An outdoor advertising recommendation system and method based on large-scale user portraits

Abstract

The present invention discloses an outdoor advertising recommendation method based on large-scale user portraits, and belongs to the technical field of computational advertising. The advertising scheduling module in the outdoor advertising recommendation method of the present invention is respectively connected to the user end, the advertising management module, and the real-time and accurate crowd analysis module. The real-time and accurate crowd analysis module performs parameter interaction with the advertising retrieval module, the user portrait module, and the user behavior module respectively. The user data mining module is respectively connected to the advertising management module, the user behavior module, and the user portrait module, and the advertising management module is connected to the advertising retrieval module. The outdoor advertising recommendation method of the present invention completes the generation of a personalized advertising list based on the user portrait query of the user's interests and historical behavior records when the user logs into the platform, and finally pushes the advertisement to the outdoor advertising terminal. The present invention has good autonomous learning ability, can effectively improve the reach rate of outdoor smart advertising, and is suitable for outdoor advertising recommendation under the background of the Internet of Things and big data.

Description

Technical Field

[0001] The present invention relates to an advertisement recommendation method, in particular to an outdoor advertisement recommendation method based on large-scale user portraits, and belongs to the technical field of computational advertising. Background Art

[0002] With the rapid development of 5G technology and the Internet of Things technology, it provides technical support and development opportunities for the online, digital and intelligent outdoor media advertising. Outdoor advertising will also achieve self-innovation in the era of the Internet of Things and fully enter the data-driven era. At the same time, with the construction of modern smart cities, outdoor LED programmatic advertising has been given a new vitality of change by the Internet of Things technology. Offline commercial display screens can not only show the commercial vitality of the city, but also realize the function of traffic monetization for city operators.

[0003] In the era of big data, the combination of big data thinking and advertising is the trend of the future. Big data thinking has three important dimensions: quantitative thinking, correlation thinking and experimental thinking. Simply put, everything can be measured (providing a reliable description of things), everything can be related (the internal connection between different things), and everything can be tested (all solutions can be verified). The combination of these ways of thinking and advertising will bring about major changes in the advertising industry. The combination of online advertising and offline advertising is more effective. Although online advertising has developed rapidly, offline advertising can be more warm and often leaves a deeper impression on people, so the combination of online advertising and offline advertising will be a better solution. In the future, with the continuous deepening and development of "Internet +", the development trend of online and offline combined advertising will become more obvious.

[0004] At present, more and more advertisers are not satisfied with cheap advertising, and begin to require advertising agencies to explore traffic, activity, communication effects, etc. Under the new requirements of advertisers, the industry still stays in the traditional practice of using the declared traffic and vehicle flow data to estimate the impression volume (CPM) and using subjective judgment to evaluate the outdoor media viewable opportunity (OTS). This practice lacks data basis, so the traditional outdoor advertising model is not suitable for outdoor advertising recommendation in the context of big data. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an outdoor advertising recommendation method based on large-scale user portraits, which has good autonomous learning ability and can effectively improve the intelligence level of outdoor advertising recommendation. It is suitable for outdoor advertising recommendation in the current context of the Internet of Things and big data.

[0006] The present invention proposes an outdoor advertising recommendation system based on large-scale user portraits, comprising an advertising scheduling module, a real-time accurate crowd analysis module, an advertising retrieval module, a user behavior module, a user portrait module, an advertising management module and a user data mining module, wherein the advertising scheduling module is respectively connected with the real-time accurate crowd analysis module and the advertising management module of the user end; the real-time accurate crowd analysis module performs parameter interaction with the advertising retrieval module, the user behavior module and the user portrait module; the user data mining module is respectively connected with the user behavior module, the user portrait module and the advertising management module; and the advertising retrieval module is connected with the advertising management module;

[0007] The advertisement scheduling module performs real-time traffic analysis based on the real-time accurate crowd analysis module to complete the environmental guidance of the entire advertisement scheduling execution;

[0008] The real-time accurate crowd analysis module is used to complete the advertisement sorting, and return the sorted advertisement numbers and scores to the advertisement scheduling module;

[0009] The advertisement retrieval module performs advertisement data indexing by acquiring the label and feature parameter information transmitted by the real-time accurate crowd analysis module 2 and the data of the advertisement management module, and returns the hit advertisement list parameters to the real-time accurate crowd analysis module;

[0010] The user behavior module obtains the category labels provided by the user data mining module and the user portrait parameters transmitted by the real-time accurate crowd analysis module, completes the query of user behavior information, and returns the user information and strategy to the real-time accurate crowd analysis;

[0011] The user portrait module completes user portrait matching through the real-time information transmitted by the real-time accurate crowd analysis module, and returns the user information to the accurate crowd analysis module;

[0012] The advertisement management module is used to store the latest advertisement delivery strategy set and provide the data set to the advertisement retrieval module for use by the user data mining module and the advertisement scheduling module;

[0013] The user data mining module can obtain data from the advertising management module in real time, match users with user portraits, and analyze and predict user behaviors.

[0014] The present invention further defines the technical solution, in the aforementioned outdoor advertising recommendation system based on large-scale user portraits, the user data mining module includes three parts: a portrait update module, a strategy update module, and a behavior flow detection module. The portrait update module and the strategy update module realize the real-time update and concurrent use of portraits and strategies by online constructing an Hbase dynamic data storage area.

[0015] In the aforementioned outdoor advertising recommendation system based on large-scale user portraits, the behavior flow detection module uses the Flume distributed log collection system to receive real-time data, and recommends advertisements by establishing a ranking model based on CTR, viewing time, positive feedback and negative feedback.

[0016] A method for outdoor advertising recommendation based on large-scale user portraits includes the following steps:

[0017] S1: Collect user profile information and user behavior information through APP log data and business data;

[0018] S2: Preprocess the collected data. The input data format is libSVM, which is expressed as follows:

[0019] y index_1:value_1index_2:value_2…index_n:value_n,

[0020] Therefore, the data format of the feature vector is converted into the libSVM format to meet the input requirements of the FM model;

[0021] S3: In view of the fact that the information features of the user portrait information, user behavior information and advertising portrait information are not always continuous category values, in order to conform to the training data format of the model, the collected data is processed by One-hot encoding, which can be expressed as follows after One-hot encoding:

[0022]

[0023] Among them, u represents the user, a represents the advertisement, and h represents the historical behavior information;

[0024] S4: The FM model is used to model all the pairwise combinations between each pair of features, and then the personalized advertising recommendation list of the user is estimated. The second-order polynomial FM model can be expressed as:

[0025]

[0026] S5: Decompose the matrix W in S4 into W = V T V, where V = (V 1 ,V 2 ,…,V n ) T ,V i =(v i1 ,v i2 ,…,v ik ), matrix V TEach row of represents the correlation between a certain user and different features, and each row of the matrix V represents the correlation between a certain feature and different advertisements. The model can be expressed as:

[0027]

[0028] in So we have:

[0029]

[0030] S6: In order to perform offline verification of the recommendation model, the following definitions are made:

[0031] Precision: The ratio of the number of correctly classified positive samples to the number of samples determined as positive samples by the classifier. Here R(u) is equivalent to the positive sample of the model, expressed as:

[0032]

[0033] Among them, R(u) is the recommendation list obtained by the recommendation model, and T(u) is the user's behavior list in the actual scenario;

[0034] Recall rate: the ratio of the number of correctly classified positive samples to the number of true positive samples. Here T(u) is equivalent to the set of true positive samples, expressed as:

[0035]

[0036] Coverage: The ratio of the number of recommended items in the recommendation system to the total number of items. For the same coverage, the quantity distribution or popularity of different advertisements may be different. In order to better describe the ability of the recommendation system to mine the long tail, it is necessary to count the number of times different advertisements appear. Information entropy is used to define coverage:

[0037]

[0038] Where p(i) is the popularity of ad i divided by the sum of the popularity of all ads;

[0039] Diversity: Measures the dissimilarity between all ads in the recommendation list. Use different similarity functions to measure the similarity of ads in the recommendation list, such as content-based similarity and collaborative filtering-based similarity. This way, you can get diversity from different angles. Specifically, it can be expressed as:

[0040]

[0041] Among them, s(i,j) is the similarity between advertisement i and advertisement j;

[0042] The overall diversity of the recommendation system can be defined as the average diversity of all users' recommendation lists:

[0043]

[0044] Novelty: The simplest method is to recommend ads that users have never seen before. However, the number of ads that each user has never seen is huge, so the average popularity of recommended ads is calculated. The lower the popularity, the more likely it is that the user will find the ad novel. Therefore, if the average popularity of the products in the recommended results is relatively low, it means that the recommended results are relatively novel.

