WO2022151649A1

WO2022151649A1 - Deep interest network-based topic recommendation method and apparatus

Info

Publication number: WO2022151649A1
Application number: PCT/CN2021/099766
Authority: WO
Inventors: 刘志杰; 陈鑫晶; 蔡淇森
Original assignee: 稿定（厦门）科技有限公司
Priority date: 2021-01-15
Filing date: 2021-06-11
Publication date: 2022-07-21
Also published as: CN113688167A; CN112800097A

Abstract

The present application discloses a deep interest network-based topic recommendation method and apparatus, a medium, and a device. The deep interest network-based topic recommendation method comprises: obtaining user information and historical click data of a user, and generating training data; obtaining, by training, a deep interest capture model; obtaining item information corresponding to an item, outputting a corresponding item vector by means of the deep interest capture model, and calculating a topic vector according to an item vector corresponding to each item; obtaining click data to be analyzed of the user, and outputting a corresponding user vector by means of the deep interest capture model; and performing similarity retrieval according to the user vector and the topic vector, determining a topic recommendation list according to the retrieval result, and pushing the topic recommendation list to the user. The application can precisely perform topic recommendation on a user without establishing labels corresponding to a topic, thereby reducing manpower and material resources consumed during topic recommendation.

Description

Method and device for topic recommendation based on deep interest network

technical field

The invention relates to the technical field of deep learning, and in particular, to a topic recommendation method based on a deep learning network, a computer-readable storage medium, a computer device, and a topic recommendation device based on a deep interest network.

Background technique

In the related art, when it is necessary to recommend corresponding topics for users, the method of portrait is often used; that is, first, based on the rules, the user's preference scores for different topics are counted; Priority is given to display to complete the recommendation of the topic; however, this method is highly dependent on the tags corresponding to the topic. In order to improve the accuracy of the topic recommendation, a lot of manpower and material resources must be spent to establish high-quality tags.

SUMMARY OF THE INVENTION

The present invention aims to solve one of the technical problems in the above technologies at least to a certain extent. Therefore, an object of the present invention is to propose a topic recommendation method based on a deep interest network, which can accurately recommend topics to users without establishing labels corresponding to topics, and reduce the cost of the topic recommendation process. Human and material resources.

A second object of the present invention is to provide a computer-readable storage medium.

The third object of the present invention is to propose a computer device.

The fourth object of the present invention is to provide a topic recommendation device based on a deep interest network.

In order to achieve the above object, the embodiment of the first aspect of the present invention proposes a topic recommendation method based on a deep interest network, including the following steps: acquiring user information and historical click data of users, and according to the user information and the historical click data data to generate training data; perform model training according to the training data to obtain a deep interest capture model; obtain item information corresponding to an item, and input the item information into the deep interest capture model to capture through the deep interest The model outputs the corresponding item vector, and calculates the thematic vector according to the item vector corresponding to each item; obtains the user's click data to be analyzed, and inputs the to-be-analyzed click data into the depth interest capture model to pass the depth The interest capture model outputs a corresponding user vector; performs similarity retrieval according to the user vector and the topic vector, determines a topic recommendation list according to the retrieval result, and pushes the topic recommendation list to the user.

According to the topic recommendation method based on the deep interest network according to the embodiment of the present invention, first, user information and historical click data of the user are obtained, and training data is generated according to the user information and the historical click data; then, according to the training data Perform model training to obtain a deep interest capture model; then, obtain item information corresponding to the item, and input the item information into the deep interest capture model, to output a corresponding item vector through the deep interest capture model, and Calculate the thematic vector according to the item vector corresponding to each item; then, obtain the user's click data to be analyzed, and input the to-be-analyzed click data into the deep interest capture model, so as to output the corresponding deep interest capture model through the deep interest capture model User vector; then, carry out similarity retrieval according to the user vector and the topic vector, and determine the topic recommendation list according to the retrieval result, and push the topic recommendation list to the user; so as to realize that there is no need to establish a label corresponding to the topic Under the premise of , it can accurately recommend topics to users, and reduce the manpower and material resources required in the process of topic recommendation.

In addition, the topic recommendation method based on the deep interest network proposed according to the above embodiments of the present invention may also have the following additional technical features:

Optionally, the user's historical click data includes item information and time information corresponding to each historical click behavior of the user, and ranking information among various historical click behaviors.

Optionally, the training data includes discrete features, continuous features, and sequence features; wherein, the discrete features include time information, user attribute information, and item classification information, and the continuous features include user historically clicked item classifications. Statistical information, the sequence feature includes the item information sequence corresponding to the user's historical click behavior.

Optionally, the training data further includes sample time characteristics, wherein generating training data according to the user information and the historical click data includes: generating a training sample according to the user information and the historical click data, and calculating The time difference between the training sample and the current time, and determining whether the time difference is greater than a preset time threshold, so as to use the judgment result as a time characteristic of the sample.

Optionally, generating training data according to the user information and the historical click data includes: counting the number of clicks corresponding to each item, and determining the probability of selecting a negative sample corresponding to each item according to the statistical result; The negative sample selection probability corresponding to the item is used for random selection of negative samples.

Optionally, determining the topic recommendation list according to the retrieval result includes: clustering topics according to the kmeas clustering algorithm to generate multiple topic categories; generating a topic list to be recommended according to the retrieval results, and according to the multiple topic categories and The sliding window breaking method is used to break up the list of topics to be recommended, so as to generate a final recommendation list of topics.

In order to achieve the above object, the embodiment of the second aspect of the present invention provides a computer-readable storage medium, which stores a topic recommendation program based on a deep interest network, and the topic recommendation program based on a deep interest network is implemented when the processor is executed. As mentioned above, the topic recommendation method based on deep interest network.

According to the computer-readable storage medium of the embodiment of the present invention, by storing the topic recommendation program based on the deep interest network, the processor implements the above-mentioned topic recommendation program based on the deep interest network when executing the topic recommendation program based on the deep interest network The recommendation method can be used to accurately recommend topics to users without establishing labels corresponding to topics, thereby reducing the manpower and material resources required in the process of topic recommendation.

In order to achieve the above object, a third aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, To achieve the above-mentioned topic recommendation method based on deep interest network.

According to the computer device of the embodiment of the present invention, the deep interest network-based topic recommendation program is stored in the memory, so that when the processor executes the deep interest network-based topic recommendation program, the above-mentioned deep interest network-based topic recommendation program is implemented. The recommendation method can be used to accurately recommend topics to users without establishing labels corresponding to topics, thereby reducing the manpower and material resources required in the process of topic recommendation.

In order to achieve the above purpose, a fourth aspect of the present invention provides a topic recommendation device based on a deep interest network, including: an acquisition module, the acquisition module is used to acquire user information and user historical click data, and according to the User information and the historical click data to generate training data; a training module, which is used for model training according to the training data to obtain a deep interest capture model; an interest capture module, which is used to acquire items corresponding item information, and input the item information into the deep interest capture model, so as to output the corresponding item vector through the deep interest capture model, and calculate the topic vector according to the item vector corresponding to each item; the interest The capture module is also used to obtain the click data to be analyzed of the user, and input the click data to be analyzed into the deep interest capture model, so as to output the corresponding user vector through the deep interest capture model; the recommendation module, the recommending The module is configured to perform similarity retrieval according to the user vector and the topic vector, determine a topic recommendation list according to the retrieval result, and push the topic recommendation list to the user.

According to the special topic recommendation device based on the deep interest network according to the embodiment of the present invention, the acquisition module is configured to acquire user information and historical click data of the user, and generate training data according to the user information and the historical click data; the training module uses Carry out model training according to the training data to obtain a deep interest capture model; the interest capture module is used to obtain the item information corresponding to the item, and input the item information into the deep interest capture model, so as to obtain the deep interest capture model through the deep interest capture module. The capture model outputs the corresponding item vector, and calculates the topic vector according to the item vector corresponding to each item; the interest capture module is also used to obtain the user's click data to be analyzed, and input the to-be-analyzed click data into the deep interest capture model, to output the corresponding user vector through the deep interest capture model; the recommendation module is used to perform similarity retrieval according to the user vector and the topic vector, and determine the topic recommendation list according to the retrieval result, and recommend the topic The list is pushed to the user; thus, it is possible to accurately recommend the topic to the user without establishing a label corresponding to the topic, and reduce the manpower and material resources required in the process of the topic recommendation.

In addition, the topic recommendation device based on the deep interest network proposed according to the above embodiments of the present invention may also have the following additional technical features:

Description of drawings

1 is a schematic flowchart of a topic recommendation method based on a deep interest network according to an embodiment of the present invention;

2 is a schematic structural diagram of a deep interest capture model according to an embodiment of the present invention;

FIG. 3 is a schematic block diagram of a topic recommendation apparatus based on a deep interest network according to an embodiment of the present invention.

