CN113888238A - Advertisement click rate prediction method and device and computer equipment - Google Patents
Advertisement click rate prediction method and device and computer equipment Download PDFInfo
- Publication number
- CN113888238A CN113888238A CN202111243596.4A CN202111243596A CN113888238A CN 113888238 A CN113888238 A CN 113888238A CN 202111243596 A CN202111243596 A CN 202111243596A CN 113888238 A CN113888238 A CN 113888238A
- Authority
- CN
- China
- Prior art keywords
- user
- interest
- advertisement
- vector
- data
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 47
- 239000013598 vector Substances 0.000 claims abstract description 73
- 238000012549 training Methods 0.000 claims abstract description 22
- 230000007246 mechanism Effects 0.000 claims abstract description 18
- 238000013528 artificial neural network Methods 0.000 claims abstract description 13
- 239000013604 expression vector Substances 0.000 claims abstract description 5
- 238000007781 pre-processing Methods 0.000 claims abstract description 3
- 230000006399 behavior Effects 0.000 claims description 95
- 230000006870 function Effects 0.000 claims description 26
- 230000008569 process Effects 0.000 claims description 15
- 238000004590 computer program Methods 0.000 claims description 13
- 238000000605 extraction Methods 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 11
- 230000003068 static effect Effects 0.000 claims description 11
- 230000000306 recurrent effect Effects 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 125000004122 cyclic group Chemical group 0.000 claims description 2
- 238000009499 grossing Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 2
- 239000010410 layer Substances 0.000 description 26
- 238000010586 diagram Methods 0.000 description 9
- 230000009471 action Effects 0.000 description 7
- 238000004088 simulation Methods 0.000 description 6
- 230000003993 interaction Effects 0.000 description 4
- 238000011161 development Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000003860 storage Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- 230000002238 attenuated effect Effects 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000006397 emotional response Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0242—Determining effectiveness of advertisements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
- G06Q30/0255—Targeted advertisements based on user history
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Accounting & Taxation (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Entrepreneurship & Innovation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Game Theory and Decision Science (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Description
技术领域technical field
本发明属于电商大数据推荐领域,特别涉及一种基于用户兴趣与时序行为的广告点击率预测方法、装置及计算机设备。The invention belongs to the field of e-commerce big data recommendation, and in particular relates to a method, device and computer equipment for predicting the click-through rate of advertisements based on user interests and timing behaviors.
背景技术Background technique
随着信息技术的发展,许多国内外互联网电商平台越来越关注在线广告系统的盈利效果,注重实现个性化、精准化的营销策略。广告点击率(CTR,Click Through Rate)是电商平台系统中最核心的指标之一,在广告推荐、网页搜索、赞助推荐等领域至关重要。点击率预测的准确度不仅会影响电商平台的收益,还会影响用户的满意度和消费体验。With the development of information technology, many domestic and foreign Internet e-commerce platforms pay more and more attention to the profitability of online advertising systems, and pay attention to the realization of personalized and precise marketing strategies. Click Through Rate (CTR) is one of the most core indicators in the e-commerce platform system, and it is very important in the fields of advertising recommendation, web search, sponsorship recommendation and so on. The accuracy of click-through rate prediction will not only affect the revenue of e-commerce platforms, but also affect user satisfaction and consumption experience.
在当前的电商平台中,尽管营销人员想知道网络访问者的反应,但是使用当前技术几乎不可能量化对网站的情感反应以及该网站对公司品牌的影响。不过,点击率却是很容易获得。点击率衡量的是页面访问者数量与该页面商品广告点击后并将其重定向到另一个页面的访问者的比例,在该页面中,他们可以购买商品或了解有关产品或服务的更多信息。通常,点击率越高,则表明该广告商品更有商业价值或是该营销活动更吸引人。大多数电商网站旨在通过点击率来调整主页商品广告的展示,做个性化推荐。In current e-commerce platforms, while marketers want to know how web visitors react, it is nearly impossible with current technology to quantify the emotional response to a website and the impact that website has on a company’s brand. However, CTR is easy to obtain. Click-through rate measures the ratio of visitors to a page to those who click on an ad for an item on that page and are redirected to another page where they can buy an item or learn more about a product or service . Generally, a higher CTR indicates that the advertised item is more commercially valuable or the marketing campaign is more appealing. Most e-commerce sites aim to adjust the display of homepage product advertisements through click-through rate and make personalized recommendations.
目前,许多国内外学者都对CTR模型展开了深入的研究,研究成果主要体现在以下几个方面:一方面,随着深度学习技术的发展,深度CTR模型逐步取代了需要人工特征工程的LR等基于机器学习的CTR模型。另一方面,一些深度CTR模型注重于特征的压缩与交互。此外,也有模型重点关注用户行为序列特征的提取。但是现阶段的广告点击率仍存在以下不足:At present, many domestic and foreign scholars have carried out in-depth research on the CTR model, and the research results are mainly reflected in the following aspects: On the one hand, with the development of deep learning technology, the deep CTR model has gradually replaced the LR that requires artificial feature engineering. Machine learning based CTR model. On the other hand, some deep CTR models focus on feature compression and interaction. In addition, there are also models that focus on the extraction of user behavior sequence features. However, the current CTR still has the following shortcomings:
1.用户历史行为序列的时效性。传统时序模型忽略了顺序行为之间的时间间隔对用户兴趣表达的影响,传统的RNN可以很好地捕捉行为序列中的顺序关系之间的依赖关系,但用户行为不仅仅是顺序关系,行为的时间间隔和行为的特点等含有更多的先验信息,这些信息对用户兴趣的表示至关重要。1. Timeliness of user historical behavior sequences. The traditional time series model ignores the influence of the time interval between sequential actions on the expression of user interest. The traditional RNN can well capture the dependencies between the sequential relationships in the behavior sequence, but the user behavior is not only the sequential relationship. Time interval and behavioral characteristics contain more prior information, which is crucial to the representation of user interest.
2.用户兴趣的泛化性和复杂性。用户的兴趣具有多样性并且有变化的趋势,用户在某一段时间内的喜好具有集中性,并且每种兴趣都有自己的演变趋势,不同种类的兴趣之间很少相互影响。2. Generalization and complexity of user interests. The interests of users are diverse and have a changing trend. The preferences of users in a certain period of time are centralized, and each interest has its own evolution trend. Different kinds of interests rarely affect each other.
3.数据特征的维度高,隐性信息量大。电商平台广告数据输入特征除了用户行为序列特征外,上下文特征、广告特征等特征之间的关系也影响着点击率预估的准确度。这些特征维度高,隐含信息大,获取它们之间的关系变得困难。3. The dimension of data features is high, and the amount of implicit information is large. In addition to user behavior sequence features, the relationship between contextual features, advertising features and other features also affects the accuracy of click-through rate estimation. These features have high dimension and large implicit information, and it becomes difficult to obtain the relationship between them.
发明内容SUMMARY OF THE INVENTION
针对上述不足,本发明提出一种广告点击率预测方法、装置及计算机设备用以解决广告点击率的预测问题。In view of the above deficiencies, the present invention provides a method, device and computer equipment for predicting the click-through rate of advertisements to solve the problem of predicting the click-through rate of advertisements.
在本发明的第一方面,本发明提供了一种广告点击率预测方法,所述方法包括:In a first aspect of the present invention, the present invention provides a method for predicting a click-through rate of an advertisement, the method comprising:
获取电商平台的用户行为数据,用户肖像数据以及广告数据;Obtain user behavior data, user portrait data and advertising data of e-commerce platforms;
对电商平台的用户行为数据进行预处理,并形成用户行为序列;Preprocess the user behavior data of the e-commerce platform and form a user behavior sequence;
将用户行为序列、用户肖像数据、广告数据分别进行编码表示,得到对应特征的嵌入向量;The user behavior sequence, user portrait data, and advertisement data are encoded and represented respectively, and the embedding vector of the corresponding feature is obtained;
输入用户行为序列特征,采用基于时间因子的Time-GRU的深度神经网络,输出用户的兴趣表示向量;Input the user behavior sequence features, use the Time-GRU deep neural network based on the time factor, and output the user's interest representation vector;
输入用户的兴趣表示向量,采用基于注意力机制的AT-GRU的深度神经网络,模拟兴趣的更新过程,并输出用户的兴趣更新向量;Input the user's interest representation vector, use the AT-GRU deep neural network based on the attention mechanism to simulate the update process of interest, and output the user's interest update vector;
输入用户肖像特征和广告特征,采用堆栈式自动编码机,提取出用户肖像特征与广告特征之间的隐形关系向量;Input user portrait features and advertisement features, and use a stacked automatic encoding machine to extract the invisible relationship vector between user portrait features and advertisement features;
将用户的兴趣表示向量和用户肖像特征与广告特征之间的隐形关系向量输入到多层感知机中进行联合训练,得到广告点击率的预测结果。The user's interest representation vector and the invisible relationship vector between the user's portrait feature and the advertisement feature are input into the multi-layer perceptron for joint training, and the prediction result of the advertisement click rate is obtained.
在本发明的第二方面,本发明还提供了一种广告点击率预测装置,所述装置包括:In a second aspect of the present invention, the present invention further provides a device for predicting click-through rate of advertisements, the device comprising:
获取模块,用于获取电商平台的用户行为数据,用户肖像数据以及广告数据;The acquisition module is used to acquire the user behavior data, user portrait data and advertisement data of the e-commerce platform;
处理模块,用于对电商平台的用户行为数据进行预处理,并形成用户行为序列;The processing module is used to preprocess the user behavior data of the e-commerce platform and form a user behavior sequence;
嵌入模块,用于将用户行为序列、用户肖像数据、广告数据分别进行编码表示,得到对应特征的嵌入向量;The embedding module is used to encode and represent the user behavior sequence, user portrait data, and advertisement data respectively, and obtain the embedding vector of the corresponding feature;
第一特征提取模块,用于输入用户行为序列特征,采用基于时间因子的Time-GRU的深度神经网络,输出用户的兴趣表示向量;The first feature extraction module is used for inputting user behavior sequence features, using a Time-GRU deep neural network based on time factor, and outputting the user's interest representation vector;
第二特征提取模块,用于输入用户的兴趣表示向量,采用基于注意力机制的AT-GRU的深度神经网络,模拟兴趣的更新过程,并输出用户的兴趣更新向量;The second feature extraction module is used to input the user's interest representation vector, adopts the AT-GRU deep neural network based on the attention mechanism, simulates the update process of the interest, and outputs the user's interest update vector;
第三特征提取模块,用于输入用户肖像特征和广告特征,采用堆栈式自动编码机,提取出用户肖像特征与广告特征之间的隐性关系向量;The third feature extraction module is used to input user portrait features and advertisement features, and uses a stacked automatic encoder to extract the implicit relationship vector between user portrait features and advertisement features;
广告点击率预测模块,用于将用户的兴趣表示向量和用户肖像特征与广告特征之间的隐形关系向量输入到多层感知机中进行联合训练,得到广告点击率的预测结果。The advertising click-through rate prediction module is used to input the user's interest representation vector and the invisible relationship vector between the user's portrait feature and the advertisement feature into the multilayer perceptron for joint training, and obtain the prediction result of the advertisement click-through rate.
在本发明的第三方面,本发明还提供了一种计算机设备,包括存储器和处理器,所述存储器中存储有计算机程序,所述计算机程序被所述处理器执行时,使得所述处理器执行如本发明第一方面所述方法的步骤。In a third aspect of the present invention, the present invention also provides a computer device, comprising a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to The steps of the method as described in the first aspect of the present invention are performed.
本发明的有益效果:Beneficial effects of the present invention:
本发明利用电商平台的用户行为数据和目标广告数据,针对用户时序行为序列背后隐藏的用户兴趣,通过对其兴趣更新过程进行更新与建模得到兴趣表示,结合其它非时序特征之间的隐形关联进行广告点击率的预测,本发明能够有效提高电商平台广告的点击率,实现精准营销和推荐的效果。The invention utilizes the user behavior data and target advertisement data of the e-commerce platform, aiming at the user interest hidden behind the user time series behavior sequence, and obtains the interest expression by updating and modeling the interest update process, and combines the invisible between other non-sequential features. The prediction of the click-through rate of advertisements is performed in association, and the present invention can effectively improve the click-through rate of advertisements on the e-commerce platform, and realize the effects of precise marketing and recommendation.
附图说明Description of drawings
图1为本发明实施例中广告点击率预测框架图;1 is a frame diagram of an advertisement click-through rate prediction in an embodiment of the present invention;
图2为本发明实施例中广告点击率预测方法流程图;2 is a flowchart of a method for predicting an advertisement click-through rate in an embodiment of the present invention;
图3为本发明构建时序模型模拟用户兴趣特征示图;FIG. 3 is a schematic diagram of constructing a time series model to simulate user interest characteristics according to the present invention;
图4为本发明构建注意力机制的兴趣更新模型示图;4 is a schematic diagram of an interest update model for constructing an attention mechanism according to the present invention;
图5为本发明引入一种非监督的特征提取方法示图;5 is a schematic diagram of an unsupervised feature extraction method introduced by the present invention;
图6为本发明是私立中广告点击率预测装置结构图。FIG. 6 is a structural diagram of an apparatus for predicting click-through rate of advertisements in a private medium according to the present invention.
具体实施方式Detailed ways
下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
图1为本发明实施例中广告点击率预测框架图,如图1所示,在本实施例的预测框架中,主要包括四个部分,首先,采集用户行为数据,用户肖像数据以及广告数据;对这些数据进行处理后得到用户行为序列、用户肖像特征和广告特征;其次,采用数据特征处理的方式,利用用户行为序列构建出兴趣更新模型,再利用兴趣模拟模型得到用户的兴趣表示向量,同时还利用特征交互模型对用户肖像数据以及广告数据进行处理,提取出隐性关系向量;然后再结合用户最终兴趣表示和非时序特征隐形关系构建出点击率预测模型;利用该点击率预测模型可以预测出广告点击率,还可以根据广告点击率完成精准的广告推送。FIG. 1 is a frame diagram of an advertisement click rate prediction in an embodiment of the present invention. As shown in FIG. 1 , in the prediction frame of this embodiment, it mainly includes four parts. First, user behavior data, user portrait data and advertisement data are collected; After processing these data, the user behavior sequence, user portrait feature and advertisement feature are obtained; secondly, the user behavior sequence is used to construct an interest update model by means of data feature processing, and the interest representation vector of the user is obtained by using the interest simulation model. It also uses the feature interaction model to process user portrait data and advertising data to extract the implicit relationship vector; and then combines the user's final interest representation and the non-sequential feature implicit relationship to build a click-through rate prediction model; using the click-through rate prediction model can predict The click-through rate of advertisements can be calculated, and accurate advertisements can also be pushed according to the click-through rate of advertisements.
图2为本发明实施例中一种广告点击率预测方法流程图,如图2所示,所述方法包括:FIG. 2 is a flowchart of a method for predicting a click-through rate of an advertisement in an embodiment of the present invention. As shown in FIG. 2 , the method includes:
S1、获取电商平台的用户行为数据,用户肖像数据以及广告数据;S1. Obtain user behavior data, user portrait data and advertisement data of the e-commerce platform;
在本发明实施例中,可以获取电商平台的一些基础数据,包括用户历史行为数据、用户肖像数据和广告数据。以淘宝电商平台在线广告展示数据集为例,这个数据集主要是记录了淘宝用户在电商平台的浏览/点击记录,包括用户行为历史、用户肖像和广告基本信息三个部分。用户行为历史数据包括用户ID、广告ID、时间以及是否点击等字段,是体现用户隐含兴趣的时序特征数据,通过对用户行为数据经过处理,可以得到用户行为序列。用户肖像数据包括用户ID、年龄、性别和购物深度等特征信息,反应了用户的基本特征信息。广告基本数据包括广告ID、商品类目ID、商品品牌ID和价格等特征信息,广告数据表明了待推荐广告的一些基础特征,是非时序数据的重要组成部分。In this embodiment of the present invention, some basic data of the e-commerce platform can be acquired, including user historical behavior data, user portrait data, and advertisement data. Take the online advertisement display dataset of Taobao e-commerce platform as an example. This dataset mainly records the browsing/clicking records of Taobao users on the e-commerce platform, including user behavior history, user portrait and basic advertising information. User behavior history data includes fields such as user ID, advertisement ID, time, and whether or not to click. It is time series characteristic data that reflects the user's implicit interest. By processing the user behavior data, the user behavior sequence can be obtained. User portrait data includes user ID, age, gender, shopping depth and other characteristic information, which reflects the basic characteristic information of users. The basic advertising data includes feature information such as advertising ID, commodity category ID, commodity brand ID, and price. The advertising data indicates some basic features of the advertisement to be recommended, and is an important part of the non-sequential data.
在本发明实施例中,对于获取数据的方法,可以从电商平台提供的数据源或直接下载现有公开数据源,这些方式都可以得到原始数据,本发明对此不作限定。In the embodiment of the present invention, for the method of acquiring data, the data source provided by the e-commerce platform or the existing public data source can be directly downloaded, and the original data can be obtained by these methods, which is not limited in the present invention.
通常获取的原始数据都是非结构化的,不能直接用于数据分析。通过简单的数据清洗可以使大部分非结构化数据结构化。例如,删除重复数据、清理无效节点如部分游客数据等。The raw data usually obtained are unstructured and cannot be directly used for data analysis. Most unstructured data can be structured through simple data cleaning. For example, delete duplicate data, clean up invalid nodes such as some visitor data, etc.
S2、对电商平台的用户行为数据进行预处理,并形成用户行为序列;S2. Preprocess the user behavior data of the e-commerce platform, and form a user behavior sequence;
在本发明实施例中,可以去除无效的用户行为数据;举个例子,一些浏览时间过短的用户行为会影响数据的有效性,本发明定义广告点击有效性的浏览阈值为25秒。当用户在一个广告页面浏览时间高于这个阈值时,则认为这个数据是有效的,否则就是无效的数据;将这些无效的数据删除。In this embodiment of the present invention, invalid user behavior data can be removed; for example, some user behaviors with too short browsing time will affect the validity of the data, and the present invention defines the browsing threshold of advertisement click validity as 25 seconds. When the browsing time of a user on an advertisement page is higher than this threshold, the data is considered to be valid, otherwise it is invalid data; the invalid data is deleted.
在本发明实施例中,可以按照用户的数量进行数据的统计,并按照每个用户的ID、浏览广告和时间戳信息拼接出原始的用户行为数据;按照这种方式形成一系列的用户行为数据。In this embodiment of the present invention, data statistics can be performed according to the number of users, and original user behavior data can be spliced out according to each user's ID, browsing advertisement, and timestamp information; in this way, a series of user behavior data can be formed. .
在本发明实施例中,可以采用多重插值的方法对统计后的用户行为数据进行补全;例如当缺失率超过15%时直接去除数据,对于不超过15%的数据值进行补充,其中的补充方式本领域技术人员可以根据实际需求进行设置。In this embodiment of the present invention, the statistical user behavior data can be completed by using multiple interpolation methods; for example, when the missing rate exceeds 15%, the data is directly removed, and the data values that do not exceed 15% are supplemented. The method can be set by those skilled in the art according to actual needs.
在本发明实施例中,构建出基于时间差的用户行为序列;并根据用户ID对用户行为数据进行分组,将用户行为数据以时间的先后顺序排序,构成用户行为序列;对于其中的每一个行为序列,使用下一个行为的时间戳与当前行为的时间戳之差作为时间因子特征。In the embodiment of the present invention, a user behavior sequence based on time difference is constructed; and the user behavior data is grouped according to the user ID, and the user behavior data is sorted in time order to form a user behavior sequence; for each behavior sequence. , using the difference between the timestamp of the next behavior and the timestamp of the current behavior as the time factor feature.
S3、将用户行为序列、用户肖像数据、广告数据分别进行编码表示,得到对应特征的嵌入向量;S3, encode and represent the user behavior sequence, user portrait data, and advertisement data respectively, and obtain the embedding vector of the corresponding feature;
在本发明实施例中,将经过处理得到的用户行为序列、广告数据和用户肖像数据进行one-hot编码,然后还可以分别对数据特征进行归一化处理。然后使用特征嵌入的方法将输入的高维稀疏特征向量转化为低维的稠密向量,得到特征数据的嵌入表示,即可输出低维稠密的用户行为序列特征、广告特征以及用户肖像特征。In the embodiment of the present invention, one-hot encoding is performed on the user behavior sequence, advertisement data, and user portrait data obtained after processing, and then the data features may be normalized respectively. Then, the input high-dimensional sparse feature vector is converted into a low-dimensional dense vector by the feature embedding method, and the embedded representation of the feature data is obtained, and the low-dimensional and dense user behavior sequence features, advertisement features and user portrait features can be output.
在本发明的优选实施例中,所述用户行为序列特征采用了时序建模,即假设用户行为序列特征是用户行为-时间序列特征二元集合U(B,ΔT),具体表示为;In a preferred embodiment of the present invention, the user behavior sequence feature adopts time series modeling, that is, it is assumed that the user behavior sequence feature is a binary set U(B, ΔT) of user behavior-time sequence feature, which is specifically expressed as;
U(B,ΔT)={(b1,δt1),(b2,δt2),…,(bn,δtn)}U(B,ΔT)={(b 1 ,δt 1 ),(b 2 ,δt 2 ),…,(b n ,δt n )}
用户行为-时间序列特征二元集合U(B,ΔT)被定义为B和ΔT构成的二元集合,表示为:User behavior-time series feature binary set U(B, ΔT) is defined as the binary set composed of B and ΔT, expressed as:
B={b1,b2,...,bn}B={b 1 ,b 2 ,...,b n }
ΔT={δt1,δt2,...,δtn|δt1=0,δti=time(bi)-time(bi-1)i>1}ΔT={δt 1 ,δt 2 ,...,δt n |δt 1 =0,δt i =time(b i )-time(b i-1 )i>1}
其中,B表示用户的历史行为序列特征集合,ΔT表示B中相邻两个用户行为对应的时间之差,其中δt1=0表示第一个序列集合对应的时间差为0。Among them, B represents the user's historical behavior sequence feature set, ΔT represents the time difference corresponding to two adjacent user behaviors in B, where δt 1 =0 represents that the time difference corresponding to the first sequence set is 0.
S4、输入用户行为序列特征,采用基于时间因子的Time-GRU的深度神经网络,输出用户的兴趣表示向量;S4. Input the user behavior sequence features, use the Time-GRU deep neural network based on the time factor, and output the user's interest representation vector;
其中,在本发明实施例中,基于时间因子的门控循环单元(Time-Gate RecurrentUnit,简称Time-GRU)采用了时间因子与GRU结合的方式;本发明中模拟用户的兴趣表示向量的过程主要是根据用户行为序列,采用时间门控循环单元学习用户的静态兴趣组状态集合,表示为Intress=Time-GRU(eu)={h'1,h'2,...,h'n}。Among them, in the embodiment of the present invention, the time factor-based gating cyclic unit (Time-Gate Recurrent Unit, Time-GRU for short) adopts the combination of time factor and GRU; the process of simulating the user's interest expression vector in the present invention mainly According to the user behavior sequence, the time-gated recurrent unit is used to learn the user's static interest group state set, which is expressed as Intres s =Time-GRU(e u )={h' 1 ,h' 2 ,...,h' n }.
其中,静态兴趣组状态集合Intress定义为对用户行为-时间序列特征二元组集合经过特征嵌入处理后,经过时序建模后输出的每一时刻的隐状态集合,其中eu表示用户行为-时间序列特征二元组集合的嵌入表示,该集合中每一个隐状态反映了这一时刻从用户行为序列中提取的用户兴趣;Time-GRU(eu)表示对嵌入向量eu采用时间门控循环单元得出的结果;h'n表示第n个隐藏兴趣状态。Among them, the static interest group state set Intres s is defined as the hidden state set at each moment output after time series modeling for the user behavior-time series feature two-tuple set after feature embedding processing, where e u represents user behavior- Embedding representation of the time series feature two-tuple set, each hidden state in the set reflects the user interest extracted from the user behavior sequence at this moment; Time-GRU(e u ) means that the embedding vector e u is time-gated The result from the recurrent unit; h' n denotes the nth hidden interest state.
图3为本发明实施例构建时序模型模拟用户兴趣特征示图,如图3所示,该结构表示的是时序模型中的Time-GRU核结构,该结构通过结合上一个行为序列的隐藏兴趣状态ht-1与当前行为序列中输入用户的行为特征it和时间因子Δt,经过时间门、更新门和重置门的作用,输出下一个隐藏兴趣状态的代表的用户兴趣ht。时间门控循环单元加强了用户行为序列中时间因子对用户兴趣的影响,它的每一个中间隐藏兴趣状态突出了某个时刻体现用户短期、高频行为的用户静态兴趣。FIG. 3 is a diagram illustrating the characteristics of a user's interest in constructing a time sequence model according to an embodiment of the present invention. As shown in FIG. 3 , the structure represents the Time-GRU core structure in the time sequence model. This structure combines the hidden interest state of the previous behavior sequence h t-1 and the input user's behavior feature i t and time factor Δt in the current behavior sequence, through the action of time gate, update gate and reset gate, output the user interest h t represented by the next hidden interest state. The time-gated recurrent unit strengthens the influence of the time factor in the user behavior sequence on the user's interest, and each of its intermediate hidden interest states highlights the user's static interest that reflects the user's short-term and high-frequency behavior at a certain moment.
在本发明的优选实施例中,所述步骤S4具体包括以下步骤:In a preferred embodiment of the present invention, the step S4 specifically includes the following steps:
S41、根据输入的时间间隔和输入特征,计算时间门权重,具体公式如下,Tg=σ(Wt[Δt,it]):其中Δt是时间因子,即当前行为的时间戳与上一个行为的时间戳之差。考虑到在序列化模型中,用户的浏览、点击等偏好行为随时间的衰减符合长尾分布,因此本发明对时间因子Δt加入对数处理,改进后的时间门权重Tt的公式如下:S41. Calculate the time gate weight according to the input time interval and input features. The specific formula is as follows, T g =σ(W t [Δt, i t ]): where Δt is the time factor, that is, the timestamp of the current behavior is the same as the previous one. The difference between the timestamps of the behavior. Considering that in the serialization model, the user's preference behaviors such as browsing and clicking are attenuated with time in accordance with the long-tailed distribution, the present invention adds logarithmic processing to the time factor Δt, and the improved formula for the time gate weight T t is as follows:
Tt=σ(Wt[log(Δt+ζ),it])T t =σ(W t [log(Δt+ζ),i t ])
当输入时间因子Δt越小时,时间门权重Tg越小,对数处理后的时间门权重Tt也越小,当前步骤保留的信息越少,上一步保留的信息越多。即用户两个相邻行为之间的时间间隔较短,两个行为之间的依赖关系就较高;When the input time factor Δt is smaller, the time gate weight T g is smaller, and the time gate weight T t after logarithmic processing is also smaller, the less information is retained in the current step, and the more information is retained in the previous step. That is, the time interval between two adjacent actions of the user is shorter, and the dependency between the two actions is higher;
S42、根据输入的特征、时间间隔和上一个状态的状态,分别更新重置门rt、更新门zt和中间隐藏兴趣状态具体公式如下所示:S42. Update the reset gate r t , the update gate z t and the intermediate hidden interest state respectively according to the input features, the time interval and the state of the previous state The specific formula is as follows:
zt=σ(Wzit+Uzht-1+bz),z t =σ(W z i t +U z h t-1 +b z ),
rt=σ(Writ+Urht-1+br),r t =σ(W r i t +U r h t-1 + br ),
其中,σ是sigmoid激活函数,ο是逐元素相乘,wz,wr,Uz,Ur,Uh∈nH×nH,nH是隐藏层的大小,nI是输入层的大小。it表示Time-GRU的输入向量,表示第t个隐藏兴趣状态的临时状态,zt是更新门(update gate),rt是重置门(reset gate)。zt和rt在sigmoid函数的映射作用下取值范围为0到1;where σ is the sigmoid activation function, ο is the element-wise multiplication, w z , w r , U z ,U r ,U h ∈n H ×n H , where n H is the size of the hidden layer and n I is the size of the input layer. i t represents the input vector of Time-GRU, represents the temporary state of the t-th hidden interest state, z t is the update gate, and r t is the reset gate. z t and r t range from 0 to 1 under the mapping of the sigmoid function;
S43、将时间门权重加入到更新门的更新策略中,具体公式如下所示:S43, adding the weight of the time gate to the update strategy of the update gate, and the specific formula is as follows:
在本发明实施例中,在门结构中新增了时间因子作为输入,同时在内部引入了时间权重作为辅助参与到更新门的更新策略中,使时间因子能够作为一个重要因素参与兴趣模拟的过程中。In the embodiment of the present invention, a time factor is added to the gate structure as an input, and a time weight is introduced internally as an auxiliary to participate in the update strategy of the update gate, so that the time factor can be used as an important factor to participate in the process of interest simulation middle.
S5、输入用户的兴趣表示向量,采用基于注意力机制的AT-GRU的深度神经网络,模拟兴趣的更新过程,并输出用户的兴趣更新向量;S5. Input the user's interest expression vector, adopt the AT-GRU deep neural network based on the attention mechanism, simulate the update process of interest, and output the user's interest update vector;
在本发明实施例中,基于注意力机制的门控循环单元(Attention-GateRecurrent Unit,简称AT-GRU)采用了注意力机制与GRU结合的方式,下来将来详细介绍该网络。In the embodiment of the present invention, the attention mechanism-based gated recurrent unit (Attention-GateRecurrent Unit, AT-GRU for short) adopts the combination of the attention mechanism and the GRU, and the network will be described in detail in the future.
图4为本发明构建注意力机制的兴趣更新模型示图,如图4所示,该结构是一个基于注意力机制的门控循环单元AT-GRU的核结构,该结构的输入部分是表示当前步骤t中经过时序模型模拟的用户静态兴趣ht与这个兴趣与目标广告相关的注意力分数αt,经过改进后的更新门的作用,模拟用户兴趣沿着与目标广告相关的过程进行更新,最后一个单元输出的就是模拟的最终兴趣兴趣状态。在本发明实施例中,所述步骤S5可以包括以下步骤:Fig. 4 is a diagram of an interest update model for constructing an attention mechanism in the present invention. As shown in Fig. 4, the structure is a kernel structure of a gated recurrent unit AT-GRU based on an attention mechanism. In step t, the user's static interest h t simulated by the time series model and the attention score α t related to this interest and the target advertisement, after the function of the improved update gate, the simulated user interest is updated along the process related to the target advertisement, The output of the last unit is the final interest state of the simulation. In this embodiment of the present invention, the step S5 may include the following steps:
S51、根据所述静态兴趣组状态集合,计算每个兴趣状态与目标广告的注意力分数集合,表示为Atns={αi|i=1,2,...T};S51. According to the static interest group state set, calculate the attention score set of each interest state and the target advertisement, which is expressed as Atns={α i |i=1,2,...T};
在本发明实施例中,αi表示第i个注意力分数,T表示兴趣状态的数量;αi定义为SI中每个兴趣状态h'i与目标广告q的一种经过权重参数分配机制计算后的相似度衡量,可以捕获模型中重要的特征。In the embodiment of the present invention, α i represents the ith attention score, and T represents the number of interest states; α i is defined as a weight parameter assignment mechanism for each interest state h' i and the target advertisement q in the SI to calculate The latter similarity measure can capture important features in the model.
其中,s(h'i,q)表示兴趣状态h'i与目标广告q经过双线性模型的相似度函数计算过后的相似度权重, Among them, s(h' i , q) represents the similarity weight calculated by the similarity function of the bilinear model between the interest state h' i and the target advertisement q,
S52、根据静态兴趣组状态集合和注意力分数集合,采用基于注意力机制的门控循环单元学习兴趣最终更新状态,表示为H=AT-GRU(Intress,Atns)={hi|i=1,2...T},T为AT-GRU中隐藏层的大小;S52. According to the static interest group state set and the attention score set, adopt the gated recurrent unit based on the attention mechanism to learn the final update state of interest, expressed as H=AT-GRU(Intres s , Atns)={hi | i = 1,2...T}, T is the size of the hidden layer in AT-GRU;
考虑到用户兴趣的动态性和泛化性,兴趣更新最终状态定义为静态兴趣组状态Intress在基于注意力机制更新策略下,经过兴趣更新模型提取出的最终兴趣表示向量。Considering the dynamics and generalization of user interests, the final state of interest update is defined as the final interest representation vector extracted by the interest update model under the update strategy based on the attention mechanism.
在本发明的优选实施例中,所述步骤S52具体包括以下步骤:In a preferred embodiment of the present invention, the step S52 specifically includes the following steps:
S521、根据步骤S51的公式,可得到每个兴趣与候选广告的相关度权重,即注意力分数ai,具体如下:S521. According to the formula of step S51, the relevance weight of each interest and the candidate advertisement, that is, the attention score a i , can be obtained, as follows:
其中,ead是来自不同类别广告字段的嵌入向量的连接,是参数矩阵,nH是隐藏兴趣状态向量的维度,nA是广告的嵌入向量的维度。注意力分数ai反映了目标广告与输入兴趣状态的之间的相关度,兴趣状态与目标广告越相关,注意力分数越大。where e ad is the concatenation of embedding vectors from ad fields of different categories, is the parameter matrix , nH is the dimension of the hidden interest state vector, and nA is the dimension of the ad’s embedding vector. The attention score a i reflects the correlation between the target advertisement and the input interest state. The more relevant the interest state and the target advertisement, the higher the attention score.
S522、根据上述步骤计算的注意力分数,引入一种基于注意力机制的门控循环单元AT-GRU,该结构能够根据注意力分数的大小来决定隐藏兴趣状态的更新力度,即与目标广告相关的兴趣状态能够更大力度地参与到最终兴趣状态的更新过程中,与目标广告不相关的兴趣能够较小甚至不参与到更新过程,具体更新策略如下:S522. According to the attention score calculated in the above steps, a gated loop unit AT-GRU based on the attention mechanism is introduced, and the structure can determine the update strength of the hidden interest state according to the size of the attention score, that is, it is related to the target advertisement. The interest status of the advertiser can participate more strongly in the update process of the final interest status, and the interests unrelated to the target advertisement can be smaller or even not involved in the update process. The specific update strategy is as follows:
r’t=σ(Wrii't+Urih't-1+bri),r' t =σ(W ri i' t +U ri h' t-1 +b ri ),
其中h't、h't-1和都是AT-GRU的隐状态,wri,Uri,Uhi∈nHi×nHi,nHi是AT-GRU隐藏层的大小,nIi是AT-GRU输入层的大小。i't表示AT-GRU的输入向量,即经过时间门控循环单元学习的用户静态兴趣,at是注意力分数,与原始的GRU结构相比,AT-GRU结构将使用注意力分数代替了原来的更新门,AT-GRU可以有效地避免因为用户兴趣的泛化性和异构性带来的兴趣漂移问题,实现从用户不断变化发展的兴趣中模拟兴趣发展更新的过程,推动最终兴趣沿着与目标广告相关的方向更新。where h' t , h' t-1 and are the hidden states of AT-GRU, w ri , U ri , U hi ∈ n Hi ×n Hi , where n Hi is the size of the AT-GRU hidden layer, and n Ii is the size of the AT-GRU input layer. i' t represents the input vector of AT-GRU, i.e. the user's static interest learned by the time-gated recurrent unit, and a t is the attention score. Compared with the original GRU structure, the AT-GRU structure will use the attention score instead of The original update gate, AT-GRU can effectively avoid the problem of interest drift caused by the generalization and heterogeneity of users' interests, realize the process of simulating the development and updating of interests from the constantly changing and developing interests of users, and promote the final interest along the Updates with directions related to targeted ads.
S6、输入用户肖像特征和广告特征,采用堆栈式自动编码机,提取出用户肖像特征与广告特征之间的隐形关系向量;S6. Inputting user portrait features and advertisement features, and using a stacked automatic encoding machine to extract an invisible relationship vector between user portrait features and advertisement features;
在本发明实施例中,根据用户肖像特征数据和目标广告特征数据,计算这些特征之间的隐性关系,表示为Rimplicit=SAE(I,P)。In the embodiment of the present invention, the implicit relationship between these features is calculated according to the user portrait feature data and the target advertisement feature data, which is expressed as R implicit =SAE(I,P).
由于目标广告和用户肖像等非时序特征之间也存在着隐性关系,因此本发明实施例定义Rimplicit为经过堆栈式编码机经过特征压缩后提取的这些特征之间的隐性关系,其中I为目标广告特征嵌入向量集合,P为用户肖像嵌入向量集合。Since there is also an implicit relationship between non-sequential features such as target advertisements and user portraits, the embodiment of the present invention defines R implicit as the implicit relationship between these features extracted after feature compression by the stack encoder, where I is the set of embedding vectors for target advertisement features, and P is the set of embedding vectors for user portraits.
在本发明实施例中,本实施例设计一个堆栈式自动编码机结构用于获取进一步的其它非时序特征的关系,其中单层自动编码机的结构如图5所示,该结构由编码器和解码器组成,分为三个部分,分别是输入层、隐藏层和输出层。编码器层在编码器函数的作用将输入层的特征转化为隐藏层的特征,然后在解码器函数的作用下将隐藏层特征转化为输出层。堆栈式自动编码机通过逐层非监督学习的预训练来初始化深度网络的参数,预训练完毕后利用训练参数,可以学习到非时序特征之间高维非线性的元素交互关系。In this embodiment of the present invention, a stacked auto-encoder structure is designed to obtain further relationships of other non-sequential features, wherein the structure of a single-layer auto-encoder is shown in FIG. 5 , which consists of an encoder and a The decoder consists of three parts, namely the input layer, the hidden layer and the output layer. The encoder layer transforms the features of the input layer into the features of the hidden layer under the action of the encoder function, and then transforms the features of the hidden layer into the output layer under the action of the decoder function. The stacked autoencoder initializes the parameters of the deep network through the pre-training of layer-by-layer unsupervised learning. After the pre-training is completed, the training parameters can be used to learn the high-dimensional nonlinear element interaction between non-temporal features.
在本发明的优选实施例中,所述步骤S6可以具体包括如下步骤:In a preferred embodiment of the present invention, the step S6 may specifically include the following steps:
S61、根据输入的用户肖像特征和目标广告特征,进行编码层的计算,编码层负责将输入层的输入数据X转化为隐层的状态H,具体公式如下:S61, according to the input user portrait feature and target advertisement feature, perform the calculation of the coding layer, and the coding layer is responsible for converting the input data X of the input layer into the state H of the hidden layer, and the specific formula is as follows:
Z=sigmoid(W1X+b1),Z=sigmoid(W 1 X+b 1 ),
其中,W1是权重矩阵,b1是第一训练偏置,是激活函数;where W 1 is the weight matrix, b 1 is the first training bias, is the activation function;
S62、类似的,进行解码层的计算,解码层将隐层的状态H转化为输出层Y,定义为:S62. Similarly, perform the calculation of the decoding layer. The decoding layer converts the state H of the hidden layer into the output layer Y, which is defined as:
Y=sigmoid(W2Z+b2),Y=sigmoid(W 2 Z+b 2 ),
其中,W2是权重矩阵,b2是第二训练偏置。where W 2 is the weight matrix and b 2 is the second training bias.
S63、进行重构误差的计算,使得输出Y与原始的X之间的误差足够小,具体公式为,S63, calculate the reconstruction error, so that the error between the output Y and the original X is small enough, and the specific formula is,
其中,W是W1和W2的组合,λ是正则化系数,可以添加惩罚因子λ来控制权重的大小,防止过拟合。Among them, W is the combination of W 1 and W 2 , λ is the regularization coefficient, and a penalty factor λ can be added to control the size of the weight and prevent overfitting.
S64、重复上述步骤S61-S63的训练过程,通过层层叠加训练的方式,得到整个堆栈式自动编码机的训练参数,然后根据这些参数学习到非时序特征之间的隐性关系。S64. Repeat the training process of the above steps S61-S63, obtain the training parameters of the entire stacked autoencoder by means of layer-by-layer training, and then learn the implicit relationship between the non-sequential features according to these parameters.
S7、将用户的兴趣表示向量和用户肖像特征与广告特征之间的隐形关系向量输入到多层感知机中进行联合训练,得到广告点击率的预测结果。S7. Input the user's interest representation vector and the invisible relationship vector between the user's portrait feature and the advertisement feature into the multilayer perceptron for joint training, and obtain the prediction result of the advertisement click rate.
在本发明实施例中,将用户的兴趣更新向量以及所述隐形关系向量进行连接,对连接后的向量进行平滑处理;分别对兴趣模拟与更新模型中的时间门控循环单元的辅助损失函数和多层感知机的预测目标损失函数进行联合训练,训练完成后得到广告点击率的预测结果。In the embodiment of the present invention, the user's interest update vector and the invisible relationship vector are connected, and the connected vector is smoothed; the auxiliary loss function of the time-gated recurrent unit in the interest simulation and update model and The prediction target loss function of the multilayer perceptron is jointly trained, and the prediction result of the advertisement click rate is obtained after the training is completed.
其中,在训练预测模型时,采用联合训练的方式,分别对Time-GRU部分和预测模型的目标损失函数进行联合训练,模型的全局损失函数表示为:Among them, when training the prediction model, the joint training method is used to jointly train the Time-GRU part and the target loss function of the prediction model. The global loss function of the model is expressed as:
L=Ltarget+λ*Laux L=L target +λ*L aux
其中λ是超参数,用户平衡兴趣的模拟和广告点击率的预测,Laux表示Time-GRU的辅助损失。本发明实施例,对目标损失函数Ltarget进行改进,即改进MLP的损失函数时,将MLP损失函数设置为带权重的均方误差,并根据现有数据中正、负样本的比例,设置损失函数的系数,改进后目标损失函数表示为:where λ is the hyperparameter, the simulation of user balance interest and the prediction of ad click-through rate, and L aux represents the auxiliary loss of Time-GRU. In the embodiment of the present invention, the target loss function L target is improved, that is, when the loss function of the MLP is improved, the MLP loss function is set as the weighted mean square error, and the loss function is set according to the ratio of positive and negative samples in the existing data. The coefficient of the improved target loss function is expressed as:
其中,Ltarget为改进后的多层感知机的目标损失函数;N1表示正样本数量;N2表示负样本数量;y是指示变量,如果该类别和样本的类别相同就是1,否则是0;p(Y=0|X)和p(Y=1|X)分别是网络输出属于标签的不同预测概率。同时,本发明还引入了辅助损失Laux,所述辅助损失表示为:Among them, L target is the target loss function of the improved multi-layer perceptron; N1 represents the number of positive samples; N2 represents the number of negative samples; y is the indicator variable, if the category and the category of the sample are the same, it is 1, otherwise it is 0; p (Y=0|X) and p(Y=1|X) are the different predicted probabilities that the network output belongs to the label, respectively. At the same time, the present invention also introduces an auxiliary loss L aux , which is expressed as:
其中,表示用户单击的第t个嵌入向量,G是整个项目集合;表示用户i在第t步点击的项目之外的样本的嵌入;是sigmoid激活函数,表示用户i在Time-GRU中的第t个隐藏兴趣状态。辅助损失使用下一个正负点击样本行为来监督当前兴趣状态的学习。辅助损失的设计引入了用户的全网行为反馈信息,同时不会引入多场景之间的点击偏差以及造成多场景耦合;从优化的角度来讲辅助损失可以在GRU的长序列建模中减少梯度反向传播的难度。Time-GRU在辅助损失的作用下,可以将每一个单元的输出的隐藏兴趣状态与下一个点击行为发生关联,从而能够更好的根据用户行为序列模拟用户的兴趣;同时,Time-GRU在时间门的更新策略下,加强用户行为序列中时间间隔对兴趣模拟的影响,使得时间间隔越短、点击越频繁的行为对用户兴趣的影响越大。in, represents the t-th embedding vector clicked by the user, and G is the entire set of items; represents the embedding of samples other than the item clicked by user i at step t; is the sigmoid activation function, represents the t-th hidden interest state of user i in Time-GRU. The auxiliary loss uses the next positive and negative click sample behavior to supervise the learning of the current state of interest. The design of the auxiliary loss introduces the user's entire network behavior feedback information, and at the same time does not introduce click deviation between multiple scenes and cause multi-scene coupling; from the perspective of optimization, the auxiliary loss can reduce the gradient in the long sequence modeling of GRU The difficulty of backpropagation. Under the action of auxiliary loss, Time-GRU can associate the output hidden interest state of each unit with the next click behavior, so as to better simulate the user's interest according to the user behavior sequence; Under the update strategy of the gate, the influence of the time interval in the user behavior sequence on the interest simulation is strengthened, so that the shorter the time interval and the more frequent clicks, the greater the influence on the user's interest.
图6是本发明实施例的一种广告点击率预测装置结构图,如图6所示,所述装置包括:FIG. 6 is a structural diagram of an apparatus for predicting an advertisement click rate according to an embodiment of the present invention. As shown in FIG. 6 , the apparatus includes:
201、获取模块,用于获取电商平台的用户行为数据,用户肖像数据以及广告数据;201. An acquisition module, used to acquire user behavior data, user portrait data and advertisement data of the e-commerce platform;
202、处理模块,用于对电商平台的用户行为数据进行预处理,并形成用户行为序列;202. A processing module, used for preprocessing the user behavior data of the e-commerce platform, and forming a user behavior sequence;
203、嵌入模块,用于将用户行为序列、用户肖像数据、广告数据分别进行编码表示,得到对应特征的嵌入向量;203. An embedding module, configured to encode and represent the user behavior sequence, user portrait data, and advertisement data respectively, to obtain an embedding vector of the corresponding feature;
204、第一特征提取模块,用于输入用户行为序列特征,采用基于时间因子的Time-GRU的深度神经网络,输出用户的兴趣表示向量;204. A first feature extraction module, used for inputting user behavior sequence features, using a Time-GRU deep neural network based on a time factor, and outputting a user's interest representation vector;
205、第二特征提取模块,用于输入用户的兴趣表示向量,采用基于注意力机制的AT-GRU的深度神经网络,模拟兴趣的更新过程,并输出用户的兴趣更新向量;205. The second feature extraction module is used for inputting the user's interest representation vector, adopts the AT-GRU deep neural network based on the attention mechanism, simulates the update process of the interest, and outputs the user's interest update vector;
206、第三特征提取模块,用于输入用户肖像特征和广告特征,采用堆栈式自动编码机,提取出用户肖像特征与广告特征之间的隐性关系向量;206. A third feature extraction module, used for inputting user portrait features and advertisement features, and extracting the implicit relationship vector between the user portrait features and the advertisement features by using a stacked automatic encoder;
207、广告点击率预测模块,用于将用户的兴趣表示向量和用户肖像特征与广告特征之间的隐性关系向量输入到多层感知机中进行联合训练,得到广告点击率的预测结果。207. The advertisement click-through rate prediction module is configured to input the user's interest representation vector and the implicit relationship vector between the user's portrait feature and the advertisement feature into the multi-layer perceptron for joint training to obtain a prediction result of the advertisement click-through rate.
在本发明的优选实施例中,本发明的一种计算机设备可以包括存储器和处理器,所述存储器中存储有计算机程序,所述计算机程序被所述处理器执行时,使得所述处理器执行如本发明所述广告点击率预测方法的步骤。计算机设备可以为终端或服务器。该计算机设备包括通过系统总线连接的处理器、存储器和网络接口。其中,存储器包括非易失性存储介质和内存储器。该计算机设备的非易失性存储介质可存储操作系统和计算机程序。该计算机程序被执行时,可使得处理器执行一种广告点击率预测方法。该计算机设备的处理器用于提供计算和控制能力,支撑整个计算机设备的运行。该内存储器中可储存有计算机程序,该计算机程序被处理器执行时,可使得处理器执行一种偏好预测方法。计算机设备的网络接口用于进行网络通信。In a preferred embodiment of the present invention, a computer device of the present invention may include a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes The steps of the method for predicting the click-through rate of an advertisement according to the present invention. The computer equipment can be a terminal or a server. The computer device includes a processor, memory, and a network interface connected by a system bus. Wherein, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device can store the operating system and the computer program. The computer program, when executed, can cause the processor to execute a method for predicting a click-through rate of an advertisement. The processor of the computer device is used to provide computing and control capabilities and support the operation of the entire computer device. A computer program may be stored in the internal memory, and when executed by the processor, the computer program may cause the processor to execute a preference prediction method. The network interface of the computer device is used for network communication.
在一个实施例中,本申请提供的广告点击率预测装置可以实现为一种计算机程序的形式,计算机程序可在上述计算机设备上运行,计算机设备的非易失性存储介质可存储组成该广告点击率预测装置的各个程序模块。比如,图6所示的获取模块、处理模块、嵌入模块、第一特征提取模块、第二特征提取模块以及广告点击率预测模块。各个程序模块所组成的计算机程序用于使该计算机设备执行本说明书中描述的本申请各个实施例的广告点击率预测方法中的步骤。In one embodiment, the apparatus for predicting the click rate of advertisements provided by the present application may be implemented in the form of a computer program, the computer program may be executed on the above-mentioned computer equipment, and the non-volatile storage medium of the computer equipment may store the advertisement clicks that constitute the advertisement clicks. Each program module of the rate prediction device. For example, the acquisition module, the processing module, the embedding module, the first feature extraction module, the second feature extraction module and the advertisement click rate prediction module shown in FIG. 6 . The computer program composed of each program module is used to make the computer device execute the steps in the advertisement click-through rate prediction method of each embodiment of the present application described in this specification.
在本发明的描述中,需要理解的是,术语“同轴”、“底部”、“一端”、“顶部”、“中部”、“另一端”、“上”、“一侧”、“顶部”、“内”、“外”、“前部”、“中央”、“两端”等指示的方位或位置关系为基于附图所示的方位或位置关系,仅是为了便于描述本发明和简化描述,而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作,因此不能理解为对本发明的限制。In the description of the present invention, it should be understood that the terms "coaxial", "bottom", "one end", "top", "middle", "the other end", "upper", "one side", "top" "," "inside", "outside", "front", "center", "both ends", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, only for the convenience of describing the present invention and The description is simplified rather than indicating or implying that the device or element referred to must have a particular orientation, be constructed and operate in a particular orientation, and therefore should not be construed as limiting the invention.
在本发明中,除非另有明确的规定和限定,术语“安装”、“设置”、“连接”、“固定”、“旋转”等术语应做广义理解,例如,可以是固定连接,也可以是可拆卸连接,或成一体;可以是机械连接,也可以是电连接;可以是直接相连,也可以通过中间媒介间接相连,可以是两个元件内部的连通或两个元件的相互作用关系,除非另有明确的限定,对于本领域的普通技术人员而言,可以根据具体情况理解上述术语在本发明中的具体含义。In the present invention, unless otherwise expressly specified and limited, terms such as "installation", "arrangement", "connection", "fixation" and "rotation" should be understood in a broad sense, for example, it may be a fixed connection or a It can be a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between the two elements, Unless otherwise clearly defined, those of ordinary skill in the art can understand the specific meanings of the above terms in the present invention according to specific situations.
尽管已经示出和描述了本发明的实施例,对于本领域的普通技术人员而言,可以理解在不脱离本发明的原理和精神的情况下可以对这些实施例进行多种变化、修改、替换和变型,本发明的范围由所附权利要求及其等同物限定。Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, and substitutions can be made in these embodiments without departing from the principle and spirit of the invention and modifications, the scope of the present invention is defined by the appended claims and their equivalents.
