge item representation sets with landlords The representation vector in is integrated into the house representation set of the user u , Landlord List Collection , the first representation set , the second representation set , House Fusion Representation Collection and the Landlord Fusion Representation Collection In the process, the user-side conversion of each item-level representation of the house is realized, which facilitates the identification of user preferences.

[0204] A5: Collect the house representations corresponding to each user , Landlord List Collection , the first representation set , the second representation set , House Fusion Representation Collection and the Landlord Fusion Representation Collection Input the user preference module to obtain each user's housing preference representation set , Landlord Preference Set , House feature related preference representation set , landlord feature-related preference representation set , Housing Fusion Preference Representation Set and landlords fused preference representation set ;

[0205] Specifically, the house representation corresponding to each user is collected , Landlord List Collection , the first representation set , the second representation set , House Fusion Representation Collection and the Landlord Fusion Representation Collection The preference networks of the user preference modules are respectively input. In this embodiment, in order to process the input representation sets respectively, the number of the user preference modules of this embodiment is not less than 6, and each preference network processes the input representation set respectively, which is specifically manifested as follows:

[0206] For the housing representation set :

[0207]

[0208]

[0209]

[0210]

[0211] For the Landlord Representation Set :

[0212]

[0213]

[0214]

[0215]

[0216] For the first representation set :

[0217]

[0218]

[0219]

[0220]

[0221] For the second representation set :

[0222]

[0223]

[0224]

[0225]

[0226] For the housing fusion representation set :

[0227]

[0228]

[0229]

[0230]

[0231] For the landlord fusion representation set :

[0232]

[0233]

[0234]

[0235]

[0236] Through the user preference module, the item-level representation sets of all users are converted into preference representation sets for each item feature.

[0237] A6: Calculate the house preference representation set of each user through the score calculation module , Landlord Preference Set , House feature related preference representation set Preference representation set related to landlord characteristics The matching score of each item-level representation vector in the corresponding item-level representation set;

[0238] Specifically, the house preference representation set of each user u is calculated by the score calculation module: With the housing representation collection The matching score of each representation vector in is expressed as:

[0239]

[0240] The score calculation module calculates the landlord preference representation set for each user u Collect with the landlord The matching score of each representation vector in is expressed as:

[0241]

[0242] The score calculation module calculates the house feature-related preference representation set of each user u With the first characterization set The matching score of each representation vector in is expressed as:

[0243]

[0244] The score calculation module calculates the house feature-related preference representation set of each user u With the first characterization set The matching score of each representation vector in is expressed as:

[0245]

[0246] Each matching score calculated by the score calculation module reflects the degree of matching of the user to each property feature, and can provide a better reference for the user to select a suitable property.

[0247] A7: Fusion preference representation set and landlords fused preference representation set , and the housing fusion representation set and the Landlord Fusion Representation Collection Input the preference fusion interpretation module for splicing, and obtain the user's interpretable preference weights for the house and the landlord through the spliced ​​representation set and ;

[0248] Specifically, the user's interpretable preference weight for the house The acquisition of is expressed as:

[0249]

[0250] In the formula, They respectively represent the parameter vectors to be learned of the first attention pooling layer in the corresponding representation interpretation module, represents the weight matrix to be learned of the first attention pooling layer in the corresponding representation interpretation module, R is the matrix identifier, and Respectively indicate that the dimensions of the matrix are dx1 and dxd; Indicates concatenation of input parameters;

[0251] Interpretable preference weights of users for houses The acquisition is:

[0252]

[0253] A8: Calculate the first item representation set respectively With the House Item Collection and the Landlord Item Representation Collection Between, and the second item representation set With the House Item Collection and the Landlord Item Representation Collection The disentanglement regularization loss between them, and the first loss of the heterogeneous preference recognition model is obtained according to the disentanglement regularization loss ;

[0254] Distance correlation is used to characterize the independence and correlation of any two pairs of different representations. When the coefficient is closer to 1, the two pairs of representations are more correlated, and when it is closer to 0, the two pairs of representations are more independent. The regularized loss for disentanglement is obtained. Since the first item representation set represents only the representations related to the house, so the first item representation set Need to be associated with the housing item representation set Related to the landlord item representation set Unrelated, correspondingly, the second item representation set Need to be associated with the landlord's listing Related to the housing item representation set Unrelated; by obtaining the corresponding regularization loss, the detangling ability of the entanglement representation decomposition module can be effectively improved.

[0255] Specifically, the first item representation set is calculated With the House Item Collection The disentanglement regularization loss between is expressed as:

[0256]

[0257] Calculate the first item representation set Listing collection with landlords The disentanglement regularization loss between is expressed as:

[0258]

[0259] Calculate the second item representation set Listing collection with landlords The disentanglement regularization loss between is expressed as:

[0260]

[0261] Calculate the second item representation set With the House Item Collection The disentanglement regularization loss between is expressed as:

[0262]

[0263] In the above formula, represents the distance covariance between two matrices, Represents the distance variance of the matrix.

[0264] Then the first loss of the heterogeneous preference identification model is The specific calculation is:

[0265]

[0266] A9: Calculate the shared item representation sets separately With the housing item representation set , Landlord Item Listing Collection , the first item representation set and the second object representation set The orthogonal constraint loss is obtained, and the second loss of the heterogeneous preference recognition model is obtained according to the orthogonal constraint loss. ;

[0267] In order to further improve the disentanglement effect of the entanglement representation decomposition module, an orthogonal constraint is introduced to calculate the corresponding loss, specifically:

[0268] Calculate the shared item representation set With the first item representation set The orthogonal constraint loss is expressed as:

[0269]

[0270] Calculate the shared item representation set The second item characterization set The orthogonal constraint loss is expressed as:

[0271]

[0272] Calculate the shared item representation set With the housing item representation set The orthogonal constraint loss is expressed as:

[0273]

[0274] Calculate the shared item representation set With the landlord's listing collection The orthogonal constraint loss is expressed as:

[0275]

[0276] In the above formula, represents the square of the F norm, Represents the transpose of a matrix;

[0277] Then the second loss of the heterogeneous preference identification model is The specific calculation is:

[0278]

[0279] A10: Based on the housing preference representation set of the corresponding user , Landlord Preference Set , House feature related preference representation set Preference representation set related to landlord characteristics The first BPR loss is calculated based on the matching scores of each item-level representation vector in the corresponding item-level representation set, and the third loss of the heterogeneous preference recognition model is obtained based on the first BPR loss. ;

[0280] Specifically, the first BPR loss of each preference representation set is calculated by the matching scores of each preference representation set of the corresponding user obtained in step A6 and each representation vector in the corresponding item-level representation set, which is specifically expressed as:

[0281]

[0282]

[0283]

[0284]

[0285] In the formula, represents the training set, represents the positive sample set of user u in the historical user set in the historical house set based on the transaction record set, represents a negative sample set of user u in the historical house set based on the transaction record set;

[0286] Then the third loss of the heterogeneous preference identification model is The calculation is as follows:

[0287]

[0288] A11: Based on the user’s interpretable preference weights for houses and landlords and , and the matching scores of the representation vectors of each item level in the corresponding item level representation set are used to calculate the second BPR loss, and the fourth loss of the heterogeneous preference recognition model is obtained according to the second BPR loss ;

[0289] Specifically, first obtain the housing fusion preference representation set of each user respectively Fusion representation set with said house The matching scores of the item-level representation vectors in , and the landlord fusion preference representation set of each user Fusion representation set with the landlord The matching score between , the specific formula is:

[0290]

[0291]

[0292] and Respectively represent the house fusion representation set and the Landlord Fusion Representation Collection The i-th representation vector in ;

[0293] Then get the user's interpretable preference weight for the house in step A7 and interpretable preference weights for landlords ; and based on the obtained user's interpretable preference weights for the house and the user's interpretable preference weights for the landlord Calculate the user preference matching score, expressed as:

[0294]

[0295] Then the corresponding second BPR loss is calculated by the following formula:

[0296]

[0297]

[0298]

[0299] In the formula, represents the training set, represents the positive sample set of user u in the historical user set in the historical house set based on the transaction record set, represents a negative sample set of user u in the historical house set based on the transaction record set;

[0300] Then the fourth loss of the heterogeneous preference identification model is The calculation is as follows:

[0301]

[0302] A12: The first loss Second loss , third loss and the fourth loss Perform weighting to obtain the final loss of the heterogeneous preference recognition model;

[0303] The specific manifestations are:

[0304]

[0305] in, are the hyperparameters controlling the first loss, the second loss, the third loss, and the fourth loss respectively.

[0306] A13: Reversely update the heterogeneous preference recognition model according to the final loss, and finally obtain the trained heterogeneous preference recognition model.

[0307] S4: Obtain a target user set and a target housing set, input the target user set and the target housing set into the trained heterogeneous preference recognition model, obtain the matching score output by the score calculation module and recommend it to the user;

[0308] As described above, each matching score output by the score calculation module reflects the degree of matching of the user with the characteristics of each housing source. The matching score is then recommended to the user to provide a reference for the user's selection of housing sources, thereby achieving housing source recommendations for the user.

[0309] S5: respectively obtaining the respective explainable weights output by the representation interpretation module and the respective explainable preference weights output by the preference fusion interpretation module, and converting the explainable weights and explainable preference weights into corresponding evaluation scores and recommending them to the landlord.

[0310] The explainable weights output by the representation interpretation module reflect the degree of influence of the user's various characteristics of the house on the house selection, and the explainable preference weights output by the preference fusion interpretation module reflect the user's preference for each house representation. Recommending them to the landlord can provide a reference for the landlord to make corresponding improvements and alleviate the landlord's anxiety.

[0311] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. An explainable online short-term rental recommendation method integrating trust and explicit preference, characterized in that: The method comprises: Obtain a historical user set, a historical property set and a transaction record set of online short-term rentals, construct a training set based on the historical user set, the historical property set and the transaction record set, and obtain a historical transaction property sequence for each user based on the transaction record set; the historical transaction property sequence represents the properties that the user has historically stayed in and their order; For all the houses in the historical house collection, obtain the house feature representation vector and the landlord feature representation vector of each house according to the house information and the landlord information of the house, and extract the entangled feature representation vector of each house according to the relevant comments of the user on the house, and construct the house feature representation set, landlord feature representation set and entangled feature representation set of all houses; Constructing a heterogeneous preference recognition model, training the heterogeneous preference recognition model according to the house feature representation set, the landlord feature representation set and the entanglement feature representation set of the training set, and obtaining the trained heterogeneous preference recognition model; The heterogeneous preference identification model includes a representation interpretation module, an entangled representation decomposition module, a representation fusion module, a user preference module, a score calculation module and a preference fusion interpretation module; The representation interpretation module is used to obtain an interpretable weight of each representation vector in the representation set according to the input representation set, and obtain an item-level representation set corresponding to the input representation set according to the interpretable weight; The entangled representation decomposition is used to disentangle the input representation set and decompose the input representation set into a plurality of corresponding item-level representation sets; The representation fusion module is used to splice and fuse several input representation sets; The user preference module is used to obtain the preference representation vector of the corresponding user according to the relationship between the representation vectors in the input representation set, and construct the preference representation set of each user; The score calculation module is used to calculate the matching score between the user's preference representation set and each item-level representation vector in the corresponding item-level representation set; The preference fusion interpretation module is used to splice the input representation set and obtain the interpretable preference weight of the spliced ​​representation set; Obtain a target user set and a target housing set, input the target user set and the target housing set into the trained heterogeneous preference recognition model, obtain the matching score output by the score calculation module and recommend it to the user; Respectively obtain the interpretable weights output by the representation interpretation module and the interpretable preference weights output by the preference fusion interpretation module, and convert the interpretable weights and interpretable preference weights into corresponding evaluation scores and recommend them to the landlord; The training of the heterogeneous preference recognition model according to the house feature representation set, the landlord feature representation set and the entanglement feature representation set of the training set specifically includes: Obtaining, through the representation interpretation module, item-level house item representation sets, landlord item representation sets, and entangled item representation sets corresponding to the house feature representation sets, landlord feature representation sets, and entangled feature representation sets; Decomposing the entangled item representation set into an item-level first item representation set related to the house, a second item representation set related to the landlord, and a shared item representation set related to both the house and the landlord by the entangled representation decomposition module; The house item representation set, the first item representation set and the shared item representation set are fused by the representation fusion module to obtain a house fused item representation set, and the landlord item representation set, the second item representation set and the shared item representation set are fused to obtain a landlord fused item representation set; Based on the house item representation set, the landlord item representation set, the first item representation set, the second item representation set, the house fusion item representation set and the landlord fusion item representation set, respectively extract the item-level house representation set, the landlord representation set, the first representation set, the second representation set, the house fusion representation set and the landlord fusion representation set corresponding to each user through the historical transaction house source sequence; Input the house representation set, landlord representation set, first representation set, second representation set, house fusion representation set and landlord fusion representation set corresponding to each user into the user preference module, and obtain the house preference representation set, landlord preference representation set, house feature-related preference representation set, landlord feature-related preference representation set, house fusion preference representation set and landlord fusion preference representation set of each user respectively; The score calculation module calculates the matching scores of the house preference representation set, the landlord preference representation set, the house feature-related preference representation set and the landlord feature-related preference representation set of each user and each item-level representation vector in the corresponding item-level representation set respectively; Input the house fusion preference representation set and the landlord fusion preference representation set, as well as the house fusion representation set and the landlord fusion representation set into the preference fusion interpretation module for splicing, and obtain the user's interpretable preference weights for the house and the landlord respectively through the spliced ​​representation sets; Respectively calculating the disentanglement regularization loss between the first item representation set and the house item representation set and the landlord item representation set, and between the second item representation set and the house item representation set and the landlord item representation set, and obtaining the first loss of the heterogeneous preference recognition model according to the disentanglement regularization loss; Respectively calculating the orthogonal constraint losses of the shared item representation set and the house item representation set, the landlord item representation set, the first item representation set, and the second item representation set, and obtaining the second loss of the heterogeneous preference recognition model according to the orthogonal constraint losses; Calculate a first BPR loss according to the matching scores of the house preference representation set, the landlord preference representation set, the house feature-related preference representation set and the landlord feature-related preference representation set of the corresponding user and each item-level representation vector in the corresponding item-level representation set, and obtain a third loss of the heterogeneous preference recognition model according to the first BPR loss; Calculate a second BPR loss according to the user's interpretable preference weights for the house and the landlord, and the matching scores of each item-level representation vector in the corresponding item-level representation set, and obtain a fourth loss of the heterogeneous preference recognition model according to the second BPR loss; Weighting the first loss, the second loss, the third loss and the fourth loss to obtain the final loss of the heterogeneous preference identification model; The heterogeneous preference recognition model is reversely updated according to the final loss, and finally the trained heterogeneous preference recognition model is obtained.

2. According to claim 1, an explainable online short-term rental recommendation method integrating trust and explicit preference, characterized in that: The characterized interpretation module includes several first attention pooling layers arranged in parallel; The first attention pooling layer is used to parse the representation vectors in the input representation set according to the input representation set and using the attention pooling mechanism, obtain the interpretable weight of each representation vector, and perform weighted concatenation on each representation vector according to the interpretable weight to obtain the item-level representation set corresponding to the input representation set; It manifests as: In the formula, Representation set representing the input The lth characterization vector in Represents the representation set The interpretable weight calculated from the lth representation vector in, represents the concatenation function, Represents the representation set The vector after weighted concatenation of all representation vectors of The output corresponds to the representation set The item-level representation set, is the corresponding input representation set The parameter vector to be learned of the first attention pooling layer represents the overall representation set; is the corresponding input representation set The parameter vector to be learned of the first attention pooling layer, , is the weight matrix to be learned, R is the matrix identifier, and Indicates that the dimensions of the matrix are dx1 and dxd respectively.

3. According to claim 1, an explainable online short-term rental recommendation method integrating trust and explicit preference, characterized in that: The user preference module includes several preference networks arranged in parallel; the preference network includes a multi-head self-attention layer and a second attention pooling layer; The multi-head self-attention layer is used to obtain the relationship between each representation vector in the input representation set by using the self-attention mechanism, and enhance the corresponding representation vector through the relationship; The second attention pooling layer is used to parse each enhanced representation vector by using the attention mechanism, obtain the preference weight of each enhanced representation vector, and perform weighted concatenation on each enhanced representation vector according to the preference weight to obtain a preference representation set of the corresponding representation set; It manifests as: In the formula, and Respectively represent the input item-level representation set for user u The i-th characterization vector and the j-th characterization vector in is the corresponding input representation set The weight of the multi-head self-attention layer in the preference network represents the relationship between the i-th representation vector and the j-th representation vector, The input representation set The enhanced representation vector of the i-th representation vector, is the corresponding input representation set The weight matrix to be learned of the multi-head attention layer in the preference network, R is the matrix identifier, Indicates that the dimension of the matrix is ​​dxd; is the corresponding input representation set The weight of the second attention pooling layer in the preference network represents the importance of the i-th and j-th enhanced representation vectors to the user u; is the corresponding input representation set The preference representation set output by the preference network; is the parameter vector to be learned in the preference network, representing the entire historical transaction house sequence; represents the concatenation function; represents the natural exponential function; Represents the transpose of a matrix; and Respectively represent the corresponding input representation set The parameter vector and weight matrix to be learned of the second attention pooling layer in the preference network, R is the matrix identifier, and Indicates that the dimensions of the matrix are dx1 and dxd respectively.

4. According to claim 1, an explainable online short-term rental recommendation method integrating trust and explicit preference, characterized in that: The house fusion preference representation set and the landlord fusion preference representation set, as well as the house fusion representation set and the landlord fusion representation set are input into the preference fusion interpretation module for splicing, and the user's interpretable preference weights for the house and the landlord are obtained through the spliced ​​representation sets, wherein the user's interpretable preference weights for the house are obtained as follows: The user's interpretable preference weights for the landlord are obtained as follows: In the formula, They represent the house fusion preference representation set and the landlord fusion preference representation set respectively, They represent the house fusion representation set and the landlord fusion representation set respectively, They represent the parameter vectors to be learned, represents the weight matrix to be learned, R is the matrix identifier, and Respectively indicate that the dimensions of the matrix are dx1 and dxd; Indicates concatenation of input parameters; Represents the transpose of a matrix.

5. According to claim 1, an explainable online short-term rental recommendation method integrating trust and explicit preference, characterized in that: The score calculation module is used to calculate the matching score between the user's preference representation set and each item-level representation in the corresponding item-level representation set. The specific formula is as follows: In the formula, represents the i-th representation vector in the item-level representation set output by the corresponding representation interpretation module or representation fusion module; represents the preference representation set of user u output by the corresponding user preference module, Represents the transpose of a matrix.

6. An explainable online short-term rental recommendation method integrating trust and explicit preference according to any one of claims 1 to 5, characterized in that: The disentanglement regularization loss between the first item representation set and the house item representation set and the landlord item representation set, and the disentanglement regularization loss between the second item representation set and the house item representation set and the landlord item representation set are calculated respectively, and the corresponding disentanglement regularization losses are calculated respectively by the following formulas: In the formula, is the first item representation set, is the second item representation set, is a collection of house item representations, A collection of listings for said landlord; represents the distance covariance between two matrices, represents the distance variance of the matrix; The first loss of the heterogeneous preference identification model is obtained according to the regularization loss, and is specifically calculated by the following formula: In the formula, Represents the first loss.

7. An explainable online short-term rental recommendation method integrating trust and explicit preference according to any one of claims 1 to 5, characterized in that: The orthogonal constraint losses of the shared item representation set and the house item representation set, the landlord item representation set, the first item representation set, and the second item representation set are calculated respectively, and the corresponding orthogonal constraint losses are calculated respectively by the following formulas: In the formula, represents the square of the F norm, is the first item representation set, is the second item representation set, is a collection of house item representations, is a collection of landlord listings, is the shared item representation set, Represents the transpose of a matrix; The second loss of the heterogeneous preference identification model is obtained according to the orthogonal constraint loss, and is specifically calculated by the following formula: In the formula, Represents the second loss.

8. The explainable online short-term rental recommendation method integrating trust and explicit preference according to claim 5, characterized in that: The first BPR loss is calculated according to the matching scores of the house preference representation set, the landlord preference representation set, the house feature-related preference representation set and the landlord feature-related preference representation set of the corresponding user and each item-level representation vector in the corresponding item-level representation set, and the corresponding first BPR loss is calculated respectively by the following formula: In the formula, represents the training set, represents the positive sample set of user u in the historical user set in the historical house set based on the transaction record set, represents a negative sample set of user u in the historical house set based on the transaction record set; The third loss of the heterogeneous preference identification model is obtained according to the first BPR loss, and is specifically calculated by the following formula: In the formula, represents the third loss.

9. The explainable online short-term rental recommendation method integrating trust and explicit preference according to claim 5, characterized in that: The second BPR loss is calculated based on the user's interpretable preference weights for the house and the landlord, and the matching scores of each item-level representation in the corresponding item-level representation set, specifically including: The matching scores of the house fusion preference representation set of each user and the representation vectors of each item level in the house fusion representation set, as well as the matching scores between the landlord fusion preference representation set of each user and the landlord fusion representation set are obtained respectively. The specific formula is: In the formula, and Respectively represent the house fusion representation set and the Landlord Fusion Representation Collection The i-th characterization vector in, and They respectively represent the house fusion preference representation set and the user-landlord fusion preference representation set obtained based on user u; Get the user's interpretable preference weights for houses and interpretable preference weights for landlords ; The user preference matching score is calculated using the following formula: The corresponding second BPR loss is calculated by the following formulas: In the formula, represents the training set, represents the positive sample set of user u in the historical user set in the historical house set based on the transaction record set, represents a negative sample set of user u in the historical house set based on the transaction record set; The fourth loss of the heterogeneous preference identification model is obtained according to the second BPR loss, and is specifically calculated by the following formula: In the formula, represents the fourth loss.

Images