Detailed ways

The following describes in detail the embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals refer to the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary, and are intended to explain the present invention and should not be construed as limiting the present invention.

In the related art, when a topic is recommended, there is a strong dependence on the tags corresponding to the topic. In order to improve the accuracy of the topic recommendation, it is necessary to spend a lot of manpower and material resources to establish high-quality tags; The special recommendation method of the network, first, obtain user information and user historical click data, and generate training data according to the user information and the historical click data; then, carry out model training according to the training data to obtain deep interest capture Then, obtain the item information corresponding to the item, and input the item information into the deep interest capture model, so as to output the corresponding item vector through the deep interest capture model, and calculate according to the item vector corresponding to each item thematic vector; then, obtain the click data to be analyzed of the user, and input the click data to be analyzed into the deep interest capture model, so as to output the corresponding user vector through the deep interest capture model; then, according to the user The similarity search is performed between the vector and the topic vector, and the topic recommendation list is determined according to the retrieval result, and the topic recommendation list is pushed to the user; thus, it is possible to accurately carry out the search for the user without establishing the label corresponding to the topic. Thematic recommendation reduces the manpower and material resources required in the process of thematic recommendation.

For better understanding of the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention may be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present invention will be more thoroughly understood, and will fully convey the scope of the present invention to those skilled in the art.

In order to better understand the above technical solutions, the above technical solutions will be described in detail below with reference to the accompanying drawings and specific embodiments.

1 is a schematic flowchart of a topic recommendation method based on a deep interest network according to an embodiment of the present invention. As shown in FIG. 1 , the topic recommendation method based on a deep interest network includes the following steps:

S101: Obtain user information and historical click data of the user, and generate training data according to the user information and historical click data.

Among them, there may be various methods for selecting the historical click data of the user.

As an example, the user's historical click data includes the user's exposure log and click behavior log, wherein the exposure log records whether an item is exposed to a user on that day, and the click behavior log records the corresponding click behavior of the user. information.

As another example, the user's historical click data includes item information corresponding to each historical click behavior of the user, time information, and ranking information among various historical click behaviors. In other words, the user's historical click data only contains the information corresponding to the user's click behavior, but not the exposure information; it should be noted that, in the actual scene, because different pages have different depths, pages of different depths are The CTR varies greatly, and pages with deeper access depths tend to have higher CTRs; therefore, in order to avoid the impact of differences in CTRs caused by pages with different depths, the historical click data does not include exposure information.

Among them, there are various ways of setting the training data.

As an example, the training data includes discrete features, continuous features, and sequence features; wherein the discrete features include time information, user attribute information, and item classification information, the continuous features include historical user clicked item classification statistics, and sequence features Including the item information sequence corresponding to the user's historical click behavior. Specifically, the discrete features include date information (for example, the day of the week, whether it is a working day, whether it is a working period, etc.), item ID, item classification ID, and item attribute ID, etc. It should be noted that, according to the discrete feature classification, more If there are few features, it can be encoded by ONE-HOT; continuous features include the number of clicks on different attributes in the user's history, such features can be directly input continuous values without processing; sequence features include the user's historical click item ID Sequence, historical click category ID sequence.

In some embodiments, the training data further includes sample time features, wherein generating the training data according to the user information and the historical click data includes: generating a training sample according to the user information and the historical click data, and calculating the difference between the training sample and the current time. time difference, and judging whether the time difference is greater than a preset time threshold, so as to use the judgment result as a sample time characteristic.

It can be understood that since the training data contains time information, that is, the contextual features include the day of the week, whether it is a working day, etc., in the training process of the model, if the training samples are fully scattered in time, the training process will also be unstable. , due to the influence of time factors, the sample will have a great impact on the model. The closer the sample is to the test date, the greater the effect it plays; therefore, the sample time feature is added to ensure the stability of the model training process. .

In some embodiments, training data is generated based on user information and historical click data, including:

The number of clicks corresponding to each item is counted, and the negative sample selection probability corresponding to each item is determined according to the statistical result, and the negative sample is randomly selected according to the negative sample selection probability corresponding to each item. It can be understood that in the training process, it is necessary to select negative samples to smoothly train the model. Among them, there are many ways to select negative samples. For example, directly selecting a preset number of negative samples from the positive samples in a random manner; preferably, the number of clicks corresponding to each item can be counted in the above-mentioned manner to determine the probability that the item is selected as a negative sample; Therefore, the more popular items have a higher probability of being selected as negative samples, which makes the final training model more accurate.

In some embodiments, in order to avoid the problem that the calculation amount of the softmax function is too large, the sequence feature is set to the form of binary classification; that is, when the input item ID is the item ID that the user clicks next, the label is 1, 0 otherwise. Specifically, assuming that the sequence of items clicked by the user is [1, 2, 3, 4, 5, 6], the sequence feature structure is shown in Table 1:

Table 1

S102, perform model training according to the training data to obtain a deep interest capturing model.

For ease of understanding, take FIG. 2 as an example, which is a schematic structural diagram of a deep interest capture model according to an embodiment of the present invention; as shown in FIG. 2 , in this embodiment, sequence features, discrete features, and continuous features are After splicing, it passes through the BatchNormalization layer, and then input to the multi-layer fully connected layer. After each layer of the fully connected layer, it will be connected to the BatchNormalization layer and the Dice activation function, and finally the user vector is obtained.

In some embodiments, during model training, the Adagrad optimizer is used, the initial learning rate is 0.1, the learning rate decays to 1/2 of the original value every 50,000 steps, and the Batch size is 128. And in order to make the model training more stable, the L2 regularization parameter will be added to the Embedding layer and the DNN layer, and the regular loss will be added to the loss function for optimization.

S103: Obtain item information corresponding to the item, and input the item information into the deep interest capture model, so as to output the corresponding item vector through the deep interest capture model, and calculate the thematic vector according to the item vector corresponding to each item.

It can be understood that each topic will contain a different number of items, and the topic is a collection of items of the same category; for example, if the topic is sports, the items corresponding to the topic may include: football, basketball, swimming, etc. Among them, there can be various ways to calculate the topic vector according to the item vector corresponding to each item; for example, after obtaining the item vector, average pooling is performed on the item vectors of all items under the topic, so as to use the pooling result as the topic thematic vector.

S104: Acquire click data of the user to be analyzed, and input the click data to be analyzed into a deep interest capture model, so as to output a corresponding user vector through the deep interest capture model.

S105: Perform similarity retrieval according to the user vector and the topic vector, determine the topic recommendation list according to the retrieval result, and push the topic recommendation list to the user.

In some embodiments, determining the topic recommendation list according to the retrieval result includes: clustering topics according to the kmeas clustering algorithm to generate multiple topic categories; generating a topic list to be recommended according to the retrieval result, and The sliding window breaking method is used to break up the list of recommended topics to generate the final recommended topic list.

It can be understood that after generating a list of topics to be recommended according to the search results, there may be multiple topics of the same category under the same window in the list of topics to be recommended, which will bring a bad experience to users; therefore, in order to protect users Experience, through the sliding window breaking method and the clustering results of the topic, the recommended topic list is broken up, so that the categories of topics under the same window are different to determine the final topic recommendation list.

Specifically, as shown in Table 2:

Table 2

As shown in Table 2, if the topic sequence obtained by user 001 is 1|2|3|4|5|6|7|8|9, the category sequence of the topic is A|A|A|B|C|B|B |D|D. Assuming that the size of the sliding window is 3, it means that the thematic categories placed in three adjacent positions do not overlap. but:

In the first step, the categories in the first sliding window are A, A, A, and the position index starts from 0. You need to break up the lists in the first and second positions, and traverse from the third position back. , the first different category is B, then the A in the first position and the B in the third position are exchanged, and the thematic category sequence becomes A|B|A|A|C|B|B|D|D , and then swap the 1st and 3rd positions of the topic id sequence.

In the second step, the categories in the first sliding window become A, B, A, then the second position needs to be processed, starting from the fourth position and traversing backwards, the first different category is the fourth position. C, so if position 2 and position 4 are exchanged, the thematic category id sequence becomes A|B|C|A|A|B|B|D|D, and the thematic Id sequence is 1|4|5|2| 3|6|7|8|9.

In the third step, the category sequence in the second sliding window is B, C, A, then no processing is required, the window continues to slide forward, and the third window is C, A, A, then A in the fourth position needs to be Processing, exchange with position 5, the thematic category ID sequence becomes A|B|C|A|B|A|B|D|D, and the theme id sequence becomes 1|4|5|2|6|3 |7|8|9, and so on, until the sequence is complete or the fragmented length reaches the threshold.

To sum up, according to the topic recommendation method based on the deep interest network according to the embodiment of the present invention, first, user information and historical click data of the user are obtained, and training data is generated according to the user information and the historical click data; then, Perform model training according to the training data to obtain a deep interest capture model; then, obtain item information corresponding to the item, and input the item information into the deep interest capture model, so as to output the corresponding deep interest capture model through the deep interest capture model and calculate the thematic vector according to the item vector corresponding to each item; then, obtain the click data to be analyzed of the user, and input the click data to be analyzed into the deep interest capture model, so as to pass the deep interest The user vector corresponding to the capture model output; then, similarity retrieval is performed according to the user vector and the topic vector, and a topic recommendation list is determined according to the retrieval result, and the topic recommendation list is pushed to the user; Under the premise of establishing labels corresponding to topics, users can accurately recommend topics to reduce the manpower and material resources required in the process of topic recommendation.

In order to realize the above-mentioned embodiments, the embodiments of the present invention provide a computer-readable storage medium on which a topic recommendation program based on a deep interest network is stored, and when the topic recommendation program based on a deep interest network is executed by a processor, the above-mentioned The topic recommendation method based on deep interest network.

In order to implement the above embodiments, the embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, the processor implements the following The above-mentioned topic recommendation method based on deep interest network.

In order to realize the above embodiment, the embodiment of the present invention proposes a topic recommendation device based on a deep interest network. As shown in FIG. 3 , the topic recommendation device based on a deep interest network includes: an acquisition module 10, a training module 20, an interest capture module module 30 and recommendation module 40 .

Wherein, the acquisition module 10 is used to acquire user information and historical click data of the user, and generate training data according to the user information and historical click data;

The training module 20 is used for model training according to the training data to obtain a deep interest capture model;

The interest capture module 30 is used to obtain the article information corresponding to the article, and the article information is input into the deep interest capture model, to output the corresponding article vector by the deep interest capture model, and calculate the thematic vector according to the article vector corresponding to each article;

The interest capture module 30 is further configured to obtain the click data to be analyzed of the user, and input the click data to be analyzed into the deep interest capture model, so as to output the corresponding user vector through the deep interest capture model;

The recommendation module 40 is configured to perform similarity retrieval according to the user vector and the topic vector, determine the topic recommendation list according to the retrieval result, and push the topic recommendation list to the user.

In some embodiments, the user's historical click data includes item information corresponding to each historical click behavior of the user, time information, and ranking information among various historical click behaviors.

It should be noted that, the above description about the topic recommendation method based on the deep interest network in FIG. 1 is also applicable to the topic recommendation apparatus based on the deep interest network, and will not be repeated here.

To sum up, according to the special topic recommendation device based on the deep interest network according to the embodiment of the present invention, an acquisition module is set to acquire user information and historical click data of users, and to generate training according to the user information and the historical click data The training module is used to perform model training according to the training data to obtain a deep interest capture model; the interest capture module is used to obtain the item information corresponding to the item, and input the item information into the deep interest capture model to obtain a deep interest capture model. The corresponding item vector is output through the deep interest capture model, and the thematic vector is calculated according to the item vector corresponding to each item; the interest capture module is also used to obtain the user's click data to be analyzed, and input the to-be-analyzed click data into the The deep interest capture model is used to output the corresponding user vector through the deep interest capture model; the recommendation module is configured to perform similarity retrieval according to the user vector and the topic vector, and determine a topic recommendation list according to the retrieval result, and The topic recommendation list is pushed to the user; thus, the topic recommendation can be accurately performed to the user without establishing a label corresponding to the topic, and the manpower and material resources required in the topic recommendation process are reduced.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

It should be noted that, in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not preclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, and third, etc. do not denote any order. These words can be interpreted as names.

Although preferred embodiments of the present invention have been described, additional changes and modifications to these embodiments may occur to those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be construed to include the preferred embodiment and all changes and modifications that fall within the scope of the present invention.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit and scope of the invention. Thus, provided that these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

In the description of the present invention, it should be understood that the terms "first" and "second" are only used for description purposes, and cannot be interpreted as indicating or implying relative importance or the number of indicated technical features. Thus, a feature defined as "first" or "second" may expressly or implicitly include one or more of that feature. In the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

In the present invention, unless otherwise expressly specified and limited, the terms "installed", "connected", "connected", "fixed" and other terms should be understood in a broad sense, for example, it may be a fixed connection or a detachable connection , or integrated; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the internal connection of the two elements or the interaction relationship between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

In the present invention, unless otherwise expressly specified and limited, a first feature "on" or "under" a second feature may be in direct contact between the first and second features, or the first and second features indirectly through an intermediary touch. Also, the first feature being "above", "over" and "above" the second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is level higher than the second feature. A first feature "below", "below" and "below" a second feature may mean that the first feature is directly below or diagonally below the second feature, or simply means that the first feature has a lower level than the second feature.

In the description of this specification, description with reference to the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples", etc., mean specific features described in connection with the embodiment or example , structure, material or feature is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the particular features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, those skilled in the art may combine and combine the different embodiments or examples described in this specification, as well as the features of the different embodiments or examples, without conflicting each other.

Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Embodiments are subject to variations, modifications, substitutions and variations.

Claims

A topic recommendation method based on deep interest network, characterized in that it includes the following steps:

Obtain user information and historical click data of the user, and generate training data according to the user information and the historical click data;

Perform model training according to the training data to obtain a deep interest capture model;

Obtain the item information corresponding to the item, and input the item information into the deep interest capture model, so as to output the corresponding item vector through the deep interest capture model, and calculate the thematic vector according to the item vector corresponding to each item;

Acquiring the click data to be analyzed of the user, and inputting the click data to be analyzed into the deep interest capture model, so as to output the corresponding user vector through the deep interest capture model;

Similarity retrieval is performed according to the user vector and the topic vector, a topic recommendation list is determined according to the retrieval result, and the topic recommendation list is pushed to the user.
The topic recommendation method based on the deep interest network according to claim 1, wherein the user's historical click data includes item information corresponding to each historical click behavior of the user, time information and the ranking among the historical click behaviors information.
The topic recommendation method based on a deep interest network according to claim 1, wherein the training data includes discrete features, continuous features and sequence features;

Wherein, the discrete features include time information, user attribute information and item classification information, the continuous features include historical user clicked item classification statistics, and the sequence features include an item information sequence corresponding to the user's historical click behavior.
The thematic recommendation method based on the deep interest network according to claim 3, wherein the training data further includes sample time characteristics, wherein generating the training data according to the user information and the historical click data includes:

Generate training samples according to the user information and the historical click data, calculate the time difference between the training samples and the current time, and judge whether the time difference is greater than a preset time threshold, so as to use the judgment result as Sample time characteristics.
The thematic recommendation method based on a deep interest network according to any one of claims 1-4, wherein generating training data according to the user information and the historical click data, comprising:

The number of clicks corresponding to each item is counted, and the negative sample selection probability corresponding to each item is determined according to the statistical result, and the negative sample is randomly selected according to the negative sample selection probability corresponding to each item.
The topic recommendation method based on the deep interest network according to any one of claims 1-4, wherein determining a topic recommendation list according to a retrieval result, comprising:

Cluster the topics according to the kmeas clustering algorithm to generate multiple topic categories;

A list of topics to be recommended is generated according to the retrieval result, and the list of topics to be recommended is broken up according to the plurality of topic categories and the sliding window breaking method, so as to generate a final recommendation list of topics.
A computer-readable storage medium, characterized in that a topic recommendation program based on a deep interest network is stored thereon, and when the topic recommendation program based on a deep interest network is executed by a processor, any one of claims 1-6 is implemented The described topic recommendation method based on deep interest network.
A computer device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, when the processor executes the program, any one of claims 1-6 is implemented. The topic recommendation method based on deep interest network described in item.
A special topic recommendation device based on a deep interest network, characterized in that it includes:

an acquisition module, which is used to acquire user information and historical click data of the user, and generate training data according to the user information and the historical click data;

a training module, which is used for model training according to the training data to obtain a deep interest capture model;

An interest capture module, the interest capture module is used to obtain the item information corresponding to the item, and input the item information into the deep interest capture model, so as to output the corresponding item vector through the deep interest capture model, and according to each The item vector corresponding to each item calculates the thematic vector;

The interest capture module is further configured to acquire click data to be analyzed of the user, and input the click data to be analyzed into the deep interest capture model, so as to output the corresponding user vector through the deep interest capture model;

A recommendation module, which is configured to perform similarity retrieval according to the user vector and the topic vector, determine a topic recommendation list according to the retrieval result, and push the topic recommendation list to the user.
The topic recommendation device based on the deep interest network according to claim 9, wherein the user's historical click data includes item information, time information and ranking information between each historical click behavior corresponding to each historical click behavior of the user .