[0045] S7: Filter out the model with the best performance in offline verification, conduct online A / B testing, and evaluate the effectiveness of online advertising recommendations.

[0046] A method for recommending outdoor advertisements based on large-scale user portraits, wherein the user portrait information includes basic user information, user social data, user consumption characteristics, and user historical purchase data; the user behavior information includes mobile user behavior information and device user behavior information; the mobile user behavior information includes online viewing, skipping, sharing, and commenting on advertisements, online advertisement viewing time, and online advertisement marking status; the device user behavior information includes offline device viewing, skipping, and offline device advertisement marking status.

[0047] Furthermore, in the aforementioned outdoor advertising recommendation method based on large-scale user portraits, the FM model uses the Adadelta algorithm to adaptively adjust the learning rate, make larger updates to uncommon parameters, and make smaller updates to common parameters, and accumulate fixed-size items without directly storing these items or setting default values.

[0048] Beneficial effects of the invention: Compared with the prior art, the invention has the following advantages:

[0049] 1. The present invention maps the problem of outdoor advertising based on large-scale user portraits into the problem of behavior prediction of large-scale online data through historical behavior information. By using a highly accurate integrated model, the category prediction of user expected behavior is performed, which improves the reach and intelligence level of the outdoor advertising recommendation system. In addition, a data storage structure and corresponding prediction algorithm dedicated to the system integration model are designed. The present invention can be applied to outdoor advertising recommendation in the context of the Internet of Things and big data.

[0050] 2. Based on the established ranking model, this method adopts a sublinear online prediction method. Compared with the traditional linear prediction method, the Embedding layer of the trained ranking model is extracted, the user features are vectorized in sequence, and then all features are weighted averagely to form the user's final Embedding; similarly, the same operation is performed on the advertisement. Then, the vector retrieval tool Faiss is used to establish an index to build an advertising recommendation list, which greatly improves the prediction speed. The prediction time is only 3% of the traditional method, which better copes with the challenges of large-scale data processing.

[0051] 3. This method separates the businesses with low real-time requirements from the real-time business system, solves the conflict between deep mining of user data through offline data analysis and real-time online service needs, alleviates the pressure on the real-time business system, and ensures the accuracy of data analysis in the process of pushing services to users in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and the embodiments in the drawings do not constitute any limitation to the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 Schematic diagram of the recommended method for outdoor advertising of the present invention.

[0054] Figure 2 Schematic diagram of the data structure of the FM model in this embodiment.

[0055] Figure 3 Schematic diagram of matrix conversion in this embodiment.

[0056] Figure 4 This is a flow chart of the outdoor advertising method of this embodiment.

[0057] In the figure: 1. Advertising scheduling module, 2. Real-time and accurate crowd analysis module, 3. Advertising retrieval module, 4. User behavior module, 5. User portrait module, 6. Advertising management module, 7. User data mining module. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, which are preferred embodiments of the present invention. It should be understood that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments; it should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0060] The embodiment of the present invention provides an outdoor advertising recommendation method for large-scale user portraits. Figure 1 As shown, it includes an advertisement scheduling module 1, a real-time accurate crowd analysis module 2, an advertisement retrieval module 3, a user behavior module 4, a user portrait module 5, an advertisement management module 6 and a user data mining module 7. Among them, the advertisement scheduling module 1 is connected with the real-time accurate crowd analysis module 2, the user end and the advertisement management module 6 respectively. The real-time accurate crowd analysis module 2 performs parameter interaction with the advertisement retrieval module 3, the user behavior module 4 and the user portrait module 5 respectively. The user data mining module 7 is connected with the user behavior module 4, the user portrait module 5 and the advertisement management module 6 respectively. The advertisement retrieval module 3 is connected with the advertisement management module 6.

