CN112215629B - Multi-target advertisement generating system and method based on construction countermeasure sample - Google Patents
Multi-target advertisement generating system and method based on construction countermeasure sample Download PDFInfo
- Publication number
- CN112215629B CN112215629B CN201910615860.9A CN201910615860A CN112215629B CN 112215629 B CN112215629 B CN 112215629B CN 201910615860 A CN201910615860 A CN 201910615860A CN 112215629 B CN112215629 B CN 112215629B
- Authority
- CN
- China
- Prior art keywords
- sample set
- layer
- type
- model
- request identifier
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0277—Online advertisement
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
Abstract
The invention provides a multi-target advertisement generating system and a method based on construction countermeasure samples, and belongs to the technical field of countermeasure neural networks. The method comprises the following steps: the method comprises the following steps: acquiring a first type sample set and a second type sample set related to user behaviors, and acquiring a countermeasure sample set with relevant characteristics of the first type sample set and the second type sample set; constructing a recall model to be trained, training the recall model to be trained by using the first type sample set, the second type sample set and the countermeasure sample set to obtain a recall model with a three-classification function, and combining the recall model with a preset filtering rule to form a generation model. The invention can ensure the relativity of the advertisement and the search request and also can consider the click preference of the user on the advertisement.
Description
Technical Field
The present invention relates to the technical field of antagonistic neural networks, in particular to a method of constructing a generative model, a method of recalling target results, a system for generating a model, a generative model, an apparatus for producing a model and a computer readable storage medium.
Background
Advertisement matching in a search scene is a typical retrieval problem and can be abstracted into three stages of generation, coarse ordering and fine ordering. The generation stage retrieves candidate advertisement sets with good matching (matching of the user searching intention and the advertisement service) from the advertisement library according to the user searching information; and in the sorting stage, candidate advertisements are comprehensively scored and truncated according to estimated click rate (ctr 2) of the advertisements, information of search income per thousand times (cpm 1) and the like, so that a final advertisement showing set is obtained.
In the traditional scheme, the generation stage mainly considers advertiser delivery rules (advertisement delivery time and region) and semantic matching degree of advertisement service and user search words, the sequencing stage predicts satisfaction degree and predicted click rate of users on advertisements according to the introduced user personalized information, and the traditional scheme has inconsistent upstream and downstream optimization targets and is a compromise design under the historic calculation performance and calculation frame background.
In the traditional scheme, the semantic matching degree of advertisement information and user search words is mainly considered in the advertisement generation stage, and the advertisement owner release rule (advertisement release time and region) comprises two modes:
1) Word-based matching: when an advertiser carries out advertisement, the advertisement business content can be described through a short sentence, and the advertisement business content can be called advertisement buying information (for example, the search request is "summer ultraviolet-proof sunglasses", and the buying information can be "sunglasses" or "ultraviolet-proof glasses"); in the conventional method, firstly, word segmentation is carried out on a user search request query to obtain a word set A; word segmentation is carried out on the advertisement buying words to obtain a word set B; screening target display advertisements according to the number and length of the intersection words of the word sets A and B as quantization indexes; in addition, synonym rewriting is performed on the word sets A and B through a technical means, so that the probability of intersection generation of the two sets is increased, and more advertisement candidates are obtained.
2) Model-based matching:
semantic relevance between search request query and advertisement buying words is quantified through a supervised learning model. And manually labeling the correlation between the search request query and the advertisement buying word, and learning the labeling data through a machine learning model, so as to predict the correlation of any one search query and any pair of buying word. The scheme modeling process mainly analyzes and mines lexical, grammatical and semantic information of query and buying word information through natural language processing means, and quantifies the correlation between two character strings. At present, the traditional machine learning model is modeled in a full-connection mode, and neurons are connected completely.
In the above conventional scheme:
1) Word-based matching: according to the scheme, the matching degree between the search request query and the advertisement is quantified based on the rule matching, the generalization and recall capability is limited, so that a large number of advertisements which are semantically related to the target search request query but are not literally related cannot be recalled, and the rendering capability of an advertisement system is limited;
2) Model-based matching: compared with 1, the scheme has a larger improvement on efficiency, but the quantitative target of the method is only the semantic relevance of the search request query and the target advertisement, personalized information such as the clicked probability of the advertisement and the preference of the user for the advertisement cannot be considered, and the method can lead to that although the relevance of the advertisement recalled by the system and the search query of the user is good, the advertisement is filtered out in the downstream sorting stage of the advertisement system due to factors such as low click rate, and the integral efficiency of the advertisement system is limited. In addition, the fully-connected modeling mode needs to establish connection between every two adjacent nerve cells of the nerve layer, a large amount of calculation is needed, the online calculation efficiency is low, and a large performance challenge is brought to an advertisement online system.
In a combined view, although the two modes have different implementation manners, the advertisement is generated around the single target of the relevance, so that the relevance of the advertisement candidate set and the search query can be guaranteed, but the estimated click rate of the advertisement and the personal preference of the user to the advertisement are not considered, so that the advertisement has reduced competitiveness in the downstream sequencing stage of the advertisement system (the sequencing score score=the estimated click rate ctr2×the advertisement bid), and the generated advertisement is filtered in the downstream flow of the system processing.
Disclosure of Invention
It is an object of the present invention to provide a multi-targeted advertisement generation system based on construction of a countermeasure sample and a method thereof.
