User cold start recommendation method based on individual and group self-adaptive meta-learning

Abstract

The application discloses a user cold start recommendation method based on individual and group adaptability meta-learning, comprising: inputting item interaction information into a related user identification component and a prior-based interest extraction component respectively; identifying related users showing the same interest for a specified item through the related user identification component, identifying related users with similar preferences and sharing similar knowledge by using a meta-path in a heterogeneous information network, obtaining prior knowledge corresponding to the found group, and realizing user group level adaptation; extracting the correlation relationship between different interests of the user and historical task knowledge through the prior-based interest extraction component, obtaining individual prior knowledge, and realizing user individual level adaptation; and guiding the global shared prior knowledge to adapt to the knowledge of the user by constructing an adapter for the user. The application has the advantages of simple implementation method, good generalization effect, high recommendation efficiency and precision, etc.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user recommendation, and particularly relates to a user cold start recommendation method based on individual and group adaptivity meta-learning. BACKGROUND

[0002] A recommendation system has the function of filtering out information irrelevant to user personalization, and can effectively alleviate the problem of information overload, and is widely deployed in various network services. The traditional recommendation system usually adopts the method of collaborative filtering, predicts the preferences of users according to the rich user-item interaction data such as clicks and evaluations of users, and then recommends personalized items to users according to the preferences of users. However, the user-item interaction of new users is usually very small, which will cause the so-called user cold start problem, and it is very challenging to learn the preferences of these new users.

[0003] In order to alleviate the user cold start problem, a common method is to integrate various types of auxiliary data into the recommendation system, for example, user profiles and side information of items, to enhance the preference representation of cold start users. However, the main limitation of this method is that the learned model may recommend the same items to users with similar features, thereby ignoring individual interests. In addition, due to the existence of privacy policies, auxiliary information is not easy to obtain.

[0004] The concept of meta-learning can solve the above problems, that is, the episodic meta-learning paradigm is used to model new users with sparse interaction data, and the prior knowledge shared globally in similar preference learning tasks (i.e., the initialization of the parameters of the model) is learned, so as to quickly adapt to the preference learning task of the cold start user on the basis of limited interaction data, thereby quickly adapting to a new user using prior knowledge and a small amount of interaction. Meta-learning is also known as learning to learn, and the purpose is to learn globally shared prior knowledge in similar learning tasks, which can quickly adapt to new tasks with a small number of labeled examples. Since meta-learning has the ability to quickly adapt, quickly adapting means quickly learning on previously unseen tasks using a small amount of data, so it can effectively alleviate the user cold start problem.

[0005] However, the recommendation setting using meta-learning in the prior art usually uses the same globally shared prior knowledge to guide the parameter initialization of all users, and does not fully utilize the prior knowledge brought by related users with similar preferences to improve the generalization ability of the meta-learning framework, so that it is not enough to distinguish the complex behavior and different interests of users, and thus the problem of poor generalization is caused. For example, Figure 1As shown in (a), traditional meta-learning recommendation methods, such as MeLU, assume that the global shared prior knowledge in all tasks, however, the global shared prior knowledge may not be sufficient to distinguish the complex behavior of users, and the fine-grained personal interests are ignored, resulting in poor generalization effect.

[0006] In order to overcome the above limitations of meta-learning recommendation, some practitioners propose a meta-learning recommendation method based on group level adaptation, which uses the information of related users with similar preferences to improve the generalization ability of the meta-learning architecture, but this kind of scheme only considers the group level adaptation problem and ignores the personal level adaptation problem, so it cannot model the personal level adaptation, and thus the actual generalization effect still needs to be improved. SUMMARY

[0007] The technical problem to be solved by the present application is that in view of the technical problems existing in the prior art, the present application provides a user cold start recommendation based on individual and group adaptive meta-learning, which can simultaneously consider group level adaptation and personal level adaptation to achieve adaptive meta-learning, and the implementation method is simple, the generalization effect is good, and the recommendation efficiency and accuracy are high.

[0008] To solve the above technical problems, the technical solution provided by the present application is:

[0009] A user cold start recommendation method based on individual and group adaptive meta-learning, comprising:

[0010] Inputting the item interaction information into a pre-constructed related user identification component RUI and a pre-constructed prior-based interest extraction component PIE respectively;

[0011] Identifying related users who show the same interest in the specified item through the related user identification component RUI, wherein a heterogeneous information network HIN is constructed by using the interaction information of all training users in the meta-training process in the related user identification component RUI, the meta-path in the heterogeneous information network is used to identify related users with similar preferences and share similar knowledge, and the prior knowledge corresponding to the found group is obtained after finding the group containing the related users through each user, to realize the adaptation at the user group level;

[0012] Extracting the correlation relationship between different interests and historical task knowledge of the user through the prior-based interest extraction component PIE, and learning the personal prior knowledge through the meta-training process to realize the adaptation at the individual level of the user;

[0013] An adapter for the user is constructed to guide the adaptation of the global shared prior knowledge to the knowledge of the user; the adapter for the user is used to predict the scores of the target user for each item according to the prior knowledge learned from the meta-training process to realize recommendation.

[0014] Further, the related path in the heterogeneous information network HIN is defined to represent the sequence of relations connecting two objects, the users appearing in the same related path have similar preferences, that is, the users appearing in the same related path are the related users, and when identifying the related users of the target user u, the related users of the target user u are obtained by collecting all related paths P starting from the target user u, calculating the correlation degree between the target user u and other users in the related path P, and selecting part of the related users with the highest correlation degree as the related user group of the target user u.

[0015] Further, the correlation degree Rel between the user u1 and the user u2 is calculated according to the following formula:

[0016]

[0017]

[0018] wherein, N (u1, e in , i n ) represents the number of neighbors of the user node u1 with the same relationship value e in and the node type i n .

[0019] Further, the related user identification component RUI further comprises a related user enhanced recommender for realizing related user recommendation, and the construction of the related user enhanced recommender comprises:

[0020] The representations of the users and the items are initialized using an embedding layer, and the features are connected, for a user u with m features e1~em, the initial embedding is defined as:

[0021]

[0022] wherein, z m represents the initial embedding of the user u, z u is a one-hot vector of the mth feature, and Z represents a corresponding feature embedding matrix;

[0023] The user embedding e u of the item is obtained by aggregating the interactions of the item:

[0024] e j =η(Mean(We u +b:j∈S u , u∈U))

[0025] wherein, S rel is the support set of the user u, Mean() is an average pooling operation, W and b are a weight matrix and a bias vector respectively, and η represents an activation function;

[0026] and get the item-based user embedding e rel :

[0027] e rel = η (Mean (We j + b: j ∈ S u , u ∈ U rel ))

[0028] Stack the user embeddings e rel to get the corresponding matrix E, and use attention operation to capture the contribution of different related users, i.e.,

[0029]

[0030] where e r represents the related user based on preference embedding;

[0031] Concatenate the user embedding e u of the item and the related user based on the preference embedding e to get the overall representation x u of the user, i.e.,

[0032]

[0033] The knowledge from related users with similar preferences is incorporated into the cold-start user representation as input to the prediction module to predict the rating of item i by user u The expression is as follows:

[0034]

[0035] where MLP represents a multi-layer perceptron, e i represents the embedding of item i.

