A cold start recommendation method and related device

Abstract

The present application provides a cold start recommendation method and related device. The method includes: respectively extracting the features of a target user, a first meta-path set, a second meta-path set, an item to be scored, and an item that has been scored, and mapping the extracted features of the first meta-path set, the second meta-path set, the item to be scored, and the item that has been scored to the same embedding space; determining a scoring offset of the item to be scored relative to the item that has been scored based on the features of the target user and the features mapped to the embedding space; determining a scoring prediction value of the target user for the item to be scored according to the scoring value of the target user for the item that has been scored and the scoring offset, so as to determine whether to recommend the item to be scored to the target user based on the scoring prediction value. Based on the features mapped to the embedding space and the features of the target user, the present application can determine an accurate scoring offset, and further, the scoring prediction value determined based on the scoring offset is more consistent with the true scoring value, and the recommendation result is more accurate.

Description

Technical Field

[0001] The present application relates to the technical field of cold start recommendation, and particularly to a cold start recommendation method and related devices. Background Art

[0002] In the big data era, due to its ability to effectively alleviate the problem of information overload, recommendation systems have been deployed in a considerable number of Internet services.

[0003] Currently, the recommendation algorithms adopted by recommendation systems complete the recommendation tasks based on existing data and the similarity of objects. However, when new objects (such as new items or new users) appear in the recommendation system, due to the sparsity of interactions between users and items, the recommendation system usually cannot work effectively at this time, resulting in the cold start problem. For example, when a newly registered user enters an e-commerce platform, the user profile of this user is almost blank, and traditional recommendation algorithms usually rely on user profile information such as the user's browsing history, friends, orders, etc. to recommend content to the user. When facing new users, traditional recommendation algorithms cannot accurately complete the recommendation.

[0004] In summary, due to the limited interactions of new objects, traditional recommendation algorithms are trapped in a dilemma of lack of information and cannot accurately recommend items to new users or recommend new items to users. Summary of the Invention

[0005] In view of this, the present application provides a cold start recommendation method and related devices, which are used to solve the problem in the prior art that items cannot be accurately recommended to new users or new items cannot be recommended to users. The technical solutions are as follows:

[0006] A cold start recommendation method includes:

[0007] Obtain a first meta-path set of a target user and an item to be scored, a second meta-path set of the target user and scored items, and the scoring values of the target user for the scored items, where the first meta-path set and the second meta-path set are sets of meta-paths starting from the corresponding items in a heterogeneous information network, and the target user is a new user who has scored a small number of items and / or the item to be scored is a new item;

[0008] Extract the features of the target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored items respectively, and map the extracted features of the first meta-path set, the second meta-path set, the item to be scored, and the scored items to the same embedding space;

[0009] Based on the features of the target user and the features mapped to the embedding space, determine the scoring offset of the item to be scored relative to the scored items;

[0010] Determine the predicted rating value of the item to be rated by the target user based on the rating value and rating offset of the rated items by the target user, so as to determine whether to recommend the item to be rated to the target user based on the predicted rating value.

[0011] Optionally, the process of extracting the features of the target user includes:

[0012] Extract the features of multiple information dimensions of the target user;

[0013] Determine the adaptive weights of the features of multiple information dimensions through a one-dimensional convolutional algorithm;

[0014] Perform weighted summation on the features of multiple information dimensions based on the adaptive weights to obtain the features of the target user extracted.

[0015] Optionally, the first meta-path set and the second meta-path set respectively include multiple meta-path subsets, and the multiple meta-path subsets are obtained by dividing the meta-paths in the corresponding meta-path set according to the meta-path types;

[0016] The process of extracting the features of the one-way meta-path set includes:

[0017] Extract the features of multiple meta-path subsets included in the meta-path set;

[0018] Sort the features of the multiple meta-path subsets included in the meta-path set in descending or ascending order of similarity, and perform multi-level one-dimensional convolutional processing on the sorted features of the multiple meta-path subsets. Among them, the convolutional parameters of each level of one-dimensional convolutional processing are different and the length of the convolutional kernel is at least 2. In the first-level one-dimensional convolutional processing, perform one-dimensional convolutional processing on the features of every two adjacent meta-path subsets in the sorted multiple meta-path subsets to obtain the fused features under the first-level one-dimensional convolutional processing. In each subsequent level of one-dimensional convolutional processing, perform one-dimensional convolutional processing on every two adjacent fused features in the fused features under the previous level of one-dimensional convolutional processing until only one fused feature remains, which is used as the features of the meta-path set.

[0019] Optionally, extracting the features of multiple meta-path subsets included in the meta-path set includes:

[0020] For each meta-path subset included in the meta-path set, extract the features of each meta-path in the meta-path subset, calculate the average value of the features of each meta-path in the meta-path subset, and use the calculated average value as the features of the meta-path subset to obtain the features of each meta-path subset included in the meta-path set.

[0021] Optionally, mapping the features of the first meta-path set, the second meta-path set, the item to be rated, and the rated items to the same embedding space includes:

[0022] Concatenate the features of the extracted target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored items to obtain the concatenated features;

[0023] Map the features of the first meta-path set, the second meta-path set, the item to be scored, and the scored items included in the concatenated features to the same embedding space.

[0024] Optionally, the scored items include multiple items, and the second meta-path set includes the second meta-path sets corresponding to each of the items included in the scored items;

[0025] Concatenate the features of the extracted target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored items to obtain the concatenated features, including:

[0026] For each item included in the scored items, concatenate the features of the extracted target user, the item to be scored, the first meta-path set, the item, and the second meta-path set corresponding to the item to obtain the concatenated features corresponding to the item;

[0027] Map the features of the first meta-path set, the second meta-path set, the item to be scored, and the scored items included in the concatenated features to the same embedding space, including:

[0028] Map the target features included in the concatenated features corresponding to each item to the same embedding space, where the target features included in the concatenated features corresponding to an item include the features of the first meta-path set, the features of the item to be scored, the features of the second meta-path set corresponding to the item, and the features of the item.