To achieve the above object, an embodiment of the present invention provides a method of constructing a generative model, the method including:
s1) acquiring a first type sample set and a second type sample set related to user behaviors, and acquiring a countermeasure sample set with relevant characteristics of the first type sample set and the second type sample set;
s2) constructing a recall model to be trained, training the recall model to be trained by using the first type sample set, the second type sample set and the countermeasure sample set to obtain a recall model with a three-classification function, and combining the recall model with a preset filtering rule to form a generation model.
Specifically, step S1) includes, before obtaining a challenge sample set having relevant characteristics of the first type sample set and the second type sample set:
s101) acquiring a sample set, and labeling a request identifier and a target result which have a corresponding relation in the sample set according to user behaviors;
s102) dividing samples in the sample set with the user behavior being in a first trigger state into a first type of sample set and dividing samples in the sample set with the user behavior being in a second trigger state into a second type of sample set, wherein the first type of sample set is provided with a first type request identifier and a first type target result, and the second type of sample set is provided with a second type request identifier and a second type target result.
Specifically, step S1) of obtaining a challenge sample set having relevant characteristics of the first type sample set and the second type sample set includes:
s103) constructing a first trigger event according to the first type request identifier and the second type target result, and constructing a second trigger event according to the second type request identifier and the first type target result, estimating a first probability of the first trigger event or estimating a second probability of the second trigger event, wherein,
the first trigger event includes a first condition and a trigger object when the first condition is satisfied, the second trigger event includes a second condition and a trigger object when the second condition is satisfied,
the first condition is that a currently selected request identifier in the first type request identifier and a currently selected target result in the second type target result form the corresponding relation,
the second condition is that the request identifier currently selected from the second type request identifiers and the target result currently selected from the first type target results form the corresponding relationship,
the trigger object when the first condition is satisfied is that the user behavior is the first trigger state when the first condition is satisfied,
The triggering object when the second condition is satisfied is that the user behavior is the first triggering state when the second condition is satisfied;
s104) acquiring the current countermeasure intensity by utilizing the first probability or the second probability in the step S103) and combining a preset weighted sampling mapping relation;
s105) according to the magnitude of the current countermeasure intensity, selectively combining a request identifier and a target result corresponding to the first probability or the second probability with a preset correlation mapping relation to obtain a current correlation predicted value, marking the request identifier and the target result corresponding to the first probability or the second probability as countermeasure samples when the current correlation predicted value meets a preset correlation threshold condition, and returning to the step S103), wherein all the countermeasure samples form a countermeasure sample set.
Specifically, the preset correlation mapping relationship in step S105) is configured to:
performing vector inner product calculation by using the word vector of the request identifier in step S103) and the word vector of the target result; or alternatively, the process may be performed,
predicting by a preset deep learning model using the feature vector of the request identifier and the feature vector of the target result in step S103).
Specifically, in step S2), a recall model to be trained is constructed, including:
s201), a first input layer, a first hidden layer and a first full connection layer for acquiring a characteristic vector of a current request identifier are formed;
s202) forming a second input layer, a second hidden layer and a second full-connection layer for acquiring a characteristic vector of a current target result, wherein the layer structures of the second input layer and the second hidden layer are independent relative to the layer structures of the first input layer and the first hidden layer, the first full-connection layer and the second full-connection layer are provided with a shared layer, and the first full-connection layer and the second full-connection layer are subjected to cross calculation in the shared layer;
s203) forming an output layer, wherein the output layer receives the result of the cross-computation in step S202) and outputs a current classification result belonging to one of three classifications corresponding to the first type sample set, the second type sample set, and the challenge sample;
s204) configuring initial parameters of each layer of neural network in the first input layer, the first hidden layer, the first full-connection layer, the second input layer, the second hidden layer, the second full-connection layer and the output layer to form a recall model to be trained.
Specifically, in step S202), the first input layer is configured to obtain a vector of the current request identifier according to the user feature classification situation and the request feature classification situation;
the second input layer is configured to obtain a vector of a current target result according to the target result feature classification.
Specifically, in step S202), the total number of layers of the second hidden layer is the same as the total number of layers of the first hidden layer;
the first hidden layer is configured to obtain a current request identifier feature vector corresponding to the vector of current request identifiers;
the second hidden layer is configured to obtain a current target result feature vector corresponding to the vector of the current target result.
Specifically, in step S2), a recall model to be trained is constructed, and further includes:
s205) selecting and configuring a relevance verification model, combining a current request identifier feature vector and a target result feature vector which are associated with the current classification result with the relevance verification model to obtain a current relevance verification value, and describing the training degree of the recall model to be trained by using the current relevance verification value.
Specifically, the generative model in step S2) has the following functions:
And combining the current classification result output by the recall model with a preset filtering rule, and selectively generating a request identifier and a target result with a correlation according to the current request identifier feature vector and the current target result feature vector.
The embodiment of the invention provides a method for recalling a target result, which comprises the following steps:
s1) acquiring a request identifier and acquiring a target result to be recalled, wherein the target result to be recalled has a corresponding relation with the request identifier;
s2) extracting the feature vector of the request identifier and the feature vector of the target result to be recalled by using a pre-trained recall model, classifying the target result to be recalled, obtaining a classification result after the classification is finished, and selectively recalling the target result in the target result to be recalled according to the relation between the classification result and a preset filtering rule.