[0036] Further, the prior-based interest extraction component PIE utilizes a multi-interest extraction layer to generate an interest vector of the user, and then organizes the interest vector into a graph structure to generate a multi-interest relationship graph MIRG to learn the relationship between the interest vectors, wherein the vertices in the multi-interest relationship graph MIRG represent different interests of the user, and the edges and corresponding edge weights are generated according to the similarity between the interests.

[0037] Further, the generation step of the multi-interest relationship graph MIRG includes:

[0038] First, a multi-interest extraction layer containing a dynamic routing algorithm is used to generate an interest vector of the user:

[0039]

[0040] where t jis the input vector, h is the encoder, E is the transformation matrix, e u is the user embedding of item i, e i represents the embedding of item i, w jk represents the weight connecting input t j and temporary output z k , and:

[0041] w jk = softmax(b jk )

[0042] where b jk is the routing log obeying Gaussian distribution;

[0043] A nonlinear squashing function is applied to the temporary output z k to obtain the interest vector C, the calculation expression of each element c k in the interest vector C is:

[0044]

[0045] After routing, all interactions are clustered into K c interest vectors C, and the value of K c of different users is adjusted according to the following formula using a heuristic rule:

[0046] K c = max(1, min(K2, log2|S u |))

[0047] where S u is the support set of user u;

[0048] The connection between any two interest vectors is regarded as an edge in the multi-interest relationship graph MIRG, and the corresponding edge weight is measured by the similarity between the two interest vectors, the calculation expression of the similarity between the interest vector c m and the interest vector c n is:

[0049]

[0050] where W c and b c are the learnable weight matrix and bias vector, and δ is a sigmoiid function.

[0051] Further, the prior-based interest extraction component PIE constructs a parameterized prior knowledge relation graph PKRG to learn the relationship between historical prior knowledge, each vertex of the prior knowledge relation graph PKRG describes a prior knowledge pattern, and edges are constructed according to the relationship between prior knowledge; an hypergraph is constructed by connecting the multiple interest relation graph IMRG of each user and the prior knowledge relation graph PKRG, and knowledge propagation is performed by using a graph neural network to propagate the most relevant prior knowledge in the prior knowledge relation graph PKRG to the multiple interest relation graph MIRG.

[0052] Further, the edge weight in the prior knowledge relation graph PKRG is defined as follows:

[0053]

[0054] wherein v m and v n are any two vertices in the prior knowledge relation graph PKRG, W v and b v are learnable parameters;

[0055] The expression of knowledge propagation by using a graph neural network is as follows:

[0056]

[0057] wherein M t denotes a message function, U t denotes a vertex update function, E t is a learnable weight matrix at t time step, W hg , H hg are adjacent matrix and feature matrix respectively.

[0058] Further, the expression of the user-specific adapter constructed is as follows:

[0059]

[0060] ω u =ωoo u

[0061] wherein FC denotes a fully connected layer activated by a sigmoid function, represents an element-wise product, ω represents prior knowledge, ω u represents specific knowledge of user u, x u represents overall representation, x u represents user representation;

[0062] The final prediction of an unobserved item is calculated as follows:

[0063]

[0064] Wherein, the MLP is provided with user-specific prior knowledge to guide the prediction process.

[0065] Further, the user-specific prior knowledge is locally updated to minimize the prediction loss of the user u by following the gradient descent steps, so as to obtain the personalized prior knowledge;

[0066]

[0067] Wherein, p is a local learning rate, L represents a prediction loss, and omega u * is personalized prior knowledge for guiding the prediction of the items in the query set Qu;

[0068] The global parameter theta is globally updated by minimizing the loss of the query set Qu, that is:

[0069]

[0070] Wherein, lambda is a global learning rate, theta = w U phi represents all parameters of the model, phi represents the remaining parameters except the prior knowledge omega, and B is the extracted task.

[0071] Compared with the prior art, the advantages of the present application are that: the personal and group adaptive meta-learning method of the present application absorbs the knowledge of individuals and groups, identifies trusted related users based on the meta-path in the heterogeneous information network through the use of a related user identification component (RUI), obtains similar knowledge, and thus realizes the group-level adaptation of users; through a prior-based interest extraction component (PIE), the relationship between different interests of the user and the historical task knowledge is explored, and thus the most relevant knowledge is obtained, realizing the individual-level adaptation of the user; at the same time, through an adapter for the user, the globally shared prior knowledge is adapted to the specific knowledge of the user, and the adaptability of the individual level and the group level can be combined, so that the two specific adaptabilities of the individual level and the group level of the user are obtained, the user-specific prior knowledge is obtained, the existing meta-learning recommendation model is improved, and better expression ability and generalization ability are achieved. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a schematic diagram of the principle comparison between the traditional meta-learning recommendation method and the method of the present application.

[0073] Figure 2 is a schematic diagram of the user interaction information network in the specific application embodiment.

[0074] Figure 3It is the implementation principle architecture diagram of the user cold start recommendation method based on individual and group adaptivity meta-learning of the embodiment.

[0075] Figure 4 It is the principle diagram of HIN and related path construction in the specific application embodiment.

[0076] Figure 5 It is the result diagram of the influence of the number of related users obtained in the specific application embodiment.

[0077] Figure 6 It is the result diagram of the influence of the number of vertices in MIRG obtained in the specific application embodiment.

[0078] Figure 7 It is the result diagram of the influence of the number of vertices in PKRG obtained in the specific application embodiment.

[0079] Figure 8 It is the result diagram of the influence of the number of local updates obtained in the specific application embodiment.

[0080] Figure 9 It is the result diagram of the influence of the size of the support set obtained in the specific application embodiment.

[0081] Figure 10 It is the result diagram of the influence of the embedding dimension obtained in the specific application embodiment.

[0082] Figure 11 It is the visualization analysis diagram obtained in the specific application embodiment. DETAILED DESCRIPTION

[0083] The present application is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present application is not limited thereby.

[0084] On the basis of solving the user cold start recommendation problem by adopting meta-learning, the present application realizes the rapid adaptation of cold start users at the individual and group levels, that is, the meta-learning has user-specific adaptivity, including the adaptation of two levels of user individual and group, learns various adaptivity for users, so that the user obtains the most relevant prior knowledge instead of the same global shared prior knowledge, so that the performance of meta-learning recommendation can be effectively improved, and the generalization ability is better, wherein not only the interest of the user is explored, but also the related user with similar historical preference is explored, so that knowledge customization and knowledge generalization can be realized.

[0085] As shown in (b) of the Figure 1 θ1 is customized by the learned prior knowledge θ, which contains the interest of task 1 (i.e. a user), and is generalized to task 3 according to the preference similarity between task 1 and task 3.Figure 2 For example, the figure shows users and their historical interactions, where users u1 and u3 have similar transferable prior knowledge for recommendations because they exhibit similar historical preferences, while user u2 has very different preferences from both u1 and u3. At the same time, each user has their own complex interests, which come from their historical interactions, such as u1 likes sports, games, and musical instruments, while u2 likes clothes and cosmetics.

[0086] As Figure 3 shown, the user cold start recommendation method based on individual and group adaptive meta-learning of the embodiment includes:

[0087] The item interaction information is respectively input to a pre-constructed related user identification component RUI and a pre-constructed prior-based interest extraction component PIE;

[0088] The related user identification component RUI identifies related users who show the same interest in the specified item. In the related user identification component RUI, a heterogeneous information network HIN is constructed by using the interaction information of all training users in the meta-training process. The meta-path in the heterogeneous information network is used to identify related users with similar preferences and shared similar knowledge. After finding the group containing the related users through each user, the prior knowledge corresponding to the found group is obtained to realize the adaptation at the user group level.

[0089] The prior-based interest extraction component PIE extracts the correlation relationship between different interests and historical task knowledge of the user, and learns the personal prior knowledge through the meta-training process to realize the adaptation at the user individual level.

[0090] An adapter for the user is constructed to guide the global shared prior knowledge to adapt to the knowledge of the user. The adapter for the user predicts the scores of the target user for each item according to the prior knowledge learned from the meta-training process to realize the recommendation.

[0091] The personal and group adaptive meta-learning method (MetaIG) of the embodiment is to combine the adaptability of the individual level and the group level by absorbing the knowledge of the individual and the group, so as to obtain the specific adaptability of the individual level and the group level of the user, obtain the specific prior knowledge of the user, and improve the existing meta-learning recommendation model to have better expression ability and generalization ability. Specifically, the embodiment first identifies the trusted related users based on the meta-path in the pre-defined heterogeneous information network through a related user identification component (RUI) to obtain similar knowledge, so as to realize the group level adaptation of the user; then, the most relevant knowledge is obtained by exploring the relationship between the different interests of the user and the historical task knowledge through a prior interest extraction component (PIE), so as to realize the individual level adaptation of the user. Then, an adapter for the user is used to guide the global shared prior knowledge to adapt to the specific knowledge of the user.

[0092] In the embodiment, it is assumed that U represents a user set, I represents an item set, and R represents an actual rating set. N{ui} is the available interaction of a specific user ui, where ui is in U, vj is in I, and r{i,j} describes the actual rating of ui to the item vj in R. |N{ui}| describes the number of interactions of ui. According to the setting of meta-learning, N{ui} = S{ui}∪Q{ui} is represented, where S{ui} is a support set containing the interaction items of ui, and Q{ui} is a query set containing the items to be predicted, so that the support set and the query set are mutually exclusive. Therefore, the task in the meta-learning setting is defined as follows in RS:

[0093] Definition of task: Given a user set U and an item set I, predict the preference rating behavior of a specific user u.

[0094] Generally, all tasks are divided into meta-training tasks T{tr} and meta-testing tasks T{te}. The support set and the query set of a task are composed of the rating items of ui. The support set only contains scarce items. During meta-training, the meta-learner learns on each support set and then updates on multiple query sets. In this way, the model can derive global shared prior knowledge (i.e., prior) among a set of users T{tr} and adapt the prior knowledge to new tasks. During meta-testing, for each task T{{ui}^{'}} in T{te}, its support set S{{ui}^{'} also contains some items rated by ui, but its query set Q{{ui}^{'} only contains the rating items to be predicted. The meta-learner can predict the ratings of the items in Q{{ui}^{'} according to S{{ui}^{'}} and the prior knowledge learned from meta-training.

[0095] The architecture of the MetaIG of the embodiment specifically includes three basic components: related user identification (RUI), prior-based interest extraction (PIE), and user-specific adapter, as shown in Figure 3 The RUI captures related user groups, enables group-level adaptation, and to some extent alleviates the data sparsity problem, and it contains a method for identifying related users and a related user-enhanced recommender for enhancing the representation of cold-start users. The PIE user obtains the relationship between the user's interest and the prior knowledge of the historical task to achieve personal-level adaptation. The user-specific adapter is used to guide the globally shared prior knowledge to adapt to the individual group's prior knowledge. Through the above MetaIG, both individual and group-specific adaptation is achieved, thereby achieving good generalization and rapid adaptation to cold-start users.

[0096] The MetaIG specifically includes three procedures: an adaptation process (implemented by the user-specific adapter in the figure), a local update, and a global update, and the adaptation process customizes the globally shared prior knowledge into user-specific knowledge. In the local update process, the meta-learner generates personalized knowledge for each task / user by using user-specific knowledge. In the global update process, the learned personalized knowledge is used to optimize all parameters. Once the model is well trained, it can easily adapt to new users, and only a few gradient descent steps are needed to achieve high-quality prediction.

[0097] In the embodiment, the related path in the heterogeneous information network HIN is defined to represent the sequence of relationships connecting two objects, and users appearing in the same related path have similar preferences, i.e., users appearing in the same related path are the related users, and when identifying the related users of the target user u, the target user u and the corresponding support set are obtained, all related paths P starting from the target user u are collected, the correlation between the target user u and other users in the related path P is calculated, and the highest correlation is selected. The part of the related users is selected as the related user group of the target user u.

[0098] Specifically, since the related users are identified through the related paths in the heterogeneous information network (HIN), which records the interactions between all training users and all training items. The above HIN is constructed by utilizing the interactions of all training users in the meta-training process. The specific definition is as follows:

[0099] Definition 1: Heterogeneous Information Network (HIN). The HIN can be defined as a graph G=(V,E,A,L), where each entity node v∈V, each connection edge e∈E, and each V conforms to the type mapping function V→A and E→L, where |A|+|L>2.

[0100] Definition 2: Relevance path. Given a path in a HIN, where each user node has the same rating to an item node, it is considered a relevance path.

[0101] As Figure 4 An example of a HIN and a relevance path is shown, where the HIN consists of two types of entity nodes (users and items) and one type of relation (rating). In this HIN, u1→i1→u3 is a relevance path, where u1 and u3 are users with similar interests because they both rated i1 the same. Obviously, u1→i2→u4 is not a relevance path because they rated i2 differently.

