Cold Start Recommendation Method, Device, Electronic Device and Storage Medium

Abstract

An embodiment of the present invention provides a cold start recommendation method, and the method includes: obtaining text vectors of each item under a first item category label, classifying the text vectors through a pre-trained classification neural network to obtain first item category information of each item; obtaining a first mapping relationship between a preset user group and a second item category label, and constructing a second mapping relationship from the second item category label to the user group according to the first mapping relationship; mapping the first item category information of each item into the user group according to the second mapping relationship, and counting the occurrence times of the first item category information of each item corresponding to each user group; based on the occurrence times, tagging the corresponding user group with the first item category label, and performing item recommendation for the corresponding user group according to the first item category label and the second item category label. The present invention can achieve the cold start of the recommendation system and improve the cold start recommendation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of big data, and in particular, to a cold start recommendation method, apparatus, electronic device, and storage medium. Background Art

[0002] With the development of big data technology, potential users can be mined through big data and the decision-making cost of users can be reduced. For example, in a recommendation system, the recommendation system can predict the future behaviors and interests of users based on the historical behavior data of users. Therefore, a large amount of user behavior data has become an important part and prerequisite of the recommendation system. It can be said that the recommendation system is driven by a large amount of user behavior data and item data. However, for many applications or websites that have just started to implement a recommendation system, there is not enough user behavior data and item data. In the case of no large amount of user behavior data, the recommendation accuracy of the designed recommendation system is relatively low, which easily makes users dissatisfied with the recommendation results and reluctant to use the recommendation system. Therefore, the existing recommendation methods of recommendation systems have the problem that they rely on a large amount of user behavior data to improve the recommendation accuracy. Summary of the Invention

[0003] An embodiment of the present invention provides a cold start recommendation method, which can perform cold start recommendation in the case of a small amount or no user behavior data, and improve the recommendation accuracy of the cold start of the recommendation system.

[0004] In a first aspect, an embodiment of the present invention provides a cold start recommendation method, and the method includes:

[0005] Obtain text vectors of each item under the first item category label, classify the text vectors through a pre-trained classification neural network, and obtain the first item category information of each item;

[0006] Obtain a first mapping relationship between a preset user group and a second item category label, and construct a second mapping relationship from the second item category label to the user group according to the first mapping relationship, where the user group is classified according to user static information;

[0007] According to the second mapping relationship, map the first item category information of each item to the user group, and count the occurrence times of the first item category information corresponding to each item in each user group;

[0008] Based on the occurrence times, assign the first item category label to the corresponding user group, and perform item recommendation for the corresponding user group according to the first item category label and the second item category label.

[0009] Optionally, the pre-trained classification neural network includes a preset number of fully connected layers, and the method further includes:

[0010] In the process of classifying the text vector by the pre-trained classification neural network, perform a fully-connected calculation on the text vector through the preset number of fully-connected layers, and use the output of the last fully-connected layer as the item content vector corresponding to the item;

[0011] Based on the item content vector, perform similar item recommendation.

[0012] Optionally, the step of tagging the corresponding user group with the first item category label based on the occurrence times includes:

[0013] Determine whether there is a user group with an occurrence times greater than or equal to a preset times value, where the preset value is related to the number of items under the first item category label;

[0014] If there is a user group with an occurrence times greater than or equal to the preset times value, then tag the user group with an occurrence times greater than or equal to the preset times value with the first item category label;

[0015] If there is no user group with an occurrence times greater than or equal to the preset times value, then tag the user group with the most occurrence times with the first item category label.

[0016] Optionally, the step of performing item recommendation for the corresponding user group according to the first item category label and the second item category label includes:

[0017] According to the static information of the target user, determine the target user group to which the target user belongs;

[0018] Sample a preset number of item category labels from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user;

[0019] Perform item recommendation for the target user through the item category labels to be recommended.

[0020] Optionally, the step of sampling the item category labels to be recommended from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user includes:

[0021] Obtain the click behavior of the target user group for each item under the candidate item category labels during recommendation, where the candidate item category labels include the first item category label and the second item category label corresponding to the target user group;

[0022] Calculate the sampling weights of each candidate item category label according to the click behavior;

[0023] Based on the sampling weights, sort the candidate item category labels, and sequentially select a preset number of item category labels from the sorted candidate item category labels as the item category labels to be recommended for the target user.

[0024] Optionally, each candidate item category label maintains a first click parameter and a second click parameter. The first click parameter is related to the number of occurrences of the click behavior of each item under the first item category label and the second item category label when recommended by the user group. The second click parameter is related to the number of non-occurrences of the click behavior of each item under the first item category label and the second item category label when recommended by the user group. Calculating the sampling weights of the corresponding item category labels of each item according to the click behavior includes:

[0025] Construct a beta distribution according to the first click parameter and the second click parameter corresponding to each candidate item category label;

[0026] According to the beta distribution, assign a random number to each candidate item category label as the corresponding sampling weight.

[0027] Optionally, the item recommendation for the target user through the item category labels to be recommended includes:

[0028] Obtain the number of occurrences m of the click behavior of the target user in the first preset time period and the item information when the click behavior occurs. The item information includes the item content vector;

[0029] According to the item content vector corresponding to each item, obtain m * k similar items as candidate items for item recommendation;

[0030] Obtain the number of recommendation times n of the target user in the second preset time period and the items that have been recommended in the number of recommendation times n;

[0031] Detect whether there is a duplication between the candidate items and the items that have been recommended;

[0032] Perform a negative correlation adjustment on the recommendation index of the candidate items that are duplicates of the items that have been recommended; and

[0033] Perform a positive correlation adjustment on the recommendation index of the candidate items that are not duplicates of the items that have been recommended;

[0034] According to the recommendation index of the candidate items, select k items from the m * k candidate items as the recommended items and recommend them to the target user.

[0035] Optionally, the method further includes:

[0036] Maintain a first recommendation list and a second recommendation list, sample k items from the first recommendation list and the second recommendation list and add them to the recommendation list, and recommend items to the target user through the recommendation list;

[0037] Among them, the first recommendation list is constructed based on the first item category label and the second item category label corresponding to the static information of the target user; and

[0038] The second recommendation list is constructed based on the item category label to be recommended corresponding to the behavior clicks of the target user.

[0039] In a second aspect, an embodiment of the present invention further provides a cold start recommendation device, and the device includes:

[0040] A first acquisition module, configured to acquire text vectors of each item under the first item category label, classify the text vectors through a pre-trained classification neural network, and obtain first item category information of each item;

[0041] A second acquisition module, configured to acquire a first mapping relationship between a preset user group and a second item category label, and construct a second mapping relationship between the second item category label and the user group according to the first mapping relationship, where the user group is classified according to user static information;

[0042] A first processing module, configured to map the first item category information of each item to the user group according to the second mapping relationship, and count the occurrence times of the first item category information of each item corresponding to each user group;

[0043] A first recommendation module, configured to label the corresponding user group with the first item category label based on the occurrence times, and recommend items to the corresponding user group according to the first item category label and the second item category label.

[0044] In a third aspect, an embodiment of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the steps in the cold start recommendation method provided by the embodiment of the present invention are implemented.

[0045] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the cold start recommendation method provided by the embodiment of the invention are implemented.

[0046] In the embodiments of the present invention, text vectors of various items under the first item category label are obtained, and the text vectors are classified by a pre-trained classification neural network to obtain the first item category information of each item; a first mapping relationship between a preset user group and a second item category label is obtained, and according to the first mapping relationship, a second mapping relationship between the second item category label and the user group is constructed, and the user group is classified according to user static information; according to the second mapping relationship, the first item category information of each item is mapped into the user group, and the occurrence times of the first item category information of each item corresponding to each user group are counted; based on the occurrence times, the first item category label is assigned to the corresponding user group, and item recommendations are made for the corresponding user group according to the first item category label and the second item category label. It is possible to assign the first item category label to the corresponding user group through the mapping relationship between the second item category label and the user group. In the whole process, only the static information of the user group and item information are used, and the cold start of the recommendation system can be realized without a large amount of user behavior data. Moreover, by classifying items through a neural network, the classification result is more accurate, and the cold start recommendation accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0048] Figure 1 is a flowchart of a cold start recommendation method provided by an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of the association between a second item category label and a user group provided by an embodiment of the present invention;

[0050] Figure 3 is a flowchart of another cold start recommendation method provided by an embodiment of the present invention;

[0051] Figure 4 is an architecture diagram of a book recommendation system provided by an embodiment of the present invention;

[0052] Figure 5 is a flowchart of an item recommendation method extracted by an embodiment of the present invention;

[0053] Figure 6 is a schematic diagram of the probability function of a beta distribution provided by an embodiment of the present invention;

[0054] Figure 7 It is a flowchart of another item recommendation method provided by an embodiment of the present invention;

[0055] Figure 8 It is a schematic structural diagram of a cold start recommendation device provided by an embodiment of the present invention;

[0056] Figure 9 It is a schematic structural diagram of another cold start recommendation device provided by an embodiment of the present invention;

[0057] Figure 10 It is a schematic structural diagram of a first recommendation module provided by an embodiment of the present invention;

[0058] Figure 11 It is a schematic structural diagram of another first recommendation module provided by an embodiment of the present invention;

[0059] Figure 12 It is a schematic structural diagram of a sampling sub-module provided by an embodiment of the present invention;

[0060] Figure 13 It is a schematic structural diagram of a calculation unit provided by an embodiment of the present invention;

[0061] Figure 14 It is a schematic structural diagram of a recommendation sub-module provided by an embodiment of the present invention;

[0062] Figure 15 It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0063] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0064] Please refer to Figure 1 , Figure 1 which is a flowchart of a cold start recommendation method provided by an embodiment of the present invention. As shown in Figure 1 , it includes the following steps:

[0065] 101. Obtain the text vectors of each item under the first item category label, classify the text vectors through a pre-trained classification neural network, and obtain the first item category information of each item.

[0066] In an embodiment of the present invention, the above-mentioned first item category label is an item category label associated with a user group. Further, the above-mentioned first item category label can be a newly added item category label or an existing item category label. Specifically, it can be an item category label that needs to be migrated to another recommendation system that does not include this item category label. For example, assume that item category label A is associated with user group a (equivalent to tagging user group a with item category label A), and item category label B is associated with user group b (equivalent to tagging user group b with item category label B). At this time, there is a new item category label C that is not associated with any user group, then this new item category label C is the above-mentioned first item category label. It should be noted that under each item category label, there can be multiple items corresponding to it. For example, if the item category label is a middle school textbook label, then there can be multiple textbooks corresponding to this middle school textbook label, such as textbooks corresponding to middle school Chinese, middle school mathematics, etc.

[0067] The text vectors of the above-mentioned various items can be obtained by vector encoding the text description information of the items through natural language processing methods. The above-mentioned text description information of the items can include information such as name, shape, color, item number, etc. The above-mentioned items can be books, articles, music, commodities, etc. Taking books as an example, the above-mentioned description information can be information such as ISBN (International Standard Book Number), book title, book introduction, book category, etc. The description information of the item can be encoded into a vector space through a natural language processing model NLP (such as a trained bert model) to obtain the text vector of the item.

[0068] The above-mentioned first item category information is used to describe the category of the corresponding item. The category of the item can be defined by the user himself. The probability of the item belonging to each category can be calculated through the above-mentioned pre-trained neural network, and the category with the highest probability is taken as the first item category information of the item.

[0069] Specifically, the above-mentioned pre-trained classification neural network can include a preset number of fully connected layers. After obtaining the text description information of the item, the text description information is encoded into a vector space through a natural language processing model NLP to obtain the initial vectors of each piece of information in the text description information, and the initial vectors of each piece of information are concatenated to obtain the text vector of the item; the text vector is input into the pre-trained classification neural network, and is calculated through a preset number of layers of fully connected layers FC (Fully Connected layers, abbreviated as FC), and the output vector of the last layer is normalized to obtain the probability of the item belonging to each category.

[0070] Among them, the calculation process of the above-mentioned fully connected layer FC can be shown as the following formula:

[0071] z = WT X + b

[0072] a = act(z)

[0073] Where W T is a parameter matrix, X is an input matrix, b is a bias, act is an activation function, and the activation function act is usually relu. The fully connected layer FC of the last layer may not pass through the activation function.

[0074] The above normalization can be the softmax function, specifically as shown in the following formula:

[0075]

[0076] Where z i is the output value of the i-th output node of the fully connected layer FC of the last layer, C is the number of output nodes, each output node outputs an output value, and z c is the output value of each node of the fully connected layer FC of the last layer, that is, the number of classification categories, exp is the exponential function, and p i is the probability of the i-th category.

