List recommendation method, system, device and storage medium based on zero-shot learning

Through the zero-learning list recommendation method, the pre-trained model is used to calculate the similarity of user needs, and the problem that the existing recommendation system cannot meet personalized needs is solved, especially in cold start scenarios, efficient and accurate recommendations are achieved List generation.

CN113139127BActive Publication Date: 2024-11-19PEKING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110481796.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-04-30
Publication Date
2024-11-19
Estimated Expiration
2041-04-30

AI Technical Summary

Technical Problem

The existing recommendation system is difficult to meet the personalized needs of users, especially in cold start scenarios, which cannot effectively create a recommendation list that meets user preferences, and requires multiple additional information acquisitions, which consumes time and human resources.

Method used

Using the zero-learning-based list recommendation method, the basic information attributes of user needs are obtained through the pre-trained zero-learning model, the similarity between the predicted sample features and the real sample features is calculated, and the most similar sample features are recommended to build a list. , solves the cold start problem and improves the accuracy and interactivity of recommendations.

Benefits of technology

It realizes the creation of a recommendation list that meets users' personalized needs from scratch, improves the accuracy and interactivity of recommendations, reduces the computing volume and memory usage, and simplifies the acquisition process of user needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113139127B_ABST
    Figure CN113139127B_ABST
Patent Text Reader

Abstract

The present invention provides a method, system, device and storage medium for list recommendation based on zero-shot learning, the method comprising: obtaining basic information attributes of a list to be created; inputting the basic information attributes into a pre-trained zero-shot learning model to obtain predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is obtained based on the mapping training between the training list and the training samples; performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the successfully matched real sample features to the list to be created to construct the list. The present invention obtains predicted sample features that have a mapping relationship with the basic information attributes of the list to be created through a zero-shot learning model, and recommends samples to the list to be created based on the predicted sample features, thereby performing personalized list recommendations based on user needs or preferences, improving the accuracy of the generated recommended list, and further improving the interactivity and interaction accuracy with the user.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a list recommendation method, system, device and storage medium based on zero-shot learning. Background Art

[0002] In recent years, the rapid development of the Internet has become more and more inseparable from human life. Human behavior on the Internet, such as purchasing, browsing, posting blogs, filling in information, etc., will be stored, forming a steady stream of data, which has led to today's big data era. If these data can be used well, it can bring many conveniences to human life in the future. There are many technologies to achieve "data creates a better life", and the recommendation system is one of them. It predicts the user's possible future interests based on the user's current preferences, so as to show the goods or projects that the user may like. For example, based on a certain product purchased by the user, the system continues to push related products or accessories to the user, which can not only increase the sales of the seller, but also provide convenience for the user to choose products; for example, based on the user's preferences, recommend works of similar styles to achieve the purpose of user retention.

[0003] Recommendation systems are usually divided into four types: collaborative filtering-based recommendations, content-based recommendations, knowledge-based recommendations, and hybrid-based recommendations. Among them: collaborative filtering recommendation algorithms usually include user- or item-based collaboration, matrix decomposition, latent factor models, etc.; content and knowledge recommendations usually introduce additional information about items or users and calculate similarities to make recommendations; hybrid-based recommendations combine the above solutions. With the increasing maturity of deep learning technology, neural network matrix decomposition, deep neural networks, generative adversarial networks (GANs), reinforcement learning, etc. have all been applied to recommendation systems, which usually have better performance than traditional methods, can efficiently process a large number of candidate products, and achieve more accurate recommendations.

[0004] Since most existing recommendation systems are based on historical data to recommend single or similar preferences and items, in order to achieve diversified recommendation services, a list recommendation method is adopted, with a collection of items as the recommended objects, and recommendations are made to users in the form of a list. However, most existing list recommendations are sorted according to the user's preference for the recommended items, and a certain number of recommended items are selected based on the sorting to form a list. This list recommendation method ignores the directionality of user demand for items, such as the actual user's taste preferences in the curry beef recipe shopping list, the actual user's style preferences in the spring travel outfit list, the actual user's interest preferences in the list of literary classics suitable for college students, and the user's preferences for music style, singers, and languages ​​in the song list of user-specified style genres, so it cannot meet the personalized needs of users.

