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

Abstract

The present invention discloses a real-time bidding sorting method, system, storage medium and electronic device for improving the conversion effect of digital mall advertisements. The method includes judging whether it is a page turning request. If it is not a page turning request, entering the advertisement sorting process; the advertisement sorting process includes: obtaining basic traffic information; filtering related products according to search keywords and filtering conditions in the basic traffic information, generating an initial advertisement queue; further filtering the advertisement queue according to the directional conditions; calculating the basic bid of the advertisements in the queue; pushing the advertisement queue into the real-time reasoning service, and calculating the click-through rate and / or conversion rate of the advertisement in real time through the model; reading the traffic distribution coefficient, calculating the chance of the advertisements in the queue, forming the final sorting, and outputting the advertisement queue; outputting the advertisement with the first ranking and popping it out from the queue. The present invention scientifically and orderly organizes algorithms and big data technologies, refines the data resources of the digital mall, and can improve the conversion and drainage effect of the advertisement resources in the station.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a real-time bidding ranking method, system, storage medium and electronic device for improving the conversion effect of digital mall advertisements. Background Art

[0002] When building a digital shopping mall platform, a traffic diversion system will be provided to sellers (stores). Through bidding services, the website traffic (advertising resources) in the mall will be purchased and diverted to the store homepage or product detail page, which will then be used to bring about conversions. Currently, most digital shopping mall platforms have weak big data and algorithm technologies and cannot effectively divert traffic by combining business data (such as user behavior, user characteristics, product characteristics, etc.).

[0003] In view of this, the present invention is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time bidding ranking method, system, storage medium and electronic device for improving the conversion effect of digital mall advertisements, so as to solve the problem that the current digital mall cannot well introduce the real-time bidding technology of digital advertisements and maximize the traffic value.

[0005] To solve the above problems, in a first aspect, an embodiment of the present invention provides a real-time bidding ranking method for improving the conversion effect of digital mall advertisements, comprising:

[0006] Determine whether it is a page turning request. If not, enter the advertisement sorting process. The advertisement sorting process includes: obtaining basic traffic information; filtering related products according to the search keywords and filtering conditions in the basic traffic information to generate an initial advertisement queue; further filtering the advertisement queue according to the targeting conditions; calculating the basic bid of the advertisements in the queue; pushing the advertisement queue into the real-time inference service, and calculating the click-through rate and / or conversion rate of the advertisements in real time through the model; reading the traffic distribution coefficient, calculating the chance of the advertisements in the queue, forming a final sorting, and outputting the advertisement queue; outputting the advertisement ranked first and popping it out of the queue.

[0007] Optionally, the real-time bidding calculation engine called in the advertisement ranking process is integrated in the following manner: constructing an initial database; the initial database includes a keyword vector library, a product vector library and a crowd vector library; constructing an unsupervised classification model; the unsupervised classification model includes a keyword classification model, a product classification model, and a crowd classification model; configuring key algorithms; the key algorithms include a keyword recommendation and / or matching algorithm, a click-through rate and / or conversion rate estimation algorithm, a traffic allocation algorithm, and a price range recommendation algorithm.

[0008] Optionally, configuring the key algorithm also includes configuring an anti-cheating algorithm, wherein the anti-cheating algorithm includes cleaning extreme data according to a general blacklist and marking abnormal click behaviors according to a click behavior blacklist.

[0009] Optionally, AB testing is also used to verify the actual effect of the constructed model or algorithm strategy.

[0010] Optionally, the constructing of the initial database and the constructing of the unsupervised classification model include: first constructing a keyword vector library, generating recommended keywords and constructing an unsupervised keyword classification model based on the keyword vector library; constructing a crowd vector library based on crowd labels; constructing an unsupervised learning crowd classification model and a crowd attribute model based on the crowd vector library; constructing a product vector library based on the unsupervised learning crowd classification model and the crowd attribute model; constructing an unsupervised learning product classification model based on the product vector library; finally, fusing the unsupervised keyword classification model, the unsupervised learning crowd classification model, the unsupervised learning product classification model, and the crowd attribute model to construct a bidding click-through rate and / or conversion rate prediction model for returning the predicted user generation probability of multiple advertising campaigns.

[0011] Optionally, the configuration method of the price range recommendation algorithm includes: obtaining at least one data among the products, categories, and keywords input by the user; selecting products or keywords corresponding to the targeted range; obtaining the product price range within the corresponding conditions, returning the result if there is data, and if there is no data: obtaining the user product keyword price range, returning the result if there is data, and if there is no data: obtaining the product classification price range, returning the result if there is data, and returning the default price range if there is no data.

[0012] Optionally, the configuration of the traffic distribution algorithm includes: obtaining the original data of the current advertising position; calculating the current score according to the scoring criteria; constructing a genetic algorithm chromosome; setting the genetic algorithm scoring criteria; setting the optimal chromosome after the genetic algorithm evolves for N generations, calculating the score based on the obtained optimal chromosome, and comparing it with the original score, using the ratio with a high score as the final distribution ratio; and providing the final distribution ratio to the advertising server for use.

[0013] In a second aspect, an embodiment of the present invention provides a real-time bidding ranking system for improving the conversion effect of digital mall advertisements, including: a judgment module, used to judge whether it is a page turning request, if it is not a page turning request, entering the advertisement ranking process; an advertisement ranking process module, used to execute the advertisement ranking process, the advertisement ranking process including: obtaining basic traffic information; filtering related products according to the search keywords and filtering conditions in the basic traffic information, and generating an initial advertisement queue; further filtering the advertisement queue according to the directional conditions; calculating the basic bid of the advertisements in the queue; pushing the advertisement queue into the real-time inference service, and calculating the click-through rate and / or conversion rate of the advertisements in real time through the model; reading the traffic distribution coefficient, calculating the chance of the advertisements in the queue, forming a final ranking, and outputting the advertisement queue; an output module, used to output the advertisement ranked first and pop up the advertisement from the queue.

[0014] In a third aspect, an embodiment of the present invention provides a storage medium having a computer program stored thereon, wherein the program implements the above method when executed by a processor.

[0015] In a fourth aspect, an embodiment of the present invention provides an electronic device, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0016] The embodiments of the present invention provide a real-time bidding sorting method, device, storage medium and electronic device for improving the conversion effect of digital mall advertisements. The method scientifically and orderly organizes algorithms and big data technologies, refines the data resources of the electronic mall, and can greatly improve the conversion and drainage effect of the advertising resources in the site. That is, a scientific and effective real-time calculation technology for advertising bidding is established, which is conducive to improving the conversion efficiency in the site, effectively increasing the transaction volume of the digital mall sellers' stores, effectively increasing the traffic revenue of the digital mall (collecting more advertising fees), and effectively improving the advertising experience of digital mall users (seeing more advertisements of interest). BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a real-time bidding ranking method for improving the conversion effect of digital mall advertisements according to an embodiment of the present invention is shown;

[0018] Figure 2 A flowchart showing a detailed process of a real-time bidding ranking method for improving the conversion effect of digital mall advertisements according to an embodiment of the present invention is shown;

[0019] Figure 3 A flowchart showing the construction of various vector libraries and classification models according to an embodiment of the present invention is shown;

[0020] Figure 4A flowchart of a recommended price range algorithm according to an embodiment of the present invention is shown;

[0021] Figure 5 A flowchart of calculating the average and standard deviation of prices in a recommended price range algorithm according to an embodiment of the present invention is shown;

[0022] Figure 6 A flowchart of constructing a keyword vector library in a keyword matching algorithm for recommended keywords on a delivery page according to an embodiment of the present invention is shown;

[0023] Figure 7 A flow chart of an oCPM traffic allocation algorithm according to an embodiment of the present invention is shown;

[0024] Figure 8 A flowchart of constructing an unsupervised learning population database and a classification model according to an embodiment of the present invention is shown;

[0025] Fig. 9 A flowchart of establishing an unsupervised learning product library and a classification model according to an embodiment of the present invention is shown;

[0026] Fig.10 A flowchart of establishing an unsupervised learning keyword classification model according to an embodiment of the present invention is shown;

[0027] Fig.11 A flowchart of the click-through rate / conversion estimation model optimization of the bidding process according to an embodiment of the present invention is shown;

[0028] Fig.12 A flow chart of crowd label calculation according to an embodiment of the present invention is shown;

[0029] Fig.13 A flow chart of an anti-cheating algorithm according to an embodiment of the present invention is shown;

[0030] Fig.14 A flowchart of configuring a universal blacklist library in an anti-cheating algorithm according to an embodiment of the present invention is shown;

[0031] Fig.15 A flowchart of configuring a click behavior blacklist library in an anti-cheating algorithm according to an embodiment of the present invention is shown;

[0032] Fig.16 A flowchart of an AB test according to an embodiment of the present invention is shown;

[0033] Fig.17 A block diagram of a computing device capable of implementing various embodiments of the present invention is shown. DETAILED DESCRIPTION

[0034] The principle and spirit of the present invention will be described below with reference to several exemplary embodiments shown in the accompanying drawings. It should be understood that these embodiments are described only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way.

[0035] Please refer to Figure 1 Combined with Figure 2 In order to solve the above problems, an embodiment of the present invention provides a real-time bidding ranking method for improving the conversion effect of digital mall advertisements, including: when an advertisement display opportunity is generated, the calculation process is as follows:

[0036] Step 10, determine whether it is a page turning request, if not, enter the advertisement sorting process.

[0037] In step 10, when the server receives a request from the user, it first determines whether it is a page turning request, so as to decide whether to recalculate or directly read the cache. If it is a page turning request, it directly reads the cache queue and jumps to step 30. If it is not a page turning request, it enters the advertisement sorting process of step 20.

[0038] Step 20, Advertisement Sorting Process

[0039] As described above, if the request from the user is not a page turning request, but other requests, the advertisement sorting process of step 20 is entered, so that when the user turns the page, the latest advertisement sorting sequence can be read in the cache, so that after turning the page, the advertisement of the product or service is displayed to the user according to the sorting result. Specifically, in one embodiment, step 20 includes the following sub-steps:

[0040] Step 201, obtaining basic traffic information.

[0041] Optionally, the basic traffic information includes user information, traffic information, browser UA information (userAgent), user features / tags, location information, search keywords, etc. User features and tags can be called from the big data analysis service. The big data analysis service is configured as a process of the digital mall server, which includes user features and tag data. When it is called in step 201, the user features and tag data can be obtained.

[0042] Step 202, filtering related products according to the search keywords and filtering conditions in the basic traffic information to generate an initial advertisement queue.

[0043] Through step 202, it is ensured that the advertisement is relevant to the user behavior. In step 202, the relevance calculation service in the digital mall server can be called, and the core module of the relevance calculation service can be configured as a relevance calculation engine, which can be configured to filter products that match the filter conditions from the advertisement data of the server based on search engine technology according to filter conditions such as keywords, and load the products into the advertisement queue.

[0044] Step 203: further filter the advertisement queue according to the targeting condition.

[0045] Optionally, the targeting conditions include keywords, categories, products, groups, etc. For example, when the user performs a keyword search or uses a filtering condition in step 202, a secondary screening or secondary filtering search control or selection control can be configured in the system for the user to further filter.

[0046] Step 204, calculating the basic bids of the advertisements in the queue.

[0047] As an example, when calculating the basic bid of an ad in a queue, the cost coefficient can be read and combined with the cost coefficient to improve the accuracy of the calculation. The cost control coefficient is provided by the PID service and is calculated in accordance with the upper and lower limits of the price fluctuation of the ad campaign.

[0048] Step 205: Push the advertisement queue to the real-time reasoning service, and use the model to calculate the click-through rate (CTR) and / or conversion rate (CVR) of the advertisement in real time. Modify the advertisement bid according to the calculation result, and sort from high to low according to the price. For example, the click-through rate and conversion rate may be used as calculation parameters for modifying the advertisement ranking. The higher the click-through rate and conversion rate, the more conducive to improving the advertisement ranking. In one embodiment, the core module of the real-time reasoning service is configured as a bidding click-through rate / conversion rate prediction model (corresponding to Figure 2 The construction method of the bidding click rate / conversion rate prediction model will be described below.

[0049] Step 206, read the traffic distribution coefficient, calculate the chance of advertisements appearing in the queue, form a final ranking, and output the advertisement queue.

[0050] Step 30: Output the advertisement ranked first and eject it from the queue. Record the advertisement in the log.

[0051] In summary, the bidding service comprehensively evaluates the data performance of the advertisement itself, traffic characteristics, product characteristics, user behavior characteristics, etc. to comprehensively estimate the advertising effect. It is a service process designed specifically for the traffic scenarios within the digital mall. It can automatically provide mall advertisers with visible improvements in marketing effects without manual intervention.

[0052] In the above step 20, at least some sub-steps need to call related algorithms or models for calculation. Below, according to some embodiments of the present invention, the algorithms and / or models that may be used in each sub-step of step 20 are introduced.

[0053] In one embodiment, it is necessary to construct an initial database, an unsupervised classification model, and configure key algorithms, which are integrated to form a real-time bidding calculation engine. When the server executes step 20, the real-time bidding calculation engine is called to implement advertisement sorting. Below, a brief introduction is first given to constructing an initial database, an unsupervised classification model, and configuring key algorithms.

[0054] <Building the initial database>

[0055] In one embodiment, the initial database includes a keyword vector library, a product vector library, and a crowd vector library.

[0056] <Building an unsupervised classification model>

[0057] In one embodiment, the unsupervised classification model includes a keyword classification model, a product classification model, and a crowd classification model.

[0058] <Configure key algorithms>

[0059] In one embodiment, the key algorithms include keyword recommendation / matching algorithm, pCTR / pCVR prediction algorithm (i.e. click-through rate / conversion rate prediction algorithm), oCPM / oCPC / oCPA traffic allocation algorithm, price range recommendation algorithm, related product recommendation algorithm, and anti-cheating algorithm.

[0060] Related terms: oCPM: Optimized Cost per Mille, which means optimized bid per thousand impressions, but in essence, it is still paid according to CPM. Adopting a more accurate click-through rate and conversion rate estimation mechanism, the ads are displayed to the users who are most likely to convert, which improves the conversion rate and reduces the conversion cost while obtaining traffic, and increases the speed of running volume faster. oCPC: Optimized Cost per Click, which means optimized click payment, but in essence, it is still paid according to CPC. The accuracy of adopting a more scientific conversion rate estimation mechanism can help advertisers improve the conversion completion rate while obtaining more high-quality traffic. Based on the advertiser's bid, the system will adjust the bid based on multi-dimensional, real-time feedback and massive data accumulated in history, and intelligently adjust the bid according to the estimated conversion rate and the competitive environment, thereby optimizing the order of ads, helping advertisers to compete for the most suitable traffic, and reducing conversion costs. oCPA: Optimized Cost per Action, which means optimized behavioral bidding, but in essence, it is still paid according to CPA. When an advertiser selects a specific optimization goal (for example, mobile app activation, website ordering) during the ad delivery process, provides the average price they are willing to pay for this delivery goal, and promptly and accurately transmits performance data, the conversion estimation model estimates the conversion value of each click to the advertiser in real time, automatically bids, and ultimately charges per click. At the same time, the conversion estimation model will be continuously and automatically optimized based on the advertiser's ad conversion data.

[0061] <Implementation effect verification>:

[0062] As described above, it is necessary to integrate the initial database, unsupervised classification model and key algorithms to form a real-time bidding calculation engine. In order to verify the calculation effect before going online, in one embodiment, ABTEST can be used to verify the effect of the real-time bidding calculation engine.

[0063] <Keyword vector library of initial database>

[0064] As described above, the initial database includes a keyword vector library. In the digital mall system, in order to attract users of the mall merchants, it is necessary to match and provide suitable keywords for purchase or clicks for the merchant users based on the target product ID. To this end, it is necessary to configure a keyword matching algorithm for the delivery page recommendation. In this embodiment, the implementation of the algorithm includes constructing a keyword vector library (i.e., a vector library), and then calculating the correlation coefficient of each product with other products using a correlation coefficient formula based on the product keyword vector library, and saving the results after sorting them by relevance. The output result returned after saving can be the query corresponding to the product keyword relevance coefficient, and the relevance coefficient can be sorted according to certain rules. For example, in some embodiments, 1 is the most relevant.

[0065] As an example, combining Figure 6 ,The keyword vector library can be constructed through the following process.

[0066] Step 1: One-hot encoding (ONE-HOT) is performed on keywords from different sources of all products as features.

[0067] In this step 1, the features produced may be at the million level, so as a preferred method, the first N (N is a positive integer) keywords can be retained for each source. In addition, a PAD can be made as a fill-in feature. If the keyword cannot be found or established, PAD can be used instead. In step 1, FM (factor decomposition machine) and DeepFM algorithms can also be used as backup algorithms for constructing features. In addition, in step 1, sources can be increased or decreased according to actual needs. As some examples, keyword sources may include: i. Effect keywords: high-frequency keywords from product sources; ii. Long-tail keywords (long-tail keywords are needed to better guide users, which can better promote user effects): 1. Low-frequency keywords from product sources; 2. Product themes, high-frequency words in content; 3. High-frequency words from product source pages; 4. User search source keywords; 5. User-selected keywords for advertising; 6. Search page product source keywords; iii. Supplement: Statistical word frequency needs to be filtered according to the STOPWORD thesaurus.

[0068] Step 2: Count the number of sources and keywords used for each product, and use the proportion of different keywords for this product as the original feature value. In step 2, the standard normalization algorithm can be used to optimize the data distribution. In addition, the range normalization and mean normalization algorithms can also be used.

[0069] Step 3: Multiply the original feature values ​​from different sources by the source weights as the final feature values. In the early stage, the source weights can be set manually, and later they can be used as a neural network model based on the user's use effect, and trained again to optimize the weights.

[0070] Step 4: The entire matrix formed through the above steps 1-3 is used as the original keyword vector library.

[0071] At this point, a basic keyword vector library has been established, which can be optimized later based on actual conditions. The algorithm used is the collaborative filtering algorithm. In addition, you can also make a selection based on actual data and effects.

[0072] <Crowd vector library of initial database, crowd classification model of unsupervised classification model>

[0073] As described above, the initial database also includes a crowd vector library, and the unsupervised classification model includes a crowd classification model. Figure 8,The crowd vector library and crowd classification model can be constructed through the following steps.

[0074] Step 1: One-hot encoding (ONE-HOT) is used as a feature for all data sources such as products and product categories. In step 1, the feature can retain the TOP N keyword compression dimension. In addition, a PAD is made as a fill-in feature. If the feature cannot be found or is not established, PAD is used instead. Alternative algorithms: FM, DeepFM.

[0075] As an example, the latitude of a data source may include: i. users accessing N products in a recent period (the period can be defined, and data within 7-14 days is generally used); ii. categories of users accessing N products in a recent period (the period can be defined, and data within 7-14 days is generally used) (generally low weight); iii. IP region; iv. product delivery region; v. mobile phone model; vi. brand; vii. average monthly consumption amount; viii. shopping time period; ix. purchasing power: L1-L5; x. frequently visited time period: the most frequent shopping time period in the past 180 days, to determine the most frequently used browsing and purchasing time period for users, and the labels are displayed according to 24 hours a day; xi. whether logged in; xii. the interval between the last visit; xiii. the interval between uploads and purchases; xiv. labels for specific groups of people.

[0076] Step 2: Count the number of times each user visits products and product categories, and use the total number of products visited by users and the proportion of individual products as the original feature value. In step 2, the preferred algorithm is standard normalization to optimize data distribution. Alternative algorithms include range normalization and mean normalization.

[0077] In step 3, the original feature values ​​of different sources are multiplied by the source weights as the final feature values. In step 3, in the early stage, the source weights can be set manually, and later they can be used as a neural network model based on the effect of user use, and trained again to optimize the weights.

[0078] Step 4: Use the above data sources as a crowd vector library and save it. Step 5: Use K-MEANS and the above data sources as input to generate user groups, that is, crowd classification models. Step 6: Save the model and provide the crowd update program call.

[0079] <Crowd Attribute Model>

[0080] In some embodiments, a crowd attribute model can also be constructed using a crowd vector library. For example, the following steps can be used to construct (combined with Fig.12 ):

[0081] Step 1: Load training samples. Step 2: Adjust sample distribution to ensure different sample sizes.

[0082] In step 2, when the number of training samples is relatively small, the content can be increased by directly randomly copying the content of relatively small classification samples. If the number of sample sorting is relatively large, the more classification samples can be reduced, so that the multi-classification samples can be consistent with the sample data of small classifications.

[0083] Step 3: Load the crowd vector library as features, learn classification through deep learning, and verify the model.

[0084] In step 3, you can use neural network algorithms. You can also use decision tree, gradient boosted decision tree, random forest, logistic regression, support vector machine and other algorithms.

[0085] Step 4: Save the model to get the crowd attribute model. The crowd attribute model is used to predict the basic attributes of the crowd. This label is a common attribute of the crowd, such as age, gender, occupation, etc. This type of label uses supervised learning for learning prediction.

[0086] In addition to the above-mentioned general attributes of the crowd, you can also design a specific crowd label: This label is defined according to the existing business. There are clear definitions and algorithms for the data, which are generated according to the end-user's access behavior and purchase records. For example: Purchasing power label definition: L1-L5 is an index of consumption power from low to high, which cannot be directly matched to the consumption amount. This label uses an algorithm to combine multiple consumption behaviors to score users, and then ranks the scores from low to high, according to the total transaction amount of 5 equal parts. The first user who contributes 20% of the transaction amount is L1, and the user who contributes 20%-40% is L2, and so on.

[0087] <Product vector library of initial database, product classification model of unsupervised classification model>

[0088] As described above, the initial database also includes a product vector library, and the unsupervised classification model includes a product classification model. Fig. 9 ,The product vector library and product classification model can be constructed through the following steps.

[0089] Step 1: One-hot encoding (ONE-HOT) is performed on all data sources such as unsupervised learning population and labeled population classification as features.

[0090] Among them, unsupervised learning crowd and label crowd classification can be obtained from the crowd classification model and crowd attribute model.

[0091] In step 1, the features can retain the TOP N keyword compression dimension. In addition, a PAD is made as a fill-in feature. If the feature cannot be found or is not established, PAD is used instead. Alternative algorithms: FM, DeepFM.

[0092] As an example, the data source may include the following dimensions:

[0093] i. The number of visits and purchases of unsupervised learning crowd labels in the recent period (the period can be defined, generally data within 7-14 days is used).

[0094] ii. The number of visits and purchases by population tags in the recent period (the period can be defined, generally data within 7-14 days is used).

[0095] iii. Purchase products simultaneously (which can be defined as purchasing products within 24 hours).

[0096] Step 2: Count the data distribution of each tag in the total tags, and take the ratio of the total number of user visits and purchases to a single tag as the original feature value.

[0097] In step 2, the preferred algorithm is standard normalization to optimize data distribution. Alternative algorithms include range normalization and mean normalization.

[0098] Step 3: Multiply the original feature values ​​from different sources by the source weights to obtain the final feature values.

[0099] In step 3, in the initial stage, the source weights can be set manually, and later on, they can be used as a neural network model based on the user's usage effect and trained again to optimize the weights.

[0100] Step 4: Save the entire matrix formed above as a product vector library.

[0101] Step 5: Use K-MEANS and the above product vector library as input to generate product grouping, that is, product classification model.

[0102] Step 6, save the model and provide the crowd updater call.

[0103] <Keyword classification model of unsupervised classification model>

[0104] As described above, the unsupervised classification model includes a keyword classification model. According to one embodiment of the present invention, combined with Fig.10 ,The keyword classification model can be constructed through the following steps.

[0105] Step 1: Read the keyword vector library.

[0106] Step 2: Use K-MEANS and the keyword vector library as input to establish a grouping model, namely, a keyword classification model.

[0107] Step 3: Save the model and provide it for other programs to call.

[0108] So far, the embodiments of the present invention have constructed various vector libraries and classification models. Figure 3 , a complete build includes:

[0109] First, we build a keyword vector library, generate recommended keywords based on the keyword vector library, and build an unsupervised keyword classification model. We build a crowd vector library based on crowd labels.

[0110] According to the crowd vector library, an unsupervised learning crowd classification model and a crowd attribute model are constructed respectively. In some preferred embodiments, a keyword unsupervised crowd classification model can also be constructed in combination with keywords to improve the accuracy of the final model.

[0111] Based on the unsupervised learning crowd classification model and crowd attribute model, a product vector library is constructed.

[0112] According to the product vector library, an unsupervised learning product classification model is constructed to generate alternative products and related products. Taking related products as an example, products with high correlation coefficients can be calculated according to the product vector library. Different weights can be set for browsing products and purchasing products for optimization.

[0113] Finally, the unsupervised keyword classification model, the keyword unsupervised population classification model, the unsupervised learning population classification model, the unsupervised learning product classification model, and the population attribute model are integrated to construct a bidding click rate / conversion rate prediction model for returning the user generation probability of multiple advertising activities. As an example, the model can be loaded into the real-time reasoning service and called in the above step 205.

[0114] When constructing a bidding click rate / conversion rate prediction model, combine Fig.11 , you can also use the following steps to optimize the model:

[0115] Step 1: Collect and organize original feature data.

[0116] As an example, the original feature data may include the following dimensions: i. Features: advertising product ID, keyword unsupervised learning label, advertising product unsupervised learning label, advertising product category, advertising product label, advertising position, unsupervised learning population label, supervised learning population label, platform, region, time period, day of the week; ii. Label: whether clicked;

[0117] Step 2: Create a training dataset.

[0118] In step 2, the original feature data obtained in step 1 is used as a training data set and as an input of the model. In addition, it can be noted in step 2 that the ratio of click users and non-click users with different features must be 50% each.

[0119] Step 3: Build the model.

[0120] In step 3, you can use the naive Bayes algorithm and the neural network algorithm. Alternative algorithms include decision tree, gradient boosting decision tree, random forest, logistic regression, support vector machine, FM, DeepFM, GBDT, etc. You can choose according to the actual data and effect, combined with computing power.

[0121] Step 4: Save the model and use it in real-time bidding.

[0122] <Product Label>

[0123] As described above, in the above embodiments, product tags may be needed when constructing various vector libraries and classification models. As an example, product tags can be designed as follows: sku_number of this event; sku keyword extracted from name; sku keyword extracted from description information; product classification information; product classification information (parent classification); number of stars in product ratings; total number of product ratings; number of likes for product ratings; number of dislikes for product ratings; total number of 1-star products; total number of 2-star products; total number of 3-star products; total number of 4-star products; total number of 5-star products; product review keywords; and the transaction prices of the product's most recent 10 evaluations.

[0124] <user tag>

[0125] As described above, in the above embodiments, user tags may be required when constructing various vector libraries and classification models. As an example, user tags can be designed as follows: id (lock the sample to be unique, and the data source can be traced back through this id); user ID; whether to click; age; gender; marriage: default, 0 married, 1 unmarried, 2 divorced, 3; whether to have children: default, 0 no, 1 yes, 2; income level / year; time of event occurrence, hour / day; day / month; week / month; Monday is 1, Sunday is 7; the last transaction, how many days from now; transaction frequency in the last natural month; transaction frequency in the current year; transaction frequency in the last 30 days; transaction frequency in the last 365 days; the amount of the most recent transaction; the average amount of the last 10 transactions; the city where the user is located (transaction user address-->order address-->billing address); country; device type; network type; transaction channel; recently browsed Product sku_number; Number of items added by the user (starting from the last time the shopping cart was emptied); The sku_number that the user has recently added to the cart; The time the user last logged in, in days; The duration of the user's last login, in seconds; The user's most recent search keyword; The user's ipv, in the last 1 day, 3 days, 7 days, 14 days, 30 days; The user's pv, in the last 1 day, 3 days, 7 days, 14 days, 30 days; The user's ipv for this sku, in the last 1 day, 3 days, 7 days, 14 days, 30 days; The user's pv for this sku, in the last 1 day, 3 days, 7 days, 14 days, 30 days; AIPL classification: default, 0A, 1I, 2P, 3L, 4; RFM classification: default, 0 important value customers, 1 general value customers, 2 important development customers, 3 general development customers, 4 important retained customers, 5 general retained customers, 6 important retained customers, 7 general retained customers, 8; Coupon users, default, 0 No, 0 Yes, 1.

[0126] <Recommended price range algorithm> (maximum and minimum values ​​are calculated offline)

[0127] The recommended price range algorithm is used to return the price range required for delivery based on the product / category / keyword selected by the user.

[0128] In one embodiment, the recommended price range algorithm includes an interface return process and a maximum and minimum general process for calculating.

[0129] Optional, combined Figure 4 , the interface return process includes:

[0130] Step 1, obtaining at least one of the product, category, and keyword data input by the user. In some embodiments, all of the above data are obtained.

[0131] Step 2: Select the corresponding targeted products or keywords.

[0132] Since the server system already has relevant products, categories, and locations as well as their historical data in the last N days (N can be configured to a reasonable value using parameters), it can be targeted based on the products, categories, and keywords entered by the user.

[0133] Step 3, obtain the product price range within the corresponding conditions, return the result if there is data, otherwise go to step 4.

[0134] Step 4, obtain the price range of the user's product keyword, return the result if there is data, otherwise go to step 5.

[0135] Step 5, get the product category price range, return the result if there is data, and return the default price range if there is no data.

[0136] The returned result includes the minimum and maximum recommended prices. In addition, the system can also verify the minimum price. If the minimum price is lower than the currently set low price, the system low price will prevail.

[0137] Optionally, the returned result can also include the average and standard deviation of the price, so that the calculation logic is more flexible at the interface level. Figure 5 , the mean and standard deviation of the price can be calculated as follows:

[0138] Step 1: Enter the LOG number (i.e. transaction price) of the last N days.

[0139] In step 1, if the LOG is too large, the data can be compressed, such as reducing it to the average transaction price of the period.

[0140] Step 2, calculate the average and standard deviation of the input transaction price. Step 3, use the 3 sigma rule to filter abnormal data. Step 4, calculate the average and standard deviation of the filtered LOG. In step 4, a neural network algorithm can be used. In addition, algorithms such as decision tree, gradient boosting decision tree, random forest, logistic regression, support vector machine, etc. can also be used.

[0141] Step 5, save the result and store it in the average value and standard deviation. In step 5, optionally, the average value plus or minus 2 standard deviations is generally used as the price range, of which 2 standard deviations can be configured. It can also be adjusted according to preference, such as if you want users to have a higher price, you can subtract one standard deviation and add 2 standard deviations as the recommended range.

[0142] <Configure oCPM traffic distribution algorithm>

[0143] In one embodiment, in combination Figure 7 ,The traffic distribution algorithm is configured through the following steps.

[0144] Step 1: construct raw data for the ad slot. Step 2: obtain the raw data of the sub-ads in the current ad slot.

[0145] In step 2, optionally, the raw data includes display rate, CTR, CPC (Cost per Click), CPM, conversion rate, conversion value, current allocation ratio, etc. The conversion value can be set by the user, or only the number of conversions or transaction price can be used. The time range for obtaining data is controlled according to the amount of data. Generally, when the data is relatively small, 10, 30, 60, or 120 minutes of data can be used for acquisition. For example, if 10 minutes of data is insufficient, 30 minutes of data can be used to supplement it, and so on.

[0146] Step 3: Calculate the current score according to the scoring criteria. The score here is the same as the genetic algorithm score.

[0147] 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 the process proceeds to step 3. If there is no current data of the ad slot, the system default data is used to supplement the data, and the process proceeds to step 3.

[0148] Step 4, construct genetic algorithm chromosomes. In step 4, you can use advertising materials as genes and use 0-100 as the gene value range. Step 5, set the genetic algorithm scoring standard. In step 5, calculate the final conversion value based on the traffic ratio allocated by the chromosome. The larger the value, the higher the score. If all gene values ​​are greater than 100, the score is 0. Prevent percentage overflow.

[0149] Step 6: After the genetic algorithm evolves for N generations, set the optimal chromosome. If the final percentage is not 100, scale it to 100. Step 7: Calculate the score of the scaled ratio and compare it with the original score. Use the ratio with the higher score as the final allocation ratio. Step 8: Save the result and provide it to the ad server.

[0150] At this point, the above genetic algorithm has been used to complete the calculation of the optimal effect of the traffic share in the ad slot. After ABtest testing, the optimized CTR is increased by 20% compared with the ordinary traffic allocation algorithm. As an example, the score calculated by the traffic allocation algorithm can be stored in the traffic allocation cache as a traffic allocation coefficient, which is called by the server in the above step 206. The traffic weight (i.e., traffic weight distribution) of each advertisement in the ad queue obtained in step 205 is determined according to the coefficient, thereby further accurately correcting the order of the ad queue.

[0151] <Configure anti-cheat algorithm>

[0152] In some embodiments, the anti-cheating algorithm is configured by the following process (combined with Fig.13) Step 1: Load the general blacklist library. Step 2: Verify whether the traffic belongs to the blacklist. If it belongs to the blacklist, add the general traffic suspicion flag to the LOG and proceed to the next process. If it does not belong to the blacklist, go to Step 3. Step 3: Verify whether it is a click behavior blacklist. Of course, this Step 3 can also be started by loading the click behavior blacklist library. If it belongs to the click behavior blacklist, add the general traffic suspicion flag to the LOG and proceed to the next process.

[0153] In the above process, the general blacklist library can be configured through the following process (combined with Fig.14 ) : Step 1: Obtain the original log. Step 2: Split the log by time. The splitting rule is based on the data volume and business requirements. In Step 2, when the data volume is large, the log can be split by time period. When the data volume is too large, it can be further split according to the page type. Step 3: Statistically calculate the required statistical values for each group, such as the number of accesses of UA, IP, device ID, etc. Step 4: Use the 3-sigma rule to filter out the outliers. Step 5: Store the outliers in the blacklist to obtain the general blacklist library. Step 6: Clean the automatically maintained blacklist. For example, the blacklist is valid for 1 - 14 days and is automatically updated when it expires.

[0154] In the above process, the click behavior blacklist library can be configured through the following process (combined with Fig.15 ) :

[0155] Step 1: Obtain the user click behavior data. Step 2: Split and statistically calculate the data of the log by time. The splitting rule is based on the data volume and business requirements. In Step 2, when the data volume is large, the log can be split by time period. When the data volume is too large, it can be further split according to the page type. Step 3: Group according to UA, IP, and device IP, and statistically calculate the average access depth and average stay time after the user clicks. Step 4: Use the 3-sigma rule to filter out the outliers. Step 5: Store the outliers in the blacklist to obtain the click behavior blacklist library. Step 6: Clean the automatically maintained blacklist. For example, the blacklist is valid for 1 - 14 days and is automatically updated when it expires.

[0156] Anti-cheat traffic filtering is mainly divided into two types of blacklists: the general blacklist and the click behavior blacklist. When a user in the blacklist is detected, add the corresponding suspicious traffic flag to the LOG. The general blacklist is used to clean up extreme data to ensure the reliability of the data source, thereby better establishing various models. It includes filtering out abnormal users such as users with extremely high accesses and crawlers. Click behavior blacklist: used to mark abnormal click behaviors, such as intensive clicks and malicious clicks.

[0157] <AB test configuration>

[0158] For the actual effect of the constructed model or algorithm strategy, in addition to the evaluation of the algorithm itself in the offline stage, AB experiments should also be conducted in the production environment (online environment) to ensure that the design is positive or meaningful in terms of optimization goals.

[0159] In some embodiments, user traffic needs to be evenly divided using a hash method to ensure consistent data distribution in each partition.

[0160] The uniformity will be different when different hash algorithms are selected, but the difference is not significant. This embodiment uses MurMurHash.

[0161] In order to avoid the error of evaluating the split separately, this embodiment adopts the split method of AABB (combined with Fig.16 ), A is the control group (control group A1, A2), B is the experimental group (experimental group B1, B2), of the total traffic, 90% of users are allocated to the control group, and 10% are allocated to the experimental group. The results are evaluated based on the statistical values ​​of the control group and the experimental group.

[0162] The embodiment of the present invention also provides a real-time bidding ranking system for improving the conversion effect of digital mall advertisements, including: a judgment module, used to judge whether it is a page turning request, if it is not a page turning request, then enter the advertisement ranking process; an advertisement ranking process module, used to execute the advertisement ranking process, the advertisement ranking process includes: obtaining basic traffic information; filtering related products according to the search keywords and filtering conditions in the basic traffic information, generating an initial advertisement queue; further filtering the advertisement queue according to the directional conditions; calculating the basic bid of the advertisements in the queue; pushing the advertisement queue into the real-time reasoning service, and calculating the click-through rate and / or conversion rate of the advertisement in real time through the model; reading the traffic distribution coefficient, calculating the chance of the advertisements in the queue, forming the final ranking, and outputting the advertisement queue; an output module, used to output the ranked first advertisement and pop up the advertisement from the queue. It should be understood that the above-mentioned various program modules have a one-to-one correspondence with the various steps described in the method embodiment, and the technical scheme described in the method embodiment can also be applied to the specific configuration of each program module. In order to avoid repetition, it will not be repeated here.

[0163] In summary, the bidding service comprehensively evaluates the data performance of the advertisement itself, traffic characteristics, product characteristics, user behavior characteristics, etc. to comprehensively estimate the advertising effect. It is a service process designed specifically for the traffic scenarios within the digital mall. It can automatically provide mall advertisers with visible improvements in marketing effects without manual intervention.

[0164] The present invention has the following characteristics: it considers the data assets of the digital mall as a whole, including transaction data, user behavior data, search data, user characteristics, and product characteristics, explores the value of data, and effectively combines them to provide support for marketing behavior; it sorts out business data one by one and forms outputs; it introduces real-time bidding technology for digital advertising to maximize the value of traffic; advertising material (product material) sorting is the most effective method to improve conversion, and the present invention uses scientific calculations and continuous optimization based on effects and data changes; it has a set of effect testing and evaluation mechanisms to rationally evaluate the actual effects of bidding methods; it has an anti-cheating mechanism to clean invalid data and malicious behaviors; and it performs the above actions under the premise of protecting user privacy.

[0165] Fig.17 A block diagram of a computing device 600 capable of implementing multiple embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, and the components shown herein, their connections and relationships, and their functions are merely exemplary and are not intended to limit implementations of the present inventions described and / or claimed herein.

[0166] like Fig.17 As shown, the device 600 includes a computing unit 601, which 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 computing 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 the bus 604.

[0167] 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, etc.; an output unit 607, such as various types of displays, speakers, etc.; a storage unit 608, such as a disk, an optical disk, etc.; and a communication unit 609, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 609 allows the device 600 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0168] The invention concept is described in detail using specific examples herein, and the description of the above embodiments is only used to help understand the core idea of ​​the invention. It should be pointed out that any obvious modification, equivalent substitution or other improvement made by a person of ordinary skill in the art without departing from the invention concept should be included in the protection scope of the present invention.

Claims

1. A real-time bidding ranking method for improving the conversion effect of digital mall advertisements, characterized in that: include: Determine whether it is a page turning request. If not, enter the advertisement sorting process; The advertisement sorting process includes: Get basic traffic information; Filter related products based on the search keywords and filter conditions in the basic traffic information to generate an initial advertising queue; Further filter the ad queue based on targeting criteria; Calculate the base bid for ads in the queue; Push the ad queue to the real-time inference service and use the model to calculate the click-through rate and / or conversion rate of the ad in real time; Read the traffic distribution coefficient, calculate the chance of ads appearing in the queue, form the final sorting, and output the ad queue; Output the first ranked advertisement and pop it from the queue; The real-time bidding calculation engine called in the advertisement ranking process is integrated in the following way: Constructing an initial database; the initial database includes a keyword vector library, a product vector library and a crowd vector library; Constructing an unsupervised classification model; the unsupervised classification model includes a keyword classification model, a product classification model, and a crowd classification model; Configure key algorithms; the key algorithms include keyword recommendation and / or matching algorithms, click-through rate and / or conversion rate estimation algorithms, traffic allocation algorithms, and price range recommendation algorithms; The configuration of the key algorithm also includes configuring an anti-cheating algorithm, wherein the anti-cheating algorithm includes cleaning extreme data according to a general blacklist and marking abnormal click behaviors according to a click behavior blacklist; The constructing of the initial database and the constructing of the unsupervised classification model include: First, we build a keyword vector library, generate recommended keywords based on the keyword vector library, and build an unsupervised keyword classification model; we build a crowd vector library based on crowd labels; According to the crowd vector library, unsupervised learning crowd classification model and crowd attribute model are constructed respectively; Build a product vector library based on unsupervised learning crowd classification model and crowd attribute model; Based on the product vector library, an unsupervised learning product classification model is constructed; Finally, the unsupervised keyword classification model, the unsupervised learning crowd classification model, the unsupervised learning product classification model, and the crowd attribute model are integrated to build a bidding click rate and / or conversion rate prediction model, which is used to return the user generation probability predicted by multiple advertising campaigns; The configuration of the traffic distribution algorithm includes: Get the original data of the current ad slot; Calculate the current score according to the scoring criteria; Construct genetic algorithm chromosomes; use advertising materials as genes and 0-100 as the gene value range; Set the genetic algorithm scoring criteria; calculate the final conversion value based on the flow ratio allocated to the chromosome. The greater the value, the higher the score. If all gene values ​​are greater than 100, the score is 0 to prevent percentage overflow; After the genetic algorithm evolves for N generations, the optimal chromosome is obtained. The score is calculated based on the obtained optimal chromosome and compared with the original score, and the proportion with a high score is used as the final allocation proportion; Providing the final allocation ratio to the ad server; When building a bidding click rate / conversion rate prediction model, the following steps are used to optimize the model: Step 1, collect and organize original feature data; The original feature data includes the following dimensions: i. Features: ad product ID, keyword unsupervised learning label, ad product unsupervised learning label, ad product category, ad product label, ad position, unsupervised learning population label, supervised learning population label, platform, region, time period, day of the week; ii. Label: whether clicked; Step 2, establish a training data set; In step 2, the original feature data obtained in step 1 is used as a training data set and as the input of the model. In step 2, it is necessary to ensure that the ratio of click users and non-click users with different features is 50% each; Step 3, build the model; In step 3, the naive Bayes algorithm and the neural network algorithm are used. The alternative algorithms include decision tree, gradient boosting decision tree, random forest, logistic regression, support vector machine, FM, DeepFM, GBDT; Step 4: Save the model and use it in real-time bidding.

2. The real-time bidding ranking method for improving the conversion effect of digital mall advertisements according to claim 1 is characterized in that: AB testing is also used to verify the actual effect of the constructed model or algorithm strategy.

3. The real-time bidding ranking method for improving the conversion effect of digital mall advertisements according to claim 1 is characterized in that: The configuration method of the price range recommendation algorithm includes: Obtain at least one data of product, category, and keyword input by the user; Select products or keywords that correspond to the targeting range; Get the product price range within the corresponding conditions. If there is data, the result is returned. If there is no data, then: Get the price range of the user's product keyword. If there is data, return the result. If there is no data, then: Get the price range of product categories. If there is data, the result will be returned. If there is no data, the default price range will be returned.

4. A real-time bidding ranking system for improving the conversion effect of digital mall advertisements, characterized in that: include: The judgment module is used to judge whether it is a page turning request. If it is not a page turning request, the advertisement sorting process is entered; The advertisement sorting process module is used to execute the advertisement sorting process, and the advertisement sorting process includes: Get basic traffic information; Filter related products based on the search keywords and filter conditions in the basic traffic information to generate an initial advertising queue; Further filter the ad queue based on targeting criteria; Calculate the base bid for ads in the queue; Push the ad queue to the real-time inference service and use the model to calculate the click-through rate and / or conversion rate of the ad in real time; Read the traffic distribution coefficient, calculate the chance of ads appearing in the queue, form the final sorting, and output the ad queue; An output module, used for outputting the advertisement ranked first and ejecting the advertisement from the queue; The real-time bidding calculation engine called in the advertisement ranking process is integrated in the following way: Constructing an initial database; the initial database includes a keyword vector library, a product vector library and a crowd vector library; Constructing an unsupervised classification model; the unsupervised classification model includes a keyword classification model, a product classification model, and a crowd classification model; Configure key algorithms; the key algorithms include keyword recommendation and / or matching algorithms, click-through rate and / or conversion rate estimation algorithms, traffic allocation algorithms, and price range recommendation algorithms; The configuration of the key algorithm also includes configuring an anti-cheating algorithm, wherein the anti-cheating algorithm includes cleaning extreme data according to a general blacklist and marking abnormal click behaviors according to a click behavior blacklist; The constructing of the initial database and the constructing of the unsupervised classification model include: First, we build a keyword vector library, generate recommended keywords based on the keyword vector library, and build an unsupervised keyword classification model; we build a crowd vector library based on crowd labels; According to the crowd vector library, unsupervised learning crowd classification model and crowd attribute model are constructed respectively; Build a product vector library based on unsupervised learning crowd classification model and crowd attribute model; Based on the product vector library, an unsupervised learning product classification model is constructed; Finally, the unsupervised keyword classification model, the unsupervised learning crowd classification model, the unsupervised learning product classification model, and the crowd attribute model are integrated to build a bidding click rate and / or conversion rate prediction model, which is used to return the user generation probability predicted by multiple advertising campaigns; The configuration of the traffic distribution algorithm includes: Get the original data of the current ad slot; Calculate the current score according to the scoring criteria; Construct genetic algorithm chromosomes; use advertising materials as genes and 0-100 as the gene value range; Set the genetic algorithm scoring criteria; calculate the final conversion value based on the flow ratio allocated to the chromosome. The greater the value, the higher the score. If all gene values ​​are greater than 100, the score is 0 to prevent percentage overflow; After the genetic algorithm evolves for N generations, the optimal chromosome is obtained. The score is calculated based on the obtained optimal chromosome and compared with the original score, and the proportion with a high score is used as the final allocation proportion; Providing the final allocation ratio to the ad server; When building a bidding click rate / conversion rate prediction model, the following steps are used to optimize the model: Step 1, collect and organize original feature data; The original feature data includes the following dimensions: i. Features: ad product ID, keyword unsupervised learning label, ad product unsupervised learning label, ad product category, ad product label, ad position, unsupervised learning population label, supervised learning population label, platform, region, time period, day of the week; ii. Label: whether clicked; Step 2, establish a training data set; In step 2, the original feature data obtained in step 1 is used as a training data set and as the input of the model. In step 2, it is necessary to ensure that the ratio of click users and non-click users with different features is 50% each; Step 3, build the model; In step 3, the naive Bayes algorithm and the neural network algorithm are used. The alternative algorithms include decision tree, gradient boosting decision tree, random forest, logistic regression, support vector machine, FM, DeepFM, GBDT; Step 4: Save the model and use it in real-time bidding.

5. A storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method according to any one of claims 1 to 3 is implemented.

6. An electronic device, comprising: one or more processors; as well as A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 3.

Images