[0061] Among them, the advertisement scheduling module 1 performs real-time traffic analysis according to the real-time accurate crowd analysis module to complete the environmental guidance of the entire advertisement scheduling execution. The real-time accurate crowd analysis module 2 is used to complete the advertisement sorting, and returns the sorted advertisement number and scoring to the advertisement scheduling module 1. The advertisement retrieval module 3 indexes the advertisement data by obtaining the parameter information such as tags and features transmitted by the real-time accurate crowd analysis module 2 and the data of the advertisement management module 6, and returns the hit advertisement list parameters to the real-time accurate crowd analysis module 2. The user behavior module 4 obtains the category tags provided by the user data mining module 7 and the user portrait parameters transmitted by the real-time accurate crowd analysis module 2, completes the query of user behavior information, and returns the user information and strategy to the real-time accurate crowd analysis 2. The user portrait module 5 completes the user portrait matching through the real-time information transmitted by the real-time accurate crowd analysis module 2, and returns the user information to the implementation accurate crowd analysis module 2. The advertisement management module 6 is used to store the latest set of advertisement delivery strategies, and provide the data set to the advertisement retrieval module 3, the user data mining module 7 and the advertisement scheduling module 1 for use.

[0062] The user data mining module 7 can obtain the data of the advertising management module 6 in real time, and analyze and predict the user's profile matching and user behavior. It specifically includes three parts: a profile update module, a strategy update module, and a behavior stream detection module. The profile update module and the strategy update module realize the real-time update and concurrent use of profiles and strategies by constructing an Hbase dynamic data storage area online. The behavior stream detection module uses the Flume distributed log collection system to receive real-time data, and recommends advertisements by establishing a sorting model based on CTR, viewing time, positive feedback, and negative feedback. As a result, each user will get a set of candidate advertisements, which will be stored using Redis to form a high-performance database in a keyword format, where the key represents a user and the value represents the user's list of advertisements to be recommended. Figure 4 As shown, this embodiment discloses an outdoor advertising recommendation method based on large-scale user portraits, such as Figure 4 As shown, the steps are as follows:

[0063] S1: Collect user portrait information and user behavior information through APP log data and business data.

[0064] S2: User portrait information includes basic user information, user social data, user consumption characteristics, and user historical purchase data; user behavior information includes mobile user behavior information and device user behavior information; mobile user behavior information includes online completion, skipping, sharing, and commenting on ads, online ad viewing time, and online ad marking status; device user behavior information includes offline device completion, skipping, and offline device ad marking status.

[0065] S3: For data preprocessing, the input data format must be libSVM, that is, expressed as follows:

[0066] y index_1:value_1index_2:value_2…index_n:value_n

[0067] Therefore, the data format of the feature vector is converted to libSVM format to meet the input requirements of the FM model.

[0068] S4: Considering that information features such as user profile information, user behavior information, and advertising profile information are not always continuous, and are mostly categorical values, in order to better conform to the training data format of the model, one-hot encoding is required, such as Figure 2 As shown. After one-hot encoding, it can be expressed as:

[0069]

[0070] Among them, u represents the user, a represents the advertisement, and h represents the historical behavior information.

[0071] S5: The FM model estimates the user's personalized advertising recommendation list by modeling all pairwise combinations between each pair of features. The second-order polynomial FM model can be expressed as:

[0072]

[0073] S6: Decompose the matrix W in S5 into W = V T V, where V = (V 1 , V 2 , …, V n ) T , V i =(v i1 , v i2 , …, v ik ). Matrix V T Each row of represents the correlation between a certain user and different features, and each row of the matrix V represents the correlation between a certain feature and different advertisements, such as Figure 3 As shown; the model can be expressed as:

[0074]

[0075] in So we have:

[0076]

[0077] S7:FM models usually use the SGD algorithm, but compared with the SGD algorithm, the Adadelta algorithm can adaptively adjust the learning rate. This method can make larger updates to uncommon parameters and smaller updates to common parameters, and it accumulates fixed-size items without directly storing them or setting default values.

[0078] S8: In order to perform offline verification of the recommendation model, the following definitions are made:

[0079] Precision: The ratio of the number of correctly classified positive samples to the number of samples determined as positive samples by the classifier, expressed as (here R(u) is equivalent to the positive sample of the model):

[0080]

[0081] Among them, R(u) is the recommendation list obtained by the recommendation model, and T(u) is the user's behavior list in the actual scenario.