[0029] Optionally, the scoring offset of the item to be scored relative to the scored items includes the scoring offset of each item included in the item to be scored relative to the scored items, and the scoring value of the target user for the scored items includes the scoring value of the target user for each item;

[0030] Determine the scoring prediction value of the target user for the item to be scored according to the scoring value of the target user for the scored items and the scoring offset, including:

[0031] Add the scoring value of the target user for each item to the scoring value of the item to be scored relative to the corresponding item to obtain the scoring prediction value with each item as the prediction benchmark;

[0032] Calculate the average value of the scoring prediction values with all the items included in the scored items as the prediction benchmarks, and use this average value as the scoring prediction value of the target user for the item to be scored.

[0033] A cold start recommendation device, including:

[0034] A basic data acquisition unit for acquiring a first meta-path set of a target user and an item to be scored, a second meta-path set of the target user and scored items, and the scoring values of the target user for the scored items. Herein, the first meta-path set and the second meta-path set are sets of meta-paths starting from the corresponding items in a heterogeneous information network, and the target user is a new user who has scored a small number of items and / or the item to be scored is a new item;

[0035] A feature extraction and mapping unit for respectively extracting the features of the target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored items, and mapping the features of the first meta-path set, the second meta-path set, the item to be scored, and the scored items to the same embedding space;

[0036] A scoring offset determination unit for determining the scoring offset of the item to be scored relative to the scored items based on the features of the target user and the features mapped to the embedding space;

[0037] A scoring value prediction unit for determining the scoring prediction value of the target user for the item to be scored according to the scoring value of the target user for the scored items and the scoring offset, so as to determine whether to recommend the item to be scored to the target user based on the scoring prediction value.

[0038] A cold start recommendation device, including a memory and a processor;

[0039] The memory is used for storing programs;

[0040] The processor is used for executing the program to implement each step of the cold start recommendation method as described in any one of the above.

[0041] A readable storage medium, on which a computer program is stored. The computer program, when executed by a processor, implements each step of the cold start recommendation method as described in any one of the above.

[0042] As can be seen from the above technical solutions, the cold start recommendation method provided by this application can obtain the first meta-path set and the second meta-path set from the heterogeneous information network, capture more semantic information under the condition of data scarcity of new users and / or new items, which is beneficial to the accuracy of score prediction. At the same time, this application can map the features of the first meta-path set, the features of the second meta-path set, the features of the item to be scored, and the features of the scored items to the same embedding space. In the same embedding space, it is easy to see the deviation of the features related to the item to be scored and the scored items respectively. The features mapped to the embedding space combined with the features of the target user can accurately determine the score offset of the item to be scored relative to the scored item. The score prediction value determined based on this score offset is more consistent with the true score value of the target user for the item to be scored. Furthermore, based on the score prediction value, items are recommended for the target user, and the recommendation result is more accurate. Description of the Drawings

[0043] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0044] Figure 1 It is a schematic diagram of the training and testing process based on the support set and the query set in meta-learning;

[0045] Figure 2 It is a schematic diagram of the process of a cold start recommendation method provided by an embodiment of this application;

[0046] Figure 3 It is a schematic diagram of the extraction process of the features of the target user provided by an embodiment of this application;

[0047] Figure 4 It is a schematic diagram of the extraction process of the features of the meta-path set provided by an embodiment of this application;

[0048] Figure 5 It is a schematic diagram of the process of realizing cold start recommendation provided by an embodiment of this application;

[0049] Figure 6 It is a schematic diagram of the feature mapping provided by an embodiment of this application;

[0050] Figure 7 It is a schematic diagram of the structure of the score prediction model provided by an embodiment of this application;

[0051] Figure 8 It is a schematic diagram of the structure of a cold start recommendation device provided by an embodiment of this application;

[0052] Figure 9 This is a hardware structure block diagram of a cold start recommendation device provided by an embodiment of the present application. Detailed implementation manners

[0053] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0054] To facilitate the description of the cold start recommendation method provided by the present application, the relevant terms applied in the present application will be explained below.

[0055] Heterogeneous Information Network (HIN): A HIN can be defined as a graph G = {V, E, O, R}, where V is the set of nodes, E is the set of edges, O is the set of node classes, and R is the set of relationship classes. Each node and edge follows a mapping relationship When at least one of the number of node classes and the number of relationship classes is greater than 2, this graph can be called a HIN. The heterogeneous information network supplements more heterogeneous information and establishes connections between different types of objects. For example, in a HIN, authors, publishers, etc. are no longer just auxiliary features of books, but a type of object. Different books can establish connections through these new objects, and this connection is called a meta-path, and the length and type of the meta-path are not limited. For example, the meaning of the path "user - book - publisher - book" is "books with the same publisher as the books evaluated by the user". This path connects objects that cannot be connected in other methods but have an objective association, thereby constructing a more complex graph from limited information.

[0056] Meta-Path: It is a path composed of multiple relationships connected end to end. In the graph G = {V, E, O, R}, a meta-path P with length l can be expressed as Considering that in this article, there is only one relationship between any two types of nodes, the meta-path P can be briefly expressed as P = O1O2O3...O l+1 . As a more complex graph structure, compared with other simple representation methods, a HIN contains richer semantic information, and meta-paths help to extract this rich semantic information.

[0057] Meta-learning: Regarding several pattern recognition tasks as the main body of learning, the aim is to learn a global experience applicable to tasks, rather than taking samples as the main body as in traditional machine learning. Therefore, meta-learning can directly transfer the experience of making recommendations for other users to target users or items, regardless of whether they are new users or items. For meta-learning, the experience abstracted from existing tasks of making recommendations for other new users can be used to quickly and simply recommend appropriate products for target users.

[0058] As Figure 1 shown, meta-learning requires a labeled dataset D = {(x, y)} and a target task paradigm F = {S, Q}, generally divided into two stages, namely meta-training and meta-testing. First, generate a specific sub-dataset D T = {D train , D test} in the dataset D, and then generate a supervised task T based on D T according to the paradigm F. The goal of each task is to predict the query set Q given the known support set S, that is, the goal of task T i is to predict the query set Qi given the known support set Si, where i = 1, 2,..., n. When performing meta-training based on D train , calculate a loss independently for each task. We can use the idea of permutation and combination to generate a large number of tasks for training on a limited dataset. Since each task only requires very few or even only one labeled sample, few-shot learning is a perfect test platform for adapting to the meta-learning idea.