Claims (10)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111243596.4A CN113888238B (en) | 2021-10-25 | 2021-10-25 | Advertisement click rate prediction method and device and computer equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111243596.4A CN113888238B (en) | 2021-10-25 | 2021-10-25 | Advertisement click rate prediction method and device and computer equipment |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113888238A true CN113888238A (en) | 2022-01-04 |
CN113888238B CN113888238B (en) | 2024-04-09 |
Family
ID=79014053
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111243596.4A Active CN113888238B (en) | 2021-10-25 | 2021-10-25 | Advertisement click rate prediction method and device and computer equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113888238B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114565436A (en) * | 2022-02-28 | 2022-05-31 | 重庆长安汽车股份有限公司 | Vehicle model recommendation system, method, device and storage medium based on time sequence modeling |
CN114662570A (en) * | 2022-03-04 | 2022-06-24 | 中国建设银行股份有限公司 | Business risk prediction method and device, computer equipment and storage medium |
CN114862511A (en) * | 2022-04-29 | 2022-08-05 | 浙江工业大学 | A deep learning-based short video live broadcast marketing task recommendation method |
CN114862477A (en) * | 2022-05-25 | 2022-08-05 | 中国科学技术大学 | Commodity purchase prediction method and system based on dynamic user click module diagram |
CN115018552A (en) * | 2022-06-28 | 2022-09-06 | 中国科学技术大学 | How to determine product click-through rate |
CN116228368A (en) * | 2023-03-15 | 2023-06-06 | 重庆邮电大学 | Advertisement click rate prediction method based on deep multi-behavior network |
WO2024140829A1 (en) * | 2022-12-27 | 2024-07-04 | 华为技术有限公司 | Data processing method and apparatus, computing device, system and readable storage medium |
WO2025077537A1 (en) * | 2023-10-09 | 2025-04-17 | 马上消费金融股份有限公司 | Click-through rate estimation method and apparatus, electronic device, storage medium, and program product |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109960759A (en) * | 2019-03-22 | 2019-07-02 | 中山大学 | Prediction method of click-through rate of recommendation system based on deep neural network |
CN112307257A (en) * | 2020-11-25 | 2021-02-02 | 中国计量大学 | Short video click rate prediction method based on multi-information node graph network |
CN112381581A (en) * | 2020-11-17 | 2021-02-19 | 东华理工大学 | Advertisement click rate estimation method based on improved Transformer |
CN112395505A (en) * | 2020-12-01 | 2021-02-23 | 中国计量大学 | Short video click rate prediction method based on cooperative attention mechanism |
CN112700274A (en) * | 2020-12-29 | 2021-04-23 | 华南理工大学 | Advertisement click rate estimation method based on user preference |
-
2021
- 2021-10-25 CN CN202111243596.4A patent/CN113888238B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109960759A (en) * | 2019-03-22 | 2019-07-02 | 中山大学 | Prediction method of click-through rate of recommendation system based on deep neural network |
CN112381581A (en) * | 2020-11-17 | 2021-02-19 | 东华理工大学 | Advertisement click rate estimation method based on improved Transformer |
CN112307257A (en) * | 2020-11-25 | 2021-02-02 | 中国计量大学 | Short video click rate prediction method based on multi-information node graph network |
CN112395505A (en) * | 2020-12-01 | 2021-02-23 | 中国计量大学 | Short video click rate prediction method based on cooperative attention mechanism |
CN112700274A (en) * | 2020-12-29 | 2021-04-23 | 华南理工大学 | Advertisement click rate estimation method based on user preference |
Non-Patent Citations (1)
Title |
---|
朱江湖: "新型电商平台广告点击率算法研究与实现", 《》, 27 May 2022 (2022-05-27) * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114565436A (en) * | 2022-02-28 | 2022-05-31 | 重庆长安汽车股份有限公司 | Vehicle model recommendation system, method, device and storage medium based on time sequence modeling |
CN114565436B (en) * | 2022-02-28 | 2024-10-29 | 重庆长安汽车股份有限公司 | Vehicle model recommendation system, method, equipment and storage medium based on time sequence modeling |
CN114662570A (en) * | 2022-03-04 | 2022-06-24 | 中国建设银行股份有限公司 | Business risk prediction method and device, computer equipment and storage medium |
CN114862511A (en) * | 2022-04-29 | 2022-08-05 | 浙江工业大学 | A deep learning-based short video live broadcast marketing task recommendation method |
CN114862477A (en) * | 2022-05-25 | 2022-08-05 | 中国科学技术大学 | Commodity purchase prediction method and system based on dynamic user click module diagram |
CN115018552A (en) * | 2022-06-28 | 2022-09-06 | 中国科学技术大学 | How to determine product click-through rate |
WO2024140829A1 (en) * | 2022-12-27 | 2024-07-04 | 华为技术有限公司 | Data processing method and apparatus, computing device, system and readable storage medium |
CN116228368A (en) * | 2023-03-15 | 2023-06-06 | 重庆邮电大学 | Advertisement click rate prediction method based on deep multi-behavior network |
WO2025077537A1 (en) * | 2023-10-09 | 2025-04-17 | 马上消费金融股份有限公司 | Click-through rate estimation method and apparatus, electronic device, storage medium, and program product |
Also Published As
Publication number | Publication date |
---|---|
CN113888238B (en) | 2024-04-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113888238A (en) | Advertisement click rate prediction method and device and computer equipment | |
CN113626719B (en) | Information recommendation method, device, equipment, storage medium and computer program product | |
CN112288471B (en) | Advertisement click rate prediction method based on user history behavior sequence | |
CN115082147B (en) | A sequence recommendation method and device based on hypergraph neural network | |
CN111080400B (en) | Commodity recommendation method and system based on gate control graph convolution network and storage medium | |
CN111753209B (en) | A Sequence Recommendation List Generation Method Based on Improved Time Series Convolutional Network | |
CN111241394B (en) | Data processing methods, devices, computer-readable storage media and electronic equipment | |
CN111667024B (en) | Content pushing method, device, computer equipment and storage medium | |
CN116089708A (en) | Agricultural knowledge recommendation method and device | |
CN115564517A (en) | Commodity recommendation method, prediction model training method and related equipment | |
CN119323442B (en) | Advertisement recommendation method and system based on click rate prediction model | |
CN119377498A (en) | Private domain traffic scheduling and content distribution method and system based on group behavior analysis | |
CN118377969A (en) | Diversity recommendation method based on multi-task graph comparison learning | |
CN119693066A (en) | A method and system for accurate advertising delivery based on artificial intelligence | |
CN115293812A (en) | A session-aware recommendation prediction method for e-commerce platforms based on long-term and short-term interests | |
CN119474552A (en) | A cultural and tourism content recommendation system that analyzes preferences for cultural and tourism attractions | |
CN115618079A (en) | Session recommendation method, device, electronic equipment and storage medium | |
CN113449182B (en) | Knowledge information personalized recommendation method and system | |
CN114493724A (en) | Multi-task learning-based bidding keyword effect estimation model construction method | |
CN117993995A (en) | E-commerce recommendation method and device based on real-time number bin and recommendation model | |
CN114329193B (en) | Click rate prediction method based on time perception interest evolution | |
CN116975426A (en) | Service data processing method, device, equipment and medium | |
CN114862477A (en) | Commodity purchase prediction method and system based on dynamic user click module diagram | |
CN114519600A (en) | Graph neural network CTR estimation algorithm fusing adjacent node variances | |
CN115701102B (en) | Content push method, device, equipment and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |