WO2019223552A1 - 文章推荐方法、装置、计算机设备及存储介质 - Google Patents

文章推荐方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2019223552A1
WO2019223552A1 PCT/CN2019/086374 CN2019086374W WO2019223552A1 WO 2019223552 A1 WO2019223552 A1 WO 2019223552A1 CN 2019086374 W CN2019086374 W CN 2019086374W WO 2019223552 A1 WO2019223552 A1 WO 2019223552A1
Authority
WO
WIPO (PCT)
Prior art keywords
article
historical
articles
attention
recommended
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/086374
Other languages
English (en)
French (fr)
Inventor
刘毅
胡澜涛
张博
夏锋
林乐宇
冯喆
陈磊
饶君
刘书凯
丘志杰
孙振龙
王良栋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Publication of WO2019223552A1 publication Critical patent/WO2019223552A1/zh
Priority to US16/945,066 priority Critical patent/US11763145B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/16—Devices for psychotechnics; Testing reaction times ; Devices for evaluating the psychological state
    • A61B5/165—Evaluating the state of mind, e.g. depression, anxiety
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/335—Filtering based on additional data, e.g. user or group profiles
    • G06F16/337—Profile generation, learning or modification
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20—Movements or behaviour, e.g. gesture recognition

Definitions

  • the present application relates to the field of data processing technology, and in particular, to an article recommendation method, device, computer device, and storage medium.
  • the article publishing platform will recommend articles to users based on historical articles that the user has read. However, when recommending articles to users, they recommend similar articles to users based on their reading history.
  • the content-based recommendation technology is to read articles based on user history, determine articles similar to the content of articles that the user has read, and recommend these similar articles to the user.
  • this application provides an article recommendation method, device, computer equipment, and storage medium to improve the diversity of article recommendation.
  • this application provides a method for article recommendation, including:
  • the computer device obtains a history reading set of the target user to be analyzed, and the history reading set includes: multiple historical articles that the target user has read at different reading moments;
  • the computer device determines, for each historical article in the plurality of historical articles, a first attention degree of each pending historical article and each of the historical articles at the time of reading before each of the historical articles. The degree is reflected in the possibility that the user is recommended to read each of the above historical articles when the user reads the above-mentioned historical articles;
  • the computer device selects at least one recommended reference article from the plurality of pending historical articles according to the first attention level of each of the pending historical articles and at least one historical article;
  • the computer device determines at least one candidate recommended article to be recommended to the target user from the set of recommendable articles according to each of the recommended reference articles.
  • this application also provides an article recommendation device, which is applied to computer equipment and includes:
  • the history obtaining unit is configured to obtain a history reading set of a target user to be analyzed, and the history reading set includes: multiple historical articles read by the target user at different reading times;
  • the first attention analysis unit is configured to determine, for each historical article in the foregoing multiple historical articles, the degree of first attention of each pending historical article before each of the historical articles and the foregoing historical article, The above-mentioned first degree of attention is reflected in the possibility that the user is recommended to read each of the above-mentioned historical articles when the user reads the above-mentioned pending historical articles;
  • the recommended reference determination unit is configured to select at least one recommended reference article from a plurality of the above-mentioned pending historical articles according to the first attention level of each of the above-mentioned pending historical articles and at least one historical article;
  • the candidate recommendation determining unit is configured to determine at least one candidate recommended article to be recommended to the target user from a set of recommendable articles according to each of the above recommended reference articles.
  • the present application also provides a computer device, including:
  • the processor is configured to execute a computer program stored in the memory
  • the memory is used to store the computer program, and the computer program is used at least:
  • the above historical reading collection includes:
  • each historical article in the above multiple historical articles determine the first attention level of each pending historical article and each of the above historical articles at the time of reading before each of the above historical articles.
  • the first attention degree is reflected in In the case where the user reads the above-mentioned pending historical articles, it is recommended that the user read each of the above-mentioned historical articles;
  • At least one candidate recommended article to be recommended to the target user is determined from the set of recommendable articles.
  • the present application also provides a storage medium.
  • a computer program is stored in the storage medium, and when the computer program is executed, the foregoing method for recommending the foregoing article is executed.
  • FIG. 1 is a schematic diagram of a system composition architecture to which an article recommendation method according to an embodiment of the present application is applied;
  • FIG. 1 is a schematic diagram of a system composition architecture to which an article recommendation method according to an embodiment of the present application is applied;
  • FIG. 2 is a schematic diagram showing a composition architecture of a computer device to which the article recommendation method according to the embodiment of the present application is applied;
  • FIG. 3 is a schematic flowchart of an article recommendation method according to an embodiment of the present application.
  • FIG. 4 is a schematic flowchart of training an attention model in an article recommendation method according to an embodiment of the present application.
  • FIG. 5 shows a schematic diagram of an article vector based on an LSTM output article in an embodiment of the present application
  • FIG. 6 shows a schematic diagram of an implementation principle for determining a recommended reference article in an embodiment of the present application
  • FIG. 7 is a schematic flowchart of another embodiment of an article recommendation method according to an embodiment of the present application.
  • FIG. 8 is a schematic diagram of an implementation principle of an article recommendation method according to an embodiment of the present application.
  • FIG. 9 is a schematic flowchart of another article recommendation method according to an embodiment of the present application.
  • FIG. 10 shows an optional interface display diagram of this embodiment
  • FIG. 11 shows another optional interface display diagram of this embodiment
  • FIG. 12 shows yet another optional interface display diagram of this embodiment
  • FIG. 13 is a schematic diagram of a composition architecture of an article recommendation device according to an embodiment of the present application.
  • FIG. 14 is a schematic diagram of another composition architecture of the article recommendation device according to the embodiment of the present application.
  • the article recommendation method in the embodiment of the present application is applicable to various article publishing platforms, to determine articles that different users need to recommend to the article publishing platform, and to increase the variety of articles recommended to users.
  • news media platforms recommend news articles to different users
  • social platforms or public accounts accessed by social platforms recommend articles to users.
  • FIG. 1 illustrates a schematic diagram of a composition architecture of an article recommendation system of the present application.
  • the article recommendation system may include: a computer device 101 and at least one article publishing server 102.
  • the article publishing server 102 is configured to publish multiple articles that can be read by users, and records articles that different users have read at different times. To make it easier to distinguish, articles that users have read are called historical articles.
  • the computer device 101 is configured to determine a recommended article to be recommended to the user based on a plurality of historical articles read by each user.
  • the computer device 101 may also send the determined information of the recommended article to be recommended to the user to the article publishing server.
  • the article recommendation system may further include a data storage server 103, and the data storage server may store user data generated by at least one article publishing server, such as a user portrait, a user's reading history, and the like.
  • the user portrait may include user attribute information such as the user ’s gender, age, education, and region.
  • the reading history may include information such as articles the user has read and sources related to the article.
  • the computer device 101 may obtain the reading history of the user from the data storage server 103 to determine the recommended article to be recommended to the user according to the reading history of the user.
  • the historical articles read by the user on which the computer device 101 is based may be derived from one or more article publishing servers in a certain article publishing platform, or may be articles published from multiple article publishing platforms. server.
  • the computer device 101 may be a server or terminal that is separately set up from the article publishing server; it may also be the same server as the article publishing server, that is, the computer device has article publishing And determine the ability to recommend articles to users.
  • FIG. 2 shows a composition diagram of the computer device of the present application.
  • the computer device 200 may include a processor 201, a memory 202, a communication interface 203, an input unit 204, a display 205, and a communication bus 206.
  • the processor 201 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), or an application-specific integrated circuit. (ASIC), off-the-shelf programmable gate array (FPGA), or other programmable logic devices.
  • CPU central processing unit
  • ASIC application-specific integrated circuit
  • DSP digital signal processor
  • ASIC application-specific integrated circuit
  • FPGA off-the-shelf programmable gate array
  • the processor may call a program stored in the memory 202, and the processor may perform operations performed on the server side in FIGS. 3 to 9 below.
  • the memory 202 is used to store one or more programs, and the programs may include program codes.
  • the program codes include computer operation instructions.
  • the memory stores at least programs for implementing the following functions:
  • the historical reading collection includes: multiple historical articles that the target user has read at different reading moments;
  • the first of each pending historical article and historical article is determined separately. Attention level, the first attention level is reflected in the possibility that users are recommended to read historical articles in the case of users reading pending historical articles;
  • At least one candidate recommended article to be recommended to the target user is determined from the set of recommendable articles.
  • the memory 202 may include a storage program area and a storage data area, where the storage program area may store an operating system, the above-mentioned programs, and applications required for at least one function (such as a sound playback function, an image playback function, and the like). Programs, etc .;
  • the storage data area can store data created during the use of computer equipment and received data to be processed, such as user reading records.
  • the memory 202 may include a high-speed random access memory, and may further include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
  • a non-volatile memory such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
  • the communication interface 203 may be an interface of a communication module.
  • the server may further include an input unit 205, such as a keyboard and the like.
  • the display 204 includes a display panel.
  • the display panel may be configured by using a liquid crystal display (Liquid Crystal Display, LCD), an organic light emitting diode (Organic Light-Emitting Diode, OLED), or the like.
  • LCD Liquid Crystal Display
  • OLED Organic Light-Emitting Diode
  • the structure of the computer equipment shown in FIG. 2 does not constitute a limitation on the computer equipment in the embodiment of the present application.
  • the computer equipment may include more or fewer components than those shown in FIG. 2, or some components may be combined. .
  • FIG. 3 it is a schematic flowchart of an article recommendation method according to an embodiment of the present application. This embodiment is introduced from a computer device side. The process includes:
  • the historical reading collection includes: multiple historical articles read by the target user at different reading moments.
  • the historical reading set includes multiple historical articles sorted in order of reading.
  • the order of reading times corresponding to the multiple historical articles can be determined.
  • the historical reading set may include multiple historical articles read by multiple target users and the reading moments corresponding to each historical article.
  • the computer device For each historical article, according to the content characteristics of the historical article, the computer device reads at least one pending historical article from the historical reading collection before the reading time of the historical article.
  • the historical article collection includes Article A, Article B, and Article C, and the order of these three articles indicates the reading order of these three articles by the user. From this, we can see that Article A and Article B are reading articles. Article C was read before, then article C may be because the user chose to read article C after reading one or two of article A and article B. Therefore, it is necessary to determine article A and article read before article C. B, and analyze the relationship between article A, article B and article C, respectively, to obtain the possibility that article A will trigger the user to read article C, and the possibility that article B will trigger the user to read article C.
  • each pending historical article that has been read before the historical article may be regarded as a pending historical article, and subsequent operations of analyzing the relationship between the pending historical article and the historical article may be performed.
  • a set number of historical articles read before the historical article can be taken as pending History articles. For example, from other historical articles read before the historical article, randomly select a set number of historical articles, use the selected set historical articles as pending historical articles, and perform subsequent operations.
  • the reading time can be before the reading time of the historical article, and the reading time and Among multiple historical articles whose duration between reading moments of the historical article does not exceed a preset duration, randomly select a preset number of historical articles as pending historical articles or select multiple historical articles closest to the reading time of the historical article as Pending history articles.
  • the number of pending historical articles corresponding to historical articles may be less than the preset number.
  • the computer device For each historical article, based on the content characteristics of the historical article and the content characteristics of at least one pending historical article corresponding to the historical article, the computer device separately determines the first attention level of each pending historical article and the historical article. .
  • the degree of attention can also be called the degree of focus, which reflects the possibility of recommending users to read another article when reading one article. At the same time, the degree of attention can also reflect the possibility of choosing to read another article by clicking on it in the case of one article. Because of the above meaning of attention, it is also called the degree of attention or focus from one article to another.
  • the degree of attention may be a score, a probability, a level, and the like, which is not limited herein.
  • the degree of attention between the pending historical article and the historical article is called the first degree of attention. Accordingly, the first degree of attention is used to reflect the possibility that the user is recommended to read the historical article when the user reads the pending historical article.
  • the computer device needs to determine the first attention level of each pending historical article and the historical article, respectively.
  • the content characteristics of the article are determined by the title of the article and the words in the body of the article.
  • the content characteristics of the article can reflect the type of article, the main idea of the article, and the specific content involved in the article. and many more. Therefore, for any two articles, you can analyze the degree of attention from one article to the other according to the content characteristics of the two articles.
  • the computer device can analyze the similarity of the content characteristics of the two articles and determine the similarity based on the similarity of the content characteristics.
  • the level of attention between the two articles For example, consider the similarity of the content characteristics of the two articles as the degree of attention between the two articles.
  • the computer equipment determines the similarity between the pending historical article and the historical article and determines the degree of attention between the pending historical article and the historical article according to the content characteristics of the pending historical article and the historical article.
  • the normalized result is a value greater than or equal to zero and less than 1, and the value can be reflected from the Article M triggers the possibility of clicking article N.
  • the computer device may be pre-trained to obtain an attention model for determining the degree of attention.
  • the attention model is obtained by training a set of historical article samples of multiple users and a sequence of labeled attention scores corresponding to each historical article sample in each historical article sample set.
  • the historical articles that train the attention model are called historical article samples.
  • the sequence of attention scores corresponding to the historical article sample of the user includes: a plurality of historical article samples that the user read before reading the historical article sample and the attention score of the historical article sample, respectively.
  • the attention score For example, by obtaining the user's historical reading log from the server of the article publishing platform, when users read different articles, they can select other articles to read by clicking, etc., so as to count the probability of triggering reading of other articles from each article, and The corresponding probability is used as the attention score.
  • the computer device may use the content characteristics of the historical article and the content characteristics of at least one pending historical article before the reading time of the historical article in the historical reading set, and use pre-training to obtain The attention model determines the attention score of each pending historical article and the historical article separately.
  • the content characteristics of the article can be represented by the article vector of the article.
  • the content characteristics of each article eg, article M and article N
  • the article vector of each article M and article N is input into the attention model to output the attention score of each article M to article N, respectively.
  • the attention degree (such as the attention score) between each pending historical article and the historical article output by the attention model is the same. Can be the normalized result.
  • the attention model is taken as an example to be obtained by training a deep neural network. For example, see FIG. 4, which illustrates a schematic flowchart of training an attention model.
  • the computer device obtains a sample set of historical articles of multiple users and a sequence of labeled attention scores corresponding to each historical article sample in each historical article sample set.
  • the labeled attention score sequence corresponding to article S4 may include: the article S4 and Attention scores for article S3, attention scores for article S4 and article S2, and attention scores for article S4 and article S1.
  • the corresponding labeled attention score sequences may include: article S3 and article S2 attention scores, article S3 and article S1 attention scores.
  • the attention score sequence may not be set.
  • each attention score in the labeled attention score sequence corresponding to the historical article sample is a normalized result according to each attention score in the attention score sequence.
  • step S402 the computer equipment inputs each historical article sample into the deep neural network for cyclic training, and performs cyclic training on the deep neural network model according to the respective attention score sequences of each historical article sample until the deep neural network The accuracy between the actual attention score sequence corresponding to each historical article output by the network and the corresponding labeled attention score sequence meets preset requirements.
  • the labeled attention score sequence corresponding to the historical article sample S m read by the user at time N is taken as an example for illustration, and the labeled attention score sequence corresponding to the historical article sample S m is Attention scores between the historical article sample S m and multiple historical article samples in the historical article sample sequence S, where the historical article sample sequence S is the multiple historical articles in the historical sample set G that were read before time N A sequence of samples.
  • the computer equipment inputs the historical article sample S N and its corresponding labeled attention score sequence A m * into the deep neural network model, and the deep neural network model can output the actual attention corresponding to the historical article sample S m
  • the force score sequence A m , the actual attention score sequence A m includes the attention score of the historical article sample S m output from the deep neural network model and each historical article sample in the historical article sample sequence S, respectively.
  • the gap between the labeled attention score sequence A m * and the actual attention score sequence A m it can be analyzed whether the accuracy of the deep neural network model meets the requirements. It can be optimized based on a preset loss function and a sequence of labeled attention scores and actual attention scores corresponding to multiple historical article samples, and it can be optimized by gradient descent until iterative convergence. For example, you can define the following loss function (ie, formula one):
  • a m * and A m respectively represent the marked attention scores corresponding to different historical article samples, and the actual attention score, and n is the total number of historical article samples used for training.
  • the computer equipment can use the gradient descent method to optimize the optimization goal until the final iteration converges, and the training ends.
  • the trained deep neural network model can be used as an attention model.
  • the trained attention model also calculates the similarity between articles based on the content characteristics of different articles, and determines the degree of article attention based on the similarity.
  • the article to a sequence determined (set) D ⁇ d 1, d 2, whil d i ?? d t-2, d t-1 ⁇ according to any one article respectively d i d t of the article
  • the degree of attention will be described as an example.
  • d i represents the sequence of the article to any article D
  • i is a natural number from 1 to t-1 is
  • t-1 is the total number of articles in article D sequence.
  • the functional relationship between the article d i and article d t 's attention score in the article sequence D in the attention model can be expressed as follows Formula two:
  • Article d i is the vector of the article
  • Article vector for article d t is the vector of the article
  • W, U, and b are parameters set in the attention model. The parameter values of these parameters are determined during the training process.
  • F represents a set functional relationship, which can also be determined during training.
  • the article is ascertained E i d i d and Article similarity between articles vector t, and is normalized by the following formula 2, the article can be obtained fraction attention d i d T of the article.
  • the computer device may segment the content in the article, and input a plurality of words segmented from the article into a pre-trained article vector model to output the article vector of the article through the article vector model.
  • the article vector model is taken as an example for long short-term memory network (LSTM).
  • LSTM long short-term memory network
  • the content of the article is segmented.
  • the N segmentation words are sequentially input into the LSTM network model, so that the article vector of the article d can be finally output through the LSTM network model.
  • FIG. 5 illustrates a schematic diagram of an LSTM network model that converts an article's LSTM to a plurality of word segmentations of the article into an article vector of the article.
  • the input in the LSTM network model is each word segmentation of the article segmentation, and the final output vector of the article is h t .
  • the target user reads a historical article, it is less likely to trigger the recommendation of another article to the target user or trigger the target user to click to read other articles.
  • the probability of recommending articles to target users based on the historical article is low.
  • the content of the historical article or other related information is of interest to the user, but it rarely gets related information.
  • a subsequent article is recommended to the target user based on the historical article, it will definitely help to expand the article read by the user and increase the diversity of article recommendation.
  • each historical article corresponds to one or more pending historical articles
  • a plurality of pending historical articles can be obtained in step S303.
  • a historical article may be used as a pending historical article for one or more other historical articles.
  • a pending historical article may separately determine the degree of attention with multiple historical articles. In this way, for each article The pending historical article can obtain the degree of attention of the pending historical article and at least one historical article. Therefore, for each pending historical article, the degree of attention between the pending historical article and the corresponding historical articles can be integrated to obtain the possibility of recommending the article to the target user based on the pending historical article.
  • article 4 you may use article 3, article 2, and article 1 as pending historical articles, and determine the degree of attention between article 3, article 2, and article 1 and article 4, respectively.
  • article 3 Article 2 and Article 1 may be considered as pending historical articles, and the degree of attention of Article 2 and Article 1 and Article 3 may be determined respectively. corresponding.
  • article 2 based on the degree of attention between article 2 and article 4, and the degree of attention between article 2 and article 3, the possibility of triggering the recommendation of the article to the user based on the article 2 is comprehensively determined.
  • any one pending historical article there may be various ways to determine the possibility of recommending an article to a user based on the pending historical article.
  • the pending history is determined according to the degree of attention of the pending historical article and at least one historical article (at least one historical article corresponding to the pending historical article).
  • the average level of attention of the article to the at least one historical article can reflect the possibility of recommending articles to the user based on the pending historical article, wherein the lower the average degree of attention, the lower the possibility of triggering an article recommendation to the user.
  • the computer device may select at least one pending historical article with a lower average degree of attention from the plurality of pending historical articles corresponding to the historical article as a recommended reference article, so that it is determined based on the recommended reference article that Recommended articles for target users. For example, a pending historical article with the lowest average level of attention is selected as a recommended reference article.
  • the first attention degree of each pending historical article can be ranked and a ranking result can be obtained. If the first attention degree is sorted from high to small, the type or content corresponding to the pending historical article corresponding to the first attention degree that is higher in the ranking result is often read by users, and the later in the ranking result The type or content of pending historical articles corresponding to the first degree of attention is not often read by users.
  • a recommended reference article may be determined according to needs, such as obtaining a pending historical article corresponding to the last first attention level in the ranking result, and using the pending historical article as a recommended reference article. Then you can recommend the types of articles you don't read often based on recommended reference articles.
  • the number of recommended reference articles specifically selected may be a certain percentage of the multiple historical articles read by the user, such as 30%.
  • the computer device selects recommended reference articles from a plurality of pending historical articles as an example to illustrate the process of determining recommended reference articles from a set of historical articles in this application. For example, referring to FIG. 6, a schematic diagram of an implementation principle for determining a recommended reference article is shown.
  • the history reading set of the target user includes article d1, article d2, article d3, article d4, article d5, and article d6 as an example.
  • three articles read before the historical article are selected as pending historical articles for each article.
  • three articles relatively close to the reading time of the article are selected as pending historical articles.
  • the pending historical articles corresponding to article d6 include: article d5, article d4, and article d3.
  • the pending historical articles corresponding to article d5 include: article d4, article d3, and article d2.
  • the pending historical articles corresponding to article d4 include: article d3, article d2, and article d1.
  • the pending historical article corresponding to article d3 is only article d2 and article d2; accordingly, the pending historical article corresponding to article d2 is article d1, and article d1 No other articles have been read before, so there is no need to analyze the corresponding pending historical articles for article d1.
  • the computer equipment needs to separately calculate the attention score between the historical article and at least one pending historical article corresponding to the historical article, so that the attention sequence corresponding to the historical article.
  • the identifier of the two articles is connected by a line “-” to indicate the attention score between the two articles.
  • the attention score between article d6 and article d5 needs to be calculated.
  • each article may be regarded as multiple other articles.
  • Pending history articles a plurality of historical articles corresponding to each article as a pending historical article, and an attention score between the pending historical article and each historical article are shown.
  • article d5 is only the pending historical article corresponding to article d6, and the attention scores d6-d5 of article d5 and article d6 can be found from the sequence of attention scores corresponding to article d6, that is, the attention of article d5 and article d6
  • the force score d6-d5 is 0.6.
  • the average attention score corresponding to article d5 is 0.6.
  • article d4 is the pending historical article of article d6 and article d5, where the attention score d6-d4 of article d4 and article d6 is 0.1; and the attention score d5-d4 of article d4 and article d5 is 0.4. It can be seen that the average attention score corresponding to the pending historical article d4 is 0.25.
  • the average attention score corresponding to the article d3 as the pending historical article is 0.4; the average attention score corresponding to the article d2 as the pending historical article is 0.33; and the article d1 as the pending historical article is corresponding.
  • the computer device may default that each pending historical article corresponds to a preset number of historical articles. If the number of historical articles corresponding to a pending historical article is less than the preset number, the preset number and The difference between the number of historical articles corresponding to the historical article supplements the difference to a default degree of attention (eg, an attention score). Then, the sum of the attention levels between the pending historical article and a preset number of historical articles is calculated separately, and the added attention degree is used as a basis for reflecting the possibility that the pending historical article has the ability to trigger a recommended article.
  • a default degree of attention eg, an attention score
  • the attention score between each pending historical article and 3 historical articles needs to be determined, and the default attention score is assumed to be 0.5.
  • Article d5 in FIG. 6 is only pending historical articles corresponding to article d6. Therefore, the number of historical articles corresponding to article d5 is less than a preset number, that is, three articles.
  • the sum of the attention levels corresponding to the article d5 may be the sum of the attention scores of the articles d5 and d6 and the two default attention scores, specifically: 0.6 + 0.5 + 0.5 is equal to 1.6.
  • the computer device determines at least one candidate recommended article to be recommended to the target user from the set of recommendable articles according to the content characteristics of each recommended reference article.
  • the recommendable article collection includes multiple articles that can be recommended to the target user.
  • the recommendable article collection can be a collection of all articles that can be published by the article publishing platform.
  • the recommended article collection may also be a collection of articles read by all users in the article publishing platform.
  • the computer device may determine the specific recommended article to be recommended to the target user in various ways. For example, in a possible implementation manner, the computer device may, based on the content characteristics of the recommended reference article, combine the article collaborative recommendation algorithm, the content-based recommendation algorithm, or the sequence-based recommendation algorithm from the set of recommendable articles, Identify at least one candidate recommended article.
  • the computer device may also determine the candidate recommendation article based on the attention recommendation strategy.
  • the attention-based recommendation strategy is: for each recommended reference article, the recommended reference article and the recommendable article can be determined according to the content characteristics of the recommended reference article and the content characteristics of each article in the set of recommendable articles. Attention level for each article in the collection. Accordingly, from the set of recommendable articles, at least one candidate recommended article with a higher degree of attention to the recommended reference article may be selected.
  • multiple candidate recommended articles may also be determined. For example, after calculating the second degree of attention of each article in the recommendable article set and at least one recommended reference article, the obtained second degree of attention can be ranked, and the higher the second degree of attention means that the user reads When a reference article is recommended, it is more likely to be read. If two candidate recommended articles are determined from the set of recommendable articles, an article with the second highest degree of attention and an article with the second highest degree of attention may be selected from the set of recommendable articles.
  • the manner in which the computer device determines the degree of attention of each article in the recommended reference article and the recommendable article collection can refer to the related introduction of determining the degree of attention, for example, the article vector of each article in the recommendable article collection can be determined. And the article vector of the recommended reference article is input into a pre-trained attention model to output the degree of attention (eg, attention score) between the recommended reference article and each article in the set of recommendable articles. Sequence of attention scores.
  • the method for determining the degree of attention in other ways described above is also applicable here, and is not repeated here.
  • the computer device analyzes the degree of attention of other historical articles and historical articles read by the user before reading the historical article, due to the attention between the two articles.
  • the degree of strength can reflect the possibility that a user is recommended to read another article in the case of reading one article. Therefore, according to the determined degree of attention between each article and other articles, it can be obtained from the reading history collection. Identify historical articles that are less likely to recommend articles to target users, and the type and content of the historical articles are not the type and content of articles that users often read. Therefore, use this historical article as a recommended reference article. Recommending candidate recommendation articles to users is helpful to increase the diversity of the candidate recommendation articles that are determined, and in turn helps to increase the variety of articles recommended to users.
  • the computer device can directly regard the candidate recommended article as an article that needs to be recommended to the target user.
  • the number of candidate recommended articles may be relatively large; moreover, in actual applications, multiple recommendation strategies may be configured, and different recommendations will be used to determine the matching of the recommended reference article.
  • Multiple candidate recommendation articles will also cause a large number of candidate recommendation articles to be determined. If these large number of candidate recommendation articles are recommended to the user, it is difficult to accurately recommend the user's articles to the user, which affects the recommendation effect.
  • At least one candidate recommended article is determined based on the degree of attention between each article in the recommendable article set and the recommended reference article, based on one or more of content recommendation algorithms, collaborative recommendation algorithms, and other methods.
  • the computer device can also calculate the attention degree of each historical article in the historical reading set and the candidate recommended article for each candidate recommended article, and get the attention between the candidate recommended article and multiple historical articles in the historical reading set. Force degree sequence. Then, the information entropy of the attention degree sequence corresponding to each candidate recommended article is calculated separately to obtain the information entropy corresponding to each candidate recommended article. Accordingly, from the at least one candidate recommendation article, at least one candidate recommendation article with less information entropy may be selected as at least one target recommendation article to be recommended to the target user.
  • the information entropy of each candidate recommended article can be sorted. The smaller the information entropy, the higher the priority of the corresponding candidate recommended article. If two target recommendation articles need to be pushed, the information entropy of the candidate recommendation articles is sorted, and the candidate recommendation article with the smallest information entropy and the candidate recommendation article with the second smallest information entropy are obtained, and the two candidate recommendation articles are used as the target. Recommend articles to users.
  • the degree of attention of each historical article and the candidate recommended article in the historical reading collection can reflect the possibility of triggering reading of the candidate recommended article by each historical article; and, the candidate recommended article and the history
  • the information entropy corresponding to the attention degree sequence between each historical article in the reading set can measure the stability or reliability of each historical article in the historical reading set triggering reading of the candidate recommended article. It can be seen that by selecting the corresponding candidate recommended article with a small information entropy as the target recommended article to be recommended, it is conducive to the maximum degree of recommending articles that the user is interested in to the user on the premise of increasing the variety of article recommendations to the user. Achieve personalized recommendations based on users.
  • the computer device determines at least one target recommendation article, it can also determine the ranking of the at least one target recommendation article, so that the subsequent article publishing platform outputs the at least one target recommendation article according to the ranking.
  • FIG. 7 illustrates another schematic flowchart of an article recommendation method according to an embodiment of the present application.
  • the method in this embodiment is described from a computer device side.
  • the method may include:
  • the computer device acquires a historical reading set of the target user to be analyzed.
  • the historical reading collection includes: multiple historical articles read by the target user at different reading moments.
  • the computer device determines the article vector of the historical article, and according to the historical reading set, the reading time is at least one pending historical article before the reading time of the historical article.
  • the respective content characteristics determine the article vector of each pending historical article.
  • the computer device inputs an article vector of at least one pending historical article corresponding to the historical article and a content vector of the historical article into a pre-trained attention model to output through the attention model. Attention score sequence.
  • the attention score sequence includes attention scores between the at least one pending historical article and the historical article, respectively.
  • the pending historical article and the attention score of the historical article may reflect the possibility that the user is recommended to read the historical article when the user reads the pending historical article.
  • the computer device determines an average attention score of the pending historical article to at least one historical article according to the attention score of the pending historical article to at least one historical article.
  • the computer device selects at least one pending historical article with a lower average attention score from the plurality of pending historical articles corresponding to the historical article as a recommended reference article.
  • steps S701 to S705 are only an implementation method for determining a recommended reference article based on the attention score between the pending historical article and the corresponding historical article.
  • steps S704 and S705 are only an implementation method for determining a recommended reference article based on the attention score between the pending historical article and the corresponding historical article.
  • the other methods mentioned in the previous embodiment are also applicable to this implementation. Examples are not repeated here.
  • steps S701 to S705 are actually selection strategies based on the attention model, that is, based on the attention model, the recommended reference article is selected from the target user's reading history set.
  • selecting recommended reference articles based on the attention model can only be a selection strategy for selecting articles to be recommended. Therefore, in addition to configuring the attention model-based selection strategy in computer equipment, You can also configure one or more other selection strategies at the same time, such as a selection strategy based on user portraits, that is, based on the user portrait of the target user, select one or more historical articles from the reading history collection as recommended reference articles.
  • the computer device can separately select the recommended reference articles through the configured multiple selection strategies and perform subsequent operations.
  • FIG. 8 a schematic diagram of an implementation principle of the article recommendation method in this application. It can be seen from FIG. 8 that there can be multiple recommendation strategies for selecting recommended reference articles. Correspondingly, according to the reading history set read by the target user and the user portrait of the user's target user, a variety of different recommendation strategies are used to determine recommended reference articles from the reading history set. In this way, based on each recommendation strategy, it is possible to Pick a recommended reference article.
  • the computer device determines the second attention of each article in the recommendable article set and the recommended reference article according to the content characteristics of the recommended reference article and the content characteristics of each article in the recommendable article set. Strength.
  • the second degree of attention between any article in the recommendable article collection and the recommended reference article indicates the possibility that the user is recommended to read the article in the recommendable article collection when reading the recommended reference article.
  • the computer device selects at least one candidate recommended article with a higher degree of second attention than the recommended reference article.
  • the attention-based recommendation strategy is actually used to determine the candidate recommendation article as an example, but it can be understood that the method of determining the candidate recommendation article through other recommendation strategies is also applicable to this implementation. example.
  • a computer device when a computer device determines a candidate recommended article based on a recommended reference article, it may only set a recommendation strategy, such as the attention-based recommendation strategy mentioned above.
  • multiple recommendation strategies may be set, for example, while setting attention-based recommendation strategies, content-based recommendation strategies, collaborative recommendation strategies, and so on.
  • the computer equipment can use each recommended reference article as a benchmark and multiple recommendation strategies to determine at least one corresponding candidate recommendation article from the set of recommendable articles; or, by combining multiple recommendation strategies, Identify at least one candidate recommended article.
  • candidate recommended articles can be selected from the set of recommended articles in order to perform subsequent preliminary selection operations.
  • the computer device separately calculates the third degree of attention of each historical article in the historical reading set and the candidate recommendation article, and obtains between the candidate recommended article and multiple historical articles in the historical reading set. Sequence of attention levels.
  • the attention degree sequence includes the third attention degree of the candidate recommendation article and a plurality of historical articles in the historical reading set.
  • the computer device separately calculates the information entropy of the attention degree sequence corresponding to each candidate recommended article, and obtains the information entropy corresponding to each candidate recommended article.
  • the information entropy corresponding to the attention degree sequence can reflect the distribution of the attention degree between the candidate recommendation article and each historical article in the historical reading set.
  • the information entropy corresponding to the attention score sequence H (A) can be calculated by the following formula:
  • a i in formula 4 is the attention score in the attention score sequence A t .
  • S710 From the at least one candidate recommendation article, select at least one candidate recommendation article with smaller information entropy as at least one target recommendation article to be recommended to the target user.
  • the computer equipment can select the corresponding candidate recommendation article with a smaller information entropy as the target recommendation article, so that the target recommendation article can be screened to meet the user's interest on the premise of satisfying the diversity.
  • the above steps S708 to S710 are based on the attention model, and from the candidate recommendation articles, a way of determining the recommended target recommendation articles is finally selected, that is, the primary selection strategy based on the attention model.
  • the preliminary selection strategy may be based on user portraits to select at least one target recommendation article that matches the user portrait from multiple candidate recommendation articles.
  • preliminary selection strategies can also be integrated on the basis of the preliminary selection strategy based on the attention model to finally screen at least one target recommendation article from the candidate recommendation articles.
  • a plurality of preliminary selection strategies can be configured in the computer device.
  • the computer device may determine at least one target recommended article from the candidate recommendation articles based on the multiple preliminary selection strategies, respectively; or, it may comprehensively determine at least one candidate recommended article by combining multiple preliminary selection strategies.
  • the computer device determines a fourth degree of attention of the at least one target recommended article and the recommended reference article, respectively.
  • the fourth degree of attention reflects the possibility that the user is recommended to read the target recommended article when the user reads the recommended reference article.
  • the computer device determines a recommendation order of the at least one target recommendation article according to a fourth degree of attention between the at least one target recommendation article and the recommended reference article.
  • the comprehensive value of the fourth attention degree of the target recommended article can be determined according to the fourth attention degree of the target recommended article and each of the recommended reference articles, and The recommendation order is determined according to the comprehensive value of each target recommendation article.
  • the higher the fourth attention degree corresponding to the target recommendation article the more likely that the user chooses to read the target recommendation article when the user reads the recommended reference article, so the corresponding fourth attention degree is higher.
  • the computer device determines the recommendation order of the at least one target recommendation article, it can send the identification and recommendation order of the at least one target recommendation article to the article publishing server.
  • the article publishing server can present the at least one target recommended article to the target user according to the recommendation order, so that the target user can See a wider range of articles that are more relevant to your interests.
  • the second attention level, the third attention level, and the fourth attention level are only for the convenience of distinguishing the attention levels between different articles. Ways of attention between articles, for example, using a pre-trained attention model to determine the above second attention, third attention, and fourth attention. For details, refer to the related examples in the previous embodiment. Introduction, will not repeat them here.
  • this application also provides an article recommendation method.
  • the above article recommendation method includes:
  • the computer device obtains a historical reading set of the target user to be analyzed.
  • the historical reading set includes: multiple historical articles that the target user has read at different reading moments;
  • the computer device determines, for each historical article in the multiple historical articles, the first attention degree of each pending historical article and each historical article before the reading of each historical article.
  • the first attention degree is reflected in In the case of users reading pending historical articles, the possibility of recommending users to read each historical article;
  • the computer device selects at least one recommended reference article from the plurality of pending historical articles according to the first attention degree of each pending historical article and at least one historical article;
  • the computer device determines at least one candidate recommended article to be recommended to the target user from the set of recommendable articles according to each recommended reference article.
  • FIG. 10 is an optional user browsing interface.
  • the user can use the terminal 1002 to log in to the client to browse articles, and FIG. 10 shows the article 1004 that the client has recommended according to the existing recommendation method.
  • Users can read some articles by clicking on the relevant article title or link.
  • some reading records will be generated, and the reading records include the articles read and the reading moments.
  • Use the reading history as the user's historical reading collection.
  • the history reading collection includes multiple historical articles that users have read.
  • the mobile phone After obtaining the above-mentioned multiple historical articles, the mobile phone will obtain the pending historical articles of each of the multiple historical articles.
  • Pending historical articles are those that were read before each of the above historical articles. For example, taking multiple historical articles including “We are the champions” and “schedule", “starring table”, and “costumes” as examples, the reading order of the four historical articles is: “We are the champions”, “ “Schedule”, “Starring table”, “Costume”.
  • the historical articles to be determined for the “Costume” are three historical articles before the “Costume”; the historical articles to be determined for the “Starring table” are two historical articles for the “Starring Table”; the historical articles for the “Schedule” are to be located A historical article before the "Schedule”; "We are champions” has no pending historical article.
  • the mobile phone determines the pending historical article of each historical article, it determines the first attention level of each pending historical article and the corresponding historical article. As shown in Table (1), the horizontal line "-" in Table 1 indicates that the first degree of attention is not included.
  • the method for determining the first degree of attention may adopt various methods mentioned above, and details are not described herein again. It can be seen that after the user reads "We are the champion", the possibility of reading the "schedule” is higher, and the possibility of reading the "starring table” is lower. For each pending historical article, such as “we are champions”, the first attention level of "we are champions” and at least one corresponding historical article may be calculated. As calculated above, the first attention level of "We are the champion” and “schedule” is 0.9, the first attention level of "We are the champion” and “actors” is 0.4, “we are the champion” and "costume” The first degree of attention is 0.2, which gives an average of 0.5. In the same way, the average of the first attention level of each historical article to be determined is obtained. The average of the first attention of the "schedule” is 0.2, and the first attention level of the "starring table” is 0.8.
  • the mobile phone determines the "schedule” with the lowest average first attention among a number of pending historical articles as a recommended reference article, and then recommends a candidate recommended article from the set of recommendable articles according to the recommended reference article "schedule". For example, if the set of recommendable articles includes “schedule time” and "director list”, "schedule time” can be used as a candidate recommended article, and when the user obtains the recommended article, the candidate recommended article is recommended to the user. As shown in FIG. 11, the user requests a recommendation article by a pull-down method in FIG. 11. As shown in FIG. 12, the mobile phone recommends “schedule time” to the user and displays it in the recommendation area 1202.
  • this application also provides an article recommendation device.
  • FIG. 13 shows a schematic structural diagram of an embodiment of an article recommendation device of the present application.
  • the device of this embodiment can be applied to the preceding computer equipment.
  • the device may include:
  • the history acquisition unit 1301 is configured to acquire a historical reading set of a target user to be analyzed, and the historical reading set includes: multiple historical articles that the target user has read at different reading times;
  • the first attention analysis unit 1302 is configured to determine, for each historical article, according to the content characteristics of the historical article and the content characteristics of at least one pending historical article whose reading time is before the reading time of the historical article in the historical reading set.
  • the first attention level of the pending historical articles and historical articles, the first attention level is reflected in the possibility that the user is recommended to read the historical articles when the user reads the pending historical articles;
  • the recommendation reference determining unit 1303 is configured to select the probability of triggering the recommendation of the article to the target user from the plurality of pending historical articles according to the first attention degree of the obtained pending historical articles and at least one historical article. At least one pending historical article as a recommended reference article;
  • the candidate recommendation determining unit 1304 is configured to determine at least one candidate recommended article to be recommended to a target user from a set of recommendable articles according to the content characteristics of each recommended reference article.
  • the first attention analysis unit may include:
  • the first attention analysis sub-unit is set according to the content characteristics of the historical article and the content characteristics of at least one pending historical article whose reading time is before the reading time of the historical article in the historical reading set, and uses a pre-trained attention model, Determine the attention score of each pending historical article and historical article separately;
  • the attention model is trained by using a set of historical article samples of multiple users and a sequence of labeled attention scores corresponding to each historical article sample in each historical article sample set.
  • the power score sequence includes: a plurality of historical article samples that the user read before reading the historical article sample and an attention score of the historical article sample, respectively.
  • the first attention analysis sub-unit may include:
  • the first vector determining subunit is set to determine the article vector of the historical article according to the content characteristics of the historical article
  • the second vector determining subunit is set to determine an article vector of each pending historical article according to the respective content characteristics of at least one pending historical article in the historical reading set before the reading time of the historical article;
  • the attention score analysis subunit is set according to the article vector of the historical article and the article vector of at least one pending historical article, and uses the attention model to determine the attention score of each pending historical article and the historical article separately.
  • the recommended reference determination unit includes:
  • the average attention analysis sub-unit is set to determine, for each pending historical article, the average attention degree of the pending historical article to at least one historical article according to the first attention level of the pending historical article and at least one historical article;
  • the reference screening subunit is set to select at least one pending historical article with a lower average degree of attention as a recommended reference article from among the plurality of pending historical articles corresponding to the multiple historical articles.
  • the candidate recommendation determination unit includes:
  • the second attention analysis unit is set for each recommended reference article. Based on the content characteristics of the recommended reference article and the content characteristics of each article in the set of recommendable articles, determine the content of each article in the set of recommended articles and the recommended reference article separately. Second attention degree, the second attention degree reflects the possibility that the user is recommended to read the articles in the set of recommendable articles when the user is reading the recommended reference article;
  • the candidate recommendation selection unit is configured to select at least one candidate recommendation article with a higher degree of second attention from the recommended reference article from the set of recommendable articles.
  • FIG. 14 illustrates another composition and structure diagram of the article recommendation device of the present application.
  • the article recommendation device of this embodiment is different from the device shown in FIG. 14 in the foregoing in that:
  • the device In addition to the history acquisition unit 1401, the first attention analysis unit 1402, the recommendation reference determination unit 1403, and the candidate recommendation determination unit 1404, the device also includes:
  • the third attention analysis unit 1405 is configured to calculate the third attention of each historical article in the historical reading set and the candidate recommendation article for each candidate recommendation article after the candidate recommendation determination unit determines at least one candidate recommendation article. Degree to obtain the attention degree sequence between the candidate recommendation article and multiple historical articles in the historical reading collection, and the third attention degree reflects the user's recommendation to read the candidate recommendation article when reading the historical article in the historical reading collection. possibility;
  • the information entropy calculation unit 1406 is configured to separately calculate the information entropy of the attention degree sequence corresponding to each candidate recommended article, to obtain the information entropy corresponding to each candidate recommended article;
  • the target recommendation determining unit 1407 is configured to select, from at least one candidate recommendation article, at least one candidate recommendation article with less information entropy as at least one target recommendation article to be recommended to a target user.
  • the device may further include:
  • the fourth attention analysis unit 1408 is configured to determine the fourth attention degree of the recommended reference article and the at least one target recommendation article, respectively.
  • the fourth attention degree reflects that the user recommends the user to read the target recommendation article while reading the recommended reference article.
  • the ranking determination unit 1409 is configured to determine a recommendation order of at least one target recommendation article according to the fourth attention degree of the recommended reference article and the at least one target recommendation article, respectively.
  • the history acquisition unit 1401, the first attention analysis unit 1402, the recommendation reference determination unit 1403, and the candidate recommendation determination unit 1404 can refer to the relevant descriptions of the previous embodiments for details.
  • a storage medium stores a computer program, and the computer program is configured to execute the steps in any one of the foregoing method embodiments when running.
  • the foregoing storage medium may be configured to store a computer program for performing the following steps:
  • the historical reading set includes: multiple historical articles that the target user has read at different reading times;
  • the first attention level reflects the possibility of recommending a user to read a historical article when the user is reading a pending historical article
  • the foregoing storage medium may be configured to store a computer program for performing the following steps:
  • the computer device acquires a historical reading set of the target user to be analyzed.
  • the historical reading set includes: multiple historical articles that the target user has read at different reading moments;
  • the computer equipment determines the first attention degree of each pending historical article and each historical article before each historical article at the time of reading.
  • the first attention degree is reflected in In the case of users reading pending historical articles, the possibility of recommending users to read each historical article;
  • the computer device selects at least one recommended reference article from the plurality of pending historical articles according to the first attention degree of each pending historical article and at least one historical article;
  • the computer device determines at least one candidate recommended article to be recommended to the target user from the set of recommendable articles according to each recommended reference article.
  • the storage media may include: a flash disk, a read-only memory (ROM), a random access device (Random Access Memory, RAM), a magnetic disk, or an optical disk.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Databases & Information Systems (AREA)
  • Psychiatry (AREA)
  • Social Psychology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Pathology (AREA)
  • Veterinary Medicine (AREA)
  • Psychology (AREA)
  • Public Health (AREA)
  • Child & Adolescent Psychology (AREA)
  • Developmental Disabilities (AREA)
  • Educational Technology (AREA)
  • Animal Behavior & Ethology (AREA)
  • Hospice & Palliative Care (AREA)
  • Surgery (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Evolutionary Biology (AREA)
  • Human Computer Interaction (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请提供了一种文章推荐方法、装置、计算机设备及存储介质，该方法包括：计算机设备获取待分析的目标用户的历史阅读集合，历史阅读集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章；计算机设备针对多篇历史文章中的每篇历史文章，确定阅读时刻在每篇历史文章之前的每篇待定历史文章与每篇历史文章的第一注意力程度；计算机设备依据每篇待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇待定历史文章中选取出至少一篇推荐参考文章；计算机设备分别根据每篇推荐参考文章，从可推荐文章集合中，确定待推荐给目标用户的至少一篇候选推荐文章。本申请的方案可以提高向用户推荐文章的多样性。

Description

文章推荐方法、装置、计算机设备及存储介质
本申请要求于2018年5月25日提交中国专利局、优先权号为2018105158698、申请名称为“文章推荐方法、装置、计算机设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及数据处理技术领域，尤其涉及一种文章推荐方法、装置、计算机设备及存储介质。
背景技术
随着网络的不断发展，用户通过互联网中各种文章发布平台(如，即时通讯平台中的各种公众号等)所能阅读到的文章数量也日益增多。
为了吸引用户阅读文章发布平台的文章，文章发布平台会根据用户阅读过的历史文章，向用户推荐文章。但是目前为用户推荐文章时，都是根据用户阅读历史，向用户推荐内容上相似的文章。如，基于内容的推荐技术是根据用户历史阅读文章，确定与用户阅读过的文章的内容相似的文章，并将这些相似文章推荐给用户。
然而，由于为用户推荐的文章与用户历史阅读过的文章的内容大多相似，会导致文章推荐的多样性较差。
发明内容
有鉴于此，本申请提供了一种文章推荐方法、装置、计算机设备及存储介质，以提高文章推荐的多样性。
为实现上述目的，一方面，本申请提供了一种文章推荐方法，包括：
计算机设备获取待分析的目标用户的历史阅读集合，上述历史阅读集合包括：上述目标用户在不同阅读时刻阅读过的多篇历史文章；
上述计算机设备针对上述多篇历史文章中的每篇历史文章，确定阅读时刻在上述每篇历史文章之前的每篇待定历史文章与上述每篇历史文章的第一注意力程度，上述第一注意力程度反映在用户阅读上述待定历史文章的情况下，推荐用户阅读上述每篇历史文章的可能性；
上述计算机设备依据每篇上述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇上述待定历史文章中选取出至少一篇推荐参考文章；
上述计算机设备分别根据每篇上述推荐参考文章，从可推荐文章集合中，确定待推荐给上述目标用户的至少一篇候选推荐文章。
又一方面，本申请还提供了一种文章推荐装置，应用在计算机设备中，包括：
历史获取单元，设置为获取待分析的目标用户的历史阅读集合，上述历史阅读集合包括：上述目标用户在不同阅读时刻阅读过的多篇历史文章；
第一注意分析单元，设置为针对上述多篇历史文章中的每篇历史文章，确定阅读时刻在上述每篇历史文章之前的每篇待定历史文章与上述每篇历史文章的第一注意力程度，上述第一注意力程度反映在用户阅读上述待定历史文章的情况下，推荐用户阅读上述每篇历史文章的可能性；
推荐参考确定单元，设置为依据每篇上述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇上述待定历史文章中选取出至少一篇推荐参考文章；
候选推荐确定单元，设置为分别根据每篇上述推荐参考文章，从可推荐文章集合中，确定待推荐给上述目标用户的至少一篇候选推荐文章。
又一方面，本申请还提供了一种计算机设备，包括：
处理器和存储器；
其中，上述处理器用于执行上述存储器中存储的计算机程序；
上述存储器用于存储上述计算机程序，上述计算机程序至少用于：
获取待分析的目标用户的历史阅读集合，上述历史阅读集合包括：
上述目标用户在不同阅读时刻阅读过的多篇历史文章；
针对上述多篇历史文章中的每篇历史文章，确定阅读时刻在上述每篇历史文章之前的每篇待定历史文章与上述每篇历史文章的第一注意力程度，上述第一注意力程度反映在用户阅读上述待定历史文章的情况下，推荐用户阅读上述每篇历史文章的可能性；
依据每篇上述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇上述待定历史文章中选取出至少一篇推荐参考文章；
分别根据每篇上述推荐参考文章，从可推荐文章集合中，确定待推荐给上述目标用户的至少一篇候选推荐文章。
又一方面，本申请还提供了一种存储介质，上述存储介质中存储有计算机程序，上述计算机程序被运行时，执行上述上述上述文章推荐方法
经以上可知，在本申请实施例中，对于用户阅读过的历史文章，会分析用户在阅读该历史文章之前阅读的其他历史文章与该历史文章的注意力程度，由于两篇文章之间的注意力程度可以反映出用户阅读一篇文章的情况下，推荐用户阅读另一篇文章的可能性，因此，依据确定出的各篇文章与其他文章之间的注意力程度，可以从阅读历史集合中确定出表征向目标用户推荐文章的可能性较低的历史文章，而该历史文章的类型及内容便不属于用户经常阅读到的文章类型及文章内容，因此，以该历史文章作为推荐参考文章，向用户推荐候选推荐文章，有利于提高确定出的候选推荐文章的多样性，进而有利于提高向用户推荐文章的多样性。
附图说明
为了更清楚地说明本申请实施例或相关技术中的技术方案，下面将对实施例或相关技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的实施例，对于本领域普通技术人员来 讲，在不付出创造性劳动的前提下，还可以根据提供的附图获得其他的附图。
图1示出了本申请实施例的文章推荐方法所适用的一种系统组成架构示意图；
图2示出了本申请实施例的文章推荐方法所适用的一种计算机设备的组成架构示意图；
图3示出了本申请实施例的文章推荐方法的一种流程示意图；
图4示出了本申请实施例的文章推荐方法中训练注意力模型的一种流程示意图；
图5示出了本申请实施例中基于LSTM输出文章的文章向量的示意图；
图6示出了本申请实施例中确定推荐参考文章的一种实现原理示意图；
图7示出了本申请实施例的文章推荐方法又一个实施例的流程示意图；
图8示出了本申请实施例的文章推荐方法的一种实现原理示意图；
图9示出了本申请实施例的另一种文章推荐方法的流程示意图；
图10示出了本实施例的一种可选的界面显示图；
图11示出了本实施例的另一种可选的界面显示图；
图12示出了本实施例的又一种可选的界面显示图；
图13示出了本申请实施例的文章推荐装置的一种组成架构示意图；
图14示出了本申请实施例的文章推荐装置的又一种组成架构示意图。
具体实施方式
本申请实施例的文章推荐方法适用于各种文章发布平台，以确定向文章发布平台中不同用户所需要推荐的文章，提高向用户推荐文章的多样性。如，新闻媒体平台针对不同用户推荐新闻文章；社交平台或者社交平台接入的公众号向用户推荐文章等。
为了便于理解，参见图1，其示出了本申请的一种文章推荐系统的组成架构示意图。
由图1可知，该文章推荐系统可以包括：计算机设备101以及至少一台文章发布服务器102。
其中，文章发布服务器102，用于发布可供用户阅读的多篇文章，并记录不同用户在不同时刻阅读过的文章。为了便于区分，将用户阅读过的文章称为历史文章。
该计算机设备101，用于分别基于每个用户阅读过的多篇历史文章，确定需要向该用户推荐的推荐文章。
该计算机设备101还可以将确定出的待推荐给用户的推荐文章的信息发送给文章发布服务器。
可选的，该文章推荐系统还可以包括数据存储服务器103，该数据存储服务器可以存储至少一台文章发布服务器所产生的用户数据，如用户画像，以及用户的阅读历史记录等等。其中，用户画像可以包括用户的性别、年龄、学历以及所处地区等用户属性信息。阅读历史记录可以包括：用户阅读过的文章以及文章相关的来源等等信息。
相应的，计算机设备101可以从数据存储服务器103获取用户的阅读历史记录，以根据用户阅读历史记录，确定需要向该用户推荐的推荐文章。
需要说明的是，计算机设备101所依据的用户阅读过的历史文章可以是来源于某一个文章发布平台中的一台或多台文章发布服务器，也可以是来源于多个文章发布平台的文章发布服务器。
其中，对于某个文章发布平台而言，计算机设备101可以是独立于文章发布服务器单独设置的服务器或者终端；也可以是与文章发布服务器为同一台服务器，也就是说，该计算机设备具备文章发布以及确定向用户推荐文章的功能。
为了便于理解该计算机设备，如，参见图2，其示出了本申请的计算机设备的一种组成示意图。
在图2中，该计算机设备200可以包括：处理器201、存储器202、通信接口203、输入单元204和显示器205和通信总线206。
处理器201、存储器202、通信接口203、输入单元204、显示器205、均通过通信总线206完成相互间的通信。
在本申请实施例中，该处理器201，可以为中央处理器(Central Processing Unit，CPU)，特定应用集成电路(application-specific integrated circuit，ASIC)，数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件等。
该处理器可以调用存储器202中存储的程序，处理器可以执行以下图3至图9中服务器侧所执行的操作。
存储器202中用于存放一个或者一个以上程序，程序可以包括程序代码，程序代码包括计算机操作指令，在本申请实施例中，该存储器中至少存储有用于实现以下功能的程序：
获取待分析的目标用户的历史阅读集合，历史阅读集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章；
针对每篇历史文章，依据历史文章的内容特征以及历史阅读集合中阅读时刻在历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，分别确定每篇待定历史文章与历史文章的第一注意力程度，第一注意力程度反映在用户阅读待定历史文章的情况下，推荐用户阅读历史文章的可能性；
依据得到的多篇待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇待定历史文章中，选取出触发向目标用户推荐文章的可能性相对较低的至少一篇待定历史文章作为推荐参考文章；
分别根据每篇推荐参考文章的内容特征，从可推荐文章集合中，确定待推荐给目标用户的至少一篇候选推荐文章。
其中，存储器202可包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、以上所提到的程序，以及至少一个功能(比如声音播放功能、图像播放功能等)所需的应用程序等；存储数据区可存储根据计 算机设备的使用过程中所创建的数据以及接收到的待处理数据，比如，用户阅读记录等。
此外，存储器202可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他易失性固态存储器件。
该通信接口203可以为通信模块的接口。
本申请中该服务器还可以包括输入单元205，如，键盘等等。
该显示器204包括显示面板。在一种可能的情况中，可以采用液晶显示器(Liquid Crystal Display，LCD)、有机发光二极管(Organic Light-Emitting Diode,OLED)等形式来配置显示面板。
当然，图2所示的计算机设备结构并不构成对本申请实施例中计算机设备的限定，在实际应用中计算机设备可以包括比图2所示的更多或更少的部件，或者组合某些部件。
结合以上共性，对本申请实施例的一种文章推荐方法进行介绍。如，参见图3，其示出了本申请实施例一种文章推荐方法的一种流程示意图，本实施例从计算机设备侧进行介绍。该过程包括：
S301，获取待分析的目标用户的历史阅读集合。
其中，历史阅读集合包括：该目标用户在不同阅读时刻阅读过的多篇历史文章。
如，历史阅读集合包括按照阅读先后顺序排序的多篇历史文章，这样，根据该多篇历史文章的先后顺序，便可以确定该多篇历史文章对应的阅读时刻的先后顺序。
又如，该历史阅读集合可以包括多篇目标用户阅读过的多篇历史文章以及每篇历史文章对应的阅读时刻。
可以理解的是，针对不同的用户，需要结合不同用户的历史阅读集合，来确定适合推荐给该用户的文章，在本申请实施例中，为了便于描述，将当前需要确定待推荐文章的用户称为待分析的目标用户。同时，为了与推荐给用户的文章进行区分，将用户阅读过的文章称为历史文章。
S302，针对每篇历史文章，依据该历史文章的内容特征，计算机设备从历史阅读集合中阅读时刻在该历史文章的阅读时刻之前的至少一篇待定历史文章。
可以理解的是，对于历史阅读集合中的任意一篇历史文章，可以认为是在该目标用户通过阅读该历史文章之前，由于阅读其他历史文章而阅读到该历史文章。而此处的其他历史文章与该历史文章之间的关系可以表征出其他历史文章与该历史文章之间的聚焦程度，即，以其他历史文章为基准，推荐用户阅读该历史文章的可能性，因此，本申请需要确定用户在阅读历史文章之前阅读的其他历史文章，并分析相应的历史文章之间的关系。
如，历史文章集合中包括文章A、文章B和文章C，且这三篇文章的顺序表征了这三篇文章被用户阅读的阅读先后顺序，由此可知，文章A和文章B是在阅读文章C之前阅读的，那么文章C有可能是由于用户阅读了文章A和文章B中的一篇或者两篇文章后，选择阅读的该文章C，因此，需要确定文章C之前阅读的文章A与文章B，并分析文章A、文章B分别与文章C的关系，以得到由文章A触发用户阅读文章C的可能性，以及由文章B触发用户阅读文章C的可能性。
其中，为了便于区分，将该历史文章之前阅读过的，用于分析与该历史文章之间关系的其他历史文章均称为待定历史文章。
在本申请实施例中，可以将该历史文章之前阅读过的每篇待定历史文章均作为待定历史文章，并执行后续分析待定历史文章与该历史文章之间关系的操作。可选的，考虑到与在该历史文章之前阅读过的文章数量较大，为了降低数据处理量，又不影响后续分析结果，可以将该历史文章之前阅读到的设定数量篇历史文章作为待定历史文章。例如，从该历史文章之前阅读的其他历史文章中，随机选取设定数量篇历史文章，以将选取出的设定篇历史文章作为待定历史文章，并执行后续操作。
考虑到与该历史文章的阅读时刻较为接近的其他历史文章，对于用户能否阅读到该历史文章的影响性较大，因此，可以从阅读时刻处于该历史文章的阅读时刻之前，且阅读时刻与该历史文章的阅读时刻之间的时长不 超过预设时长的多篇历史文章中，随机选取预设数量篇历史文章作为待定历史文章或者选取与该历史文章的阅读时刻最近的多篇历史文章作为待定历史文章。
需要说明的是，考虑到历史阅读集合中，会存在某些历史文章之前阅读的文章的数量不足预设数量，例如，历史阅读集合中阅读时刻处于第二位的历史文章之前仅仅有一篇文章，因此，在该种情况中，历史文章对应的待定历史文章的数量可以少于该预设数量。
S303，针对每篇历史文章，依据该历史文章的内容特征以及该历史文章对应的至少一篇待定历史文章的内容特征，计算机设备分别确定每篇待定历史文章与该历史文章的第一注意力程度。
其中，该注意力程度也可以称为聚焦程度，反映的是阅读一篇文章的情况下，推荐用户阅读另一篇文章的可能性。同时，注意力程度也可以反映出在用户一篇文章的情况下，通过点击等方式选择阅读另一篇文章的可能性。正是由于注意力程度的以上含义，其也被称为从一篇文章到另一篇文章的注意力程度或者聚焦程度。
在本申请实施例中，注意力程度可以是一个分值、概率、等级等等，在此不加以限制。
其中，为了便于与后续其他文章之间的注意力程度进行区分，将待定历史文章与历史文章之间的注意力程度称为第一注意力程度。相应的，该第一注意力程度用于反映在用户阅读该待定历史文章的情况下，推荐用户阅读该历史文章的可能性。在该步骤S303中，计算机设备需要分别确定每篇待定历史文章各自与该历史文章的第一注意力程度。
可以理解的是，对于任意文章，文章的内容特征由该文章的标题及文章的正文中的字词决定，文章的内容特征可以反映出文章类型、文章的主题思想及文章所涉及到的具体内容等等。因此，对于任意两篇文章，可以依据这两篇文章各自的内容特征，来分析从其中一篇文章到另一篇文章的注意力程度。
在一种可能的实现方式中，对于任意两篇文章，基于这两篇文章各自 的内容特征，计算机设备可以分析这两篇文章的内容特征的相似性，并依据内容特征的相似性，确定这两篇文章之间的注意力程度。如，将这两篇文章的内容特征的相似性，作为这两篇文章之间的注意力程度。
相应的，计算机设备依据待定历史文章与历史文章的内容特征，确定待定历史文章与历史文章的相似性，确定待定历史文章与历史文章之间的注意力程度。
特别的，大部分情况需要确定一个文章集合中各篇文章M与某一篇文章N的注意力程度的情况，而且可能需要比较不同文章集合中的文章M与该篇文章M的注意力程度之间的大小关系。在该种情况下，为了便于比较各篇文章M与该篇文章N之间的注意力程度，对于每个文章集合，可以在分别确定出该文章集合中每篇文章M与该篇文章N的内容特征的相似性之后，根据该文章集合中各篇文章M各自对应的相似性，分别对该文章集合中每篇文章M与文章N的相似性进行归一化，并将每篇文章M与该文章N之间的相似性对应的归一化结果作为反映注意力程度的评价指标。
其中，计算机设备按照上面的方式，对一篇文章M与文章N的内容特征的相似性进行归一化之后，归一化结果都是一个大于等于零小于1的数值，通过该数值可以反映从该文章M触发点击文章N的可能性。
在又一种可能的实现方式中，计算机设备可以预先训练得到用于确定注意力程度的注意力模型。该注意力模型为利用多个用户的历史文章样本集合以及每个历史文章样本集合中每篇历史文章样本各自对应的已标注的注意力分数序列训练得到的。其中，为了便于区分，将训练注意力模型的历史文章称为历史文章样本。
其中，对于一个用户而言，该用户的历史文章样本对应的注意力分数序列包括：用户阅读该历史文章样本之前阅读的多篇历史文章样本分别与该历史文章样本的注意力分数。如，通过从文章发布平台的服务器中获取用户的历史阅读日志，得到用户阅读不同文章时，通过点击等方式选择阅读的其他文章，从而统计出从各篇文章触发阅读其他文章的概率，并将相 应的概率作为注意力分数。
相应的，在该步骤S303中，计算机设备可以依据历史文章的内容特征以及该历史阅读集合中阅读时刻在该历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，并利用预先训练得到的注意力模型，分别确定每篇该待定历史文章与该历史文章的注意力分数。
可选的，文章的内容特征可以通过文章的文章向量表示。相应的，为了确定某一篇或者多篇文章M到另一篇文章N的注意力分数，可以先分别确定每篇文章(如，文章M及文章N)的内容特征，确定出该文章的文章向量。然后，将各篇文章M以及文章N的文章向量输入到该注意力模型中，以分别输出每篇文章M到文章N的注意力分数。
可以理解的是，为了便于后续比较不同待定历史文章推荐出其他历史文章的可能性的大小，该注意力模型输出的各个待定历史文章与历史文章之间的注意力程度(如注意力分数)同样可以为归一化后的结果。
其中，依据历史样本集合训练注意力模型的过程可以有多种，本申请对此不加以限制。为了便于理解，以注意力模型为通过对深度神经网络进行训练得到的为例进行说明。如，参见图4，其示出了训练注意力模型的一种流程示意图。
在S401部分中，计算机设备获取多个用户的历史文章样本集合以及每个历史文章样本集合中各篇历史文章样本各自对应的已标注的注意力分数序列。
如，对于某个用户的历史文章样本集合中包括用户先后阅读的文章S1，文章S2、文章S3和文章S4，其中，对于文章S4对应的已标注的注意力分数序列可以包括：该文章S4与文章S3的注意力分数，文章S4与文章S2的注意力分数、该文章S4与文章S1的注意力分数。而对于文章S3而言，其对应的已标注的注意力分数序列可以包括：文章S3与文章S2的注意力分数，文章S3与文章S1的注意力分数。对于文章S2和S1这种阅读时刻处于其阅读时刻之前的文章仅有一篇或者没有文章的情况，可以不设置注意力分数序列。
可以理解的是，历史文章样本对应的已标注的注意力分数序列中每个注意力分数为依据该注意力分数序列中各个注意力分数归一化之后的结果。
在步骤S402部分中，计算机设备分别将各篇历史文章样本输入到深度神经网络中进行循环训练，并依据各篇历史文章样本各自的注意力分数序列对深度神经网络模型进行循环训练，直至深度神经网络输出的各篇历史文章对应的实际注意力分数序列分别与相应的已标注的注意力分数序列之间的准确度满足预设要求。
如，以历史样本集合G中，用户在N时刻阅读的历史文章样本S m对应的已标注的注意力分数序列为例进行说明，历史文章样本S m对应的已标注的注意力分数序列为该历史文章样本S m分别与历史文章样本序列S中多篇历史文章样本之间的注意力分数，其中，历史文章样本序列S为该历史样本集合G中在N时刻之前被阅读的多篇历史文章样本所组成的序列。
相应的，历史文章样本S m对应的已标注的注意力分数序列可以表示为：A m *＝{a 1 *，a 2 *，……，a m-1 *}，其中，a j *表示历史文章样本S m与历史文章样本序列S中历史文章样本S j之间的注意力分数，j表示从1到m-1的自然数，m-1表示历史文章样本序列S中历史样本的总个数。
计算机设备将该历史文章样本S N及其对应的已标注的注意力分数序列A m *输入到该深度神经网络模型中，可以该深度神经网络模型可以输出该历史文章样本S m对应的实际注意力分数序列A m，该实际注意力分数序列A m包括该深度神经网络模型输出的该历史文章样本S m分别与历史文章样本序列S中各篇历史文章样本的注意力分数。
这样，通过比较已标注的注意力分数序列A m *与实际注意力分数序列A m的差距，可以分析该深度神经网络模型的准确度是否符合要求。可以依据预设的损失函数，以及多个历史文章样本分别对应的已标注的注意力分数序列和实际注意力分数序列，并采用梯度下降法来进行优化，直至迭代收敛。如，可以定义如下的损失函数(即公式一)：
Figure PCTCN2019086374-appb-000001
其中，m取值不同的是，A m *，A m分别表示不同历史文章样本对应的已标注的注意力分数，与实际注意力分数，n为训练用的历史文章样本的总个数。
相应的，假设优化目标为minL(A，A *)，则计算机设备可以使用梯度下降的方法对该优化目标进行优化，直至最终迭代收敛，则训练结束。训练出的深度神经网络模型可以作为注意力模型。
其中，训练得到的注意力模型同样是根据不同文章的内容特征，来计算文章之间的相似度，并基于相似度，确定文章的注意力程度。
举例说明，以需要确定一个文章序列(集合)D＝{d 1，d 2，……d i……d t-2，d t-1}中任一篇文章d i分别与文章d t的注意力程度为例进行说明。其中，d i表示文章序列D中任意一篇文章，i为从1到t-1的自然数，t-1为文章序列D中文章的总数量。为了确定文章序列D中任意一篇文章d i与文章d t之间的注意力程度，注意力模型中该文章序列D中文章d i与文章d t的注意力分数的函数关系可以表示为如下公式二：
Figure PCTCN2019086374-appb-000002
其中，
Figure PCTCN2019086374-appb-000003
其中，
Figure PCTCN2019086374-appb-000004
为文章d i的文章向量，
Figure PCTCN2019086374-appb-000005
为文章d t的文章向量。W，U，b均为注意力模型中设定的参数，这些参数的参数值在训练过程中确定。F表示一种设定的函数关系，该函数关系同样可以在训练过程中确定。
如，
Figure PCTCN2019086374-appb-000006
其中，e i求取的是文章d i与文章d t之间文章向量的相似性，而通过公式二进行归一化之后，可以得到文章d i与文章d t的注意力分数。
其中，对于任意一篇文章，基于文章的内容特征确定该文章的文章向量的方式也可以有多种。可选的，计算机设备可以是将该文章中的内容进行分词，并将该文章分词出的多个词语输入到预先训练得到的文章向量模型中，以通过文章向量模型输出该文章的文章向量。
如，以文章向量模型为长短期记忆网络(Long Short-Term Memory，LSTM)为例进行说明。为了确定文章d的文章向量，首先，将文章的内容进行分词，假设文章d分词出N个词，则文章可以表示为d＝{w 1,w 2……w N}，其中，w i表示文章d中分词出的第i个分词，i为从1到N的自然数。然后，将这N个分词依次输入到该LSTM网络模型中，从而可以通过该LSTM网络模型最终输出该文章d的文章向量。
为了便于理解，可以参见图5，其示出了LSTM网络模型将文章的LSTM将文章的多个分词转换为文章的文章向量的一种示意图。
由该图5可以看出，LSTM网络模型中的输入为文章分词出的各个分词，最终输出为文章的文章向量为h t。
S304，依据得到的多篇该待定历史文章分别与至少一篇历史文章的注意力程度，从该多篇待定历史文章中，选取出触发向该目标用户推荐文章的可能性相对较低的至少一篇待定历史文章作为推荐参考文章。
可以理解的是，在目标用户的历史阅读集合中，如果目标用户阅读某篇历史文章的情况下，触发向目标用户推荐其他文章或者说触发目标用户点击阅读其他文章的可能性较低，则说明在当前时刻之前，基于该篇历史文章向目标用户推荐文章的概率较低。而由于该篇历史文章存在于历史阅读集合中，所以该篇历史文章的内容或者其他相关信息是用户感兴趣的，但是却很少获取到相关信息的文章。在该种情况下，如果后续基于该篇历史文章向该目标用户推荐文章，必然有利于扩展用户所阅读到的文章，增加文章推荐的多样性。
举例说明，假设目标用户的历史阅读集合中，存在属于时尚类、文艺类、新闻资讯类、科学类等类型的文章，而对于属于同一类型或者内容相似的两篇或者多篇文章而言，通过前面步骤，确定出属于同一类型的两篇 或多篇文章之间的注意力程度会相对较高。因此，如果确定某篇待定历史文章与多个历史文章之间的注意力程度相对较高，则说明该篇待定历史文章触发推荐文章的可能性高，而且历史阅读集合中与该待定历史文章的类型或者内容相似的文章也相对较多；反之，如果某篇待定历史文章与其他文章之间的注意力程度较低，则说明目标用户阅读过该篇待定历史文章相关类型或者内容的文章较少，因此，可以以该篇待定历史文章为推荐参考文章，以便增加向该目标用户推荐文章的多样性。
可以理解的是，由于每篇历史文章均对应的一个或者多个待定历史文章，因此，在该步骤S303中可以得到多篇待定历史文章。同时，一篇历史文章可能会作为一篇或者多篇其他历史文章的待定历史文章，相应的，一篇待定历史文章可能会分别确定出与多篇历史文章的注意力程度，这样，针对每篇待定历史文章，可以得到该待定历史文章与至少一篇历史文章的注意力程度。因此，针对每篇待定历史文章，综合该待定历史文章与对应的各篇历史文章之间的注意力程度，可以得到基于待定历史文章向目标用户推荐文章的可能性。
如，对于文章4，可能会分别将文章3、文章2和文章1作为待定历史文章，并分别确定文章3、文章2以及文章1与文章4之间的注意力程度，同时，对于文章3也可能会将文章2和文章1作为待定历史文章，并分别确定出文章2和文章1与文章3的注意力程度。相应的。对于文章3，仅仅确定出文章3与文章4之间的注意力程度，而该注意力程度可以用于反映基于文章3触发向用户推荐文章的可能性。而对于文章2，则依据文章2与文章4之间的注意力程度，以及文章2与文章3的注意力程度，来综合确定基于该文章2，触发向用户推荐文章的可能性。
其中，对于任意一篇待定历史文章，确定基于该待定历史文章向用户推荐文章的可能性的方式可以有多种。
如，在一种实现方式中，针对每篇待定历史文章，根据该待定历史文章分别与至少一篇历史文章(该待定历史文章对应的至少一篇历史文章)的注意力程度，确定该待定历史文章与该至少一篇历史文章的平均注意力 程度。相应的，该平均注意力程度可以反映基于该待定历史文章，向用户推荐文章的可能性，其中，该平均注意力程度越低，则触发向用户推荐文章的可能性也越低。
计算机设备可以从多篇该历史文章对应的多篇待定历史文章中，选取出平均注意力程度较低的至少一篇待定历史文章作为推荐参考文章，以便后续基于该推荐参考文章，确定需要向该目标用户推荐的文章。如，选取平均注意力程度最低的待定历史文章作为推荐参考文章。
例如，在对多篇待定历史文章中的每一篇待定历史文章都确定出一个第一注意力程度之后，可以对每一篇待定历史文章的第一注意力程度进行排序并获得排序结果。若是对第一注意力程度由大到小进行排序，则排序结果中越靠前的第一注意力程度对应的待定历史文章所对应的类型或内容是用户经常阅读的，而排序结果中越靠后的第一注意力程度对应的待定历史文章的类型或内容是用户不常阅读到的。此时可以根据需要确定推荐参考文章，如获取排序结果中最后一个第一注意力程度对应的待定历史文章，并将该待定历史文章作为推荐参考文章。则可以根据推荐参考文章向用户推荐不常阅读的文章类型。具体选择的推荐参考文章的数量可以为用户所阅读的多篇历史文章的一定百分比，如30％。
为了便于理解，以基于待定历史文章的平均注意力程度，计算机设备从多篇待定历史文章中选取推荐参考文章为例，对本申请从历史文章集合中确定出推荐参考文章的过程进行举例说明。如，参见图6，其示出了确定推荐参考文章的一种实现原理示意图。
在图6中以目标用户的历史阅读集合中包括文章d1、文章d2、文章d3、文章d4、文章d5和文章d6为例。在图6中以每篇文章选取在该历史文章之前阅读的三篇文章作为待定历史文章为例，同时，为了便于描述，选取与文章的阅读时刻相对较近的三篇文章作为待定历史文章。
如，图6中“待定历史文章选取”部分中，文章d6对应的待定历史文章包括：文章d5、文章d4和文章d3。文章d5对应的待定历史文章包括：文章d4、文章d3和文章d2。文章d4对应的待定历史文章包括：文 章d3、文章d2和文章d1。由于在文章d3和文章d2之前被阅读的文章都不足三篇，因此，文章d3对应的待定历史文章只有文章d2和文章d2；相应的，文章d2对应的待定历史文章为文章d1，而文章d1之前不存在其他被阅读的文章，所以针对文章d1不需要分析其对应的待定历史文章。
相应的，针对每篇历史文章，计算机设备需要分别计算该历史文章与该历史文章对应的至少一篇待定历史文章之间的注意力分数，从而该历史文章对应的注意力序列。参见图6中从上到下第一个箭头指向的部分图像，在图6中该部分，通过连线“-”连接两篇文章的标识来表示出两篇文章之间的注意力分数。如图6中所示，对于文章d6，需要计算文章d6与文章d5之间的注意力分数，在图6中其表示为d6-d5，注意力分数为0.6；相应的，文章d6与文章d4之间的注意力分数d6-d4为0.1；文章d6与文章d3之间的注意力分数为d6-d3为0.3。对于其他文章d5、d4、d3、d2也类似，具体可以参见图6中“注意力分数序列”部分。
结合图6中从上到下的第二个箭头指向的部分图像中示出的了待定历史文章与历史文章之间的对应关系，由此可以看出，每篇文章可能会作为多篇其他文章的待定历史文章。在该部分中示出了作为待定历史文章的每篇文章所对应的多篇历史文章，及该待定历史文章与各篇历史文章之间的注意力分数。如，文章d5仅仅是文章d6对应的待定历史文章，而该文章d5与文章d6的注意力分数d6-d5可以从文章d6对应的注意力分数序列中查出，即文章d5与文章d6的注意力分数d6-d5为0.6。相应的，文章d5对应的平均注意力分数为0.6。
又如，文章d4作为文章d6以及文章d5的待定历史文章，其中，文章d4与文章d6的注意力分数d6-d4为0.1；而文章d4与文章d5的注意力分数d5-d4为0.4，由此可知，该待定历史文章d4对应的平均注意力分数为0.25。
与此类似，由图6可以看出，得到作为待定历史文章的文章d3对应的平均注意力分数为0.4；作为待定历史文章的文章d2对应的平均注意力0.33；作为待定历史文章的文章d1对应的平均注意力分数为0.56。
可见，对于所有的待定历史文章，文章d4对应的平均注意力分数最低，则将文章d4确定为推荐参考文章，以便后续以文章d4为依据，确定需要推荐给目标用户的候选推荐文章。
在另一种实现方式，计算机设备可以默认每篇待定历史文章对应着预设数量篇历史文章，如果某篇待定历史文章对应的历史文章的数量不足该预设数量，而预设数量与根据待定历史文章对应的历史文章的数量之间的差值，补充该差值个默认注意力程度(如，注意力分数)。然后，分别计算待定历史文章与预设数量篇历史文章之间的注意力程度的加和，并将加和后的注意力程度作为反映该待定历史文章具备触发推荐文章的可能性的依据。
举例说明，仍以图6为例，假设需要确定出每篇待定历史文章与3篇历史文章之间的注意力分数，并假设默认注意力分数为0.5。在该图6中文章d5仅仅是文章d6对应的待定历史文章，因此，文章d5对应的历史文章的数量不足预设数量，即3篇。在该种情况下，文章d5对应的注意力程度的加和可以为文章d5与文章d6的注意力分数与两个默认注意力分数加和，具体为：0.6+0.5+0.5等于1.6。而文章b3对应的历史文章有三篇，即为文章b6、文章b5和文章b4，在该种情况下，可以直接将文章b3分别与这两篇历史文章的注意力分数加和即可，即文章b3对应的加和后的注意力分数为：0.3+0.5+0.4＝1.2。由此可以看出，由文章b3触发推荐文章的可能性，小于由文章b5触发推荐文章的可能性。
S305，计算机设备分别根据每篇该推荐参考文章的内容特征，从可推荐文章集合中，确定待推荐给该目标用户的至少一篇候选推荐文章。
其中，可推荐文章集合中包括能够推荐给该目标用户的多篇文章，如针对一个文章发布平台，该可推荐文章集合可以为文章发布平台所能发布的所有文章的集合；又如，该可推荐文章集合也可以是文章发布平台中所有用户阅读过的文章所构成的集合，当然，还可能有其他可能情况，在此不加以限制。
其中，针对每篇推荐参考文章，基于该推荐参考文章，计算机设备确 定待推荐给该目标用户的候选推荐文章的具体方式可以有多种。如，在一种可能的实现方式中，计算机设备可以依据推荐参考文章的内容特征，并结合文章协同推荐算法、基于内容的推荐算法或者基于序列的推荐算法等，从该可推荐文章集合中，确定出至少一篇候选推荐文章。
可选的，为了使得确定出的候选推荐文章能够与该待推荐参考文章的类型或者内容的相关程度更高，计算机设备还可以基于注意力的推荐策略，确定候选推荐文章。其中，基于注意力的推荐策略为：针对每篇推荐参考文章，可以依据该推荐参考文章的内容特征以及可推荐文章集合中每篇文章的内容特征，分别确定该推荐参考文章与该可推荐文章集合中每篇文章的注意力程度。相应的，可以从该可推荐文章集合中，选取出与该推荐参考文章的注意力程度较高的至少一篇候选推荐文章。
在确定候选推荐文章时，也可以确定出多篇候选推荐文章。例如，在计算可推荐文章集中每篇文章与至少一篇的推荐参考文章的第二注意力程度后，可以对得到的第二注意力程度进行排序，第二注意力程度越高则表示用户阅读推荐参考文章后，阅读该篇文章的可能性越大。若是从可推荐文章集合中确定出两篇候选推荐文章，则可以从可推荐文章集合中选择第二注意力程度最大的文章与第二注意力程度第二大的文章。
其中，计算机设备确定推荐参考文章与可推荐文章集合中各篇文章的注意力程度的方式可以参考前面确定注意力程度的相关介绍，如，可以将可推荐文章集合的各篇文章的文章向量，以及该推荐参考文章的文章向量输入到预先训练出的注意力模型中，以输出该推荐参考文章与该可推荐文章集合中各篇文章之间的注意力程度(如，注意力分数)所构成的注意力分数序列。当然，前面描述的通过其他方式确定注意力程度的方式也同样适用于此处，在此不再赘述。
可见，在本申请实施例中，对于用户阅读过的历史文章，计算机设备会分析用户在阅读该历史文章之前阅读的其他历史文章与该历史文章的注意力程度，由于两篇文章之间的注意力程度可以反映出用户阅读一篇文章的情况下，推荐用户阅读另一篇文章的可能性，因此，依据确定出的各 篇文章与其他文章之间的注意力程度，可以从阅读历史集合中确定出表征向目标用户推荐文章的可能性较低的历史文章，而该历史文章的类型及内容便不属于用户经常阅读到的文章类型及文章内容，因此，以该历史文章作为推荐参考文章，向用户推荐候选推荐文章，有利于提高确定出的候选推荐文章的多样性，进而有利于提高向用户推荐文章的多样性。
可以理解的是，基于推荐参考文章，确定出候选推荐文章之后，计算机设备可以直接将候选推荐文章作为需要推荐给该目标用户的文章。然而，确定出的考虑到候选推荐文章的数量有可能会相对较大；而且，在实际应用中，可能会配置多种推荐策略，并分别依据不同的推荐策略，确定与该推荐参考文章匹配的多篇候选推荐文章，也会使得确定出的候选推荐文章的数量较大，如果将这些大量的候选推荐文章都推荐给用户，很难实现精准向用户推荐用户感兴趣的文章，从而影响到推荐效果。
可选的，在依据可推荐文章集合中各文章与推荐参考文章之间的注意力程度，基于内容推荐算法，协同推荐算法等方式中的一种或者多种，确定出至少一个候选推荐文章之后，计算机设备还可以针对每篇候选推荐文章，分别计算历史阅读集合中每篇历史文章与该候选推荐文章的注意力程度，得到候选推荐文章与该历史阅读集合中多篇历史文章之间的注意力程度序列。然后，分别计算每篇候选推荐文章对应的注意力程度序列的信息熵，得到每篇候选推荐文章各自对应的信息熵。相应的，可以从该至少一篇候选推荐文章中，选取出信息熵较小的至少一篇候选推荐文章作为待推荐给该目标用户的至少一篇目标推荐文章。
可选地，在确定出每一篇候选推荐文章的信息熵后，可以对信息熵进行排序，信息熵越小，则对应的候选推荐文章越优先推送。如需要推送两篇目标推荐文章时，则对候选推荐文章的信息熵进行排序，并获取信息熵最小的候选推荐文章与信息熵第二小的候选推荐文章，将该两篇候选推荐文章作为目标推荐文章推送给用户。
可以理解的是，历史阅读集合中每篇历史文章与该候选推荐文章的注 意力程度可以反映出，由各篇历史文章触发阅读该候选推荐文章的可能性；而，该候选推荐文章与该历史阅读集合中各篇历史文章之间的注意力程度序列对应的信息熵可以衡量该历史阅读集合中各篇历史文章触发阅读到该候选推荐文章的稳定性或者说可靠性。可见，通过选取对应的信息熵较小的候选推荐文章作为待推荐的目标推荐文章，有利于在增加向用户推荐文章多样性的前提下，最大程度的保证向用户推荐用户感兴趣的文章，以实现基于用户进行个性化推荐。
计算机设备确定出至少一篇目标推荐文章之后，还可以确定这至少一篇目标推荐文章的排序，以便后续文章发布平台按照该排序输出该至少一篇目标推荐文章。
如，参见图7，其示出了本申请实施例一种文章推荐方法的又一种流程示意图，本实施例的方法从计算机设备侧描述，该方法可以包括：
S701，计算机设备获取待分析的目标用户的历史阅读集合。
其中，该历史阅读集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章。
S702，针对每篇历史文章，依据该历史文章的内容特征，计算机设备确定该历史文章的文章向量，并依据该历史阅读集合中阅读时刻在该历史文章的阅读时刻之前的至少一篇待定历史文章各自的内容特征，分别确定每篇待定历史文章的文章向量。
S703，针对每篇历史文章，计算机设备将该历史文章对应的至少一篇待定历史文章的文章向量以及该历史文章的内容向量输入到预先训练出的注意力模型中，以通过该注意力模型输出注意力分数序列。
其中，该注意力分数序列包括该至少一篇待定历史文章分别与该历史文章之间的注意力分数。待定历史文章与该历史文章的注意力分数可以反映在用户阅读待定历史文章的情况下，推荐用户阅读该历史文章的可能性。
S704，针对每篇该待定历史文章，计算机设备根据该待定历史文章分别到至少一篇历史文章的注意力分数，确定该待定历史文章到至少一篇历史文章的平均注意力分数。
S705，计算机设备从多篇该历史文章对应的多篇待定历史文章中，选取出平均注意力分数较低的至少一篇待定历史文章作为推荐参考文章。
以上步骤S701到S705可以参见前面的相关介绍，在此不再赘述。其中，步骤S704和S705仅仅是基于待定历史文章与相应的历史文章之间的注意力分数，确定推荐参考文章的一种实现方式，对于前面实施例提到的其他方式，也同样适用于本实施例，在此不再赘述。
可以理解的是，以上步骤S701到S705实际上是基于注意力模型的选取策略，即，基于注意力模型，从目标用户的阅读历史集合中选取推荐参考文章。但是可以理解的是，在实际应用中，基于注意力模型，来选取推荐参考文章可以仅仅是选取待推荐文章的一种选取策略，因此，在计算机设备中除了配置基于注意力模型的选取策略之后，还可以同时配置其他一种或多种选取策略，如基于用户画像的选取策略，即，基于目标用户的用户画像，从该阅读历史集合中选取出一篇或多篇历史文章作为推荐参考文章。相应的，计算机设备可以分别通过配置的多种选取策略，分别选取出的推荐参考文章，并执行后续的操作。
如，参见图8，本申请中文章推荐方法的一种实现原理示意图。由图8可以看出，用于选取推荐参考文章的推荐策略可以有多种。相应的，依据目标用户阅读的阅读历史集合以及用户的目标用户的用户画像，并分别采用多种不同的推荐策略，从阅读历史集合中确定推荐参考文章，这样，基于每种推荐策略可以均可以选取出一篇推荐参考文章。
S706，针对每篇推荐参考文章，计算机设备依据该推荐参考文章的内容特征以及可推荐文章集合中每篇文章的内容特征，分别确定可推荐文章集合中每篇文章与推荐参考文章的第二注意力程度。
其中，可推荐文章集合中任意一篇文章与推荐参考文章之间第二注意力程度表征用户在阅读推荐参考文章的情况下，推荐用户阅读该可推荐文章集合中该文章的可能性。
S707，从该推荐文章集合中，计算机设备选取出与该荐参考文章的第二注意力程度较高的至少一篇候选推荐文章。
可以理解的是，文章与推荐参考文章的注意力程度越高，则说明文章与该推荐参考文章的类型及内容的关联性越高，则该文章属于用户阅读过的但是又不经常阅读过的文章类型，将该文章选取为候选推荐文章，则既能贴合用户的阅读需求，又有可能增加用户阅读的文章的多样性。
在以上步骤S706至S707中实际上是基于注意力的推荐策略，来确定候选推荐文章为例进行说明，但是可以理解的通过其他的推荐策略，来确定候选推荐文章的方式也同样适用于本实施例。
可以理解的是，在实际应用中，计算机设备基于推荐参考文章，确定候选推荐文章时，可以仅仅设置一种推荐策略，如上面提到的基于注意力的推荐策略。在实际应用中，可能会设置多种推荐策略，如，设置基于注意力的推荐策略的同时，还会设置基于内容的推荐策略，以及基于协同的推荐策略等等。如图8，推荐策略也可以配置有多种。相应的，计算机设备可以分别以各篇推荐参考文章为基准，并分别多种推荐策略，从可推荐文章集合中确定出相应的至少一篇候选推荐文章；或者是，综合多种推荐策略，来确定出至少一篇候选推荐文章。如图8所示，依据多种推荐策略以及确定出的推荐参考文章，可以从推荐文章集合中选取候选推荐文章，以便执行后续的初选操作。
S708，针对每篇候选推荐文章，计算机设备分别计算该历史阅读集合中每篇历史文章与该候选推荐文章的第三注意力程度，得到候选推荐文章与该历史阅读集合中多篇历史文章之间的注意力程度序列。
其中，该注意力程度序列中包括候选推荐文章分别与该历史阅读集合中多篇历史文章的第三注意力程度。
S709，计算机设备分别计算每篇候选推荐文章对应的注意力程度序列的信息熵，得到每篇候选推荐文章各自对应的信息熵。
其中，该注意力程度序列对应的信息熵可以反映出该候选推荐文章与历史阅读集合中各篇历史文章之间注意力程度的分布情况。候选推荐文章对应的注意力程度序列的信息熵越小，则说明根据用户阅读的历史阅读集合，确定出用户对该候选推荐文章的感兴趣程度越高。
其中，计算注意力程度序列的信息熵的方式可以有多种，本申请对此不加以限制。
为了便于理解，以一种计算信息熵的方式进行介绍。如，假设候选推荐文章与历史文章集合的注意力程度序列为一个注意力分数序列，其表示为A t＝{a 1,a 2……a t}，则该注意力分数序列对应的信息熵H(A)可以通过如下公式计算得到：
Figure PCTCN2019086374-appb-000007
其中，公式四中的a i为注意力分数序列A t中的注意力分数。
S710，从该至少一篇候选推荐文章中，选取出信息熵较小的至少一篇候选推荐文章作为待推荐给该目标用户的至少一篇目标推荐文章。
计算机设备通过选取对应的信息熵较小的候选推荐文章作为目标推荐文章，可以使得筛选出的目标推荐文章在满足多样性的前提下，更符合用户的兴趣。
可以理解的是，以上步骤S708到S710是基于注意力模型，从候选推荐文章中，确定最终筛选推荐的目标推荐文章的方式，即，基于注意力模型的初选策略。但是可以理解的是，通过其他初选策略，从多篇候选推荐文章中最终筛选出至少一篇目标推荐文章也同样适用于本实施例。如，初选策略可以为基于用户画像，来从多篇候选推荐文章中选取与用户画像最为匹配的至少一篇目标推荐文章。
在实际应用中，还可以在基于注意力模型的初选策略的基础上，综合其他初选策略，以最终从候选推荐文章中筛选出至少一篇目标推荐文章。如图8所示，计算机设备中可以配置多种初选策略。相应的，计算机设备可以分别基于该多种初选策略，从候选推荐文章中确定出至少一篇目标推荐文章；或者是，结合多种初选策略，来综合确定出至少一篇候选推荐文章。
S711，针对任意一篇推荐参考文章，计算机设备确定该至少一篇目标推荐文章分别与推荐参考文章的第四注意力程度。
其中，第四注意力程度反映用户在阅读该推荐参考文章的情况下，推荐用户阅读该目标推荐文章的可能性。
S712，依据该至少一篇目标推荐文章分别与推荐参考文章之间的第四注意力程度，计算机设备确定该至少一篇目标推荐文章的推荐顺序。
如，目标推荐文章对应的第四注意力程度越高，该目标推荐文章的排序越靠前。特别的，当存在多篇推荐参考文章的情况下，可以依据目标推荐文章分别与各篇推荐参考文章的第四注意力程度，来确定该目标推荐文章的第四注意力程度的综合值，并根据各个目标推荐文章的该综合值，来确定推荐顺序。
其中，目标推荐文章对应的第四注意力程度越高，则在用户阅读该推荐参考文章的情况下，选择阅读该目标推荐文章的可能性越大，因此，对应的第四注意力程度较高的目标推荐文章的排在较为靠前的顺序上，越有利于用户点击该目标推荐文章。
可以理解的是，计算机设备确定出该至少一篇目标推荐文章的推荐顺序之后，可以将该至少一篇目标推荐文章的标识以及推荐顺序发送给文章发布服务器。相应的，在目标用户访问文章发布服务器，如，登录文章发布服务器阅读文章等，则该文章发布服务器可以依据该推荐顺序，向该目标用户展现该至少一篇目标推荐文章，以使得目标用户可以看到更为广泛类型且较为符合自身兴趣的文章。
可以理解的是，第二注意力程度，第三注意力程度、第四注意力程度仅仅是为了便于区分不同文章之间的注意力程度，但是其求取过程都可以参见前面介绍的用于确定文章之间的注意力程度的方式，如，利用预先训练得到的注意力模型来确定以上的第二注意力程度、第三注意力程度及第四注意力程度，具体可以参见前面实施例的相关介绍，在此不再赘述。
另一方面，本申请还提供了一种文章推荐方法，可选地，如图9所示，上述文章推荐方法包括：
S902，计算机设备获取待分析的目标用户的历史阅读集合，历史阅读 集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章；
S904，计算机设备针对多篇历史文章中的每篇历史文章，确定阅读时刻在每篇历史文章之前的每篇待定历史文章与每篇历史文章的第一注意力程度，第一注意力程度反映在用户阅读待定历史文章的情况下，推荐用户阅读每篇历史文章的可能性；
S906，计算机设备依据每篇待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇待定历史文章中选取出至少一篇推荐参考文章；
S908，计算机设备分别根据每篇推荐参考文章，从可推荐文章集合中，确定待推荐给目标用户的至少一篇候选推荐文章。
以下将上述文章推荐方法应用到具体的推荐过程中进行说明。计算机设备为手机。如图10所示，图10为一种可选的用户的浏览界面，用户可以使用终端1002登录客户端浏览文章，图10中显示有客户端根据现有的推荐方法已经推荐的文章1004。用户可以通过点击相关的文章标题或者链接阅读一些文章。在用户的阅读过程中，会产生一些阅读记录，阅读记录包括阅读的文章与阅读时刻。将阅读记录作为用户的历史阅读集合。历史阅读集合中包括了用户阅读过的多篇历史文章。
在获取到上述多篇历史文章之后，手机会获取多篇历史文章中每一篇历史文章的待定历史文章。待定历史文章为阅读时刻在上述每一篇历史文章之前的文章。例如，以多篇历史文章包括“我们是冠军”与“赛程表”、“主演表”、“古装”为例，四篇历史文章的阅读顺序从先到后为：“我们是冠军”、“赛程表”、“主演表”、“古装”。“古装”的待定历史文章为位于“古装”之前的三篇历史文章；“主演表”的待定历史文章为位于“主演表”之前的两篇历史文章；“赛程表”的待定历史文章为位于“赛程表”之前的一篇历史文章；“我们是冠军”没有待定历史文章。手机确定出每一篇历史文章的待定历史文章之后，确定每一篇待定历史文章与对应的历史文章的第一注意力程度。如表(1)，表1中的横线“-”表示不包含第一注意力程度。
表(1)
Figure PCTCN2019086374-appb-000008
确定第一注意力程度的方法可以采用上述提到的各种方法，在此不再赘述。可见，在用户阅读“我们是冠军”后，阅读“赛程表”的可能性较高，而阅读“主演表”的可能性较低。而对于每一篇待定历史文章，如“我们是冠军”，可以计算“我们是冠军”和与其对应的至少一篇的历史文章的第一注意力程度。如上述计算的“我们是冠军”与“赛程表”的第一注意力程度为0.9、“我们是冠军”与“演员表”的第一注意力程度为0.4、“我们是冠军”与“古装”的第一注意力程度为0.2，得到一个平均值为0.5。采用同样的方法，得到每一篇待定历史文章的第一注意力程度的平均值，“赛程表”的第一注意力平均值为0.2，“主演表”的第一注意力程度为0.8。
手机将多篇待定历史文章中，第一注意力平均值最低的“赛程表”确定为推荐参考文章，然后根据推荐参考文章“赛程表”从可推荐文章集合中推荐出一篇候选推荐文章。如，可推荐文章集合中包括“赛程时间”、“导演表”，则可以将“赛程时间”作为候选推荐文章，在用户获取推荐文章时，将候选推荐文章推荐给用户。如图11所示，图11中用户通过下拉的方法请求推荐文章，如图12所示，手机将“赛程时间”推荐给用户，显示在推荐区1202中。
另一方面，本申请还提供了一种文章推荐装置。如，参见图13，其示出了本申请一种文章推荐装置一个实施例的组成结构示意图，本实施例的装置可以应用于前面的计算机设备，该装置可以包括：
历史获取单元1301，设置为获取待分析的目标用户的历史阅读集合，历史阅读集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章；
第一注意分析单元1302，设置为针对每篇历史文章，依据历史文章的内容特征以及历史阅读集合中阅读时刻在历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，分别确定每篇待定历史文章与历史文章的第一注意力程度，第一注意力程度反映在用户阅读待定历史文章的情况下，推荐用户阅读历史文章的可能性；
推荐参考确定单元1303，设置为依据得到的多篇待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇待定历史文章中，选取出触发向目标用户推荐文章的可能性相对较低的至少一篇待定历史文章作为推荐参考文章；
候选推荐确定单元1304，设置为分别根据每篇推荐参考文章的内容特征，从可推荐文章集合中，确定待推荐给目标用户的至少一篇候选推荐文章。
可选的，第一注意分析单元可以包括：
第一注意分析子单元，设置为依据历史文章的内容特征以及历史阅读集合中阅读时刻在历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，并利用预先训练得到的注意力模型，分别确定每篇待定历史文章与历史文章的注意力分数；
其中，注意力模型为利用多个用户的历史文章样本集合以及每个历史文章样本集合中每篇历史文章样本各自对应的已标注的注意力分数序列训练的到的，其中，历史文章样本的注意力分数序列包括：用户阅读该历史文章样本之前阅读的多篇历史文章样本分别与该历史文章样本的注意力分数。
可选的，第一注意力分析子单元可以包括：
第一向量确定子单元，设置为依据历史文章的内容特征，确定历史文章的文章向量；
第二向量确定子单元，设置为依据历史阅读集合中阅读时刻在历史文章的阅读时刻之前的至少一篇待定历史文章各自的内容特征，分别确定每篇待定历史文章的文章向量；
注意分数分析子单元，设置为根据历史文章的文章向量以及至少一篇待定历史文章各自的文章向量，并利用注意力模型，分别确定每篇待定历史文章与历史文章的注意力分数。
在一种可能的实现方式中，推荐参考确定单元，包括：
平均注意分析子单元，设置为针对每篇待定历史文章，根据待定历史文章分别与至少一篇历史文章的第一注意力程度，确定待定历史文章到至少一篇历史文章的平均注意力程度；
参考筛选子单元，设置为从多篇历史文章对应的多篇待定历史文章中，选取出平均注意力程度较低的至少一篇待定历史文章作为推荐参考文章。
在一种可能的实现方式中，候选推荐确定单元，包括：
第二注意分析单元，设置为针对每篇推荐参考文章，依据推荐参考文章的内容特征以及可推荐文章集合中每篇文章的内容特征，分别确定可推荐文章集合中每篇文章与推荐参考文章的第二注意力程度，第二注意力程度反映用户在阅读推荐参考文章的情况下，推荐用户阅读可推荐文章集合中文章的可能性；
候选推荐选取单元，设置为从可推荐文章集合中，选取出与推荐参考文章的第二注意力程度较高的至少一篇候选推荐文章。
如，参见图14，其示出了本申请的文章推荐装置的又一种组成结构示意图，本实施例的文章推荐装置与前面图14所示装置的不同之处在于：
该装置除了包括，历史获取单元1401，第一注意分析单元1402，推荐参考确定单元1403和候选推荐确定单元1404之外，还包括：
第三注意分析单元1405，设置为在候选推荐确定单元确定出至少一篇候选推荐文章之后，针对每篇候选推荐文章，分别计算历史阅读集合中每篇历史文章与候选推荐文章的第三注意力程度，得到候选推荐文章与历史阅读集合中多篇历史文章之间的注意力程度序列，第三注意力程度反映用户在阅读历史阅读集合中的历史文章的情况下，推荐用户阅读候选推荐文章的可能性；
信息熵计算单元1406，设置为分别计算每篇候选推荐文章对应的注意力程度序列的信息熵，得到每篇候选推荐文章各自对应的信息熵；
目标推荐确定单元1407，设置为从至少一篇候选推荐文章中，选取出信息熵较小的至少一篇候选推荐文章作为待推荐给目标用户的至少一篇目标推荐文章。
可选的，该装置还可以包括：
第四注意分析单元1408，设置为分别确定推荐参考文章与至少一篇目标推荐文章的第四注意力程度，第四注意力程度反映用户在阅读推荐参考文章的情况下，推荐用户阅读目标推荐文章的可能性；
排序确定单元1409，设置为依据推荐参考文章分别与至少一篇目标推荐文章的第四注意力程度，确定至少一篇目标推荐文章的推荐顺序。
其中，历史获取单元1401，第一注意分析单元1402，推荐参考确定单元1403和候选推荐确定单元1404具体可以参见前面实施例的相关介绍，在此不在赘述。
根据本申请的实施例的又一方面，还提供了一种存储介质，该存储介质中存储有计算机程序，其中，该计算机程序被设置为运行时执行上述任一项方法实施例中的步骤。
可选地，在本实施例中，上述存储介质可以被设置为存储用于执行以下步骤的计算机程序：
S1，获取待分析的目标用户的历史阅读集合，历史阅读集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章；
S2，针对每篇历史文章，依据历史文章的内容特征以及历史阅读集合中阅读时刻在历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，分别确定每篇待定历史文章与历史文章的第一注意力程度，第一注意力程度反映在用户阅读待定历史文章的情况下，推荐用户阅读历史文章的可能性；
S3，依据得到的多篇待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇待定历史文章中，选取出触发向目标用户推荐文章的可能性相对较低的至少一篇待定历史文章作为推荐参考文章；
S4，分别根据每篇推荐参考文章的内容特征，从可推荐文章集合中，确定待推荐给目标用户的至少一篇候选推荐文章。
或者，在本实施例中，上述存储介质可以被设置为存储用于执行以下步骤的计算机程序：
S1，计算机设备获取待分析的目标用户的历史阅读集合，历史阅读集合包括：目标用户在不同阅读时刻阅读过的多篇历史文章；
S2，计算机设备针对多篇历史文章中的每篇历史文章，确定阅读时刻在每篇历史文章之前的每篇待定历史文章与每篇历史文章的第一注意力程度，第一注意力程度反映在用户阅读待定历史文章的情况下，推荐用户阅读每篇历史文章的可能性；
S3，计算机设备依据每篇待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇待定历史文章中选取出至少一篇推荐参考文章；
S4，计算机设备分别根据每篇推荐参考文章，从可推荐文章集合中，确定待推荐给目标用户的至少一篇候选推荐文章。
可选地，在本实施例中，本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令终端设备相关的硬件来完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括：闪存盘、只读存储器(Read-Only Memory，ROM)、随机存取器(Random Access Memory，RAM)、磁盘或光盘等。
需要说明的是，本说明书中的各个实施例均采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似的部分互相参见即可。对于装置类实施例而言，由于其与方法实施例 基本相似，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
最后，还需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括要素的过程、方法、物品或者设备中还存在另外的相同要素。
对所公开的实施例的上述说明，使本领域技术人员能够实现或使用本申请。对这些实施例的多种修改对本领域技术人员来说将是显而易见的，本文中所定义的一般原理可以在不脱离本申请的精神或范围的情况下，在其它实施例中实现。因此，本申请将不会被限制于本文所示的这些实施例，而是要符合与本文所公开的原理和新颖特点相一致的最宽的范围。
以上仅是本申请的可选实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本申请原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本申请的保护范围。
工业实用性
在本申请实施例中，对于用户阅读过的历史文章，会分析用户在阅读该历史文章之前阅读的其他历史文章与该历史文章的注意力程度，由于两篇文章之间的注意力程度可以反映出用户阅读一篇文章的情况下，推荐用户阅读另一篇文章的可能性，因此，依据确定出的各篇文章与其他文章之间的注意力程度，可以从阅读历史集合中确定出表征向目标用户推荐文章的可能性较低的历史文章，而该历史文章的类型及内容便不属于用户经常阅读到的文章类型及文章内容，因此，以该历史文章作为推荐参考文章， 向用户推荐候选推荐文章，有利于提高确定出的候选推荐文章的多样性，进而有利于提高向用户推荐文章的多样性。

Claims (15)

  1. 一种文章推荐方法，包括：
    计算机设备获取待分析的目标用户的历史阅读集合，所述历史阅读集合包括：所述目标用户在不同阅读时刻阅读过的多篇历史文章；
    所述计算机设备针对所述多篇历史文章中的每篇历史文章，确定阅读时刻在所述每篇历史文章之前的每篇待定历史文章与所述每篇历史文章的第一注意力程度，所述第一注意力程度反映在用户阅读所述待定历史文章的情况下，推荐用户阅读所述每篇历史文章的可能性；
    所述计算机设备依据每篇所述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇所述待定历史文章中选取出至少一篇推荐参考文章；
    所述计算机设备分别根据每篇所述推荐参考文章，从可推荐文章集合中，确定待推荐给所述目标用户的至少一篇候选推荐文章。
  2. 根据权利要求1所述的文章推荐方法，其中，所述计算机设备针对所述多篇历史文章中的每篇历史文章，确定阅读时刻在所述每篇历史文章之前的每篇待定历史文章与所述每篇历史文章的第一注意力程度包括：
    所述计算机设备依据所述每篇历史文章的内容特征以及所述历史阅读集合中阅读时刻在所述每篇历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，并利用预先训练得到的注意力模型，分别确定每篇所述待定历史文章与所述每篇历史文章的注意力分数；
    其中，所述注意力模型为利用多个用户的历史文章样本集合以及每个所述历史文章样本集合中每篇历史文章样本各自对应的已标注的注意力分数序列训练得到的，其中，历史文章样本的注意力分数序列包括：用户阅读该历史文章样本之前阅读的多篇历史文章样本分别与该历史文章样本的注意力分数。
  3. 根据权利要求2所述的文章推荐方法，其中，所述计算机设备利用预先训练得到的注意力模型，分别确定每篇所述待定历史文章与所述历史 文章的注意力分数，包括：
    所述计算机设备依据所述历史文章的内容特征，确定所述历史文章的文章向量；
    所述计算机设备依据所述历史阅读集合中阅读时刻在所述历史文章的阅读时刻之前的至少一篇待定历史文章各自的内容特征，分别确定每篇待定历史文章的文章向量；
    所述计算机设备根据所述历史文章的文章向量以及所述至少一篇待定历史文章各自的文章向量，并利用所述注意力模型，分别确定每篇待定历史文章与所述历史文章的注意力分数。
  4. 根据权利要求1至3任一项所述的文章推荐方法，其中，所述计算机设备依据每篇所述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇所述待定历史文章中选取出至少一篇推荐参考文章包括：
    所述计算机设备针对每篇所述待定历史文章，根据所述待定历史文章分别与至少一篇历史文章的第一注意力程度，确定所述待定历史文章到所述至少一篇历史文章的平均注意力程度；
    所述计算机设备对所述待定历史文章的所述平均注意力程度由小到大进行排序，并获取与排序结果中前X个平均注意力程度对应的X篇待定历史文章，并将所述X篇待定历史文章确定为推荐参考文章，其中，所述X根据所述多篇历史文章的数量确定。
  5. 根据权利要求1所述的文章推荐方法，其中，所述计算机设备分别根据每篇所述推荐参考文章，从可推荐文章集合中，确定待推荐给所述目标用户的至少一篇候选推荐文章，包括：
    所述计算机设备针对每篇推荐参考文章，依据所述推荐参考文章的内容特征以及可推荐文章集合中每篇文章的内容特征，分别确定所述可推荐文章集合中每篇文章与所述推荐参考文章的第二注意力程度，所述第二注意力程度反映用户在阅读所述推荐参考文章的情况下，推荐用户阅读所述可推荐文章集合中文章的可能性；
    所述计算机设备对所述可推荐文章集合中的每篇文章的第二注意 力程度由大到小进行排序，得到排序结果，并从所述可推荐文章集合中获取与所述排序结果中前Y个第二注意力程度对应的Y篇文章，将所述Y篇文章确定为候选推荐文章，其中，所述Y根据所述可推荐文章集合中的文章的数量确定。
  6. 根据权利要求1或5所述的文章推荐方法，其中，所述计算机设备确定出至少一篇候选推荐文章之后，还包括：
    所述计算机设备针对每篇候选推荐文章，分别计算所述历史阅读集合中每篇历史文章与所述候选推荐文章的第三注意力程度，得到所述候选推荐文章与所述历史阅读集合中多篇历史文章之间的注意力程度序列，所述第三注意力程度反映用户在阅读所述历史阅读集合中的历史文章的情况下，推荐用户阅读所述候选推荐文章的可能性；
    所述计算机设备分别计算每篇候选推荐文章对应的注意力程度序列的信息熵，得到每篇候选推荐文章各自对应的信息熵；
    所述计算机设备对所述每篇候选推荐文章的信息熵由小到大进行排序，并获取排序结果中前Z个信息熵所对应的Z篇候选推荐文章，将所述Z篇候选推荐文章确定为待推荐给所述目标用户的目标推荐文章，其中，所述Z根据所述候选推荐文章的数量确定。
  7. 根据权利要求6所述的文章推荐方法，其中，在所述计算机设备对所述每篇候选推荐文章的信息熵由小到大进行排序，并获取排序结果中前Z个信息熵所对应的Z篇候选推荐文章，将所述Z篇候选推荐文章确定为待推荐给所述目标用户的目标推荐文章之后，还包括：
    所述计算机设备分别确定所述推荐参考文章与至少一篇目标推荐文章的第四注意力程度，所述第四注意力程度反映用户在阅读所述推荐参考文章的情况下，推荐用户阅读所述目标推荐文章的可能性；
    依据所述推荐参考文章分别与所述至少一篇目标推荐文章的第四注意力程度，确定所述至少一篇目标推荐文章的推荐顺序。
  8. 一种文章推荐装置，应用在计算机设备中，包括：
    历史获取单元，设置为获取待分析的目标用户的历史阅读集合， 所述历史阅读集合包括：所述目标用户在不同阅读时刻阅读过的多篇历史文章；
    第一注意分析单元，设置为针对所述多篇历史文章中的每篇历史文章，确定阅读时刻在所述每篇历史文章之前的每篇待定历史文章与所述每篇历史文章的第一注意力程度，所述第一注意力程度反映在用户阅读所述待定历史文章的情况下，推荐用户阅读所述每篇历史文章的可能性；
    推荐参考确定单元，设置为依据每篇所述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇所述待定历史文章中选取出至少一篇推荐参考文章；
    候选推荐确定单元，设置为分别根据每篇所述推荐参考文章，从可推荐文章集合中，确定待推荐给所述目标用户的至少一篇候选推荐文章。
  9. 根据权利要求8所述的文章推荐装置，其中，所述第一注意分析单元，包括：
    第一注意分析子单元，设置为依据所述每篇历史文章的内容特征以及所述历史阅读集合中阅读时刻在所述每篇历史文章的阅读时刻之前的至少一篇待定历史文章的内容特征，并利用预先训练得到的注意力模型，分别确定每篇所述待定历史文章与所述每篇历史文章的注意力分数；
    其中，所述注意力模型为利用多个用户的历史文章样本集合以及每个所述历史文章样本集合中每篇历史文章样本各自对应的已标注的注意力分数序列训练的到的，其中，历史文章样本的注意力分数序列包括：用户阅读该历史文章样本之前阅读的多篇历史文章样本分别与该历史文章样本的注意力分数。
  10. 根据权利要求8或9所述的文章推荐装置，其中，所述推荐参考确定单元，包括：
    平均注意分析子单元，设置为针对每篇所述待定历史文章，根据 所述待定历史文章分别与至少一篇历史文章的第一注意力程度，确定所述待定历史文章到所述至少一篇历史文章的平均注意力程度；
    参考筛选子单元，设置为对所述待定历史文章的所述平均注意力程度由小到大进行排序，并获取与排序结果中前X个平均注意力程度对应的X篇待定历史文章，并将所述X篇待定历史文章确定为推荐参考文章，其中，所述X根据所述多篇历史文章的数量确定。
  11. 根据权利要求8所述的文章推荐装置，其中，所述候选推荐确定单元，包括：
    第二注意分析单元，设置为针对每篇推荐参考文章，依据所述推荐参考文章的内容特征以及可推荐文章集合中每篇文章的内容特征，分别确定所述可推荐文章集合中每篇文章与所述推荐参考文章的第二注意力程度，所述第二注意力程度反映用户在阅读所述推荐参考文章的情况下，推荐用户阅读所述可推荐文章集合中文章的可能性；
    候选推荐选取单元，设置为对所述可推荐文章集合中每一篇文章的第二注意力程度由大到小进行排序，得到排序结果，并从所述可推荐文章集合中获取与所述排序结果中前Y个第二注意力程度对应的Y篇文章，将所述Y篇文章确定为候选推荐文章，其中，所述Y根据所述可推荐文章集合中的文章的数量确定。
  12. 根据权利要求8或11所述的文章推荐装置，其中，还包括：
    第三注意分析单元，设置为在所述候选推荐确定单元确定出至少一篇候选推荐文章之后，针对每篇候选推荐文章，分别计算所述历史阅读集合中每篇历史文章与所述候选推荐文章的第三注意力程度，得到所述候选推荐文章与所述历史阅读集合中多篇历史文章之间的注意力程度序列，所述第三注意力程度反映用户在阅读所述历史阅读集合中的历史文章的情况下，推荐用户阅读所述候选推荐文章的可能性；
    信息熵计算单元，设置为分别计算每篇候选推荐文章对应的注意力程度序列的信息熵，得到每篇候选推荐文章各自对应的信息熵；
    目标推荐确定单元，设置为对所述每篇候选推荐文章的信息熵由 小到大进行排序，并获取排序结果中前Z个信息熵所对应的Z篇候选推荐文章，将所述Z篇候选推荐文章确定为待推荐给所述目标用户的目标推荐文章，其中，所述Z根据所述候选推荐文章的数量确定。
  13. 根据权利要求12所述的文章推荐装置，其中，还包括：
    第四注意分析单元，设置为分别确定所述推荐参考文章与至少一篇目标推荐文章的第四注意力程度，所述第四注意力程度反映用户在阅读所述推荐参考文章的情况下，推荐用户阅读所述目标推荐文章的可能性；
    排序确定单元，设置为依据所述推荐参考文章分别与所述至少一篇目标推荐文章的第四注意力程度，确定所述至少一篇目标推荐文章的推荐顺序。
  14. 一种计算机设备，包括：
    处理器和存储器；
    其中，所述处理器用于执行所述存储器中存储的计算机程序；
    所述存储器用于存储计算机程序，所述计算机程序至少用于：
    获取待分析的目标用户的历史阅读集合，所述历史阅读集合包括：所述目标用户在不同阅读时刻阅读过的多篇历史文章；
    针对所述多篇历史文章中的每篇历史文章，确定阅读时刻在所述每篇历史文章之前的每篇待定历史文章与所述每篇历史文章的第一注意力程度，所述第一注意力程度反映在用户阅读所述待定历史文章的情况下，推荐用户阅读所述每篇历史文章的可能性；
    依据每篇所述待定历史文章分别与至少一篇历史文章的第一注意力程度，从多篇所述待定历史文章中选取出至少一篇推荐参考文章；
    分别根据每篇所述推荐参考文章，从可推荐文章集合中，确定待推荐给所述目标用户的至少一篇候选推荐文章。
  15. 一种存储介质，所述存储介质中存储有计算机程序，所述计算机程序被运行时，执行所述权利要求1至8任一项中所述的文章推荐方法。
PCT/CN2019/086374 2018-05-25 2019-05-10 文章推荐方法、装置、计算机设备及存储介质 Ceased WO2019223552A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/945,066 US11763145B2 (en) 2018-05-25 2020-07-31 Article recommendation method and apparatus, computer device, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810515869.8 2018-05-25
CN201810515869.8A CN110598086B (zh) 2018-05-25 2018-05-25 文章推荐方法、装置、计算机设备及存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/945,066 Continuation US11763145B2 (en) 2018-05-25 2020-07-31 Article recommendation method and apparatus, computer device, and storage medium

