WO2016091114A1 - 一种信息推送方法和装置 - Google Patents

一种信息推送方法和装置 Download PDF

Info

Publication number
WO2016091114A1
WO2016091114A1 PCT/CN2015/096245 CN2015096245W WO2016091114A1 WO 2016091114 A1 WO2016091114 A1 WO 2016091114A1 CN 2015096245 W CN2015096245 W CN 2015096245W WO 2016091114 A1 WO2016091114 A1 WO 2016091114A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
pushed
model
life cycle
conversion rate
Prior art date
Application number
PCT/CN2015/096245
Other languages
English (en)
French (fr)
Inventor
杨志雄
吕韬
Original Assignee
阿里巴巴集团控股有限公司
杨志雄
吕韬
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 阿里巴巴集团控股有限公司, 杨志雄, 吕韬 filed Critical 阿里巴巴集团控股有限公司
Publication of WO2016091114A1 publication Critical patent/WO2016091114A1/zh

Links

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor

Definitions

  • the present application relates to the field of computer applications, and in particular, to an information push method and apparatus.
  • the search server On the e-commerce website, users can get their own target of interest or preference by entering search keywords in the search box in the station. Specifically, the search server first searches for an object matching the search keyword input by the user based on a matching algorithm (such as a text matching algorithm), then sorts the searched object, and finally displays the searched object on the website page according to the sorting. On, for the user to choose their favorite target object.
  • a matching algorithm such as a text matching algorithm
  • the e-commerce website can also push other objects similar to the searched object to the user.
  • the push server first obtains an object similar to the searched object according to the searched object and according to a correlation algorithm (for example, calculating a behavior correlation and text correlation between the object and the object through a collaborative filtering algorithm), And as a candidate object, then the searched object is removed from the candidate object, and the remaining candidate object is used as the object to be pushed, and then the object to be pushed is sorted, and finally the push object is displayed on the website according to the order of sorting. On the page.
  • a correlation algorithm for example, calculating a behavior correlation and text correlation between the object and the object through a collaborative filtering algorithm
  • the push server can generally push only a limited number of push objects to the user. Therefore, the push server selects the top N push objects (N is a positive integer) in the order of sorting, and then displays the top N push objects on the website's page.
  • the push server sorts the push objects in descending order of relevance.
  • this sorting method only considers the correlation between the push object and the searched object, and the push object itself is probably not the target object that the user pays attention to or likes, and the user therefore needs to reconstruct the search keyword to search again, and The process of repeated searches not only reduces the user experience, but also excessively consumes the resources of the search server and the push server.
  • the embodiment of the present application provides an information pushing method and apparatus, so as to reduce the possibility of repeated user search as much as possible, improve the user experience, and save resources of the search server and the push server.
  • An information push method includes:
  • the attribute feature Retrieving the attribute feature from the session to which the current behavior belongs in response to the current behavior of the user, and inputting the attribute feature into a preset user behavior life cycle model, outputting the user's current stage in the user behavior life cycle , wherein the user behavior life cycle includes a target object ambiguous phase, a target object explicit selection phase, and a target object locking phase;
  • the user behavior life cycle model is a continuous model obtained by training the GBRT model.
  • the hit probability model is a discrete model obtained by training a logistic regression LR model.
  • the preset hit probability model corresponding to the user currently in the user behavior life cycle includes: an exposure click conversion rate model, a click favorite conversion rate model, and a click order conversion rate model;
  • the attribute feature extracted from the object to be pushed and the user is input to a hit probability model preset and corresponding to a stage in which the user is currently in the user behavior life cycle, and the hit probability of the object to be pushed is output.
  • the attribute features extracted from the object to be pushed and the user are input into the exposure click conversion rate model, and the object to be pushed is outputted.
  • the conversion rate between the number of impressions and the number of clicks;
  • the attribute feature extracted from the object to be pushed and the user is input into the click collection conversion rate model, and the object to be pushed is outputted.
  • the conversion rate between the number of clicks and the number of people being watched;
  • the attribute features extracted from the object to be pushed and the user are input into the click order conversion rate model, and the object to be pushed is outputted.
  • the selecting at least one object to be pushed according to the order in which the hit probability of the object to be pushed is from the largest to the smallest comprises:
  • At least one candidate is selected according to the conversion rate between the number of exposed objects to be pushed and the number of clicked objects.
  • At least one candidate is selected according to a conversion ratio between the number of clicked objects to be pushed and the number of attentiond objects Object
  • At least one candidate is selected according to the conversion rate between the number of clicked objects to be pushed and the number of orders to be pushed. Object.
  • the object to be pushed is obtained locally, and the method further includes:
  • the searched object is extracted from the candidate object, and the remaining candidate object is used as the object to be pushed.
  • An information pushing device includes:
  • a life cycle determining unit configured to extract an attribute feature from a session to which the current behavior belongs in response to a current behavior of the user, and input the attribute feature into a preset user behavior life cycle
  • the user is currently in a phase in a user behavior life cycle, wherein the user behavior life cycle includes a target object ambiguous phase, a target object explicit selection phase, and a target object locking phase;
  • a hit probability determining unit configured to input the attribute feature extracted from the object to be pushed and the user into a preset hit probability model corresponding to a stage in the user behavior life cycle of the user, and output the to-be-pushed The hit probability of the object, in which the hit probability models corresponding to different stages are different;
  • the object to be pushed is selected to select at least one object to be pushed according to the hit probability of the object to be pushed from the largest to the smallest;
  • the pushing unit is configured to push the selected object to be pushed.
  • the user behavior life cycle model is a continuous model obtained by training the GBRT model.
  • the hit probability model is a discrete model obtained by training a logistic regression LR model.
  • the preset hit probability model corresponding to the user currently in the user behavior life cycle includes: an exposure click conversion rate model, a click favorite conversion rate model, and a click order conversion rate model;
  • the hit probability determining unit includes a first determining subunit, a second determining subunit, and a third determining subunit;
  • a first determining subunit configured to: when the user is currently in the user object life cycle, the attribute feature extracted from the object to be pushed and the user is input into the exposure click conversion rate model, Outputting a conversion rate between the number of exposures of the object to be pushed and the number of clicks;
  • a second determining subunit configured to: when the user currently selects a stage for the target object in a stage in the user behavior life cycle, input the attribute feature extracted from the object to be pushed and the user into the click conversion conversion rate model, Outputting a conversion rate between the number of clicks of the object to be pushed and the number of attentions;
  • a third determining subunit configured to input the attribute feature extracted from the object to be pushed and the user to the click when the user is currently in the target object locking phase in the user behavior life cycle
  • the conversion rate between the number of clicked objects to be pushed and the number of orders to be pushed is output.
  • the to-be-pushed object selection unit includes a first selection subunit, a second selection subunit, and a third selection subunit; wherein
  • a first selecting subunit configured to: when the user is currently in the user activity life cycle, the target object ambiguous phase, according to the conversion rate between the number of exposed objects to be pushed and the number of clicked objects Select at least one candidate object in a small order;
  • a second selection sub-unit configured to: when the user currently explicitly selects a phase for the target object in a phase in the user behavior life cycle, according to a conversion rate between the number of clicked objects to be pushed and the number of objects to be pushed Select at least one candidate object in a small order;
  • a third selecting sub-unit configured to: when the user is currently in the target object locking phase of the user behavior life cycle, according to the conversion rate between the number of clicked objects to be pushed and the number of orders to be pushed from the large Select at least one candidate object in a small order.
  • the object to be pushed is obtained locally, and the device further includes:
  • a similarity calculation unit configured to calculate a similarity value between each object and the searched object according to the correlation algorithm
  • a candidate object selecting unit configured to select at least one object in order of similarity values as a candidate object
  • the culling unit is configured to extract the searched object from the candidate object, and use the remaining candidate object as the object to be pushed.
  • the user can be determined at the stage of the user's behavior life cycle.
  • different hit probability models are used to determine the hit probability of each object to be pushed.
  • the present application provides the push mode that meets the current needs to the users at different stages, so that the push is performed.
  • the object pushed by the method is more likely to be the target object of the user's preference, thereby reducing the possibility of the user repeatedly searching and improving the user experience. Save resources on search servers and push servers.
  • FIG. 1 schematically illustrates an exemplary application scenario in which embodiments of the present application may be implemented
  • FIG. 2 is a flow chart schematically showing an information push method of the present application
  • FIG. 3 is a flow chart schematically showing a method for obtaining an object to be pushed in the present application
  • FIG. 4 is a schematic diagram showing the operation of a push flow in the present application.
  • FIG. 5 is a block diagram showing the structure of an information pushing apparatus in the present application.
  • FIG. 6 is a schematic block diagram showing a structure of a hit probability determining unit in the present application.
  • FIG. 7 is a schematic structural block diagram of a to-be-pushed object selection unit in the present application.
  • Fig. 8 is a block diagram showing the structure of another information pushing apparatus in the present application.
  • Fig. 1 schematically illustrates an exemplary application scenario in which embodiments of the present application may be implemented.
  • the client 10 submits the search keyword 11 input by the user to the search server 20, and the search server 20 searches for an object matching the search keyword 11 based on the search algorithm, and sorts the searched object 21 and feeds it back to the client 10.
  • the push server 30 acquires the searched object 21 from the search server 20, obtains an object similar to the searched object 21 according to the correlation algorithm, and as a candidate object, then extracts the searched object 21 from the candidate object, and the remaining The candidate candidate is the object to be pushed.
  • the behavior may be a click behavior for a certain searched object
  • the push server 30 sorts each object to be pushed based on the behavior, and sorts according to the At least one object to be pushed is selected in the order from first to last, and finally the selected object to be pushed 31 is pushed to the client 10.
  • Search server 20 and push server 30 can be web servers, too Can be an APP server.
  • an "object" is an item.
  • search server 20 and push server 30 can be the same server.
  • the inventor of the present application found in the research that when a user conducts a session with an e-commerce website, the user's demand for the push mode is different when the user is at different stages of the session process. Therefore, it is necessary to provide users in different stages with a push mode that meets their current needs, so that the object pushed by the push mode is more likely to be the target object of the user's preference.
  • a session between a user and an e-commerce website is taken as a user behavior life cycle, and the user behavior life cycle is divided into three different stages, and for any one stage, there is a corresponding hit probability model.
  • the hit probability of each object to be pushed can be obtained, so that the order of the push object to be compared with the hit probability is a sorting method more in line with the user requirements at the stage, and the object pushed according to the sort is more likely to be The target object that the user likes.
  • FIG. 2 is a flow chart schematically showing a method for pushing information according to the present application.
  • the method may be implemented. The method includes the following steps:
  • Step 201 In response to the current behavior of the user, extract an attribute feature from the session to which the current behavior belongs, and input the attribute feature into a preset user behavior life cycle model, and output the user current life cycle of the user behavior.
  • the stage in which the user behavior life cycle includes a target object ambiguous phase, a target object explicit selection phase, and a target object locking phase.
  • Step 202 Input an attribute feature extracted from the object to be pushed and the user into a preset hit probability model corresponding to a stage in the user behavior life cycle of the user, and output the The hit probability of the object to be pushed, wherein the hit probability models corresponding to different stages are different.
  • Step 203 Select at least one object to be pushed according to the hit probability of the object to be pushed from the largest to the smallest.
  • Step 204 Push the selected object to be pushed.
  • the life cycle of user behavior is divided into the following three stages: target object ambiguous stage, target object explicit selection stage and target object locking stage, and different stages will correspond to different hit probability. model.
  • the inventor of the present invention found in the research that at the beginning, the user often does not know what the target object he really likes. At this time, the user is in the ambiguous stage of the target object, and at this stage, the user mainly browses through the browsing. The process gradually clarifies the target objects that you really like. For example, the user wants to buy a dress, but the user does not explicitly want what style of dress.
  • the user enters the target object explicit selection stage, in which the user needs to compare and analyze a plurality of candidate target objects, and select a target object that suits his or her preference, and this stage generally lasts for a long time. For example, if a user explicitly wants a bohemian dress, a comparative analysis of bohemian dresses in various patterns, colors, materials, and prices.
  • the user After a large number of comparative analysis, the user gradually locks to a target object, and the user enters the target object locking phase. For example, after locking in a bohemian dress, the user will pay more attention to the seller's word of mouth and the user's evaluation of the bohemian dress until the last order.
  • the following is a session between the user and the e-commerce site (generally, if a user does not re-site with the site within the predetermined time after the first interaction with the site) Any interaction is considered to be the end of a session, otherwise, a session is considered to continue.
  • the scheduled time can be 30 minutes.
  • search keywords may be entered multiple times, and after each search keyword is entered, it is possible Multiple behaviors are triggered (ie, triggering a sequence of behaviors), such as click behavior, adding a shopping cart behavior, adding a favorite behavior, or placing an order. Therefore, the behavior of all input search keywords in a session and the sequence of actions triggered after each input of the search keyword can be extracted and divided as follows:
  • the term is the smallest unit word obtained after the word segmentation of the search keyword, and may be a noun or an adjective.
  • the search keyword "Korean Slim Dress” has three terms: Korean, Slim and Dress.
  • the predetermined time can be any time, for example, 15 days.
  • a discrete user behavior lifecycle model can be obtained by training LR (Logistic Regression) or SVM (Support Vector Machine) models.
  • the target is a continuous value, and determines which phase of the user's behavior life cycle the user is currently in according to the interval in which the indication value of the behavior falls.
  • the user behavior life cycle model is a continuous model obtained by training the GBRT (Gradient Boost Regression Tree) model.
  • a marker value interval [0, 2] can be divided into three regions.
  • the indicated value of the behavior falls within the interval [0, 0.9)
  • the user is in the ambiguous phase of the target object, and the indication value of the behavior falls.
  • the [0.9, 1.5] interval it indicates that the user is in the target object clear selection phase.
  • the indicated value of the behavior falls within the (1.5, 2) interval, the user is in the target object locking phase.
  • the target object explicit selection phase is an important process of connecting the target object ambiguous phase and the target object locking phase.
  • the user's intent ie, which target object is specifically locked
  • the target object clear selection phase the user finally enters the target locking phase, completes the order, or does not enter the target locking phase, leaving the website, is the cumulative result of the behavior sequence triggered by the user in the target object clear selection phase.
  • the rate of change of the user's intention caused by different behavior types is different in the target object explicit selection phase.
  • each behavior may be given different weight values; then the number of times of each behavior in the behavior sequence of the stage is counted, and weighted and summed according to its weight; The span of the marked value interval of the stage and the weighted summation value of all the behaviors determine the amount of behavior change caused by each behavior to the user's intention; finally, according to the behavior change amount of each behavior, the marked value of each behavior is calculated.
  • add is the behavior of adding a shopping cart (or favorites)
  • click is a click behavior
  • the user behavior life cycle model After training to obtain the user behavior life cycle model, it is possible to estimate which phase of the user behavior life cycle the user is currently based on the user's real-time input search keywords and real-time behavior sequences. For example, after the user inputs a search keyword "dress”, if one of the behaviors in the sequence of behaviors is clicking on the object 1, and the calculated value of the click behavior is calculated according to the user behavior life cycle model, the estimated value is 0.7. The user is now at the target stage of the user's life cycle. If the user's behavior after clicking the behavior of the object 1 is to add the object 2 to the shopping cart, and calculate the flag value of the added shopping cart behavior according to the user behavior life cycle model, the user is estimated to be at this time. The target object in the user's life cycle is clearly selected.
  • the hit probability model outputs the hit probability of each object to be pushed, that is, each object to be pushed is just the user's preference. The probability of the target object.
  • the hit probability model corresponding to the ambiguous stage of the target object is an exposure click conversion rate model
  • the hit probability model corresponding to the target object explicit selection stage is a click favorite conversion rate model
  • the hit probability model corresponding to the target object locking stage is a click. Order conversion rate model.
  • the attribute characteristics extracted when training different hit probability models are also different.
  • the conversion probability between the number of exposed objects to be pushed and the number of clicked objects is the target training hit probability model
  • the main attribute features extracted during the training include the object The number of orders placed, the image quality of the object, and whether the object is in the user's preferred category.