[0059] Support set: The user himself / herself, the items that the user has rated, and the user's ratings for these items.

[0060] Query set: Items that the user has not rated, and the goal of the task is to predict the user's ratings for these items.

[0061] Metric learning, a branch under meta-learning, studies how to learn a distance function on a specific task so that this distance function can help algorithms based on neighbors (such as KNN, k-means, etc.) achieve better performance.

[0062] Cold start recommendation: Let U and I represent the sets of users and items respectively, and the partially unknown matrix R represents the ratings of user u for item i, that is, R = {r u，i | u ∈ U, i ∈ I}. The goal of the recommendation system is to predict the unknown ratings in the matrix R. If a specific user ux has only very few known values {r u,i : u = u x, i ∈ I}, which means that user ux is a new user who has made very few ratings. This scenario is user cold - start. Similarly, if an item i has been rated a limited number of times, it can be regarded as a new item, and this situation is item cold - start. When both the user and the item are new, it is user - item cold - start.

[0063] This application provides a cold - start recommendation method. Optionally, this cold - start recommendation method can run on a server or a server cluster. Next, the cold - start recommendation method provided by this application will be introduced in detail through the following embodiments.

[0064] Please refer to Figure 2 , which shows a schematic flowchart of the cold - start recommendation method provided by the embodiments of this application. This cold - start recommendation method may include:

[0065] Step S101, obtain the first meta - path set between the target user and the item to be rated, the second meta - path set between the target user and the rated items, and the rating values of the target user for the rated items.

[0066] Among them, the first meta - path set and the second meta - path set are sets of meta - paths starting from the corresponding items in the heterogeneous information network. Specifically, the first meta - path set is the set of meta - paths starting from the item to be rated in the heterogeneous information network, and the second meta - path set is the set of meta - paths starting from the rated items in the heterogeneous information network.

[0067] Taking the users including UserName1 and UserName2, the items including Books 1 - 4, where the authors of Books 2 - 4 are all Author 1, and the ratings including: UserName1 rated Book 2 with 4 points, Book 1 with 2 points, UserName2 rated Book 2 with 5 points, and Book 3 with 4 points as an example. Assuming the target user is UserName1 and the rated item is Book 2, the second meta - path set starting from Book 2 includes: {UserName1 - Book 2, UserName1 - Book 2 - UserName2 - Book 3, UserName1 - Book 2 - Author 1 - Book 3, UserName1 - Book 2 - Author 1 - Book 4}.

[0068] In this application, the target user is a new user who has rated a small number of items and / or the item to be rated is a new item. Among them, when the target user is a new user, this application solves the problem of recommending items for the target user under user cold - start. When the item to be rated is a new item, this application solves the problem of recommending items for the target user under item cold - start.

[0069] It should be noted that the "item" in the embodiments of the present application is a general term for things that can determine user preferences through score values, and the present application does not make specific limitations. For example, the "item" can be movies, books, articles, food, etc.

[0070] For the convenience of subsequent description, in the embodiments of the present application, the target user is represented by u, the item to be scored is represented by iq, the first meta-path set between the target user and the item to be scored is represented by P iq The score value of the target user for the item to be scored is represented by r u,iq The scored item is represented by is, the second meta-path set between the target user and the scored item is represented by P is The score value of the target user for the scored item is represented by r u,is It is represented.

[0071] Step S102: Extract the features of the target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored item respectively, and map the features of the extracted first meta-path set, second meta-path set, item to be scored, and scored item to the same embedding space.

[0072] Specifically, the process of extracting features in this step can be implemented by a feature extraction module. The feature extraction module includes a feature extraction network FN. In the present application, the target user u, the scored item is, the second meta-path set P is , the item to be scored iq, and the first meta-path set P iq are input into FN, and FN can extract the features (feature map) of the target user u, the first meta-path set P iq , the second meta-path set P is , the item to be scored iq, and the scored item is respectively.

[0073] Considering that traditional metric learning methods always default that the query sample (i.e., the item to be scored) exists in the support set (i.e., the target user, the scored item, the second meta-path set, and the score value of the target user for the scored item). For example, if the true score value of the target user for the item to be scored is 4 points, then the present application defaults that the score value of the target user for the scored item includes the score of 4 points. In this case, traditional metric learning methods directly calculate the feature similarity between the query sample and each support sample, and take the category corresponding to the support sample with the highest similarity as the query sample category.

[0074] However, in the cold start recommendation scenario, it is usually the case that the evaluations of a user, especially a new user, are likely not to cover the entire range of values. That is to say, there may be samples in the query set that do not belong to any of the categories in the support set. For example, for a rating value range from 1 to 5, a new user has only two rating records, which are 1 and 4 respectively. Now, to predict the rating value of an item to be rated, with the true rating value being 3, but there is no category of 3 in the support set, the accuracy of rating prediction using traditional metric learning methods is relatively low.

[0075] The inventor of this case thought that the feature mapping in the same embedding space could be carried out by setting an offset module (feature embedding module and offset module). Specifically, the features of the target user, the first meta-path set, the second meta-path set, the item to be rated, and the rated items extracted are input into the offset module, and in the offset module, the features of the first meta-path set, the features of the second meta-path set, the features of the item to be rated, and the features of the rated items are all mapped to the same embedding space.

[0076] In an optional embodiment, the process of "mapping the features of the first meta-path set, the second meta-path set, the item to be rated, and the rated items to the same embedding space" in this step may include: splicing the features of the target user, the first meta-path set, the second meta-path set, the item to be rated, and the rated items extracted to obtain the spliced features, and mapping the features of the first meta-path set, the second meta-path set, the item to be rated, and the rated items included in the spliced features to the same embedding space.

[0077] Here, the formula used for feature splicing is as follows in formula (1).

[0078]

[0079] In the formula, F u,is,iq represents the spliced features, c(·,·,…) represents splicing the features of the target user, the first meta-path set, the second meta-path set, the item to be rated, and the rated items at the channel level, f u (u) represents the features of the target user, f i (is) represents the features of the rated item is, f p (P is ) represents the features of the second meta-path set P is , f i (iq) represents the features of the item to be rated iq, f p (P iq ) represents the features of the first meta-path set P iq , and M represents the total number of rated items.

