CN114997927A

CN114997927A - Real-time bidding sorting method, system, storage medium and electronic device for improving advertisement conversion effect of digital mall

Info

Publication number: CN114997927A
Application number: CN202210740638.3A
Authority: CN
Inventors: 赵征
Original assignee: Beijing Quyun Wanwei Information Technology Co ltd
Current assignee: Beijing Quyun Wanwei Information Technology Co ltd
Priority date: 2022-06-28
Filing date: 2022-06-28
Publication date: 2022-09-02

Abstract

The invention discloses a real-time bidding sorting method, a real-time bidding sorting system, a storage medium and an electronic device for improving the advertisement conversion effect of a digital mall, wherein the method comprises the steps of judging whether the advertisement is a page turning request or not, and entering an advertisement sorting flow if the advertisement is not the page turning request; the advertisement sequencing process comprises the following steps: acquiring basic flow information; filtering related products according to search keywords and filtering conditions in the basic flow information to generate an initial advertisement queue; further filtering the ad queue according to targeting conditions; calculating the advertisement base bid in the queue; pushing the advertisement queue into a real-time reasoning service, and calculating the click rate and/or the conversion rate of the advertisement in real time through the model; reading the flow distribution coefficient, calculating the advertisement occurrence opportunities in the queue, forming final sequencing and outputting an advertisement queue; the sorted first ad is output and popped from the queue. The invention scientifically and orderly organizes algorithms and big data technology, refines data resources of the digital mall and can improve the conversion and drainage effect of advertisement resources in the station.

Description

Real-time bidding sorting method, system, storage medium and electronic device for improving advertisement conversion effect of digital mall

Technical Field

The invention relates to the technical field of Internet, in particular to a real-time bidding sorting method, a real-time bidding sorting system, a storage medium and electronic equipment for improving the advertisement conversion effect of a digital mall.

Background

When the digital mall platform is built, a set of drainage system is provided for a seller (shop), website flow (advertisement resources) in the mall is purchased through bidding service, and drainage is performed to a shop home page or a commodity detail page, so that the digital mall platform is used during conversion. At present, most of digital mall platform big data and algorithm technology are weak, and business data (such as user behaviors, user characteristics, commodity characteristics and the like) cannot be effectively guided in a combined mode.

Therefore, the invention is especially provided.

Disclosure of Invention

The invention aims to provide a real-time bidding sorting method, a real-time bidding sorting system, a storage medium and electronic equipment for improving the advertisement conversion effect of a digital mall, and solves the problem that the current digital mall cannot well introduce the real-time bidding technology of digital advertisements and maximize the flow value.

In order to solve the above problem, in a first aspect, an embodiment of the present invention provides a real-time bidding sorting method for improving an advertisement conversion effect of a digital mall, including:

judging whether the request is a page turning request or not, and if not, entering an advertisement sequencing process; the advertisement sequencing process comprises the following steps: acquiring basic flow information; filtering related products according to search keywords and filtering conditions in the basic flow information to generate an initial advertisement queue; further filtering the ad queue according to the targeting condition; calculating the advertisement base bid in the queue; pushing the advertisement queue into a real-time reasoning service, and calculating the click rate and/or the conversion rate of the advertisement in real time through the model; reading the flow distribution coefficient, calculating the advertisement occurrence opportunities in the queue, forming final sequencing and outputting an advertisement queue; the ordered first ad is output and popped from the queue.

Optionally, the real-time bid calculation engine invoked in the advertisement ranking process is integrated in the following manner: constructing an initial database; the initial database comprises a keyword vector library, a product vector library and a crowd vector library; constructing an unsupervised classification model; the unsupervised classification model comprises a keyword classification model, a product classification model and a crowd classification model; configuring a key algorithm; the key algorithms comprise keyword recommendation and/or matching algorithms, click rate and/or conversion rate estimation algorithms, flow distribution algorithms and price interval recommendation algorithms.

Optionally, in configuring the key algorithm, configuring an anti-cheating algorithm is further included, where the anti-cheating algorithm includes clearing the client data according to a general blacklist, and marking an abnormal click behavior according to a click behavior blacklist.

Optionally, an AB test is further used to verify the actual effect of the constructed model or algorithm strategy.

Optionally, the constructing an initial database and the constructing an unsupervised classification model include: firstly, a keyword vector library is established, and recommended keywords and an unsupervised keyword classification model are generated according to the keyword vector library; constructing a crowd vector library according to the crowd label; respectively constructing an unsupervised learning crowd classification model and a crowd attribute model according to a crowd vector library; constructing a product vector library based on an unsupervised learning population classification model and a population attribute model; constructing an unsupervised learning product classification model according to a product vector library; and finally, fusing according to the unsupervised keyword classification model, the unsupervised learning crowd classification model, the unsupervised learning product classification model and the crowd attribute model, thereby constructing a bidding click rate and/or conversion rate prediction model for returning a plurality of advertisement activities to predict the generation probability of the user.

Optionally, the configuration method of the price interval recommendation algorithm includes: acquiring at least one data of a product, a category and a keyword input by a user; selecting products or keywords corresponding to the orientation range; and (3) obtaining a product price interval in a corresponding condition, if data exist, returning a result, and if no data exist, then: acquiring a keyword price interval of a user product, if data exists, returning a result, and if no data exists, then: and obtaining a product classification price interval, if the data exists, returning a result, and if the data does not exist, returning to a default price interval.

Optionally, the configuration of the flow allocation algorithm includes: acquiring original data of a current advertisement space; calculating the current score according to the scoring standard; constructing a genetic algorithm chromosome; setting a genetic algorithm scoring standard; setting and obtaining an optimal chromosome after N generations of genetic algorithm evolution, calculating scores according to the obtained optimal chromosome, comparing the scores with original scores, and using the proportion with high scores as a final distribution proportion; the final allocation ratio is provided to the ad server for use.

In a second aspect, an embodiment of the present invention provides a real-time bid sorting system for improving advertisement conversion effect in a digital mall, including: the judging module is used for judging whether the request is a page turning request or not, and if not, entering an advertisement sequencing process; an advertisement sequencing flow module, configured to execute the advertisement sequencing flow, where the advertisement sequencing flow includes: acquiring basic flow information; filtering related products according to search keywords and filtering conditions in the basic flow information to generate an initial advertisement queue; further filtering the ad queue according to targeting conditions; calculating the advertisement base bid in the queue; pushing the advertisement queue into a real-time reasoning service, and calculating the click rate and/or the conversion rate of the advertisement in real time through the model; reading the flow distribution coefficient, calculating the advertisement occurrence opportunities in the queue, forming final sequencing and outputting an advertisement queue; and the output module is used for outputting the ordered first advertisements and popping up the advertisements from the queue.

In a third aspect, an embodiment of the present invention provides a storage medium, on which a computer program is stored, where the program is executed by a processor to implement the method described above.

In a fourth aspect, an embodiment of the present invention provides an electronic device, where the electronic device includes: one or more processors; and a storage device for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to implement the above-described method.

The real-time bidding sorting method, the device, the storage medium and the electronic equipment for improving the advertisement conversion effect of the digital mall, provided by the embodiment of the invention, have the advantages that the scientific and ordered organization algorithm and the big data technology are adopted, the data resources of the electronic mall are refined, the conversion and drainage effects of the advertisement resources in the station can be greatly improved, namely, the scientific and effective advertisement bidding real-time calculation technology is established, the conversion efficiency in the station is favorably improved, the transaction amount of stores of digital mall sellers is effectively improved, the flow income of the digital mall is effectively improved (more advertisement cost is charged), and the advertisement experience of users in the digital mall is effectively improved (more interested advertisements are seen).

Drawings

FIG. 1 illustrates a flow diagram of a method for real-time bid ranking to enhance the conversion of digital mall advertisements, according to an embodiment of the invention;

FIG. 2 is a flow diagram illustrating a detailed process of a real-time bid ranking method for enhancing the conversion effectiveness of digital mall advertisements according to an embodiment of the present invention;

FIG. 3 illustrates a flow diagram of the construction of various vector libraries and classification models, according to an embodiment of the invention;

FIG. 4 shows a flowchart of a recommended price interval algorithm according to an embodiment of the invention;

FIG. 5 illustrates a flowchart of the calculation of the mean and standard deviation of prices in the recommended price interval algorithm according to an embodiment of the present invention;

FIG. 6 shows a flow chart of construction of a keyword vector library in a delivered page recommended keyword matching algorithm according to an embodiment of the present invention;

FIG. 7 shows a flowchart of an oCPM traffic distribution algorithm according to an embodiment of the present invention;

FIG. 8 illustrates a flow chart for building an unsupervised learning population library and classification model according to an embodiment of the present invention;

FIG. 9 illustrates a flow chart for building an unsupervised learning product library and classification model according to an embodiment of the invention;

FIG. 10 illustrates a flow diagram for building an unsupervised learning keyword classification model according to an embodiment of the invention;

FIG. 11 illustrates a bid process click-through/conversion prediction model optimization flow diagram according to an embodiment of the present invention;

FIG. 12 illustrates a crowd tag computation flow diagram according to an embodiment of the invention;

FIG. 13 illustrates an anti-cheating algorithm flow diagram according to an embodiment of the present invention;

FIG. 14 is a flowchart illustrating a general blacklist repository configuration in an anti-cheating algorithm according to an embodiment of the present invention;

FIG. 15 illustrates a flowchart of a configuration of a blacklist library of click behaviors in an anti-cheating algorithm, according to an embodiment of the present invention;

FIG. 16 shows an AB test flow diagram in accordance with an embodiment of the present invention;

FIG. 17 illustrates a block diagram of a computing device capable of implementing various embodiments of the invention.

Detailed Description

The principles and spirit of the present invention will be described with reference to a number of exemplary embodiments shown in the drawings. It should be understood that these embodiments are described only to enable those skilled in the art to better understand and to implement the present invention, and are not intended to limit the scope of the present invention in any way.

Referring to fig. 1 in conjunction with fig. 2, to solve the above problem, an embodiment of the present invention provides a real-time bidding sorting method for improving advertisement conversion effect of a digital mall, including: when the advertisement showing opportunity is generated, the calculation flow is as follows:

and step 10, judging whether the request is a page turning request, and if not, entering an advertisement sequencing process.

In step 10, when receiving a request sent by a user, the server first determines whether the request is a page turning request, and determines whether to recalculate or directly read the cache, and if the request is a page turning request, directly reads the cache queue and jumps to step 30, and if the request is not a page turning request, the advertisement sorting process in step 20 is performed.

Step 20, advertisement sorting process

As described above, if the request from the user is not a page turning request, and other requests, the advertisement sorting process in step 20 is entered, so that the user can read the latest advertisement sorting sequence in the cache when turning the page, and then the advertisement of the goods or services is presented to the user according to the sorting result after turning the page. Specifically, in one embodiment, step 20 includes the following sub-steps:

step 201, obtaining basic traffic information.

Optionally, the basic traffic information includes user information, traffic information, information of a browser ua (user agent), user characteristics/tags, location information, search keywords, and the like, and the user characteristics and the tags may be called from a big data analysis service. The big data analytics service is configured as a process of a digital mall server, which includes user features and tag data that are retrieved when invoked in step 201.

Step 202, filtering related products according to the search keywords and the filtering conditions in the basic traffic information, and generating an initial advertisement queue.

The advertisements are guaranteed to be relevant to the user's behavior, via step 202. In step 202, a relevance computation service in the digital mall server may be called, and a core module of the relevance computation service may be configured as a relevance computation engine, which may be configured to filter products matching the filtering condition from the advertisement data of the server according to the filtering condition such as a keyword based on a search engine technology, and load the products into an advertisement queue.

The ad queue is further filtered according to targeting conditions, step 203.

Optionally, the targeting conditions include keywords, categories, products, groups, and the like. For example, when the user performs a keyword search or uses a filtering condition in step 202, a secondary filtering search control, or a selection control may be configured in the system for further filtering by the user.

At step 204, the advertisement base bids in the queue are calculated.

As one example, in calculating the advertisement base bids in the queue, the cost coefficients can be read and calculated in conjunction with the cost coefficients to improve the accuracy of the calculation. The cost control coefficient is provided by PID service, and complies with the price fluctuation upper and lower limits of the advertisement activity when calculating.

And step 205, pushing the advertisement queue into a real-time reasoning service, and calculating the click rate (CTR) and/or the conversion rate (CVR) of the advertisement in real time through the model. And correcting the advertisement bids according to the calculation result, and sorting the advertisement bids from high to low according to the price. For example, the click rate and the conversion rate may be used as calculation parameters for advertisement ranking correction, and the higher the click rate and the conversion rate, the better the advertisement ranking is promoted. In one embodiment, the core module of the real-time reasoning service is configured as a bid click-rate/conversion rate prediction model (corresponding to the "behavior prediction modeling" in FIG. 2). The bid click-through rate/conversion rate prediction model is constructed as described below.

And step 206, reading the flow distribution coefficient, calculating the advertisement occurrence opportunities in the queue, forming final sequencing and outputting the advertisement queue.

And step 30, outputting the sorted first advertisement and popping up the advertisement from the queue. And recording the log.

In conclusion, the bidding service comprehensively evaluates the data performance, the flow characteristic, the product characteristic, the behavior characteristic of the user and the like of the advertisement, and is a service flow specially designed for the flow scene in the digital mall station. Visual promotion of marketing effectiveness for mall advertisers can be automated without human intervention.

In step 20, at least some of the sub-steps require the invocation of an associated algorithm or model for the calculation. In the following, an algorithm and/or model that may be used in each sub-step of step 20 is described, according to some embodiments of the present invention.

In one embodiment, an initial database, an unsupervised classification model, and a configuration key algorithm are required to be constructed and integrated to form a real-time bidding calculation engine, and the server calls the real-time bidding calculation engine to realize advertisement ranking when executing step 20. First, the initial database construction, the unsupervised classification model construction and the key algorithm configuration are briefly introduced.

< construction of initial database >

In one embodiment, the initial database includes a keyword vector library, a product vector library, a crowd vector library.

< construction of unsupervised Classification model >

In one embodiment, the unsupervised classification model includes a keyword classification model, a product classification model, a crowd classification model.

< configuration Key Algorithm >

In one embodiment, the key algorithms include a keyword recommendation/matching algorithm, a pCTR/pCVR prediction algorithm (i.e., click-through rate/conversion rate prediction algorithm), an opcm/opcpc/opca flow distribution algorithm, a price interval recommendation algorithm, a related product recommendation algorithm, and an anti-cheating algorithm.

Related terms: oCPM: the abbreviation of Optimized Cost per Mille, i.e., optimizing thousands of presentation bids, is again paid per cpm per se. And a more accurate click rate and conversion rate pre-estimation mechanism is adopted, the advertisement is displayed to the user who is most easy to generate conversion, the conversion rate is improved, the conversion cost is reduced and the running quantity is faster while the flow is acquired. oCPC: the abbreviation of Optimized Cost per Click, i.e., Optimized pay-per-Click, is also paid per cpc per se. The accuracy of a more scientific conversion rate pre-estimation mechanism is adopted, so that an advertiser can be helped to obtain more high-quality flow and improve the conversion completion rate. The system can intelligently and dynamically adjust the bids according to the estimated conversion rate and the competitive environment based on the mass data of multi-dimensional, real-time feedback and historical accumulation on the basis of the bids of advertisers, further optimize the advertisement sequencing, help the advertisers to bid the most suitable flow and reduce the conversion cost. oCPA: the abbreviation of Optimized Cost per Action, i.e., optimization behavior bid, is again paid per cpa per se. When an advertiser selects a specific optimization target (such as activation of mobile application and ordering of a website) in an advertisement putting process, providing an average price willing to be paid for the advertisement putting target, timely and accurately returning effect data, estimating the conversion value of each click to the advertiser in real time by means of a conversion estimation model, automatically bidding, and finally deducting fees according to the click; meanwhile, the conversion pre-estimation model can be continuously and automatically optimized according to the advertisement conversion data of the advertiser.

< verification of effect >:

as described above, the initial database, the unsupervised classification model and the key algorithm need to be integrated together to form a real-time bid calculation engine, and in order to verify the calculation effect before online, in one embodiment, abdest can be used for effect verification of the real-time bid calculation engine.

< keyword vector library of initial database >

As previously described, the initial database includes a library of keyword vectors. In the digital mall system, in order to guide the mall user, keywords suitable for purchase or click-through are matched and provided for the mall user according to the target product ID. In this embodiment, the implementation of the algorithm includes constructing a keyword vector library (i.e., a vector library), then calculating a correlation coefficient between each product and other products using a correlation coefficient formula according to the product keyword vector library, and sorting and storing the results according to the correlation. The output returned after saving may be a query corresponding to a product keyword relevance coefficient, which may be ordered according to a certain rule, e.g., in some embodiments 1 is the most relevant.

As an embodiment, in conjunction with fig. 6, the keyword vector library may be constructed by the following process.

Step 1, manufacturing a thermal encoding (ONE-HOT) by using keywords of different sources of all products as characteristics.

In this step 1, the features that are fabricated may be in the millions, so as a preferred way, the first N (N is a positive integer) bits of keys may be reserved for each source. In addition, a PAD can be made as a complementary feature. PAD can be used instead if no keywords can be found or established. In step 1, FM (factorization) and deep FM algorithms may also be used as backup algorithms for constructing features. In addition, in step 1, the sources can be increased or decreased according to the actual needs. As some examples, the keyword sources may include: i. effect class keyword: high-frequency keywords of product sources; long-tailed keywords (long-tailed keywords are needed for better drainage to the user, which may better promote user effectiveness): 1. low-frequency keywords of product sources; 2. product theme, content high frequency words; 3. high-frequency words of a product source page; 4. searching source keywords by a user; 5. the advertisement delivery user self-selects keywords; 6. searching page product source keywords; supplement: the statistical word frequency needs to be filtered according to the STOPWORD lexicon.

And 2, counting the use quantity of each source and keyword of each product, and taking the ratio of different keyword quantities of the product as an original characteristic value. In step 2, a standard normalization algorithm may be used to perform data distribution optimization, and in addition, a range normalization and mean-removal normalization algorithm may also be used.

And 3, multiplying the original characteristic values of different sources by the source weight to serve as final characteristic values. In the initial stage, the source weight can be set manually, and in the later stage, the source weight can be used as a neural network model according to the effect used by the user, and the neural network model is trained again to optimize the weight.

And 4, taking the whole matrix formed in the step 1-3 as an original keyword vector library.

Therefore, a basic keyword vector library is established, optimization can be performed according to actual conditions, the used algorithm is a collaborative filtering algorithm, and selection can be performed according to actual data and effects.

< population vector library of initial database, population classification model of unsupervised classification model >

As described above, the initial database further includes a population vector library, and the unsupervised classification model includes a population classification model. According to an embodiment of the present invention, in conjunction with fig. 8, the crowd vector library and the crowd classification model may be constructed as follows.

Step 1, making a thermal encoding (ONE-HOT) by using data sources such as all products and product classifications as characteristics. In step 1, the feature may preserve the key compression latitude of TOP N. In addition, a PAD is made as a bit-filling feature. If no features can be found or not established PAD is used instead. Alternative algorithms FM, DeepFM.

As one example, the latitude of the data source may include: i. recent periods (a period may be defined, typically using data in 7-14 days) of user access to N products; a recent period (a period may be defined, typically using data within 7-14 days) user access to a classification of N products (typically low weight); ip territory; product receiving territory; v. the model of the mobile phone; brand name; average monthly spending amounts; a shopping time period; ix, purchasing power: L1-L5; frequent visit period: judging the most frequently used browsing and purchasing time periods of the user in the most frequently used purchasing time period of the last 180 days, and displaying the labels according to 24 hours a day; whether logging in; xii time last visit interval; upload purchase interval time; specific population label.

And 2, counting the number of times of products visited by each user and product classification, and taking the ratio of the total number of the products visited by the user to a single product as an original characteristic value. In step 2, the algorithm preferentially used is standard normalization, and data distribution optimization is performed. Alternative algorithms include range normalization, and de-mean normalization.

And 3, multiplying the original characteristic values of different sources by the source weight to serve as final characteristic values. In step 3, in the initial stage, the source weight can be set manually, and in the later stage, the source weight can be used as a neural network model according to the effect used by the user, and training is performed again to optimize the weight.

And 4, taking the data sources as a crowd vector library and storing the crowd vector library. And 5, generating a user group, namely a crowd classification model, by using the K-MEANS and the data source as input. And 6, saving the model to provide the calling of the crowd updating program.

< model of attributes of population >

In some embodiments, the crowd attribute model may also be constructed using a crowd vector library. It can be constructed, for example, by the following steps (in conjunction with FIG. 12):

step 1, loading training samples. And 2, adjusting sample distribution to ensure different sample quantity.

In step 2, the content can be increased by directly and randomly copying the contents of fewer classified samples when the number of training samples is small. If the number of the samples to be sorted is large, a large number of classified samples can be reduced, so that the sample data of the multi-classified sample and the sample data of the small classification can be kept consistent.

And 3, loading a crowd vector library as features, performing advanced repair classification through deep learning, and verifying the model.

In step 3, a neural network algorithm may be used, and in addition, algorithms such as a decision tree, a gradient boosting decision tree, a random forest, a logistic regression, a support vector machine, and the like may be used.

And 4, storing the model to obtain the crowd attribute model. The crowd attribute model is used for predicting the basic attributes of the crowd, and the tag is a general attribute of the crowd. Such as: age, gender, occupation, etc. Such labels employ supervised learning for learning predictions.

In addition to the crowd-common attributes described above, crowd-specific tags may be designed: this label is defined in terms of existing services. The data has definite definition and algorithm, and is generated according to the access behavior and the purchase record of the end user. Such as: definition of purchasing power label: L1-L5 is an index of wasted effort from low to high, and cannot be directly matched with the amount of consumption. The label is that a plurality of consumption behaviors are combined by an algorithm to score the users, then the scores are queued from low to high, the first user contributing 20 percent of the deal amount is L1, the 20 to 40 percent of the users are L2 according to 5 equal parts of the total deal amount, and the like.

< product vector library of initial database, product classification model of unsupervised classification model >

As previously stated, the initial database also includes a library of product vectors, and the unsupervised classification model includes a product classification model. According to one embodiment of the present invention, in conjunction with FIG. 9, the product vector library and the product classification model may be constructed as follows.

Step 1, making a thermal code (ONE-HOT) from data sources such as all unsupervised learning population and label population classification as characteristics.

The unsupervised learning population and the label population classification can be obtained from a population classification model and a population attribute model.

In step 1, the feature may preserve the key compression latitude of TOP N. In addition, a PAD is made as a bit-filling feature. PAD is used instead if no features can be found or are not established. Alternative algorithms FM, DeepFM.

As one example, the data sources may include the following latitudes:

i. recent periods (periods can be defined, data within 7-14 days is generally used) unsupervised learning population tag visit volume, purchase volume.

A recent period (a period may be defined, typically using data within 7-14 days) crowd tag visit amount, purchase amount.

Buy products simultaneously (which may be defined as purchasing products within 24 hours).

And 2, counting the data distribution of each label in the total labels, and taking the ratio of the total purchase quantity of the user access to the single label as an original characteristic value.

In step 2, the algorithm preferentially used is standard normalization, and data distribution optimization is performed. Alternative algorithms include range normalization, and de-mean normalization.

And 3, multiplying the original characteristic values of different sources by the source weight to serve as final characteristic values.

In step 3, the source weight can be set manually in the initial stage, and in the later stage, the source weight can be used as a neural network model according to the effect of the user, and training is performed again to optimize the weight.

And 4, taking the formed whole matrix as a product vector library and storing the product vector library.

And 5, generating a product group, namely a product classification model, by using the K-MEANS and the product vector library as input.

And 6, saving the model to provide the calling of the crowd updating program.

< keyword Classification model of unsupervised Classification model >

As described above, the unsupervised classification model includes a keyword classification model. According to an embodiment of the present invention, in conjunction with FIG. 10, a keyword classification model may be constructed by the following steps.

Step 1, reading a keyword vector library.

And 2, establishing a grouping model, namely a keyword classification model, by using the K-MEANS and the keyword vector library as input.

And 3, saving the model to provide other program calls.

So far, the embodiment of the invention has been implemented for constructing each vector library and classification model. In general, in one embodiment, in conjunction with FIG. 3, the complete construction includes:

firstly, a keyword vector library is constructed, recommended keywords are generated according to the keyword vector library, and an unsupervised keyword classification model is constructed. And constructing a crowd vector library according to the crowd label.

And respectively constructing an unsupervised learning population classification model and a population attribute model according to the population vector library. In some preferred embodiments, a keyword unsupervised population classification model can be constructed by combining keywords, so as to improve the accuracy of the final model.

And constructing a product vector library based on the unsupervised learning population classification model and the population attribute model.

And constructing an unsupervised learning product classification model according to the product vector library to generate a substitute product and a related product. Taking the related product as an example, the product with high correlation coefficient can be calculated according to the product vector library. Wherein different weights can be set for optimization according to browsing products and purchasing products.

And finally, fusing the unsupervised keyword classification model, the keyword unsupervised crowd classification model, the unsupervised learning product classification model and the crowd attribute model to construct a bidding click rate/conversion rate prediction model for returning a plurality of advertisement activities to predict the generation probability of the user. As one example, the models may be loaded into a real-time reasoning service, invoked in step 205 above.

When building a bidding click rate/conversion rate prediction model, in conjunction with fig. 11, the following steps may also be employed to optimize the model:

step 1, collecting and sorting original characteristic data.

As an example, the raw feature data may include the following latitude: i. is characterized in that: advertisement product ID, keyword unsupervised learning label, advertisement product classification, advertisement product label, advertisement space, unsupervised learning crowd label, supervised learning crowd label, platform, region, time period, week; ii, a label: whether to click or not;

and 2, establishing a training data set.

In step 2, the original feature data obtained in step 1 is used as a training data set as an input of the model, and in addition, it can be noted that in step 2, it is required to ensure that the proportion of clicking users and non-clicking users with different features respectively accounts for 50%.

And 3, establishing a model.

In step 3, a naive bayes algorithm and a neural network algorithm can be used, and the alternative algorithms include a decision tree, a gradient boosting decision tree, a random forest, a logistic regression, a support vector machine, an FM, a deep FM, a GBDT, and the like. The selection can be specifically carried out according to actual data and effects and combining calculation force.

And 4, maintaining the model and using the model in real-time bidding.

< product Label >

As described above, in the above embodiments, when constructing each vector library and classification model, a product label may be needed, and as an example, the product label may be designed as follows: the current event sku _ number; extracting sku keywords from the name; extracting sku keywords from the description information; commodity classification information; commodity classification information (parent classification); grading the star number of the commodity; total number of goods scoring; grading praise number of the commodities; grading and stepping number of commodities; total number of 1 star of the commodity; 2 total stars of the commodity; total number of 3 stars in the commodity; 4 stars of the commodity; 5 stars total for the commodity; a commodity comment keyword; the transaction price of the commodity was evaluated for the last 10 times.

< user tag >

As mentioned above, in the above embodiments, when constructing each vector library and classification model, a user tag may be needed, and as an example, the user tag may be designed as follows: id (it is sufficient to lock the sample uniquely and by this id the data source can be traced back); a user ID; whether to click or not; age; sex; marriage: default, 0 married, 1 unmarked, 2 outliers, 3; whether there are children: default, 0 none, 1 none, 2; income level/year; time of occurrence of event, hours/day; day/month; week/month; monday is 1, Sunday is 7; the last transaction, how many days away from the present; frequency of transactions in the last natural month; the frequency of transactions in the year; frequency of transactions of the last 30 days; frequency of transactions over the last 365 days; the last transaction amount; average amount of last 10 transactions; the city where the user is located (trade user address, order address and bill address); the country to which it belongs; a device type; a network type; a transaction channel; recently browsing the commodity sku _ number; the number of user buys (since the last time the shopping cart was emptied); sku _ number recently purchased by the user; the last login time and day of the user; the stay time of the user logging in last time is second; a user recently searches for a keyword; ipv of the user for the last 1, 3, 7, 14, 30 days; user's pv, last 1 day, 3 days, 7 days, 14 days, 30 days; the user had the latest 1, 3, 7, 14, 30 days for the ski ipv; the user has the latest 1, 3, 7, 14, 30 days for the sku's pv; AIPL classification: default, 0A,1I,2P,3L, 4; RFM classification: default, 0 important value customer, 1 general value customer, 2 important development customer, 3 general development customer, 4 important maintenance customer, 5 general maintenance customer, 6 important saving customer, 7 general saving customer, 8; coupon user, default, no 0, yes 0, 1.

< recommended price interval Algorithm > (maximum and minimum values are calculated off-line)

And the recommended price interval algorithm is used for returning the price interval required by putting according to the product/category/keyword selected by the user.

In one embodiment, the recommended price interval algorithm includes an interface return flow and a calculate maximum minimum common flow.

Optionally, with reference to fig. 4, the interface return flow includes:

step 1, acquiring at least one data of a product, a category and a keyword input by a user. In some embodiments, all of the above data is acquired.

And 2, selecting products or keywords corresponding to the orientation range.

Since the server system already has the relevant products, categories and locations and their historical data for the last N days (N can be configured at a reasonable value using the parameters). It is possible to orient according to the product, category and keyword entered by the user.

And 3, obtaining the product price interval in the corresponding condition, if the data exists, returning the result, and if the data does not exist, entering the step 4.

And 4, acquiring a price interval of the keywords of the user product, returning a result if data exist, and entering the step 5 if no data exist.

And 5, acquiring a product classification price interval, returning a result if data exist, and returning a default price interval if no data exist.

The returned result includes a suggested price minimum and a suggested price maximum. In addition, the system may also verify a minimum price, subject to the system low price if the minimum price is lower than the current set low price.

Optionally, the returned result may further include an average value and a standard deviation of the price, and such calculation logic is more flexible at the interface layer. In connection with fig. 5, the mean and standard deviation of the price can be calculated as follows:

step 1, inputting LOG number (namely bargain price) of the last N days.

In step 1, data may be compressed if the LOG is too large, e.g., reduced to an average bargain for the time period.

And 2, calculating the average number and the standard deviation of the input bargaining prices. And 3, filtering abnormal data by adopting a 3 sigma rule. And 4, calculating the average value and the standard deviation of the filtered LOG. In step 4, a neural network algorithm may be used, and in addition, algorithms such as a decision tree, a gradient boosting decision tree, a random forest, a logistic regression, a support vector machine, and the like may be adopted.

And 5, storing the result into the average value and the standard deviation. In step 5, optionally, the average value plus or minus 2 standard deviations is generally adopted as the price interval, wherein 2 standard deviations can be configured. The recommendation range can also be adjusted according to the tendency, if the user wants to have higher price, the standard deviation can be reduced by one, and 2 standard deviations are added to be used as the recommendation range.

< configure oCPM traffic distribution Algorithm >

In one embodiment, in conjunction with FIG. 7, the flow allocation algorithm is configured by the following steps.

Step 1, constructing original data aiming at the advertisement space. And 2, acquiring original data of the sub-advertisements of the current advertisement space.

In step 2, optionally, the raw data includes presentation rate, CTR, CPC (Cost per Click), CPM, conversion rate, conversion value, current distribution ratio, and the like. The conversion value can be set by a user, or only the conversion times or the bargaining price can be used. And controlling the data acquisition time range according to the data amount. When the general data is less, the data acquisition can be carried out for 10, 30, 60 and 120 minutes. For example, a 10 minute data deficit is supplemented with 30 minute data, and so on.

And 3, calculating the current score according to the scoring standard, wherein the score is the same as the score of the genetic algorithm.

It should be noted that, in step 2, if there is no original data, the current data of the ad slot is used to supplement the data, and then step 3 is entered. If there is no current data for the ad slot either, then system default data is used to supplement, proceeding to step 3.

And 4, constructing a genetic algorithm chromosome. In step 4, the advertisement material can be used as a gene, and 0-100 can be used as a value range of the gene. And 5, setting a genetic algorithm scoring standard. In step 5, the final conversion value is calculated based on the chromosome assigned flow ratio, the higher the score with about a large value. If all gene values are greater than 100, the score is 0. Preventing percent overflow.

And 6, setting and acquiring an optimal chromosome after the genetic algorithm evolves N generations, and scaling according to the target 100 if the final percentage is not 100. And 7, calculating the scaled proportion, and comparing the scaled proportion with the original score. The high scoring ratio is used as the final allocation ratio. And 8, storing the result and providing the result to the advertisement server for use.

Therefore, the calculation of the optimal effect of the traffic proportion in the advertisement space is completed through the genetic algorithm, and the optimized CTR is improved by 20% through the ABtest compared with the ordinary traffic distribution algorithm. As an example, the score calculated by the traffic allocation algorithm may be stored in the traffic allocation cache as a traffic allocation coefficient, and in the above step 206, the score is called by the server, and the traffic weight (i.e., the traffic weight allocation) of each advertisement in the advertisement queue obtained in step 205 is determined according to the coefficient, so as to further accurately correct the ordering of the advertisement queue.

< configuration anti-cheating Algorithm >

In some embodiments, the anti-cheating algorithm is configured by the following flow (in conjunction with FIG. 13). Step 1, loading a universal black list library. And 2, checking whether the flow belongs to a blacklist, if so, adding the general flow suspicious identification into the LOG, and simultaneously entering the next flow. If not, go to step 3. And 3, checking whether the step is a click behavior blacklist, and certainly, the step 3 can also be started by loading a click behavior blacklist library. And if the flow belongs to the click behavior blacklist, adding the general flow suspicious identification into the LOG, and simultaneously entering the next flow.

In the above flow, the general blacklist library may be configured by the following flow (in conjunction with fig. 14): step 1, acquiring an original log. And 2, splitting the log according to time, wherein the splitting rule is based on the data volume and the service requirement. In step 2, the data volume is large, and the log can be split according to time periods. And when the data volume is too large, splitting can be performed on the page type. And step 3, counting the required statistical values of each group, such as access times of UA, IP, equipment ID and the like. And 4, filtering the abnormal value by using 3 sigma rule. And 5, storing the abnormal value into a blacklist to obtain a universal blacklist library. And 6, cleaning the automatically maintained blacklist, for example, the validity period of the blacklist is 1-14 days, and the expiration is automatically updated.

In the above flow, the click behavior blacklist library may be configured by the following flow (in conjunction with fig. 15):

step 1, obtaining user click behavior data. And 2, splitting the log into statistical data according to time, wherein the splitting rule is based on the data volume and the service requirement. In step 2, the data volume is large, and the log can be split according to time periods. And when the data volume is too large, splitting can be performed on the page type. And step 3, grouping according to UA, IP and equipment IP, and counting the average access depth and average residence time after the user clicks. And 4, filtering the abnormal value by using 3 sigma rule. And 5, storing the abnormal value into a blacklist to obtain a click behavior blacklist library. And 6, cleaning the automatically maintained blacklist, for example, the validity period of the blacklist is 1-14 days, and the expiration is automatically updated.

Anti-cheating traffic filtering is mainly divided into two blacklists, namely a general blacklist and a click behavior blacklist. And when the users are found to be blacklisted, adding corresponding suspicious traffic identifications into the LOG. The universal blacklist is used for cleaning data of the extreme end to ensure the reliability of data sources, so that each model is better established. The method comprises the steps of filtering abnormal users such as ultrahigh access users and crawlers. And (3) clicking a behavior blacklist: to mark anomalous click behavior. Such as dense clicks, malicious clicks.

< AB test configuration >

For the actual effect of the constructed model or algorithm strategy, in addition to the evaluation itself against the algorithm itself at the off-line stage, an AB experiment should also be performed on the production environment (on-line environment). To ensure that the design is forward, or meaningful, on the optimization goal.

In some embodiments, user traffic needs to be equally divided in a hash manner, so as to ensure that data distribution under each partition is consistent.

Different hash algorithms are selected, the uniformity is different, but the difference is not large, and the MurMurHash is adopted in the embodiment.

In order to avoid errors in separate evaluation of the split flow, this example adopts an AABB split flow method (see fig. 16), where a is a control group (control groups a1 and a2) and B is an experimental group (experimental groups B1 and B2), and 90% of all flow rates are assigned to the control group and 10% to the experimental group. And evaluating the result according to the statistic value of the control group and the statistic value of the experimental group.

The embodiment of the invention also provides a real-time bidding sorting system for improving the advertisement conversion effect of the digital mall, which comprises the following steps: the judging module is used for judging whether the request is a page turning request or not, and if not, entering an advertisement sequencing process; an advertisement sequencing flow module, configured to execute the advertisement sequencing flow, where the advertisement sequencing flow includes: acquiring basic flow information; filtering related products according to search keywords and filtering conditions in the basic flow information to generate an initial advertisement queue; further filtering the ad queue according to targeting conditions; calculating the advertisement base bid in the queue; pushing the advertisement queue into a real-time reasoning service, and calculating the click rate and/or the conversion rate of the advertisement in real time through the model; reading the flow distribution coefficient, calculating the advertisement occurrence opportunities in the queue, forming final sequencing and outputting an advertisement queue; and the output module is used for outputting the ordered first advertisements and popping up the advertisements from the queue. It should be understood that, the above program modules and the steps described in the method embodiments have a one-to-one correspondence relationship, and the technical solution described in the method embodiments may also be applied to the specific configuration of each program module, and in order to avoid repetition, the details are not described here again.

The invention has the following characteristics: the data assets of the digital mall are integrally considered, the data assets comprise transaction data, user behavior data, search data, user characteristics and commodity characteristics, data value is mined and effectively combined, and support is provided for marketing behaviors; combing the service data one by one and forming an output object; a real-time bidding technology of digital advertisements is introduced to maximize the flow value; the sequencing of the advertisement materials (commodity materials) is the most effective method for improving conversion, and the method is continuously optimized through scientific calculation according to the effect and data change; the system has a set of effect testing and evaluating mechanism, and reasonably evaluates the actual effect of the bidding method; the anti-cheating mechanism is provided, and invalid data and malicious behaviors are cleaned; and on the premise of protecting the privacy of the user, performing the actions.

FIG. 17 illustrates a block diagram of a computing device 600 capable of implementing multiple embodiments of the present invention. The electronic devices are intended to represent various forms of digital computers, and the components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not intended to limit implementations of the inventions described and/or claimed herein.

As shown in fig. 17, the apparatus 600 includes a computing unit 601 that can perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM)602 or a computer program loaded from a storage unit 608 into a Random Access Memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the device 600 can also be stored. The calculation unit 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input/output (I/O) interface 605 is also connected to bus 604.

A number of components in the device 600 are connected to the I/O interface 605, including: an input unit 606 such as a keyboard, a mouse, or the like; an output unit 607 such as various types of displays, speakers, and the like; a storage unit 608, such as a magnetic disk, optical disk, or the like; and a communication unit 609 such as a network card, modem, wireless communication transceiver, etc. The communication unit 609 allows the device 600 to exchange information/data with other devices via a computer network such as the internet and/or various telecommunication networks.

The inventive concept is explained in detail herein using specific examples, which are only provided to help understanding the core idea of the present invention. It should be understood that any obvious modifications, equivalents and other improvements made by those skilled in the art without departing from the spirit of the present invention are included in the scope of the present invention.

Claims

1. The real-time bidding sorting method for improving the advertisement conversion effect of the digital mall is characterized by comprising the following steps of:

judging whether the request is a page turning request or not, and if not, entering an advertisement sequencing process;

the advertisement sequencing process comprises the following steps:

acquiring basic flow information;

filtering related products according to search keywords and filtering conditions in the basic flow information to generate an initial advertisement queue;

further filtering the ad queue according to targeting conditions;

calculating the advertisement base bid in the queue;

pushing the advertisement queue into a real-time reasoning service, and calculating the click rate and/or the conversion rate of the advertisement in real time through the model;

reading the flow distribution coefficient, calculating the advertisement occurrence opportunities in the queue, forming final sequencing and outputting an advertisement queue;

the sorted first ad is output and popped from the queue.

2. The method for real-time bid sorting for improving the advertisement conversion effect of the digital mall according to claim 1, wherein a real-time bid calculation engine called in the advertisement sorting process is integrated by the following means:

constructing an initial database; the initial database comprises a keyword vector library, a product vector library and a crowd vector library;

constructing an unsupervised classification model; the unsupervised classification model comprises a keyword classification model, a product classification model and a crowd classification model;

configuring a key algorithm; the key algorithms comprise a keyword recommendation and/or matching algorithm, a click rate and/or conversion rate pre-estimation algorithm, a flow distribution algorithm and a price interval recommendation algorithm.

3. The method according to claim 2, wherein in configuring the key algorithm, configuring an anti-cheating algorithm further comprises cleaning up endpoint data according to a general blacklist and marking abnormal click behavior according to a click behavior blacklist.

4. The real-time bidding sorting method for improving the advertisement conversion effect of the digital mall according to claim 2, wherein an AB test is further adopted to verify the actual effect of the constructed model or algorithm strategy.

5. The method for real-time bid ranking for improving advertisement conversion efficiency of a digital mall of claim 3, wherein the constructing the initial database and the constructing the unsupervised classification model comprise:

firstly, a keyword vector library is established, and recommended keywords and an unsupervised keyword classification model are generated according to the keyword vector library; constructing a crowd vector library according to the crowd label;

respectively constructing an unsupervised learning crowd classification model and a crowd attribute model according to a crowd vector library;

constructing a product vector library based on an unsupervised learning population classification model and a population attribute model;

constructing an unsupervised learning product classification model according to the product vector library;

and finally, fusing according to the unsupervised keyword classification model, the unsupervised learning crowd classification model, the unsupervised learning product classification model and the crowd attribute model, thereby constructing a bidding click rate and/or conversion rate prediction model for returning a plurality of advertisement activities to predict the generation probability of the user.

6. The real-time bidding sorting method for improving the advertisement conversion effect of the digital mall according to claim 2, wherein the method for configuring the price interval recommendation algorithm comprises:

acquiring at least one data of a product, a category and a keyword input by a user;

selecting products or keywords corresponding to the orientation range;

obtaining product price intervals in corresponding conditions, if data exist, returning results, and if no data exist, then:

obtaining a price interval of a keyword of a user product, if the price interval has data, returning a result, and if the price interval has no data, then:

and obtaining a product classification price interval, if the data exist, returning a result, and if the data do not exist, returning to a default price interval.

7. The real-time bidding sorting method for improving the advertisement conversion effect of the digital mall according to claim 2, wherein the configuration of the traffic distribution algorithm comprises:

acquiring original data of a current advertisement space;

calculating the current score according to the scoring standard;

constructing a genetic algorithm chromosome;

setting a genetic algorithm scoring standard;

the genetic algorithm evolves for N generations and then sets and obtains the optimal chromosome,

calculating a score according to the obtained optimal chromosome, comparing the score with the original score, and using the proportion of high score as a final distribution proportion;

the final allocation ratio is provided to the ad server for use.

8. The real-time bidding sorting system for improving the advertisement conversion effect of the digital mall is characterized by comprising the following steps of:

the judging module is used for judging whether the request is a page turning request or not, and if not, entering an advertisement sequencing process;

an advertisement sorting flow module, configured to execute the advertisement sorting flow, where the advertisement sorting flow includes:

acquiring basic flow information;

further filtering the ad queue according to the targeting condition;

calculating the advertisement base bid in the queue;

and the output module is used for outputting the ordered first advertisements and popping up the advertisements from the queue.

9. A storage medium, characterized in that a computer program is stored thereon, which program, when being executed by a processor, carries out the method of any one of claims 1-7.

10. An electronic device, the electronic device comprising:

one or more processors; and

storage means for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to carry out the method of any one of claims 1-7.