An embodiment of the present invention provides a system for generating a model, the system including:
a processing engine for performing the aforementioned method.
The embodiment of the invention provides a generation model, which comprises the following steps:
a recall model having a multi-classification function for obtaining a request identifier, for obtaining a classification result according to the request identifier and the multi-classification function, and for recalling a first set of target results corresponding to the request identifier, wherein the multi-classification function is a three-class classification function and at least one-class classification function is used for determining a correlation of the request identifier and target results within the first set of target results;
The preset filtering model is provided with preset filtering rules;
the preset filtering model is used for screening the first target result set according to the classification result and the preset filtering rule and generating a second target result set for returning corresponding to the request identifier after screening is completed.
In yet another aspect, an embodiment of the present invention provides an apparatus for producing a model, comprising:
at least one processor;
a memory coupled to the at least one processor;
wherein the memory stores instructions executable by the at least one processor, the at least one processor implementing the aforementioned methods by executing the memory-stored instructions.
In yet another aspect, embodiments of the present invention provide a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the foregoing method.
In yet another aspect, embodiments of the present invention also provide a computer program product comprising a computer program which, when executed by a processor, implements the aforementioned method.
Corresponding to the above, the invention constructs the countermeasures describing the correlation by obtaining the samples of the user behaviors (such as clicking and non-clicking), so that the recalled and classified target results can be fully used for improving the pruning efficiency of the generation stage and the funnel efficiency of the whole system;
according to the invention, through the pre-estimated probability and the threshold value and combining with the preset weighted sampling relation, the sample which can be used for judging the relevant characteristics in the two types of samples can be selected, and then the selected sample is combined with the preset correlation mapping relation to judge the relevant characteristics, so that whether the sample can be used as an countermeasure sample or not can be obtained, and an countermeasure sample construction means with relevant characteristics is provided, so that the triggering state of the user behavior and the relevant characteristics under the triggering state are considered, and the error filtering in the generation stage is remarkably reduced, and the filtering efficiency is improved.
Additional features and advantages of embodiments of the invention will be set forth in the detailed description which follows.
Drawings
The accompanying drawings are included to provide a further understanding of embodiments of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain, without limitation, the embodiments of the invention. In the drawings:
FIG. 1 is a schematic diagram of a main system framework according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a correlation judgment model in an authentication model according to an embodiment of the present invention;
fig. 3 is a schematic diagram of a correlation verification model according to an embodiment of the present invention.
Detailed Description
The following describes the detailed implementation of the embodiments of the present invention with reference to the drawings. It should be understood that the detailed description and specific examples, while indicating and illustrating the invention, are not intended to limit the invention.
Example 1
Referring to fig. 1, an embodiment of the present invention provides a method for constructing a generative model, the method including:
s1) acquiring a first type sample set and a second type sample set related to user behaviors, and acquiring a countermeasure sample set with relevant characteristics of the first type sample set and the second type sample set;
s2) constructing a recall model to be trained, training the recall model to be trained by using the first type sample set, the second type sample set and the countermeasure sample set to obtain a recall model with a three-classification function, and combining the recall model with a preset filtering rule to form a generation model;
specifically, step S1) includes, before obtaining a challenge sample set having relevant characteristics of the first type sample set and the second type sample set:
S101) acquiring a sample set, and labeling a request identifier and a target result which have a corresponding relation in the sample set according to user behaviors;
s102) dividing samples in the sample set with the user behavior being in a first trigger state into a first type of sample set and dividing samples in the sample set with the user behavior being in a second trigger state into a second type of sample set, wherein the first type of sample set is provided with a first type request identifier and a first type target result, and the second type of sample set is provided with a second type request identifier and a second type target result;
the user may be a user who issues a search request; the user behavior is a click behavior, for example, the user clicking on an advertisement in a searched target result (advertisement); the first trigger state may be a state in which a trigger click occurs; the first class of sample sets (clicks) may be clicked class of sample sets (class 1 of class identifier), the second class of sample sets (non-clicks) may be non-clicked class of sample sets (class 0 of class identifier), and the countersample sets may be clicked but poorly correlated sample sets (class 2 of class identifier) according to the correlation characteristics; the request identifier may be an identifier classified by a character string case corresponding to a search request of the user (with respect to a search request query performed by the user) or an identifier classified by a user characteristic case performing the current search (with respect to a user usr performing the search request), and the target result may be advertisements, each of which has a unique identification code, advertisement content, and metadata such as an advertisement title, an advertisement landing page, and an advertisement home merchant identification code.