[0102] In meta-learning, the adaptability at the user group level enables relevant users with similar preferences to share similar knowledge, which helps to improve the generalization ability of the model. However, due to the scarcity of user interaction data in the user cold start scenario, there are few direct neighbors, and the information of users one hop away can be helpful. The users connected indirectly are called relevant users. Therefore, it is necessary to identify these relevant users to achieve group-level adaptation. To this end, the embodiment defines a relevance path, which is a sequence of relations connecting two objects, belonging to a meta-path on the defined heterogeneous information network. It is assumed that users appearing in the same relevance path have similar preferences because they have the same rating for an item, and such users are defined as the proposed relevant users. The relevance path can be specifically expressed as p{r}=u→i→u, where u is in U, r{ui is in R, and i is in I. In the example of the meta-learning, the dark path shows that users u1 and u3 both interacted with item i_1 and gave it a rating of 5, indicating that u1 and u3 may have similar preferences (both like i1). In contrast, u1 and u2 also interacted with the same item i2, but since they gave different scores, they expressed different interests in the item. Therefore, u1, u3 and u2, u4 are relevant users who may have similar preferences, while u1 and u2 are not. It should be noted that in the embodiment, only two-hop neighbors (i.e. the length of the relevance path is 2) are considered to avoid introducing too much noise due to too far distance. Figure 4 In the example of meta-learning, the dark path shows that users u1 and u3 both interacted with item i_1 and gave it a rating of 5, indicating that u1 and u3 may have similar preferences (both like i1). In contrast, u1 and u2 also interacted with the same item i2, but since they gave different scores, they expressed different interests in the item. Therefore, u1, u3 and u2, u4 are relevant users who may have similar preferences, while u1 and u2 are not. It should be noted that in the embodiment, only two-hop neighbors (i.e. the length of the relevance path is 2) are considered to avoid introducing too much noise due to too far distance.

[0103] Further, a measurement method is set up to evaluate the relevance strength between users based on the relevance path. It is assumed that u1 is connected to u2 through p{r n}=u1→i{n} and u2, the relevance between them is strengthened. The relevance Rel between user u1 and user u2 is calculated as follows:

[0104]

[0105]

[0106] where N(u1, e in , i n ) represents the number of neighbors of user node u1 with the same relationship value e in and node type i n , indicating the divergence of user preference. The larger N(u{1},e{in},i{n}) is, the smaller P(u{1}|e{in},i{n}) is, which means that the preference of user u1 for item i{n} is less valued. sum{n=1}^{n} means that when the same two users appear in multiple same correlation paths, their correlations will be superimposed instead of being covered, and the minimum n is 1.

[0107] Given a user u and her support set, collect all correlation paths starting from user u, denoted as P, and then calculate the correlation of u with other users in P through formula (1) and formula (2). Finally, select the top-ranked relevant users as the relevant user group of u, denoted as U{rel}, for subsequent model use.

[0108] In this embodiment, the relevant user identification component RUI further comprises constructing a correlation user enhanced recommender to implement correlation user recommendation, and the construction steps of the correlation user enhanced recommender comprise:

[0109] The representation of users and items is initialized using an embedding layer, and each feature embedding is connected. For a user u with m features e1~em, the initial embedding is defined as: m

[0110]

[0111] e m =Z T z m (4)

[0112] where, represents the initial embedding of user u, z m is a one-hot vector of the mth feature, and Z represents the corresponding feature embedding matrix;

[0113] This embodiment integrates the historical interaction of users and the information of the corresponding relevant user group through the support set of users, to capture the preferences reflected in the items and the relationship between the current user and others. Specifically, the user embedding e u of the item is obtained by aggregating the interactions of the item:

[0114] e u =η(Mean(We j +b:j∈S u， u∈U)) (5)

[0115] where, S​u is the support set of user u, Mean() is the average pooling operation, W and b are the weight matrix and bias vector respectively, and represents the activation function (LeakyReLU is used specifically);

[0116] and get the item-based user embedding e rel of each related user of user u

[0117] e rel = η(Mean(We j +b: j e S u , u e U rel )) (6)

[0118] where e rel and e u have the same shape. In addition, when user content is available, e re1 and e u may be further connected with their corresponding initial embeddings .

[0119] Stack the user embeddings e rel to get the corresponding matrix E, and use the attention operation to capture the contribution of different related users, i.e.

[0120]

[0121] where e r represents the related user based on the preference embedding.

[0122] Concatenate the user embedding e u of the item and the related user e based on the preference embedding to get the overall representation x u of the user, i.e.

[0123]

[0124] Considering the overall preference embedding x u of the user and the embedding e i of the item i, the knowledge from related users with similar preferences is incorporated into the cold-start user representation as input to the prediction module to predict the rating of the user u for the item i The expression is as follows:

[0125]

[0126] where MLP represents a multi-layer perceptron, and e i represents the embedding of the item i.

[0127] Then minimize the following loss for user u to understand the preferences of user u and optimize the parameters, i.e.

[0128]

[0129] where θ includes all the parameters in the model, Du represents a set of items to be predicted, r ui is the actual rating of item i by user u.

[0130] To explore the relationship between the user's interest and the underlying prior knowledge of the historical tasks, realize the adaptation at the individual level, and enhance the task representation of the model, the embodiment is based on the idea of MIND, uses a multi-interest extraction layer in the prior-based interest extraction component PlE to generate the user's interest vector, and then organizes the interest vector into a graph structure to generate a multi-interest relationship graph MIRG to learn the relationship between the interest vectors, wherein the vertices in the multi-interest relationship graph MIRG represent different interests of the user, and the edges and the corresponding edge weights are generated according to the similarity between the interests. The underlying knowledge obtained from the historical tasks is used to strengthen the task representation, so as to better propagate the relevant prior knowledge to the user's interest vector.

[0131] Given a user containing a set of historical interactions, each of which may represent the interest of user u, first explore the relationship between these interests, which is represented by a graph, i.e., a multi-interest relationship graph (MIRG), wherein the vertices in the MIRG represent different interests of the user. At the same time, the edges and the corresponding edge weights are generated according to the similarity between the interests. Specifically, first use a multi-interest extraction layer containing a dynamic routing algorithm to generate the user's interest vector, as follows.