[0077] In the embodiments of the present invention, a training set can be used to train the classification neural network. The above training set includes sample text vectors corresponding to each item category and real categories. The sample text vectors are input into the classification neural network, calculated through a preset number of fully connected layers FC, and the output vector of the last layer is normalized to obtain the probability that an item belongs to each category. According to the probability of the category, the predicted category of the sample text vector is obtained. The error loss between the predicted category and the real category is calculated through a loss function, and based on this error loss, backpropagation is performed, and the parameters of the classification neural network are adjusted by the method of gradient descent to obtain a classification neural network with accurate item category classification.

[0078] 102. Obtain the first mapping relationship between the preset user group and the second item category label, and construct the second mapping relationship between the second item category label and the user group according to the first mapping relationship.

[0079] In the embodiments of the present invention, the above user group can be classified according to user static information, and the static information can be information such as gender and age. For example, it can be divided into a male user group and a female user group according to gender, and can be divided into a user group over 60 years old, a user group from 35 to 60 years old, a user group from 22 to 35 years old, a user group from 13 to 22 years old, a user group from 7 to 13 years old, and a user group from 0 to 7 years old according to age.

[0080] The above second item category label can be a label manually associated and mapped to the user group. For example, please refer to Figure 2 ,Figure 2 It is a schematic diagram showing the association between the second item category label and the user group provided by an embodiment of the present invention. As Figure 2 shown, when the item label is "elderly", it can be associated and mapped with the user group over 60 years old; when the item label is "middle-aged", "man", "history", "politics", "finance", etc., it can be associated and mapped with the male user group aged 35 to 60 years old; when the item label is "middle-aged", "woman", "family", "health preservation", it can be associated and mapped with the female user group aged 35 to 60 years old; when the item label is "youth", "male", "biography", "society", "investment", etc., it can be associated and mapped with the male user group aged 22 to 35 years old; when the item label is "youth", "female", "emotion", "travel", "food", etc., it can be associated and mapped with the female user group aged 22 to 35 years old; when the item label is "teenager", "boy", "youth", "growth", etc., it can be associated and mapped with the male user group aged 13 to 22 years old; when the item label is "teenager", "girl", "campus", "love", etc., it can be associated and mapped with the female user group aged 13 to 22 years old; when the item label is "juvenile", "boy", etc., it can be associated and mapped with the male user group aged 7 to 13 years old; when the item label is "juvenile", "girl", etc., it can be associated and mapped with the female user group aged 7 to 13 years old; when the item label is "childhood", "enlightenment", it can be associated and mapped with the user group aged 0 to 7 years old.

[0081] Through the first mapping relationship between the user group and the second item category label, when the user group is determined, the corresponding second item category label can be determined. For example, if the user group is the female user group aged 22 to 35 years old, then according to the first mapping relationship, the corresponding second item category label can be determined as "youth", "female", "emotion", "travel", "food", etc. Correspondingly, the second mapping relationship between the second item category label and the user group can be constructed. When the item category label is determined, the corresponding user group can be determined. For example, if the item category label is "youth", the corresponding user groups can be determined as the female user group aged 22 to 35 years old and the male user group aged 22 to 35 years old, and then the second mapping relationship between the second item category label and the user group can be determined.

[0082] It should be noted that the first item category label does not belong to the second item category label. Therefore, the first item category label can be regarded as a new item category label. The first item category information can be the item information of the second item category label. For example, if the first item label is "success", the item information under the corresponding item label includes A, B, and C, and the item information A can also be the item information under the investment label in the second item category label at the same time, B can also be the item information under the finance label in the second item category label at the same time, and C can also be the item information under the biography label in the second item category label at the same time.

[0083] 103. According to the second mapping relationship, map the first item category information of each item to user groups, and count the occurrence times of the first item category information corresponding to each item for each type of user group.

[0084] In the embodiments of the present invention, the second item category label corresponding to each item can be determined according to the first item category information of each item, and then, according to the second mapping relationship between the second item category label and user groups, the user group corresponding to the first item category information can be determined.

[0085] For example, the first item category label is success, the first item information under the corresponding item label includes A, B, and C, and the first item information A can also be the item information A under the investment label in the second item category label at the same time. The user group corresponding to the investment label is male users aged 22 to 35 years old, that is, the user group corresponding to the first item information A is male users aged 22 to 35 years old; the first item information B can also be the item information under the financial management label in the second item category label at the same time. The user group corresponding to the financial management label is male users aged 35 to 60 years old, that is, the user group corresponding to the first item information B is male users aged 35 to 60 years old; the first item information C can also be the item information under the biography label in the second item category label at the same time. The user group corresponding to the biography label is male users aged 22 to 35 years old, that is, the user group corresponding to the first item information C is male users aged 22 to 35 years old.

[0086] The occurrence times of the first item category information of each item for each type of user group above can be further understood as the number of items belonging to each type of user group in the first item category label. Taking the above example for illustration, the first item category label is success, the first item information under the corresponding item label includes A, B, and C. The user group corresponding to the first item information A is male users aged 22 to 35 years old, the user group corresponding to the first item information B is male users aged 35 to 60 years old, and the user group corresponding to the first item information A is male users aged 22 to 35 years old. It can be seen that under the first item category label, the occurrence times of male users aged 22 to 35 years old are 2, and that of male users aged 35 to 60 years old is 1.

[0087] 104. Based on the occurrence times, assign the first item category label to the corresponding user group, and perform item recommendation for the corresponding user group according to the first item category label and the second item category label.

[0088] In the embodiments of the present invention, the above-mentioned occurrence times refer to the occurrence times of each user group under the first item category label.

[0089] Further, it is possible to determine whether there is a user group whose occurrence times are greater than or equal to a preset number value, and the preset value is related to the number of items under the first item category label; for example, it is possible to determine whether there is a user group whose occurrence times are greater than or equal to half of the number of items under the first item category label.

[0090] If there is a user group whose occurrence times are greater than or equal to the preset number value, then assign the first item category label to the user group whose occurrence times are greater than or equal to the preset number value; if there is no user group whose occurrence times are greater than or equal to the preset number value, then assign the first item category label to the user group with the most occurrence times.