[0005] In addition, for users who use the list recommendation system for the first time, projects that join the recommendation system for the first time, or recommendation systems that are put into use for the first time, it is difficult for the recommendation system to deal with the cold start problem well. Cold start is usually divided into three scenarios: new user cold start, new project cold start, and new system cold start. During the cold start, whether it is a new user, a new project, or a new system, since they are all used for the first time, there is no historical data related to them in the recommendation system, such as the interest preferences of new users, the rating data or usage data of the new system, and the historical behavior data of the new system. Therefore, in order to enable the recommendation system to predict preferences and make recommendations, it is necessary to obtain additional information in advance, such as using questionnaires, performance data from other platforms, or social information data to obtain additional information about users, and using questionnaires to investigate the usage intention preferences of the project to obtain additional information about the items, so as to make predictions and recommendations for the first use based on the pre-acquired additional information. However, this method of predicting and recommending based on pre-acquired additional information requires multiple acquisitions of additional information when facing different cold start objects, which is time-consuming and wastes human resources. Summary of the invention

[0006] The present invention provides a list recommendation method, system, device and storage medium based on zero-shot learning, which is used to solve the defect in the prior art that a recommendation list cannot be created from scratch based on the personalized needs of users. It can select matching items according to user needs to form a recommendation list sorted according to user needs, realize user personalized list recommendation, and facilitate the creation of a recommendation list that meets the personalized needs of users from scratch.

[0007] The present invention provides a list recommendation method based on zero-shot learning, comprising: obtaining basic information attributes of a list to be created; inputting the basic information attributes into a pre-trained zero-shot learning model to obtain predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between a training list and training samples; performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the successfully matched real sample features to construct a list in the list to be created.

[0008] According to the present invention, a list recommendation method based on zero-shot learning is provided, in which the predicted sample features are matched with the real sample features for similarity, and samples corresponding to the successfully matched real sample features are recommended to the list to be created to construct the list, including: selecting the real sample features that are closest to the predicted sample features; and using the selected closest real sample features as recommended samples, and recommending them to the list to be created for list construction.

[0009] According to a list recommendation method based on zero-shot learning provided by the present invention, when the real sample feature that performs similarity matching with the predicted sample feature belongs to the same real list, finding the real sample feature that is closest to the predicted sample feature includes: calculating the Euclidean distance or cosine similarity between the predicted sample feature and the real sample feature; sorting the Euclidean distance according to the distance size, or sorting the cosine similarity according to the similarity size; according to the sorting, selecting the real sample feature with the smallest Euclidean distance or the largest cosine similarity as the closest real sample feature.

[0010] According to a list recommendation method based on zero-shot learning provided by the present invention, when the real sample features that are matched with the predicted sample features for similarity belong to different real lists, finding the real sample features that are closest to the predicted sample features includes: calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; sorting the Euclidean distances according to the distance size, or sorting the cosine similarities according to the similarity size; according to the sorting, counting the number of occurrences of the same real sample feature among multiple real sample features with the smallest Euclidean distance or the largest cosine similarity, and taking the real sample feature with the largest number of occurrences as the real sample feature that is closest to the predicted sample feature.

[0011] According to a list recommendation method based on zero-shot learning provided by the present invention, the obtaining of basic information attributes of the list to be created includes: determining the basic information of the list to be created according to user needs or preferences; performing feature extraction based on the basic information, and converting the extracted features into vectors; and splicing the converted vectors to form basic information attributes.

[0012] According to a list recommendation method based on zero-shot learning provided by the present invention, training the zero-shot learning model includes: extracting training list attributes and training sample features from a pre-constructed training database; performing mapping training on the zero-shot learning model using the training list attributes and the training sample features; calculating a loss function based on the mapping between the training list attributes and the training sample features; if the loss function converges, stopping the training; otherwise, continuing to train the loss function.

[0013] According to a list recommendation method based on zero-shot learning provided by the present invention, the loss function is calculated based on the mapping between the training list attributes and the training sample features, including: calculating the distance between the training list attributes after being mapped to the training sample features and the real sample features, as the first distance loss; calculating the distance between the training sample features after being mapped to the training list attributes and the real list attributes, as the second distance loss; calculating the distance between the mapping from the training list attributes to the training sample features and the mapping from the training sample features to the training list attributes, recorded as mapping loss; calculating the regularization loss in the mapping process; and summing the first distance loss, the second distance loss, the mapping loss and the regularization loss, recorded as the loss function.