Publications (1)

Publication Number Publication Date
WO2019223552A1 true WO2019223552A1 (zh) 2019-11-28

Family

ID=68616542

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/086374 Ceased WO2019223552A1 (zh) 2018-05-25 2019-05-10 文章推荐方法、装置、计算机设备及存储介质

Country Status (3)

Country Link
US (1) US11763145B2 (zh)
CN (1) CN110598086B (zh)
WO (1) WO2019223552A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111046286A (zh) * 2019-12-12 2020-04-21 腾讯科技（深圳）有限公司 一种对象推荐方法、装置、以及计算机存储介质
CN111311351A (zh) * 2020-01-15 2020-06-19 上海风秩科技有限公司 文案推荐方法、装置、计算机设备和可读存储介质
CN111506820A (zh) * 2020-04-30 2020-08-07 喜大（上海）网络科技有限公司 推荐模型、方法、装置、设备及存储介质
CN111506821A (zh) * 2020-04-30 2020-08-07 喜大（上海）网络科技有限公司 推荐模型、方法、装置、设备及存储介质
CN111866724A (zh) * 2019-12-31 2020-10-30 北京嘀嘀无限科技发展有限公司 用户需求推荐方法、装置、计算机设备和可读存储介质
CN114108227A (zh) * 2020-08-27 2022-03-01 青岛海尔洗衣机有限公司 洗衣机的控制方法、装置、洗衣机及存储介质
CN114297367A (zh) * 2021-11-22 2022-04-08 深圳价值在线信息科技股份有限公司 文档信息推送方法、系统、服务器、终端设备及存储介质

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112541846B (zh) * 2020-12-22 2022-11-29 山东师范大学 一种基于注意力机制的高校选修课混合推荐方法及系统
CN113655934B (zh) * 2021-08-30 2024-07-16 咪咕数字传媒有限公司 电子文件显示方法、装置、计算设备及计算机存储介质
CN113886694B (zh) * 2021-09-30 2024-07-23 挂号网（杭州）科技有限公司 内容推荐方法、设备和存储介质
CN114491295B (zh) * 2022-04-14 2022-06-24 北京创新乐知网络技术有限公司 一种基于安卓的网络社区资源推荐方法及装置
CN115292477B (zh) * 2022-07-18 2024-04-16 盐城天眼察微科技有限公司 推送相似文章判定方法和装置、及存储介质和电子设备
CN116521936B (zh) * 2023-06-30 2023-09-01 云南师范大学 一种基于用户行为分析的课程推荐方法、装置及存储介质
CN117194792B (zh) * 2023-09-18 2024-03-01 广州道然信息科技有限公司 基于角色预测的儿童绘本推荐方法及系统
CN117563104B (zh) * 2023-11-07 2024-05-14 好心情健康产业集团有限公司 一种用于训练注意力的设备
CN119128282B (zh) * 2024-11-15 2025-04-04 湖北中文在线数字出版有限公司 一种基于人工智能的书籍推荐方法及系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103488714A (zh) * 2013-09-11 2014-01-01 杭州东信北邮信息技术有限公司 一种基于社交网络的图书推荐方法和系统
CN103559262A (zh) * 2013-11-04 2014-02-05 北京邮电大学 基于社区的作者及其学术论文推荐系统和推荐方法
CN103678620A (zh) * 2013-12-18 2014-03-26 国家电网公司 一种基于用户历史行为特征的知识文档推荐方法
US20150120712A1 (en) * 2013-03-15 2015-04-30 Yahoo! Inc. Customized News Stream Utilizing Dwelltime-Based Machine Learning
US20180107645A1 (en) * 2016-10-13 2018-04-19 SkywriterRX, Inc. Book analysis and recommendation

Family Cites Families (35)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102654859B (zh) * 2011-03-01 2014-04-23 北京彩云在线技术开发有限公司 一种歌曲推荐方法及系统
US8260117B1 (en) * 2011-07-26 2012-09-04 Ooyala, Inc. Automatically recommending content
GB2499395A (en) * 2012-02-14 2013-08-21 British Sky Broadcasting Ltd Search method
CN107330124A (zh) * 2013-03-11 2017-11-07 曹华诚 内容推荐方法
US10148375B2 (en) * 2013-03-14 2018-12-04 The Directv Group, Inc. Method and system for displaying recommended content such as movies associated with a cluster
US9336210B2 (en) * 2013-07-15 2016-05-10 Google Inc. Determining a likelihood and degree of derivation among media content items
JPWO2015136584A1 (ja) * 2014-03-11 2017-04-06 パナソニックＩｐマネジメント株式会社 推薦画像表示システム、推薦画像出力装置、および、推薦画像出力方法
US20150339693A1 (en) * 2014-05-21 2015-11-26 Qbeats Inc. Determination of initial value for automated delivery of news items
US20150339692A1 (en) * 2014-05-21 2015-11-26 Qbeats Inc. Automated volatility score adjustment of access charge curve for automated delivery of news items
CN104090919B (zh) * 2014-06-16 2017-04-19 华为技术有限公司 推荐广告的方法及广告推荐服务器
CN105989045A (zh) * 2015-02-04 2016-10-05 华为技术有限公司 一种劝导内容推荐的方法、装置及系统
US9864951B1 (en) * 2015-03-30 2018-01-09 Amazon Technologies, Inc. Randomized latent feature learning
US20160353235A1 (en) * 2015-06-01 2016-12-01 Accenture Global Services Limited Location-based order recommendations
US10341701B2 (en) * 2015-04-21 2019-07-02 Edge2020 LLC Clustering and adjudication to determine a recommendation of multimedia content
RU2632100C2 (ru) * 2015-09-28 2017-10-02 Общество С Ограниченной Ответственностью "Яндекс" Способ и сервер создания рекомендованного набора элементов
RU2629638C2 (ru) * 2015-09-28 2017-08-30 Общество С Ограниченной Ответственностью "Яндекс" Способ и сервер создания рекомендуемого набора элементов для пользователя
US20170169040A1 (en) * 2015-12-15 2017-06-15 Le Holdings (Beijing) Co., Ltd. Method and electronic device for recommending video
US20170188102A1 (en) * 2015-12-23 2017-06-29 Le Holdings (Beijing) Co., Ltd. Method and electronic device for video content recommendation
US10341728B2 (en) * 2015-12-30 2019-07-02 Sling Media L.L.C. Media systems for temporally and contextually relevant recommendations
US10257308B2 (en) * 2016-05-23 2019-04-09 Facebook, Inc. Recommending a group to a user of a social networking system based on affinities of the user for members of the group
RU2632132C1 (ru) * 2016-07-07 2017-10-02 Общество С Ограниченной Ответственностью "Яндекс" Способ и устройство для создания рекомендаций содержимого в системе рекомендаций
US10515423B2 (en) * 2016-08-02 2019-12-24 Microsoft Technology Licensing, Llc Shareability score
US10671680B2 (en) * 2016-08-25 2020-06-02 Microsoft Technology Licensing, Llc Content generation and targeting using machine learning
US10832130B2 (en) * 2016-09-19 2020-11-10 Google Llc Recommending a document for a user to access
US10015124B2 (en) * 2016-09-20 2018-07-03 Google Llc Automatic response suggestions based on images received in messaging applications
US11488270B2 (en) * 2016-12-07 2022-11-01 Tata Consultancy Services Limited System and method for context and sequence aware recommendation
US10846617B2 (en) * 2017-05-12 2020-11-24 Adobe Inc. Context-aware recommendation system for analysts
US11138507B2 (en) * 2017-09-28 2021-10-05 Applied Materials Israel Ltd. System, method and computer program product for classifying a multiplicity of items
CN107766547A (zh) * 2017-10-31 2018-03-06 掌阅科技股份有限公司 电子书推荐方法、电子设备及计算机存储介质
RU2699574C2 (ru) * 2017-11-24 2019-09-06 Общество С Ограниченной Ответственностью "Яндекс" Способ и сервер для представления элемента рекомендуемого содержимого пользователю
US11604968B2 (en) * 2017-12-11 2023-03-14 Meta Platforms, Inc. Prediction of next place visits on online social networks
FR3087921A1 (fr) * 2018-10-31 2020-05-01 Amadeus S.A.S. Systèmes et méthodes de recommandation utilisant des modèles d'apprentissage automatique en cascade
CN109815368A (zh) * 2018-12-10 2019-05-28 百度在线网络技术（北京）有限公司 资源推荐方法、装置、设备及计算机可读存储介质
US20200380540A1 (en) * 2019-05-30 2020-12-03 HG Insights, Inc. Predicting the probability of a product purchase
CN112100504B (zh) * 2020-11-03 2021-09-10 北京达佳互联信息技术有限公司 内容推荐方法、装置、电子设备及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150120712A1 (en) * 2013-03-15 2015-04-30 Yahoo! Inc. Customized News Stream Utilizing Dwelltime-Based Machine Learning
CN103488714A (zh) * 2013-09-11 2014-01-01 杭州东信北邮信息技术有限公司 一种基于社交网络的图书推荐方法和系统
CN103559262A (zh) * 2013-11-04 2014-02-05 北京邮电大学 基于社区的作者及其学术论文推荐系统和推荐方法
CN103678620A (zh) * 2013-12-18 2014-03-26 国家电网公司 一种基于用户历史行为特征的知识文档推荐方法
US20180107645A1 (en) * 2016-10-13 2018-04-19 SkywriterRX, Inc. Book analysis and recommendation

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111046286A (zh) * 2019-12-12 2020-04-21 腾讯科技（深圳）有限公司 一种对象推荐方法、装置、以及计算机存储介质
CN111866724A (zh) * 2019-12-31 2020-10-30 北京嘀嘀无限科技发展有限公司 用户需求推荐方法、装置、计算机设备和可读存储介质
CN111311351A (zh) * 2020-01-15 2020-06-19 上海风秩科技有限公司 文案推荐方法、装置、计算机设备和可读存储介质
CN111311351B (zh) * 2020-01-15 2023-09-01 上海风秩科技有限公司 文案推荐方法、装置、计算机设备和可读存储介质
CN111506820A (zh) * 2020-04-30 2020-08-07 喜大（上海）网络科技有限公司 推荐模型、方法、装置、设备及存储介质
CN111506821A (zh) * 2020-04-30 2020-08-07 喜大（上海）网络科技有限公司 推荐模型、方法、装置、设备及存储介质
CN111506820B (zh) * 2020-04-30 2023-05-09 喜大（上海）网络科技有限公司 推荐模型、方法、装置、设备及存储介质
CN111506821B (zh) * 2020-04-30 2023-06-20 喜大（上海）网络科技有限公司 推荐模型、方法、装置、设备及存储介质
CN114108227A (zh) * 2020-08-27 2022-03-01 青岛海尔洗衣机有限公司 洗衣机的控制方法、装置、洗衣机及存储介质
CN114297367A (zh) * 2021-11-22 2022-04-08 深圳价值在线信息科技股份有限公司 文档信息推送方法、系统、服务器、终端设备及存储介质

Also Published As

Publication number Publication date
US11763145B2 (en) 2023-09-19
CN110598086A (zh) 2019-12-20
CN110598086B (zh) 2020-11-24
US20200364506A1 (en) 2020-11-19

Similar Documents

Publication Publication Date Title
WO2019223552A1 (zh) 文章推荐方法、装置、计算机设备及存储介质
CN110569496B (zh) 实体链接方法、装置及存储介质
CN110909182B (zh) 多媒体资源搜索方法、装置、计算机设备及存储介质
CN109783671B (zh) 一种以图搜图的方法、计算机可读介质及服务器
CN101551806B (zh) 一种个性化网址导航的方法和系统
WO2020108608A1 (zh) 搜索结果处理方法、装置、终端、电子设备及存储介质
US9996504B2 (en) System and method for classifying text sentiment classes based on past examples
US11947590B1 (en) Systems and methods for contextualized visual search
CN107391680A (zh) 内容推荐方法、装置和设备
CN102855256B (zh) 用于确定网站评价信息的方法、装置及设备
CN114491294B (zh) 基于图神经网络的数据推荐方法及装置、电子设备
US11403339B2 (en) Techniques for identifying color profiles for textual queries
WO2016086724A1 (zh) 一种确定候评项的质量信息的方法与装置
US8825641B2 (en) Measuring duplication in search results
US20150309988A1 (en) Evaluating Crowd Sourced Information Using Crowd Sourced Metadata
CN112487242A (zh) 用于识别视频的方法、装置、电子设备及可读存储介质
CN110209928A (zh) 一种信息推荐方法、装置及存储介质
CN115712780A (zh) 一种基于云计算和大数据的信息推送方法及装置
CN114329206B (zh) 标题生成方法和装置、电子设备、计算机可读介质
CN113849688B (zh) 资源处理方法、资源处理装置、电子设备以及存储介质
CN115374850A (zh) 一种同行企业智能发现方法、装置、电子设备及介质
CN115422485A (zh) 信息发送方法、装置、电子设备和存储介质
CN112818221B (zh) 实体的热度确定方法、装置、电子设备及存储介质
KR101958555B1 (ko) 검색 결과 제공 장치 및 방법
CN111651643B (zh) 候选内容的处理方法及相关设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19806404

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19806404

Country of ref document: EP

Kind code of ref document: A1