WO2020114145A1 - 内容推送方法及存储介质、计算机设备 - Google Patents

内容推送方法及存储介质、计算机设备 Download PDF

Info

Publication number
WO2020114145A1
WO2020114145A1 PCT/CN2019/113566 CN2019113566W WO2020114145A1 WO 2020114145 A1 WO2020114145 A1 WO 2020114145A1 CN 2019113566 W CN2019113566 W CN 2019113566W WO 2020114145 A1 WO2020114145 A1 WO 2020114145A1
Authority
WO
WIPO (PCT)
Prior art keywords
value
vector
content
pushed
rate
Prior art date
Application number
PCT/CN2019/113566
Other languages
English (en)
French (fr)
Inventor
聂照昌
Original Assignee
广州市百果园信息技术有限公司
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 广州市百果园信息技术有限公司 filed Critical 广州市百果园信息技术有限公司
Publication of WO2020114145A1 publication Critical patent/WO2020114145A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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

Definitions

  • This application relates to the technical field of data processing, and in particular, to a content pushing method, a storage medium, and a computer device.
  • CTR Click-through rate
  • the collected features are often large-scale discrete and sparse.
  • the model needs to use a certain feature of the commodity, it needs to use a tens of thousands of vectors to represent the commodity. feature.
  • the CTR estimates that there may be a strong correlation between features in the scene, such as clothing and gender, and these two features obviously have a strong correlation.
  • the clothes are specifically represented as "skirts" and the gender as "female", the probability of user clicks will increase.
  • the vast majority of recommended system engineers need to dig out strongly related feature combinations.
  • mining effective feature combinations requires a lot of manual work, also requires certain cross-domain knowledge, and needs to be tried continuously. It is a very heavy task for the recommendation system engineer.
  • the present application proposes a content push method, storage medium, and computer equipment to realize the automatic mining of content push related features to push content to users, improve the accuracy of content push, and reduce the amount of labor.
  • a content push method includes: extracting multiple related features of content to be pushed; inputting the multiple related features into a click-through rate prediction model based on feature combinations to obtain an estimated click-through rate of the content to be pushed; wherein, The click-rate prediction model based on feature combination is used to combine the plurality of related features, and the estimated click-through rate of the content to be pushed is determined according to the relevance of the combined related features; State the estimated click rate of the content to be pushed and push the content to the user.
  • the content to be pushed includes short video content to be pushed;
  • the extracting a plurality of related features of the content to be pushed includes: extracting a plurality of user characteristics and a plurality of the short video content to be pushed Short video content features;
  • the inputting the plurality of related features into a click-through rate prediction model based on a combination of features to obtain the estimated click-through rate of the content to be pushed includes: combining the multiple user characteristics and the multiple Short video content features are input into the click rate prediction model based on the combination of features to obtain the estimated click rate of the short video content to be pushed;
  • the estimated click rate according to the content to be pushed is pushed to the user
  • the content includes: pushing the short video content to the user according to the estimated click rate of the short video content to be pushed.
  • the click-through rate prediction model based on feature combination includes: an input layer for uniquely encoding the plurality of related features to obtain a uniquely hot vector; and a logistic regression component for Logistic regression operation is performed on the one-hot vector output from the input layer to obtain a first operation value; a nested module is used to convert the high-dimensional sparse discretization feature of the one-hot vector into a low-dimensional continuous value feature to obtain a low-dimensional A continuous value vector; an inner product component, used to perform a vector inner product of the low-dimensional continuous value vector output by the nested module to obtain a second operation value; a splicing module, used to output all the values output by the nested module The low-dimensional continuous value vector is subjected to vector splicing to obtain a splicing vector; a hidden layer module is used to input the splicing vector output by the splicing module into the deep neural network hidden layer to obtain a third operation value; an estimated click rate calculation module , Used to calculate the estimated click
  • calculating the estimated click rate of the content to be pushed based on the first operation value, the second operation value, and the third operation value includes: calculating the first operation value Value, the second operation value and the third operation value are spliced to perform a normalization operation to obtain an estimated click rate of the content to be pushed.
  • the first calculated value, the second calculated value, and the third calculated value are spliced to perform a normalized operation to obtain an estimated click rate of the content to be pushed, including : Obtain the first weight corresponding to the first calculated value, the second weight corresponding to the second calculated value, and the third weight corresponding to the third calculated value; multiply the first calculated value by the first A weight to obtain a first value; multiplying the second operation value by the second weight to obtain a second value; multiplying the third operation value by the third weight to obtain a third value; A value, the second value, and the third value are accumulated to obtain an estimated click rate of the content to be pushed.
  • the vector inner product of the low-dimensional continuous value vectors output by the nested module to obtain a second operation value includes: obtaining the weight corresponding to each low-dimensional continuous value vector; Each low-dimensional continuous value is multiplied by a corresponding weight to perform the inner product of the vector to obtain the second operation value.
  • the vector splicing of the low-dimensional continuous value vectors output by the nesting module to obtain a splicing vector includes: obtaining weights corresponding to the low-dimensional continuous value vectors; After multiplying the low-dimensional continuous value vector by the corresponding weight, vector accumulation is performed to obtain the splicing vector.
  • the click-through rate prediction model based on feature combination includes: an input layer for uniquely encoding the plurality of related features to obtain a uniquely hot vector; and a logistic regression component for Logistic regression operation is performed on the one-hot vector output from the input layer to obtain a first operation value; a nested module is used to convert the high-dimensional sparse discretization feature of the one-hot vector into a low-dimensional continuous value feature to obtain a low-dimensional Continuous value vector; a vector multiplying component, which is used to perform vector multiplication of the low-dimensional continuous value vector output by the nested module by matrix multiplication through parallel calculation to obtain a result value vector; obtain the result value vector The value of the upper triangle, the values of the upper triangle are accumulated, and the accumulated result value is used as the second operation value; the splicing module is used to perform vector splicing on the low-dimensional continuous value vector output by the nested module to obtain Stitching vector; hidden layer module, which is used to input the stitching vector output by the stitching module
  • a computer device including: one or more processors; a memory; one or more application programs, wherein the one or more application programs are stored in the memory and configured to be configured by the one or more Executed by one processor, and the one or more application programs are configured to execute the content pushing method according to any one of the foregoing embodiments.
  • the content pushing method extracts multiple related features of the content to be pushed, enters the multiple related features into a click-through rate prediction model based on the combination of features, and combines multiple related features through the click-through rate prediction model , And analyze the relevance of the combined related features, and determine the estimated click-through rate of the content to be pushed according to the relevance, so that the content can be pushed to the user according to the estimated click-through rates of multiple content to be pushed. It may be that, among the content to be pushed, the push content with a higher estimated click rate is selected, and the part of the push content is pushed to the user. Therefore, the accuracy of pushing content to the user can be improved. Moreover, this method does not require manual selection of combined features, which can reduce the amount of labor.
  • FIG. 1 is a flowchart of a method in an embodiment of a content pushing method provided by this application;
  • FIG. 3 is a structural block diagram of an embodiment of a click rate prediction model based on feature combinations provided in this application;
  • FIG. 4 is a structural block diagram of a specific embodiment of a click rate prediction model based on feature combinations provided by this application;
  • FIG. 5 is a schematic diagram of an embodiment in which two vectors are multiplied in a vector multiplication component provided by this application;
  • FIG. 6 is a schematic structural diagram of an embodiment of a computer device provided by this application.
  • the present application provides a content pushing method, which is used to directionally push relevant content to users based on the estimated click rate of the content to be pushed by the user, so as to improve the accuracy of pushing content.
  • a content pushing method which is used to directionally push relevant content to users based on the estimated click rate of the content to be pushed by the user, so as to improve the accuracy of pushing content.
  • Click Rate Prediction Predict the probability of "when a user recommends a certain information/advertisement and other content to the user, the user will click to push the content".
  • Feature combination A composite feature formed by combining individual features (multiplication or Cartesian product). Among them, feature combination helps to represent the nonlinear relationship between features.
  • Embedding A method of converting high-dimensional sparse discretized features into low-dimensional continuous-value features.
  • One-hot vector only one sparse discrete vector with a value of 1 and all other values being 0.
  • the existing click-rate prediction deep neural model when embedding is nested, for an input feature, only a low-dimensional continuous-value feature vector is generated to be combined with other features.
  • the existing feature 1 is clothing
  • feature 2 is gender
  • feature 3 is age.
  • the degree of association between clothing and age, clothing and gender is different.
  • the vector corresponding to clothes is (0.3, 0.4)
  • the vector corresponding to gender is (x1, y1)
  • the vector corresponding to gender is (x2, y2)
  • the dot product of the feature vector is used as the value of the importance of the feature combination
  • the weights provided by the clothes are all 0.5 (that is, the modulus of the vector), which is the same for age and gender.
  • this application proposes a new click-through rate prediction neural network model to cross-features, deep mining the correlation between cross-features, and finally output the probability value of the click, based on the probability value to determine the user Click product probability.
  • the present application provides a content pushing method, including the following steps:
  • the server extracts multiple related features of the content to be pushed. among them.
  • the content to be pushed may be products, advertisements, short videos and the like.
  • the plurality of related features may include content features of the content to be pushed and non-native features associated with the content to be pushed.
  • multiple related features may include characteristics of a user who views the short video and characteristics of the short video itself (such as the number of viewers, the number of likes, video tags, etc.).
  • the click rate prediction model based on feature combination is used to combine multiple input features and analyze the correlation between the combined features, and determine the content to be pushed by the user according to the correlation of the features Estimated clickthrough rate.
  • the server inputs a plurality of relevant features of the content to be pushed into the predictive click rate model based on the combination of features, and obtains the estimated click rate of the content to be pushed according to the correlation of the combined features.
  • the click-rate prediction model based on feature combination includes: input layer 10, logistic regression component 20, nesting module 30, inner product component 40, splicing module 50, hidden layer module 60 And the estimated click rate calculation module 70.
  • the input layer 10 is used to uniquely encode multiple related features to obtain a uniquely hot vector. Specifically, the input layer 10 receives a plurality of related features, performs one-hot encoding on the plurality of related features, and obtains a plurality of one-hot vectors. Each one-hot vector is a sparse discrete vector in which there is only one value in the vector and all other values are 0.
  • the logistic regression component 20 is used to perform a logistic regression operation on the one-hot vector output by the input layer 10 to obtain the first calculated value. Therefore, simple and effective independent features of the content to be pushed can be tapped.
  • the nesting module 30 is used to convert the high-dimensional sparse discretization features of the one-hot vector output by the input layer 10 into low-dimensional continuous-value features to obtain low-dimensional continuous-value vectors. After one unique heat vector is input into the nesting module 30, multiple low-dimensional continuous value vectors are obtained.
  • the inner product component 40 is used to perform a vector inner product of the low-dimensional continuous value vector output by the nesting module 30 to obtain a second operation value.
  • the inner product of vectors can be regarded as a way of combining features provided in this application. In this part, the inner product of the low-dimensional continuous value vectors can be intuitively expressed as the combination relationship between the two features.
  • the low-dimensional continuous value vector output from the nested module is subjected to a vector inner product to obtain a second operation value, which includes: obtaining the weight corresponding to each low-dimensional continuous value vector; multiplying each low-dimensional continuous value vector Perform the inner product of the vectors with the corresponding weights to obtain the second calculated value.
  • each low-dimensional continuous value vector corresponds to a weight
  • the weight is automatically displayed during the model training process.
  • the server obtains the weight corresponding to each updated low-dimensional continuous value vector every time, multiplies each low-dimensional continuous value by the corresponding weight, and performs a vector inner product to obtain the second operation value.
  • the splicing module 50 is used for vector splicing of the low-dimensional continuous value vectors output by the nesting module 30 to obtain a splicing vector.
  • vector stitching can also be seen as a way of feature combination.
  • the results are input into a multi-layer neural network to further explore the deep feature combination relationship.
  • the low-dimensional continuous value vectors output by the nested modules are subjected to vector splicing to obtain a spliced vector, which includes: obtaining the weight corresponding to the low-dimensional continuous value vector; multiplying each low-dimensional continuous value vector by the corresponding weight After that, vector accumulation is performed to obtain the stitching vector.
  • each low-dimensional continuous value vector corresponds to a weight
  • the weight is automatically displayed during the model training process.
  • the server obtains the weight corresponding to each updated low-dimensional continuous value vector every time, multiplies each low-dimensional continuous value by the corresponding weight, and then performs vector accumulation, and then obtains the stitching vector after accumulation.
  • the hidden layer module 60 is used to input the splicing vector output by the splicing module 50 into the deep neural network hidden layer to obtain a third operation value.
  • the deep neural network may be Deep and Wide, FNN (Factorisation-machine supported Neural Networks), PNN (Factorisation-machine supported Neural Networks Product-based Neural Networks), DeepFM (Deep Factorization-machine) and so on.
  • the estimated click rate calculation module 70 is used to calculate the estimated content to be pushed based on the first calculated value output by the logistic regression component 20, the second calculated value output by the inner product component 40, and the third calculated value output by the hidden layer module 60 CTR.
  • calculating the estimated click rate of the content to be pushed based on the first calculated value, the second calculated value, and the third calculated value including: splicing the first calculated value, the second calculated value, and the third calculated value to return One operation, get the estimated click rate of the content to be pushed.
  • the splicing described here may be to accumulate the first calculated value, the second calculated value and the third calculated value.
  • the first weight corresponding to the first operation value, the second weight corresponding to the second operation value, and the third weight corresponding to the third operation value are obtained; the first operation value is multiplied by the first weight , Get the first value; multiply the second operation value by the second weight to get the second value; multiply the third operation value by the third weight to get the third value; accumulate the first value, the second value and the third value To get the estimated click-through rate of the content to be pushed. That is, in the cumulative calculation process, different calculation values correspond to different weights. After multiplying the calculated value by its corresponding weight and then accumulating, the accumulated result value finally obtained is the estimated click rate of the content to be pushed.
  • the original input one-hot vector is a plurality of discretized one-hot vectors (one-hot vector).
  • the solid points indicate a value of 1, and the hollow points indicate a value of zero.
  • a logistics regression (LR) component 200 is designed in the model.
  • each one-hot vector will generate a set of continuous-valued feature vectors.
  • different feature vectors are used. Assuming clothes, age, and gender as examples, clothes will generate two continuous-value feature vectors (0.3, 0.4) and (0.6, 0.8). When combining clothing and age, use (0.3, 0.4), when the weight of clothing is 0.5; when combining clothing and gender, use (0.6, 0.8), and then the weight of clothing is 1.0.
  • the server further performs the inner product 400 of the low-dimensional continuous value vector 300, specifically the inner product between the vectors, in order to mine simple and effective feature combination methods.
  • the server also performs vector splicing on the low-dimensional continuous value vector 300, and inputs the splicing result into a multi-layer neural network hidden layer (Hidden Layers) 600 to further mine complex feature combination methods to improve the learning ability of the model.
  • Hidden Layers multi-layer neural network hidden layer
  • the output of the LR 400, the output of the inner product 400 of the vector, and the output of the multi-layer hidden layer (Hidden Layers) 600 are spliced together to perform a normalized operation 700 and then output the corresponding estimated click rate.
  • vector inner product and vector splicing can be regarded as the way of feature combination. After vector splicing, the input to the multi-layer neural network can mine the deep feature combination relationship, and the vector inner product can intuitively represent the combination relationship between the two features.
  • the weighted edges (the connecting line between the input layer 100 and the RL 200 in the drawing, the connecting line between the input layer 100 and the low-dimensional continuous value vector 300, and the low-dimensional continuous value vector 300 and Vector inner product 400 connecting line, low dimension continuous value vector 300 and vector stitching 500 connecting line, vector stitching 500 and hidden layer (Hidden Layers) 600 connecting line and RL 200, vector inner product 400, hidden layer (Hidden Layers) 600 and the connection line of the normalization operation 700 respectively) are the universal connection edges of the neural network (the weights of each edge are different, and the weights are automatically updated during the model training process).
  • Edges with a weight of 1 (RL 200, vector inner product 400, hidden layer (Hidden Layers 600) and the connection line of the normalization operation 700) will not be updated, and the edge is always 1, which can be embodied as a direct addition form .
  • the weighted edge means that before the operation, the corresponding vector or value needs to be multiplied by the corresponding weight before the subsequent operation.
  • the one-hot vector input at the bottom can include the characteristics of the user (such as age, gender, city, etc.), and the characteristics of the commodity (commodity category, commodity price, historical purchase quantity of the commodity, etc.). These features are used as the input of the click-through rate prediction model based on the combination of features.
  • the model automatically crosses user features and product features, and deeply mines the correlation between the two. Finally, the model outputs a probability value that the user clicks on the product, that is, the estimated click rate value, and determines the probability that the user clicks on the product according to the probability value.
  • the server optimizes the CTR prediction model based on the feature combination.
  • the optimized CTR prediction model based on the feature combination includes the input layer 10, the logistic regression component 20, the nesting module 30, and the splicing module 50.
  • the hidden layer module 60 and the estimated click rate calculation module 70, and the original inner product component 40 is optimized and modified.
  • the inner product component 40 is replaced with a vector multiplication component.
  • the vector multiplication component is used to perform vector multiplication on the low-dimensional continuous values output by the two nested modules through parallel calculation to obtain the result value vector; to obtain the upper triangle value of the result value vector and to perform the upper triangle value Accumulate, use the accumulated result value as the second operation value. Specifically, as shown in FIG.
  • the operation of inner product between vectors is changed into matrix multiplication, and then the three values of the upper triangle are obtained by taking the upper triangle, and the three values are accumulated to obtain the second operation value.
  • the GPU's own parallel computing method can greatly reduce the training time of the model.
  • step S300 includes: obtaining estimated click-through rates of a plurality of content to be pushed, sorting the content to be pushed according to the estimated click-through rate of each content to be pushed from high to low, and obtaining preset presets The amount of content to be pushed, and push the preset amount of content to be pushed to the user.
  • the content to be pushed whose estimated click rate is greater than a preset value is obtained, and the content to be pushed whose estimated click rate is greater than the preset value is pushed to the user.
  • the content pushing method extracts multiple related features of the content to be pushed, enters the multiple related features into a click-through rate prediction model based on the combination of features, and combines multiple related features through the click-through rate prediction model , And analyze the relevance of the combined related features, and determine the estimated click-through rate of the content to be pushed according to the relevance, so that the content can be pushed to the user according to the estimated click-through rates of multiple content to be pushed. It may be that, among the content to be pushed, the push content with a higher estimated click rate is selected, and the part of the push content is pushed to the user. Therefore, the accuracy of pushing content to the user can be improved. Moreover, this method does not require manual selection of combined features, which can reduce the amount of labor.
  • Step S100 includes:
  • Step S200 includes:
  • S210 Input multiple user features and multiple short video content features into a click-through rate estimation model based on feature combinations to obtain an estimated click-through rate of short video content to be pushed.
  • Step S300 includes:
  • S310 Push the short video content to the user according to the estimated click rate of the short video content to be pushed.
  • the content to be pushed is short video content to be pushed.
  • the server extracts the relevant features of the short video content to be pushed, such as multiple user features and multiple short video content features.
  • the relevant characteristics may include user characteristics of the target user, such as user gender, user age, user occupation, and user's permanent residence.
  • Related features may also include the characteristics of the short video content itself, such as the number of viewers of the short video content, the number of likes, and video tags.
  • Multiple relevant features of short video content to be pushed are input into a click-through rate prediction model based on feature combination, and multiple user features and multiple short video content features are combined and crossed through the model to mine the correlation of related features after combination , So as to output the estimated click rate value of the short video content to be pushed.
  • the short video content is personalized and pushed to the user according to the estimated click rate value.
  • the application also provides a storage medium.
  • a computer program is stored on the storage medium; when the computer program is executed by the processor, the content pushing method provided by any of the foregoing embodiments is implemented.
  • the storage medium may be a memory.
  • internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory.
  • the external memory may include hard disks, floppy disks, ZIP disks, U disks, magnetic tapes, etc.
  • the storage media disclosed in this application include but are not limited to these types of memories.
  • the memory disclosed in this application is only an example and not a limitation.
  • a computer device includes: one or more processors; memory; one or more application programs.
  • One or more application programs are stored in the memory and configured to be executed by one or more processors, and the one or more application programs are configured to execute the content pushing method provided by any of the foregoing embodiments.
  • the computer equipment provided in this embodiment may be a server, a personal computer, and a network equipment.
  • the device includes devices such as a processor 603, a memory 605, an input unit 607, and a display unit 609.
  • the memory 605 may be used to store application programs 601 and various functional modules.
  • the processor 603 runs the application programs 601 stored in the memory 605 to execute various functional applications and data processing of the device.
  • the memory may be internal memory or external memory, or include both internal memory and external memory.
  • the internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory.
  • ROM read-only memory
  • PROM programmable ROM
  • EPROM electrically programmable ROM
  • EEPROM electrically erasable programmable ROM
  • flash memory or random access memory.
  • the external memory may include hard disks, floppy disks, ZIP disks, U disks, magnetic tapes, etc.
  • the memories disclosed in this application include but are not limited to these types of memories.
  • the memory disclosed in this application is only an example and not a limitation.
  • the input unit 607 is used to receive an input of a signal and a keyword input by a user.
  • the input unit 607 may include a touch panel and other input devices.
  • the touch panel can collect the user's touch operations on or near it (such as the user's operation on the touch panel or near the touch panel using any suitable objects or accessories such as fingers, stylus, etc.), and according to the preset
  • the program drives the corresponding connection device; other input devices may include but are not limited to one or more of a physical keyboard, function keys (such as playback control keys, switch keys, etc.), trackball, mouse, joystick, etc.
  • the display unit 609 can be used to display information input by the user or information provided to the user and various menus of the computer device.
  • the display unit 609 may take the form of a liquid crystal display, an organic light-emitting diode, or the like.
  • the processor 603 is the control center of the computer equipment. It uses various interfaces and lines to connect the various parts of the entire computer. Various functions and processing data.
  • the device includes one or more processors 603, and one or more memories 605, and one or more applications 601.
  • One or more application programs 601 are stored in the memory 605 and configured to be executed by one or more processors 603, and the one or more application programs 601 are configured to execute the content pushing method provided by the above embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module.
  • the above integrated modules may be implemented in the form of hardware or software function modules. If the above integrated modules are implemented in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.
  • a person of ordinary skill in the art may understand that all or part of the steps to implement the above-described embodiments may be completed by hardware, or may be completed by a program instructing related hardware.
  • the program may be stored in a computer-readable storage medium, and the storage medium may include Memory, magnetic disk or optical disk, etc.
  • each functional unit in each embodiment of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units may be integrated into one module.
  • the above integrated modules may be implemented in the form of hardware or software function modules.

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • User Interface Of Digital Computer (AREA)

Abstract

本申请提供一种内容推送方法及存储介质、计算机设备，所述方法包括：提取待推送内容的多个相关特征；将所述多个相关特征输入基于特征组合的点击率预估模型，得到所述待推送内容的预估点击率；其中，所述基于特征组合的点击率预估模型用于将所述多个相关特征进行组合，根据组合后的所述相关特征的关联性，确定出所述待推送内容的预估点击率；根据所述待推送内容的预估点击率，向用户推送内容。该方法可提高向用户推送内容的精准度。并且，该方法无需进行人工筛选组合特征，可减少人工量。

Description

内容推送方法及存储介质、计算机设备
本申请要求于2018年12月07日提交的申请号为201811497834.2、发明名称为“内容推送方法及存储介质、计算机设备”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，特别涉及一种内容推送方法及存储介质、计算机设备。
背景技术
随着互联网技术的快速发展，人类进入了大数据时代。这给互联网用户和互联网平台带来了新的挑战。对于用户而言，从大量信息中获取自己感兴趣的信息的难度逐渐增加。对于平台而言，如何给每位用户推荐个性化的信息/广告并以此获取收益也逐渐变成了平台的关键任务之一。因此，个性化推荐技术受到了越来越多的关注。点击率(Click-through Rate，CTR)预估是个性化推荐中的一项核心关键技术，通过机器学习方法建立模型，用于预测“如果给一位用户推荐某个信息/广告，该用户会点击的概率”。一个好的点击率预估模型可以大大提升互联网平台的收益。
业界CTR预估的场景中，采集的特征往往都是大规模离散稀疏化的。比如在淘宝的场景中，商品的数量成千上万，每一个商品会有对应的id，假设模型需要使用到商品某个特征，则需要用一个成千上万维的向量去表示商品的该特征。而CTR预估场景中特征之间可能存在比较强的相关性，比如衣服和性别，这两个特征明显存在一个很强的关联关系。当衣服具体表现为“裙子”，性别为“女”时，用户点击的概率会提高。目前在工业界，绝大多数推荐系统工程师需要挖掘出强相关的特征组合。但是挖掘有效的特征组合，需要大量的人工工作，也需要一定的跨领域知识，并且需要不断地尝试，对于推荐系统工程师来说是一个很繁重的任务。
发明内容
本申请提出一种内容推送方法及存储介质、计算机设备，实现自动化挖掘内容推送的关联特征以向用户进行内容推送，提高内容推送的精准度，并减少人工量。
本申请提供以下方案：
一种内容推送方法，包括：提取待推送内容的多个相关特征；将所述多个相关特征输入基于特征组合的点击率预估模型，得到所述待推送内容的预估点击率；其中，所述基于特征组合的点击率预估模型用于将所述多个相关特征进行组合，根据组合后的所述相关特征的关联性，确定出所述待推送内容的预估点击率；根据所述待推送内容的预估点击率，向用户推送内容。
在一实施例中，所述待推送内容包括待推送的短视频内容；所述提取待推送内容的多个相关特征，包括：提取所述待推送的短视频内容的多个用户特征以及多个短视频内容特征；所述将所述多个相关特征输入基于特征组合的点击率预估模型，得到所述待推送内容的预估点击率，包括：将所述多个用户特征以及所述多个短视频内容特征输入所述基于特征组合的点击率预估模型，得到所述待推送的短视频内容的预估点击率；所述根据所述待推送内容的预估点击率，向用户推送内容，包括：根据所述待推送的短视频内容的预估点击率，向用户推送短视频内容。
在一实施例中，所述基于特征组合的点击率预估模型包括：输入层，用于对所述多个相关特征进行独热编码，得到独热向量；逻辑回归组件，用于对所述输入层输出的所述独热向量进行逻辑回归运算，得到第一运算值；嵌套模块，用于将所述独热向量的高维度稀疏离散化特征转换成低维度连续值特征，得到低维度连续值向量；内积组件，用于将所述嵌套模块输出的所述低维度连续值向量进行向量内积，得到第二运算值；拼接模块，用于将所述嵌套模块输出的所述低维度连续值向量进行向量拼接，得到拼接向量；隐藏层模块，用于将所述拼接模块输出的所述拼接向量输入深度神经网络隐藏层，得到第三运算值；预估点击率运算模块，用于根据所述第一运算值、所述第二运算值及所述第三运算值计算出所述待推送内容的预估点击率。
在一实施例中，所述根据所述第一运算值、所述第二运算值及所述第三运算值计算出所述待推送内容的预估点击率，包括：将所述第一运算值、所述第二运算值及所述第三运算值拼接后进行归一化运算，得到所述待推送内容的预估点击率。
在一实施例中，所述将所述第一运算值、所述第二运算值及所述第三运算值拼接后进行归一化运算，得到所述待推送内容的预估点击率，包括：获取所述第一运算值对应的第一权重、所述第二运算值对应的第二权重、所述第三运算值对应的第三权重；将所述第一运算值乘以所述第一权重，得到第一值；所述第二运算值乘以所述第二权重，得到第二值；所述第三运算值乘以所述第三权重，得到第三值；将所述第一值、所述第二值和所述第三值进行累加，得到所述待推送内容的预估点击率。
在一实施例中，所述将所述嵌套模块输出的所述低维度连续值向量进行向量内积，得到第二运算值，包括：获取每个低维度连续值向量对应的权重；将所述每个低维度连续值乘以对应的权重后进行所述向量内积，得到所述第二运算值。
在一实施例中，所述将所述嵌套模块输出的所述低维度连续值向量进行向量拼接，得到拼接向量，包括：获取所述低维度连续值向量对应的权重；将所述每个低维度连续值向量乘以对应的权重后，进行向量累加，得到所述拼接向量。
在一实施例中，所述基于特征组合的点击率预估模型包括：输入层，用于对所述多个相关特征进行独热编码，得到独热向量；逻辑回归组件，用于对所述输入层输出的所述独热向量进行逻辑回归运算，得到第一运算值；嵌套模块，用于将所述独热向量的高维度稀疏离散化特征转换成低维度连续值特征，得到低维度连续值向量；向量相乘组件，用于通过并行计算方式，采用矩阵乘法对所述嵌套模块输出的所述低维度连续值向量进行向量相乘，得到结果值向量；获取所述结果值向量上三角的值，将所述上三角的值进行累加，将累加结果值作为第二运算值；拼接模块，用于将所述嵌套模块输出的所述低维度连续值向量进行向量拼接，得到拼接向量；隐藏层模块，用于将所述拼接模块输出的所述拼接向量输入深度神经网络隐藏层，得到第三运算值；预估点击率运算模块，用于根据所述第一运算值、所述第二运算值及所述第三运算值计算出所述待推送内容的预估点击率。
一种存储介质，其上存储有计算机程序；所述计算机程序适于由处理器加载并执行上述任一实施例所述的内容推送方法。
一种计算机设备，其包括：一个或多个处理器；存储器；一个或多个应用程序，其中所述一个或多个应用程序被存储在所述存储器中并被配置为由所述 一个或多个处理器执行，所述一个或多个应用程序配置用于执行根据上述任一实施例所述的内容推送方法。
上述实施例提供的内容推送方法，提取待推送内容的多个相关特征，将该多个相关特征输入基于特征组合的点击率预估模型中，通过点击率预估模型将多个相关特征进行组合，并分析组合后的相关特征的关联性，根据关联性确定出待推送内容的预估点击率，从而可以根据多个待推送内容的预估点击率向用户推送内容。可以是，从多个待推送内容中筛选出预估点击率较高的推送内容，并向用户推送该部分的推送内容。因此，可提高向用户推送内容的精准度。并且，该方法无需进行人工筛选组合特征，可减少人工量。
本申请附加的方面和优点将在下面的描述中部分给出，这些将从下面的描述中变得明显，或通过本申请的实践了解到。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例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一些实施例，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，还可以根据这些附图获得其他的附图。
图1为本申请提供的一种内容推送方法的一实施例中的方法流程图；
图2为本申请提供的一种内容推送方法的另一实施例中的方法流程图；
图3为本申请提供的基于特征组合的点击率预估模型的一实施例中的结构框图；
图4为本申请提供的基于特征组合的点击率预估模型的一具体实施例中的结构框图；
图5为本申请提供的向量相乘组件中两向量相乘的一实施例中的示意图；
图6为本申请提供的一种计算机设备的一实施例中的结构示意图。
具体实施方式
下面详细描述本申请的实施例，本申请实施例的示例在附图中示出，其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的，仅用于解释本申请，而不能解释为对本申请的限制。
本技术领域技术人员可以理解，除非特意声明，这里使用的单数形式“一”、“一个”、“所述”和“该”也可包括复数形式，这里使用的“第一”、“第二”仅用于区别同一技术特征，并不对该技术特征的顺序和数量等加以限定。应该进一步理解的是，本申请的说明书中使用的措辞“包括”是指存在特征、整数、步骤、操作、元件和/或组件，但是并不排除存在或添加一个或多个其他特征、整数、步骤、操作、元件、组件和/或它们的组。
本技术领域技术人员可以理解，除非另外定义，这里使用的所有术语(包括技术术语和科学术语)，具有与本申请所属领域中的普通技术人员的一般理解相同的意义。还应该理解的是，诸如通用字典中定义的那些术语，应该被理解为具有与现有技术的上下文中的意义一致的意义，并且除非像这里一样被特定定义，否则不会用理想化或过于正式的含义来解释。
本申请提供一种内容推送方法，用于根据用户对待推送内容的预估点击率，定向向用户推送相关内容，以提高推送内容的精准度。以下先对内容推送方法的相关背景进行先导性说明：
点击率预估：预测“如果向某位用户推荐某个信息/广告等内容时，该用户会点击推送内容”的概率。
特征组合：通过将单独的特征进行组合(相乘或求笛卡尔积等方式)而形成的合成特征。其中，特征组合有助于表示特征之间的非线性关系。
嵌套(embedding)：将高维度稀疏离散化的特征转换成低维度连续值特征的方法。
独热向量(one-hot vector)：只有一个值为1，其他值均为0的稀疏离散向量。
现有的点击率预估深度神经模型，在嵌套embedding时，针对一个输入的特征，只会生成一个低维度的连续值特征向量，以与其他特征做组合。假设已有特征1为衣服，特征2为性别，特征3为年龄。其中，衣服与年龄，衣服与性别的关联程度是不一样的。假设衣服对应的向量为(0.3,0.4)，性别对应的向量为(x1,y1)，性别对应的向量为(x2,y2)，用特征向量的点积作为特征组合的重要性的值，则衣服提供的权重都是0.5(即向量的模)，该值对于年龄和性别都是一样的。但显然，衣服与年龄的组合，和衣服与性别的组合，是完全不一样的两种方式。针对这个问题，本申请提出了一种新的点击率预估神经网络模型，以将特征进行交叉，深度挖掘交叉后特征之间的关联性，最终输出点击的概率 值，根据该概率值判定用户点击商品概率。
在一实施例中，如图1所示，本申请提供内容推送方法，包括以下步骤：
S100，提取待推送内容的多个相关特征。
在本实施例中，服务器提取待推送内容的多个相关特征。其中。待推送内容可以是商品、广告、短视频等内容。多个相关特征可以包括待推送内容本身特性的内容特征以及与该待推送内容关联的非本身特性的特征。如，待推送内容为短视频内容时，其多个相关特征可包括观看短视频的用户特征以及短视频本身特征(如观看人数、点赞数、视频标签等)。
S200，将多个相关特征输入基于特征组合的点击率预估模型，得到待推送内容的预估点击率；其中，基于特征组合的点击率预估模型用于将多个相关特征进行组合，根据组合后的相关特征的关联性，确定出待推送内容的预估点击率。
在本实施例中，基于特征组合的点击率预估模型为用于将输入的多个特征进行特征组合，并分析组合后特征之间的关联性，根据特征的关联性确定出用户对待推送内容的预估点击率。服务器将待推送内容的多个相关特征输入该基于特征组合的点击率预估模型中，根据组合后特征的相关性得到待推送内容的预估点击率。
在一实施例中，如图3所示，基于特征组合的点击率预估模型包括：输入层10、逻辑回归组件20、嵌套模块30、内积组件40、拼接模块50、隐藏层模块60以及预估点击率运算模块70。
输入层10用于对多个相关特征进行独热编码，得到独热向量。具体地，输入层10接收多个相关特征，将该多个相关特征进行独热编码，得到多个独热向量。每个独热向量为向量内只有一个值为1，其他值均为0的稀疏离散向量。逻辑回归组件20用于对输入层10输出的独热向量进行逻辑回归运算，得到第一运算值。因此，可挖掘待推送内容的简单有效的独立特征。嵌套模块30用于将输入层10输出的独热向量的高维度稀疏离散化特征转换成低维度连续值特征，得到低维度连续值向量。其中，一个独热向量输入嵌套模块30后，得到多个低维度连续值向量。
内积组件40用于将嵌套模块30输出的低维度连续值向量进行向量内积，得到第二运算值。向量内积可看作本申请提供的特征组合的一种方式。此部分对低维度连续值向量进行向量内积，可直观地表示两个特征之间的组合关系。 在一实施方式中，将嵌套模块输出的低维度连续值向量进行向量内积，得到第二运算值，包括：获取每个低维度连续值向量对应的权重；将每个低维度连续值乘以对应的权重后进行向量内积，得到第二运算值。在该实施方式中，每个低维度连续值向量均对应一个权重，并且该权重在模型训练过程中会自动更显。服务器每次获取更新后的每个低维度连续值向量对应的权重，将每个低维度连续值乘以对应的权重后进行向量内积，得到第二运算值。
拼接模块50用于将嵌套模块30输出的低维度连续值向量进行向量拼接，得到拼接向量。此处，向量拼接也可看成特征组合的方式。通过向量拼接之后，将结果输入多层的神经网络，可更进一步挖掘深度的特征组合关系。在一实施方式中，将嵌套模块输出的低维度连续值向量进行向量拼接，得到拼接向量，包括：获取低维度连续值向量对应的权重；将每个低维度连续值向量乘以对应的权重后，进行向量累加，得到拼接向量。在该实施方式中，每个低维度连续值向量均对应一个权重，并且该权重在模型训练过程中会自动更显。服务器每次获取更新后的每个低维度连续值向量对应的权重，将每个低维度连续值乘以对应的权重后进行向量累加，累加后得到拼接向量。
隐藏层模块60用于将拼接模块50输出的拼接向量输入深度神经网络隐藏层，得到第三运算值。此处，深度神经网络可以是Deep and Wide，FNN(Factorisation-machine supported Neural Networks)，PNN(Factorisation-machine supported Neural NetworksProduct-based Neural Networks)，DeepFM(Deep Factorisation-machine)等。预估点击率运算模块70用于根据逻辑回归组件20输出的第一运算值、内积组件40输出的第二运算值及隐藏层模块60输出的第三运算值计算出待推送内容的预估点击率。
进一步地，根据第一运算值、第二运算值及第三运算值计算出待推送内容的预估点击率，包括：将第一运算值、第二运算值及第三运算值拼接后进行归一化运算，得到待推送内容的预估点击率。此处所述的拼接，可以是将第一运算值、第二运算值及第三运算值进行累加。具体地，在一实施方式中，获取第一运算值对应的第一权重、第二运算值对应的第二权重、第三运算值对应的第三权重；将第一运算值乘以第一权重，得到第一值；第二运算值乘以第二权重，得到第二值；第三运算值乘以第三权重，得到第三值；将第一值、第二值和第三值进行累加，得到待推送内容的预估点击率。也即是，累加计算过程中，不同的运算值对应不同的权重。将运算值乘以其对应的权重之后，再进行累加， 最终得到的累加结果值即为待推送内容的预估点击率。
以下基于上述基于特征组合的点击率预估模型给出一个具体实施例，以进一步说明该基于特征组合的点击率预估模型。具体参见图4所示：
参见图4，由底向上看模型的结构图：
最底部的是输入层100，原始输入的独热向量是多个离散化one-hot向量(独热向量)，实体点表示值为1，空心点都表示值为0。为了挖掘简单有效的独立特征，在模型中设计了一个logistics regression(LR)的组件200。
往上一层，对one-hot向量进行嵌套embedding，得到低维度连续值向量300。需要说明的是，每一个one-hot向量，会生成一组连续值特征向量，在与不同的特征做组合(此处指向量拼接或者向量内积)时，使用不同的特征向量。假设以衣服、年龄、性别为例，衣服会生成两个连续值特征向量(0.3,0.4)和(0.6,0.8)。衣服和年龄组合时，使用(0.3,0.4)，此时衣服的权重为0.5；衣服和性别组合时，使用(0.6,0.8)，此时衣服的权重为1.0。
服务器进一步对低维度连续值向量300进行向量内积400，具体为进行向量间的内积，以挖掘简单有效的特征组合方式。服务器还会对低维度连续值向量300进行向量拼接，将拼接结果输入到多层神经网络隐藏层(Hidden Layers)600中，进一步挖掘复杂的特征组合方式，提升模型的学习能力。最终，将LR 400的输出，向量内积400的输出，多层隐藏层(Hidden Layers)600的输出，拼接一起，做一个归一化运算700后输出对应的预估点击率。其中，向量内积和向量拼接都可以看成特征组合的方式。向量拼接之后将输入到多层的神经网络，可挖掘深度的特征组合关系，而向量内积可以很直观地表示两个特征之间的组合关系。
需要说明的是，如图4所示，带权重的边(附图中输入层100与RL 200的连接线、输入层100与低维度连续值向量300的连接线、低维度连续值向量300与向量内积400连接线、低维度连续值向量300与向量拼接500的连接线、向量拼接500与隐藏层(Hidden Layers)600的连接线以及RL 200、向量内积400、隐藏层(Hidden Layers)600分别与归一化运算700的连接线等)为神经网络的普遍的连接边(每条边的权重不一样，在模型训练的过程中，自动更新权重)。权重为1的边(RL 200、向量内积400、隐藏层(Hidden Layers)600分别与归一化运算700的连接线)不会更新，恒为1的边，具体可以体现为直接相加形式。带权重的边表示在运算之前，对应的向量或者值需乘以对应的权重后再进 行后续的运算。
以下提供一个具体的应用场景，以说明上述基于特征组合的点击率预估模型的应用：
以电商为例，底部输入的one-hot向量可以包括用户的特征(比如年龄、性别、所在城市等)，商品的特征(商品类别、商品价格、商品历史购买量等)。将这些特征作为基于特征组合的点击率预估模型的输入，模型在训练过程中自动将用户特征与商品特征进行交叉，深度挖掘两者之间的关联性。最终，模型输出一个用户点击商品的概率值，也即是预估点击率值，根据该概率值来判定用户点击该商品的概率。
在一实施例中，服务器对基于特征组合的点击率预估模型进行优化，优化后的基于特征组合的点击率预估模型包括输入层10、逻辑回归组件20、嵌套模块30、拼接模块50、隐藏层模块60以及预估点击率运算模块70，并将原来的内积组件40进行优化修改。具体将内积组件40替换为向量相乘组件。向量相乘组件用于通过并行计算方式，采用矩阵乘法对二嵌套模块输出的低维度连续值进行向量相乘，得到结果值向量；获取结果值向量上三角的值，将上三角的值进行累加，将累加结果值作为第二运算值。具体地，参见图5所示，将向量间内积的操作，变成矩阵乘法，然后取上三角的方式获取上三角的三个值，将该三个值进行累加之后得到第二运算值。利用GPU自带的并行计算方式，可以大大减少模型的训练时间。
S300，根据待推送内容的预估点击率，向用户推送内容。
在本实施例中，服务器获取多个待推送内容的预估点击率，并根据每个待推送内容的预估点击率确定出向用户进行推送的推送内容。在一实施例中，步骤S300包括：获取多个待推送内容的预估点击率，根据每个待推送内容的预估点击率从高到低对待推送内容进行排序，获取排序靠前的预设数量的待推送内容，并向用户推送该预设数量的待推送内容。或者，获取预估点击率大于预设值的待推送内容，向用户推送该预估点击率大于预设值的待推送内容。
上述实施例提供的内容推送方法，提取待推送内容的多个相关特征，将该多个相关特征输入基于特征组合的点击率预估模型中，通过点击率预估模型将多个相关特征进行组合，并分析组合后的相关特征的关联性，根据关联性确定出待推送内容的预估点击率，从而可以根据多个待推送内容的预估点击率向用户推送内容。可以是，从多个待推送内容中筛选出预估点击率较高的推送内容， 并向用户推送该部分的推送内容。因此，可提高向用户推送内容的精准度。并且，该方法无需进行人工筛选组合特征，可减少人工量。
在一实施例中，如图2所示，待推送内容包括待推送的短视频内容。步骤S100，包括：
S110，提取待推送的短视频内容的多个用户特征以及多个短视频内容特征。
步骤S200，包括：
S210，将多个用户特征以及多个短视频内容特征输入基于特征组合的点击率预估模型，得到待推送的短视频内容的预估点击率。
步骤S300，包括：
S310，根据待推送的短视频内容的预估点击率，向用户推送短视频内容。
在该实施例中，待推送内容为待推送的短视频内容。服务器提取待推送的短视频内容的相关特征，如多个用户特征以及多个短视频内容特征。具体地，相关特征可包括目标用户的用户特征，如用户性别、用户年龄、用户职业以及用户常驻地等。相关特征还可包括短视频内容本身特征，如短视频内容的观看人数、点赞人数、视频标签等。将待推送的短视频内容的多个相关特征输入基于特征组合的点击率预估模型中，通过模型将多个用户特征以及多个短视频内容特征进行组合交叉，挖掘组合后相关特征的关联性，从而输出该待推送短视频内容的预估点击率值。最终，根据该预估点击率值向用户进行个性化推送短视频内容。
本申请还提供一种存储介质。该存储介质上存储有计算机程序；该计算机程序被处理器执行时，实现上述任一实施例提供的内容推送方法。该存储介质可以是存储器。例如，内存储器或外存储器，或者包括内存储器和外存储器两者。内存储器可以包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦写可编程ROM(EEPROM)、快闪存储器、或者随机存储器。外存储器可以包括硬盘、软盘、ZIP盘、U盘、磁带等。本申请所公开的存储介质包括但不限于这些类型的存储器。本申请所公开的存储器只作为例子而非作为限定。
本申请还提供一种计算机设备。一种计算机设备包括：一个或多个处理器；存储器；一个或多个应用程序。其中一个或多个应用程序被存储在存储器中并被配置为由一个或多个处理器执行，一个或多个应用程序配置用于执行上述任一实施例提供的内容推送方法。
图6为本申请一实施例中的计算机设备的结构示意图。本实施例提供的计算机设备可以是服务器、个人计算机以及网络设备。如图6所示，设备包括处理器603、存储器605、输入单元607以及显示单元609等器件。本领域技术人员可以理解，图6示出的设备结构器件并不构成对所有设备的限定，可以包括比图示更多或更少的部件，或者组合某些部件。存储器605可用于存储应用程序601以及各功能模块，处理器603运行存储在存储器605的应用程序601，从而执行设备的各种功能应用以及数据处理。存储器可以是内存储器或外存储器，或者包括内存储器和外存储器两者。内存储器可以包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦写可编程ROM(EEPROM)、快闪存储器、或者随机存储器。外存储器可以包括硬盘、软盘、ZIP盘、U盘、磁带等。本申请所公开的存储器包括但不限于这些类型的存储器。本申请所公开的存储器只作为例子而非作为限定。
输入单元607用于接收信号的输入，以及接收用户输入的关键字。输入单元607可包括触控面板以及其它输入设备。触控面板可收集用户在其上或附近的触摸操作(比如用户使用手指、触笔等任何适合的物体或附件在触控面板上或在触控面板附近的操作)，并根据预先设定的程序驱动相应的连接装置；其它输入设备可以包括但不限于物理键盘、功能键(比如播放控制按键、开关按键等)、轨迹球、鼠标、操作杆等中的一种或多种。显示单元609可用于显示用户输入的信息或提供给用户的信息以及计算机设备的各种菜单。显示单元609可采用液晶显示器、有机发光二极管等形式。处理器603是计算机设备的控制中心，利用各种接口和线路连接整个电脑的各个部分，通过运行或执行存储在存储器605内的软件程序和/或模块，以及调用存储在存储器内的数据，执行各种功能和处理数据。
在一实施方式中，设备包括一个或多个处理器603，以及一个或多个存储器605，一个或多个应用程序601。其中一个或多个应用程序601被存储在存储器605中并被配置为由一个或多个处理器603执行，一个或多个应用程序601配置用于执行以上实施例提供的内容推送方法。
此外，在本申请各个实施例中的各功能单元可以集成在一个处理模块中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。上述集成的模块如果以软件功能模块的形式实现并作为独立的产品 销售或使用时，也可以存储在一个计算机可读取存储介质中。
本领域普通技术人员可以理解实现上述实施例的全部或部分步骤可以通过硬件来完成，也可以通过程序来指令相关的硬件完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括存储器、磁盘或光盘等。
以上所述仅是本申请的部分实施方式，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。
应该理解的是，在本申请各实施例中的各功能单元可集成在一个处理模块中，也可以各个单元单独物理存在，也可以两个或两个以上单元集成于一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。

Claims (10)

  1. 一种内容推送方法，其特征在于，包括：
    提取待推送内容的多个相关特征；
    将所述多个相关特征输入基于特征组合的点击率预估模型，得到所述待推送内容的预估点击率；其中，所述基于特征组合的点击率预估模型用于将所述多个相关特征进行组合，根据组合后的所述相关特征的关联性，确定出所述待推送内容的预估点击率；
    根据所述待推送内容的预估点击率，向用户推送内容。
  2. 根据权利要求1所述的方法，其特征在于，所述待推送内容包括待推送的短视频内容；所述提取待推送内容的多个相关特征，包括：提取所述待推送的短视频内容的多个用户特征以及多个短视频内容特征；
    所述将所述多个相关特征输入基于特征组合的点击率预估模型，得到所述待推送内容的预估点击率，包括：将所述多个用户特征以及所述多个短视频内容特征输入所述基于特征组合的点击率预估模型，得到所述待推送的短视频内容的预估点击率；
    所述根据所述待推送内容的预估点击率，向用户推送内容，包括：根据所述待推送的短视频内容的预估点击率，向用户推送短视频内容。
  3. 根据权利要求1所述的方法，其特征在于，所述基于特征组合的点击率预估模型包括：
    输入层，用于对所述多个相关特征进行独热编码，得到独热向量；
    逻辑回归组件，用于对所述输入层输出的所述独热向量进行逻辑回归运算，得到第一运算值；
    嵌套模块，用于将所述独热向量的高维度稀疏离散化特征转换成低维度连续值特征，得到低维度连续值向量；
    内积组件，用于将所述嵌套模块输出的所述低维度连续值向量进行向量内积，得到第二运算值；
    拼接模块，用于将所述嵌套模块输出的所述低维度连续值向量进行向量拼接，得到拼接向量；
    隐藏层模块，用于将所述拼接模块输出的所述拼接向量输入深度神经网络隐藏层，得到第三运算值；
    预估点击率运算模块，用于根据所述第一运算值、所述第二运算值及所述第三运算值计算出所述待推送内容的预估点击率。
  4. 根据权利要求3所述的方法，其特征在于，所述根据所述第一运算值、所述第二运算值及所述第三运算值计算出所述待推送内容的预估点击率，包括：
    将所述第一运算值、所述第二运算值及所述第三运算值拼接后进行归一化运算，得到所述待推送内容的预估点击率。
  5. 根据权利要求4所述的方法，其特征在于，所述将所述第一运算值、所述第二运算值及所述第三运算值拼接后进行归一化运算，得到所述待推送内容的预估点击率，包括：
    获取所述第一运算值对应的第一权重、所述第二运算值对应的第二权重、所述第三运算值对应的第三权重；
    将所述第一运算值乘以所述第一权重，得到第一值；所述第二运算值乘以所述第二权重，得到第二值；所述第三运算值乘以所述第三权重，得到第三值；
    将所述第一值、所述第二值和所述第三值进行累加，得到所述待推送内容的预估点击率。
  6. 根据权利要求3所述的方法，其特征在于，所述将所述嵌套模块输出的所述低维度连续值向量进行向量内积，得到第二运算值，包括：
    获取每个低维度连续值向量对应的权重；
    将所述每个低维度连续值乘以对应的权重后进行所述向量内积，得到所述第二运算值。
  7. 根据权利要求3所述的方法，其特征在于，所述将所述嵌套模块输出的所述低维度连续值向量进行向量拼接，得到拼接向量，包括：
    获取所述低维度连续值向量对应的权重；
    将所述每个低维度连续值向量乘以对应的权重后，进行向量累加，得到所述拼接向量。
  8. 根据权利要求1所述的方法，其特征在于，所述基于特征组合的点击率预估模型包括：
    输入层，用于对所述多个相关特征进行独热编码，得到独热向量；
    逻辑回归组件，用于对所述输入层输出的所述独热向量进行逻辑回归运算，得到第一运算值；
    嵌套模块，用于将所述独热向量的高维度稀疏离散化特征转换成低维度连续值特征，得到低维度连续值向量；
    向量相乘组件，用于通过并行计算方式，采用矩阵乘法对所述嵌套模块输出的所述低维度连续值向量进行向量相乘，得到结果值向量；获取所述结果值向量上三角的值，将所述上三角的值进行累加，将累加结果值作为第二运算值；
    拼接模块，用于将所述嵌套模块输出的所述低维度连续值向量进行向量拼接，得到拼接向量；
    隐藏层模块，用于将所述拼接模块输出的所述拼接向量输入深度神经网络隐藏层，得到第三运算值；
    预估点击率运算模块，用于根据所述第一运算值、所述第二运算值及所述第三运算值计算出所述待推送内容的预估点击率。
  9. 一种存储介质，其特征在于，其上存储有计算机程序；所述计算机程序适于由处理器加载并执行上述权利要求1至8中任一项所述的内容推送方法。
  10. 一种计算机设备，其特征在于，其包括：
    一个或多个处理器；
    存储器；
    一个或多个应用程序，其中所述一个或多个应用程序被存储在所述存储器中并被配置为由所述一个或多个处理器执行，所述一个或多个应用程序配置用于执行根据权利要求1至8任一项所述的内容推送方法。
PCT/CN2019/113566 2018-12-07 2019-10-28 内容推送方法及存储介质、计算机设备 WO2020114145A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811497834.2 2018-12-07
CN201811497834.2A CN109359247B (zh) 2018-12-07 2018-12-07 内容推送方法及存储介质、计算机设备