[0014] The present invention also provides a list recommendation system based on zero-shot learning, comprising: a data acquisition module, which acquires basic information attributes of a list to be created; a prediction module, which inputs the basic information attributes into a pre-trained zero-shot learning model to obtain predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between a training list and training samples; a matching module, which performs similarity matching between the predicted sample features and the real sample features, and recommends samples corresponding to the successfully matched real sample features to the list to be created to construct the list.

[0015] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned methods for recommending lists based on zero-shot learning are implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described methods for recommending lists based on zero-shot learning.

[0017] The zero-shot learning-based list recommendation method, system, device and storage medium provided by the present invention obtain predicted sample features that have a mapping relationship with the basic information attributes of the list to be created through a zero-shot learning model, so as to preliminarily obtain the user's needs or preferences, and then calculate the similarity between the predicted sample features and the real sample features, and recommend the samples corresponding to the predicted sample features with the highest similarity to the list to be created, thereby performing personalized list recommendations based on user needs or preferences, improving the accuracy of the generated recommendation list, and further improving the interactivity and interaction accuracy with the user; in addition, based on the basic information attributes provided by the user, it is convenient to create a recommendation list from scratch, so as to further generate the recommendation list in real time and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is one of the flow charts of the list recommendation method based on zero-shot learning provided by the present invention;

[0020] Figure 2 It is a schematic diagram of a process of forming basic information attributes of a list recommendation method based on zero-shot learning provided by the present invention;

[0021] Figure 3 It is a flowchart of training the zero-shot learning model of the list recommendation method based on zero-shot learning provided by the present invention;

[0022] Figure 4 It is a schematic diagram of a flow chart of forming training sample features in a list recommendation method based on zero-shot learning provided by the present invention;

[0023] Figure 5 This is the second flow chart of the list recommendation method based on zero-shot learning provided by the present invention;

[0024] Figure 6 It is a schematic diagram of the training loss function and test accuracy of the list recommendation method model based on zero-shot learning provided by the present invention;

[0025] Figure 7 It is a schematic diagram comparing the test accuracy of the DEM model and the DeepSAE model provided by the present invention;

[0026] Figure 8 It is a structural diagram of a list recommendation system based on zero-shot learning provided by the present invention;

[0027] Fig. 9 is a schematic diagram of the structure of an electronic device provided by the present invention;

[0028] Reference numerals:

[0029] 1: data acquisition module; 2: zero-time learning module; 3: matching module; 161: processor; 162: communication interface; 163: memory; 164: communication bus. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Figure 1 A flowchart of a list recommendation method based on zero-shot learning of the present invention is shown, and the method comprises:

[0032] S01, obtaining basic information attributes of the list to be created;

[0033] S02, inputting the basic information attributes into a pre-trained zero-shot learning model to obtain the predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between the training list and the training samples;

[0034] S03, performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the real sample features that have been successfully matched to the list to be created to build a list.

[0035] It should be noted that the S0N in this specification does not represent the order of the list recommendation method based on zero-shot learning. Figure 2-Figure 7 The list recommendation method based on zero-shot learning of the present invention is described in detail.

[0036] Step S01, obtaining basic information attributes of the list to be created.

[0037] In this embodiment, obtaining the basic information attributes of the list to be created includes: determining the basic information of the list to be created according to the user's needs or preferences; extracting attributes based on the basic information and converting the extracted attributes into vectors; and concatenating the converted vectors to form the basic information attributes. Figure 2 When concatenating vectors, vectors with the same label can be summed, and vectors with different labels can be multiplied by preset coefficients for concatenation. The extracted attributes can be converted into vectors using a Chinese corpus. The vector dimension can be 200 dimensions. The specific dimension and preset coefficient can be set according to actual needs, which will not be described here.

[0038] In addition, the basic information of the list to be created includes: a label of the list to be created, a name of the list to be created, and a language description for introducing the list to be created. The list to be created in this embodiment can be determined according to user needs or preferences, so as to quickly understand the user's needs or preferences, and then to recommend a list suitable for the user's needs or preferences, thereby realizing user personalized list recommendation and avoiding the situation where the list is inaccurate due to lack of user needs or preference information when creating a list from scratch.

[0039] Step S02, inputting the basic information attributes into a pre-trained zero-shot learning model to obtain the predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between the training list and the training samples.

[0040] In this embodiment, the predicted sample features are obtained based on the zero-shot learning model and are mapped to the basic information attributes of the list to be created. The predicted sample features can be labels. For example, when the basic information attributes are at least one of the labels such as playlist name, playlist label, playlist description, and playlist creator, the predicted sample features can be the labels of the songs, such as the song name, singer, release time, song features, related works or characters, music style / important elements, and lyrics keywords.

[0041] Before inputting the basic information attributes into the pre-trained zero-shot learning model, it is necessary to train the zero-shot learning model. The training of the zero-shot learning model includes: extracting the training list attributes and training sample features from the pre-built training database; using the training list attributes and training sample features to map the zero-shot learning model and output the predicted mapping label; calculating the loss function based on the mapping between the training list attributes and the training sample features; if the loss function converges, stop training; otherwise, continue training the loss function. It should be noted that during the training process, the adam optimization algorithm can be used for gradient descent, and the corresponding step size can be set to 0.00005. The optimization algorithm can be selected according to actual needs, and this application does not make further restrictions on this. In this embodiment, the deep semantic autoencoder DeepSAE model is used to map the training list attributes and training sample features.

[0042] Specifically, the loss function is calculated based on the mapping between the training list attributes and the training sample features, including: calculating the distance between the training list attributes mapped to the training sample features and the real sample features as the first distance loss L α→x ; Calculate the distance between the training sample features mapped to the training list attributes and the real list attributes as the second distance loss L x→α ; Calculate the distance between the mapping from the training list attributes to the training sample features and the mapping from the training sample features to the training list attributes, recorded as the mapping loss Lw ; Calculate the regularization loss L during the mapping process R ; and based on the sum of the first distance loss, the second distance loss, the mapping loss and the regularization loss, it is recorded as the loss function L, refer to Figure 3 .

[0043] For example, if for multiple training list attributes and multiple training sample features, if the training list attributes are mapped to the training sample features to form mappings W1 and W2; the training sample features are mapped to the training list attributes to form mappings W3 and W4; then the first distance loss L α→x , the second distance loss L x→α , mapping loss L w And the loss function L are expressed as:

[0044]

[0045] in, Represented as the i-th training sample feature of the training, Represents the training list attribute of the training list to which the feature of the i-th training sample belongs; f1 and f2 are both activation functions. It should be noted that different activation functions can be selected according to the combination of different data sets. For example, for the AwA data set, tanh is used as the activation function f1 and f2; for the bird image CUB data set, ReLU is used as the activation function f1 and tanh is used as the activation function f2, which can obtain better results; for the scene image SUN data set and aPY data set, ReLU is used as the activation function f1 and f2; for the music data set, ReLU is used as the activation function f1 and tanh is used as the activation function f2.

[0046] It should be noted that before extracting the training list attributes and training sample features from the pre-constructed training database, the training database is constructed. Constructing the training database includes: obtaining real samples and real lists that have a mapping relationship with the real samples as training samples and training lists; converting the training lists and training samples into training list attributes and training sample features through word2vec embedding; and constructing the training database based on the training list attributes and training sample features. The method of forming the training sample features and the method of forming the training list attributes can refer to the above-mentioned method of forming the basic information attributes, which will not be repeated here.

[0047] For example, the list includes a first language label x1, a second list label x2, a third list label x3, a first list description y1, and a second list description y2, then the training list attribute formed = {μ1·x1, μ2·x2, μ3·x3, μ4·(y1+y2)}, where μ1, μ2, μ3, and μ4 are preset coefficients, and x1, x2, x3, y1, and y2 are all represented as vectors.

[0048] The song contains singer label X1, chorus label X2, singer gender label X3, first style element label X3, second style element label X4, song feature label X5, year label X6, related film, television or character label X7, lyrics keyword label X8, lyrics hidden meaning label X9, etc., then the training sample features formed = {τ1·(X1+X2+X3), τ2·x1, τ3·x2+τ4·x3+τ5·X3+τ6·X4+τ7·X5, τ8·X6, τ9·X7, τ 10 ·(X8+ X9)}, where, τ1, τ2,···, τ 10 are preset coefficients. X1, X2, X3, X3, X4, X5, X6, X7, X8, and X9 are all represented as vectors. Figure 4 .

[0049] Step S03, performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the real sample features that have been successfully matched to the list to be created to construct a list.

[0050] In this embodiment, the predicted sample features are matched with the real sample features for similarity, and the samples corresponding to the successfully matched real sample features are recommended to the list to be created to build the list, including: selecting the real sample features that are closest to the predicted sample features; and recommending the selected closest real sample features as recommended samples to the list to be created for list construction. By finding the closest real sample features, the real samples that are most similar to the predicted samples are obtained as recommended samples to build the recommended list, thereby reducing the amount of calculation, improving the running speed, and reducing the memory usage.

[0051] In a possible implementation, when the real sample features that are matched with the predicted sample features for similarity belong to the same real list, the real sample features that are closest to the predicted sample features are found, including: calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; sorting the Euclidean distances according to the distance size, or sorting the cosine similarities according to the similarity size; based on the sorting, selecting the real sample feature with the smallest Euclidean distance or the largest cosine similarity as the real sample feature that is closest to the predicted sample feature. It should be noted that the number of real sample features selected as the closest neighbors can be one or more, and the specific number can be determined based on the Euclidean distance calculation result or the cosine similarity calculation result.

[0052] In another possible implementation, when the real sample features that are matched with the predicted sample features for similarity belong to different real lists, since the categories of the multiple lists to be created are different, when selecting the real sample features that are closest to the predicted sample features, the K nearest neighbor algorithm K-NN can be used for selection, specifically including: calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; sorting the Euclidean distance according to the distance size, or sorting the cosine similarity according to the similarity size; based on the sorting, counting the number of occurrences of the same real sample feature among multiple real sample features with the smallest Euclidean distance or the largest cosine similarity, and taking the real sample feature with the largest number of occurrences as the real sample feature closest to the predicted sample feature.

[0053] The following uses the playlist as a list and the songs as examples to illustrate the DeepSAE model of the deep semantic autoencoder. Figure 5 .

[0054] First, the zero-shot learning model is trained. In the process of training the zero-shot learning model, the training set and the test set are first divided based on the training list and training examples. For example, 21 playlists are selected, of which 14 playlists are used as training playlists, each of which contains 50 songs, and 7 playlists are used as test playlists, each of which contains 30 songs, for a total of 910 songs, covering multiple languages, as well as pop, rock, folk, rap, ancient style and other styles.

[0055] It should be noted that since the basic information of some user-defined playlists is not particularly accurate in the process of forming basic information attributes, in order to improve the accuracy of feature extraction, the labels of the selected 21 playlists were fine-tuned to accurately extract the playlist labels.

[0056] Secondly, extract the text features of the training songs and convert them into vectors as training song features, where the text features include label information such as singer, genre, release time, lyrics keywords, etc. Then extract the training playlist features of the playlist where the training songs are located. The training playlist features are user-set labels including language, genre, scene, emotion, and the user's linguistic description of the playlist such as the playlist title and playlist introduction. Map the training song features to the training playlist features, and determine whether to end the training based on whether the loss function converges. Reference Figure 6, according to the comparison between the recommended song list and the test song list, it is shown that when the loss function converges, the test accuracy is good. It can be seen that in this embodiment, whether the model training is completed can be judged based on whether the loss function converges. In addition, the deep embedding model DEM model can be trained based on the training songs and the test songs, and its training results can be compared with the training results of the DeepSAE model. It can be seen that the test accuracy of the DeepSAE model is better. Figure 7 ,in Figure 7 The number after "Test Accuracy" is the number of playlists tested.

[0057] After the training is completed, the basic information of the test playlist, such as the playlist label, playlist name, and language description used to introduce the playlist, is obtained and converted into basic information attributes in vector form. The basic information attributes are input into the trained zero-shot learning model to obtain the predicted song features, and the similarity between the predicted song features and the real song features is calculated to output the song most similar to the predicted song features, and the song is recommended to the playlist to be created to form a recommended playlist.

[0058] In summary, the present invention obtains predicted sample features that have a mapping relationship with the basic information attributes of the list to be created through a zero-shot learning model, so as to preliminarily obtain the user's needs or preferences, and then calculates the similarity between the predicted sample features and the real sample features, and recommends the samples corresponding to the predicted sample features with the highest similarity to the list to be created, thereby performing personalized list recommendations based on user needs or preferences, improving the accuracy of the generated recommendation list, and further improving the interactivity and interaction accuracy with the user; in addition, based on the basic information attributes provided by the user, it is convenient to create a recommendation list from scratch, so as to further generate the recommendation list in real time and accurately.