[0080] After splicing according to the above formula (1) in this step, then map f i (is), f p (P is ), f i (iq) and f p (P iq ) in the combined features to the same embedding space.

[0081] In this step, map the features of the first meta-path set, the features of the second meta-path set, the features of the item to be scored, and the features of the scored items to the same embedding space. It is easy to see the differences between the item to be scored and the scored items in the embedding space, which is beneficial to determining the scoring offset in the subsequent steps.

[0082] Step S103: Determine the scoring offset of the item to be scored relative to the scored items based on the features of the target user and the features mapped to the embedding space.

[0083] The process of determining the scoring offset bias in this application can be implemented by an offset module. The offset module contains an offset calculation function R, which can determine the scoring offset of the item to be scored relative to the scored items based on the features of the target user and the features mapped to the embedding space.

[0084] Here, the process of the offset calculation function R determining the scoring offset of the item to be scored relative to the scored items based on the features of the target user and the features mapped to the embedding space can be implemented by the following formula (2).

[0085]

[0086] In the formula, b u,is,iq represents the scoring offset of the item iq to be scored by the target user u relative to the scored item is (i.e., the above bias), and F′ u,is,iq represents the combined features after feature mapping.

[0087] It should be noted that the way of first feature splicing and then feature mapping given in the embodiments of this application is only an optional implementation manner. In addition, feature mapping can be performed first and then feature splicing, and this application does not make specific limitations on this.

[0088] Step S104: Determine the scoring prediction value of the target user for the item to be scored according to the scoring value of the target user for the scored items and the scoring offset, so as to determine whether to recommend the item to be scored to the target user based on the scoring prediction value.

[0089] Specifically, the rating offset characterizes the distance between the rating value of the item to be rated by the target user and the rating values of the items already rated by the target user. Combining this distance with the rating values of the items already rated by the target user can determine the rating value of the item to be rated by the target user. To distinguish the determined rating value from the true rating value of the item to be rated by the target user, the embodiment of this application defines the determined rating value as the rating prediction value.

[0090] By adopting the above steps S101 to S104, the embodiment of this application can determine the rating prediction values corresponding to all items to be rated recommended to the target user, and then recommend one or more items to be rated with the highest rating prediction values to the target user.

[0091] The cold start recommendation method provided by this application can obtain the first meta-path set and the second meta-path set from the heterogeneous information network, can capture more semantic information under the condition of data scarcity of new users and / or new items, which is beneficial to the accuracy of rating prediction; at the same time, this application can map the features of the first meta-path set, the features of the second meta-path set, the features of the items to be rated, and the features of the items already rated to the same embedding space. In the same embedding space, it is easy to see the deviation of the features related to the items to be rated and the items already rated respectively. The features mapped to the embedding space combined with the features of the target user can accurately determine the rating offset of the item to be rated relative to the item already rated. The rating prediction value determined based on this rating offset is more consistent with the true rating value of the item to be rated by the target user, and then recommend items to the target user based on the rating prediction value, and the recommendation result is more accurate.

[0092] In some embodiments of this application, the process of extracting the features of the target user in the foregoing step S102 is introduced.

[0093] It can be understood that the target user may have multiple features such as an identity identification number (ID), age, gender, etc. In the prior art, generally, the average of different features is taken as the features of the target user, treating each feature equally; there are also some that set weights for the features, hoping that the attention mechanism can more effectively fuse different features. However, for different users, the weights of each feature remain fixed. However, the inventor of this case found through in-depth research that in practical applications, for different users, the importance of each feature is different. If the weights of each feature remain fixed, it may lead to the weakening of important features of some users.

[0094] To avoid this situation, the inventors of this case proposed a method of using one-dimensional convolution to dynamically set adaptive weights for different features. To this end, the process of extracting the features of the target user may include: extracting the features of multiple information dimensions of the target user, determining the adaptive weights of the features of multiple information dimensions through the one-dimensional convolution algorithm, and performing weighted summation on the features of multiple information dimensions based on the adaptive weights to obtain the features of the extracted target user.

[0095] As Figure 3 shown, taking the features of multiple information dimensions including the ID feature u[Id], gender feature u[Gender], occupation feature u[Occupation], and age feature u[Age] as an example, this application can determine the respective adaptive weights of u[Id], u[Gender], u[Occupation], and u[Age] through the one-dimensional convolution algorithm. In the case of filtering using a convolution kernel with a length of 1, the feature f u (u) of the target user can be expressed as:

[0096] f u (u) = w1u[Id] + w2u[Gender] + w3u[Occupation] + w4u[Age] Equation (3)

[0097] In the formula, w1 represents the adaptive weight of the ID feature u[Id], w2 represents the adaptive weight of the gender feature u[Gender], w3 represents the adaptive weight of the occupation feature u[Occupation], and w4 represents the adaptive weight of the age feature u[Age]. The specific values of these four weights may be different when the users are different.

[0098] As introduced in the foregoing embodiments, the process of extracting features in step S102 can be implemented by the feature extraction network FN in the feature extraction module. In this application, the goal of f u in FN is to learn the convolution model during training, and in this way, the process of enabling FN to establish attention is given, so that FN can adaptively focus on important features.

[0099] It should be noted that the embodiments of this application do not limit the length of the convolution kernel and the number of convolution layers. If the process of determining the adaptive weights is to be made more complex and non-linear, this application can also increase the number of convolution layers or increase the length of the convolution kernel.

[0100] The embodiments of this application can extract different auxiliary features of the object attentively by setting adaptive weights for different users, making the determined features of the target user more targeted.

[0101] In some other embodiments of the present application, the process of extracting the features of the first meta-path set and the second meta-path set in the foregoing step S102 is introduced and explained.

[0102] First, a brief introduction to the composition of the first meta-path set and the second meta-path set is given.

[0103] In the embodiments of the present application, the first meta-path set and the second meta-path set respectively contain multiple meta-path subsets, and the multiple meta-path subsets are obtained by dividing the meta-paths in the corresponding meta-path set according to the meta-path types.

[0104] For example, taking the first meta-path set as an example, there are multiple meta-paths starting from the item to be scored in the heterogeneous information network. For example, there are multiple meta-paths of the UBUB type, multiple meta-paths of the UB type, and multiple meta-paths of the UBAB type. Then, in the present application, multiple meta-paths of the UB type can be used as a meta-path subset, multiple meta-paths of the UBUB type can be used as another meta-path subset, and multiple meta-paths of the UBAB type can be used as yet another meta-path subset. Here, U represents a user, B represents an item, and A represents an author. For example, UB represents User-Book (specifically, the target user - the book to be scored in the present application), UBAB represents User-Book-User-Book (specifically, the target user - the book to be scored - other users who have scored the book to be scored - other books scored by other users), and UBAB represents User-Book-Author-Book (specifically, the target user - the book to be scored - the author of the book to be scored - other books written by the author).

[0105] As can be seen from the above introduction, the information contained in the meta-path is relatively complex, and there are both differences and connections between different types of meta-paths. In the prior art, the method based on HIN is to take the average value of the features of all the meta-paths in the meta-path set as the basis for parameter update, which not only ignores the rich semantic information therein but also may introduce noise.

[0106] Therefore, the inventor of this case found through in-depth research that for the three meta-path subsets UB, UBUB, and UBAB, obviously, compared with UBAB, UB and UBUB have stronger relevance because they are composed of the same kind of relationship. Similarly, due to the same length, UBAB and UBUB are more similar than UB.

[0107] To capture this similarity, the inventors of this case thought that when organizing the set of meta-paths, adjacent sub-sets of similar meta-paths should be made adjacent, and then one-dimensional convolution (the one-dimensional convolution here is different from the one-dimensional convolution when extracting the features of the target user) should be performed on the sub-sets of meta-paths with stronger correlation (i.e., higher similarity) to extract the rich semantic information between the sub-sets of meta-paths with high similarity. After one round of one-dimensional convolution, one layer of prototypes of the similar meta-path sub-sets is extracted. Following a similar principle, one-dimensional convolution is sequentially performed on the sub-sets of meta-paths with weak correlation, and by extracting the prototypes of similar meta-paths layer by layer, the semantic information between the meta-paths is comprehensively considered.

[0108] Based on this, the process of extracting the features of any one of the first meta-path set and the second meta-path set in this embodiment may include: extracting the features of multiple meta-path sub-sets included in the meta-path set; sorting the features of the multiple meta-path sub-sets included in the meta-path set in ascending or descending order of similarity, and performing multi-level one-dimensional convolution processing on the sorted features of the multiple meta-path sub-sets. Among them, the convolution parameters of each level of one-dimensional convolution processing are different and the length of the convolution kernel is at least 2. In the first-level one-dimensional convolution processing, one-dimensional convolution processing is performed on the features of every two adjacent meta-path sub-sets among the sorted multiple meta-path sub-sets to obtain the fused features under the first-level one-dimensional convolution processing. In each subsequent level of one-dimensional convolution processing, one-dimensional convolution processing is performed on every two adjacent fused features among the fused features under the previous-level one-dimensional convolution processing until only one fused feature remains as the feature of the meta-path set.

[0109] Optionally, the process of "extracting the features of multiple meta-path sub-sets included in the meta-path set" may include: for each meta-path sub-set included in the meta-path set, extracting the features of each meta-path in the meta-path sub-set, calculating the average value of the features of each meta-path in the meta-path sub-set, and using the calculated average value as the feature of the meta-path sub-set to obtain the features of each meta-path sub-set included in the meta-path set.

[0110] See Figure 4 As shown, taking a meta-path set including four meta-path sub-sets UB, UBUB, UBAB, and UBAC as an example, the result after sorting by similarity is as Figure 4As shown in the figure, the process of performing multi-level one-dimensional convolution processing on the features of each meta-path subset in the meta-path set includes: First, perform one-dimensional convolution processing with a convolution kernel length of at least 2 on the features of UB and the features of UBUB, and perform one-dimensional convolution processing with a convolution kernel length of at least 2 on the features of UBUB and the features of UBAB, and perform one-dimensional convolution processing with a convolution kernel length of at least 2 on the features of UBAB and the features of UBAC. The above three one-dimensional convolution processes obtain three fused features under the first-level one-dimensional convolution processing; then, perform second-level one-dimensional convolution processing on the three fused features obtained under the first-level one-dimensional convolution processing according to Figure 4 to obtain two fused features under the second-level one-dimensional convolution processing; finally, perform third-level one-dimensional convolution processing on the two fused features obtained under the second-level one-dimensional convolution processing according to Figure 4 to obtain a fused feature under the third-level one-dimensional convolution processing. Since there is only one fused feature at this time, no fourth-level one-dimensional convolution processing is performed.

[0111] It should be noted that the convolution parameters of the above three-level one-dimensional convolution processing are all different.

[0112] The embodiments of the present application abstract similar meta-paths based on evidence, strengthen the meta-paths that appear multiple times, and these paths often represent the obvious tendencies of users, avoiding the phenomenon of feature weakening caused by blind summarization.

[0113] In some other embodiments of the present application, the cases where the scored items include one item and multiple items are described.

[0114] In the case where the scored item includes one item, the present application can directly add the scoring offset of the item to be scored relative to this item to the scoring value of the target user for this item, and use the sum value as the scoring prediction value of the target user for the item to be scored.

[0115] Considering that only one item is used as the basis for scoring prediction, in the case where the scoring offset deviates, the accuracy of the scoring prediction value will be reduced (of course, the accuracy is still higher than that of the prior art). For this reason, preferably, the scored items can include multiple items.

[0116] In the case where the scored items include multiple items, as shown in Figure 5 the scored items include is1, is2, and is3. The second meta-path set of the target user and is1 is P is1 , the second meta-path set of the target user and is2 is P is2 , the second meta-path set of the target user and is3 is P is3 . When it is necessary to estimate the score of the item to be scored iq, u, is1, P is1 , is2, P is2, is3, P is3 , iq and the first meta-path set P iq are input into FN, and FN extracts the features f of each piece of information respectively u , f i and f q .

[0117] The extracted features are sent to the offset module. In the offset module, first, feature splicing and feature mapping are performed according to step S102, and then score offset calculation is performed according to step S103

[0118] It should be noted that when the rated items include multiple items, the aforementioned second meta-path set includes the second meta-path sets corresponding to each item included in the rated items respectively

[0119] In this case, the process of the aforementioned step S102 "splice the features of the target user, the first meta-path set, the second meta-path set, the item to be rated, and the rated items to obtain the spliced features" may include: for each item included in the rated items, splice the features of the target user, the item to be rated, the first meta-path set, this item, and the second meta-path set corresponding to this item to obtain the spliced features corresponding to this item

[0120] That is, in this embodiment, for each item included in the rated items, splicing can be performed according to the aforementioned formula (1) (the features of the target user, the features of each item, the features of the second meta-path set corresponding to each item, the features of the item to be rated, and the features of the first meta-path set are spliced in one channel. Therefore, the number of items included in the rated items determines the number of channels for splicing in this application

[0121] Correspondingly, the process of the aforementioned "map the features of the first meta-path set, the second meta-path set, the item to be rated, and the rated items included in the spliced features to the same embedding space" may include: map the target features included in the spliced features corresponding to each item to the same embedding space, where the target features included in the spliced features corresponding to an item include the features of the first meta-path set, the features of the item to be rated, the features of the second meta-path set corresponding to this item, and the features of this item

[0122] Taking the item to be rated as iq and the rated items including is1 and is2 as an example, the schematic diagram of feature mapping is as Figure 6 shown, and the upper left of the arrow shows the features of the item to be rated iq and the corresponding first meta-path set P iq the features of the rated item is1 and the corresponding second meta-path set P is1 the features of the rated item is2 and the corresponding second meta-path set P is2Obviously, the iq is not the same as the score of any of the supporting samples (i.e., the scored items). In this application, the above feature can be mapped to the embedding space in the lower right corner via the feature extraction network FN.

[0123] See Figure 6 , in this application, the features of each item included in the scored items and the corresponding second meta-path set are mapped separately. For this reason, in step S103, with the assistance of the features of the target user, the offset calculation function R can calculate the score offset bias of each item included in the item to be scored relative to the scored items. For example, Figure 6 In the case where the scored items include two items, the score offset bias determined in this application includes 2.

[0124] It should be noted that the score offset calculated in this application is a signed quantity. A positive number represents that the target user's score value for the item to be scored is higher than that of the scored item, and a negative number represents lower than the scored item.

[0125] In this application, the score value of the target user for the scored items includes the score value of the target user for each item. Then, the process of step S104 "determine the score prediction value of the target user for the item to be scored according to the score value of the target user for the scored items and the score offset" may include: adding the score value of the target user for each item to the score value of the item to be scored relative to the corresponding item, specifically as shown in formula (4), to obtain the score prediction value based on each item as the prediction benchmark; calculating the average value of the score prediction values based on all items included in the scored items, and this average value is used as the score prediction value of the target user for the item to be scored.

[0126] modification(r u,is ,b u,is,iq )=r u,is +b u,is,iq Formula (4)

[0127] In the formula, modification(r u,is ,b u,is,iq ) represents the score prediction value based on each item as the prediction benchmark.

[0128] In this application, for each item included in the scored items, a score prediction value is calculated according to the above formula (4), and then the average value of the score prediction values corresponding to all items is calculated. This average value is the score prediction value of the target user for the item to be scored finally determined in this application.

[0129] In summary, this application combines the metric learning method with the HIN, gives the metric the practical meaning of a score offset, and attempts to dynamically calculate the metric under different scored items, which can indirectly predict the classification of samples that do not originally exist in the support set, improving the accuracy of the score prediction for items to be scored and thus enhancing the accuracy of recommendations.

[0130] In a possible implementation, this application can design the above-mentioned feature extraction module and offset module into an end-to-end score prediction model. The model training uses the MSE (Mean Squared Error) as the loss function. By giving the metric a more practical meaning, this way transforms the direct similarity metric into an indirect metric, endowing the model with stronger robustness; the flexible use of one-dimensional convolution allows this application to simply and elegantly fuse the rich information contained in the HIN while avoiding introducing too much noise.

[0131] To verify the prediction results of the model provided by this application, the inventors of this case have conducted sufficient practices on two datasets (the publicly available book review dataset and the movie review dataset).

[0132] The score value range of the publicly available book review dataset is from 1 to 5. In this dataset, considering the situation where users and items have only one feature and three meta-paths, a relatively simple task is set to mainly examine whether the model of this application can be used for cold start recommendation.

[0133] The movie review dataset is widely used as a stable benchmark dataset for recommendations, containing user ratings for movies with values from 1 to 5. In contrast, this dataset is more complex. This test considers more complex user and item features, sets four meta-paths, and the task is also more difficult, aiming to study how accurate the score prediction of the model is.

[0134] When testing the construction of the support set and the query set, consider users with scores more than 13 and less than 100 and the items they evaluated. Select 10 evaluations as the query set and another 5 evaluations as the support set to construct a classic 5-way few-shot learning task, which can simulate new users with only 3 to 5 scoring records as much as possible and give full play to the advantages of metric learning. When testing, select users and items that were included in the dataset relatively late as new users and new items, accounting for about 20%. The training tasks are all composed of existing users and items. Four test tasks, namely user cold-start (UC), item cold-start (IC), user and item cold-start (UIC), and non-cold-start (NC), are constructed according to the problem settings described above. In particular, if the score of a query sample does not exist in the support set in a task, this task is regarded as a difficult task because traditional metric learning methods cannot make correct predictions anyway.

[0135] The model implementation structure is as Figure 7 shown, and some parameters need to be set according to the specific business scenario. See Figure 7 shown, FN is a convolutional neural network composed of five convolutional modules and a first fully connected module.

[0136] Among them, the first convolutional module, the second convolutional module, and the third convolutional module are used to extract the features of the target user u. The third convolutional module is a basic one-dimensional convolutional module with a kernel size of 3, which is used to perform user feature fusion according to the aforementioned formula (3), and the convolutional output is one channel. On the basis of the third convolutional module, this application adds two convolutional modules (i.e., the first convolutional module and the second convolutional module) that do not change the number of features to extract the features of multiple information dimensions of the target user and enhance the non-linearity. Of course, adding two convolutional modules is only an example, and the number of convolutional modules in this application can be increased or decreased according to the actual situation.

[0137] In order to extract the features of the item to be scored, the scored items, the first meta-path set, and the second meta-path set, and ensure that these four types of features can be mapped to the same embedding space to calculate the score offset, a first fully connected module shared by the item and the meta-path is set. The fourth convolutional module is used to extract the features of the meta-path and obtain the features of each meta-path subset based on the extracted features of the meta-path. The fifth convolutional module is used to Figure 4 fuse the features of multiple meta-path subsets included in the meta-path set to obtain the features of the meta-path set.

[0138] It should be noted that the number of filters in the fourth convolution module and the fifth convolution module is determined by the dimensionality of the meta-path features; the convolution kernel size is set according to the number of initial meta-path prototypes. In order to output one prototype within two layers of convolution, the setting for the public book review dataset should be (2, 2), while for the movie review dataset with four initial prototypes, this setting is (2, 3).

[0139] The offset module R consists of a feature concatenation module and a second fully connected module. The feature concatenation module is used to perform feature concatenation according to formula (1), and the second fully connected module is used to map the target features included in the concatenated features to the same embedding space and calculate the score offset according to formula (2) after mapping.

[0140] The model provided by this application is compared with three existing technologies. The three existing technologies are traditional methods (FM and NeuMF), HIN-based methods (mp2vec and HERec), and meta-learning-based methods (MeLU and MetaHIN). In particular, MetaHIN is also an HIN-based method. For non-HIN methods, other types of nodes except items are degenerated into item features. For non-meta-learning methods, all training tasks and test tasks are expanded into training sets and test sets.

[0141] Two commonly used evaluation metrics are applied when evaluating the effects. One of them is the mean absolute error (MAE), which can measure the accuracy of score prediction, and the smaller the value, the more accurate the prediction. The other is the normalized discounted cumulative accuracy at rank K (nDCG@K), which represents the prediction performance of the top K rankings, and the larger the value, the better the performance. In this paper, the value of K is set to 3 and 5.

[0142] When facing the more complex and difficult situation of the movie review dataset, the performance of other baselines has declined to varying degrees, and the traditional methods have declined the most severely. However, thanks to the adaptive fusion of user features and meta-path prototype extraction in this application, the model provided by this application has achieved the best performance, and the improvement is more obvious in more difficult situations. Moreover, the model of this application pays more attention to the relationship between samples rather than the direct mapping between sample features and scores. Therefore, it has achieved a higher accuracy in ranking, which shows that the model of this application has application value in sequence recommendation that pays more attention to the fine ranking of descending scores. In addition, the efforts made by the model of this application to improve the situation of lack of information due to data scarcity are fruitful, and the improvement of information richness has significantly improved the training speed of the model of this application.

[0143] Experiments on two datasets show that the model of this application has achieved state-of-the-art performance in various situations, with significant improvements in complex and difficult situations, and is particularly suitable for cold-start recommendation in sequential recommendation scenarios.

[0144] The present application also provides a cold start recommendation device. Figure 8 , shows a schematic diagram of the structure of the cold start recommendation device provided in an embodiment of the present application, such as Figure 8 As shown, the cold start recommendation device may include: a basic data acquisition unit 801, a feature extraction and mapping unit 802, a score offset determination unit 803 and a score value prediction unit 804.

[0145] The basic data acquisition unit 801 is used to obtain a first meta-path set between a target user and an item to be rated, a second meta-path set between a target user and an item that has been rated, and a rating value of the target user for the rated item, wherein the first meta-path set and the second meta-path set are sets of meta-paths starting from the corresponding item in a heterogeneous information network, and the target user is a new user who has rated a small number of items and / or the item to be rated is a new item.

[0146] The feature extraction and mapping unit 802 is used to respectively extract features of the target user, the first meta-path set, the second meta-path set, the items to be rated, and the rated items, and map the extracted features of the first meta-path set, the second meta-path set, the items to be rated, and the rated items to the same embedding space.

[0147] The rating offset determination unit 803 is used to determine the rating offset of the to-be-rated item relative to the rated item based on the target user's features and the features mapped to the embedding space.

[0148] The rating value prediction unit 804 is used to determine the target user's rating prediction value for the to-be-rated item based on the target user's rating value and rating offset for the rated item, so as to determine whether to recommend the to-be-rated item to the target user based on the rating prediction value.

[0149] The working principle of the cold start recommendation device provided in the present application is the same as the working principle of the aforementioned cold start recommendation method. For details, please refer to the above introduction and will not be repeated here.

[0150] The embodiment of the present application also provides a cold start recommendation device. Optionally, Figure 9 The hardware structure diagram of the cold start recommended equipment is shown in Figure 9 , the hardware structure of the cold start recommendation device may include: at least one processor 901, at least one communication interface 902, at least one memory 903 and at least one communication bus 904;

[0151] In an embodiment of the present application, the number of the processor 901, the communication interface 902, the memory 903, and the communication bus 904 is at least one, and the processor 901, the communication interface 902, and the memory 903 complete communication with each other through the communication bus 904;

[0152] The processor 901 may be a central processing unit (CPU), or a specific integrated circuit (ASIC) (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention, etc.;

[0153] The memory 903 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;

[0154] Among them, the memory 903 stores a program, and the processor 901 can call the program stored in the memory 903. The program is used for:

[0155] Obtain a first meta-path set of a target user and an item to be scored, a second meta-path set of the target user and an item that has been scored, and the score value of the target user for the item that has been scored. Among them, the first meta-path set and the second meta-path set are sets of meta-paths starting from the corresponding item in the heterogeneous information network, and the target user is a new user who has scored a small number of items and / or the item to be scored is a new item;

[0156] Extract the features of the target user, the first meta-path set, the second meta-path set, the item to be scored, and the item that has been scored respectively, and map the extracted features of the first meta-path set, the second meta-path set, the item to be scored, and the item that has been scored to the same embedding space;

[0157] Based on the features of the target user and the features mapped to the embedding space, determine the score offset of the item to be scored relative to the item that has been scored;

[0158] According to the score value of the target user for the item that has been scored and the score offset, determine the score prediction value of the target user for the item to be scored, so as to determine whether to recommend the item to be scored to the target user based on the score prediction value.

[0159] Optionally, the refinement function and the extension function of the program can be referred to the above description.