[0091] For example, the first item category label is "success", the first item information under the corresponding item label includes A, B, and C. The user group corresponding to the first item information A is male users aged 22 to 35, the user group corresponding to the first item information B is male users aged 35 to 60, and the user group corresponding to the first item information A is male users aged 22 to 35. It can be seen that under the first item category label, the occurrence times of the male user group aged 22 to 35 are 2, and the occurrence times of the male user group aged 35 to 60 are 1. Then, the male user group aged 22 to 35 can be assigned the item category label of "success".

[0092] In the embodiments of the present disclosure, it is possible to recommend items under the corresponding item category label to the corresponding user group according to the first item category label and the second item category label of the user group. Specifically, according to the static information of the target user, determine the target user group to which the target user belongs, and recommend a preset number of items to the target user according to the first item category label and the second item category label corresponding to the target user group.

[0093] In an embodiment of the present invention, text vectors of each item under the first item category label are obtained, and the text vectors are classified by a pre-trained classification neural network to obtain the first item category information of each item; a first mapping relationship between a preset user group and a second item category label is obtained, and according to the first mapping relationship, a second mapping relationship from the second item category label to the user group is constructed, and the user group is classified according to user static information; according to the second mapping relationship, the first item category information of each item is mapped into the user group, and the occurrence times of the first item category information corresponding to each item for each type of user group are counted; based on the occurrence times, the first item category label is assigned to the corresponding user group, and item recommendations are made for the corresponding user group according to the first item category label and the second item category label. It is possible to assign the first item category label to the corresponding user group through the mapping relationship between the second item category label and the user group. In the whole process, only the static information of the user group and item information are used, and the cold start of the recommendation system can be realized without a large amount of user behavior data. Moreover, the items are classified by a neural network, making the classification result more accurate and improving the recommendation accuracy.

[0094] It should be noted that generally, in a recommendation system based on item recommendation, only the text vectors of the items need to be obtained and stored. When a target user clicks on an item, the text vector of this item will be taken out and the similarity is calculated with the text vectors of other items, and the k most similar items are selected and recommended to the target user.

[0095] Optionally, please refer to Figure 3 , Figure 3 which is a flowchart of another cold start recommendation method in an embodiment of the present invention. In an embodiment of the present invention, the above-mentioned pre-trained classification neural network may include a preset number of fully connected layers. As Figure 3 shown, it includes the following steps:

[0096] 301. Obtain the text description information of each item.

[0097] 302. Encode the text description information of each item into a vector space through a natural language processing method to obtain the text vectors of each item.

[0098] 303. Input the text vector of the item into the pre-trained classification neural network. During the process of classifying the text vector by the pre-trained classification neural network, perform a full connection calculation on the text vector through a preset number of fully connected layers, and use the output of the last fully connected layer as the item content vector corresponding to the item.

[0099] In an embodiment of the present invention, in step 302 above, the trained bert (Bidirectional Encoder Representation from Transformers) model can be used to encode the text description information into a vector space. In step 303 above, after the calculation result is output by the last fully connected layer, the calculation result is not calculated by an activation function, and the calculation result output by the last fully connected layer is directly used as the item content vector. It can be understood that the content classification information is implicitly contained in the item content vector. When calculating the similarity of item content vectors for items of the same category, the metric distance is closer than when directly calculating the similarity using text vectors, thereby improving the accuracy of recommendations.

[0100] Specifically, taking the above item as a book for example, please refer to Figure 4 , Figure 4 which is an architecture diagram of a book recommendation system provided by an embodiment of the present invention. As shown in Figure 4 , it includes an input layer, a vector encoding layer (Embedding layer), a fully connected layer FC, and a classification task layer (softmax). Among them, the input of the input layer is the text description information of the book. The text description information of the book may include a book introduction, a book title, a book category, a book ISBN, etc. The above text description information is encoded by the vector encoding layer and spliced into a text vector, and the text vector is input into the fully connected layer FC. The fully connected layer FC includes multiple fully connected layer sub-networks. Through the output of the last fully connected layer sub-network, the book content vector is obtained, and the book content vector is stored in the database.

[0101] When training the fully connected layer FC and the classification task layer, the book recommendation system further includes a loss calculation layer loss. The fully connected layer FC and the classification task layer are trained through a training set. The training set includes m training samples and corresponding true categories. The training samples are text vectors of items. After the sample text vectors pass through the fully connected layer FC and the classification task layer, the predicted category of the text vector is obtained. The error between the predicted category and the true category is calculated by the loss function in the loss calculation layer loss, and the parameters in the fully connected layer FC are adjusted according to the error. Specifically, the above loss function can be shown as the following formula:

[0102]

[0103] where m is the number of samples, p i is the predicted category corresponding to the i-th sample, yi is the true category corresponding to the i-th sample, and C is the number of output nodes.

[0104] 304. Perform similar item recommendations based on item content vectors.

[0105] When making book recommendations, when the target user clicks on a book, the book content vector of that book can be obtained, and the similarity is calculated with all book content vectors in the database. Then, the k books corresponding to the book content vectors with the highest similarity are selected and recommended to the target user. Specifically, the above similarity calculation can be shown as the following formula:

[0106]

[0107] Among them, the above A and B are the item content vectors of two items. Compared with the existing method of calculating the similarity of text vectors, which requires calculating the component similarities of text vectors first (such as name similarity, introduction similarity, etc.) and then summing up the component similarities, the similarity between two vectors can be directly calculated based on the item content vector without calculating the component similarities. The similarity calculation of the existing text vectors can be shown as the following formula:

[0108]

[0109] Among them, the above A and B are the text vectors of two items, and the above A i and B i are the vector components of A and B respectively, and the above n is the number of micro components.

[0110] In the embodiments of the present invention, the text vectors are classified by a classification neural network to obtain high-level semantic vectors (item content vectors) with implicit classification information. Through the item content vectors, when calculating the similarity, the classification information implicit in the classification content can be considered, and the similarity metrics of items in the same category are closer. Furthermore, more similar items can be recommended to the target user, improving the accuracy of the recommendation.

[0111] Optionally, please refer to Figure 5 , Figure 5 which is a flowchart of an item recommendation method extracted in the embodiments of the present invention. As Figure 5 shown, the item recommendation method includes the following steps:

[0112] 501. According to the static information of the target user, determine the target user group to which the target user belongs.

[0113] In the embodiments of the present invention, the above static information of the target user can be information such as the user's gender and age. For example, if the target user is a 24-year-old female, the target user group to which the target user belongs is the female user group aged 22 to 35.

[0114] 502. Sample a preset number of item category labels from the first item category labels and the second item category labels corresponding to the target user group as the item category labels to be recommended for the target user.

[0115] In the embodiment of the present invention, the first item category labels corresponding to the target user group are newly added item category labels, and the second item category labels are original item category labels. In some possible embodiments, the first item category labels are empty. In this case, it is equivalent to only sampling the second item category labels.

[0116] The preset number is set in advance. For example, if the target user group to which the target user belongs is a female user group aged 22 to 35, and the corresponding item category labels are youth, female, emotion, travel, food, etc., and the preset number is 3, then 3 item category labels will be sampled from youth, female, emotion, travel, food as the item category labels to be recommended for the target user.

[0117] Optionally, the click behavior of the target user group for each item under the candidate item category labels during recommendation can be obtained. The candidate item category labels include the first item category labels and the second item category labels corresponding to the target user group. Further, the candidate item category labels can be understood as all item category labels corresponding to the target user group. It should be noted that when there are no corresponding first item category labels in the target user group, the first item category labels can be not considered. According to the click behavior, calculate the sampling weights of each candidate item category label. Based on the sampling weights, sort the candidate item category labels, and sequentially select a preset number of item category labels from the sorted candidate item category labels as the item category labels to be recommended for the target user.

[0118] Further, in the embodiment of the present invention, each candidate item category label maintains a first click parameter and a second click parameter. The first click parameter is related to the number of occurrences of the click behavior of the user group for each item under the first item category labels and the second item category labels during recommendation. The second click parameter is related to the number of non-occurrences of the click behavior of the user group for each item under the first item category labels and the second item category labels during recommendation. Furthermore, according to the first click parameter and the second click parameter corresponding to each candidate item category label, construct a beta distribution; according to the beta distribution, assign a random number to each candidate item category label as the corresponding sampling weight.

[0119] It can be understood that in existing recommendation methods, one or several item category labels are sampled from all item category labels. Usually, a uniform distribution is used, which means that the sampling probability of each item category label is the same, indicating that the degree of interest of this group of people in all item category labels is the same. However, in reality, the degree of interest of the population in all item category labels of this population is different and usually conforms to a long-tail distribution. The population has a relatively high degree of interest in several item category labels among all item category labels of this population, and a general or relatively low degree of interest in other item category labels. Therefore, in the embodiments of the present invention, the sampling weight of each item category label is dynamically adjusted, and the sampling of the item category label is determined according to the sampling weight.

[0120] Specifically, the beta distribution is the density function of the conjugate prior distribution of the Bernoulli distribution and the binomial distribution. The probability density formula can be shown as the following formula:

[0121]

[0122] Where Γ here represents the gamma function, α is the first click parameter of the beta distribution in the embodiments of the present invention, β is the second click parameter of the beta distribution in the embodiments of the present invention, and u is the mean value of the random variable x. It can be understood that the beta distribution maintains a set of parameters (α, β). The first click parameter α represents the number of times an event occurs (the number of times a click behavior occurs), and the second click parameter β represents the number of times an event does not occur (the number of times a click behavior does not occur). Figure 6 is a schematic diagram of the probability function of a beta distribution provided by the embodiments of the present invention. As Figure 6 shown, there are many shapes of the beta distribution, but they are all within the 0-1 interval. Therefore, the beta distribution can describe various shapes (events) within the 0-1 interval. For example, it can describe the sampling probability of item category labels. In the embodiments of the present invention, for each item category label corresponding to the user group, a set of parameters of the beta distribution is maintained, that is, the first click parameter α and the second click parameter β. When a recommended item of this item category label is recommended to a user, if the user clicks on the recommended item, then the first click parameter α is incremented by 1. If the user does not click on the recommended item, then the second click parameter β is incremented by 1.

[0123] Furthermore, based on the above beta distribution, the Thompson sampling algorithm is combined to update the first click parameter and the second click parameter of the user, so that the item category label with a high number of click behavior occurrences has a high sampling probability, thereby maximizing the expected click of the click behavior. It can be continuously optimized through the click behavior feedback of the user, so as to learn the item type preferences of this user group, and thus determine the corresponding item category label.

[0124] For each candidate item category label, extract the first click parameter α and the second click parameter β, and generate a random number using the beta distribution. The candidate item category labels can be sorted according to the magnitude of the above random number, and the top j item category labels can be selected as the item category labels to be recommended for the target user.

[0125] 503. Perform item recommendation for the target user based on the item category labels to be recommended.

[0126] In an embodiment of the present invention, k items can be selected from the item category labels to be recommended as the recommended items, and the k items can be recommended to the target user.

[0127] Optionally, please refer to Figure 7 , Figure 7 which is a flowchart of another item recommendation method provided by an embodiment of the present invention. As shown in Figure 7 , it includes the following steps:

[0128] 701. Obtain the number of occurrences m of the click behavior of the target user in the first preset time period and the item information when the click behavior occurs.

[0129] In an embodiment of the present invention, the above item information includes an item content vector. The above first preset time period can be 1 month, 1 quarter, etc. Through the occurrence of the above click behavior, the items of interest to the target user can be determined. The number of occurrences m of the click behavior indicates that the target user is interested in m recommended items.

[0130] 702. m*k similar items can be obtained as candidate items for item recommendation according to the item content vector corresponding to each item.

[0131] In an embodiment of the present invention, each of the above items is one of the m recommended items of interest to the target user. The above item content vector can be obtained from a database or can be obtained according to the text vector of the item and the above classification neural network.

[0132] 703. Obtain the number of recommendation times n of the target user in the second preset time period and the items that have been recommended in the number of recommendation times n.

[0133] In an embodiment of the present invention, the above second preset time period can be the same as or different from the first preset time period. For example, if the first preset time period is 1 month, the second preset time period can be 1 month or half a month. The number of items that have been recommended in the number of recommendation times n is n*k (k items are recommended each time).

[0134] 704. Detect whether there are duplicates between the candidate items and the items that have been recommended.

[0135] In an embodiment of the present invention, the similarity between a candidate item and a recommended item can be calculated based on the item content vector of the candidate item and the item content vector of the recommended item, and it is determined whether there is a duplicate according to the similarity. For example, if the similarity between a candidate item and a recommended item is greater than a preset threshold, it indicates that the candidate item is a duplicate.

[0136] 705. Make a negatively correlated adjustment to the recommendation index for candidate items that are duplicates of the recommended items.

[0137] In an embodiment of the present invention, the above-mentioned negatively correlated adjustment of the recommendation index can be understood as reducing the recommendation index of the candidate item, such as subtracting 0.1 from the recommendation index, etc. The above-mentioned recommendation index can be expressed as a recommendation probability. The lower the recommendation index, the lower the recommendation probability.

[0138] 706. Make a positively correlated adjustment to the recommendation index for candidate items that are not duplicates of the recommended items.

[0139] In an embodiment of the present invention, the above-mentioned positively correlated adjustment of the recommendation index can be understood as increasing the recommendation index of the candidate item. For example, the recommendation index of the candidate item is adjusted to the highest recommendation index, and the highest recommendation index is 1.

[0140] 707. Select k items from m*k candidate items as recommended items and recommend them to the target user according to the recommendation index of the candidate items.

[0141] In an embodiment of the present invention, the k candidate items with the highest recommendation index can be selected as recommended items and recommended to the target user.

[0142] In a possible way, the above-mentioned recommendation index can be converted into a recommendation probability. For example, through normalization, the recommendation indexes of all candidate items are normalized to obtain a recommendation probability. According to this recommendation probability, k candidate items are randomly selected as recommended items and recommended to the target user. It should be noted that the higher the recommendation probability, the higher the probability of being randomly selected.

[0143] In an embodiment of the present invention, by detecting whether there is a duplicate between a candidate item and a recommended item, and making a downward adjustment to the recommendation index of the duplicate candidate item, the recommendation rate of the duplicate candidate item is reduced, so as to control the exposure during the recommendation process of the item, and avoid the decline of the user experience caused by repeatedly recommending an item.

[0144] Optionally, the first recommendation list and the second recommendation list can be maintained. Specifically, the first recommendation list of the target user can be constructed and recommended according to the item category labels, or the second recommendation list of the target user can be constructed and recommended according to the click behavior of the target user. In the embodiments of the present invention, k items can be sampled from the above first recommendation list and second recommendation list and added to the recommendation list, and item recommendations are made for the target user through the above recommendation list; wherein, the above first recommendation list is constructed based on the first item category label and the second item category label corresponding to the static information of the above target user; and the above second recommendation list is constructed based on the item category label to be recommended corresponding to the click behavior of the above target user. It can be understood that the final recommendation list will be sampled from the first recommendation list and the second recommendation list. In order to better fit the user's interests, it is more likely to sample from the second list. At the same time, in order to ensure the diversity of the recommendation results, it will also sample from the first recommendation list with a lower probability to obtain the corresponding recommendation list. The above recommendation list includes k or more candidate items, and finally k items are recommended to the user through the recommendation list.

[0145] It should be noted that the cold start recommendation method provided in the embodiments of the present invention can be applied to devices such as mobile phones, monitors, computers, and servers that can perform cold start recommendations.

[0146] Please refer to Figure 8 , Figure 8 which is a schematic structural diagram of a cold start recommendation device provided in the embodiments of the present invention. As Figure 8 shown, the device includes:

[0147] The first acquisition module 801 is configured to acquire text vectors of each item under the first item category label, classify the text vectors through a pre-trained classification neural network, and obtain the first item category information of each item;

[0148] The second acquisition module 802 is configured to acquire a first mapping relationship between a preset user group and a second item category label, and construct a second mapping relationship between the second item category label and the user group according to the first mapping relationship, and the user group is classified according to user static information;

[0149] The first processing module 803 is configured to map the first item category information of each item to the user group according to the second mapping relationship, and count the occurrence times of the first item category information of each item corresponding to each user group;

[0150] The first recommendation module 804 is used to label the corresponding user groups with the first item category label based on the occurrence times, and perform item recommendations for the corresponding user groups according to the first item category label and the second item category label.

[0151] Optionally, as Figure 9 shown, the pre-trained classification neural network includes a preset number of fully connected layers, and the device further includes:

[0152] The second processing module 805 is used to perform a fully connected calculation on the text vector through the preset number of fully connected layers during the process of classifying the text vector by the pre-trained classification neural network, and use the output of the last fully connected layer as the item content vector of the corresponding item;

[0153] The second recommendation module 806 is used to perform similar item recommendations based on the item content vector.

[0154] Optionally, as Figure 10 shown, the first recommendation module 804 includes:

[0155] The first judgment sub-module 8041 is used to judge whether there is a user group with an occurrence times greater than or equal to a preset times value, and the preset value is related to the number of items under the first item category label;

[0156] The first labeling module 8042 is used to label the user group with an occurrence times greater than or equal to the preset times value with the first item category label if there is such a user group;

[0157] The second labeling module 8043 is used to label the user group with the most occurrence times with the first item category label if there is no user group with an occurrence times greater than or equal to the preset times value.

[0158] Optionally, as Figure 11 shown, the first recommendation module 804 further includes:

[0159] The second judgment sub-module 8044 is used to judge the target user group to which the target user belongs according to the static information of the target user;

[0160] The sampling sub-module 8045 is used to sample a preset number of item category labels from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user;

[0161] The recommendation sub-module 8046 is used to perform item recommendations for the target user through the item category labels to be recommended.

[0162] Optionally, as Figure 12 shown, the sampling sub-module 8045 includes:

[0163] A first obtaining unit 80451, configured to obtain the click behavior of the target user group on each item under the candidate item category label during recommendation, where the candidate item category label includes a first item category label and a second item category label corresponding to the target user group;

[0164] A calculation unit 80452, configured to calculate the sampling weight of each candidate item category label according to the click behavior;

[0165] A first selection unit 80453, configured to sort the candidate item category labels based on the sampling weight, and sequentially select a preset number of item category labels from the sorted candidate item category labels as the item category labels to be recommended for the target user.

[0166] Optionally, as Figure 13 shown, each candidate item category label maintains a first click parameter and a second click parameter. The first click parameter is related to the number of occurrences of the click behavior of the user group on each item under the first item category label and the second item category label during recommendation, and the second click parameter is related to the number of non-occurrences of the click behavior of the user group on each item under the first item category label and the second item category label during recommendation. The calculation unit 80452 includes:

[0167] A construction subunit 804521, configured to construct a beta distribution according to the first click parameter and the second click parameter corresponding to each candidate item category label;

[0168] An allocation subunit 804522, configured to allocate a random number to each candidate item category label as the corresponding sampling weight according to the beta distribution.

[0169] Optionally, as Figure 14 shown, the recommendation sub-module 8046 includes:

[0170] A second obtaining unit 80461, configured to obtain the number of occurrences m of the click behavior of the target user in the first preset time period and the item information when the click behavior occurs, where the item information includes an item content vector;

[0171] A third obtaining unit 80462, configured to obtain m*k similar items as candidate items for item recommendation according to the item content vector corresponding to each item;

[0172] The fourth acquisition unit 80463 is configured to acquire the number of recommendations n of the target user in the second preset time period and the items that have been recommended in the number of recommendations n;

[0173] The detection unit 80464 is configured to detect whether there is a duplication between the candidate item and the item that has been recommended;

[0174] The first adjustment unit 80465 is configured to perform a negatively correlated adjustment on the recommendation index of the candidate item that duplicates the item that has been recommended; and

[0175] The second adjustment unit 80466 is configured to perform a positively correlated adjustment on the recommendation index of the candidate item that does not duplicate the item that has been recommended;

[0176] The second selection unit 80467 is configured to select k items from the m×k candidate items as the recommended items according to the recommendation index of the candidate items and recommend them to the target user.

[0177] Optionally, the first recommendation module 804 is further configured to maintain a first recommendation list and a second recommendation list, sample k items from the first recommendation list and the second recommendation list and add them to the recommendation list, and perform item recommendation for the target user through the recommendation list;

[0178] Wherein, the first recommendation list is constructed based on the first item category label and the second item category label corresponding to the static information of the target user; and

[0179] The second recommendation list is constructed based on the item category label to be recommended corresponding to the behavior clicks of the target user.

[0180] It should be noted that the cold start recommendation device provided by the embodiments of the present invention can be applied to devices such as mobile phones, monitors, computers, and servers that can perform cold start recommendations.

[0181] The cold start recommendation device provided by the embodiments of the present invention can implement each process implemented by the cold start recommendation method in the above method embodiments and can achieve the same beneficial effects. To avoid repetition, details are not described herein again.

[0182] See Figure 15 , Figure 15 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. As Figure 15 shown, it includes: a memory 1502, a processor 1501, and a computer program stored on the memory 1502 and executable on the processor 1501. Among them:

[0183] The processor 1501 is configured to call the computer program stored in the memory 1502 and execute the following steps:

[0184] Obtain the text vectors of each item under the first item category label, and classify the text vectors through a pre-trained classification neural network to obtain the first item category information of each item;

[0185] Obtain the first mapping relationship between the preset user groups and the second item category label, and construct the second mapping relationship from the second item category label to the user groups according to the first mapping relationship, where the user groups are classified according to user static information;

[0186] According to the second mapping relationship, map the first item category information of each item to the user groups, and count the occurrence times of the first item category information of each item corresponding to each user group;

[0187] Based on the occurrence times, label the corresponding user groups with the first item category label, and perform item recommendation for the corresponding user groups according to the first item category label and the second item category label.

[0188] Optionally, the pre-trained classification neural network includes a preset number of fully connected layers, and the processor 1501 also executes the following:

[0189] In the process of classifying the text vectors through the pre-trained classification neural network, perform full connection calculation on the text vectors through the preset number of fully connected layers, and use the output of the last fully connected layer as the item content vector of the corresponding item;

[0190] Based on the item content vector, perform similar item recommendation.

[0191] Optionally, the processor 1501 executes the step of labeling the corresponding user groups with the first item category label based on the occurrence times, including:

[0192] Judge whether there is a user group with an occurrence times greater than or equal to a preset times value, where the preset value is related to the number of items under the first item category label;

[0193] If there is a user group with an occurrence times greater than or equal to the preset times value, then label the user group with an occurrence times greater than or equal to the preset times value with the first item category label;

[0194] If there is no user group with an occurrence times greater than or equal to the preset times value, then label the user group with the most occurrence times with the first item category label.

[0195] Optionally, the processor 1501 executes the step of performing item recommendation for the corresponding user groups according to the first item category label and the second item category label, including:

[0196] Based on the static information of the target user, determine the target user group to which the target user belongs;

[0197] Sample a preset number of item category labels from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user;

[0198] Recommend items to the target user through the item category labels to be recommended.

[0199] Optionally, the processor 1501 executes sampling the item category labels to be recommended from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user, including:

[0200] Obtain the click behavior of the target user group for each item under the candidate item category labels, where the candidate item category labels include the first item category label and the second item category label corresponding to the target user group;

[0201] Calculate the sampling weights of each candidate item category label according to the click behavior;

[0202] Based on the sampling weights, sort the candidate item category labels, and sequentially select a preset number of item category labels from the sorted candidate item category labels as the item category labels to be recommended for the target user.

[0203] Optionally, each candidate item category label maintains a first click parameter and a second click parameter. The first click parameter is related to the number of times of click behavior when the user group recommends each item under the first item category label and the second item category label, and the second click parameter is related to the number of times of non - occurrence of click behavior when the user group recommends each item under the first item category label and the second item category label. The processor 1501 executes calculating the sampling weights of the corresponding item category labels of each item according to the click behavior, including:

[0204] Construct a beta distribution according to the first click parameter and the second click parameter corresponding to each candidate item category label;

[0205] Assign a random number to each candidate item category label as the corresponding sampling weight according to the beta distribution.

[0206] Optionally, the processor 1501 executes recommending items to the target user through the item category labels to be recommended, including:

[0207] Obtain the number of occurrences m of the click behavior of the target user in the first preset time period and the item information when the click behavior occurs, where the item information includes an item content vector;

[0208] According to the item content vector corresponding to each item, obtain m*k similar items as candidate items for item recommendation;

[0209] Obtain the number of recommended times n of the target user in the second preset time period and the items that have been recommended among the number of recommended times n;

[0210] Detect whether there is a duplication between the candidate items and the recommended items;

[0211] Perform a negative correlation adjustment on the recommendation index of the candidate items that are duplicates of the recommended items; and

[0212] Perform a positive correlation adjustment on the recommendation index of the candidate items that are not duplicates of the recommended items;

[0213] Select k items from the m*k candidate items as recommended items according to the recommendation index of the candidate items and recommend them to the target user.