[0059] The following is a description of a list recommendation system based on zero-time learning provided by the present invention. The list recommendation system based on zero-time learning described below and the list recommendation method based on zero-time learning described above can be referred to each other.

[0060] refer to Figure 8 , Figure 8 A list recommendation system based on zero-shot learning is shown, the system comprising:

[0061] Data acquisition module 1, obtains basic information attributes of the list to be created;

[0062] Zero-time learning module 2, inputs the basic information attributes acquired by the data acquisition module into a pre-trained zero-time learning module to obtain the predicted sample features output by the zero-time learning module; wherein the zero-time learning module is trained based on the mapping between the training list and the training samples;

[0063] The matching module 3 performs similarity matching between the predicted sample features output by the zero-time learning module and the real sample features, and recommends samples corresponding to the successfully matched real sample features to the list to be created to construct a list.

[0064] In this embodiment, the data acquisition module 1 includes a data acquisition unit, a data conversion unit and a vector splicing unit, wherein the data acquisition unit determines the basic information of the list to be created according to user needs or preferences; the data conversion unit extracts features of the basic information and converts the extracted features into vectors; the vector splicing unit splices the converted vectors to form basic information attributes.

[0065] The zero-shot learning module 2 includes a feature extraction unit, a mapping unit and a loss function calculation unit, wherein: the feature extraction unit extracts training list attributes and training sample features from a pre-constructed training database; the mapping unit performs mapping training on the training list attributes and the training sample features; the loss function calculation unit calculates the loss function based on the mapping between the training list attributes and the training sample features; if the loss function converges, the training is stopped; otherwise, the loss function continues to be trained.

[0066] The matching module 3 includes a matching unit and a recommendation unit. The matching unit performs similarity matching between the predicted sample features output by the zero-time learning module and the real sample features. The recommendation unit recommends the samples corresponding to the successfully matched real sample features to the list to be created to construct the list.

[0067] Fig. 9 An example of a physical structure diagram of an electronic device is shown in FIG. Fig. 9 As shown, the electronic device may include: a processor 171, a communication interface 172, a memory 173 and a communication bus 174, wherein the processor 171, the communication interface 172 and the memory 173 communicate with each other through the communication bus 174. The processor 171 may call the logic instructions in the memory 173 to execute a list recommendation method based on zero-time learning, the method comprising: obtaining basic information attributes of a list to be created; inputting the basic information attributes into a pre-trained zero-time learning model to obtain predicted sample features output by the zero-time learning model; wherein the zero-time learning model is obtained by training based on the mapping between the training list and the training samples; performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the successfully matched real sample features to the list to be created to construct the list.

[0068] In addition, the logic instructions in the above-mentioned memory 173 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0069] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the list recommendation method based on zero-shot learning provided by the above methods, and the method includes: obtaining basic information attributes of the list to be created; inputting the basic information attributes into a pre-trained zero-shot learning model to obtain predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between the training list and the training samples; performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the successfully matched real sample features to the list to be created to construct the list.

[0070] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the above-mentioned zero-shot learning-based list recommendation methods, the methods comprising: obtaining basic information attributes of the list to be created; inputting the basic information attributes into a pre-trained zero-shot learning model to obtain predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between the training list and the training samples; performing similarity matching between the predicted sample features and the real sample features, and recommending samples corresponding to the successfully matched real sample features to the list to be created to construct the list.

[0071] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0072] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A list recommendation method based on zero-shot learning, characterized in that: include: Get the basic information attributes of the list to be created; Input the basic information attributes into a pre-trained zero-shot learning model to obtain the predicted sample features output by the zero-shot learning model; wherein the zero-shot learning model is trained based on the mapping between the training list and the training samples, and the predicted sample features are obtained based on the zero-shot learning model and have a mapping relationship with the basic information attributes of the list to be created; The predicted sample features are matched with the real sample features for similarity, and the samples corresponding to the real sample features that are successfully matched are recommended to the list to be created to build a list; The basic information attributes of the list to be created are obtained, including: Determine basic information of the list to be created according to user needs or preferences, wherein the basic information of the list to be created includes a label of the list to be created, a name of the list to be created, and a language description for introducing the list to be created; Perform feature extraction based on the basic information and convert the extracted features into vectors; The converted vectors are concatenated to form basic information attributes; The similarity matching of the predicted sample features with the real sample features and recommending samples corresponding to the successfully matched real sample features to the to-be-created list to build a list includes: Selecting the real sample feature that is closest to the predicted sample feature; The selected nearest real sample features are used as recommended samples and recommended to the list to be created for list construction; When the real sample features that are matched with the predicted sample features by similarity belong to the same real list, finding the real sample features that are closest to the predicted sample features includes: Calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; Sorting the Euclidean distances according to the distances, or sorting the cosine similarities according to the similarities; According to the sorting, selecting the real sample feature with the smallest Euclidean distance or the largest cosine similarity as the real sample feature closest to the predicted sample feature; When the real sample features that are matched with the predicted sample features by similarity belong to different real lists, finding the real sample features that are closest to the predicted sample features includes: Calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; Sorting the Euclidean distances according to the distances, or sorting the cosine similarities according to the similarities; According to the sorting, the number of occurrences of the same real sample feature among multiple real sample features with the smallest Euclidean distance or the largest cosine similarity is counted, and the real sample feature with the largest number of occurrences is used as the real sample feature closest to the predicted sample feature.

2. The list recommendation method based on zero-shot learning according to claim 1, characterized in that: Training the zero-shot learning model includes: Extracting training list attributes and training sample features from a pre-built training database; Performing mapping training on the zero-shot learning model using the training list attributes and the training sample features, and outputting a predicted mapping label; Calculate the loss function based on the mapping between the training list attributes and the training sample features; If the loss function converges, the training is stopped; otherwise, the training of the loss function continues.

3. The list recommendation method based on zero-shot learning according to claim 2, characterized in that: The loss function is calculated based on the mapping between the training list attributes and the training sample features, including: Calculating the distance between the training list attribute mapped to the training sample feature and the real sample feature as the first distance loss; Calculating the distance between the training sample feature mapped to the training list attribute and the real list attribute feature as a second distance loss; Calculate the distance between the mapping from the training list attribute to the training sample feature and the mapping from the training sample feature to the training list attribute, recorded as mapping loss; Calculate the regularization loss during the mapping process; And based on the sum of the first distance loss, the second distance loss, the mapping loss and the regularization loss, it is recorded as a loss function.

4. A list recommendation system based on zero-shot learning, characterized in that: include: Data acquisition module, to obtain the basic information attributes of the list to be created; A prediction module, inputting the basic information attributes into a pre-trained zero-time learning model to obtain prediction sample features output by the zero-time learning model; wherein the zero-time learning model is trained based on the mapping between the training list and the training samples, and the prediction sample features are obtained based on the zero-time learning model and have a mapping relationship with the basic information attributes of the list to be created; A matching module performs similarity matching between the predicted sample features and the real sample features, and recommends samples corresponding to the real sample features that have been successfully matched to the list to be created to construct a list; The data acquisition module comprises: A data acquisition unit, which determines basic information of the list to be created according to user needs or preferences, wherein the basic information of the list to be created includes a label of the list to be created, a name of the list to be created, and a language description for introducing the list to be created; A data conversion unit extracts features from the basic information and converts the extracted features into vectors; A vector concatenation unit concatenates the converted vectors to form basic information attributes; The matching module is used for: Selecting the real sample feature that is closest to the predicted sample feature; The selected nearest real sample features are used as recommended samples and recommended to the list to be created for list construction; When the real sample features that are similarity matched with the predicted sample features belong to the same real list, the matching module is further used to: Calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; Sorting the Euclidean distances according to the distances, or sorting the cosine similarities according to the similarities; According to the sorting, selecting the real sample feature with the smallest Euclidean distance or the largest cosine similarity as the real sample feature closest to the predicted sample feature; When the real sample features that are matched with the predicted sample features by similarity belong to different real lists, the matching module is further used to: Calculating the Euclidean distance or cosine similarity between the predicted sample features and the real sample features; Sorting the Euclidean distances according to the distances, or sorting the cosine similarities according to the similarities; According to the sorting, the number of occurrences of the same real sample feature among multiple real sample features with the smallest Euclidean distance or the largest cosine similarity is counted, and the real sample feature with the largest number of occurrences is used as the real sample feature closest to the predicted sample feature.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the list recommendation method based on zero-shot learning as claimed in any one of claims 1 to 3 are implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the list recommendation method based on zero-shot learning are implemented as claimed in any one of claims 1 to 3.