Specifically, step S1) of obtaining a challenge sample set having relevant characteristics of the first type sample set and the second type sample set includes:
s103) constructing a first trigger event according to the first type request identifier and the second type target result, constructing a second trigger event according to the second type request identifier and the first type target result, estimating a first probability of the first trigger event or estimating a second probability of the second trigger event, wherein the first trigger event comprises a first condition and a trigger object when the first condition is met, the second trigger event comprises a second condition and a trigger object when the second condition is met, the first condition is that a currently selected request identifier in the first type request identifier and a currently selected target result in the second type target result form the corresponding relation, the second condition is that a currently selected request identifier in the second type request identifier and a currently selected target result in the first type target result form the corresponding relation, the trigger object when the first condition is met is that the user behavior is the first trigger state, and the second condition is that the user behavior is the second trigger state when the first condition is met;
S104) acquiring the current countermeasure intensity by utilizing the first probability or the second probability in the step S103) and combining a preset weighted sampling mapping relation;
s105) according to the magnitude of the current countermeasure intensity, selectively combining a request identifier and a target result corresponding to the first probability or the second probability with a preset correlation mapping relation to obtain a current correlation predicted value, marking the request identifier and the target result corresponding to the first probability or the second probability as countermeasure samples when the current correlation predicted value meets a preset correlation threshold condition, returning to the step S103), and forming all the countermeasure samples into a countermeasure sample set (Generative adversial sample);
the first probability and the second probability can be estimated click rate, the estimated click rate can be obtained through an estimation model, and the estimation model can be built in the recall model and trained together with the recall model; randomly selecting a request identifier and a target result from the first type sample set and the second type sample set; the step S104) and the step S105) may be performed by an anti-sampling process (Adversial sampling), for example, constructing an authentication Model (Judge Model) having a preset weighted sampling mapping relationship and a preset correlation mapping relationship, and performing the step S104) and the step S105 by the authentication Model), where the selected sample of the authenticated Model may be called as a displayed Data, and in some embodiments, the selected sample may be updated and added and deleted, for example, the selected sample may be added with the correlation sample according to a synonym rewriting manner; specifically, the mapping relationship among the current challenge strength sample_sample, the estimated click rate ECR and the preset weighted sampling is as follows:
Wherein, alpha is a weighting coefficient used for configuration and bias is a bias parameter used for configuration, the alpha and bias determine the window size of weighted sampling, and the configuration process of alpha and bias at least needs to ensure (alpha is ECR+bias) to be positive or zero;
specifically, the preset correlation mapping relationship in step S105) is configured to:
performing vector inner product calculation by using the word vector of the request identifier in step S103) and the word vector of the target result; or alternatively, the process may be performed,
predicting by using the feature vector of the request identifier in the step S103) and the feature vector of the target result through a preset deep learning model;
as shown in FIG. 2, the preset relevance mapping relationship may be a relevance judgment model, such as a pre-trained deep neural network model (which may be used in an offline training process) or a model constructed according to the inner product of word vectors (which may be used in an online use process), and the relevance judgment model may include mutually independent request feature extraction layer structures<query_input>,<query_hidden layer a,b>(a＝1,2,3,4,5)]And advertisement feature extractionLayer structure [<ad_input>,<Ad_hidden layer c,b>(c＝1,2,3,4,5)]The request feature extraction layer structure may have a request input layer<query_input>And request feature extraction hidden layer<query_hidden layer a,b>The advertisement feature extraction layer structure can be provided with an advertisement input layer<ad_input>And advertisement feature extraction hidden layer <Ad_hidden layer c,b>B is offset and also comprises a network computing layerThe network computing layer performs matrix inner product computation to obtain a correlation prediction value (release) and a correlation judgment model forms a double-tower layer structure; specifically, a sample (a request identifier and a target result corresponding to the first probability or the second probability, namely a search request and an advertisement corresponding to the first probability or the second probability) corresponding to the current countermeasure intensity meeting an intensity threshold condition is selected, and in the offline training process, a search request feature vector and an advertisement feature vector are extracted to be scored through a correlation judgment model, and the sample lower than a preset correlation threshold value is marked as 2 types; considering the online use process, the word vector of the sample can be obtained first, the Euclidean distance of the word vector is calculated, the correlation is judged to be poor when the Euclidean distance is larger than a preset distance threshold value, and the correlation is judged to be good when the Euclidean distance is smaller than or equal to the preset distance threshold value; in some implementations, the search request in the selected sample can be associated with the buying information corresponding to the advertisement, and the correlation between the search request and the buying information can be judged, so that the accuracy of the correlation judgment of the countersample can be increased.
Specifically, in step S2), a recall model to be trained is constructed, including:
s201) forming a first input layer, a first hidden layer and a first fully connected layer for obtaining a feature vector of a current request identifier, wherein the vectors include word vectors and feature vectors having a certain dimension;
s202) forming a second input layer, a second hidden layer and a second full-connection layer for acquiring a characteristic vector of a current target result, wherein the layer structures of the second input layer and the second hidden layer are independent relative to the layer structures of the first input layer and the first hidden layer, the first full-connection layer and the second full-connection layer are provided with a shared layer, and the first full-connection layer and the second full-connection layer are subjected to cross calculation in the shared layer;
s203) forming an output layer, wherein the output layer receives the result of the cross-computation in step S202) and outputs a current classification result belonging to one of three classifications corresponding to the first type sample set, the second type sample set, and the challenge sample;
s204) configuring initial parameters of each layer of neural network in the first input layer, the first hidden layer, the first full-connection layer, the second input layer, the second hidden layer, the second full-connection layer and the output layer to form a recall model to be trained;
The input layer of the Recall Model (Recall Model) extracts information from both sides, a user side (usr_input) and an advertisement side (ad_input); the three dimensional characteristics of the user characteristics, the search request characteristics and the advertisement characteristics are used, the user characteristics can comprise characteristics such as user browser information and user network name identification, the search request characteristics can comprise characteristics such as search string length and proper nouns, and the advertisement characteristics can comprise proper nouns in titles, buying word information and the like; extracting user characteristics and search request characteristics at the user side and extracting advertisement characteristics at the advertisement side at the input layer; independent layers (usr_hidden layer1-7 and ad_hidden layer 1-7) are used for feature extraction and feature vector formation, and the last hidden layer usr_hidden layer7 or ad_hidden layer7 also has offset (b) and position (position) for feature vector specific structure configuration; the initial parameters may be learning rate, sampling rate, and neuron core size; in the full connection layer, the feature vector usr_vec i and the feature vector ad_vec i in the full connection 1 perform cross computation (i=0, 1, 2) of the feature vector, for example, matrix inner product computation, to obtain a full connection 2, and the value usr_ad i exists in the full connection 2; the output layer can be a normalized exponential function, the classification result can be class 0, class 1 and class 2, the recall model (as shown on the right side of fig. 1) also forms a neural network model with a double-tower structure, the top of the tower is an output layer, the bottom of the tower is an input layer, and the middle of the tower is a hidden layer and a full-connection layer (full-connection 1 and full-connection 2) in sequence;
Jth neuron z in output layer of recall model j The presence is:
wherein K and K are positive integers,to describe the jth neuron z j Is a function of the exponent of (a).
The double-tower model reduces the number of connections between neurons, so that the calculation cost is reduced, and compared with a general full-connection model, the model effect is not reduced, but is greatly improved from the view point of the model training effect; the method is characterized in that under the model design of a double-tower structure, the characteristics in the user side tower and the advertisement side tower can be fully exchanged and learned respectively, the tower top obtains the high-order characteristics of the two main body towers, when the two tower top vectors are in full-connection operation, the high-order characteristics in two aspects can be crossed and learned more efficiently (the characteristic exchange is directly carried out in a full-connection mode relative to the low-order characteristics), so that the model learning effect is greatly improved compared with a general full-connection model.
Specifically, in step S202), the first input layer is configured to obtain a vector of the current request identifier according to the user feature classification situation and the request feature classification situation;
the second input layer is configured to obtain a vector of a current target result according to a target result feature classification situation, and according to a user feature classification situation, a request feature classification situation and a target result feature classification situation, the user feature can be classified by quantization, a search request of quantization and an advertisement classification by quantization (wherein, the feature classification of the advertisement is classified and defined on the basis of combining the user feature classification and the request feature classification), at least 20 features are constructed, and the specific feature classification is shown in table 1 (ID is an object identity identifier, URL is a link);
TABLE 1 characteristic classification Condition Table
Specifically, in step S202), the total number of layers of the second hidden layer is the same as the total number of layers of the first hidden layer, for example, hidden layers of 7 layers in total;
the first hidden layer is configured to obtain a current request identifier feature vector corresponding to the vector of current request identifiers;
the second hidden layer is configured to obtain a current target result feature vector corresponding to the vector of the current target result.
Specifically, in step S2), a recall model to be trained is constructed, and further includes:
s205) selecting and configuring a relevance verification model, combining a current request identifier feature vector and a target result feature vector which are related to the current classification result with the relevance verification model to obtain a current relevance verification value, and describing the training degree of the recall model to be trained by utilizing the current relevance verification value (score);
as shown in fig. 3, a correlation check model may be provided in a check flow after the recall model is outputted, and may be used to maintain consistency of the system as a whole and describe indirect correlation, and the correlation check model may include an independent request extraction layer structure [ < usr_input >, < usr_hidden layer m, b > (m=1, 2,3,4, 5) ] and a bid information extraction layer structure [ < bid_input >, < bid_hidden layer n, b > (n=1, 2,3,4, 5) ], and an output layer for calculating an output vector of the request extraction layer structure and an output vector of the bid information extraction layer structure, the bid information being obtained by correlating advertisements corresponding to the search request (after screening using the recall model classification result).
Specifically, the generative model in step S2) has the following functions:
combining the current classification result output by the recall model with a preset filtering rule, and selectively generating a request identifier and a target result with a correlation according to the current request identifier feature vector and the current target result feature vector;
the query-advertisement candidate set is directly generated, and compared with the traditional triggering mode based on advertisement keywords, the efficiency is remarkably improved;
the preset filtering rule can be set to reject advertisements corresponding to class 2 in the classification result, the preset filtering rule can be set in a preset filtering model, and the preset filtering model can also be a pre-trained neural network model.
When training is completed and the model is used online, the embodiment of the invention also provides a method for recalling the target result, which comprises the following steps:
s1) acquiring a request identifier and acquiring a target result to be recalled, wherein the target result to be recalled has a corresponding relation with the request identifier;
s2) extracting the feature vector of the request identifier and the feature vector of the target result to be recalled by using a pre-trained recall model, classifying the target result to be recalled, obtaining a classification result after the classification is finished, and selectively recalling the target result in the target result to be recalled according to the relation between the classification result and a preset filtering rule.
The embodiment of the invention also provides a system for generating a model, which comprises:
a processing engine for performing the aforementioned method.
The embodiment of the invention also provides a generation model, which comprises the following steps:
a recall model having a multi-classification function for obtaining a request identifier, for obtaining a classification result according to the request identifier and the multi-classification function, and for recalling a first set of target results corresponding to the request identifier, wherein the multi-classification function is a three-class classification function and at least one-class classification function is used for determining a correlation of the request identifier and target results within the first set of target results;
the preset filtering model is provided with preset filtering rules;
the preset filtering model is used for screening the first target result set according to the classification result and the preset filtering rule and generating a second target result set for returning corresponding to the request identifier after screening is completed.
The present invention provides systems and methods for constructing and implementing generative models; the invention can learn the relativity of the advertisement and the search request query and learn the estimated click rate of the candidate advertisement in the current search, thereby not only ensuring the relativity of the advertisement and the search query, but also considering the click preference of the user on the advertisement; the estimated click rate (ctr 2) is a very important attribute in commercial rendering, and the advertisement system integrally sorts advertisements according to ctr2 bid.
The foregoing details of the optional implementation of the embodiment of the present application have been described in detail with reference to the accompanying drawings, but the embodiment of the present application is not limited to the specific details of the foregoing implementation, and various simple modifications may be made to the technical solution of the embodiment of the present application within the scope of the technical concept of the embodiment of the present application, and these simple modifications all fall within the protection scope of the embodiment of the present application.
In addition, the specific features described in the above embodiments may be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, various possible combinations of embodiments of the present application are not described in detail.
Those skilled in the art will appreciate that all or part of the steps in implementing the methods of the embodiments described above may be implemented by a program stored in a storage medium, including instructions for causing a single-chip microcomputer, chip or processor (processor) to perform all or part of the steps of the methods of the embodiments described herein. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In addition, any combination of various embodiments of the present invention may be performed, so long as the concept of the embodiments of the present invention is not violated, and the disclosure of the embodiments of the present invention should also be considered.
Claims (11)
1. A method of constructing a generative model, the method comprising:
s1) acquiring a first type sample set and a second type sample set related to user behaviors, and acquiring a countermeasure sample set with relevant characteristics of the first type sample set and the second type sample set;
s2) constructing a recall model to be trained, training the recall model to be trained by using the first type sample set, the second type sample set and the countermeasure sample set to obtain a recall model with a three-classification function, and combining the recall model with a preset filtering rule to form a generation model;
step S1) comprises, before obtaining a challenge sample set having relevant characteristics of the first and second type sample sets:
s101) acquiring a sample set, and labeling a request identifier and a target result which have a corresponding relation in the sample set according to user behaviors;
s102) dividing samples in the sample set with the user behavior being in a first trigger state into a first type of sample set and dividing samples in the sample set with the user behavior being in a second trigger state into a second type of sample set, wherein the first type of sample set is provided with a first type request identifier and a first type target result, and the second type of sample set is provided with a second type request identifier and a second type target result;
Step S1) of obtaining a challenge sample set having relevant features of the first and second sample sets, comprising:
s103) constructing a first trigger event according to the first type request identifier and the second type target result, constructing a second trigger event according to the second type request identifier and the first type target result, and estimating a first probability of the first trigger event or estimating a second probability of the second trigger event;
s104) acquiring the current countermeasure intensity by utilizing the first probability or the second probability in the step S103) and combining a preset weighted sampling mapping relation;
s105) according to the magnitude of the current countermeasure intensity, selectively combining a request identifier and a target result corresponding to the first probability or the second probability with a preset correlation mapping relation to obtain a current correlation predicted value, marking the request identifier and the target result corresponding to the first probability or the second probability as countermeasure samples when the current correlation predicted value meets a preset correlation threshold condition, and returning to the step S103), and forming all the countermeasure samples into a countermeasure sample set;
Step S2) constructing a recall model to be trained, which comprises the following steps:
s201), a first input layer, a first hidden layer and a first full connection layer for acquiring a characteristic vector of a current request identifier are formed;
s202) forming a second input layer, a second hidden layer and a second full-connection layer for acquiring a characteristic vector of a current target result, wherein the layer structures of the second input layer and the second hidden layer are independent relative to the layer structures of the first input layer and the first hidden layer, the first full-connection layer and the second full-connection layer are provided with a shared layer, and the first full-connection layer and the second full-connection layer are subjected to cross calculation in the shared layer;
s203) forming an output layer, wherein the output layer receives the result of the cross-computation in step S202) and outputs a current classification result belonging to one of three classifications corresponding to the first type sample set, the second type sample set, and the challenge sample;
s204) configuring initial parameters of each layer of neural network in the first input layer, the first hidden layer, the first full-connection layer, the second input layer, the second hidden layer, the second full-connection layer and the output layer to form a recall model to be trained.
2. The method of constructing a generative model according to claim 1, wherein the preset correlation mapping relationship in step S105) is configured to:
performing vector inner product calculation by using the word vector of the request identifier in step S103) and the word vector of the target result; or alternatively, the process may be performed,
predicting by a preset deep learning model using the feature vector of the request identifier and the feature vector of the target result in step S103).
3. The method of constructing a generative model as claimed in claim 1, wherein in step S202), the first input layer is configured to obtain a vector of current request identifiers according to user feature classification cases and request feature classification cases;
the second input layer is configured to obtain a vector of a current target result according to the target result feature classification.
4. A method of constructing a generative model as claimed in claim 3, wherein in step S202), the total number of layers of the second hidden layer is the same as the total number of layers of the first hidden layer;
the first hidden layer is configured to obtain a current request identifier feature vector corresponding to the vector of current request identifiers;
The second hidden layer is configured to obtain a current target result feature vector corresponding to the vector of the current target result.
5. The method of constructing a generative model as claimed in claim 1, wherein constructing a recall model to be trained in step S2) further comprises:
s205) selecting and configuring a relevance verification model, combining a current request identifier feature vector and a target result feature vector which are associated with the current classification result with the relevance verification model to obtain a current relevance verification value, and describing the training degree of the recall model to be trained by using the current relevance verification value.
6. The method of constructing a generative model as claimed in claim 1, wherein the generative model in step S2) has the following functions:
and combining the current classification result output by the recall model with a preset filtering rule, and selectively generating a request identifier and a target result with a correlation according to the current request identifier feature vector and the current target result feature vector.
7. A method of recalling target results using a pre-trained recall model, the method comprising:
S1) acquiring a request identifier and acquiring a target result to be recalled, wherein the target result to be recalled has a corresponding relation with the request identifier;
s2) extracting the feature vector of the request identifier and the feature vector of the target result to be recalled by using a pre-trained recall model, classifying the target result to be recalled, obtaining a classification result after the classification is finished, and selectively recalling the target result in the target result to be recalled according to the relation between the classification result and a preset filtering rule;
the pre-trained recall model is obtained through the following steps:
s1) acquiring a first type sample set and a second type sample set related to user behaviors, and acquiring a countermeasure sample set with relevant characteristics of the first type sample set and the second type sample set; step S1) comprises, before obtaining a challenge sample set having relevant characteristics of the first and second type sample sets: s101) acquiring a sample set, and labeling a request identifier and a target result which have a corresponding relation in the sample set according to user behaviors; s102) dividing samples in the sample set with the user behavior being in a first trigger state into a first type of sample set and dividing samples in the sample set with the user behavior being in a second trigger state into a second type of sample set, wherein the first type of sample set is provided with a first type request identifier and a first type target result, and the second type of sample set is provided with a second type request identifier and a second type target result;
Step S1) of obtaining a challenge sample set having relevant features of the first and second sample sets, comprising:
s103) constructing a first trigger event according to the first type request identifier and the second type target result, constructing a second trigger event according to the second type request identifier and the first type target result, and estimating a first probability of the first trigger event or estimating a second probability of the second trigger event;
s104) acquiring the current countermeasure intensity by utilizing the first probability or the second probability in the step S103) and combining a preset weighted sampling mapping relation;
s105) according to the magnitude of the current countermeasure intensity, selectively combining a request identifier and a target result corresponding to the first probability or the second probability with a preset correlation mapping relation to obtain a current correlation predicted value, marking the request identifier and the target result corresponding to the first probability or the second probability as countermeasure samples when the current correlation predicted value meets a preset correlation threshold condition, and returning to the step S103), and forming all the countermeasure samples into a countermeasure sample set;
S2) constructing a recall model to be trained, and training the recall model to be trained by using the first type sample set, the second type sample set and the countermeasure sample set to obtain a recall model with a three-classification function; step S2) constructing a recall model to be trained, which comprises the following steps:
s201), a first input layer, a first hidden layer and a first full connection layer for acquiring a characteristic vector of a current request identifier are formed;
s202) forming a second input layer, a second hidden layer and a second full-connection layer for acquiring a characteristic vector of a current target result, wherein the layer structures of the second input layer and the second hidden layer are independent relative to the layer structures of the first input layer and the first hidden layer, the first full-connection layer and the second full-connection layer are provided with a shared layer, and the first full-connection layer and the second full-connection layer are subjected to cross calculation in the shared layer;
s203) forming an output layer, wherein the output layer receives the result of the cross-computation in step S202) and outputs a current classification result belonging to one of three classifications corresponding to the first type sample set, the second type sample set, and the challenge sample;
S204) configuring initial parameters of each layer of neural network in the first input layer, the first hidden layer, the first full-connection layer, the second input layer, the second hidden layer, the second full-connection layer and the output layer to form a recall model to be trained.
8. A system for generating a model, the system comprising:
a processing engine for performing the method of any of claims 1 to 7.
9. A generative model apparatus, wherein the generative model in the generative model apparatus is constructed by the method for constructing a generative model as claimed in any one of claims 1 to 6; the generating model comprises the following steps:
a recall model having a multi-classification function for obtaining a request identifier, for obtaining a classification result according to the request identifier and the multi-classification function, and for recalling a first set of target results corresponding to the request identifier, wherein the multi-classification function is a three-class classification function and at least one-class classification function is used for determining a correlation of the request identifier and target results within the first set of target results;
The preset filtering model is provided with preset filtering rules;
the preset filtering model is used for screening the first target result set according to the classification result and the preset filtering rule and generating a second target result set for returning corresponding to the request identifier after screening is completed.
10. An apparatus for producing a model, comprising:
at least one processor;
a memory coupled to the at least one processor;
wherein the memory stores instructions executable by the at least one processor, the at least one processor implementing the method of any one of claims 1 to 7 by executing the instructions stored by the memory.
11. A computer readable storage medium storing computer instructions which, when run on a computer, cause the computer to perform the method of any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910615860.9A CN112215629B (en) | 2019-07-09 | 2019-07-09 | Multi-target advertisement generating system and method based on construction countermeasure sample |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910615860.9A CN112215629B (en) | 2019-07-09 | 2019-07-09 | Multi-target advertisement generating system and method based on construction countermeasure sample |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112215629A CN112215629A (en) | 2021-01-12 |
| CN112215629B true CN112215629B (en) | 2023-09-01 |
Family
ID=74047367
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910615860.9A Active CN112215629B (en) | 2019-07-09 | 2019-07-09 | Multi-target advertisement generating system and method based on construction countermeasure sample |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112215629B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113627496A (en) * | 2021-07-27 | 2021-11-09 | 交控科技股份有限公司 | Method, device, electronic equipment and readable storage medium for predicting fault of turnout switch machine |
| CN113704623B (en) * | 2021-08-31 | 2024-04-16 | 平安银行股份有限公司 | Data recommendation method, device, equipment and storage medium |
| CN116757748B (en) * | 2023-08-14 | 2023-12-19 | 广州钛动科技股份有限公司 | Advertisement click prediction method based on random gradient attack |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273978A (en) * | 2017-05-25 | 2017-10-20 | 清华大学 | A kind of production of three models game resists the method for building up and device of network model |
| CN107766577A (en) * | 2017-11-15 | 2018-03-06 | 北京百度网讯科技有限公司 | A kind of public sentiment monitoring method, device, equipment and storage medium |
| CN109241319A (en) * | 2018-09-28 | 2019-01-18 | 百度在线网络技术（北京）有限公司 | A kind of picture retrieval method, device, server and storage medium |
| CN109766991A (en) * | 2019-01-14 | 2019-05-17 | 电子科技大学 | A kind of artificial intelligence optimization's system and method using antagonistic training |
| CN109800785A (en) * | 2018-12-12 | 2019-05-24 | 中国科学院信息工程研究所 | One kind is based on the relevant data classification method of expression and device certainly |
| CN109857845A (en) * | 2019-01-03 | 2019-06-07 | 北京奇艺世纪科技有限公司 | Model training and data retrieval method, device, terminal and computer readable storage medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10007866B2 (en) * | 2016-04-28 | 2018-06-26 | Microsoft Technology Licensing, Llc | Neural network image classifier |
-
2019
- 2019-07-09 CN CN201910615860.9A patent/CN112215629B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273978A (en) * | 2017-05-25 | 2017-10-20 | 清华大学 | A kind of production of three models game resists the method for building up and device of network model |
| CN107766577A (en) * | 2017-11-15 | 2018-03-06 | 北京百度网讯科技有限公司 | A kind of public sentiment monitoring method, device, equipment and storage medium |
| CN109241319A (en) * | 2018-09-28 | 2019-01-18 | 百度在线网络技术（北京）有限公司 | A kind of picture retrieval method, device, server and storage medium |
| CN109800785A (en) * | 2018-12-12 | 2019-05-24 | 中国科学院信息工程研究所 | One kind is based on the relevant data classification method of expression and device certainly |
| CN109857845A (en) * | 2019-01-03 | 2019-06-07 | 北京奇艺世纪科技有限公司 | Model training and data retrieval method, device, terminal and computer readable storage medium |
| CN109766991A (en) * | 2019-01-14 | 2019-05-17 | 电子科技大学 | A kind of artificial intelligence optimization's system and method using antagonistic training |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112215629A (en) | 2021-01-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598206B (en) | Text semantic recognition method and device, computer equipment and storage medium | |
| CN109493166B (en) | Construction method for task type dialogue system aiming at e-commerce shopping guide scene | |
| Xu et al. | Hierarchical emotion classification and emotion component analysis on Chinese micro-blog posts | |
| CN107992531A (en) | News personalization intelligent recommendation method and system based on deep learning | |
| CN112215629B (en) | Multi-target advertisement generating system and method based on construction countermeasure sample | |
| CN110287314B (en) | Long text reliability assessment method and system based on unsupervised clustering | |
| CN106294618A (en) | Searching method and device | |
| CN112989208B (en) | Information recommendation method and device, electronic equipment and storage medium | |
| CN111985228A (en) | Text keyword extraction method and device, computer equipment and storage medium | |
| CN109829154B (en) | Personality prediction method based on semantics, user equipment, storage medium and device | |
| CN113326374B (en) | Short text emotion classification method and system based on feature enhancement | |
| CN111783903A (en) | Text processing method, text model processing method and device and computer equipment | |
| CN111538846A (en) | Third-party library recommendation method based on mixed collaborative filtering | |
| CN112819024B (en) | Model processing method, user data processing method and device and computer equipment | |
| CN114077661A (en) | Information processing apparatus, information processing method, and computer readable medium | |
| CN114357151A (en) | Processing method, device and equipment of text category identification model and storage medium | |
| CN110598126B (en) | Cross-social network user identity recognition method based on behavior habits | |
| Khanday et al. | Nnpcov19: artificial neural network-based propaganda identification on social media in covid-19 era | |
| CN110910235A (en) | Method for detecting abnormal behavior in credit based on user relationship network | |
| CN114491079A (en) | Knowledge graph construction and query method, device, equipment and medium | |
| CN113486143A (en) | User portrait generation method based on multi-level text representation and model fusion | |
| CN112685656A (en) | Label recommendation method and electronic equipment | |
| CN116401368A (en) | Intention recognition method and system based on topic event analysis | |
| CN113095073B (en) | Corpus tag generation method and device, computer equipment and storage medium | |
| CN116150353A (en) | Training method for intention feature extraction model, intention recognition method and related device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |