Cold start user recommendation method, system and device based on heterogeneous transfer learning

Abstract

The application discloses a cold start user recommendation method, system and device based on heterogeneous transfer learning, which comprises the following steps: using feature enhancement technology to combine source domain and target domain data into an enhanced matrix, then using an autoencoder to map the enhanced matrix to a latent feature space, performing feature matching and distribution alignment processing in the space, and finally reconstructing user features. Then, a solving model is established according to the reconstructed user features, that is, the solving model can be solved through matrix feature decomposition, thereby improving the recommendation accuracy for cold start users. In addition, the application also extends the model to a neural network, and the effectiveness of the model is verified through experiments on multiple real world data sets, and the performance index is excellent. The application can effectively utilize auxiliary source domain information to assist the recommendation task of the target domain, thereby providing more accurate and personalized recommendations for cold start users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of recommendation, in particular to a cold start user recommendation method, system and device based on heterogeneous transfer learning. BACKGROUND

[0002] Recommendation systems play a crucial role in e-commerce, social media, and content distribution platforms. They enhance user experience by analyzing users' behavior and preferences to provide personalized recommendations. However, a long-standing challenge is how to handle cold start users, i.e., new users without historical interaction data in the system.

[0003] In related technologies, cold start users lack historical interaction data, making it difficult for recommendation models to learn their preferences. This limits the ability of recommendation systems to serve new users. Meanwhile, user features come from different domains (such as social networks and purchase history), which may have distribution differences between domains, making it difficult to directly integrate. This heterogeneity increases the complexity of extracting useful information from different data sources. In addition, existing models have difficulty quickly adapting and providing recommendations when new users or new items appear. This indicates that existing systems have limitations in handling dynamic changes in user preferences and item characteristics. SUMMARY

[0004] The purpose of the present application is to provide a cold start user recommendation method, system and device based on heterogeneous transfer learning, to solve the problem of inaccurate recommendation for cold start users in related recommendation technology.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides a cold start user recommendation method based on heterogeneous transfer learning, comprising:

[0007] obtaining first information data of a user, wherein the first information data of the user includes social relationship data of warm users and cold start users and historical interaction data of warm users;

[0008] performing feature enhancement processing on the first information data of the user to obtain second information data of the user;

[0009] performing spatial mapping processing, feature matching processing and distribution alignment processing on the second information data of the user in sequence using an autoencoder or a neural network-based autoencoder to obtain third information data of the user;

[0010] constructing a target recommendation model according to the third information data of the user;

[0011] generating a recommendation list for a cold start user according to the target recommendation model.

[0012] Further, the first information data of the user is subjected to feature enhancement processing to obtain second information data of the user, comprising:

[0013] Data feature enhancement is performed by zero padding, and the formula is as follows:

[0014]

[0015] wherein X is the second information data of the user, U w is the social relationship data of the warm user, H is the historical interaction data of the warm user, U c is the social relationship data of the cold start user.

[0016] Further, the second information data of the user is subjected to space mapping processing, feature matching processing and distribution alignment processing in sequence by using an autoencoder to obtain third information data of the user, and the details are as follows:

[0017] The second information data X of the user is mapped from the original space to a low-dimensional latent space by using an autoencoder, and then unknown interaction of the cold start user is obtained by using the reconstruction of the autoencoder, wherein a loss function in the space mapping processing process is constructed as follows:

[0018]

[0019] s.t.,W T W=I

[0020]

[0021] wherein I is a unit matrix, X is the second information data of the user, and ||·|| F is the Euclidean norm, W is a mapping matrix, W T is the transpose of W, and Tr(·) is the trace of the mapping matrix.

[0022] A loss function in the feature matching processing process is constructed as follows:

[0023]

[0024] wherein, is the loss function in the feature matching, W s and W t are mapping matrices, and is the transpose of the mapping matrix, m is the number of items, and d is the feature dimension of the user social relationship.

[0025] A loss function in the distribution alignment processing process is constructed as follows:

[0026]

[0027] n = n w + n c

[0028] wherein, is the loss function of distribution alignment processing, u i and h j are the i-th and j-th vectors of U and H respectively, n is the number of users, n w is the number of warm users, n c is the number of cold start users.

[0029] Further, the target recommendation model is constructed according to the third information data of the user, and the specific formula is as follows:

[0030]

[0031] A = λB + γM - XX T

[0032] f(W) = Tr(W T AW) - Tr(Φ T (W T W-I))

[0033]

[0034] Wherein, λ and γ are hyperparameters, used to balance the weight between the three loss functions; is the overall loss function, f(W) is the Lagrange function, and Φ is a symmetric matrix.