[0082] Recall rate: The ratio of the number of correctly classified positive samples to the number of true positive samples, expressed as (here T(u) is equivalent to the set of true positive samples):

[0083]

[0084] Coverage: The ratio of the number of recommended product sets to the total number of product sets. For the same coverage, the number distribution (popularity) of different advertisements can be different. In order to better describe the ability of the recommendation system to mine the long tail, it is necessary to count the number of times different advertisements appear. This patent uses information entropy to define coverage:

[0085]

[0086] Here, p(i) is the popularity of ad i divided by the sum of the popularity of all ads.

[0087] Diversity: Measures the dissimilarity between all ads in the recommendation list. Different similarity functions can be used to measure the similarity of ads in the recommendation list, such as content-based similarity and collaborative filtering-based similarity, so that diversity from different angles can be obtained. Specifically expressed as:

[0088]

[0089] Among them, s(i,j) is the similarity between advertisement i and advertisement j;

[0090] The overall diversity of the recommendation system can be defined as the average diversity of all users' recommendation lists:

[0091]

[0092] Novelty: The simplest method is to recommend ads that users have never seen before. However, the number of ads that each user has never seen is huge, so the average popularity of recommended ads is generally calculated. The lower the popularity, the more likely it is that the user will find the ad novel. Therefore, if the average popularity of the products in the recommended results is relatively low, it means that the recommended results are relatively novel.

[0093] S9: Filter out the model with the best performance in offline verification, conduct online A / B testing, and evaluate the effectiveness of online advertising recommendations.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An outdoor advertising recommendation system based on large-scale user portraits, comprising an advertising scheduling module (1), a real-time accurate crowd analysis module (2), an advertising retrieval module (3), a user behavior module (4), a user portrait module (5), an advertising management module (6) and a user data mining module (7), characterized in that: The advertisement scheduling module (1) is respectively connected to the real-time accurate crowd analysis module (2) and the advertisement management module (6) of the user end; the real-time accurate crowd analysis module (2) performs parameter interaction with the advertisement retrieval module (3), the user behavior module (4) and the user portrait module (5); the user data mining module (7) is respectively connected to the user behavior module (4), the user portrait module (5) and the advertisement management module (6); and the advertisement retrieval module (3) is connected to the advertisement management module (6); The advertisement scheduling module (1) performs real-time traffic analysis based on the real-time accurate crowd analysis module to complete the environmental guidance of the entire advertisement scheduling execution; The real-time accurate crowd analysis module (2) is used to complete the advertisement sorting, and return the sorted advertisement numbers and scores to the advertisement scheduling module (1); The advertisement retrieval module (3) performs advertisement data indexing by acquiring the label and feature parameter information transmitted by the real-time accurate crowd analysis module (2) and the data of the advertisement management module (6), and returns the hit advertisement list parameters to the real-time accurate crowd analysis module (2); The user behavior module (4) obtains the category label provided by the user data mining module (7) and the user portrait parameter transmitted by the real-time accurate crowd analysis module (2), completes the query of user behavior information, and returns the user information and strategy to the real-time accurate crowd analysis module (2); The user portrait module (5) completes user portrait matching through the real-time information transmitted by the real-time accurate crowd analysis module (2), and returns the user information to the real-time accurate crowd analysis module (2); The advertisement management module (6) is used to store the latest advertisement delivery strategy set and provide the set to the advertisement retrieval module (3) for use by the user data mining module (7) and the advertisement scheduling module (1); The user data mining module (7) is capable of acquiring data from the advertising management module (6) in real time, and matching user portraits and analyzing and predicting user behaviors; the user data mining module (7) includes three parts: a portrait update module, a strategy update module, and a behavior flow detection module; the portrait update module and the strategy update module realize real-time updating and concurrent use of portraits and strategies by constructing an Hbase dynamic data storage area online; the behavior flow detection module receives real-time data using a Flume distributed log collection system, and recommends advertisements by establishing a ranking model based on CTR, viewing time, positive feedback, and negative feedback.

2. The outdoor advertising recommendation system based on large-scale user portraits according to claim 1 is characterized by: User portrait information includes basic user information, user social data, user consumption characteristics, and user historical purchase data; the user behavior information includes mobile user behavior information and device user behavior information; the mobile user behavior information includes online viewing, skipping, sharing, commenting on ads, online ad viewing time, and online ad marking status information; Device-side user behavior information includes offline device completion and ad skipping information, and offline device ad marking status information.

Images