[0214] Optionally, the processor 1501 also executes including:

[0215] Maintain a first recommendation list and a second recommendation list, sample k items from the first recommendation list and the second recommendation list and add them to the recommendation list, and perform item recommendation for the target user through the recommendation list;

[0216] Wherein, the first recommendation list is constructed based on the first item category label and the second item category label corresponding to the static information of the target user; and

[0217] The second recommendation list is constructed based on the item category label to be recommended corresponding to the behavioral clicks of the target user.

[0218] It should be noted that the above electronic device can be a mobile phone, a monitor, a computer, a server, etc. that can be applied to cold start recommendation.

[0219] The electronic device provided by the embodiments of the present invention can implement each process implemented by the cold start recommendation method in the above method embodiments, and can achieve the same beneficial effects. To avoid repetition, it will not be elaborated here.

[0220] The embodiments of the present invention also provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements each process of the cold start recommendation method provided by the embodiments of the present invention, and can achieve the same technical effects. To avoid repetition, it will not be elaborated here.

[0221] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above various methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0222] The above-disclosed are only the preferred embodiments of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope covered by the present invention.

Claims

1. A cold start recommendation method, characterized in that, it includes the following steps: Obtain the text vectors of each item under the first item category label, classify the text vectors through a pre-trained classification neural network, and obtain the first item category information of each item; Obtain the first mapping relationship between the preset user group and the second item category label, and construct the second mapping relationship from the second item category label to the user group according to the first mapping relationship. The user group is classified according to user static information. The first item category label is a new item category label, and the first item category information is the item information of the second item category label; According to the second mapping relationship, map the first item category information of each item to the user group, and count the occurrence times of the first item category information corresponding to each item in each user group; Based on the occurrence times, label the corresponding user group with the first item category label, and perform item recommendation for the corresponding user group according to the first item category label and the second item category label.

2. The method according to claim 1, characterized in that, The pre-trained classification neural network includes a preset number of fully connected layers, and the method further includes: During the process of classifying the text vectors through the pre-trained classification neural network, perform a full connection calculation on the text vectors through the preset number of fully connected layers, and use the output of the last fully connected layer as the item content vector of the corresponding item; Based on the item content vector, perform similar item recommendation.

3. The method according to claim 2, characterized in that, The step of labeling the corresponding user group with the first item category label based on the occurrence times includes: Judge whether there is a user group with an occurrence times greater than or equal to a preset times value, and the preset value is related to the number of items under the first item category label; If there is a user group with an occurrence times greater than or equal to the preset times value, then label the user group with an occurrence times greater than or equal to the preset times value with the first item category label; If there is no user group with an occurrence times greater than or equal to the preset times value, then label the user group with the most occurrence times with the first item category label.

4. The method according to claim 3, characterized in that, The step of performing item recommendation for the corresponding user group according to the first item category label and the second item category label includes: According to the static information of the target user, judge the target user group to which the target user belongs; Sample a preset number of item category labels from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user; Perform item recommendation for the target user through the item category labels to be recommended.

5. The method according to claim 4, characterized in that, The step of sampling the item category labels to be recommended from the first item category label and the second item category label corresponding to the target user group as the item category labels to be recommended for the target user includes: Obtain the click behavior of the target user group on each item under the candidate item category labels, where the candidate item category labels include the first item category label and the second item category label corresponding to the target user group; Calculate the sampling weights of each candidate item category label according to the click behavior; Based on the sampling weights, sort the candidate item category labels, and sequentially select a preset number of item category labels from the sorted candidate item category labels as the item category labels to be recommended for the target user.

6. The method according to claim 5, wherein, Each candidate item category label maintains a first click parameter and a second click parameter. The first click parameter is related to the number of occurrences of the click behavior of the user group on each item under the first item category label and the second item category label during recommendation. The second click parameter is related to the number of non-occurrences of the click behavior of the user group on each item under the first item category label and the second item category label during recommendation. The calculating the sampling weights of the corresponding item category labels of each item according to the click behavior includes: Construct a beta distribution according to the first click parameter and the second click parameter corresponding to each candidate item category label; According to the beta distribution, assign a random number to each candidate item category label as the corresponding sampling weight.

7. The method according to claim 6, wherein, The performing item recommendation for the target user through the item category labels to be recommended includes: Obtain the number of occurrences m of the click behavior of the target user in the first preset time period and the item information when the click behavior occurs. The item information includes the item content vector; According to the item content vector corresponding to each item, obtain m * k similar items as candidate items for item recommendation; Obtain the number of recommendations n of the target user in the second preset time period and the items that have been recommended in the number of recommendations n; Detect whether there is a duplicate between the candidate items and the items that have been recommended; Perform a negatively correlated adjustment of the recommendation index on the candidate items that are duplicates of the items that have been recommended; and Perform a positively correlated adjustment of the recommendation index on the candidate items that are not duplicates of the items that have been recommended; According to the recommendation index of the candidate items, select k items from the m * k candidate items as recommended items to recommend to the target user.

8. The method according to any one of claims 5 to 7, wherein, The method further includes: Maintain a first recommendation list and a second recommendation list, sample k items from the first recommendation list and the second recommendation list and add them to the recommendation list, and perform item recommendation for the target user through the recommendation list; wherein, the first recommendation list is constructed based on the first item category label and the second item category label corresponding to the static information of the target user; and The second recommendation list is constructed based on the item category labels to be recommended corresponding to the behavior clicks of the target user.

9. A cold start recommendation device, wherein, The device includes: A first acquisition module, configured to acquire text vectors of each item under a first item category label, classify the text vectors through a pre-trained classification neural network, and obtain first item category information of each item; A second acquisition module, configured to acquire a first mapping relationship between a preset user group and a second item category label, and construct a second mapping relationship between the second item category label and the user group according to the first mapping relationship, where the user group is classified according to user static information, the first item category label is a new item category label, and the first item category information is item information of the second item category label; A first processing module, configured to map the first item category information of each item to the user group according to the second mapping relationship, and count the occurrence times of the first item category information of each item corresponding to each type of user group; A first recommendation module, configured to label the corresponding user group with the first item category label based on the occurrence times, and perform item recommendation for the corresponding user group according to the first item category label and the second item category label.

10. An electronic device, characterized in that it includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the computer program, the steps in the cold start recommendation method according to any one of claims 1 to 8 are implemented.

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

Images