[0132]

[0133]

[0134] where t j is the input vector (j∈Su), h is the encoder, E is the transformation matrix, e u is the user embedding of item, e i represents the embedding of item i, w jk represents the weight connecting the input t j and the temporary output z k , and the weight is calculated by a soft key function, that is:

[0135] w jk =softmax(b jk ) (13)

[0136] where b jk is the routing log obeying Gaussian distribution;

[0137] The temporary output z kThe interest vectors C are obtained by applying a non-linear squashing function to the vector C k The computational expression of c

[0138]

[0139] After routing, all interactions are clustered into K c interest vectors C. Since different users can have different numbers of interest vectors, the present embodiment uses a heuristic rule to adjust the value of K c for different users according to the following formula:

[0140] K c = max(1, min(K c , log2|S u |)) (15)

[0141] where S u is the support set of user u.

[0142] After the above steps, all interest vectors C are obtained for a user, which are used as vertices in the MIRG. The edges and corresponding edge weights are further defined.

[0143] The connection between any two interest vectors is regarded as an edge in the multi-interest relation graph MIRG, and the corresponding edge weight is measured by the similarity between the two interest vectors. The computational expression of the similarity between interest vector c m and interest vector c n is:

[0144]

[0145] where W c and b c are the learnable weight matrix and bias vector, and δ is a sigmoid function that normalizes the weights between 0 and 1. Thus, the formal representation of the MIRG is: where denotes a set of vertices, and A describes the adjacency matrix, which reflects the similarity between interests.

[0146] The user can be enabled to absorb the most relevant knowledge by using the basic knowledge obtained from the historical tasks, which can improve the expressiveness and generalization of the model. The embodiment further extracts a component based on prior interest PIE by constructing a parameterized graph, i.e., a prior knowledge relation graph (PKRG), to memorize and organize these structured knowledge, to learn the relationship between the historical prior knowledge, so that the MIRG of the cold start user can use the PKRG to obtain the most relevant prior knowledge, further promote the adaptation process, and improve the generalization ability of the model. Each vertex in the prior knowledge relation graph PKRG describes a prior knowledge mode, and the edge is constructed according to the relationship between the previous knowledge; by connecting the multiple interest relation graph IMRG of each user and the prior knowledge relation graph PKRG to construct a hypergraph, and using a graph neural network for knowledge propagation, the most relevant prior knowledge in the prior knowledge relation graph PKRG is propagated to the multiple interest relation graph MIRG.

[0147] Specifically, in order to extract and organize knowledge from the historical learning process, the embodiment constructs and maintains a parameterized prior knowledge relation graph G{pr}(PKRG) to store the prior knowledge and their relationships. For the PKRG, the representation matrix of all vertices is defined as where each vertex describes a prior knowledge mode (for example: the knowledge of how to predict the preference of a user who has bought different types of clothes is learned by a meta learner), and the edge is automatically constructed to reflect the relationship between the previous knowledge. When a new user comes, the relationship prior knowledge in the previous task can be effectively and automatically identified by referring to the prior knowledge stored in the PKRG, which is beneficial to the training of the new user and realizes the individual level adaptation of MetalG. Specifically, the vertex feature matrix of the PKRG is defined as where the vertex representation v k is automatically learned and updated during meta training, because the input of the model is a constantly changing user sequence. It should be noted that the vertex representation is randomly initialized, which allows a variety of prior knowledge to be encoded in the PKRG.

[0148] The embodiment specifically defines the edge weight in the prior knowledge relation graph PKRG as follows:

[0149]

[0150] where v m and v n are any two vertices in the prior knowledge relation graph PKRG, W v and b v are learnable parameters.

[0151] For simplicity, the embodiment represents the PKRG as

[0152] The embodiment obtains two graphs for a user: an interest-multiple relationship graph (IMRG) and a priori knowledge relationship graph (PKRG). To implement the IMRG, the most relevant priori knowledge can be absorbed by querying the PKRG. Further, by connecting the IMRG G{hg} and the PKRG G{pr} of each user, a hypergraph G{ir} is constructed, so that the query mechanism required in the meta-training process can be implemented, and the representations of the two graphs are optimized.

[0153] Specifically, the combination of the vertices in the IMRG G{ir} and the PKRG G{pr} constitutes the vertices in the hypergraph; the edges of the IMRG G{ir} and the PKRG G{pr} are also retained in the hypergraph. Then, the G{ir} and the G{ir} are connected, the connection therebetween is regarded as an edge of the hypergraph, and the corresponding edge weight is obtained by calculating the similarity between the corresponding two vertices.

[0154] Specifically, for each vertex cm in the G{ir} and each vertex vn in the G{pr}, the edge weight is which is defined as follows:

[0155]

[0156] where <,> represents the Euclidean distance, and Kv is the number of vertices representing Kv kinds of priori knowledge patterns in the G{pr}. Therefore, the definition of the hypergraph G{hg} is expressed as:

[0157]

[0158] where, and H hg respectively represent the adjacency matrix and the feature matrix of the hypergraph G{hg}. After obtaining the hypergraph G{hg}, it is hoped that the most relevant priori knowledge in the PKRG can be propagated to the MIRG, which is a message passing problem. The embodiment utilizes a graph neural network (GNN) to implement the required knowledge propagation mechanism, which follows a message passing neural network (MPNN).

[0159] Specifically, the expression for knowledge propagation by using the graph neural network GNN is:

[0160]

[0161]

[0162] where M t represents a message function, U t represents a vertex update function, E t is a learnable weight matrix at the t time step, and W hgH hg These are the neighbor matrix and the characteristic matrix, respectively.

[0163] After T time steps of GNN computation, the knowledge propagation interest representation G{iir} of MIRG is obtained, which serves as the top Kc row of HT, represented as follows. The final user / task representation of knowledge dissemination, xc, can be obtained as follows:

[0164]

[0165] This embodiment achieves high-quality prediction and good generalization by combining user-level adaptation and group-level adaptation. Group-level adaptation enables closely related users to share similar knowledge to improve generalization, while user-level adaptation acquires the most relevant knowledge to enhance personalization. Since the overall user representation xu reflects the relationships between users, while the user representation xc of knowledge propagation reflects the individualization of u.

[0166] This embodiment sets up a user-specific adapter to customize user-specific prior knowledge from globally shared prior knowledge. All parameters of MetaIG are represented as θ = w∪φ, where ω represents prior knowledge (i.e., the parameters of the meta-learner), and the rest are φ (e.g., embedding lookup tables). The constructed user-specific adapter expression is:

[0167]

[0168] ω u =ωοo u (twenty four)

[0169] Where FC represents a fully connected layer activated by the sigmoid function, ○ represents element-wise multiplication, and ω represents prior knowledge. u x represents specific knowledge of user u. u Representing the overall characterization, x u This represents the user's characteristics.

[0170] After the above steps, the globally shared prior knowledge ω is adjusted to user u's specific knowledge ωu. Therefore, the user obtains relevant knowledge from related users with similar preferences and from historical learning processes.

[0171] In this embodiment, the final prediction for an unobserved item is calculated using the following formula:

[0172]

[0173] MLPs, in particular, incorporate user-specific prior knowledge to guide the prediction process.

[0174] The local update is similar to the optimization of a common neural network, which is based on the support set to minimize the prediction loss, so that the initialization of the parameters converges to a good local optimum. After obtaining the specific prior knowledge of user u, the embodiment locally updates the user-specific prior knowledge through several gradient descent steps to minimize the prediction loss of u, thereby obtaining personalized prior knowledge.

[0175]

[0176] where ρ is the local learning rate, L represents the prediction loss, and ω u * personalized prior knowledge for guiding the prediction of items in the query set Qu;

[0177] The purpose of global update is to optimize all parameters of the model. In the training phase, a batch of tasks B is extracted from the meta-training task T{tr}. For each task, a step of gradient descent is taken to optimize all parameters according to the sum of the losses of the query set Q_u. The specific global parameter θ is globally updated by minimizing the loss of the query set Qu, that is:

[0178]

[0179] where λ is the global learning rate.

[0180] To verify the effectiveness of the present application, extensive experiments were conducted on three public and popular benchmark datasets: Yelp, MovieLens and DouBan Book, and were used for two recommendation scenarios: user cold start and non-cold start. The user cold start is constructed according to the prior art (MeLU and MetaHIN). Specifically, first, according to the joining time (or the first action time) of the user, the users of each dataset are divided into two groups (old users and new users). Then, each dataset is divided into meta-training and meta-testing datasets. The former only includes old users, while the latter only contains cold start users, which is referred to as the user cold start scenario (UC). The non-cold start scenario randomly selects 10% of the meta-training tasks as the traditional recommendation scenario, which is referred to as the non-cold start scenario (NC). Three popular indicators are used to evaluate the proposed MetaIIG and existing competitive methods, which are MAE, RMSE and nDCG@5, and the present MetaIG is comprehensively compared with three types of traditional baselines, (1) traditional methods including FM, NeuMF, GC-MC and DropoutNet; (2) traditional meta-learning methods including DropoutNet, MetaEmb and MeLU; (3) meta-learning methods that adapt to specific tasks including MAMO, CMML and CBML.

[0181] In the experiment, grid search is specifically used to adjust the hyperparameters. The default Xavier normal distribution is used to initialize all parameters, and the adaptive moment estimation (Adam) is used to optimize MetaIG, and each of the two MLP layers for prediction has 64 nodes. The embedding size of users, items and corresponding features is fixed at 32. The number of local updates is fixed at 1, and the batch size is fixed at 64 by default. For Yelp and DouBan Book, the number of vertices in PKRG is empirically set to 6, and for MovieLens, it is set to 8. For Yelp and DouBan Book, the number of vertices in MIRG is empirically set to 4, and for MovieLens, it is set to 6. For the three datasets, the number of related users is empirically fixed at 6. The global and local learning rates are set to 0.0005 and 0.001 for Yelp, 0.0005 and 0.005 for MovieLens, and 0.001 and 0.001 for DouBan Book. The sensitivity of some important hyperparameters will be discussed in the subsequent experimental section.

[0182] The performance of the model and other methods is empirically compared on the three datasets in the user cold start (UC) scenario and the non-cold start (NC) scenario, and the results are shown in Table 1. From the results, it can be obtained that:

[0183] (1) User cold start scenario: The upper half of Table 1 corresponds to the results in the user cold start scenario. On all datasets, MetaIG of the application always produces the best performance among all methods. For example, in terms of MAE, RMSE and nDCG@5, MetaIG of the application improves by an average of 8.88%, 9.76% and 5.70% in Yelp, and by an average of 11.35%, 12.29% and 7.57% in MovieLens. The traditional meta-learning methods (i.e. MetaEmb, MeLU, CMML and CBML) perform better than the traditional methods (i.e. FM, NeuMF, GC-MC and DropoutNet), which shows that the meta-learning framework can effectively alleviate the user cold start recommendation problem in most cases. In addition, in most cases, the traditional meta-learning methods are less competitive than the improved meta-learning methods (i.e. MAMO, CMML and CBML). This is because the improved meta-learning methods learn the adaptability of specific tasks, improve the expressiveness and generality of the model, which also verifies the advantage of learning task adaptation in the meta-learning model.

[0184] In addition, in the improved meta-learning method, the performance of MAMO is the worst, which may be because it relies heavily on user profile information, and the data set used cannot provide high-quality user profile information for it. The performance of CBML is the best, but it still performs worse than the MetaIG of the present application in the two scenarios, because CBML only considers the adaptability of the user group level of the traditional clustering method, while the MetaIG of the present application further explores the adaptability of the individual level, captures the knowledge most relevant to the user, and improves the expression ability of the model.

[0185] In summary, the performance of the traditional recommendation method is the worst, while the meta-learning recommendation method of the present application for learning specific tasks performs the best in the user cold start scenario with sparse interaction data.

[0186] (2) Non-cold start scenario. The performance of the MetaIG of the present application in the non-cold start scenario is shown in Table 1. It can be seen that the performance of the MetaIG of the present application is still better than all the baselines, which further confirms the robustness of the MetaIG. Since the data set evaluated is very sparse (for example, the sparsity of Douban book is 99.71%), even in the non-cold start scenario. Moreover, since the MetaIG of the present application adopts two user-oriented adaptabilities, it can alleviate the sparsity problem and achieve good generalization. In addition, it can also be seen that the performance of the baselines in the two recommendation scenarios is unstable, while the MetaIG of the present application achieves stable state-of-the-art performance in both recommendation scenarios. Finally, in the non-cold start scenario, the overall performance improvement of the MetaIG of the present application over the baselines is lower than that in the user cold start scenario, because the degree of data sparsity is relatively less severe.

[0187] Table 1: Test results in user cold start (UC) scenario and non-cold start (NC) scenario

[0188]

[0189] To verify the effectiveness of different components in the MetaIG of the present application, the following ablation experiments are further performed in this embodiment. This embodiment takes the performance in the UC scenario as an example, and the trend in the NC scenario is similar. The comparison results are shown in Table 2, wherein (I) "MetaIG" represents the complete model proposed; (II) "CBML" represents the best baseline; (III) "MetaIG-IPRM" represents the MetaIG without interest priority modeling; (IV) "MetaIG-AIRU" represents the MetaIG without the method of identifying relevant users; (V) "MetaIG-MIRG" represents the MetaIG of the present application without the multi-interest relationship graph, which uses the average pooling method to aggregate each interaction and feeds it back to the PKRG for interaction; and (VI) "MetaIG-PKRG" represents the MetaIG of the present application without the prior knowledge relationship graph, which uses the average pooling method to aggregate each interaction as a task representation.

[0190] Firstly, the interest-priority modeling (IPRM) component is deleted to verify its effectiveness. The results show that the performance of MetaIG-IPRM decreases significantly, which proves the effectiveness of IPRM and also means that it is effective to improve the representation ability of the model by exploring the historical task knowledge. Next, in order to verify the effectiveness of the method of identifying relevant users, it is deleted and only the use of IPRM to learn the user-specific adaptability is considered. Similarly, it is observed that the performance of AIRU decreases, which shows that the group-level adaptability learned by AIRU helps to improve the generalization ability of the model. The complete model MetaIG achieves the best performance, which shows that it is effective to combine the group-level adaptability and the individual-level adaptability as the user-specific adaptability of the recommendation model on the basis of meta-learning.

[0191] In addition, the performance of Meta-IPRM considering only AIRU is better than CBML, which shows that the designed method of identifying relevant users is better than the traditional clustering algorithm, because the traditional clustering algorithm is difficult to play a role in the user cold start scenario, because there is little labeled data. Finally, it is observed that MetaIG-AIRU performs better than MetaIG-IPRM, which shows that the designed individual-level adaptability performs better than the group adaptability. In other words, the performance improvement brought by fully absorbing the most relevant knowledge from the historical task knowledge is better than the performance improvement obtained from the relevant users. It is speculated that this may be because IPRM can capture more complex relationships between different knowledge.

[0192] The effectiveness of the graph structure in IPRM can be verified by the ablation experiment described above, and the experimental results clearly verify that the multi-interest relationship graph (MIRG) can help simulate different interests of users, thereby promoting performance improvement. In addition, the performance decline of MetaIG-PKRG verifies the necessity of the proposed prior knowledge relationship graph (PKRG).

[0193] Table 2: Ablation experiment results table

[0194]

[0195] The present embodiment further conducts parameter sensitivity experiments on three data sets to explore the influence of parameters on the recommendation performance. Here, only MAE and nDCG@5 in the UC scenario are taken as examples, and other trends are similar.

[0196] (1) Influence of the number of related users.

[0197] As shown in Figure 5 , the MetaIG of the present application realizes adaptation at the user group level by utilizing the similar knowledge of related users, and only the most relevant users are selected as the input of the model to verify how the number of related users will affect the performance of the model. As shown in Figure 5 , it can be obtained that for all data sets, the performance (MAE and nDCG@5) of the model rapidly increases from 0 to 6, reaches the best at 6, and then slightly decreases with the increase of the number, because more related users can introduce too much noise.

[0198] (2) Influence of the number of vertices in MIRG.

[0199] Since the MetaIG of the present application utilizes the multi-interest layer to extract the Kc interests of users, the present embodiment evaluates the effect of different numbers of vertices (i.e. Kc) in MIRG. As shown in Figure 6 , for the Yelp and Douban book data sets, the MetaIG of the present application reaches the best performance at Kc=4, and for the MovieLens data set, the best performance is reached at Kc=6. Due to the scarcity of user interaction data, a smaller Kc value can be preferred.

[0200] (3) Influence of the number of vertices in PKRG.

[0201] The prior knowledge relationship graph (PKRG) plays a crucial role in the MetaIG, and the present embodiment further evaluates the model performance when the number of vertices (i.e. Kv) in PKRG varies from 0 to 14. The corresponding results are as shown in Figure 7As shown, the trend of Kv is similar to that of Kc, except that in Yelp and Doubanbook datasets, MetaIG achieves the best performance when Kv = 6, and in MovieLens dataset, Kv = 8. For this phenomenon, it is considered that 6 or so vertices in PKRG can be sufficient to capture the potential relationship between historical task knowledge. In summary, if there is a larger dataset and more complex relationship, more vertices Kc and Kv may be needed.

[0202] (4) The impact of the number of local updates.

[0203] This embodiment further qualitatively analyzes the impact of the number of local updates on the performance of the model in the meta-training process. As shown in Figure 8 , the MAE performance of MetaIG in the UC and NC scenarios is plotted, and the number of local updates ranges from 0 to 5. As can be seen from the figure, the performance is best when there is one local update, and then the MAE gradually deteriorates as the number of local updates increases, which may be due to overfitting of the support set.

[0204] (5) The impact of the size of the support set.

[0205] In order to study the robustness of MetaIG, this embodiment obtains the performance changes of four methods (i.e. NeuMF, MeLU, CBML and MetaIG) when the size of the support set (i.e. |Su|) changes from 5 to 90. As shown in Figure 9 , as the size of the support set increases (i.e. more training data), all methods gradually achieve better MAE performance. However, when the support set becomes smaller, the performance of the MetaIG of the present application decreases the least among all methods, indicating that MetaIG has good robustness, especially in the user cold start scenario.

[0206] (6) The impact of embedding dimension.

[0207] This embodiment further analyzes the performance of different embedding dimensions from 16 to 256, as shown in Figure 10 . For the three datasets, the model performance is best at 32, which means that a larger embedding size can reduce the performance of the model and cause a large amount of calculation, and too small an embedding size can make the model difficult to learn effectively.

[0208] (7) Visual analysis.

[0209] This embodiment further gives a visual example on MovieLens, as shown in Figure 11As shown, as an explanation of the interest-prior relationship modeling method. In the figure, "P" represents different prior knowledge from the meta-training process, "Interest" represents the interest of a specific user. The darker the color, the higher the similarity weight. It can be seen that different interests can activate different prior knowledge with different similarity weights. In this way, the model can capture more personalized prior knowledge of users to adapt to global prior knowledge, improve the generalization ability of the meta-learning-based recommendation framework, and promote the improvement of recommendation performance.

[0210] The present application constructs an adaptive meta-learning method based on structured knowledge (MetaIG) for user cold start recommendation. The method consists of two user-specific adaptations, including user group-level adaptation and individual-level adaptation, to obtain the most relevant structured knowledge for different user interests, thereby achieving high-quality prediction and good generalization ability of the recommendation model. The adaptation at the group level enables related users with similar preferences to share similar prior knowledge locally, thereby enabling generalization for these users. The adaptation at the individual level enables users to obtain the most relevant knowledge from the historical learning process. At the same time, by constructing an adapter for the user, the user-specific prior knowledge is combined with the globally shared prior knowledge, which can significantly improve the expressiveness and robustness of the model.

[0211] The above is only the preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with the preferred embodiment, it is not intended to limit the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application, in accordance with the technical essence of the present application, shall fall within the scope of protection of the technical solution of the present application.

Claims

1. A user cold start recommendation method based on individual and group adaptive meta-learning, characterized in that, The method comprises the following steps: inputting the item interaction information into a pre-constructed related user identification component RUI and a pre-constructed prior-based interest extraction component PIE respectively; identifying related users showing the same interest in the specified item through the related user identification component RUI, wherein a heterogeneous information network HIN is constructed by utilizing the interaction information of all training users in the meta-training process in the related user identification component RUI, a meta-path in the heterogeneous information network HIN is used to identify related users having similar preferences and sharing similar knowledge, and after a group containing related users is found for each user, the prior knowledge corresponding to the found group is obtained to realize the adaptation at the user group level; the correlation between different interests and historical task knowledge of the user is extracted through the prior-based interest extraction component PIE, and the individual prior knowledge is learned through the meta-training process to realize the adaptation at the individual level of the user; an adapter for the user is constructed to guide the global shared prior knowledge to adapt to the knowledge of the user; the adapter for the user is used to predict the score of each item of the target user according to the prior knowledge learned from the meta-training process to realize the recommendation; in the prior-based interest extraction component PIE, a multi-interest extraction layer is used to generate an interest vector of the user, and then the interest vector is organized into a graph structure to generate a multi-interest relationship graph MIRG to learn the relationship between the interest vectors, wherein the vertices in the multi-interest relationship graph MIRG represent different interests of the user, and the edges and corresponding edge weights are generated according to the similarity between the interests; the generation step of the multi-interest relationship graph MIRG comprises: first, a multi-interest extraction layer containing a dynamic routing algorithm is used to generate an interest vector of the user: wherein, t j is an input vector, h is an encoder, E is a transformation matrix, e u is a user embedding for item e i denotes an item i embedding, w jk represents the weights connecting the input t j and the temporary output z k . wherein, b jk is the log of the route subject to a Gaussian distribution; For the temporary output z k The interest vector C is obtained by applying a non-linear squashing function to the elements of the vector C c k The computational expression for z is: After routing, all interactions are clustered into K c interest vectors C, using a heuristic rule to adjust the values of different users' interest vectors C according to the following formula: K c interest vectors C according to the following formula: wherein, S u support set for user u; The connection between any two interest vectors is regarded as an edge in a multi-interest relation graph MIRG, and the corresponding edge weight is measured by the similarity between the two interest vectors. c m The similarity between the interest vectors c n The calculation expression of the similarity between the interest vectors is: wherein, W c and b c are learnable weight matrices and bias vectors, and δ is a sigmoid function. in the prior-based interest extraction component PIE, a parameterized prior knowledge relationship graph PKRG is constructed to learn the relationship between the historical prior knowledge, each vertex in the prior knowledge relationship graph PKRG describes a prior knowledge mode, and the edges are constructed according to the relationship between the prior knowledge; a hypergraph is constructed by connecting the multi-interest relationship graph MIRG and the prior knowledge relationship graph PKRG of each user, and a graph neural network is used for knowledge propagation to propagate the most relevant prior knowledge in the prior knowledge relationship graph PKRG to the multi-interest relationship graph MIRG.

2. The individual and group self-adaptive meta-learning based user cold start recommendation method according to claim 1, characterized in that, In the heterogeneous information network HIN, a related path for representing a sequence of relationships connecting two objects is defined, and users appearing in the same related path have similar preferences, i.e. users appearing in the same related path are the related users, and when identifying the related users of the target user u, the following steps are included: obtaining the target user u and the corresponding support set, collecting all related paths P starting from the target user u, calculating the correlation degree between the target user u and other users in the related path P, and selecting the highest correlation degree part of the related users as the related user group of the target user u. 3.The individual and group self-adaptive meta-learning based user cold start recommendation method according to claim 1, characterized in that, The correlation between user u1 and user u2 is calculated according to the following formula Rel : where N(u1, e in , i n ) represents the number of neighbors of user node u1 with the same relationship value e in and node type i n . 4.The individual and group self-adaptive meta-learning based user cold start recommendation method according to claim 1, characterized in that, The related user identification component RUI further comprises a correlation user enhanced recommender for realizing correlation user recommendation, and the construction steps of the correlation user enhanced recommender comprise: Using embedding layers to initialize the representation of users and items and concatenate the respective feature embeddings, for m features e 1～ e m The initial embedding for a user u is defined as: ； wherein, denotes the initial embedding of a user u, z m is a one-hot vector of the mth feature, and Z represents the corresponding feature embedding matrix; Obtaining a user embedding e for a project by aggregating interactions of the project u : where S u is the support set of user u, Mean() is the average pooling operation, W and b are the weight matrix and bias vector respectively, and η represents the activation function. and obtaining an item-based user embedding e for each relevant user of the user u rel : Embedding e of the user rel Stacking up gives the corresponding matrix E r and the attention operation captures the contribution of different relevant users, i.e.: where e r represents a related user based on a preference embedding; Embedding users of the project e u And concatenating the related users based on the preference embedding e to get the overall representation of the user x u That is: incorporating knowledge from related users with similar preferences into the cold-start user representation as input to the prediction module, predicting a rating of an item i by a user u , the expression is as follows: wherein MLP represents a multilayer perceptron, e i represents an item i embedding.

5. The individual and group self-adaptive meta-learning based user cold start recommendation method according to claim 1, characterized in that, The edge weights in the prior knowledge relationship graph PKRG are defined according to the following formula: wherein, v m and v n are any two vertices of a priori knowledge relation graph (PKRG), W v and b v are learnable parameters; The expression for knowledge propagation using a graph neural network is: wherein, M t denotes a message function, U t denotes a vertex update function, E t is t a learnable weight matrix for the time step, W hg , H hg are an adjacency matrix and a feature matrix, respectively. 6.The user cold start recommendation method based on individual and group self-adaptive meta-learning according to any one of claims 1-5, characterized in that, The constructed expression of the adapter for the user is: where FC denotes a fully connected layer activated by a sigmoid function, denotes an element-wise product, ω denotes prior knowledge, ω u denotes specific knowledge of the user u, x u denotes an overall representation, x c denotes a user representation; The final prediction for an unobserved item is calculated according to the following formula: wherein, MLP with user-specific prior knowledge to guide the prediction process.

7. The individual and group self-adaptive meta-learning based user cold start recommendation method according to claim 6, characterized in that, The personalized prior knowledge is obtained by locally updating the user-specific prior knowledge to minimize the prediction loss of the user u according to the following gradient descent steps: wherein, p is a local learning rate, L denotes the prediction loss, in the meta-testing process, ω u * is the personalized prior knowledge to guide the prediction of the items in the query set Qu; The global parameters θ are updated globally by minimizing the loss of the query set Q u, i.e. where λ is the global learning rate, denotes all parameters of the model, φ denotes the rest of the parameters except the prior knowledge ω, B is the task of extraction.

Images