[0160] The embodiment of the present application further provides a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the cold start recommendation method as described above is implemented.

[0161] Optionally, the refinement function and the extension function of the program can be referred to the above description.

[0162] Finally, it should be noted that, in this article, relational terms such as and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprises a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0163] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0164] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cold start recommendation method, characterized in that, Including: Obtain a first meta-path set of a target user and an item to be scored, a second meta-path set of the target user and scored items, and the scoring values of the scored items by the target user. Wherein, the first meta-path set and the second meta-path set are sets of meta-paths starting from the corresponding items in a heterogeneous information network, and the target user is a new user who has scored a small number of items and / or the item to be scored is a new item; Extract the features of the target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored items respectively, and map the features of the extracted first meta-path set, second meta-path set, item to be scored, and scored items to the same embedding space; Based on the features of the target user and the features mapped to the embedding space, determine the scoring offset of the item to be scored relative to the scored items; According to the scoring value of the scored items by the target user and the scoring offset, determine the scoring prediction value of the target user for the item to be scored, so as to determine whether to recommend the item to be scored to the target user based on the scoring prediction value; Wherein, the process of extracting the features of the target user includes: Extract the features of multiple information dimensions of the target user; Determine the adaptive weights of the features of the multiple information dimensions through a one-dimensional convolution algorithm; Perform weighted summation on the features of the multiple information dimensions based on the adaptive weights to obtain the features of the extracted target user; Wherein, the first meta-path set and the second meta-path set respectively include multiple meta-path subsets, and the multiple meta-path subsets are obtained by dividing the meta-paths in the corresponding meta-path set according to the meta-path types; The process of extracting the features of a meta-path set includes: Extract the features of multiple meta-path subsets included in this meta-path set; Sort the features of the multiple meta-path subsets included in this meta-path set in descending or ascending order of similarity, and perform multi-level one-dimensional convolution processing on the sorted features of the multiple meta-path subsets. Wherein, the convolution parameters of each level of one-dimensional convolution processing are different and the convolution kernel length is at least 2. In the first-level one-dimensional convolution processing, perform one-dimensional convolution processing on the features of every two adjacent meta-path subsets among the sorted multiple meta-path subsets to obtain the fused features under the first-level one-dimensional convolution processing. In each subsequent level of one-dimensional convolution processing, perform one-dimensional convolution processing on every two adjacent fused features in the fused features under the previous level of one-dimensional convolution processing until only one fused feature remains, which is used as the feature of this meta-path set.

2. The cold start recommendation method according to claim 1, wherein The extraction of the features of multiple meta-path subsets included in this meta-path set includes: For each meta-path subset included in this meta-path set, extract the features of each meta-path in this meta-path subset, calculate the average value of the features of each meta-path in this meta-path subset, and use the calculated average value as the feature of this meta-path subset, so as to obtain the features of each meta-path subset included in this meta-path set.

3. The cold start recommendation method according to claim 1, characterized in that Mapping the extracted first meta-path set, second meta-path set, features of the item to be scored, and features of the scored items into the same embedding space includes: Concatenating the features of the extracted target user, first meta-path set, second meta-path set, item to be scored, and scored items to obtain concatenated features; Mapping the features of the first meta-path set, second meta-path set, item to be scored, and scored items included in the concatenated features into the same embedding space.

4. The cold start recommendation method according to claim 3, wherein The scored items include multiple items, and the second meta-path set includes second meta-path sets corresponding to each item included in the scored items; The concatenating the features of the extracted target user, first meta-path set, second meta-path set, item to be scored, and scored items to obtain concatenated features includes: For each item included in the scored items, concatenating the features of the extracted target user, item to be scored, first meta-path set, this item, and the second meta-path set corresponding to this item to obtain the concatenated features corresponding to this item; The mapping the features of the first meta-path set, second meta-path set, item to be scored, and scored items included in the concatenated features into the same embedding space includes: Mapping the target features included in the concatenated features corresponding to each item into the same embedding space, where the target features included in the concatenated features corresponding to an item include the features of the first meta-path set, the features of the item to be scored, the features of the second meta-path set corresponding to this item, and the features of this item.

5. The cold start recommendation method according to claim 4, characterized in that The scoring offset of the item to be scored relative to the scored items includes the scoring offset of each item included in the item to be scored relative to the scored items, and the scoring value of the target user for the scored items includes the scoring value of the target user for each item; Determining the scoring prediction value of the target user for the item to be scored according to the scoring value of the target user for the scored items and the scoring offset includes: Adding the scoring value of the target user for each item to the scoring value of the item to be scored relative to the corresponding item to obtain a scoring prediction value with each item as the prediction benchmark; Calculating the average value of the scoring prediction values with all items included in the scored items as the prediction benchmark, and taking this average value as the scoring prediction value of the target user for the item to be scored.

6. A cold start recommendation device for implementing the cold start recommendation method according to any one of claims 1 to 5, characterized in that Includes: A basic data acquisition unit for acquiring the first meta-path set of the target user and the item to be scored, the second meta-path set of the target user and the scored items, and the scoring value of the target user for the scored items, where the first meta-path set and the second meta-path set are sets of meta-paths starting from the corresponding items in the heterogeneous information network, and the target user is a new user who has scored a small number of items and / or the item to be scored is a new item; A feature extraction and mapping unit, configured to extract the features of the target user, the first meta-path set, the second meta-path set, the item to be scored, and the scored items respectively, and map the features of the first meta-path set, the second meta-path set, the item to be scored, and the scored items to the same embedding space; A scoring offset determination unit, configured to determine the scoring offset of the item to be scored relative to the scored items based on the features of the target user and the features mapped to the embedding space; A scoring value prediction unit, configured to determine the predicted scoring value of the target user for the item to be scored according to the scoring value of the target user for the scored items and the scoring offset, so as to determine whether to recommend the item to be scored to the target user based on the predicted scoring value.

7. A cold start recommendation device, characterized in that, Comprising a memory and a processor; The memory is configured to store programs; The processor is configured to execute the programs to implement the steps of the cold start recommendation method according to any one of claims 1 to 5.

8. A readable storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by the processor, the steps of the cold start recommendation method according to any one of claims 1 to 5 are implemented.

Images