Publications (1)

Publication Number Publication Date
WO2020114145A1 true WO2020114145A1 (zh) 2020-06-11

Family

ID=65331724

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/113566 WO2020114145A1 (zh) 2018-12-07 2019-10-28 内容推送方法及存储介质、计算机设备

Country Status (2)

Country Link
CN (1) CN109359247B (zh)
WO (1) WO2020114145A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359247B (zh) * 2018-12-07 2021-07-06 广州市百果园信息技术有限公司 内容推送方法及存储介质、计算机设备
CN111597429A (zh) * 2019-02-21 2020-08-28 北京京东尚科信息技术有限公司 网络资源推送方法、装置及存储介质
CN109919670A (zh) * 2019-02-27 2019-06-21 重庆金窝窝网络科技有限公司 广告点击概率的预测方法、装置、服务器和存储介质
CN109948062A (zh) * 2019-03-29 2019-06-28 广州市百果园网络科技有限公司 一种目标匹配方法、装置、服务器、系统及存储介质
CN110442810B (zh) * 2019-08-08 2023-06-13 广州华建工智慧科技有限公司 一种基于DeepFM推荐算法的移动端BIM模型智能缓存方法
CN110502715B (zh) * 2019-08-28 2023-07-14 腾讯科技（深圳）有限公司 点击概率的预测方法及装置
CN111611488B (zh) * 2020-05-21 2022-09-30 腾讯科技（深圳）有限公司 基于人工智能的信息推荐方法、装置、电子设备
CN111339433B (zh) * 2020-05-21 2020-08-21 腾讯科技（深圳）有限公司 基于人工智能的信息推荐方法、装置、电子设备
CN111967599B (zh) * 2020-08-25 2023-07-28 百度在线网络技术（北京）有限公司 用于训练模型的方法、装置、电子设备及可读存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107613022A (zh) * 2017-10-20 2018-01-19 广州优视网络科技有限公司 内容推送方法、装置及计算机设备
CN108470027A (zh) * 2018-03-30 2018-08-31 广州优视网络科技有限公司 色情应用识别方法和装置、计算设备和存储介质
CN109359247A (zh) * 2018-12-07 2019-02-19 广州市百果园信息技术有限公司 内容推送方法及存储介质、计算机设备

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101751376B (zh) * 2009-12-30 2012-03-21 中国人民解放军国防科学技术大学 利用cpu和gpu协同工作对三角线性方程组求解的加速方法
US9170836B2 (en) * 2013-01-09 2015-10-27 Nvidia Corporation System and method for re-factorizing a square matrix into lower and upper triangular matrices on a parallel processor
CN105095625B (zh) * 2014-05-14 2018-12-25 阿里巴巴集团控股有限公司 点击率预估模型建立方法、装置及信息提供方法、系统
CN105046515B (zh) * 2015-06-26 2017-12-22 深圳市腾讯计算机系统有限公司 广告排序方法及装置
CN105701068B (zh) * 2016-02-19 2018-06-19 南京大学 基于分时复用技术的cholesky矩阵求逆系统
CN106339510B (zh) * 2016-10-28 2019-12-06 北京百度网讯科技有限公司 基于人工智能的点击预估方法及装置
CN107145518B (zh) * 2017-04-10 2020-04-14 同济大学 一种社交网络下基于深度学习的个性化推荐系统
CN108053050A (zh) * 2017-11-14 2018-05-18 广州优视网络科技有限公司 点击率预估方法、装置、计算设备及存储介质

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107613022A (zh) * 2017-10-20 2018-01-19 广州优视网络科技有限公司 内容推送方法、装置及计算机设备
CN108470027A (zh) * 2018-03-30 2018-08-31 广州优视网络科技有限公司 色情应用识别方法和装置、计算设备和存储介质
CN109359247A (zh) * 2018-12-07 2019-02-19 广州市百果园信息技术有限公司 内容推送方法及存储介质、计算机设备

Also Published As

Publication number Publication date
CN109359247A (zh) 2019-02-19
CN109359247B (zh) 2021-07-06

Similar Documents

Publication Publication Date Title
WO2020114145A1 (zh) 内容推送方法及存储介质、计算机设备
Onar et al. A new hesitant fuzzy QFD approach: an application to computer workstation selection
US10783361B2 (en) Predictive analysis of target behaviors utilizing RNN-based user embeddings
CN111815415B (zh) 一种商品推荐方法、系统及设备
US20220156302A1 (en) Implementing a graphical user interface to collect information from a user to identify a desired document based on dissimilarity and/or collective closeness to other identified documents
US10102277B2 (en) Bayesian visual interactive search
CN111339415B (zh) 一种基于多交互式注意力网络的点击率预测方法与装置
KR102472572B1 (ko) 사용자 의도 프로파일링 방법 및 이를 위한 장치
KR101778679B1 (ko) 딥러닝을 이용하여 텍스트 단어 및 기호 시퀀스를 값으로 하는 복수 개의 인자들로 표현된 데이터를 자동으로 분류하는 방법 및 시스템
US11055765B2 (en) Classification of query text to generate relevant query results
WO2017190610A1 (zh) 目标用户定向方法、装置和计算机存储介质
CN112163676B (zh) 多任务服务预测模型训练方法、装置、设备以及存储介质
JP2021118000A (ja) アイテム推奨方法、システム、プログラム、電子機器及び記憶媒体
WO2017088496A1 (zh) 一种搜索推荐方法、装置、设备及计算机存储介质
US20150120432A1 (en) Graph-based ranking of items
WO2019099913A1 (en) Aspect pre-selection using machine learning
CN114663197A (zh) 商品推荐方法及其装置、设备、介质、产品
CN114240555A (zh) 训练点击率预测模型和预测点击率的方法和装置
CN112288483A (zh) 用于训练模型的方法和装置、用于生成信息的方法和装置
CN114862480A (zh) 广告投放定向方法及其装置、设备、介质、产品
US11017452B2 (en) Concerted learning and multi-instance sequential prediction tree
Mostofi et al. Real-estate price prediction with deep neural network and principal component analysis
JP7204903B2 (ja) 情報プッシュ方法、装置、機器及び記憶媒体
Lee Extraction of competitive factors in a competitor analysis using an explainable neural network
CN115907928A (zh) 商品推荐方法及其装置、设备、介质

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: 19892994

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: 19892994

Country of ref document: EP

Kind code of ref document: A1