WO2020211616A1 - 用于处理用户交互信息的方法和装置 - Google Patents

用于处理用户交互信息的方法和装置 Download PDF

Info

Publication number
WO2020211616A1
WO2020211616A1 PCT/CN2020/081317 CN2020081317W WO2020211616A1 WO 2020211616 A1 WO2020211616 A1 WO 2020211616A1 CN 2020081317 W CN2020081317 W CN 2020081317W WO 2020211616 A1 WO2020211616 A1 WO 2020211616A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
information
interaction
feature
category
Prior art date
Application number
PCT/CN2020/081317
Other languages
English (en)
French (fr)
Inventor
钟雨
杜睿桓
崔波
寿如阳
林战刚
陈茜
奈尔哈里克什·萨西库马尔
Original Assignee
北京沃东天骏信息技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京沃东天骏信息技术有限公司 filed Critical 北京沃东天骏信息技术有限公司
Priority to US17/604,283 priority Critical patent/US20220198487A1/en
Publication of WO2020211616A1 publication Critical patent/WO2020211616A1/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
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/04—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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • 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
    • 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
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities

Definitions

  • the embodiments of the present disclosure relate to the field of computer technology, and in particular to a method and device for processing user interaction information.
  • a typical application is to apply a predictive model to predict the user's preferences based on the history of the user's interaction with the object (such as browsing product information, purchasing products), and then predict the probability of the user performing related operations.
  • the embodiments of the present disclosure propose methods and devices for processing user interaction information.
  • the embodiments of the present disclosure provide a method for processing user interaction information, the method comprising: obtaining a set of user interaction information associated with a preset interaction operation, wherein the user interaction information includes the type of the interaction object Target information and brand information, user attribute information, and interactive operation time information corresponding to the brand of the interactive object; generate corresponding user interaction features based on the collection of user interaction information; based on user interaction features and pre-trained presets
  • the operation probability generation model determines the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • the pre-trained preset operation probability generation model includes a long and short-term memory network, a first fully connected network, a second fully connected network, and a third fully connected network.
  • the above-mentioned user interaction information further includes the display position of the interactive object
  • the user's interaction characteristics include an interactive operation feature matrix, a user attribute feature vector, a category feature vector, and a brand feature vector
  • the foregoing collection based on user interaction information Generating the corresponding user's interaction feature includes: generating the corresponding user's initial interaction feature matrix according to the user interaction information, wherein the elements in the initial interaction feature matrix are used to characterize the interaction features corresponding to the brand of the interactive object , The row number and column number of the element of the initial interactive operation feature matrix are used to identify the operation time of the interactive operation corresponding to the brand of the interactive object and the display position of the interactive object; the user’s initial interactive operation feature matrix is converted into the corresponding The two-dimensional matrix is used as the corresponding user's interactive operation feature matrix; to obtain the user attribute feature vector generated based on the user attribute information in the user interaction information; to obtain the information associated with the category of the interactive object in the user interaction information to generate The category
  • the aforementioned pre-trained preset operation probability generation model includes at least one preset operation probability generation sub-model corresponding to a category; and the aforementioned preset operation probability generation model based on user interaction characteristics and pre-trained, Determining the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information includes: inputting the user interaction feature generated according to the set of user interaction information into the interaction object class corresponding to the input interaction feature The preset operation probability generation sub-model that matches the purpose is generated to generate the probability that the user performs the target operation associated with the brand of the interaction object in the user interaction information corresponding to the input interaction feature.
  • the above-mentioned user interaction feature generated according to the set of user interaction information is input into the preset operation probability generation sub-model that matches the category of the interaction object corresponding to the input interaction feature, and the interaction with the input is generated
  • the probability that the user performs the target operation associated with the brand of the interaction object in the user interaction information corresponding to the feature includes: inputting the user interaction feature matrix generated according to the set of user interaction information to the input corresponding to the user interaction feature
  • the long and short-term memory network in the generation sub-model of the preset operation probability matching the category generates the corresponding first hidden feature;
  • the user attribute feature vector generated according to the set of user interaction information is input into the interaction feature with the input user
  • the first fully connected network in the generation sub-model of the preset operation probability matching the corresponding category generates the corresponding second hidden feature;
  • the category feature vector and brand feature vector generated according to the collection of user interaction information are input into The preset operation probability that matches the category corresponding to the input user’s interaction feature generates the second fully connected network
  • the above-mentioned preset operation probability generation sub-model is trained and generated by the following steps: obtaining a training sample set, where the training sample includes sample user interaction information and sample annotation information corresponding to the sample user interaction information, and sample user interaction information Including the category information and brand information of the interactive object, user attribute information, and the operation time information of the interactive operation corresponding to the brand of the interactive object.
  • the sample labeling information is used to characterize whether the sample user performs the interaction with the corresponding sample user interaction information.
  • the category in the user interaction information of each sample in the training sample set is consistent; based on the sample user interaction information of the training sample set, the corresponding user sample interaction characteristics are generated; the generated user sample interaction characteristics As input, the sample label information corresponding to the input sample interaction feature is used as the expected output, and the preset operation probability generation sub-model corresponding to the category of the interactive object in the sample user interaction information is trained.
  • the above-mentioned user interaction information is obtained in the following manner: in response to determining that the user interaction information does not include the user identification, extracting the terminal device identification from the user interaction information; acquiring at least one candidate user identification associated with the terminal device identification, Associate user interaction information with at least one candidate user identifier.
  • the foregoing generating corresponding user interaction characteristics based on the set of user interaction information further includes: according to the category group to which the category indicated by the category information of the interactive object in the user interaction information belongs, dividing the user The user interaction information in the interaction information set is divided into at least one interaction information group, where the category group is divided based on the correlation between the categories indicated by the category information; each interaction is determined according to the user interaction information in the interaction information group The number of category types of interactive objects in the information group; the interactive operation characteristic matrix of several users in the category types of each interactive information group is synthesized into a new interactive operation characteristic matrix of the user corresponding to each interactive information group; Several user attribute feature vectors of categories in the information group are synthesized into a user attribute feature matrix corresponding to each interactive information group.
  • the aforementioned preset operation probability generation model corresponds to the category group; and the aforementioned preset operation probability generation model based on the user's interaction characteristics and pre-training determines the interaction object in the corresponding user interaction information.
  • the probability of the target operation associated with the brand includes: according to the category group corresponding to the synthesized new interactive operation characteristic matrix, input the new interactive operation characteristic matrix into the long-short-term memory network in the corresponding preset operation probability generation model , Generate a new first hidden feature; input the synthesized user attribute feature matrix into the first fully connected network in the corresponding preset operation probability generation model to generate a new second hidden feature; combine the generated class
  • the target feature vector and the brand feature vector are input to the second fully connected network in the corresponding preset operation probability generation model to generate the third hidden feature; the new first hidden feature and the new second hidden feature are generated
  • the feature and the third hidden feature are input to the third fully connected network in the corresponding preset operation probability generation model, and the user corresponding to the input user's interaction feature is generated
  • the method further includes: generating and pushing inventory adjustment information corresponding to the interaction object according to the generated probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • an embodiment of the present disclosure provides an apparatus for processing user interaction information.
  • the apparatus includes: an acquiring unit configured to acquire a set of user interaction information associated with a preset interaction operation, wherein the user interaction
  • the information includes category information and brand information of the interactive object, user attribute information, and operation time information of the interactive operation corresponding to the brand of the interactive object;
  • the first generating unit is configured to generate corresponding user information based on the collection of user interaction information Interaction feature;
  • the determining unit is configured to determine the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information based on the user's interaction feature and a pre-trained preset operation probability generation model.
  • the aforementioned pre-trained preset operation probability generation model includes a long and short-term memory network, a first fully connected network, a second fully connected network, and a third fully connected network.
  • the above-mentioned user interaction information further includes the display position of the interactive object, and the interaction characteristics of the user include an interactive operation feature matrix, a user attribute feature vector, a category feature vector, and a brand feature vector;
  • the above-mentioned first generating unit includes: A generating module is configured to generate a corresponding user's initial interactive operation characteristic matrix according to user interaction information, wherein the elements in the initial interactive operation characteristic matrix are used to characterize the interactive operation characteristics corresponding to the brand of the interactive object, and the initial interactive operation The row number and column number where the elements of the feature matrix are located are used to identify the operating time of the interactive operation corresponding to the brand of the interactive object and the display position of the interactive object; the conversion module is configured to convert the user’s initial interactive operation feature matrix into The corresponding two-dimensional matrix is used as the corresponding user interaction feature matrix; the first obtaining module is configured to obtain the user attribute feature vector generated based on the user attribute information in the user interaction information; the second obtaining module is configured to obtain The category feature vector generated
  • the aforementioned pre-trained preset operation probability generation model includes at least one preset operation probability generation sub-model corresponding to a category; the aforementioned determination unit is further configured to: generate users based on a collection of user interaction information.
  • the interaction feature is input to the preset operation probability generation sub-model that matches the category of the interaction object corresponding to the input interaction feature, and the generated interaction feature corresponding to the input interaction feature is associated with the brand of the interaction object in the user interaction information. Probability of target operation.
  • the above-mentioned determining unit includes: a second generating module configured to input the user's interactive operation feature matrix generated according to the set of user interaction information into a category that matches the input user's interactive feature.
  • the long and short-term memory network in the preset operation probability generation sub-model generates the corresponding first hidden feature
  • the third generation module is configured to input the user attribute feature vector generated according to the set of user interaction information to the input user
  • the first fully connected network in the sub-model that matches the preset operation probability of the category corresponding to the interactive feature is generated, and the corresponding second hidden feature is generated
  • the fourth generation module is configured to generate a set of user interaction information
  • the category feature vector and brand feature vector of is input to the second fully connected network in the preset operation probability generation sub-model that matches the category corresponding to the input user’s interaction feature to generate the corresponding third hidden feature; 5.
  • a generating module configured to input the generated first, second, and third hidden features into a preset operation probability generator that matches the category corresponding to the input user’s interaction feature
  • the third fully connected network in the model generates the probability that the user corresponding to the input user's interaction feature performs the target operation associated with the brand of the interaction object in the user interaction information.
  • the above-mentioned preset operation probability generation sub-model is trained and generated by the following steps: obtaining a training sample set, where the training sample includes sample user interaction information and sample annotation information corresponding to the sample user interaction information, and sample user interaction information Including the category information and brand information of the interactive object, user attribute information, and the operation time information of the interactive operation corresponding to the brand of the interactive object.
  • the sample labeling information is used to characterize whether the sample user performs the interaction with the corresponding sample user interaction information.
  • the category in the user interaction information of each sample in the training sample set is consistent; based on the sample user interaction information of the training sample set, the corresponding user sample interaction characteristics are generated; the generated user sample interaction characteristics As input, the sample label information corresponding to the input sample interaction feature is used as the expected output, and the preset operation probability generation sub-model corresponding to the category of the interactive object in the sample user interaction information is trained.
  • the above-mentioned user interaction information is obtained in the following manner: in response to determining that the user interaction information does not include the user identification, extracting the terminal device identification from the user interaction information; acquiring at least one candidate user identification associated with the terminal device identification, Associate user interaction information with at least one candidate user identifier.
  • the above-mentioned first generating unit further includes: a division module configured to group the user interaction information in the user interaction information collection according to the category group to which the category information of the interaction object in the user interaction information belongs The user interaction information is divided into at least one interaction information group, wherein the category group is divided based on the correlation between the categories indicated by the category information; the determining module is configured to determine each group based on the user interaction information in the interaction information group The number of category types of interactive objects in the interactive information group; the first synthesis module is configured to synthesize the interactive operation characteristic matrix of the users of the category types in each interactive information group into the new user's corresponding user in each interactive information group The second synthesis module is configured to synthesize several user attribute feature vectors of categories in each interactive information group into a user attribute feature matrix corresponding to each interactive information group.
  • the foregoing preset operation probability generation model corresponds to a category group; the foregoing determination unit includes: a fifth generation module configured to combine the new category group corresponding to the synthesized new interactive operation feature matrix
  • the interactive operation feature matrix is input to the long and short-term memory network in the corresponding preset operation probability generation model to generate a new first hidden feature;
  • the sixth generation module is configured to input the synthesized user attribute feature matrix to the corresponding
  • the first fully connected network in the preset operation probability generation model generates a new second hidden feature;
  • the seventh generation module is configured to input the generated category feature vector and brand feature vector to the corresponding preset Operate the second fully connected network in the probability generation model to generate the third hidden feature;
  • the eighth generation module is configured to generate the new first hidden feature, the new second hidden feature, and the third hidden feature.
  • the composite feature is input to the third fully connected network in the corresponding preset operation probability generation model, and the probability of the user corresponding to the input user interaction feature performing the target operation associated with the brand of the
  • the device further includes: a second generating unit configured to generate an inventory adjustment corresponding to the interactive object according to the generated probability of the user performing the target operation associated with the brand of the interactive object in the corresponding user interaction information Information and push.
  • a second generating unit configured to generate an inventory adjustment corresponding to the interactive object according to the generated probability of the user performing the target operation associated with the brand of the interactive object in the corresponding user interaction information Information and push.
  • the embodiments of the present disclosure provide an electronic device that includes: one or more processors; a storage device on which one or more programs are stored; when one or more programs are Multiple processors execute, so that one or more processors implement the method described in any implementation manner of the first aspect.
  • the embodiments of the present disclosure provide a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the method as described in any implementation manner in the first aspect is implemented.
  • the method and apparatus for processing user interaction information provided by the embodiments of the present disclosure first obtain a collection of user interaction information associated with preset interaction operations, where the user interaction information includes category information and brand information of the interaction object, and user Attribute information and operation time information of the interactive operation corresponding to the brand of the interactive object; then, based on the collection of user interaction information, generate the corresponding user interaction characteristics; then, based on the user’s interaction characteristics and pre-trained preset operation probability generation
  • the model determines the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • the probability of the user performing the target operation associated with the brand of the interactive object in the corresponding user interaction information is determined according to the user interaction information, thereby providing reliable data support for e-commerce stocking and inventory management decisions.
  • FIG. 1 is an exemplary system architecture diagram in which an embodiment of the present disclosure can be applied
  • Fig. 2 is a flowchart of an embodiment of a method for processing user interaction information according to the present disclosure
  • Fig. 3 is a schematic diagram of an application scenario of a method for processing user interaction information according to an embodiment of the present disclosure
  • FIG. 4 is a flowchart of another embodiment of a method for processing user interaction information according to the present disclosure
  • Fig. 5 is a schematic structural diagram of an embodiment of an apparatus for processing user interaction information according to the present disclosure
  • Fig. 6 is a schematic structural diagram of an electronic device suitable for implementing embodiments of the present disclosure.
  • FIG. 1 shows an exemplary architecture 100 in which the method for processing user interaction information or the apparatus for processing user interaction information of the present disclosure can be applied.
  • the system architecture 100 may include terminal devices 101, 102, 103, a network 104, and a server 105.
  • the network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105.
  • the network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables.
  • the terminal devices 101, 102, 103 interact with the server 105 through the network 104 to receive or send messages and so on.
  • Various communication client applications such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social platform software, can be installed on the terminal devices 101, 102, and 103.
  • the terminal devices 101, 102, and 103 may be hardware or software.
  • the terminal devices 101, 102, 103 can be various electronic devices that have a display screen and support user interaction, including but not limited to smart phones, tablet computers, e-book readers, laptop computers, and desktop computers and many more.
  • the terminal devices 101, 102, and 103 are software, they can be installed in the electronic devices listed above. It can be implemented as multiple software or software modules (for example, to provide distributed services), or as a single software or software module. There is no specific limitation here.
  • the server 105 may be a server that provides various services, for example, a background server that provides support for the content of user interaction objects displayed on the terminal devices 101, 102, and 103.
  • the background server can analyze and process the received user interaction information that characterizes the interaction between the user and the interaction object, and generate processing results (such as the probability of the user performing a preset operation).
  • the server can be hardware or software.
  • the server can be implemented as a distributed server cluster composed of multiple servers, or as a single server.
  • the server is software, it can be implemented as multiple software or software modules (for example, to provide distributed services), or as a single software or software module. There is no specific limitation here.
  • the method for processing user interaction information provided by the embodiments of the present disclosure is generally executed by the server 105, and correspondingly, the device for processing user interaction information is generally set in the server 105.
  • terminal devices, networks, and servers in FIG. 1 are merely illustrative. According to implementation needs, there can be any number of terminal devices, networks and servers.
  • the method for processing user interaction information includes the following steps:
  • Step 201 Obtain a set of user interaction information associated with a preset interaction operation.
  • the executor of the method for processing user interaction information can obtain the collection of user interaction information associated with the preset interaction operation through a wired or wireless connection.
  • the aforementioned preset interactive operations can be set according to actual applications.
  • the aforementioned preset interactive operations may include, but are not limited to, at least one of the following: browsing interactive objects, clicking on the interactive objects, submitting an order for purchasing commodities in the interactive objects, and searching for commodity brand keywords.
  • the user interaction information may be information related to the aforementioned preset interaction operation.
  • the user interaction information may include category information and brand information of the interaction object, user attribute information, and operation time information of the interaction operation corresponding to the brand information of the interaction object.
  • the category information of the aforementioned interactive object can be used to describe the category to which the interactive object belongs.
  • the above category information may include category names.
  • the above category information may also include, but is not limited to, at least one of the following: the number of brands included under the category, Herfindahl-Hirschman Index (HHI).
  • HHI Herfindahl-Hirschman Index
  • the above-mentioned brand information of the interactive object may be information used to evaluate the value of the brand of the interactive object.
  • the above-mentioned brand information may include a brand name.
  • the above-mentioned brand information may also include but is not limited to at least one of the following: brand price index, market share, market share ranking, and the average value of product ratings belonging to the brand.
  • the above-mentioned user attribute information can be used to construct a user profile.
  • the aforementioned user attribute information may include, but is not limited to, at least one of the following: age, gender, occupation, and purchasing power score.
  • the operation time information of the interaction operation corresponding to the brand of the interaction object may be used to represent the time for performing the preset interaction operation on the interaction object of the brand indicated by the brand information.
  • the aforementioned interaction object may be an advertisement of a certain commodity.
  • the above-mentioned execution subject may obtain the user interaction information collection in various ways.
  • the above-mentioned execution subject may obtain a collection of user interaction information from a database server.
  • each piece of information in the above-mentioned user interaction information set may represent that a user has performed a preset interaction operation on an interaction object describing a product belonging to a certain category and a certain brand at a certain moment.
  • the above-mentioned execution subject may first obtain the original data that characterizes the interaction between the user and the interactive object from the terminal (the terminals 101, 102, 103 shown in FIG. 1).
  • the aforementioned raw data may include the category and brand of the commodity indicated by the interactive object, user identification information, and operation time information of the interactive operation.
  • the above-mentioned user identification information may include, but is not limited to, at least one of the following: user identification (identification, ID), and terminal device identification used by the user.
  • the aforementioned terminal device identifier may include, but is not limited to, at least one of the following: UDID (Unique Device Identifier) of the terminal device, and identifier of the browser client (for example, browser Cookie).
  • UDID Unique Device Identifier
  • the aforementioned raw data may be "sports shoes, brand Q, user ID: abc, 2019.3.239: 15: 20".
  • the above-mentioned executive body can also obtain the corresponding category and brand from the database server preset to store a large amount of category information and brand information according to the category and brand of the commodity indicated by the acquired interactive object.
  • the category information and brand information of the user can obtain user interaction information.
  • the number of categories indicated by the above category information may be 200, for example.
  • the user interaction information may also be obtained in the following manner: in response to determining that the user interaction information does not include the user identification, the terminal device identification is extracted from the user interaction information. After that, at least one candidate user identification associated with the terminal device identification can be acquired.
  • the user interaction information may include the device identifier of the terminal device used by the user who initiated the interactive operation, and the candidate user identifier may include the user identifier that has logged in on the terminal device indicated by the terminal device identifier. Then, the user interaction information can be associated with the aforementioned at least one candidate user identifier.
  • the aforementioned user interaction information can be associated with each of the aforementioned candidate user IDs, or can be associated with a designated candidate user ID in the candidate user IDs.
  • the aforementioned designated candidate user identification may be determined according to the operation time information of the interactive operation in the user interactive information. For example, the device identifier "XXX" of the terminal device used by the user is extracted from the user interaction information.
  • the above-mentioned execution subject may also associate the category information and brand information of the interactive object in the interactive information that does not contain the user identification, user attribute information, and the operation time information of the interactive operation corresponding to the brand information of the interactive object with the associated The interaction information of the candidate user identification is merged.
  • Step 202 Based on the set of user interaction information, generate corresponding user interaction characteristics.
  • the above-mentioned execution subject may generate corresponding user interaction characteristics in various ways.
  • the user's interaction characteristics are used to characterize the interaction between the user and a specific interactive object.
  • the aforementioned specific interaction object may be an advertisement of a product belonging to a certain category and a certain brand.
  • the above-mentioned specific interaction object may be designated in advance, for example, it may be designated as an advertisement of a product belonging to category A x brand.
  • the aforementioned specific interaction object may also be determined according to actual applications, for example, it may be determined as the interaction object with the highest number of clicks or views.
  • the above-mentioned execution subject may first extract the interaction situation between the user and the specific interaction object from the user interaction information set obtained in step 201. For example, extract the time when user X clicks on an advertisement of a product belonging to category B and brand P. Afterwards, it is possible to extract the advertising time for the user X to click on products belonging to the same category but of different brands (for example category B, brand T) within a preset time interval (for example, 1 hour, 3 hours, 1 day). Then, the category information and brand information corresponding to the category and brand of the extracted commodity are converted into word vectors. For example, the category information and brand information can be converted into corresponding serial numbers according to a preset correspondence table. Then, the user attribute information of user X can be converted into a word vector. Finally, the converted word vectors are combined to form user interaction characteristics.
  • the user interaction information may also include the display position of the interaction object, and the interaction characteristics of the user may include an interaction operation feature matrix, a user attribute feature vector, a category feature vector, and a brand feature vector.
  • the display position of the aforementioned interactive object may be characterized by the display position of the advertisement on the webpage.
  • the above-mentioned execution subject can generate corresponding user interaction features according to the following steps:
  • the first step is to generate the corresponding user's initial interactive operation characteristic matrix according to the user's interactive information.
  • the above-mentioned executive body can generate user interaction information corresponding to the category information and brand information of the interaction object in the user interaction information, user attribute information, and operation time information of the interaction operation corresponding to the brand information of the interaction object
  • the initial interactive operation feature matrix can be used to characterize the interactive operation features corresponding to the brand of the interactive object.
  • the row number and column number where the elements of the initial interactive operation feature matrix are located may be used to identify the operation time of the interactive operation corresponding to the brand of the interactive object and the display position of the interactive object, respectively.
  • the elements in the above-mentioned initial interactive operation feature matrix can be used to represent at least one of the following: used to indicate the number of interactions with interactive objects of commodities belonging to the same brand, and used to indicate interactions with commodities belonging to the same category but of different brands (ie The number of interactions with interactive objects of competing products is used to indicate the number of interactions with interactive objects of commodities belonging to different categories.
  • the user interaction information includes R time periods (for example, hours, days) and K advertisement display positions.
  • the aforementioned corresponding user's initial interactive operation characteristic matrix may be a matrix of dimensions ⁇ T, K, 3>.
  • each element in the above matrix can be a vector of length 3.
  • the values of the 3 components included in the vector in the t (1 ⁇ t ⁇ R) row and kth (1 ⁇ k ⁇ K) column can respectively represent the user's click or click in the t-th period of time corresponding to the user interaction information. Browsing the number of times that the product indicated by the advertisement in the user interaction information displayed on the k-th ad slot belongs to the same brand, belongs to the same category, different brands, and belongs to different categories.
  • the interactive information may include category information of the category A to which the product XX belongs, brand information of the brand Q to which the product XX belongs, female users, and the operating time is 2019.3.129:05.
  • the daily time period can be divided into three time periods: 0: 00-10: 00, 10:00-18:00, and 18:00-24:00.
  • the elements in the first row and the first column of the initial interactive operation feature matrix can be (3, 5, 0).
  • the above vector can indicate that in the period of 0: 00-10: 00, for the product advertisement of brand Q located in the first advertising slot, the user has clicked 3 times; for the first advertising slot belonging to category A, The user has clicked on a product advertisement that does not belong to brand Q for 5 times in total; for a product advertisement that does not belong to category A in the first ad slot, the user has not clicked.
  • the above K may be 15, for example, and different values of k may represent the kth day, for example.
  • the second step is to convert the user's initial interactive operation feature matrix into a corresponding two-dimensional matrix as the corresponding user's interactive operation feature matrix.
  • the execution subject may perform dimensional conversion according to the initial interactive operation characteristic matrix obtained in the first step, and generate a two-dimensional matrix as the corresponding user interactive operation characteristic matrix.
  • the above-mentioned executive body may convert the initial interactive operation feature matrix with dimensions ⁇ T, K, 3> into a two-dimensional matrix with dimensions ⁇ T, K ⁇ 3>, and use the above-mentioned two-dimensional matrix as the corresponding user's Interactive operation feature matrix.
  • the third step is to obtain the user attribute feature vector generated based on the user attribute information in the user interaction information.
  • the above-mentioned executive body may obtain the user attribute feature vector from a local or communication connected electronic device.
  • the aforementioned user attribute feature vector may be generated based on user attribute information in user interaction information.
  • the above generating method can be flexibly selected according to user attribute information.
  • the user attribute information may be "age: 22".
  • the above-mentioned user attribute feature vector may be a numerical value corresponding to the age, or may be a one-hot encoding (One-Hot Encoding) corresponding to the age group to which the above-mentioned age belongs.
  • the fourth step is to obtain the category feature vector generated based on the information associated with the category of the interactive object in the user interaction information and the brand feature vector generated based on the information associated with the brand of the interactive object in the user interaction information .
  • the above-mentioned executive body may obtain category feature vectors and brand feature vectors from local or communication-connected electronic devices.
  • the above category feature vector may be generated based on the information associated with the category of the interactive object in the user interaction information.
  • the above-mentioned brand feature vector may be generated based on information associated with the brand of the interaction object in the user interaction information.
  • the foregoing generating method may be similar to the foregoing method of generating the user attribute feature vector in the third step, and will not be repeated here.
  • Step 203 Based on the user's interaction characteristics and the pre-trained preset operation probability generation model, determine the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • the above-mentioned execution subject can determine that the user performs a target operation associated with the brand of the interaction object in the corresponding user interaction information The probability.
  • the above-mentioned target operation may include, but is not limited to, purchasing commodities of the brand indicated by the interaction object.
  • the above-mentioned execution subject may input the user interaction characteristics generated in step 202 into a pre-trained preset operation probability generation model.
  • the aforementioned pre-trained preset operation probability generation model may be various sequence models (Sequence Models) pre-trained through machine learning methods. For example, Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Bidirectional Recurrent Neural Network (Bidirectional RNN), etc.
  • RNN Recurrent Neural Network
  • LSTM Long Short-Term Memory
  • Bidirectional Recurrent Neural Network Bidirectional Recurrent Neural Network
  • the aforementioned pre-trained preset operation probability generation model may include a long and short-term memory network, a first fully connected network, a second fully connected network, and a third fully connected network.
  • the aforementioned pre-trained preset operation probability generation model may include at least one preset operation probability generation sub-model corresponding to a category. Because the user's interaction characteristics can be generated based on user interaction information; the user interaction information can include category information of interactive objects; and the category information can indicate the corresponding category. Therefore, the above-mentioned executor can input the user's interaction feature into the preset operation probability generation sub-model that matches the category corresponding to the interaction feature, thereby generating the interaction between the user and the user interaction information corresponding to the input interaction feature. The probability of the target operation associated with the brand of the object.
  • the above-mentioned execution subject may generate the probability of the user performing the target operation associated with the brand of the interaction object in the user interaction information corresponding to the input interaction feature according to the following steps:
  • the first step is to input the user's interactive operation feature matrix generated according to the collection of user interaction information into the long and short-term memory network in the preset operation probability generation sub-model that matches the category corresponding to the input user's interaction feature, and generate The corresponding first hidden feature.
  • the number of hidden states H of the first hidden feature is usually set between 128 and 512. For example, 256 can be selected.
  • the second step is to input the user attribute feature vector generated according to the set of user interaction information into the first fully connected network in the preset operation probability generation sub-model that matches the category corresponding to the input user interaction feature, and generate the corresponding The second hidden feature.
  • the activation function of the above-mentioned first fully connected network may, for example, adopt a tanh function.
  • the third step is to input the category feature vector and brand feature vector generated according to the set of user interaction information into the second full connection in the preset operation probability generation sub-model that matches the category corresponding to the input user interaction feature Network, generate the corresponding third hidden feature.
  • the activation function of the above-mentioned second fully connected network may, for example, adopt a tanh function.
  • the fourth step is to input the generated first, second, and third hidden features into the preset operation probability generation sub-model that matches the category corresponding to the input user’s interaction feature
  • the third fully connected network generates the probability that the user corresponding to the input user's interaction feature performs the target operation associated with the brand of the interaction object in the user interaction information.
  • the activation function of the third fully connected network may, for example, adopt a tanh function.
  • the aforementioned preset operation probability generation sub-model can be trained and generated through the following steps:
  • the first step is to obtain a collection of training samples.
  • Each training sample in the training sample set may include sample user interaction information and sample label information corresponding to the sample user interaction information.
  • the sample user interaction information may include category information and brand information of the interaction object, user attribute information, and operation time information of the interaction operation corresponding to the brand information of the interaction object.
  • the sample labeling information can be used to characterize whether the sample user performs a target operation associated with the brand of the interaction object in the corresponding sample user interaction information.
  • the categories indicated by the category information in the user interaction information of each sample in the training sample set are consistent. For example, the categories indicated by the category information in the training sample set are all "women's clothing".
  • training samples can be obtained in a variety of ways.
  • the execution subject that obtains the training sample set may obtain historical data in a manner as in step 201, and generate sample user interaction information based on the obtained historical data.
  • the execution subject may determine the sample annotation information corresponding to the historical data that characterizes the execution of the target operation associated with the brand of the interaction object in the corresponding sample user interaction information (for example, purchase of the brand of goods) as 1, and the above target operation is not performed
  • the sample label information corresponding to the historical data is determined to be 0.
  • the sample user interaction information can be stored in association with the corresponding sample label information, and finally training samples are obtained.
  • a large number of training samples are formed through a large amount of historical data, and then a training sample set is formed.
  • the executive body that obtains the training sample set can be random
  • the sampling method retains the same number of negative cases as the positive cases.
  • the second step is to generate corresponding user sample interaction characteristics based on the sample user interaction information of the training sample set.
  • the executor for training the preset operation probability generation sub-model is based on the sample user interaction information of the training sample set obtained in the second step, which can be generated by a method similar to step 202 in the foregoing embodiment The sample interaction characteristics of the corresponding users will not be repeated here.
  • the third step is to take the generated user's sample interaction features as input, and the sample label information corresponding to the input sample interaction features as the desired output, and train to obtain the preset corresponding to the category of the interactive object in the sample user interaction information Operation probability generation sub-model.
  • the executor used to train the preset operation probability generation sub-model can input the generated sample interaction features into the initial model to obtain the brand association of the sample user execution and the interaction object in the corresponding sample user interaction information Probability of target operation. Then, a preset loss function can be used to calculate the degree of difference between the obtained probability and the sample annotation information corresponding to the input sample interaction feature. Afterwards, based on the obtained degree of difference, the network parameters of the initial model are adjusted, and the training ends when the preset training end conditions are met. Finally, the initial model obtained by training can be determined as the preset operation probability generation sub-model corresponding to the category corresponding to the sample interaction feature. Further, the network parameters (for example, the weights between network neurons) of the sub-model generated by the preset operation probability obtained by training can be used to determine the information in the user interaction information that has a greater impact on the final generated probability.
  • the network parameters for example, the weights between network neurons
  • the above loss function may adopt a logarithmic loss function.
  • the foregoing preset training end conditions may include but are not limited to at least one of the following: training time exceeds the preset duration; training times exceeds the preset number of times; the calculated difference degree is less than the preset difference threshold; the accuracy rate on the test set reaches The preset accuracy threshold; the coverage on the test set reaches the preset coverage threshold.
  • the foregoing method of adjusting the network parameters of the initial model may include but is not limited to at least one of the following: BP (Back Propagation) algorithm, SGD (Stochastic Gradient Descent) algorithm.
  • the above-mentioned execution subject may also generate inventory adjustment information corresponding to the interaction object according to the generated probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information And push.
  • the above-generated probability can be used to indicate the likelihood of a user to purchase a certain brand of goods.
  • the above-mentioned executive body may also calculate the number of generated probabilities greater than a preset threshold.
  • the above-mentioned execution subject may also generate inventory adjustment information for adjusting the commodities of the corresponding brand according to the above-mentioned calculated number.
  • the above-mentioned execution subject may also send the generated inventory adjustment information to the target device.
  • the aforementioned target device may be a device that controls the distribution of goods in a logistics warehouse. This can provide support for commodity stocking and inventory management.
  • FIG. 3 is a schematic diagram of an application scenario of the method for processing user interaction information according to an embodiment of the present disclosure.
  • the user uses the terminal device 301 to log in to the user account X and then browse the e-commerce webpage.
  • the log 303 of the backend web server 302 records that the user account X has clicked on advertisements of different content at different times.
  • the backend web server 302 can obtain information related to user accounts and advertisements of products belonging to a certain category and a certain brand according to the content recorded in the log 303.
  • the gender corresponding to user account X is "female”
  • the purchasing power index of the user account is 73
  • the number of brands included in category B is 18, and the comprehensive ranking of brand P under category B is the third.
  • a collection 304 of user interaction information is formed.
  • the background web server 302 can generate the interaction feature 306 corresponding to the user account X according to the preset correspondence table 305.
  • the above interactive features can indicate that female user X with a purchasing power index of 73 has clicked on advertisements for products belonging to category B and P brands, advertisements for products belonging to category B and T brands, and advertisements for products belonging to category B and P brands within one hour. Advertising of goods.
  • category B corresponds to number 20
  • brand P corresponds to number 24
  • brand T corresponds to number 28.
  • the back-end web server 302 can input the aforementioned interactive features 306 into a pre-trained preset operation probability generation model, and generate a probability 307 that the user purchases products belonging to the B category and P brands as 0.84.
  • one of the related technologies usually predicts future purchase behaviors based on historical records of users' clicks, browses, purchases, etc., while ignoring the impact of specific interaction objects (such as advertising clicks or exposure).
  • One of the related technologies usually uses shallow models such as Logistic Regression (LR) and Gradient Boosting Decision Tree (GBDT), which results in the inability to capture the time series characteristics of user behavior.
  • LR Logistic Regression
  • GBDT Gradient Boosting Decision Tree
  • the influence of different interaction behaviors on the probability that the user performs a target operation (for example, purchasing a product belonging to a certain category and a certain brand). Since the user interaction information collection contains the time of the interaction, the relevant time series features can be extracted, making the prediction more accurate. Furthermore, it is also possible to guide e-commerce stocking and inventory management according to the obtained probability of the user performing a target operation (for example, purchasing a product belonging to a certain category and a certain brand).
  • FIG. 4 shows a flow 400 of another embodiment of a method for processing user interaction information.
  • the process 400 of the method for processing user interaction information includes the following steps:
  • Step 401 Obtain a set of user interaction information associated with a preset interaction operation.
  • Step 402 Generate a corresponding user's initial interactive operation feature matrix according to user interaction information.
  • Step 403 Obtain a user attribute feature vector generated based on the user attribute information in the user interaction information.
  • Step 404 Obtain a category feature vector generated based on the information associated with the category of the interaction object in the user interaction information and a brand feature vector generated based on the information associated with the brand of the interaction object in the user interaction information.
  • step 401 is consistent with step 201 in the foregoing embodiment.
  • step 201 also applies to step 401.
  • step 402 to step 404 please refer to the description of step 202 in the foregoing embodiment. Repeat it again.
  • Step 405 According to the category group to which the category indicated by the category information of the interaction object in the user interaction information belongs, divide the user interaction information in the user interaction information set into at least one interaction information group.
  • the above-mentioned execution subject may divide the user interaction information in the user interaction information set into at least one interaction information group according to the category group to which the category indicated by the category information of the interaction object in the user interaction information belongs. .
  • the category group can be divided based on the correlation between categories indicated by the category information.
  • the correlation between the above categories can be preset, or it can be determined according to whether the similarity is greater than a preset correlation threshold.
  • the size of the preset correlation threshold often affects the number of categories included in the category group. In practice, the number of categories is often related to the parameters of the preset operating probability generation sub-model.
  • R the total number of time periods included in the user interaction information
  • K the total number of advertising display positions
  • H the number of hidden states of the first hidden feature
  • Step 406 Determine the category number of interactive objects in each interactive information group according to the user interaction information in the interactive information group.
  • the above-mentioned execution subject may determine the category number of interactive objects in each interactive information group.
  • the categories indicated by the category information of the interactive objects in the interactive information group are different.
  • Step 407 Synthesize the interactive operation characteristic matrix of several users of the category types in each interactive information group into a new interactive operation characteristic matrix of the user corresponding to each interactive information group.
  • the above-mentioned execution subject can use matrix conversion to synthesize a new interactive operation feature matrix from the interactive operation feature matrix of several users in the category.
  • the number of category types of interactive objects in a category group may be M.
  • the user interaction operation feature matrix X may be a two-dimensional matrix with a dimension of ⁇ T, K ⁇ 3>.
  • F x can be a block matrix with dimensions of ⁇ K ⁇ 3, M ⁇ K ⁇ 3>, and each block is a square matrix of ⁇ K ⁇ 3, K ⁇ 3>, a total of M blocks.
  • the i-th square matrix of the aforementioned block matrix is the identity matrix, and the other square matrices are zero matrices.
  • Step 408 Combine several user attribute feature vectors of category types in each interactive information group into a user attribute feature matrix corresponding to each interactive information group.
  • the above-mentioned execution subject can use matrix conversion to synthesize a user attribute feature matrix of several user attribute feature vectors of the category.
  • the number of category types of interactive objects in a category group may be M.
  • the user attribute feature vector U can be a vector with dimension N 2 .
  • F U can be a block matrix with dimensions of ⁇ N 2 , M ⁇ N 2 >, each block is a square matrix of ⁇ N 2 , N 2 >, a total of M blocks.
  • the i-th (1 ⁇ i ⁇ M) category in the category group the i-th square matrix of the aforementioned block matrix is the identity matrix, and the other square matrices are zero matrices.
  • Step 409 Based on the user's interaction characteristics and the pre-trained preset operation probability generation model, determine the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • step 409 may be the same as the step 203 in the foregoing embodiment, and the above description of step 203 is also applicable to step 409, which will not be repeated here.
  • the output of the foregoing preset operation probability generation model may be a vector with the same dimension as the number of square matrix blocks (for example, M) in the input user interaction feature.
  • each element of the aforementioned output vector is a probability output result corresponding to each category corresponding to the input user's interaction feature.
  • the above-mentioned execution subject may generate the probability of the user performing the target operation associated with the brand of the interaction object in the user interaction information corresponding to the input interaction feature according to the following steps:
  • the first step according to the category group corresponding to the synthesized new interactive operation feature matrix, input the new interactive operation feature matrix into the long short-term memory network in the corresponding preset operation probability generation model to generate a new first implicit ⁇ Feature;
  • the second step is to input the synthesized user attribute feature matrix into the first fully connected network in the corresponding preset operation probability generation model to generate a new second hidden feature
  • the third step is to input the generated category feature vector and brand feature vector into the second fully connected network in the corresponding preset operation probability generation model to generate the third hidden feature;
  • the fourth step is to input the generated new first hidden feature, new second hidden feature, and third hidden feature into the third fully connected network in the corresponding preset operation probability generation model to generate and input
  • the user’s interaction feature corresponds to the probability that the user performs the target operation associated with the brand of the interaction object in the user interaction information.
  • the training of the foregoing preset operation probability generation model is similar to the description of step 203 in the foregoing embodiment.
  • the difference includes that after the sample interaction feature is generated based on the training sample, the dimensions of the sample interaction feature are converted in the manner described above in steps 406 to 409, and the predictions corresponding to the input sample interaction feature dimensions are obtained through machine learning.
  • the operation probability generation model Suppose the operation probability generation model. Therefore, the preset operation probability generation model obtained by the above training corresponds to the category group to which the category corresponding to the input sample interaction feature belongs.
  • parallel training of the preset operation probability generation model corresponding to the category group including multiple related categories can greatly save computing resources and increase the speed of model training.
  • the above-mentioned execution subject may also generate inventory adjustment information corresponding to the interaction object according to the generated probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information And push.
  • the process 400 of the method for processing user interactive information in this embodiment embodies the division of interactive information groups according to the correlation between categories and the generation of new interactive operations corresponding to each interactive information group. Steps of feature matrix and user attribute feature matrix. Therefore, the solution described in this embodiment can use the features corresponding to multiple categories to generate a matrix conversion when the number of categories indicated by the category information contained in the set of user interaction information is large (for example, greater than 10).
  • the new interactive features corresponding to the category group can be entered into the preset operation probability generation model to obtain results corresponding to multiple categories in the category group.
  • the preset operation probability generation model corresponding to the input user's new interaction feature realizes parallel training of models of multiple related categories, thereby effectively saving computing resources and improving the training efficiency of the model.
  • the present disclosure provides an embodiment of a device for processing user interaction information.
  • the device embodiment corresponds to the method embodiment shown in FIG.
  • the device can be specifically applied to various electronic devices.
  • the apparatus 500 for processing user interaction information includes an acquiring unit 501, a first generating unit 502, and a determining unit 503.
  • the acquiring unit 501 is configured to acquire a collection of user interaction information associated with preset interaction operations.
  • the user interaction information includes category information and brand information of the interaction object, user attribute information, and interaction operations corresponding to the brand of the interaction object
  • the first generating unit 502 is configured to generate corresponding user interaction characteristics based on a collection of user interaction information
  • the determining unit 503 is configured to generate based on user interaction characteristics and pre-trained preset operation probabilities
  • the model determines the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • the specific processing of the acquiring unit 501, the first generating unit 502, and the determining unit and the technical effects brought by them can be referred to in the corresponding embodiment in FIG. 2 respectively.
  • the related description of step 201, step 202 and step 203 will not be repeated here.
  • the aforementioned pre-trained preset operation probability generation model may include a long and short-term memory network, a first fully connected network, a second fully connected network, and a third fully connected network.
  • the aforementioned user interaction information may also include the display position of the interaction object, and the interaction characteristics of the user include an interaction feature matrix, a user attribute feature vector, a category feature vector, and a brand feature vector;
  • the above-mentioned first generation unit 502 may include: a first generation module (not shown in the figure), a conversion module (not shown in the figure), a first acquisition module (not shown in the figure), and a second acquisition module (not shown in the figure) Not shown).
  • the above-mentioned first generating module may be configured to generate a corresponding user's initial interactive operation characteristic matrix according to user interaction information, and the elements in the initial interactive operation characteristic matrix may be used to characterize the interactive operation characteristic corresponding to the brand of the interactive object
  • the row number and column number of the element of the initial interactive operation feature matrix can be used to identify the operation time of the interactive operation corresponding to the brand of the interactive object and the display position of the interactive object, respectively.
  • the aforementioned conversion module may be configured to convert the user's initial interactive operation characteristic matrix into a corresponding two-dimensional matrix as the corresponding user's interactive operation characteristic matrix.
  • the above-mentioned first obtaining module may be configured to obtain a user attribute feature vector generated based on user attribute information in user interaction information.
  • the above-mentioned second acquiring module may be configured to acquire category feature vectors generated based on information associated with the category of the interactive object in the user interaction information and information based on the brand associated with the interactive object in the user interaction information And the generated brand feature
  • the pre-trained preset operation probability generation model may include at least one preset operation probability generation sub-model corresponding to a category; the determination unit 503 may be further configured to: The user’s interaction characteristics generated according to the set of user interaction information are input into the preset operation probability generation sub-model that matches the category of the interaction object corresponding to the input interaction characteristics, and the user executes and executes the sub-model corresponding to the input interaction characteristics. The probability of the target operation associated with the brand of the interactive object in the user interaction information.
  • the above determining unit 503 may include: a second generation module (not shown in the figure), a third generation module (not shown in the figure), and a fourth generation module (not shown in the figure). Not shown in), the fifth generation module (not shown in the figure).
  • the above-mentioned second generating module may be configured to input the user's interactive operation feature matrix generated according to the set of user interaction information into the preset operation probability generation sub-model that matches the category corresponding to the input user's interactive feature The long and short-term memory network in, generates the corresponding first implicit feature.
  • the above-mentioned third generation module may be configured to input the user attribute feature vector generated according to the set of user interaction information into the first preset operation probability generation sub-model matching the category corresponding to the input user interaction feature.
  • the fully connected network generates the corresponding second hidden feature.
  • the above-mentioned fourth generation module may be configured to input the category feature vector and brand feature vector generated according to the set of user interaction information into the preset operation probability generation sub-model that matches the category corresponding to the input user interaction feature
  • the second fully connected network in, generates the corresponding third hidden feature.
  • the above-mentioned fifth generation module may be configured to input the generated first, second, and third hidden features into a preset operation that matches the category corresponding to the input user’s interactive feature
  • the third fully connected network in the probability generation sub-model generates the probability that the user corresponding to the input user's interaction feature performs the target operation associated with the brand of the interaction object in the user interaction information.
  • the aforementioned preset operation probability generation sub-model can be trained and generated by the following steps: Obtain a training sample set, where the training sample can include sample user interaction information and corresponding to sample user interaction information Sample label information, sample user interaction information can include the category information and brand information of the interactive object, user attribute information, and operation time information of the interactive operation corresponding to the brand of the interactive object.
  • the sample label information can be used to characterize whether the sample user performs The target operation associated with the brand of the interaction object in the corresponding sample user interaction information, the category in the user interaction information of each sample in the training sample set is consistent; based on the sample user interaction information of the training sample set, the corresponding user sample can be generated Interactive features; take the generated user's sample interaction features as input, and the sample label information corresponding to the input sample interaction features as the expected output, and train to obtain preset operations corresponding to the categories of the interactive objects in the sample user interaction information Probabilistically generate sub-models.
  • the above-mentioned user interaction information may be obtained in the following manner: in response to determining that the user interaction information does not contain the user identification, extract the terminal device identification from the user interaction information; obtain the terminal device identification The associated at least one candidate user identification associates the user interaction information with the at least one candidate user identification.
  • the above-mentioned first generating unit 502 may further include: a dividing module (not shown in the figure), a determining module (not shown in the figure), and a first synthesis module (not shown in the figure). Not shown), the second synthesis module (not shown in the figure).
  • the above-mentioned dividing module may be configured to divide the user interaction information in the user interaction information set into at least one interaction information group according to the category group to which the category indicated by the category information of the interaction object in the user interaction information belongs. , Where the category group can be divided based on the correlation between categories indicated by the category information.
  • the above determining module may be configured to determine the category number of interactive objects in each interactive information group based on user interaction information in the interactive information group.
  • the above-mentioned first synthesis module may be configured to synthesize the interactive operation characteristic matrices of several users of category types in each interactive information group into a new interactive operation characteristic matrix of the user corresponding to each interactive information group.
  • the above-mentioned second synthesis module may be configured to synthesize several user attribute feature vectors of category types in each interactive information group into a user attribute feature matrix corresponding to each interactive information group.
  • the foregoing preset operation probability generation model may correspond to a category group.
  • the above determination unit 503 may include: a fifth generation module (not shown in the figure), a sixth generation module (not shown in the figure), a seventh generation module (not shown in the figure), and an eighth generation module (not shown in the figure) Not shown).
  • the above-mentioned fifth generation module may be configured to input the new interactive operation feature matrix into the long- and short-term memory in the corresponding preset operation probability generation model according to the category group corresponding to the synthesized new interactive operation feature matrix.
  • the network generates a new first hidden feature.
  • the above-mentioned sixth generation module may be configured to input the synthesized user attribute feature matrix into the first fully connected network in the corresponding preset operation probability generation model to generate a new second hidden feature.
  • the aforementioned seventh generation module may be configured to input the generated category feature vector and brand feature vector into the second fully connected network in the corresponding preset operation probability generation model to generate the third implicit feature.
  • the above-mentioned eighth generation module may be configured to input the generated new first hidden feature, new second hidden feature, and third hidden feature into the third comprehensive model in the corresponding preset operation probability generation model.
  • the network is connected to generate the probability that the user corresponding to the input user's interaction feature performs the target operation associated with the brand of the interaction object in the user interaction information.
  • the apparatus 500 for processing user interaction information may further include: a second generating unit (not shown in the figure), configured to execute and correspond to the generated user
  • the probability of the target operation associated with the brand of the interactive object in the user interactive information generates and pushes the inventory adjustment information corresponding to the interactive object.
  • the apparatus provided by the above-mentioned embodiment of the present disclosure obtains a collection of user interaction information associated with a preset interactive operation through the obtaining unit 501, wherein the user interaction information includes category information and brand information of the interactive object, user attribute information, and interaction
  • the operation time information of the interactive operation corresponding to the brand of the object after that, the first generating unit 502 generates the corresponding user interaction characteristics based on the collection of user interaction information; then, the determining unit 503 is based on the user’s interaction characteristics and pre-trained presets
  • the operation probability generation model determines the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information. It is possible to determine the probability of the user performing the target operation associated with the brand of the interactive object in the corresponding user interaction information according to the user interaction information, thereby providing data support for e-commerce decision-making.
  • FIG. 6 shows a schematic structural diagram of an electronic device (for example, the server in FIG. 1) 600 suitable for implementing embodiments of the present disclosure.
  • the server shown in FIG. 6 is only an example, and should not bring any limitation to the function and scope of use of the embodiments of the present disclosure.
  • the electronic device 600 may include a processing device (such as a central processing unit, a graphics processor, etc.) 601, which can be loaded into a random access device according to a program stored in a read-only memory (ROM) 602 or from a storage device 608.
  • the program in the memory (RAM) 603 executes various appropriate actions and processing.
  • the RAM 603 also stores various programs and data required for the operation of the electronic device 600.
  • the processing device 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604.
  • An input/output (I/O) interface 605 is also connected to the bus 604.
  • the following devices can be connected to the I/O interface 605: including input devices 606 such as touch screen, touch panel, keyboard, mouse, etc.; including output devices 607 such as liquid crystal display (LCD, Liquid Crystal Display), speakers, vibrators, etc. ; Including storage devices 608 such as tapes, hard disks, etc.; and communication devices 609.
  • the communication device 609 may allow the electronic device 600 to perform wireless or wired communication with other devices to exchange data.
  • FIG. 6 shows an electronic device 600 having various devices, it should be understood that it is not required to implement or have all the illustrated devices. It may alternatively be implemented or provided with more or fewer devices. Each block shown in FIG. 6 can represent one device, or can represent multiple devices as needed.
  • the process described above with reference to the flowchart can be implemented as a computer software program.
  • the embodiments of the present disclosure include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart.
  • the computer program may be downloaded and installed from the network through the communication device 609, or installed from the storage device 608, or installed from the ROM 602.
  • the processing device 601 the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.
  • the computer-readable medium described in the embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two.
  • the computer-readable storage medium may be, for example, but not limited to, an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), optical fiber, portable compact disk read only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • the computer-readable storage medium may be any tangible medium that contains or stores a program, and the program may be used by or in combination with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in a baseband or as a part of a carrier wave, and a computer-readable program code is carried therein. This propagated data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • the computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium.
  • the computer-readable signal medium may send, propagate, or transmit the program for use by or in combination with the instruction execution system, apparatus, or device .
  • the program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to: wire, optical cable, RF (Radio Frequency), etc., or any suitable combination of the foregoing.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device; or it may exist alone without being assembled into the electronic device.
  • the above-mentioned computer-readable medium carries one or more programs.
  • the electronic device obtains a collection of user interaction information associated with a preset interaction operation, wherein the user interaction The information includes category information and brand information of the interactive object, user attribute information, and operation time information of the interactive operation corresponding to the brand of the interactive object; based on the collection of user interaction information, the corresponding user interaction characteristics are generated; based on the user's interaction characteristics And a pre-trained preset operation probability generation model to determine the probability of the user performing the target operation associated with the brand of the interaction object in the corresponding user interaction information.
  • the computer program code for performing the operations of the embodiments of the present disclosure can be written in one or more programming languages or a combination thereof, the programming languages including object-oriented programming languages such as Java, Smalltalk, C++, It also includes a conventional procedural programming language, such as "C" language or similar programming languages.
  • the program code can be executed entirely on the user's computer, partly on the user's computer, executed as an independent software package, partly on the user's computer and partly executed on a remote computer, or entirely executed on the remote computer or server.
  • the remote computer can be connected to the user's computer through any kind of network including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, using an Internet service provider to pass Internet connection).
  • LAN local area network
  • WAN wide area network
  • each block in the flowchart or block diagram can represent a module, program segment, or part of code, and the module, program segment, or part of code contains one or more for realizing the specified logical function Executable instructions.
  • the functions marked in the block may also occur in a different order from the order marked in the drawings. For example, two blocks shown in succession can actually be executed substantially in parallel, or they can sometimes be executed in the reverse order, depending on the functions involved.
  • each block in the block diagram and/or flowchart, and the combination of the blocks in the block diagram and/or flowchart can be implemented by a dedicated hardware-based system that performs the specified functions or operations Or it can be realized by a combination of dedicated hardware and computer instructions.
  • the units involved in the embodiments described in the present disclosure can be implemented in software or hardware.
  • the described unit may also be provided in the processor, for example, it may be described as: a processor including an acquiring unit, a first generating unit, and a determining unit.
  • the names of these units do not constitute a limitation on the unit itself under certain circumstances.
  • the acquisition unit can also be described as a unit that "acquires a collection of user interaction information associated with a preset interaction operation, where the user The interactive information includes category information and brand information of the interactive object, user attribute information, and operation time information of the interactive operation corresponding to the brand of the interactive object".

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Accounting & Taxation (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Algebra (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • User Interface Of Digital Computer (AREA)

Abstract

本公开的实施例公开了用于处理用户交互信息的方法和装置。该方法的一具体实施方式包括：获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；基于用户交互信息的集合，生成对应的用户的交互特征；基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。该实施方式实现了根据用户交互信息确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，从而可以为电商的备货及库存管理决策提供数据支持。

Description

用于处理用户交互信息的方法和装置
本专利申请要求于2019年04月15日提交的、申请号为201910300614.4、发明名称为″用于处理用户交互信息的方法和装置″的中国专利申请的优先权，该申请的全文以引用的方式并入本申请中。
技术领域
本公开的实施例涉及计算机技术领域，具体涉及用于处理用户交互信息的方法和装置。
背景技术
随着人工智能技术的发展，机器学习模型在电商领域得到越来越广泛的应用。一种典型的应用是根据用户与对象的交互(例如浏览商品信息、购买商品)历史，应用预测模型来预测用户的偏好，进而预测用户执行相关操作的概率。
发明内容
本公开的实施例提出了用于处理用户交互信息的方法和装置。
第一方面，本公开的实施例提供了一种用于处理用户交互信息的方法，该方法包括：获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；基于用户交互信息的集合，生成对应的用户的交互特征；基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
在一些实施例中，预先训练的预设操作概率生成模型包括长短期记忆网络、第一全连接网络、第二全连接网络和第三全连接网络。
在一些实施例中，上述用户交互信息还包括交互对象的展现位置，用户的交互特征包括交互操作特征矩阵、用户属性特征向量、类目特征向量和品牌特征向量；以及上述基于用户交互信息的集合，生成对应的用户的交互特征，包括：根据用户交互信息，生成对应的用户的初始交互操作特征矩阵，其中，初始交互操作特征矩阵中的元素用于表征与交互对象的品牌对应的交互操作特征，初始交互操作特征矩阵的元素所在的行号和列号分别用于标识与交互对象的品牌对应的交互操作的操作时间和交互对象的展现位置；将用户的初始交互操作特征矩阵转换为对应的二维矩阵作为对应的用户的交互操作特征矩阵；获取基于用户交互信息中的用户属性信息而生成的用户属性特征向量；获取基于与用户交互信息中的交互对象的类目相关联的信息而生成的类目 特征向量和基于与用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
在一些实施例中，上述预先训练的预设操作概率生成模型包括至少一个与类目对应的预设操作概率生成子模型；以及上述基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，包括：将根据用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在一些实施例中，上述将根据用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率，包括：将根据用户交互信息的集合生成的用户的交互操作特征矩阵输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的长短期记忆网络，生成对应的第一隐合特征；将根据用户交互信息的集合生成的用户属性特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第一全连接网络，生成对应的第二隐合特征；将根据用户交互信息的集合生成的类目特征向量和品牌特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第二全连接网络，生成对应的第三隐合特征；将所生成的第一隐合特征、第二隐合特征和第三隐合特征输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在一些实施例中，上述预设操作概率生成子模型通过如下步骤训练生成：获取训练样本集合，其中，训练样本包括样本用户交互信息和与样本用户交互信息对应的样本标注信息，样本用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息，样本标注信息用于表征样本用户是否执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作，训练样本集合中的各样本用户交互信息中的类目一致；基于训练样本集合的样本用户交互信息，生成对应的用户的样本交互特征；将所生成的用户的样本交互特征作为输入，将与输入的样本交互特征对应的样本标注信息作为期望输出，训练得到与样本用户交互信息中的交互对象的类目对应的预设操作概率生成子模型。
在一些实施例中，上述用户交互信息通过如下方式获取：响应于确定用户交互信 息中不包含用户标识，从用户交互信息中提取出终端设备标识；获取终端设备标识关联的至少一个候选用户标识，将用户交互信息关联至至少一个候选用户标识。
在一些实施例中，上述基于用户交互信息的集合，生成对应的用户的交互特征，还包括：根据用户交互信息中的交互对象的类目信息所指示的类目所属的类目组，将用户交互信息集合中的用户交互信息划分为至少一个交互信息组，其中，类目组基于类目信息所指示的类目之间的相关性划分；根据交互信息组中的用户交互信息，确定各交互信息组中的交互对象的类目种类数；将各交互信息组中的类目种类数个用户的交互操作特征矩阵合成为各交互信息组对应的用户的新的交互操作特征矩阵；将各交互信息组中的类目种类数个用户属性特征向量合成为各交互信息组对应的用户属性特征矩阵。
在一些实施例中，上述预设操作概率生成模型与类目组对应；以及上述基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，包括：根据所合成的新的交互操作特征矩阵对应的类目组，将新的交互操作特征矩阵输入至对应的预设操作概率生成模型中的长短期记忆网络，生成新的第一隐合特征；将所合成的用户属性特征矩阵输入至对应的预设操作概率生成模型中的第一全连接网络，生成新的第二隐合特征；将所生成的类目特征向量和品牌特征向量输入至对应的预设操作概率生成模型中的第二全连接网络，生成第三隐合特征；将所生成的新的第一隐合特征、新的第二隐合特征和第三隐合特征输入至对应的预设操作概率生成模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中的交互对象的品牌关联的目标操作的概率。
在一些实施例中，该方法还包括：根据所生成的用户执行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成交互对象对应的库存调整信息并推送。
第二方面，本公开的实施例提供了一种用于处理用户交互信息的装置，该装置包括：获取单元，被配置成获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；第一生成单元，被配置成基于用户交互信息的集合，生成对应的用户的交互特征；确定单元，被配置成基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
在一些实施例中，上述预先训练的预设操作概率生成模型包括长短期记忆网络、 第一全连接网络、第二全连接网络和第三全连接网络。
在一些实施例中，上述用户交互信息还包括交互对象的展现位置，用户的交互特征包括交互操作特征矩阵、用户属性特征向量、类目特征向量和品牌特征向量；上述第一生成单元包括：第一生成模块，被配置成根据用户交互信息，生成对应的用户的初始交互操作特征矩阵，其中，初始交互操作特征矩阵中的元素用于表征与交互对象的品牌对应的交互操作特征，初始交互操作特征矩阵的元素所在的行号和列号分别用于标识与交互对象的品牌对应的交互操作的操作时间和交互对象的展现位置；转换模块，被配置成将用户的初始交互操作特征矩阵转换为对应的二维矩阵作为对应的用户的交互操作特征矩阵；第一获取模块，被配置成获取基于用户交互信息中的用户属性信息而生成的用户属性特征向量；第二获取模块，被配置成获取基于与用户交互信息中的交互对象的类目相关联的信息而生成的类目特征向量和基于与用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
在一些实施例中，上述预先训练的预设操作概率生成模型包括至少一个与类目对应的预设操作概率生成子模型；上述确定单元进一步被配置成：将根据用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在一些实施例中，上述确定单元包括：第二生成模块，被配置成将根据用户交互信息的集合生成的用户的交互操作特征矩阵输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的长短期记忆网络，生成对应的第一隐合特征；第三生成模块，被配置成将根据用户交互信息的集合生成的用户属性特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第一全连接网络，生成对应的第二隐合特征；第四生成模块，被配置成将根据用户交互信息的集合生成的类目特征向量和品牌特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第二全连接网络，生成对应的第三隐合特征；第五生成模块，被配置成将所生成的第一隐合特征、第二隐合特征和第三隐合特征输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在一些实施例中，上述预设操作概率生成子模型通过如下步骤训练生成：获取训练样本集合，其中，训练样本包括样本用户交互信息和与样本用户交互信息对应的样 本标注信息，样本用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息，样本标注信息用于表征样本用户是否执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作，训练样本集合中的各样本用户交互信息中的类目一致；基于训练样本集合的样本用户交互信息，生成对应的用户的样本交互特征；将所生成的用户的样本交互特征作为输入，将与输入的样本交互特征对应的样本标注信息作为期望输出，训练得到与样本用户交互信息中的交互对象的类目对应的预设操作概率生成子模型。
在一些实施例中，上述用户交互信息通过如下方式获取：响应于确定用户交互信息中不包含用户标识，从用户交互信息中提取出终端设备标识；获取终端设备标识关联的至少一个候选用户标识，将用户交互信息关联至至少一个候选用户标识。
在一些实施例中，上述第一生成单元还包括：划分模块，被配置成根据用户交互信息中的交互对象的类目信息所指示的类目所属的类目组，将用户交互信息集合中的用户交互信息划分为至少一个交互信息组，其中，类目组基于类目信息所指示的类目之间的相关性划分；确定模块，被配置成根据交互信息组中的用户交互信息，确定各交互信息组中的交互对象的类目种类数；第一合成模块，被配置成将各交互信息组中的类目种类数个用户的交互操作特征矩阵合成为各交互信息组对应的用户的新的交互操作特征矩阵；第二合成模块，被配置成将各交互信息组中的类目种类数个用户属性特征向量合成为各交互信息组对应的用户属性特征矩阵。
在一些实施例中，上述预设操作概率生成模型与类目组对应；上述确定单元包括：第五生成模块，被配置成根据所合成的新的交互操作特征矩阵对应的类目组，将新的交互操作特征矩阵输入至对应的预设操作概率生成模型中的长短期记忆网络，生成新的第一隐合特征；第六生成模块，被配置成将所合成的用户属性特征矩阵输入至对应的预设操作概率生成模型中的第一全连接网络，生成新的第二隐合特征；第七生成模块，被配置成将所生成的类目特征向量和品牌特征向量输入至对应的预设操作概率生成模型中的第二全连接网络，生成第三隐合特征；第八生成模块，被配置成将所生成的新的第一隐合特征、新的第二隐合特征和第三隐合特征输入至对应的预设操作概率生成模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中的交互对象的品牌关联的目标操作的概率。
在一些实施例中，该装置还包括：第二生成单元，被配置成根据所生成的用户执行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成交互对象对应的库存调整信息并推送。
第三方面，本公开的实施例提供了一种电子设备，该电子设备包括：一个或多个处理器；存储装置，其上存储有一个或多个程序；当一个或多个程序被一个或多个处理器执行，使得一个或多个处理器实现如第一方面中任一实现方式描述的方法。
第四方面，本公开的实施例提供了一种计算机可读介质，其上存储有计算机程序，该程序被处理器执行时实现如第一方面中任一实现方式描述的方法。
本公开的实施例提供的用于处理用户交互信息的方法和装置，首先获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；而后，基于用户交互信息的集合，生成对应的用户的交互特征；之后，基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。实现了根据用户交互信息确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，从而可以为电商的备货及库存管理决策提供可靠的数据支持。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本公开的其它特征、目的和优点将会变得更明显：
图1是本公开的一个实施例可以应用于其中的示例性系统架构图；
图2是根据本公开的用于处理用户交互信息的方法的一个实施例的流程图；
图3是根据本公开的实施例的用于处理用户交互信息的方法的一个应用场景的示意图；
图4是根据本公开的用于处理用户交互信息的方法的又一个实施例的流程图；
图5是根据本公开的用于处理用户交互信息的装置的一个实施例的结构示意图；
图6是适于用来实现本公开的实施例的电子设备的结构示意图。
具体实施方式
下面结合附图和实施例对本公开作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅仅用于解释相关发明，而非对该发明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与有关发明相关的部分。
需要说明的是，在不冲突的情况下，本公开中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本公开。
图1示出了可以应用本公开的用于处理用户交互信息的方法或用于处理用户交互信息的装置的示例性架构100。
如图1所示，系统架构100可以包括终端设备101、102、103，网络104和服务器105。网络104用以在终端设备101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型，例如有线、无线通信链路或者光纤电缆等等。
终端设备101、102、103通过网络104与服务器105交互，以接收或发送消息等。终端设备101、102、103上可以安装有各种通讯客户端应用，例如网页浏览器应用、购物类应用、搜索类应用、即时通信工具、邮箱客户端、社交平台软件等。
终端设备101、102、103可以是硬件，也可以是软件。当终端设备101、102、103为硬件时，可以是具有显示屏并且支持用户交互的各种电子设备，包括但不限于智能手机、平板电脑、电子书阅读器、膝上型便携计算机和台式计算机等等。当终端设备101、102、103为软件时，可以安装在上述所列举的电子设备中。其可以实现成多个软件或软件模块(例如用来提供分布式服务)，也可以实现成单个软件或软件模块。在此不做具体限定。
服务器105可以是提供各种服务的服务器，例如为终端设备101、102、103上显示的用户交互对象的内容提供支持的后台服务器。后台服务器可以对接收到的表征用户与交互对象进行交互的用户交互信息进行分析处理，并生成处理结果(如用户执行预设操作的概率)。
需要说明的是，服务器可以是硬件，也可以是软件。当服务器为硬件时，可以实现成多个服务器组成的分布式服务器集群，也可以实现成单个服务器。当服务器为软件时，可以实现成多个软件或软件模块(例如用来提供分布式服务)，也可以实现成单个软件或软件模块。在此不做具体限定。
需要说明的是，本公开的实施例所提供的用于处理用户交互信息的方法一般由服务器105执行，相应地，用于处理用户交互信息的装置一般设置于服务器105中。
应该理解，图1中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端设备、网络和服务器。
继续参考图2，示出了根据本公开的用于处理用户交互信息的方法的一个实施例的流程200。该用于处理用户交互信息的方法包括以下步骤：
步骤201，获取与预设交互操作关联的用户交互信息的集合。
在本实施例中，用于处理用户交互信息的方法的执行主体(如图1所示的服务器105)可以通过有线或无线连接的方式获取与预设交互操作关联的用户交互信息的集合。其中，上述预设交互操作可以根据实际应用而设定。上述预设交互操作可以包括但不限于以下至少一项：浏览交互对象，点击交互对象，提交购买交互对象中的商品 的订单，搜索商品品牌关键词。用户交互信息可以是上述预设交互操作相关的信息。在这里，用户交互信息可以包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌信息对应的交互操作的操作时间信息。上述交互对象的类目信息可以用于描述交互对象所属的类目。上述类目信息可以包括类目名称。上述类目信息也可以包括但不限于以下至少一项：类目下包括的品牌数目，赫芬达尔一赫希曼指数(Herfindahl-Hirschman Index，HHI)。上述交互对象的品牌信息可以是用于评估交互对象的品牌的价值的信息。上述品牌信息可以包括品牌名称。上述品牌信息也可以包括但不限于以下至少一项：品牌价格指数，市场占有率，市场占有率排名，属于该品牌的商品评分的平均值。上述用户属性信息可以用于构建用户画像(user profile)。上述用户属性信息可以包括但不限于以下至少一项：年龄，性别，职业，购买力得分。上述与交互对象的品牌对应的交互操作的操作时间信息可以用于表征对品牌信息所指示的品牌的交互对象执行上述预设交互操作的时间。作为示例，上述交互对象可以是某商品的广告。
在本实施例中，上述执行主体可以通过各种方式获取用户交互信息集合。作为示例，上述执行主体可以从数据库服务器获取用户交互信息集合。其中，上述用户交互信息集合中的每条信息可以表征某用户在某时刻对描述属于某类目、某品牌的商品的交互对象执行了预设交互操作。作为又一示例，上述执行主体可以首先从终端(如图1所示的终端101、102、103)获取表征用户与交互对象进行交互操作的原始数据。其中，上述原始数据可以包括交互对象所指示的商品的类目和品牌、用户标识信息和交互操作的操作时间信息。上述用户标识信息可以包括但不限于以下至少一项：用户标识(identification，ID)，用户所使用的终端设备标识。其中，上述终端设备标识可以包括但不限于以下至少一项：终端设备的UDID(Unique Device Identifier，唯一设备标识码)，浏览器客户端的标识(例如浏览器Cookie)。例如，上述原始数据可以是″运动鞋，品牌Q，用户ID：abc，2019.3.239：15：20″。之后，上述执行主体还可以根据所获取的交互对象所指示的商品的类目和品牌，从预设存储有大量类目信息和品牌信息的数据库服务器中获取与上述所获取的类目和品牌对应的类目信息和品牌信息，从而得到用户交互信息。实践中，上述类目信息所指示的类目数目例如可以为200。
在本实施例的一些可选的实现方式中，用户交互信息还可以通过如下方式获取：响应于确定用户交互信息中不包含用户标识，从用户交互信息中提取出终端设备标识。之后，可以获取终端设备标识关联的至少一个候选用户标识。其中，用户交互信息中可以包括发起交互操作的用户所使用的终端设备的设备标识，候选用户标识可以包括 在该终端设备标识所指示的终端设备上登录过的用户标识。而后，可以将用户交互信息关联至上述至少一个候选用户标识。在这些实现方式中，可以将上述用户交互信息关联至上述每一个候选用户标识，也可以关联至候选用户标识中的指定候选用户标识。其中，上述指定候选用户标识可以根据用户交互信息中的交互操作的操作时间信息而确定。例如，从用户交互信息中提取出用户所使用的终端设备的设备标识″XXX"。之后，根据上述用户交互信息中交互操作的操作时间信息所指示的操作时间，检索上述操作时间前后在设备标识″XXX″所指示的设备上登录过的账号，将登录时间距上述操作时间的时间间隔最小的账号确定为指定候选用户标识。进而，上述执行主体还可以将不包含用户标识的交互信息中的交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌信息对应的交互操作的操作时间信息与所关联至的候选用户标识的交互信息合并。
步骤202，基于用户交互信息的集合，生成对应的用户的交互特征。
在本实施例中，基于用户交互信息的集合，上述执行主体可以通过各种方式生成对应的用户的交互特征。其中，用户的交互特征用于表征用户与特定交互对象的交互情况。上述特定交互对象可以是属于某类目某品牌的商品的广告。上述特定交互对象可以预先指定，例如可以指定为属于A类目x品牌的商品的广告。上述特定交互对象也可以根据实际应用而确定，例如可以确定为点击量或浏览量最高的交互对象。
作为示例，上述执行主体首先可以从步骤201所获取的用户交互信息集合中提取用户与特定交互对象的交互情况。例如，提取用户X点击属于类目B、品牌P的商品的广告的时间。之后，可以提取用户X在预设时间间隔(例如1小时、3小时、1天)内点击属于同类目但不同品牌(例如类目B、品牌T)的商品的广告时间。然后，将所提取的广告中的商品所属于的类目和品牌对应的类目信息和品牌信息转换为词向量。例如可以根据预设的对应关系表，将类目信息和品牌信息转换成对应的编号。而后，可以将用户X的用户属性信息转换为词向量。最后，将转换得到的各词向量进行组合，形成用户的交互特征。
在本实施例的一些可选的实现方式中，用户交互信息还可以包括交互对象的展现位置，用户的交互特征可以包括交互操作特征矩阵、用户属性特征向量、类目特征向量和品牌特征向量。其中，上述交互对象的展现位置可以采用广告在网页上的展现位次来表征。
在这些实现方式中，上述执行主体可以按照如下步骤生成对应的用户的交互特征：
第一步，根据用户交互信息，生成对应的用户的初始交互操作特征矩阵。
在这些实现方式中，上述执行主体可以根据用户交互信息中的交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌信息对应的交互操作的操作时间信息，生成用户交互信息对应的初始交互操作特征矩阵。其中，上述初始交互操作特征矩阵中的元素可以用于表征与交互对象的品牌对应的交互操作特征。上述初始交互操作特征矩阵的元素所在的行号和列号可以分别用于标识与交互对象的品牌对应的交互操作的操作时间和交互对象的展现位置。上述初始交互操作特征矩阵中的元素例如可以用于表征以下至少一项：用于指示与属于同一品牌的商品的交互对象交互的次数，用于指示与属于同一类目但不同品牌的商品(即竞品)的交互对象交互的次数，用于指示与属于不同类目的商品的交互对象交互的次数。
作为示例，用户交互信息中包括R个时段(例如可以是小时、天)和K个广告展现位置。上述对应的用户的初始交互操作特征矩阵可以是<T，K，3>维度的矩阵。其中，上述矩阵中的每个元素可以是一个长度为3的向量。例如，第t(1≤t≤R)行、第k(1≤k≤K)列的向量所包括的3个分量的值可以分别表征用户交互信息对应的用户在第t个时段内点击或浏览在第k个广告位上展示的该用户交互信息中的广告所指示的商品属于同品牌的商品，属于同类目、不同品牌的商品，属于不同类目的商品的次数。例如，交互信息可以包括商品XX所属的类目A的类目信息、商品XX所属的品牌Q的品牌信息、女性用户、操作时间是2019.3.129：05。可以将每天的时段分为0：00-10：00、10：00-18：00、18：00-24：00三个时段。初始交互操作特征矩阵的第1行、第1列的元素可以为(3，5，0)。那么，上述向量可以表示在0：00-10：00这一时段，对于位于第1广告位的品牌Q的商品广告，该用户累计点击3次；对于位于第1广告位的属于类目A、不属于品牌Q的商品广告，该用户累计点击5次；对于位于第1广告位的不属于类目A的商品广告，该用户没有点击过。实践中，上述K例如可以是15，上述k的不同取值例如可以表征第k天。
第二步，将用户的初始交互操作特征矩阵转换为对应的二维矩阵作为对应的用户的交互操作特征矩阵。
在这些实现方式中，上述执行主体可以根据上述第一步所得到的初始交互操作特征矩阵进行维度转换，生成二维矩阵作为对应的用户的交互操作特征矩阵。
作为示例，上述执行主体可以将维度为<T，K，3>的初始交互操作特征矩阵转换为维度为<T，K×3>的二维矩阵，并将上述二维矩阵作为对应的用户的交互操作特征矩阵。
第三步，获取基于用户交互信息中的用户属性信息而生成的用户属性特征向量。
在这些实现方式中，上述执行主体可以从本地或通信连接的电子设备获取用户属 性特征向量。其中，上述用户属性特征向量可以基于用户交互信息中的用户属性信息而生成。上述生成方法可以根据用户属性信息而灵活选择。例如，用户属性信息可以是″年龄：22"。上述用户属性特征向量可以是年龄对应的数值，也可以是与上述年龄所属于的年龄段对应的独热编码(One-Hot Encoding)。
第四步，获取基于与用户交互信息中的交互对象的类目相关联的信息而生成的类目特征向量和基于与用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
在这些实现方式中，上述执行主体可以从本地或通信连接的电子设备获取类目特征向量和品牌特征向量。其中，上述类目特征向量可以基于与用户交互信息中的交互对象的类目相关联的信息而生成。上述品牌特征向量可以基于与用户交互信息中的交互对象的品牌相关联的信息而生成。上述生成方法可以与上述第三步生成用户属性特征向量的方法类似，此处不再赘述。
步骤203，基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
在本实施例中，基于预先训练的预设操作概率生成模型和步骤202所生成的用户的交互特征，上述执行主体可以确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。其中，上述目标操作可以包括但不限于购买交互对象所指示的品牌的商品。
作为示例，上述执行主体可以将步骤202所生成的用户的交互特征输入至预先训练的预设操作概率生成模型。其中，上述预先训练的预设操作概率生成模型可以是通过机器学习方法而预先训练的各种序列模型(Sequence Models)。例如，循环神经网络(Recurrent Neural Network，RNN)、长短期记忆网络(Long Short-Term Memory，LSTM)、双向递归神经网络(Bidirectional RNN)等。
在本实施例的一些可选的实现方式中，上述预先训练的预设操作概率生成模型可以包括长短期记忆网络、第一全连接网络、第二全连接网络和第三全连接网络。
在这些实现方式中，上述预先训练的预设操作概率生成模型可以包括至少一个与类目对应的预设操作概率生成子模型。由于用户的交互特征可以基于用户交互信息而生成；用户交互信息可以包括交互对象的类目信息；类目信息可以指示对应的类目。因此，上述执行主体可以将用户的交互特征输入至与该交互特征对应的类目相匹配的预设操作概率生成子模型，从而生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
可选地，上述执行主体可以按照如下步骤生成输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率：
第一步，将根据用户交互信息的集合生成的用户的交互操作特征矩阵输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的长短期记忆网络，生成对应的第一隐合特征。
在这些实现方式中，上述第一隐合特征的隐合状态数H通常设置为128～512之间。例如，可以选为256。
第二步，将根据用户交互信息的集合生成的用户属性特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第一全连接网络，生成对应的第二隐合特征。
在这些实现方式中，上述第一全连接网络的激活函数例如可以采用tanh函数。第三步，将根据用户交互信息的集合生成的类目特征向量和品牌特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第二全连接网络，生成对应的第三隐合特征。
在这些实现方式中，上述第二全连接网络的激活函数例如可以采用tanh函数。
第四步，将所生成的第一隐合特征、第二隐合特征和第三隐合特征输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在这些实现方式中，上述第三全连接网络的激活函数例如可以采用tanh函数。
在这些实现方式中，上述预设操作概率生成子模型可以通过如下步骤训练生成：
第一步，获取训练样本集合。训练样本集合中的每个训练样本可以包括样本用户交互信息和与样本用户交互信息对应的样本标注信息。其中，样本用户交互信息可以包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌信息对应的交互操作的操作时间信息。样本标注信息可以用于表征样本用户是否执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作。上述训练样本集合中的各样本用户交互信息中的类目信息所指示的类目一致。例如，上述训练样本集合中的类目信息所指示的类目都是″女装″。
在这些实现方式中，训练样本可以通过多种方式得到。作为示例，可以由获取训练样本集合的执行主体通过如步骤201的方式获取历史数据，并根据所获取的历史数据生成样本用户交互信息。之后，执行主体可以将表征执行与对应的样本用户交互信 息中交互对象的品牌关联的目标操作(例如购买该品牌的商品)的历史数据对应的样本标注信息确定为1，将未执行上述目标操作的历史数据对应的样本标注信息确定为0。而后，可以将样本用户交互信息与对应的样本标注信息关联存储，最终得到训练样本。通过大量的历史数据形成大量的训练样本，进而组成训练样本集合。
可选地，实践中，如果样本标注信息为1的训练样本(即正例)的数目小于样本标注信息为0的训练样本(即负例)的数目，获取训练样本集合的执行主体可以按照随机抽样的方法保留与正例数目相同的负例。
第二步，基于训练样本集合的样本用户交互信息，生成对应的用户的样本交互特征。
在这些实现方式中，用于训练预设操作概率生成子模型的执行主体基于上述第二步所获取的训练样本集合的样本用户交互信息，可以通过与前述实施例中的步骤202类似的方法生成对应的用户的样本交互特征，此处不再赘述。
第三步，将所生成的用户的样本交互特征作为输入，将与输入的样本交互特征对应的样本标注信息作为期望输出，训练得到与样本用户交互信息中的交互对象的类目对应的预设操作概率生成子模型。
在这些实现方式中，用于训练预设操作概率生成子模型的执行主体可以将所生成的样本交互特征输入至初始模型，得到样本用户执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作的概率。然后，可以利用预设的损失函数计算所得到的概率与输入的样本交互特征对应的样本标注信息之间的差异程度。之后，基于所得到的差异程度，调整初始模型的网络参数，并在满足预设的训练结束条件的情况下结束训练。最后，可以将训练得到的初始模型确定为与样本交互特征对应的类目对应的预设操作概率生成子模型。进一步地，还可以根据训练得到的预设操作概率生成子模型的网络参数(例如网络神经元之间的权值)来确定用户交互信息中对最终所生成的概率影响较大的信息。
需要说明的是，上述损失函数可以采用对数损失函数。上述预设的训练结束条件可以包括但不限于以下至少一项：训练时间超过预设时长；训练次数超过预设次数；计算所得的差异程度小于预设的差异阈值；测试集上的准确率达到预设的准确率阈值；测试集上的覆盖率达到预设的覆盖率阈值。上述调整初始模型的网络参数的方法可以包括但不限于以下至少一项：BP(Back Propagation，反向传播)算法，SGD(Stochastic Gradient Descent，随机梯度下降)算法。
在本实施例的一些可选的实现方式中，上述执行主体还可以根据所生成的用户执 行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成交互对象对应的库存调整信息并推送。
在这些实现方式中，上述所生成的概率可以用于指示用户购买某品牌的商品的可能性。上述执行主体还可以计算所生成的概率大于预设阈值的数目。之后，上述执行主体还可以根据上述计算出的数目生成用于调整对应品牌的商品的库存调整信息。可选地，上述执行主体还可以将所生成的库存调整信息发送至目标设备。其中，上述目标设备可以是物流仓库中控制商品分发的设备。从而可以为商品备货和库存管理提供支持。
继续参见图3，图3是根据本公开的实施例的用于处理用户交互信息的方法的应用场景的一个示意图。在图3的应用场景中，用户使用终端设备301登录用户账号X后浏览电商网页。后台网页服务器302的日志303中记录了用户账号X在不同时间点击了不同内容的广告。后台网页服务器302可以根据日志303中记录的内容、获取与用户账号和属于某类目某品牌的商品的广告的信息。例如用户账号X对应的性别为″女″，用户账号的购买力指数为73，B类目下包括的品牌数目为18，P品牌在所属类目B下的综合排名为第3位。从而形成用户交互信息的集合304。之后，后台网页服务器302可以根据预设的对应关系表305，生成了用户账号X对应的交互特征306。上述交互特征可以表征购买力指数为73的女性用户X在一小时内，依次点击过属于B类目P品牌的商品的广告、属于B类目T品牌的商品的广告、属于B类目P品牌的商品的广告。其中，B类目对应编号20，P品牌对应编号24，T品牌对应编号28。而后，后台网页服务器302可以将上述交互特征306输入至预先训练的预设操作概率生成模型，生成用户购买属于B类目P品牌的商品的概率307为0.84。
目前，相关技术之一通常是基于用户的点击、浏览、购买等历史记录对未来购买行为进行预测，而忽略了特定交互对象(例如广告的点击或曝光)带来的影响。相关技术之一通常还采用逻辑回归模型(Logistic Regression，LR)、梯度提升树(Gradient Boosting Decision Tree，GBDT)等浅层模型，导致无法捕捉用户行为的时间序列特征。而本公开的上述实施例提供的方法，通过获取与预设交互操作关联的用户交互信息的集合，实现了可以综合考虑用户对交互对象(例如属于不同类目和不同品牌的商品的广告)的不同交互行为对用户执行目标操作(例如购买属于某类目某品牌的商品)的概率大小的影响。由于用户交互信息集合中包含了交互行为的时间，从而可以提取相关时间序列特征，使得预测更加准确。进而，还可以根据所得到的用户执行目标操作(例如购买属于某类目某品牌的商品)的概率大小来指导电商的备货和库存管理。
进一步参考图4，其示出了用于处理用户交互信息的方法的又一个实施例的流程400。该用于处理用户交互信息的方法的流程400，包括以下步骤：
步骤401，获取与预设交互操作关联的用户交互信息的集合。
步骤402，根据用户交互信息，生成对应的用户的初始交互操作特征矩阵。
步骤403，获取基于用户交互信息中的用户属性信息而生成的用户属性特征向量。
步骤404，获取基于与用户交互信息中的交互对象的类目相关联的信息而生成的类目特征向量和基于与用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
上述步骤401与前述实施例中的步骤201一致，上文针对步骤201的描述也适用于步骤401，步骤402至步骤404的具体实现方式可以对应参考前述实施例中步骤202的描述，此处不再赘述。
步骤405，根据用户交互信息中的交互对象的类目信息所指示的类目所属的类目组，将用户交互信息集合中的用户交互信息划分为至少一个交互信息组。
在本实施例中，上述执行主体可以根据用户交互信息中的交互对象的类目信息所指示的类目所属的类目组，将用户交互信息集合中的用户交互信息划分为至少一个交互信息组。其中，类目组可以基于类目信息所指示的类目之间的相关性划分。上述类目之间的相关性可以预先设定，也可以根据相似度是否大于预设相关阈值来确定。通常，预设相关阈值的大小往往会影响到类目组中包括的类目数目。实践中，类目数目往往与预设操作概率生成子模型的参数有关。例如，如前述实施例中R(用户交互信息中包括的时段的总数)、K(广告展现位置的总数)、H(第一隐合特征的隐合状态数)的含义，令每个类目组中包括的类目数目的平均值为
Figure PCTCN2020081317-appb-000001
通常使得
Figure PCTCN2020081317-appb-000002
的值不超过100,000,000。
步骤406，根据交互信息组中的用户交互信息，确定各交互信息组中的交互对象的类目种类数。
在本实施例中，对于步骤406所划分的至少一个交互信息组，上述执行主体可以确定各交互信息组中的交互对象的类目种类数。通常，作为同一批预设操作概率生成模型的输入，上述交互信息组中的交互对象的类目信息所指示的类目各不相同。
步骤407，将各交互信息组中的类目种类数个用户的交互操作特征矩阵合成为各交互信息组对应的用户的新的交互操作特征矩阵。
在本实施例中，对于每个类目组，上述执行主体可以利用矩阵转换，将类目种类数个用户的交互操作特征矩阵合成一个新的交互操作特征矩阵。
作为示例，某类目组中的交互对象的类目种类数可以为M。用户交互操作特征矩阵X可以为维度为<T，K×3>的二维矩阵。上述执行主体可以令新的交互操作特征矩阵X′＝X×F X。其中，F x可以为<K×3，M×K×3>维度的分块矩阵，每块为<K×3，K×3>的方阵，共M块。对于该类目组中的第i(1≤i≤M)个类目，上述分块矩阵的第i个方阵为单位矩阵，其他方阵为零矩阵。
步骤408，将各交互信息组中的类目种类数个用户属性特征向量合成为各交互信息组对应的用户属性特征矩阵。
在本实施例中，对于每个类目组，上述执行主体可以利用矩阵转换，将类目种类数个用户属性特征向量合成一个用户属性特征矩阵。
作为示例，某类目组中的交互对象的类目种类数可以为M。用户属性特征向量U可以为维度为N 2的向量。上述执行主体可以令用户属性特征矩阵U′＝U×F U。其中，F U可以为<N 2，M×N 2>维度的分块矩阵，每块为<N 2，N 2>的方阵，共M块。对于该类目组中的第i(1≤i≤M)个类目，上述分块矩阵的第i个方阵为单位矩阵，其他方阵为零矩阵。
步骤409，基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
在本实施例中，上述步骤409可以与前述实施例中的步骤203一致，上文针对步骤203的描述也适用于步骤409，此处不再赘述。
需要说明的是，由于作为预设操作概率生成模型的输入的用户的交互特征的维度发生变化，所以预设操作概率生成模型的网络参数也相应地发生变化。上述预设操作概率生成模型的输出可以为维度与输入的用户的交互特征中的方阵块数(例如M)相同的向量。其中，上述输出的向量的每个元素为输入的用户的交互特征对应的各个类目所对应的概率输出结果。
在本实施例的一些可选的实现方式中，上述执行主体可以按照如下步骤生成输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率：
第一步，根据所合成的新的交互操作特征矩阵对应的类目组，将新的交互操作特征矩阵输入至对应的预设操作概率生成模型中的长短期记忆网络，生成新的第一隐合特征；
第二步，将所合成的用户属性特征矩阵输入至对应的预设操作概率生成模型中的第一全连接网络，生成新的第二隐合特征；
第三步，将所生成的类目特征向量和品牌特征向量输入至对应的预设操作概率生成模型中的第二全连接网络，生成第三隐合特征；
第四步，将所生成的新的第一隐合特征、新的第二隐合特征和第三隐合特征输入至对应的预设操作概率生成模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中的交互对象的品牌关联的目标操作的概率。
需要说明的是，上述预设操作概率生成模型的训练与前述实施例中的步骤203的描述类似。不同之处包括，基于训练样本生成样本交互特征之后，采用如前述步骤406至步骤409的方式对样本交互特征的维度进行转换，通过机器学习的方法训练得到与输入的样本交互特征维度对应的预设操作概率生成模型。从而，上述训练得到的预设操作概率生成模型与输入的样本交互特征对应的类目所属的类目组相对应。在类目数目较大时，将与包括多个相关类目的类目组对应的预设操作概率生成模型进行并行训练，可以极大地节约计算资源，提升模型训练速度。实践中，在类目数目为200时，为每个类目单独训练预设操作概率生成子模型需要至少200台4卡GPU(Graphics Processing Unit，图形处理器)机器，训练时间超过24小时。而采用将上述200个类目划分为7个相关类目组、并行训练7个与类目组对应的预设操作概率生成模型，仅需要7个4卡GPU机器。
在本实施例的一些可选的实现方式中，上述执行主体还可以根据所生成的用户执行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成交互对象对应的库存调整信息并推送。
在这些实现方式中，具体实现方式可以对应参考前述实施例中的相关描述，此处不再赘述。
从图4中可以看出，本实施例中的用于处理用户交互信息的方法的流程400体现了根据类目之间的相关性划分交互信息组并生成各交互信息组对应的新的交互操作特征矩阵和用户属性特征矩阵的步骤。由此，本实施例描述的方案可以在用户交互信息的集合所包含的类目信息所指示的类目的数量较大(例如大于10)时，利用多个类目对应的特征经矩阵转换生成类目组对应的新的交互特征，输入预设操作概率生成模型后可以得到与类目组中多个类目对应的结果。此外，与输入的用户的新的交互特征对应的预设操作概率生成模型实现了多个相关类目的模型并行训练，从而有效地节约了计算资源，提高了模型的训练效率。
进一步参考图5，作为对上述各图所示方法的实现，本公开提供了用于处理用户交互信息的装置的一个实施例，该装置实施例与图2所示的方法实施例相对应，该装 置具体可以应用于各种电子设备中。
如图5所示，本实施例提供的用于处理用户交互信息的装置500包括获取单元501、第一生成单元502和确定单元503。其中，获取单元501，被配置成获取与预设交互操作关联的用户交互信息的集合，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；第一生成单元502，被配置成基于用户交互信息的集合，生成对应的用户的交互特征；确定单元503，被配置成基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
在本实施例中，用于处理用户交互信息的装置500中：获取单元501、第一生成单元502和确定单元的具体处理及其所带来的技术效果可分别参考图2对应实施例中的步骤201、步骤202和步骤203的相关说明，在此不再赘述。
在本实施例的一些可选的实现方式中，上述预先训练的预设操作概率生成模型可以包括长短期记忆网络、第一全连接网络、第二全连接网络和第三全连接网络。
在本实施例的一些可选的实现方式中，上述用户交互信息还可以包括交互对象的展现位置，用户的交互特征包括交互操作特征矩阵、用户属性特征向量、类目特征向量和品牌特征向量；上述第一生成单元502可以包括：第一生成模块(图中未示出)、转换模块(图中未示出)、第一获取模块(图中未示出)、第二获取模块(图中未示出)。其中，上述第一生成模块，可以被配置成根据用户交互信息，生成对应的用户的初始交互操作特征矩阵，初始交互操作特征矩阵中的元素可以用于表征与交互对象的品牌对应的交互操作特征，初始交互操作特征矩阵的元素所在的行号和列号可以分别用于标识与交互对象的品牌对应的交互操作的操作时间和交互对象的展现位置。上述转换模块，可以被配置成将用户的初始交互操作特征矩阵转换为对应的二维矩阵作为对应的用户的交互操作特征矩阵。上述第一获取模块，可以被配置成获取基于用户交互信息中的用户属性信息而生成的用户属性特征向量。上述第二获取模块，可以被配置成获取基于与用户交互信息中的交互对象的类目相关联的信息而生成的类目特征向量和基于与用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
在本实施例的一些可选的实现方式中，上述预先训练的预设操作概率生成模型可以包括至少一个与类目对应的预设操作概率生成子模型；上述确定单元503可以进一步被配置成：将根据用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在本实施例的一些可选的实现方式中，上述确定单元503可以包括：第二生成模块(图中未示出)、第三生成模块(图中未示出)、第四生成模块(图中未示出)、第五生成模块(图中未示出)。其中，上述第二生成模块，可以被配置成将根据用户交互信息的集合生成的用户的交互操作特征矩阵输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的长短期记忆网络，生成对应的第一隐合特征。上述第三生成模块，可以被配置成将根据用户交互信息的集合生成的用户属性特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第一全连接网络，生成对应的第二隐合特征。上述第四生成模块，可以被配置成将根据用户交互信息的集合生成的类目特征向量和品牌特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第二全连接网络，生成对应的第三隐合特征。上述第五生成模块，可以被配置成将所生成的第一隐合特征、第二隐合特征和第三隐合特征输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
在本实施例的一些可选的实现方式中，上述预设操作概率生成子模型可以通过如下步骤训练生成：获取训练样本集合，其中，训练样本可以包括样本用户交互信息和与样本用户交互信息对应的样本标注信息，样本用户交互信息可以包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息，样本标注信息可以用于表征样本用户是否执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作，训练样本集合中的各样本用户交互信息中的类目一致；基于训练样本集合的样本用户交互信息，可以生成对应的用户的样本交互特征；将所生成的用户的样本交互特征作为输入，将与输入的样本交互特征对应的样本标注信息作为期望输出，训练得到与样本用户交互信息中的交互对象的类目对应的预设操作概率生成子模型。
在本实施例的一些可选的实现方式中，上述用户交互信息可以通过如下方式获取：响应于确定用户交互信息中不包含用户标识，从用户交互信息中提取出终端设备标识；获取终端设备标识关联的至少一个候选用户标识，将用户交互信息关联至至少一个候选用户标识。
在本实施例的一些可选的实现方式中，上述第一生成单元502还可以包括：划分模块(图中未示出)、确定模块(图中未示出)、第一合成模块(图中未示出)、第二合成模块(图中未示出)。其中，上述划分模块，可以被配置成根据用户交互信息中的交 互对象的类目信息所指示的类目所属的类目组，将用户交互信息集合中的用户交互信息划分为至少一个交互信息组，其中，类目组可以基于类目信息所指示的类目之间的相关性划分。上述确定模块，可以被配置成根据交互信息组中的用户交互信息，确定各交互信息组中的交互对象的类目种类数。上述第一合成模块，可以被配置成将各交互信息组中的类目种类数个用户的交互操作特征矩阵合成为各交互信息组对应的用户的新的交互操作特征矩阵。上述第二合成模块，可以被配置成将各交互信息组中的类目种类数个用户属性特征向量合成为各交互信息组对应的用户属性特征矩阵。
在本实施例的一些可选的实现方式中，上述预设操作概率生成模型可以与类目组对应。上述确定单元503可以包括：第五生成模块(图中未示出)、第六生成模块(图中未示出)、第七生成模块(图中未示出)、第八生成模块(图中未示出)。其中，上述第五生成模块，可以被配置成根据所合成的新的交互操作特征矩阵对应的类目组，将新的交互操作特征矩阵输入至对应的预设操作概率生成模型中的长短期记忆网络，生成新的第一隐合特征。上述第六生成模块，可以被配置成将所合成的用户属性特征矩阵输入至对应的预设操作概率生成模型中的第一全连接网络，生成新的第二隐合特征。上述第七生成模块，可以被配置成将所生成的类目特征向量和品牌特征向量输入至对应的预设操作概率生成模型中的第二全连接网络，生成第三隐合特征。上述第八生成模块，可以被配置成将所生成的新的第一隐合特征、新的第二隐合特征和第三隐合特征输入至对应的预设操作概率生成模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中的交互对象的品牌关联的目标操作的概率。
在本实施例的一些可选的实现方式中，用于处理用户交互信息的装置500还可以包括：第二生成单元(图中未示出)，被配置成根据所生成的用户执行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成交互对象对应的库存调整信息并推送。
本公开的上述实施例提供的装置，通过获取单元501获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；之后，第一生成单元502基于用户交互信息的集合，生成对应的用户的交互特征；而后，确定单元503基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。实现了根据用户交互信息确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，从而可以为电商决策提供数据支持。
下面参考图6，下面参考图6，其示出了适于用来实现本公开的实施例的电子设备(例如图1中的服务器)600的结构示意图。图6示出的服务器仅仅是一个示例，不应对本公开的实施例的功能和使用范围带来任何限制。
如图6所示，电子设备600可以包括处理装置(例如中央处理器、图形处理器等)601，其可以根据存储在只读存储器(ROM)602中的程序或者从存储装置608加载到随机访问存储器(RAM)603中的程序而执行各种适当的动作和处理。在RAM 603中，还存储有电子设备600操作所需的各种程序和数据。处理装置601、ROM 602以及RAM603通过总线604彼此相连。输入/输出(I/O)接口605也连接至总线604。
通常，以下装置可以连接至I/O接口605：包括例如触摸屏、触摸板、键盘、鼠标等的输入装置606；包括例如液晶显示器(LCD，Liquid Crystal Display)、扬声器、振动器等的输出装置607；包括例如磁带、硬盘等的存储装置608；以及通信装置609。通信装置609可以允许电子设备600与其他设备进行无线或有线通信以交换数据。虽然图6示出了具有各种装置的电子设备600，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。图6中示出的每个方框可以代表一个装置，也可以根据需要代表多个装置。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信装置609从网络上被下载和安装，或者从存储装置608被安装，或者从ROM 602被安装。在该计算机程序被处理装置601执行时，执行本公开的实施例的方法中限定的上述功能。
需要说明的是，本公开的实施例所述的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开的实施例中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开的实施例中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号， 其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(Radio Frequency，射频)等等，或者上述的任意合适的组合。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备：获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；基于用户交互信息的集合，生成对应的用户的交互特征；基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
可以以一种或多种程序设计语言或其组合来编写用于执行本公开的实施例的操作的计算机程序代码，所述程序设计语言包括面向对象的程序设计语言一诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言一诸如″C″语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)-连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开的各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，该模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本公开的实施例中所涉及到的单元可以通过软件的方式实现，也可以通过硬件的方式来实现。所描述的单元也可以设置在处理器中，例如，可以描述为：一种处理器，包括获取单元、第一生成单元、确定单元。其中，这些单元的名称在某种情况下并不构成对该单元本身的限定，例如，获取单元还可以被描述为″获取与预设交互操作关联的用户交互信息的集合的单元，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息″。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开的实施例中所涉及的发明范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离上述发明构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开的实施例中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。

Claims (22)

  1. 一种用于处理用户交互信息的方法，包括：
    获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；
    基于所述用户交互信息的集合，生成对应的用户的交互特征；
    基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
  2. 根据权利要求1所述的方法，其中，所述预先训练的预设操作概率生成模型包括长短期记忆网络、第一全连接网络、第二全连接网络和第三全连接网络。
  3. 根据权利要求2所述的方法，其中，所述用户交互信息还包括交互对象的展现位置，用户的交互特征包括交互操作特征矩阵、用户属性特征向量、类目特征向量和品牌特征向量；以及
    所述基于所述用户交互信息的集合，生成对应的用户的交互特征，包括：
    根据用户交互信息，生成对应的用户的初始交互操作特征矩阵，其中，初始交互操作特征矩阵中的元素用于表征与交互对象的品牌对应的交互操作特征，初始交互操作特征矩阵的元素所在的行号和列号分别用于标识与交互对象的品牌对应的交互操作的操作时间和交互对象的展现位置；
    将所述用户的初始交互操作特征矩阵转换为对应的二维矩阵作为对应的用户的交互操作特征矩阵；
    获取基于所述用户交互信息中的用户属性信息而生成的用户属性特征向量；
    获取基于与所述用户交互信息中的交互对象的类目相关联的信息而生成的类目特征向量和基于与所述用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
  4. 根据权利要求3所述的方法，其中，所述预先训练的预设操作概率生成模型包括至少一个与类目对应的预设操作概率生成子模型；以及
    所述基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，包括：
    将根据所述用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对 应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
  5. 根据权利要求4所述的方法，其中，所述将根据所述用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率，包括：
    将根据所述用户交互信息的集合生成的用户的交互操作特征矩阵输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的长短期记忆网络，生成对应的第一隐含特征；
    将根据所述用户交互信息的集合生成的用户属性特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第一全连接网络，生成对应的第二隐含特征；
    将根据所述用户交互信息的集合生成的类目特征向量和品牌特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第二全连接网络，生成对应的第三隐含特征；
    将所生成的第一隐含特征、第二隐含特征和第三隐含特征输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
  6. 根据权利要求5所述的方法，其中，所述预设操作概率生成子模型通过如下步骤训练生成：
    获取训练样本集合，其中，训练样本包括样本用户交互信息和与样本用户交互信息对应的样本标注信息，样本用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息，样本标注信息用于表征样本用户是否执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作，所述训练样本集合中的各样本用户交互信息中的类目一致；
    基于所述训练样本集合的样本用户交互信息，生成对应的用户的样本交互特征；
    将所生成的用户的样本交互特征作为输入，将与输入的样本交互特征对应的样本标注信息作为期望输出，训练得到与样本用户交互信息中的交互对象的类目对应的预设操作概率生成子模型。
  7. 根据权利要求1所述的方法，其中，所述用户交互信息通过如下方式获取：
    响应于确定所述用户交互信息中不包含用户标识，从所述用户交互信息中提取出 终端设备标识；
    获取所述终端设备标识关联的至少一个候选用户标识，将所述用户交互信息关联至至少一个所述候选用户标识。
  8. 根据权利要求3所述的方法，其中，所述基于所述用户交互信息的集合，生成对应的用户的交互特征，还包括：
    根据用户交互信息中的交互对象的类目信息所指示的类目所属的类目组，将所述用户交互信息集合中的用户交互信息划分为至少一个交互信息组，其中，所述类目组基于类目信息所指示的类目之间的相关性划分；
    根据交互信息组中的用户交互信息，确定各交互信息组中的交互对象的类目种类数；
    将各交互信息组中的类目种类数个用户的交互操作特征矩阵合成为各交互信息组对应的用户的新的交互操作特征矩阵；
    将各交互信息组中的类目种类数个用户属性特征向量合成为各交互信息组对应的用户属性特征矩阵。
  9. 根据权利要求8所述的方法，其中，所述预设操作概率生成模型与类目组对应；以及
    所述基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率，包括：
    根据所合成的新的交互操作特征矩阵对应的类目组，将所述新的交互操作特征矩阵输入至对应的预设操作概率生成模型中的长短期记忆网络，生成新的第一隐含特征；
    将所合成的用户属性特征矩阵输入至所述对应的预设操作概率生成模型中的第一全连接网络，生成新的第二隐含特征；
    将所生成的类目特征向量和品牌特征向量输入至所述对应的预设操作概率生成模型中的第二全连接网络，生成第三隐含特征；
    将所生成的新的第一隐含特征、新的第二隐含特征和第三隐含特征输入至所述对应的预设操作概率生成模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中的交互对象的品牌关联的目标操作的概率。
  10. 根据权利要求1-9任一所述的方法，其中，所述方法还包括：
    根据所生成的用户执行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成所述交互对象对应的库存调整信息并推送。
  11. 一种用于处理用户交互信息的装置，包括：
    获取单元，被配置成获取与预设交互操作关联的用户交互信息的集合，其中，用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息；
    第一生成单元，被配置成基于所述用户交互信息的集合，生成对应的用户的交互特征；
    确定单元，被配置成基于用户的交互特征和预先训练的预设操作概率生成模型，确定用户执行与对应的用户交互信息中的交互对象的品牌关联的目标操作的概率。
  12. 根据权利要求11所述的装置，其中，所述预先训练的预设操作概率生成模型包括长短期记忆网络、第一全连接网络、第二全连接网络和第三全连接网络。
  13. 根据权利要求12所述的装置，其中，所述用户交互信息还包括交互对象的展现位置，用户的交互特征包括交互操作特征矩阵、用户属性特征向量、类目特征向量和品牌特征向量；
    所述第一生成单元包括：
    第一生成模块，被配置成根据用户交互信息，生成对应的用户的初始交互操作特征矩阵，其中，初始交互操作特征矩阵中的元素用于表征与交互对象的品牌对应的交互操作特征，初始交互操作特征矩阵的元素所在的行号和列号分别用于标识与交互对象的品牌对应的交互操作的操作时间和交互对象的展现位置；
    转换模块，被配置成将所述用户的初始交互操作特征矩阵转换为对应的二维矩阵作为对应的用户的交互操作特征矩阵；
    第一获取模块，被配置成获取基于所述用户交互信息中的用户属性信息而生成的用户属性特征向量；
    第二获取模块，被配置成获取基于与所述用户交互信息中的交互对象的类目相关联的信息而生成的类目特征向量和基于与所述用户交互信息中的交互对象的品牌相关联的信息而生成的品牌特征向量。
  14. 根据权利要求13所述的装置，其中，所述预先训练的预设操作概率生成模型包括至少一个与类目对应的预设操作概率生成子模型；
    所述确定单元进一步被配置成：
    将根据所述用户交互信息的集合生成的用户的交互特征输入至与输入的交互特征对应的交互对象的类目相匹配的预设操作概率生成子模型，生成与输入的交互特征对应的、用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
  15. 根据权利要求14所述的装置，其中，所述确定单元包括：
    第二生成模块，被配置成将根据所述用户交互信息的集合生成的用户的交互操作特征矩阵输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的长短期记忆网络，生成对应的第一隐含特征；
    第三生成模块，被配置成将根据所述用户交互信息的集合生成的用户属性特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第一全连接网络，生成对应的第二隐含特征；
    第四生成模块，被配置成将根据所述用户交互信息的集合生成的类目特征向量和品牌特征向量输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第二全连接网络，生成对应的第三隐含特征；
    第五生成模块，被配置成将所生成的第一隐含特征、第二隐含特征和第三隐含特征输入至与输入的用户的交互特征对应的类目相匹配的预设操作概率生成子模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中交互对象的品牌关联的目标操作的概率。
  16. 根据权利要求15所述的装置，其中，所述预设操作概率生成子模型通过如下步骤训练生成：
    获取训练样本集合，其中，训练样本包括样本用户交互信息和与样本用户交互信息对应的样本标注信息，样本用户交互信息包括交互对象的类目信息和品牌信息、用户属性信息和与交互对象的品牌对应的交互操作的操作时间信息，样本标注信息用于表征样本用户是否执行与对应的样本用户交互信息中交互对象的品牌关联的目标操作，所述训练样本集合中的各样本用户交互信息中的类目一致；
    基于所述训练样本集合的样本用户交互信息，生成对应的用户的样本交互特征；
    将所生成的用户的样本交互特征作为输入，将与输入的样本交互特征对应的样本标注信息作为期望输出，训练得到与样本用户交互信息中的交互对象的类目对应的预设操作概率生成子模型。
  17. 根据权利要求11所述的装置，其中，所述用户交互信息通过如下方式获取：
    响应于确定所述用户交互信息中不包含用户标识，从所述用户交互信息中提取出终端设备标识；
    获取所述终端设备标识关联的至少一个候选用户标识，将所述用户交互信息关联至至少一个所述候选用户标识。
  18. 根据权利要求13所述的装置，其中，所述第一生成单元还包括：
    划分模块，被配置成根据用户交互信息中的交互对象的类目信息所指示的类目所 属的类目组，将所述用户交互信息集合中的用户交互信息划分为至少一个交互信息组，其中，所述类目组基于类目信息所指示的类目之间的相关性划分；
    确定模块，被配置成根据交互信息组中的用户交互信息，确定各交互信息组中的交互对象的类目种类数；
    第一合成模块，被配置成将各交互信息组中的类目种类数个用户的交互操作特征矩阵合成为各交互信息组对应的用户的新的交互操作特征矩阵；
    第二合成模块，被配置成将各交互信息组中的类目种类数个用户属性特征向量合成为各交互信息组对应的用户属性特征矩阵。
  19. 根据权利要求18所述的装置，其中，所述预设操作概率生成模型与类目组对应；
    所述确定单元包括：
    第五生成模块，被配置成根据所合成的新的交互操作特征矩阵对应的类目组，将所述新的交互操作特征矩阵输入至对应的预设操作概率生成模型中的长短期记忆网络，生成新的第一隐含特征；
    第六生成模块，被配置成将所合成的用户属性特征矩阵输入至所述对应的预设操作概率生成模型中的第一全连接网络，生成新的第二隐含特征；
    第七生成模块，被配置成将所生成的类目特征向量和品牌特征向量输入至所述对应的预设操作概率生成模型中的第二全连接网络，生成第三隐含特征；
    第八生成模块，被配置成将所生成的新的第一隐含特征、新的第二隐含特征和第三隐含特征输入至所述对应的预设操作概率生成模型中的第三全连接网络，生成与输入的用户的交互特征对应的用户执行与用户交互信息中的交互对象的品牌关联的目标操作的概率。
  20. 根据权利要求11-19任一所述的装置，其中，所述装置还包括：
    第二生成单元，被配置成根据所生成的用户执行与对应的用户交互信息中交互对象的品牌关联的目标操作的概率，生成所述交互对象对应的库存调整信息并推送。
  21. 一种电子设备，包括：
    一个或多个处理器；
    存储装置，其上存储有一个或多个程序；
    当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如权利要求1-10中任一所述的方法。
  22. 一种计算机可读介质，其上存储有计算机程序，其中，该程序被处理器执行 时实现如权利要求1-10中任一所述的方法。
PCT/CN2020/081317 2019-04-15 2020-03-26 用于处理用户交互信息的方法和装置 WO2020211616A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/604,283 US20220198487A1 (en) 2019-04-15 2020-03-26 Method and device for processing user interaction information

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910300614.4 2019-04-15
CN201910300614.4A CN111768218A (zh) 2019-04-15 2019-04-15 用于处理用户交互信息的方法和装置

Publications (1)

Publication Number Publication Date
WO2020211616A1 true WO2020211616A1 (zh) 2020-10-22

Family

ID=72718318

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/081317 WO2020211616A1 (zh) 2019-04-15 2020-03-26 用于处理用户交互信息的方法和装置

Country Status (3)

Country Link
US (1) US20220198487A1 (zh)
CN (1) CN111768218A (zh)
WO (1) WO2020211616A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11704374B2 (en) * 2021-01-30 2023-07-18 Walmart Apollo, Llc Systems and methods for personalizing search engine recall and ranking using machine learning techniques
CN112948238B (zh) * 2021-02-01 2023-05-02 成都信息工程大学 推荐系统的多样性的量化方法

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105678578A (zh) * 2016-01-05 2016-06-15 重庆邮电大学 一种基于网购行为数据的用户品牌偏好度量方法
CN106779074A (zh) * 2017-01-22 2017-05-31 腾云天宇科技（北京）有限公司 一种商场品牌组合预测方法及预测服务器
US20180232774A1 (en) * 2017-02-16 2018-08-16 International Business Machines Corporation Contextual relevance brand promotion
CN108711075A (zh) * 2018-05-22 2018-10-26 阿里巴巴集团控股有限公司 一种产品推荐方法和装置
CN109190808A (zh) * 2018-08-15 2019-01-11 拍拍信数据服务（上海）有限公司 用户行为预测方法、装置、设备及介质
CN109272373A (zh) * 2018-08-02 2019-01-25 阿里巴巴集团控股有限公司 一种基于计算机的品牌推荐方法
CN109509056A (zh) * 2018-10-16 2019-03-22 平安科技（深圳）有限公司 基于对抗网络的商品推荐方法、电子装置及存储介质

Family Cites Families (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7996264B2 (en) * 2000-05-15 2011-08-09 Avatizing, Llc System and method for consumer-selected advertising and branding in interactive media
US8301482B2 (en) * 2003-08-25 2012-10-30 Tom Reynolds Determining strategies for increasing loyalty of a population to an entity
US20090006156A1 (en) * 2007-01-26 2009-01-01 Herbert Dennis Hunt Associating a granting matrix with an analytic platform
US20060277103A1 (en) * 2005-01-26 2006-12-07 Magee, Llc Systems and methods for personalized product promotion
US20080270363A1 (en) * 2007-01-26 2008-10-30 Herbert Dennis Hunt Cluster processing of a core information matrix
CN101329674A (zh) * 2007-06-18 2008-12-24 北京搜狗科技发展有限公司 一种提供个性化搜索的系统和方法
CN101206752A (zh) * 2007-12-25 2008-06-25 北京科文书业信息技术有限公司 电子商务网站相关商品推荐系统及其方法
CN101408964B (zh) * 2008-11-25 2016-03-30 阿里巴巴集团控股有限公司 电子商务网站的前台类目调整方法及装置
US10269021B2 (en) * 2009-04-20 2019-04-23 4-Tell, Inc. More improvements in recommendation systems
US20100268661A1 (en) * 2009-04-20 2010-10-21 4-Tell, Inc Recommendation Systems
CN102479366A (zh) * 2010-11-25 2012-05-30 阿里巴巴集团控股有限公司 一种商品推荐方法及系统
US20130197968A1 (en) * 2011-09-24 2013-08-01 Elwha LLC, a limited liability corporation of the State of Delaware Behavioral fingerprinting with retail monitoring
CN102426686A (zh) * 2011-09-29 2012-04-25 南京大学 一种基于矩阵分解的互联网信息产品推荐方法
CN104809626A (zh) * 2015-03-17 2015-07-29 徐邑江 一种基于用户信用评估的个性化商品推荐方法
CN104820879A (zh) * 2015-05-27 2015-08-05 北京京东尚科信息技术有限公司 一种用户行为信息的分析方法和装置
CN106327227A (zh) * 2015-06-19 2017-01-11 北京航天在线网络科技有限公司 一种信息推荐系统及信息推荐方法
US10210453B2 (en) * 2015-08-17 2019-02-19 Adobe Inc. Behavioral prediction for targeted end users
US10089675B1 (en) * 2015-10-20 2018-10-02 Amazon Technologies, Inc. Probabilistic matrix factorization system based on personas
US10157351B1 (en) * 2015-10-20 2018-12-18 Amazon Technologies, Inc. Persona based data mining system
CN105447724B (zh) * 2015-12-15 2022-04-05 腾讯科技（深圳）有限公司 内容项推荐方法及装置
CN108153791B (zh) * 2016-12-02 2023-04-25 阿里巴巴集团控股有限公司 一种资源推荐方法和相关装置
CN106779985A (zh) * 2017-02-24 2017-05-31 武汉奇米网络科技有限公司 一种个性化商品排序的方法及系统
CN107105031A (zh) * 2017-04-20 2017-08-29 北京京东尚科信息技术有限公司 信息推送方法和装置
CN107944913B (zh) * 2017-11-21 2022-03-22 重庆邮电大学 基于大数据用户行为分析的高潜在用户购买意向预测方法
US20190205736A1 (en) * 2017-12-29 2019-07-04 Intel Corporation Compute optimization mechanism for deep neural networks
CN108053295A (zh) * 2017-12-29 2018-05-18 广州品唯软件有限公司 一种商品品牌排序的方法和装置
US20190205905A1 (en) * 2017-12-31 2019-07-04 OneMarket Network LLC Machine Learning-Based Systems and Methods of Determining User Intent Propensity from Binned Time Series Data
CN109509054B (zh) * 2018-09-30 2023-04-07 平安科技（深圳）有限公司 海量数据下商品推荐方法、电子装置及存储介质
CN109492687A (zh) * 2018-10-31 2019-03-19 北京字节跳动网络技术有限公司 用于处理信息的方法和装置
CN109495552A (zh) * 2018-10-31 2019-03-19 北京字节跳动网络技术有限公司 用于更新点击率预测模型的方法和装置
CN109460513B (zh) * 2018-10-31 2021-01-08 北京字节跳动网络技术有限公司 用于生成点击率预测模型的方法和装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105678578A (zh) * 2016-01-05 2016-06-15 重庆邮电大学 一种基于网购行为数据的用户品牌偏好度量方法
CN106779074A (zh) * 2017-01-22 2017-05-31 腾云天宇科技（北京）有限公司 一种商场品牌组合预测方法及预测服务器
US20180232774A1 (en) * 2017-02-16 2018-08-16 International Business Machines Corporation Contextual relevance brand promotion
CN108711075A (zh) * 2018-05-22 2018-10-26 阿里巴巴集团控股有限公司 一种产品推荐方法和装置
CN109272373A (zh) * 2018-08-02 2019-01-25 阿里巴巴集团控股有限公司 一种基于计算机的品牌推荐方法
CN109190808A (zh) * 2018-08-15 2019-01-11 拍拍信数据服务（上海）有限公司 用户行为预测方法、装置、设备及介质
CN109509056A (zh) * 2018-10-16 2019-03-22 平安科技（深圳）有限公司 基于对抗网络的商品推荐方法、电子装置及存储介质

Also Published As

Publication number Publication date
US20220198487A1 (en) 2022-06-23
CN111768218A (zh) 2020-10-13

Similar Documents

Publication Publication Date Title
CN109460513B (zh) 用于生成点击率预测模型的方法和装置
US10650311B2 (en) Suggesting resources using context hashing
WO2018192491A1 (zh) 信息推送方法和装置
CN109783730A (zh) 产品推荐方法、装置、计算机设备和存储介质
CN109460514A (zh) 用于推送信息的方法和装置
CN111581926B (zh) 文案生成方法、装置、设备和计算机可读存储介质
EP4242955A1 (en) User profile-based object recommendation method and device
CN113239275B (zh) 信息推送方法、装置、电子设备和存储介质
US11677693B2 (en) Systems and methods for generating dynamic conversational responses using ensemble prediction based on a plurality of machine learning models
WO2020211616A1 (zh) 用于处理用户交互信息的方法和装置
CN116894711A (zh) 商品推荐理由生成方法及其装置、电子设备
CN113761348A (zh) 一种信息推荐方法、装置、电子设备和存储介质
CN113379449A (zh) 多媒体资源召回方法、装置、电子设备以及存储介质
CN111209351A (zh) 对象关系预测、对象推荐方法及装置、电子设备、介质
CN110555747A (zh) 确定目标用户的方法和装置
CN117057855A (zh) 一种数据处理方法及相关装置
CN116186541A (zh) 一种推荐模型的训练方法及装置
CN111797622B (zh) 用于生成属性信息的方法和装置
CN113792952A (zh) 用于生成模型的方法和装置
CN113743973A (zh) 分析市场热点趋势的方法和装置
CN117633165B (zh) 一种智能ai客服对话引导方法
US20240152512A1 (en) Machine learning for dynamic information retrieval in a cold start setting
CN117611282A (zh) 训练点击率预测模型和点击率预测的方法和装置
CN117745382A (zh) 对象推荐方法和装置
CN118013117A (zh) 对象推荐方法、装置、电子设备及可读存储介质

Legal Events

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

Ref document number: 20791607

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: 20791607

Country of ref document: EP

Kind code of ref document: A1