[0035] Further, the second information data of the user is sequentially subjected to spatial mapping processing, feature matching processing and distribution alignment processing based on a neural network autoencoder to obtain third information data of the user, and the specific formula is as follows:

[0036] Obtain the reconstructed data in the low-dimensional space The formula is as follows:

[0037]

[0038]

[0039] Y (l+1) = σ(W (l+1) Y (l) )

[0040]

[0041] X (0) = X

[0042] where σ(·) is a nonlinear ReLU activation function, is the network parameter between (l) and (l+1) layers, X (l) and X (l+1) are the representations in (l) and (l+1) layers, respectively; q(·) is an encoder; W (l+1) is the network parameter between (l) and (l+1) layers; p(·) is a decoder;

[0043] The loss function in the spatial mapping processing process is constructed, and the formula is as follows:

[0044]

[0045] where g1(·,·) is a mean square error;

[0046] The loss function in the feature matching processing process is constructed, and the formula is as follows:

[0047]

[0048] where g2(·,·) is a mean square error;

[0049] The loss function in the distribution alignment processing process is constructed, and the formula is as follows:

[0050]

[0051] where g3(·,·) is an MMD loss.

[0052] Further, the target recommendation model is constructed according to the third information data of the user, and the specific construction is as follows:

[0053] The overall loss function of the target recommendation model is constructed, and the formula is as follows:

[0054]

[0055] where, is the overall loss function;

[0056] The gradient of the overall loss function is calculated, and then the parameters are updated by an optimizer.

[0057] Further, the method further comprises:

[0058] The results of the recommendation list are evaluated, and the evaluation indexes include calculation accuracy, recall rate, normalized discount cumulative gain, and hit rate.

[0059] Further, the evaluation of the result of the recommendation list includes: calculation accuracy, recall rate, normalized discounted cumulative gain and hit rate, and the specific calculation formula is as follows:

[0060] The calculation accuracy evaluation formula is as follows:

[0061]

[0062] Wherein, N is the number of cold start users, R i is the real preference item of the i-th cold start user, and is the top-K item list recommended for the i-th cold start user;

[0063] The recall rate evaluation formula is as follows:

[0064]

[0065] The normalized discounted cumulative gain evaluation formula is as follows:

[0066]

[0067] Wherein, r i is the relevance score of the item located at position i, and IDCG is set to ensure that perfect ranking results in a normalized discounted cumulative gain value of 1;

[0068] The hit rate evaluation formula is as follows:

[0069]

[0070] Wherein, HR@K is defined as the total number of correctly recommended items divided by the total number of all user preference items.

[0071] In a second aspect, the application also provides a cold start user recommendation system based on heterogeneous transfer learning, comprising:

[0072] A data acquisition module is configured to acquire first information data of a user, wherein the first information data of the user includes social relationship data of warm users and cold start users and historical interaction data of warm start users;

[0073] A feature matching module is configured to perform feature enhancement processing on the first information data of the user to obtain second information data of the user;

[0074] A data processing module is configured to sequentially perform spatial mapping processing, feature matching processing and distribution alignment processing on the second information data of the user by using an autoencoder or a neural network-based autoencoder to obtain third information data of the user;

[0075] A constructing module is configured to construct a target recommendation model according to third information data of the user;

[0076] A recommendation list generating module is configured to generate a recommendation list of the cold-start user according to the target recommendation model.

[0077] In a third aspect, the present application further provides a computer electronic device including a memory and a processor, the memory stores a computer program, and the processor implements the steps of the cold-start user recommendation method based on heterogeneous transfer learning according to any one of the above aspects when executing the computer program.

[0078] The cold-start user recommendation method, system and device based on heterogeneous transfer learning provided by the present application have the following beneficial effects:

[0079] 1. The present application redefines the cold-start user recommendation problem as a heterogeneous domain transfer learning problem, and combines feature matching and distribution alignment processing methods, and the proposed model effectively solves the cold-start user recommendation problem and improves the performance and robustness of the recommendation system.

[0080] 2. The present application verifies the superior performance of the model in terms of precision, recall, NDCG and hit rate on multiple real data sets, and provides more accurate and personalized recommendation services for users.

[0081] 3. The present application verifies the superior performance of the model in terms of precision, recall, NDCG and hit rate on multiple real data sets, and provides more accurate and personalized recommendation services for users. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 FIG. 1 is a flowchart of a cold-start user recommendation method based on heterogeneous transfer learning in an embodiment of the present application;

[0083] Figure 2 FIG. 2 is a structural diagram of a cold-start user recommendation system based on heterogeneous transfer learning in an embodiment of the present application;

[0084] Figure 3 FIG. 3 is a structural diagram of a computer electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0086] It should be noted that when an element is referred to as being "on" another element, it can be directly on the other element or intervening elements can also be present. In contrast, when an element is referred to as being "directly on" another element, there are no intervening elements present. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements can be present. By way of non-limiting example, when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements can be present. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0087] In this application, unless otherwise specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0088] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "multiple" is two or more, unless otherwise explicitly specified.

[0089] The terms used in one or more embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present application. The singular forms "a", "said" and "the" used in one or more embodiments of the present application are also intended to include plural forms, unless the context clearly indicates otherwise.

[0090] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms used in the specification of the template are only for the purpose of describing the specific embodiments, and are not intended to limit the present application. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0091] It should be understood that, although the terms first, second, etc. can be employed in describing various information in one or more embodiments, such information should not be limited to these terms. These terms are only used to differentiate one piece of information from another. For example, a first can be termed a second, and, similarly, a second can be termed a first, without departing from the scope of one or more embodiments. The word "if' can be interpreted to mean "when" or "upon" or "in response to" depending on the context.

[0092] Currently, traditional recommendation systems mainly rely on collaborative filtering and content-based filtering techniques. These methods perform well when dealing with users with rich historical data. However, for cold-start users, these systems struggle to generate accurate recommendations due to the lack of sufficient historical interaction information. In addition, existing systems also face challenges when dealing with heterogeneous data sources, such as social relationships and purchase history, as they often require large amounts of labeled data and complex feature engineering.

[0093] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail in the specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes will not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0094] Please refer to Figure 1 The cold-start user recommendation method based on heterogeneous transfer learning provided by the embodiments of the present application at least includes the following steps:

[0095] S10, obtaining first information data of a user, wherein the first information data of the user includes social relationship data of warm users and cold-start users and historical interaction data of the warm users.

[0096] It can be understood that, in the present embodiment, the first information data of the user includes warm user data and cold-start user data. As a preferred embodiment, in the present embodiment, 90% of the users are randomly selected for training, i.e. warm users, and 10% of the users are randomly selected for testing, i.e. cold-start users.

[0097] It should be noted that heterogeneous transfer learning aims to utilize knowledge from other source domains to assist the learning task of the target domain. The main challenge of domain adaptation is to reduce the domain gap between the source domain and the target domain. Existing works tend to employ methods such as maximum mean discrepancy, optimal transport, and adversarial loss to estimate the distribution difference between the two domains.

[0098] S20, performing feature enhancement processing on the first information data of the user to obtain second information data of the user.

[0099] Specifically, in this example, data feature enhancement is performed by zero padding means, and the formula is as follows:

[0100]

[0101] where X is the second information data of the user, U w is the social relationship data of the warm user, H is the historical interaction data of the warm user, U c is the social relationship data of the cold start user.

[0102] S30, the second information data of the user is sequentially processed by a self-encoder or a neural network-based self-encoder for spatial mapping, feature matching and distribution alignment, to obtain the third information data of the user.

[0103] Specifically, in the process of processing data, the present application includes two processing methods, as follows:

[0104] First, the second information data of the user is sequentially processed by a self-encoder for spatial mapping, feature matching and distribution alignment, to obtain the third information data of the user, and the specific processing process is as follows:

[0105] S301, spatial mapping processing: using a self-encoder to map the feature-enhanced data from the original space to a low-dimensional latent space, and then obtaining the unknown interaction of the cold start user through the reconstructed , specifically:

[0106]

[0107] where X is the feature-enhanced data, U w is the social relationship of the warm user; H is the historical interaction of the warm user, U c is the social relationship of the cold start user, m is the number of items, d is the feature dimension of the user social relationship, and n is the number of users.

[0108] First, define the mapping matrix: the goal is to learn a mapping matrix to obtain the reconstructed Thus, the unknown interaction of the cold start user is indirectly obtained, and these values are regarded as the ratings of the user to the items, and the recommendation is made based on these ratings.

[0109] Second, construct the reconstruction loss function: the goal is to minimize the square of the Euclidean (Frobenius) norm between the original and the reconstructed

[0110]

[0111] st,W T W = I

[0112] Where I is the identity matrix, X is the user's second information data, and ∥·∥ F It is the Euclidean norm, and W is the mapping matrix. T This is the transpose of W. It's important to note that there are two reasons for imposing an orthogonal constraint on W. First, this constraint prevents degradation, such as all transformed samples collapsing back to the origin. Second, this constraint ensures that the transformed data has no redundant information across any dimension.

[0113] Finally, the loss function for reconstructing X is optimized: According to the definition of the Frobenius 2-norm, the loss function for reconstructing X can be optimized as follows:

[0114]

[0115] Where Tr(·) is the trace of the mapping matrix.

[0116] S302, Feature matching processing: In the latent space, feature matching is performed on the social relationships and historical interactions of warm users to reduce the difference between the two representations.

[0117] First, feature matching is defined as follows: for warm users' social relationships and their historical interactions, but these relationships are heterogeneous, so it is necessary to seek to minimize their differences in the latent space;

[0118] Secondly, define two transformation mappings: the goal is to utilize the two mapping matrices respectively. and Warm up users' social relationships U w Map their historical interactions H to a potential public space to make them similar to each other;

[0119] Next, we construct the feature matching loss function: the goal is to minimize the loss function obtained through W. s and W t The transformed social relationships and historical interactions are expressed as the square of the Frobenius norm in the latent space. Therefore, we have the following loss function for feature matching:

[0120]

[0121] Finally, the loss function for feature matching is optimized: first, W... s and W t Combined into W, the feature matching loss function can then be optimized according to the definition of the Frobenius 2-norm:

[0122]

[0123] where, is the loss function in feature matching, W s and W t are mapping matrices, and is the transpose of mapping matrix.

[0124] S303, distribution alignment processing: using the maximum mean difference (MMD) method to estimate the distribution difference between the source domain (user social relationship U) and the target domain (user historical interaction H), and minimizing these differences to reduce the gap between domains.

[0125] First, the definition of distribution alignment is as follows: for user social relationship U (including warm users U w and cold start users U c ) and user historical interaction H are representations from two different feature spaces, which exhibit heterogeneous attributes. Therefore, the user social relationship U is regarded as the source domain, and the user historical interaction H is regarded as the target domain. MMD is used for heterogeneous transfer learning, in other words, to reduce the distribution difference between the source domain and the target domain in the latent space.

[0126] Second, define two mapping matrices: the goal is to use two mapping matrices and to minimize the distribution difference between user social relationship U (source domain) and its historical interaction H (target domain) in the latent space, so that they are closer.

[0127] Next, construct the distribution alignment loss function: the goal is to use MMD to minimize the distribution difference between user social relationship U (source domain) and historical interaction H (target domain) in the latent space. Therefore, the distribution alignment loss function is as follows:

[0128]

[0129] where u i and h j are the i-th and j-th vectors of U and H, n is the number of users, n w is the number of warm users.

[0130] Finally, optimize the loss function of distribution alignment: first, combine W s and W t into W, then according to the definition of Euclidean (Frobenius) 2 norm, the distribution alignment loss function can be optimized as:

[0131]

[0132] wherein, is a loss function of distribution alignment processing.

[0133] Secondly, the second information data of the user is sequentially subjected to spatial mapping processing, feature matching processing and distribution alignment processing by using a neural network-based autoencoder to obtain third information data of the user, and the specific processing process is as follows:

[0134] S311, the spatial mapping processing is as follows: first, the feature expanded X is encoded by a layer of neural network, and the following is realized:

[0135]

[0136] wherein σ(·) is a nonlinear ReLU activation function, is a network parameter between (l) and (l+1) layers, X (l) and X (l+1) are representations in (l) and (l+1) layers respectively, and X (0) = X.

[0137] Secondly, the feature expanded X is encoded by a multi-layer neural network to obtain a latent representation Y with a dimension of K, and the following is realized:

[0138]

[0139] wherein q(·) is an encoder including multiple layers S6.1.

[0140] Next, the latent representation Y is decoded by a layer of neural network, and the following is realized:

[0141] Y (l+1) = σ(W (l+1) Y (l) )

[0142] wherein W (l+1) is a network parameter between (l) and (l+1) layers.

[0143] Then, the latent representation Y is decoded by a multi-layer neural network, and finally the reconstructed

[0144]

[0145] wherein p(·) is a decoder including multiple layers S6.3.

[0146] Finally, the reconstructed data is made as similar as possible to the original data X, and the reconstruction loss function is defined as follows:

[0147]

[0148] where g1(·,·) is the mean square error.

[0149] S312, the feature matching process is as follows: calculating the feature matching loss based on the latent representation Y, so that the social relationship U of the warm user and its historical interaction H are similar in the latent representation, which is implemented as follows: w

[0150]

[0151] where g2(·,·) is the mean square error.

[0152] S313, the distribution alignment process is as follows: calculating the distribution alignment loss based on the latent representation Y by calculating MMD, so that the distribution difference between the social relationship U (source domain) of the heterogeneous user and its historical interaction H (target domain) is reduced in the latent representation, which is implemented as follows:

[0153]

[0154] where g3(·,·) is the MMD loss.

[0155] S40, constructing a target recommendation model according to the third information data of the user.

[0156] It can be understood that in the embodiment, the third information of the user includes two kinds: the first kind is the third information data of the user corresponding to the data processing in step S30 by using the autoencoder, which is specifically as follows:

[0157] S401, overall target function and optimization: ensuring the performance of the recommendation system when processing cold start users, combining the autoencoder conversion, feature matching and distribution alignment into an overall target function, and solving the overall target function by Lagrange;

[0158] Overall loss function: comprehensive step, design the overall loss function of the model:

[0159]

[0160] Where λ and γ are hyperparameters used to balance the weights between the three loss functions;

[0161] S402, rewriting the overall loss function: combining B in feature matching and M in distribution alignment into A after feature expansion XX T , which can be rewritten as:

[0162]

[0163] Where A = λB + γM-XX T ; ​

[0164] S403, convert to Lagrange function: by introducing a symmetric matrix Φ as the Lagrange multiplier, the Lagrange function is as follows:

[0165] f(W) = Tr(W T AW)-Tr(Φ T (W T W-I))

[0166] S404, derivative of Lagrange function: derivative of Lagrange function and set its derivative to zero

[0167]

[0168] S405, convert to solve eigenvalue problem: this equation AW = WΦ shows that the column vectors of W are the eigenvectors of matrix A, and Φ is related to eigenvalues. Therefore, the problem is converted to solve the k smallest eigenvectors of matrix A.

[0169] The second is to process the third information data of the corresponding user in step S30 by using a neural network-based autoencoder, which is specifically as follows:

[0170] S411, construct the overall loss function:

[0171]

[0172] Wherein, is the overall loss function.

[0173] S412, calculate the gradient of the overall loss function;

[0174] S413, update the parameters through an optimizer (such as Adam, SGD, etc.), and then obtain a local optimal solution.

[0175] S50, generate a recommendation list for a cold start user according to the target recommendation model.

[0176] In this embodiment, the recommendation list for the cold start user is generated based on the autoencoder (FDMA) and the neural network-based autoencoder (D-FDMA) respectively.

[0177] Exemplarily, taking the autoencoder to generate the recommendation list as an example:

[0178] S501, obtain the reconstructed data (i.e. the third information data of the user), the model learns the mapping matrix W, which is specifically as follows

[0179]

[0180] Wherein, and respectively represent the reconstructed social relationship of the warm user, the reconstructed social relationship of the cold-start user and the reconstructed historical interaction of the warm user.

[0181] S502, obtain the score of the cold-start user to the item: obtain the score of the cold-start user to the item through a sigmoid activation function:

[0182]

[0183] S503, generate a recommendation list: obtain the score of the cold-start user to the item For generating top-k recommendations, each cold-start user gets the corresponding k items with the highest score.

[0184] In an embodiment of the present application, the cold-start user recommendation method based on heterogeneous transfer learning further comprises:

[0185] S60, evaluate the results of the recommendation list, and the evaluation indexes include: calculation accuracy, recall rate, normalized discount cumulative gain and hit rate.

[0186] Specifically, the following indexes are used to evaluate the performance of the recommendation system:

[0187] S601, precision (Precision): the proportion of correctly recommended items (i.e. preferred items existing in the recommendation list) in the total number of all recommended items, averaged over all samples:

[0188]

[0189] Wherein, N is the number of cold-start users, R i is the real preferred item of the i-th cold-start user, and is the top-K item list recommended for the i-th cold-start user;

[0190] S602, recall rate (Recall): the recall rate is the proportion of the number of correctly recommended items to the total number of recommended items, i.e. the number of preferred items, which can be calculated by the following formula:

[0191]

[0192] S603, normalized discount cumulative gain (NDCG): a recommendation quality index considering item ranking. For NDCG, we calculate NDCG@K as:

[0193]

[0194] Wherein, r iis the relevance score of the item at position i, IDCG is set to ensure that perfect ranking results in NDCG (Normalized Discounted Cumulative Gain) value of 1. Specifically, we set r i to 1 for the preferred item and 0 for others;

[0195] S604, HitRate (HR): the proportion of the recommendation list containing at least one item that the user actually prefers, HR@K is defined as the total number of correctly recommended items divided by the total number of all user preferred items:

[0196]

[0197] It should be noted that the present application uses the Flickr dataset for the user cold start recommendation problem, and the Flickr contains data collected from a popular personal photo management and sharing website. Users in the system can add tags to photos and subscribe to photos using tags. In our experiment, 195 interest tags are selected as items, and 8000 users are selected to construct the learning task. Similar to BlogCatalog, user features are generated from their social networks;

[0198] The Top-10 prediction results in Flickr are compared with JCTN (Tingting Liang, Congying Xia, Yuyu Yin, and Philip S Yu. 2020. Joint Training Capsule Network for Cold Start Recommendation. In ACM SIGIR Conference on Research and Development in Information Retrieval. 1769-1772.), AADA (Hanrui Wu, Jinyi Long, Nuosi Li, Dahai Yu, and Michael K Ng. 2023. Adversarial Auto-encoder Domain Adaptation for Cold-start Recommendation with Positive and Negative Hypergraphs. ACM Transactions on Information Systems 41, 2 (2023), 1-25.), and FMM (Hanrui Wu, Nuosi Li, Ka Ho Kwok, Xuheng Cai, Jia Zhang, Jinyi Long, and Michael K. Ng. 2024. Feature Matching Machine for Cold-Start Recommendation. IEEE Transactions on Services Computing 17, 1 (2024), 98-112.);

[0199] The experimental results are shown in Table 1, and algorithm 1 and algorithm 2 respectively correspond to the verification results FDMA and D-FDMA of the algorithm provided in the application; algorithm 3 corresponds to the verification result of JCTN; algorithm 4 corresponds to the verification result of AADA; and algorithm 5 corresponds to the verification result of FMM; it can be seen from Table 1 that the algorithm provided in the application is better than other algorithms in most evaluation indexes by considering the label correlation performance;

[0200] Table 1

[0201] Precision Recall NDCG Hit rate Algorithm 1 19.42±0.4 61.39±1.3 51.74±1.1 61.47±1.2 Algorithm 2 19.12±0.2 59.89±0.5 50.51±0.8 60.20±0.8 Algorithm 3 11.12±0.3 36.82±1.4 29.30±0.7 35.19±1.0 Algorithm 4 14.91±0.3 47.94±0.9 37.53±1.6 47.10±1.0 Algorithm 5 16.56±0.4 52.16±1.1 41.22±1.0 52.10±1.4

[0202] The heterogenous transfer learning cold start user recommendation method provided in the application has the beneficial effects that:

[0203] 1. The application redefines the cold-start user recommendation problem as a heterogeneous domain transfer learning problem, and combines feature matching and distribution alignment processing methods, the proposed model effectively solves the cold-start user recommendation problem, and improves the performance and robustness of the recommendation system.

[0204] 2. The application verifies the superior performance of the model on precision, recall, NDCG and hit rate on multiple real data sets, and provides more accurate and personalized recommendation services for users.

[0205] 3. The application verifies the superior performance of the model on precision, recall, NDCG and hit rate on multiple real data sets, and provides more accurate and personalized recommendation services for users.

[0206] Please refer to Figure 2 The application also provides a cold-start user recommendation system 200 based on heterogeneous transfer learning, comprising:

[0207] The data acquisition module 201 is configured to acquire first information data of a user, wherein the first information data of the user comprises social relationship data of warm users and cold-start users and historical interaction data of warm-start users.

[0208] The feature matching module 202 is configured to perform feature enhancement processing on the first information data of the user to obtain second information data of the user.

[0209] The data processing module 203 is configured to sequentially perform spatial mapping processing, feature matching processing and distribution alignment processing on the second information data of the user using an autoencoder or a neural network-based autoencoder to obtain third information data of the user.

[0210] The construction module 204 is configured to construct a target recommendation model according to the third information data of the user.

[0211] The recommendation list generation module 205 is configured to generate a recommendation list for a cold-start user according to the target recommendation model.

[0212] Please refer to Figure 3 The computer electronic device 300 provided in the embodiment of the application comprises a memory 303 and a processor 302, the memory 303 stores a computer program, and the processor 302 implements the steps of the cold-start user recommendation method based on heterogeneous transfer learning according to any one of the above embodiments when executing the computer program.

[0213] Specifically, the electronic device 300 comprises a transceiver 301, a bus interface and a processor 302, the processor 302 is configured to acquire first information data of a user, wherein the first information data of the user comprises social relationship data of a warm user and a cold start user and historical interaction data of the warm user, perform feature enhancement processing on the first information data of the user to obtain second information data of the user, perform spatial mapping processing, feature matching processing and distribution alignment processing on the second information data of the user in sequence by using a self-encoder or a neural network-based self-encoder to obtain third information data of the user, construct a target recommendation model according to the third information data of the user, and generate a recommendation list of the cold start user according to the target recommendation model.

[0214] In the embodiments of the present application, the electronic device 300 further comprises a memory 303. Figure 3 The bus architecture can comprise any number of interconnecting buses and bridges, and the various circuitry representative of the processor(s) 302 and the memory 303 are linked together by the bus architecture, which, can, in turn, be linked to various other circuits including a peripheral device(s), voltage regulators, and power management circuits, all of which are well known in the art, thus, further description of these individual components is not deemed necessary for the purposes of this disclosure. The bus interface provides an interface to the bus architecture. The transceiver 301 can be a plurality of elements, including a transmitter and a receiver, which provide a means by which the various entities communicate over a transmission medium. The processor 302 is responsible for managing the bus architecture and general processing, and the memory 303 can store data used by the processor 302 in executing its operations.

[0215] In the embodiments, the computer readable storage medium can be a non-volatile storage medium or a volatile storage medium. For example, the computer storage medium can include, but is not limited to, a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0216] In all the examples shown and described herein, any specific values should be interpreted as merely exemplary and not as a limitation, and thus, other examples of the example embodiments can have different values.

[0217] It should be noted that similar reference numbers and letters refer to similar items throughout the drawings, and thus, once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings.

[0218] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can also be implemented by other means. The apparatus embodiments described above are only illustrative, for example, the flowcharts and structural diagrams in the drawings show the possible implementation architecture, function and operation of the apparatus, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logic function. It should also be noted that in alternative implementation, the functions noted in the block can also occur in different order from that noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the structural diagram and / or flowchart, and the combination of blocks in the structural diagram and / or flowchart, can be implemented by a dedicated hardware-based system for executing the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0219] In addition, each functional module or unit in the embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0220] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing an end device (which can be a smart phone, a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application.

[0221] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A cold start user recommendation method based on heterogeneous transfer learning, characterized in that, The method comprises the following steps: obtaining first information data of a user, wherein the first information data of the user comprises social relationship data of warm users and cold-start users and historical interaction data of the warm users; performing feature enhancement processing on the first information data of the user to obtain second information data of the user, comprising: performing data feature enhancement through zero padding, and the formula is as follows: ； wherein, is second information data for the user, is social relationship data for the warm user, is historical interaction data for the warm user, is social relationship data for the cold start user; The second information data of the user is sequentially subjected to spatial mapping processing, feature matching processing and distribution alignment processing by using a self-encoder or a neural network-based self-encoder to obtain third information data of the user, including: using a self-encoder to map the second information data of the user to a low-dimensional latent space from the original space to the low-dimensional latent space, and then reconstructed by the self-encoder to obtain the unknown interaction of the cold start user, wherein the loss function in the spatial mapping processing process is constructed: ； ； wherein I is the identity matrix, X is the second information data of the user, ||·||F is the Euclidean norm, W is a mapping matrix, W T is the transpose of W, Tr(·) is the trace of the mapping matrix; constructing a loss function in the feature matching processing process; ； ； ， ； wherein, is a loss function in feature matching, and is a mapping matrix, is a transpose of the mapping matrix, is a number of items, is a feature dimension of user social relations; constructing a loss function in the distribution alignment processing process; ； ； ； ； wherein, is the loss function for the distribution alignment process, and are the i-th and j-th vectors of and respectively, is the number of users, is the number of warm users, is the number of cold-start users; constructing a target recommendation model according to the third information data of the user; generating a recommendation list of the cold-start user according to the target recommendation model. 2.The cold start user recommendation method based on heterogeneous transfer learning according to claim 1, wherein, constructing a target recommendation model according to the third information data of the user, and the specific formula is as follows: ； ； ； ； ； wherein, and is a hyperparameter used to balance the weights between the three loss functions; is the overall loss function, is the Lagrangian function, is a symmetric matrix. 3.The cold start user recommendation method based on heterogeneous transfer learning according to claim 1, wherein, performing spatial mapping processing, feature matching processing and distribution alignment processing on the second information data of the user based on a neural network autoencoder in sequence to obtain third information data of the user, and the specific process is as follows: Acquiring reconstructed data in a low-dimensional space , the formula is as follows: ； ； ； ； ； wherein, is a non-linear ReLU activation function, is a network parameter between (l) and (l + 1) layers, and are representations in (l) and (l + 1) layers, respectively; is an encoder; is a network parameter between (l) and (l + 1) layers; is a decoder; constructing a loss function in the spatial mapping processing process, and the formula is as follows: ； wherein is the mean square error; constructing a loss function in the feature matching processing process, and the formula is as follows: ； wherein is the mean square error; constructing a loss function in the distribution alignment processing process, and the formula is as follows: ； wherein, is the MMD loss. 4.The cold start user recommendation method based on heterogeneous transfer learning according to claim 3, characterized in that, constructing a target recommendation model according to the third information data of the user, and the specific process is as follows: constructing a total loss function of the target recommendation model, and the formula is as follows: ； wherein, is the overall loss function; calculating the gradient of the total loss function, and then updating the parameters through an optimizer. 5.The cold start user recommendation method based on heterogeneous transfer learning according to claim 1, characterized in that, The method further comprises: evaluating the results of the recommendation list, and the evaluation indexes include calculation accuracy, recall rate, normalized discount cumulative gain and hit rate.

6. The heterogeneous transfer learning based cold start user recommendation method according to claim 5, characterized in that, evaluating the results of the recommendation list, and the evaluation indexes include calculation accuracy, recall rate, normalized discount cumulative gain and hit rate, and the specific calculation formula is as follows: the calculation accuracy evaluation formula is as follows: ； in, It represents the number of users during cold starts. It is the first The actual preferred items of a cold-start user, and This is the top-ranked list of items recommended for the nth user during a cold start. the recall rate evaluation formula is as follows: ； the normalized discount cumulative gain evaluation formula is as follows: ； wherein is the relevance score of the item at position i, IDCG is set to ensure that a perfect ranking results in a normalized discounted cumulative gain value of 1; the hit rate evaluation formula is as follows: ； wherein, is defined as the total number of items recommended correctly divided by the total number of items preferred by all users.

7. A cold start user recommendation system based on heterogeneous transfer learning, characterized in that, The method comprises the following steps: a data acquisition module is configured to obtain first information data of a user, wherein the first information data of the user comprises social relationship data of warm users and cold-start users and historical interaction data of the warm users; a feature matching module is configured to perform feature enhancement processing on the first information data of the user to obtain second information data of the user, comprising: performing data feature enhancement through zero padding, and the formula is as follows: ； wherein, is second information data for the user, is social relationship data for the warm user, is historical interaction data for the warm user, is social relationship data for the cold start user; The data processing module is used for sequentially performing spatial mapping processing, feature matching processing and distribution alignment processing on the second information data of the user by using a self-encoder or a neural network-based self-encoder to obtain third information data of the user, comprising: using a self-encoder to perform spatial mapping processing on the second information data of the user to obtain the first information data of the user from the original space to the low-dimensional potential space, and then reconstructed by the self-encoder to obtain the unknown interaction of the cold start user, wherein the loss function in the spatial mapping processing process is constructed: ； ； wherein I is the identity matrix, X is the second information data of the user, ||·||F is the Euclidean norm, W is a mapping matrix, W T is the transpose of W, Tr(·) is the trace of the mapping matrix; constructing a loss function in the feature matching processing process; ； ； ， ； wherein, is a loss function in feature matching, and is a mapping matrix, is a transpose of the mapping matrix, is a number of items, is a feature dimension of user social relations; constructing a loss function in the distribution alignment processing process; ； ； ； ； wherein, is a loss function for distribution alignment processing, and are the i-th and j-th vectors of and respectively, is the number of users, is the number of warm users, is the number of cold-start users; a construction module is configured to construct a target recommendation model according to the third information data of the user; a recommendation list generation module is configured to generate a recommendation list of a cold-start user according to the target recommendation model.

8. A computer electronic device, comprising: The method comprises the following steps: a memory and a processor, the memory stores a computer program, and the processor implements the steps of the cold-start user recommendation method based on heterogeneous transfer learning in any one of claims 1-6 when executing the computer program.

Images