  • the conversion rate between the number of clicked objects to be pushed and the number of attentions is the target training hit probability model.
  • the main attribute features extracted during training include the user Preferred style, preferred price, and material of the object.
  • the target training hit probability model is based on the conversion rate between the number of clicked objects to be pushed and the number of orders to be pushed, and the main attribute features extracted during training include the favorable rate of the object, Features such as merchant level and credit rating.
  • the conversion rate between the number of exposures and the number of clicks the number of clicks / the number of exposures
  • the conversion rate between the number of clicks and the number of people being touched the number of people being watched / the number of clicks
  • the number of clicks and the number of clicks Conversion rate between single quantities number of orders placed / number of clicks.
  • attribute features of the training hit probability models listed above are only schematic, and in addition to extracting these attribute features, other attribute features may also be extracted.
  • the conversion rate between the number of exposed objects and the number of clicks to be pushed is in descending order.
  • Selecting at least one candidate object when the user currently explicitly selects a stage for the target object in a stage in the user behavior life cycle, the conversion rate between the number of clicked objects and the number of attentiond objects to be pushed is from large to The small order selects at least one candidate object; when the user is currently in the user object life cycle, the target object locking phase is according to the conversion rate between the number of clicked objects to be pushed and the number of orders to be pushed.
  • the hit probability model is a discrete model obtained by training the LR model.
  • the object to be pushed can be obtained by the search server 20, and the push server 30 acquires the object to be pushed from the search server 20.
  • FIG. 3 is a flow chart schematically showing a method for obtaining an object to be pushed in the present application.
  • the method is obtained by the push server 30, and the method can include the following steps:
  • Step 301 Calculate a similarity value between each object and the searched object according to a correlation algorithm.
  • Step 302 Select at least one object in descending order of similarity values as a candidate object
  • Step 302 Extract the searched object from the candidate object, and use the remaining candidate object as the object to be pushed.
  • FIG. 4 is a schematic diagram showing the operation of a push flow in the present application.
  • each object to be pushed may be displayed in the recommended area at the bottom of the search result display page, so as to push the object to be pushed to the search user.
  • the user For a user who is active on the e-commerce network, whenever a behavior is triggered, the user can be determined at the stage of the user's behavior life cycle. When at different stages, different hit probability models are used to determine the hit probability of each object to be pushed. In order to finally sort the objects to be pushed in descending order of hit probability, and select the first N bits to push. Since the user has different requirements for the push mode when the user is in different stages in the user behavior life cycle, the present application provides the push mode that meets the current needs to the users at different stages, so that the push is performed.
  • the object pushed by the method is more likely to be the target object that the user likes, thereby reducing the possibility of the user repeatedly searching, improving the user experience, and saving resources of the search server and the push server.
  • FIG. 5 is a schematic structural block diagram of an embodiment of an information pushing apparatus in the present application.
  • the apparatus includes: a life cycle determining unit 501, and a hit probability determining unit. 502.
  • the internal structure and connection relationship will be further described below in conjunction with the working principle of the device.
  • the life cycle determining unit 501 is configured to extract an attribute feature from a session to which the current behavior belongs in response to a current behavior of the user, and input the attribute feature into a preset user behavior life cycle model, and output the user Currently in the user behavior life cycle, wherein the user behavior life cycle includes a target object ambiguity phase, a target object explicit selection phase, and a target object locking phase;
  • the hit probability determining unit 502 is configured to input the attribute feature extracted from the object to be pushed and the user into a preset hit probability model corresponding to a stage in the user behavior life cycle of the user, and output the waiting The hit probability of the push object, wherein the hit probability models corresponding to different stages are different;
  • the to-be-pushed object selection unit 503 is configured to select at least one object to be pushed according to the hit probability of the object to be pushed from the largest to the smallest;
  • the pushing unit 504 is configured to push the selected object to be pushed.
  • the user behavior lifecycle model is a continuous model obtained by training a GBRT model.
  • the hit probability model is a discrete model obtained by training a logistic regression LR model.
  • the hit probability model preset and corresponding to the user currently in the user behavior life cycle includes: an exposure click conversion rate model, a click favorite conversion rate model, and a click. Order conversion rate model;
  • the hit probability determining unit 502 includes a first determining subunit 5021, a second determining subunit 5022, and a third determining subunit 5023;
  • the first determining sub-unit 5021 is configured to input the attribute features extracted from the object to be pushed and the user into the exposure click conversion rate model when the user is currently in the target object ambiguous stage in the user behavior life cycle , outputting a conversion rate between the number of exposures of the object to be pushed and the number of clicks;
  • a second determining sub-unit 5022 configured to: when the user currently explicitly selects a phase for the target object in a phase in the user behavior life cycle, the attribute feature extracted from the object to be pushed and the user is input Entering into the click-to-store conversion rate model, outputting a conversion rate between the number of clicks of the object to be pushed and the number of objects to be pushed;
  • the third determining subunit 5023 is configured to: when the user is currently in the target object locking phase of the user behavior life cycle, input the attribute feature extracted from the object to be pushed and the user into the click order conversion rate model. And outputting a conversion rate between the number of clicked objects to be pushed and the number of orders placed.
  • the to-be-pushed object selection unit 503 includes a first selection sub-unit 5031, a second selection sub-unit 5032, and a third selection sub-unit 5033;
  • a first selecting subunit configured to: when the user is currently in the user activity life cycle, the target object ambiguous phase, according to the conversion rate between the number of exposed objects to be pushed and the number of clicked objects Select at least one candidate object in a small order;
  • a second selection sub-unit configured to: when the user currently explicitly selects a phase for the target object in a phase in the user behavior life cycle, according to a conversion rate between the number of clicked objects to be pushed and the number of objects to be pushed Select at least one candidate object in a small order;
  • a third selecting sub-unit configured to: when the user is currently in the target object locking phase of the user behavior life cycle, according to the conversion rate between the number of clicked objects to be pushed and the number of orders to be pushed from the large Select at least one candidate object in a small order.
  • the object to be pushed is obtained locally, as shown in FIG. 8 (FIG. 8 only shows the added portion and the added portion and the device shown in FIG. Connection relationship), the device also includes:
  • the similarity calculation unit 801 is configured to calculate a similarity value between each object and the searched object according to the correlation algorithm
  • the candidate object selecting unit 802 is configured to select at least one object as a candidate object according to the similarity value from the largest to the smallest;
  • the culling unit 803 is configured to remove the searched object from the candidate object, and use the remaining candidate object as the object to be pushed.
  • the present application provides the push mode that meets the current needs to the users at different stages, so that the push is performed.
  • the object pushed by the method is more likely to be the target object that the user likes, thereby reducing the possibility of the user repeatedly searching, improving the user experience, and saving resources of the search server and the push server.
  • the disclosed systems, devices, and methods may be implemented in other manners.
  • the device embodiments described above are merely illustrative.
  • the division of the unit is only a logical function division.
  • there may be another division manner for example, multiple units or components may be combined. Or it can be integrated into another system, or some features can be ignored or not executed.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, device or unit, and may be in an electrical, mechanical or other form.
  • the units described as separate components may be or may be physically separated, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of the embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
  • the above integrated unit can be implemented in the form of hardware and can be implemented in the form of a software functional unit.
  • the program can be stored in a computer readable storage medium, which, when executed, can include the flow of an embodiment of the methods described above.
  • the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • General Engineering & Computer Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Databases & Information Systems (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种信息推送方法，包括：响应于用户的当前行为，从当前行为所属的会话中提取属性特征，输入到预置的用户行为生命周期模型，输出用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段（201）；将从待推送对象和用户中提取的属性特征输入到预置的且与用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出待推送对象的命中概率，其中不同阶段对应的命中概率模型不同（202）；按照待推送对象的命中概率从大到小的顺序选取至少一个待推送对象（203）；对选取的待推送对象进行推送（204）。所述方法可以尽可能地降低用户反复搜索的可能性，提升用户体验的同时，也节约搜索服务器以及推送服务器的资源。还公开了一种信息推送装置。

Description

一种信息推送方法和装置 技术领域
本申请涉及计算机应用领域，特别是涉及一种信息推送方法和装置。
背景技术
在电子商务网站上，用户可以通过在站内的搜索框中输入搜索关键词来获得自己关注或喜好的目标对象。具体地，搜索服务器先基于匹配算法(如文本匹配算法)搜索与用户输入的搜索关键词所匹配的对象，然后将搜索到的对象进行排序，最后按照排序将搜索到的对象展示在网站的页面上，以供用户从中选择自己喜好的目标对象。
除了向用户提供搜索到的对象之外，电子商务网站还可以向用户推送与搜索到的对象相似的其它对象。具体地，推送服务器先以搜索到的对象为基准，根据相关性算法(例如，通过协同过滤算法计算对象与对象之间的行为相关性和文本相关性)得到与搜索到的对象相似的对象，并作为候选对象，然后从候选对象中剔除出搜索到的对象，并将剩余下来的候选对象作为待推送对象，再将待推送对象进行排序，最后按照排序的先后顺序将推送对象展示在网站的页面上。
在实现本申请的过程中，本申请的发明人发现现有技术中至少存在如下问题：与搜索服务器向用户提供搜索到的对象不同，推送服务器一般只能向用户推送有限个数的推送对象。因此，推送服务器会按照排序的顺序选取排在前N位的推送对象(N为正整数)，然后将排在前N位的推送对象展示在网站的页面上。
在现有技术中，推送服务器是按照相关性从高到低的顺序对推送对象进行排序的。但是，这种排序方式只考虑到了推送对象与搜索到的对象之间的相关性，而推送对象本身很可能不是用户关注或喜好的目标对象，用户因此需要重新构建搜索关键词再次进行搜索，而反复搜索的过程不仅会降低用户体验，也会过度地消耗搜索服务器以及推送服务器的资源。
发明内容
为了解决上述技术问题，本申请实施例提供了一种信息推送方法和装置，以尽可能地降低用户反复搜索的可能性，提升用户体验的同时，也节约搜索服务器以及推送服务器的资源。
本申请实施例公开了如下技术方案：
一种信息推送方法，包括：
响应于用户的当前行为，从所述当前行为所属的会话中提取属性特征，并将所述属性特征输入到预置的用户行为生命周期模型，输出所述用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段；
将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，其中不同阶段对应的命中概率模型不同；
按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象；
对选取的待推送对象进行推送。
优选的，所述用户行为生命周期模型是对GBRT模型进行训练所得到的连续模型。
优选的，所述命中概率模型是对逻辑回归LR模型进行训练所得到的离散模型。
优选的，所述预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型包括：曝光点击转化率模型、点击收藏转化率模型和点击下单转化率模型；
则所述将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，包括：
当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，将从待推送对象和用户中提取的属性特征输入到曝光点击转化率模型中，输出所述待推送对象的被曝光数量与被点击数量之间的转化率；
或者，
当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，将从待推送对象和用户中提取的属性特征输入到点击收藏转化率模型中，输出所述待推送对象的被点击数量与被关注数量之间的转化率；
或者，
当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，将从待推送对象和用户中提取的属性特征输入到点击下单转化率模型中，输出所述待推送对象的被点击数量与被下单数量之间的转化率。
优选的，所述按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象包括：
当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，按照所述待推送对象的被曝光数量与被点击数量之间的转换率从大到小的顺序选取至少一个候选对象；
或者，
当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，按照所述待推送对象的被点击数量与被关注数量之间的转换率从大到小的顺序选取至少一个候选对象；
或者，
当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，按照所述待推送对象的被点击数量与被下单数量之间的转换率从大到小的顺序选取至少一个候选对象。
优选的，所述待推送对象是在本地获得的，所述方法还包括：
根据相关性算法计算各个对象与搜索到的对象之间的相似度值；
按照相似度值从大到小的顺序选取至少一个对象并作为候选对象；
从所述候选对象中剔除出搜索到的对象，将剩余下的候选对象作为待推送对象。
一种信息推送装置，包括：
生命周期确定单元，用于响应于用户的当前行为，从所述当前行为所属的会话中提取属性特征，并将所述属性特征输入到预置的用户行为生命周期 模型中，输出所述用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段；
命中概率确定单元，用于将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，其中不同阶段对应的命中概率模型不同；
待推送对象选取单元，用于按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象；
推送单元，用于对选取的待推送对象进行推送。
优选的，所述用户行为生命周期模型是对GBRT模型进行训练所得到的连续模型。
优选的，所述命中概率模型是对逻辑回归LR模型进行训练所得到的离散模型。
优选的，所述预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型包括：曝光点击转化率模型、点击收藏转化率模型和点击下单转化率模型；
所述命中概率确定单元包括第一确定子单元、第二确定子单元和第三确定子单元；其中，
第一确定子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，将从待推送对象和用户中提取的属性特征输入到曝光点击转化率模型中，输出所述待推送对象的被曝光数量与被点击数量之间的转化率；
第二确定子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，将从待推送对象和用户中提取的属性特征输入到点击收藏转化率模型中，输出所述待推送对象的被点击数量与被关注数量之间的转化率；
第三确定子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，将从待推送对象和用户中提取的属性特征输入到点击 下单转化率模型中，输出所述待推送对象的被点击数量与被下单数量之间的转化率。
优选的，所述待推送对象选取单元包括第一选取子单元、第二选取子单元和第三选取子单元；其中，
第一选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，按照所述待推送对象的被曝光数量与被点击数量之间的转换率从大到小的顺序选取至少一个候选对象；
第二选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，按照所述待推送对象的被点击数量与被关注数量之间的转换率从大到小的顺序选取至少一个候选对象；
第三选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，按照所述待推送对象的被点击数量与被下单数量之间的转换率从大到小的顺序选取至少一个候选对象。
优选的，所述待推送对象是在本地获得的，所述装置还包括：
相似度计算单元，用于计算根据相关性算法计算各个对象与搜索到的对象之间的相似度值；
候选对象选取单元，用于按照相似度值从大到小的顺序选取至少一个对象并作为候选对象；
剔除单元，用于从所述候选对象中剔除出搜索到的对象，将剩余下的候选对象作为待推送对象。
由上述实施例可以看出，与现有技术相比，本申请的优点在于：
对于在电子商务网上活跃的用户来说，每当触发一个行为时，就可以确定该用户在用户行为生命周期中所处的阶段。当处于不同的阶段时，会利用不同的命中概率模型来确定各个待推送对象的命中概率。以便最后按照命中概率从大到小的顺序对各个待推送对象进行排序，并选取前N位进行推送。由于在用户行为生命周期中，当用户处于不同的阶段时，该用户对于推送方式的需求是不同的，因此，本申请向处于不同阶段的用户提供符合其当前需求的推送方式，使得以该推送方式所推送的对象更有可能是用户喜好的目标对象，从而尽可能地降低用户反复搜索的可能性，提升用户体验的同时，也 节约搜索服务器以及推送服务器的资源。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1示意性地示出了本申请的实施方式可以在其中实施的示例性应用场景；
图2示意性地示出了本申请一种信息推送方法的流程图；
图3示意性地示出了本申请中一种获得待推送对象的方法的流程图；
图4示意性地示出了本申请中一种推送流程的操作示意图；
图5示意性地示出了本申请中一种信息推送装置的结构框图；
图6示意性地示出了本申请中一种命中概率确定单元的结构框图；
图7示意性地示出了本申请中一种待推送对象选取单元的结构框图；
图8示意性地示出了本申请中另一种信息推送装置的结构框图。
具体实施方式
首先参考图1，图1示意性地示出了本申请的实施方式可以在其中实施的示例性应用场景。其中，客户端10将用户输入的搜索关键词11提交给搜索服务器20，搜索服务器20基于搜索算法搜索与搜索关键词11匹配的对象，并将搜索到的对象21排序后反馈给客户端10。推送服务器30从搜索服务器20获取搜索到的对象21，依据相关性算法得到与搜索到的对象21相似的对象，并作为候选对象，然后从候选对象中剔除出搜索到的对象21，并将剩余下来的候选对象作为待推送对象。当用户在客户端10上触发任意一个行为时，例如，该行为可以是针对某一个搜索到的对象的一次点击行为，推送服务器30基于该行为对各个待推送对象进行排序，并按照排序中从先到后的顺序选取至少一个待推送对象，最后将选取的待推送对象31推送给客户端10。搜索服务器20和推送服务器30可以为web服务器，也 可以为APP服务器。在电子商务网站上，“对象”即为商品。本领域技术人员可以理解，图1所示的示意图仅是本发明的实施方式可以在其中得以实现的一个示例。本发明实施方式的应用范围不受到该框架任何方面的限制。例如，搜索服务器20和推送服务器30可以为同一个服务器。
本申请的发明人在研究中发现，当用户与电子商务网站进行一次会话(session)时，在用户处于该会话过程的不同阶段时，该用户对于推送方式的需求是不同的。因此，需要向处于不同阶段的用户提供符合其当前需求的推送方式，使得以该推送方式所推送的对象更有可能是用户喜好的目标对象。在本申请中，将用户与电子商务网站进行的一次会话作为一个用户行为生命周期，并将该用户行为生命周期划分为三个不同的阶段，针对任意一个阶段，都有其对应的命中概率模型，根据该命中概率模型可以获得各个待推送对象的命中概率，使得以该命中概率对待推送对象进行的排序是更符合该阶段用户需求的排序方式，根据该排序所推送的对象也更有可能是用户喜好的目标对象。
为使本申请的上述目的、特征和优点能够更加明显易懂，下面结合附图对本申请实施例进行详细描述。
方法实施例
请参阅图2，图2示意性地示出了本申请一种信息推送方法的流程图，例如，该方法可以有执行，该方法包括以下步骤：
步骤201：响应于用户的当前行为，从所述当前行为所属的会话中提取属性特征，并将所述属性特征输入到预置的用户行为生命周期模型，输出所述用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段。
步骤202：将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述 待推送对象的命中概率，其中不同阶段对应的命中概率模型不同。
步骤203：按照所述待推送对象的命中概率从大到小的顺序选区至少一个待推送对象。
步骤204：对选取的待推送对象进行推送。
在本申请中，需要以离线的方式训练得到用户行为生命周期模型和命中概率模型。其中，在用户行为生命周期模型中，将用户行为的生命周期分为以下三个阶段：目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段，而不同的阶段会对应不同的命中概率模型。
下面先说明用户行为生命周期模型的训练并建立的方式。
本发明的发明人在研究中发现，在最开始的时候，用户往往并不清楚自己真正喜好的目标对象是什么，此时用户处于目标对象不明确阶段，在该阶段，用户主要通过浏览闲逛的过程逐步明确自己真正喜好的目标对象。例如，用户想要购买一条连衣裙，但是用户没有明确想要什么风格的连衣裙。
当明确目标对象时，用户进入到目标对象明确选择阶段，在该阶段，用户需要对众多的候选目标对象进行比较分析，并从中选择出符合自己喜好的目标对象，这个阶段一般也会持续很久。例如，用户明确想要波西米亚风格的连衣裙，就会对各种图案、颜色、材质和价格等的波西米亚风格的连衣裙进行比较分析。
用户在经过大量的比较分析后，逐步锁定到某一个目标对象上，此时用户进入到目标对象锁定阶段。例如，用户在锁定到某一款波西米亚风格的连衣裙后，会进一步关注卖家的口碑和已购买用户对该款波西米亚风格的连衣裙的评价等，直至到最后的下单为止。
在将用户行为生命周期划分为三个阶段之后，下面就要在用户与电子商务网站之间的一次会话中(一般情况下，如果一个用户在与网站首次交互后的预定时间内没有再与网站进行任何交互，则认为一次会话结束，否则，就认为一次会话仍在继续。如，预定时间可以为30分钟)，根据用户的每个行为确定用户在每个行为出现时应该具体划分到哪个阶段。
对于一个用户来说，在其与电子商务网站之间的一次会话中，可能会先后多次地输入搜索关键词，并且，在每一次输入搜索关键词之后，都有可能 随之触发多个行为(即触发一个行为序列)，如，点击行为、加入购物车行为、加入收藏夹行为或下单行为等。因此，可以将一个会话中的所有输入搜索关键词的行为以及每次输入搜索关键词之后所触发的行为序列提取出来，并将其进行如下划分：
1、从第一次输入搜索关键词的行为开始，把第一次将对象加入到购物车(或收藏夹)的行为之前的所有行为划分到目标对象不明确阶段。
2、从第一次将对象加入到购物车(或收藏夹)的行为开始，把下单行为前最后一次将对象加入到购物车(或收藏夹)之前的所有行为划分到目标对象明确选择阶段。
3、从最后一次将对象加入到购物车(或收藏夹)的行为开始，把到下单行为为止的所有行为划分到目标对象锁定阶段。
根据以上的划分方式，就可以在训练样本(即会话)中的每个用户行为发生时确定出该用户处于用户行为生命周期中的哪个阶段。下面再以一个训练样本中的一个搜索关键词及其对应的行为序列为基础，说明需要从训练样本中提取的属性特征。当然，需要说明的是，以下属性特征仅仅示意性的，除了可以提取以下属性特征之外，还可以提取其它的属性特征。
Figure PCTCN2015096245-appb-000001
Figure PCTCN2015096245-appb-000002
需要说明的是，term是对搜索关键词进行分词后所得到的最小单位词，可以是名词，也可以是形容词。例如，搜索关键词“韩版修身连衣裙”中共有三个term：韩版、修身和连衣裙。预定时间可以为任意一个时间，例如，15天。
另外，还需要说明的是，在提取属性特征时，之所以选择term_session_action_num、offer_seq_action_num以及offer_session_action_num这三个在线特征作为数值特征而不是哑元特征来使用，是考虑到用户的意图应该是随着行为序列的推进而不断明确的过程，对最终是否能够达成下单，都是有很重要的影响的。
target作为行为的标示值，既可以是连续值，也可以是离散值。如果是离散值，可以根据行为的标示值直接确定用户当前处于用户行为生命周期中的哪个阶段。例如，当target＝0时，说明处于目标对象不确定阶段，当target＝1 时，说明用户处于目标对象明确选择阶段，当target＝2时，说明用户处于目标对象锁定阶段。
如果target采用的是离散值，可以通过对LR(逻辑回归，Logistic Regression)或SVM(支持向量机，Support Vector Machine)模型进行训练来获得一个离散的用户行为生命周期模型。
在本申请的一个优选实施方式中，target是连续值，并根据行为的指示值所落入的区间来确定用户当前处于用户行为生命周期中的哪个阶段。并且，用户行为生命周期模型是对GBRT(Gradient Boost Regression Tree)模型进行训练所获得的连续模型。
例如，可以将一个标示值区间[0，2]划分为3个区域，当行为的指示值落入到[0，0.9)区间时，说明用户处于目标对象不明确阶段，当行为的指示值落入到[0.9，1.5]区间时，说明用户处于目标对象明确选择阶段，当行为的指示值落入到(1.5，2]区间时，说明用户处于目标对象锁定阶段。
另外，在本申请中，目标对象明确选择阶段是连接目标对象不明确阶段和目标对象锁定阶段的重要过程。随着这个过程中行为序列的变化，用户意图(即，具体锁定哪个目标对象)也在逐渐地发生变化。经过目标对象明确选择阶段，用户最终是进入到了目标锁定阶段，完成了下单，还是没有进入到目标锁定阶段，离开了网站，都是用户在目标对象明确选择阶段所触发的行为序列的累积结果。
因此，作为一种优选的实施方式，在目标对象明确选择阶段，不同的行为类型所引起用户意图的变化率是不同的。在计算该阶段内各个行为的标示值时，可以先给每种行为赋予不同的权重值；然后统计该阶段的行为序列中每种行为的次数，并根据其权重进行加权求和；再根据该阶段的标示值区间的跨度以及所有行为的加权求和值，确定每个行为对用户意图带来的行为变化量；最后根据各个行为的行为变化量，计算出各个行为的标示值。
例如，在一个训练样本的目标对象明确选择阶段，共发生如下行为序列：
add、click、click、add、click、add、click
其中，“add”为加入购物车(或收藏夹)行为、“click”为点击行为。
除去作为起点的1个add，共有4个click和2个add。假设add与click的权重分别为3和1，则所有行为的加权求和值为：4×1+3×2＝10。假设该阶段的标示值的区间是[0.9，1.5]，每个click带来的行为变化量是(1.5-0.9)/10×1＝0.06，每个add的行为变化是0.06×3＝0.18。那么，最终的各个行为的标示值依次为：
0.9、0.96、1.02、1.2、1.26、1.44、1.5
在训练得到用户行为生命周期模型之后，就可以根据用户实时输入的搜索关键词以及实时的行为序列，估测出用户当前处于用户行为生命周期的哪个阶段。例如，用户输入一个搜索关键词“连衣裙”之后，如果在其行为序列中的某一个行为是点击对象1，并且根据用户行为生命周期模型计算出该点击行为的标示值为0.7，则估测出用户此时处于用户生命周期中的目标对象目明确阶段。如果用户在该点击对象1的行为之后的某一个行为是将对象2加入到购物车，并且根据用户行为生命周期模型计算出该加入购物车行为的标示值为1.3，则估测出用户此时处于用户生命周期中的目标对象明确选择阶段。
下面再说明命中概率模型的训练并建立的方式。
由于不同的阶段对应不同的命中概率模型，因此，需要为每个阶段分别训练一个命中概率模型，命中概率模型输出的是各个待推送对象的命中概率，即，各个待推送对象恰好为用户喜好的目标对象的概率。
其中，与目标对象不明确阶段对应的命中概率模型为曝光点击转化率模型，与目标对象明确选择阶段对应的命中概率模型为点击收藏转化率模型，与目标对象锁定阶段对应的命中概率模型为点击下单转化率模型。并且，在训练不同的命中概率模型时所提取的属性特征也不同。
也就是说，当处于目标对象不明确阶段时，是以待推送对象的被曝光数量与被点击数量之间的转化率为目标训练命中概率模型的，训练时所提取的主要属性特征包括对象的下单数量、对象的图片质量以及对象是否在用户的偏好类目中等。当处于目标对象明确选择阶段时，是以待推送对象的被点击数量与被关注(被关注包括被加入到收藏夹和被加入到购物车)数量之间的转化率为目标训练命中概率模型的，训练时所提取的主要属性特征包括用户 偏好的风格、偏好价位以及对象的材质等。当处于目标对象锁定阶段时，是以待推送对象的被点击数量与被下单数量之间的转化率为目标训练命中概率模型的，在训练时所提取的主要属性特征包括对象的好评率、商家等级和信用度等特征。
其中，被曝光数量与被点击数量之间的转化率＝被点击数量/被曝光数量，被点击数量与被关注数量之间的转化率＝被关注数量/被点击数量，被点击数量与被下单数量之间的转化率＝被下单数量/被点击数量。
需要说明的是，以上列举的训练各个命中概率模型的属性特征仅仅示意性的，除了可以提取这些属性特征之外，还可以提取其它的属性特征。
相应的，当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，是按照所述待推送对象的被曝光数量与被点击数量之间的转换率从大到小的顺序选取至少一个候选对象；当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，是按照所述待推送对象的被点击数量与被关注数量之间的转换率从大到小的顺序选取至少一个候选对象；当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，是按照所述待推送对象的被点击数量与被下单数量之间的转换率从大到小的顺序选取至少一个候选对象。
在本申请的一个优选实施方式中，命中概率模型是对LR模型进行训练所得到的离散模型。
在本申请中，可以由搜索服务器20获得待推送对象，推送服务器30从搜索服务器20获取待推送对象。
在本申请的一种优选实施方式中，待推送对象是由推送服务器30在本地获得的。请参阅图3，图3示意性地示出了本申请中一种获得待推送对象的方法的流程图。例如，该方法由推送服务器30获得的，该方法可以包括以下步骤：
步骤301、根据相关性算法计算各个对象与搜索到的对象之间的相似度值；
步骤302、按照相似度值从大到小的顺序选取至少一个对象并作为候选对象；
步骤302、从所述候选对象中剔除出搜索到的对象，将剩余下的候选对象作为待推送对象。
可以理解的，在本申请中，通过离线的方式预先建立了用户行为生命周期模型以及与该用户行为生命周期中的每个阶段分别对应的命中概率模型之后，就需要在线地根据用户的当前行为来确定当前会话中的搜索关键词序列和行为序列，并从中提取属性特征，然后输入到用户行为生命周期中，从而估测用户当前处于用户行为生命周期中的哪个阶段，进而就可以利用与该阶段对应的命中概率模型来估测各个待推送对象的命中概率，最后按照命中概率从大到小的顺序对待推送对象进行排序，并选取前N位进行推送。请参阅图4，图4示意性地示出了本申请中一种推送流程的操作示意图。
另外，还需要说明的是，在获得待推送对象后，可以在搜索结果展示页面最下方的推荐区域中展示各个待推送对象，以实现将待推送对象推送给搜索用户。
由上述实施例可以看出，与现有技术相比，本申请的优点在于：
对于在电子商务网上活跃的用户来说，每当触发一个行为时，就可以确定该用户在用户行为生命周期中所处的阶段。当处于不同的阶段时，会利用不同的命中概率模型来确定各个待推送对象的命中概率。以便最后按照命中概率从大到小的顺序对各个待推送对象进行排序，并选取前N位进行推送。由于在用户行为生命周期中，当用户处于不同的阶段时，该用户对于推送方式的需求是不同的，因此，本申请向处于不同阶段的用户提供符合其当前需求的推送方式，使得以该推送方式所推送的对象更有可能是用户喜好的目标对象，从而尽可能地降低用户反复搜索的可能性，提升用户体验的同时，也节约搜索服务器以及推送服务器的资源。
装置实施例
与上述一种信息推送方法相对应，本申请实施例还提供了一种信息推送装置。请参阅图5，图5示意性地示出了本申请中一种信息推送装置的一个实施例的结构框图，该装置包括：生命周期确定单元501、命中概率确定单元 502、待推送对象选取单元503以及推送单元504。下面结合该装置的工作原理进一步介绍其内部结构以及连接关系。
生命周期确定单元501，用于响应于用户的当前行为，从所述当前行为所属的会话中提取属性特征，并将所述属性特征输入到预置的用户行为生命周期模型中，输出所述用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段；
命中概率确定单元502，用于将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，其中不同阶段对应的命中概率模型不同；
待推送对象选取单元503，用于按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象；
推送单元504，用于对选取的待推送对象进行推送。
在本申请的一个优选实施方式中，所述用户行为生命周期模型是对GBRT模型进行训练所得到的连续模型。
在本申请的另一个优选实施方式中，所述命中概率模型是对逻辑回归LR模型进行训练所得到的离散模型。
在本申请的另一个优选实施方式中，所述预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型包括：曝光点击转化率模型、点击收藏转化率模型和点击下单转化率模型；
如图6所示，命中概率确定单元502包括第一确定子单元5021、第二确定子单元5022和第三确定子单元5023；其中，
第一确定子单元5021，用于当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，将从待推送对象和用户中提取的属性特征输入到曝光点击转化率模型中，输出所述待推送对象的被曝光数量与被点击数量之间的转化率；
第二确定子单元5022，用于当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，将从待推送对象和用户中提取的属性特征输 入到点击收藏转化率模型中，输出所述待推送对象的被点击数量与被关注数量之间的转化率；
第三确定子单元5023，用于当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，将从待推送对象和用户中提取的属性特征输入到点击下单转化率模型中，输出所述待推送对象的被点击数量与被下单数量之间的转化率。
在本申请的另一个优选实施方式中，如图7所示，待推送对象选取单元503包括第一选取子单元5031、第二选取子单元5032和第三选取子单元5033；其中，
第一选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，按照所述待推送对象的被曝光数量与被点击数量之间的转换率从大到小的顺序选取至少一个候选对象；
第二选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，按照所述待推送对象的被点击数量与被关注数量之间的转换率从大到小的顺序选取至少一个候选对象；
第三选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，按照所述待推送对象的被点击数量与被下单数量之间的转换率从大到小的顺序选取至少一个候选对象。
在本申请的另一个优选实施方式中，所述待推送对象是在本地获得的，如图8所示(图8仅示出了增加的部分以及增加的部分与图5所示装置之间的连接关系)，该装置还包括：
相似度计算单元801，用于计算根据相关性算法计算各个对象与搜索到的对象之间的相似度值；
候选对象选取单元802，用于按照相似度值从大到小的顺序选取至少一个对象并作为候选对象；
剔除单元803，用于从所述候选对象中剔除出搜索到的对象，将剩余下的候选对象作为待推送对象。
由上述实施例可以看出，与现有技术相比，本申请的优点在于：
对于在电子商务网上活跃的用户来说，每当触发一个行为时，就可以确 定该用户在用户行为生命周期中所处的阶段。当处于不同的阶段时，会利用不同的命中概率模型来确定各个待推送对象的命中概率。以便最后按照命中概率从大到小的顺序对各个待推送对象进行排序，并选取前N位进行推送。由于在用户行为生命周期中，当用户处于不同的阶段时，该用户对于推送方式的需求是不同的，因此，本申请向处于不同阶段的用户提供符合其当前需求的推送方式，使得以该推送方式所推送的对象更有可能是用户喜好的目标对象，从而尽可能地降低用户反复搜索的可能性，提升用户体验的同时，也节约搜索服务器以及推送服务器的资源。
所述领域的技术人员可以清楚地了解到，为了描述的方便和简洁，上述描述的系统、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述到的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性、机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，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单元中，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单元中。上述集成的单元既可以采用硬件的形式实现，可以采用软件功能单元的形式实现。
需要说明的是，本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的 程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，所述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory，ROM)或随机存储记忆体(Random Access Memory，RAM)等。
以上对本申请所提供的一种信息推送方法和装置进行了详细介绍，本文中应用了具体实施例对本申请的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本申请的方法及其核心思想；同时，对于本领域的一般技术人员，依据本申请的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本申请的限制。

Claims (12)

  1. 一种信息推送方法，其特征在于，包括：
    响应于用户的当前行为，从所述当前行为所属的会话中提取属性特征，并将所述属性特征输入到预置的用户行为生命周期模型，输出所述用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段；
    将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，其中不同阶段对应的命中概率模型不同；
    按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象；
    对选取的待推送对象进行推送。
  2. 根据权利要求1所述的方法，其特征在于，所述用户行为生命周期模型是对GBRT模型进行训练所得到的连续模型。
  3. 根据权利要求1所述的方法，其特征在于，所述命中概率模型是对逻辑回归LR模型进行训练所得到的离散模型。
  4. 根据权利要求1所述的方法，其特征在于，所述预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型包括：曝光点击转化率模型、点击收藏转化率模型和点击下单转化率模型；
    则所述将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，包括：
    当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，将从待推送对象和用户中提取的属性特征输入到曝光点击转化率模型中，输出所述待推送对象的被曝光数量与被点击数量之间的转化率；
    或者，
    当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，将从待推送对象和用户中提取的属性特征输入到点击收藏转化率模型中，输出所述待推送对象的被点击数量与被关注数量之间的转化率；
    或者，
    当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，将从待推送对象和用户中提取的属性特征输入到点击下单转化率模型中，输出所述待推送对象的被点击数量与被下单数量之间的转化率。
  5. 根据权利要求4所述的方法，其特征在于，所述按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象包括：
    当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，按照所述待推送对象的被曝光数量与被点击数量之间的转换率从大到小的顺序选取至少一个候选对象；
    或者，
    当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，按照所述待推送对象的被点击数量与被关注数量之间的转换率从大到小的顺序选取至少一个候选对象；
    或者，
    当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，按照所述待推送对象的被点击数量与被下单数量之间的转换率从大到小的顺序选取至少一个候选对象。
  6. 根据权利要求1至5中任意一项所述的方法，其特征在于，所述待推送对象是在本地获得的，所述方法还包括：
    根据相关性算法计算各个对象与搜索到的对象之间的相似度值；
    按照相似度值从大到小的顺序选取至少一个对象并作为候选对象；
    从所述候选对象中剔除出搜索到的对象，将剩余下的候选对象作为待推送对象。
  7. 一种信息推送装置，其特征在于，包括：
    生命周期确定单元，用于响应于用户的当前行为，从所述当前行为所属的会话中提取属性特征，并将所述属性特征输入到预置的用户行为生命周期模型中，输出所述用户当前在用户行为生命周期中的阶段，其中所述用户行为生命周期包括目标对象不明确阶段、目标对象明确选择阶段和目标对象锁定阶段；
    命中概率确定单元，用于将从待推送对象和用户中提取的属性特征输入到预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型中，输出所述待推送对象的命中概率，其中不同阶段对应的命中概率模型不同；
    待推送对象选取单元，用于按照所述待推送对象的命中概率从大到小的顺序选取至少一个待推送对象；
    推送单元，用于对选取的待推送对象进行推送。
  8. 根据权利要求7所述的装置，其特征在于，所述用户行为生命周期模型是对GBRT模型进行训练所得到的连续模型。
  9. 根据权利要求7所述的装置，其特征在于，所述命中概率模型是对逻辑回归LR模型进行训练所得到的离散模型。
  10. 根据权利要求7所述的装置，其特征在于，所述预置的且与所述用户当前在用户行为生命周期中的阶段对应的命中概率模型包括：曝光点击转化率模型、点击收藏转化率模型和点击下单转化率模型；
    所述命中概率确定单元包括第一确定子单元、第二确定子单元和第三确定子单元；其中，
    第一确定子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，将从待推送对象和用户中提取的属性特征输入到曝光点击转化率模型中，输出所述待推送对象的被曝光数量与被点击数量之间的转化率；
    第二确定子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，将从待推送对象和用户中提取的属性特征输入到点击收藏转化率模型中，输出所述待推送对象的被点击数量与被关注数量之间的转化率；
    第三确定子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，将从待推送对象和用户中提取的属性特征输入到点击下单转化率模型中，输出所述待推送对象的被点击数量与被下单数量之间的转化率。
  11. 根据权利要求10所述的装置，其特征在于，所述待推送对象选取单 元包括第一选取子单元、第二选取子单元和第三选取子单元；其中，
    第一选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象不明确阶段时，按照所述待推送对象的被曝光数量与被点击数量之间的转换率从大到小的顺序选取至少一个候选对象；
    第二选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象明确选择阶段时，按照所述待推送对象的被点击数量与被关注数量之间的转换率从大到小的顺序选取至少一个候选对象；
    第三选取子单元，用于当所述用户当前在用户行为生命周期中的阶段为目标对象锁定阶段时，按照所述待推送对象的被点击数量与被下单数量之间的转换率从大到小的顺序选取至少一个候选对象。
  12. 根据权利要求6至11中任意一项所述的装置，其特征在于，所述待推送对象是在本地获得的，所述装置还包括：
    相似度计算单元，用于计算根据相关性算法计算各个对象与搜索到的对象之间的相似度值；
    候选对象选取单元，用于按照相似度值从大到小的顺序选取至少一个对象并作为候选对象；
    剔除单元，用于从所述候选对象中剔除出搜索到的对象，将剩余下的候选对象作为待推送对象。
PCT/CN2015/096245 2014-12-10 2015-12-03 一种信息推送方法和装置 WO2016091114A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201410757445.4A CN105740268B (zh) 2014-12-10 2014-12-10 一种信息推送方法和装置
CN201410757445.4 2014-12-10

Publications (1)

Publication Number Publication Date
WO2016091114A1 true WO2016091114A1 (zh) 2016-06-16

Family

ID=56106700

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/096245 WO2016091114A1 (zh) 2014-12-10 2015-12-03 一种信息推送方法和装置

Country Status (2)

Country Link
CN (1) CN105740268B (zh)
WO (1) WO2016091114A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107800746A (zh) * 2016-09-07 2018-03-13 百度在线网络技术（北京）有限公司 消息推送方法及装置
CN111191142A (zh) * 2018-11-14 2020-05-22 腾讯科技（深圳）有限公司 一种电子资源推荐方法、装置和可读介质
CN111222038A (zh) * 2018-11-26 2020-06-02 北京嘀嘀无限科技发展有限公司 基于生命周期的数据处理方法、装置及电子设备
CN113409084A (zh) * 2017-10-19 2021-09-17 创新先进技术有限公司 模型训练方法、基于模型的用户行为预测方法及装置
CN113763107A (zh) * 2021-01-26 2021-12-07 北京沃东天骏信息技术有限公司 一种对象信息推送方法、装置、设备及存储介质

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107665225B (zh) * 2016-07-29 2022-01-28 北京京东尚科信息技术有限公司 信息推送方法和装置
CN106503131A (zh) * 2016-10-19 2017-03-15 北京小米移动软件有限公司 获取兴趣信息的方法及装置
CN107230136A (zh) * 2017-05-31 2017-10-03 合肥亿迈杰软件有限公司 一种基于大数据的购物排序推送方法
CN107169842A (zh) * 2017-05-31 2017-09-15 合肥亿迈杰软件有限公司 一种基于商品数据的电子商务数据筛选系统
CN107784390A (zh) * 2017-10-19 2018-03-09 北京京东尚科信息技术有限公司 用户生命周期的识别方法、装置、电子设备及存储介质
CN109840788B (zh) * 2017-11-27 2021-11-02 北京京东尚科信息技术有限公司 用于分析用户行为数据的方法及装置
CN110110203B (zh) * 2018-01-11 2023-04-28 腾讯科技（深圳）有限公司 资源信息推送方法及服务器、资源信息展示方法及终端
CN110866207B (zh) * 2018-08-28 2024-04-09 阿里巴巴集团控股有限公司 数据处理方法、装置和机器可读介质
CN111787042B (zh) * 2019-09-19 2022-09-06 北京京东尚科信息技术有限公司 用于推送信息的方法和装置
CN113781149A (zh) * 2021-01-22 2021-12-10 北京沃东天骏信息技术有限公司 信息推荐方法、装置、计算机可读存储介质及电子设备
CN113312913B (zh) * 2021-07-30 2021-10-08 北京惠每云科技有限公司 一种病例书的切分方法、装置、电子设备及可读存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101556603A (zh) * 2009-05-06 2009-10-14 北京航空航天大学 一种用于对检索结果重新排序的协同检索方法
US20100114803A1 (en) * 2008-10-30 2010-05-06 Ae-Kyeung Moon Apparatus and method for modeling user's service use pattern
CN102346894A (zh) * 2010-08-03 2012-02-08 阿里巴巴集团控股有限公司 推荐信息的输出方法、系统及服务器
CN103679494A (zh) * 2012-09-17 2014-03-26 阿里巴巴集团控股有限公司 商品信息推荐方法及装置
CN104809637A (zh) * 2015-05-18 2015-07-29 北京京东尚科信息技术有限公司 计算机实现的商品推荐方法及系统

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8060463B1 (en) * 2005-03-30 2011-11-15 Amazon Technologies, Inc. Mining of user event data to identify users with common interests
US20120143718A1 (en) * 2010-12-03 2012-06-07 Choicestream, Inc. Optimization of a web-based recommendation system
CN102956009B (zh) * 2011-08-16 2017-03-01 阿里巴巴集团控股有限公司 一种基于用户行为的电子商务信息推荐方法与装置
CN102609860A (zh) * 2012-01-20 2012-07-25 彭立发 一种适用于电子商务的商品与信息分类推荐方法及系统
CN103246661B (zh) * 2012-02-07 2016-08-10 阿里巴巴集团控股有限公司 可视化用户行为收集系统及其方法
CN103377242B (zh) * 2012-04-25 2016-06-22 Tcl集团股份有限公司 用户行为分析方法、分析预测方法及电视节目推送系统
CN102999588A (zh) * 2012-11-15 2013-03-27 广州华多网络科技有限公司 一种多媒体应用的推荐方法和系统
CN103927347A (zh) * 2014-04-01 2014-07-16 复旦大学 一种基于用户行为模型和蚁群聚类的协同过滤推荐算法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100114803A1 (en) * 2008-10-30 2010-05-06 Ae-Kyeung Moon Apparatus and method for modeling user's service use pattern
CN101556603A (zh) * 2009-05-06 2009-10-14 北京航空航天大学 一种用于对检索结果重新排序的协同检索方法
CN102346894A (zh) * 2010-08-03 2012-02-08 阿里巴巴集团控股有限公司 推荐信息的输出方法、系统及服务器
CN103679494A (zh) * 2012-09-17 2014-03-26 阿里巴巴集团控股有限公司 商品信息推荐方法及装置
CN104809637A (zh) * 2015-05-18 2015-07-29 北京京东尚科信息技术有限公司 计算机实现的商品推荐方法及系统

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107800746A (zh) * 2016-09-07 2018-03-13 百度在线网络技术（北京）有限公司 消息推送方法及装置
CN107800746B (zh) * 2016-09-07 2022-01-21 百度在线网络技术（北京）有限公司 消息推送方法及装置
CN113409084A (zh) * 2017-10-19 2021-09-17 创新先进技术有限公司 模型训练方法、基于模型的用户行为预测方法及装置
CN111191142A (zh) * 2018-11-14 2020-05-22 腾讯科技（深圳）有限公司 一种电子资源推荐方法、装置和可读介质
CN111191142B (zh) * 2018-11-14 2023-03-14 腾讯科技（深圳）有限公司 一种电子资源推荐方法、装置和可读介质
CN111222038A (zh) * 2018-11-26 2020-06-02 北京嘀嘀无限科技发展有限公司 基于生命周期的数据处理方法、装置及电子设备
CN113763107A (zh) * 2021-01-26 2021-12-07 北京沃东天骏信息技术有限公司 一种对象信息推送方法、装置、设备及存储介质
CN113763107B (zh) * 2021-01-26 2024-05-24 北京沃东天骏信息技术有限公司 一种对象信息推送方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN105740268B (zh) 2019-04-09
CN105740268A (zh) 2016-07-06

Similar Documents

Publication Publication Date Title
WO2016091114A1 (zh) 一种信息推送方法和装置
TWI609278B (zh) Method and system for recommending search words
TWI508011B (zh) Category information providing method and device
WO2015188699A1 (zh) 推荐项目的方法和装置
WO2018014759A1 (zh) 一种聚类数据表的展现方法、装置和系统
CN106682012B (zh) 商品对象信息搜索方法及装置
TW201911080A (zh) 搜索方法、搜索伺服器和搜索系統
US9727906B1 (en) Generating item clusters based on aggregated search history data
CN106055661B (zh) 基于多Markov链模型的多兴趣资源推荐方法
US9324102B2 (en) System and method to retrieve relevant inventory using sketch-based query
JP2015526831A (ja) 製品識別子のラベル付けおよび製品のナビゲーション
WO2017088496A1 (zh) 一种搜索推荐方法、装置、设备及计算机存储介质
CN110163703B (zh) 一种分类模型建立方法、文案推送方法和服务器
TW201342289A (zh) 資訊提供方法、網頁伺服器以及網頁瀏覽器
TW201401088A (zh) 搜索方法和裝置
CN111191133B (zh) 业务搜索处理方法、装置及设备
CN112488781A (zh) 搜索推荐方法、装置、电子设备及可读存储介质
CN114896517A (zh) 一种商品推荐方法、系统、设备及存储介质
CN111967924A (zh) 商品推荐方法、商品推荐装置、计算机设备和介质
US9552425B2 (en) System and method for determining query aspects at appropriate category levels
CN107169837B (zh) 用于辅助搜索的方法、装置、电子设备及计算机可读介质
JP6944320B2 (ja) 情報処理装置、情報処理方法、およびプログラム
CN112784061A (zh) 知识图谱的构建方法、装置、计算设备及存储介质
WO2018000533A1 (zh) 用于提供搜索推荐信息的方法和装置
Ismail et al. Implementation of naive bayes algorithm with particle swarm optimization in classification of dress recommendation

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: 15866848

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: 15866848

Country of ref document: EP

Kind code of ref document: A1