CN106960359B

CN106960359B - Full-automatic bidding optimization method and system based on streaming calculation

Info

Publication number: CN106960359B
Application number: CN201710074933.9A
Authority: CN
Inventors: 阮备军; 朱建秋; 钱肖鲁
Original assignee: Shanghai Zhizi Information Technology Co ltd
Current assignee: Shanghai Zhizi Information Technology Co ltd
Priority date: 2017-02-10
Filing date: 2017-02-10
Publication date: 2021-09-03
Anticipated expiration: 2037-02-10
Also published as: CN106960359A

Abstract

The invention discloses an automatic bidding optimization method and system for real-time bidding advertisements (Real Time Bidding, RTB) in the field of Internet advertisements. After the creative materials, transaction channels, visitor value, basic needs and constraints are determined, there is no need for manual tuning work. It can automatically optimize the bidding strategy according to market fluctuations and maximize the advertiser's income. The present invention reads RTB logs (including advertisement exposure, clicks and certain agreed behaviors, such as registration, orders, etc.), and adopts a partition-based streaming calculation method to quickly quantify various factors that affect the value of advertisement exposure opportunities without going back to the history. data, nor the need to regularly collect training data, build and update predictive models. The bid function is able to incorporate the quantitative value and pricing weights of all factors. According to the information provided by the budget consumption rhythm control algorithm, it can automatically predict exposure opportunities for different value levels and formulate different pricing weights, so as to dynamically adjust the bidding price, optimize the advertiser's income, and realize automatic bidding.

Description

Full-automatic bidding optimization method and system based on streaming calculation

Technical Field

The invention relates to automatic Bidding strategy optimization of Real Time Bidding (RTB) in the field of internet advertisement, in particular to a method and a system for automatically evaluating the value and weight of an advertisement exposure opportunity and automatically optimizing the yield of an advertiser according to a limited advertisement budget by adopting a streaming computing technology.

Background

Programmed advertisement purchasing depends on a data analysis technology, audiences are accurately positioned, automatic advertisement putting, evaluation and optimization can be realized on the basis, and the method becomes an important trend of online advertisements at home and abroad. Real-time bid advertising is the primary form of programmatic advertising. Through the RTB ad exchange, advertisers bid in real time for each ad exposure opportunity (presentation opportunity). Once the bid is successful, the advertisement of the successor is immediately presented to the current visitor.

A Demand Side Platform (DSP) solves the problem of how advertisers purchase ad exposure in real time at a exchange. The method can simplify the purchasing mode and improve the purchasing efficiency.

An efficient and easy-to-use DSP must be able to automatically optimize bidding strategies to minimize human involvement, with the goal of maximizing profitability (clicks, orders, or other actions) under specified constraints. To achieve this goal, it needs to be solved from the following two aspects:

a. a pricing mechanism: and automatically and quickly adjusting pricing according to the budget consumption condition and the value of the exposure opportunity, and pursuing the maximized income. b. Budget consumption control: the method can automatically smooth budget consumption according to market fluctuation, and cover as many exposure opportunities with high cost performance as possible from the time dimension.

The disadvantages of the current common methods of related research and systems are:

1. parameters of the estimation and optimization pricing mechanism depend on large-batch historical delivery data, and the reaction is insensitive. If the market and the budget fluctuate, the trial and estimation are repeated again, and the practical application value is lacked.

2. There is no solution to how to combine automatically, coordinating three key mechanisms: the advertisement exposure opportunity valuation, pricing and consumption rhythm control are optimized towards the direction of the maximization of the index together, and the full-automatic bidding is realized.

Drawings

FIG. 1 is an overall architecture of a bid engine for a Smart cloud Open DSP implemented according to the present invention, which is an example of one embodiment of the present invention. The engine structure comprises all processes forming a full-automatic bidding optimization mechanism.

FIG. 2 shows four major categories of factors that affect the value of an advertisement exposure opportunity. The present invention determines the price level based on the value of the exposure opportunity. The estimate of the exposure opportunity depends on the value of each type of influencing factor.

FIG. 3 illustrates that the present invention employs a budget consumption control process to guarantee consumption within each budget consumption window, satisfying the conditions: a. the total consumption does not exceed the budget total; b. the consumptions are distributed as evenly as possible within the window period. The whole calculation process is executed once every fixed time.

Fig. 4 shows a random traffic filtering sample method implemented by a version of the Open DSP. This is an example of an embodiment of the invention to aid in explaining the traffic filtering methods supported by the present invention.

Figure 5 shows the calculation of a bid for each campaign based on the current bid request and based on the bid function, for which the winner wins. This is the implementation process of a version of the Open DSP bidding engine. This is one example of an inventive bidding process that is used to assist in explaining the principles of the present invention.

FIG. 6A is a depiction of a pricing optimization process. It regularly adjusts pricing weights based on flow value and consumption rhythm control parameters, with the goal of maximizing activity revenue (clicks, registrations, or orders). This process is performed every fixed time.

FIG. 6B is a portion of a pricing optimization process that determines the magnitude of adjustment for each value segment based on the determined direction of weight adjustment and searches for weight adjusted end point value segments.

Disclosure of Invention

The solution is to solve the technical shortcomings of the related research and system. The method is suitable for flow calculation in the aspects of valuation, pricing and consumption rhythm control, the DSP can optimize while throwing, large-batch historical throwing data does not need to be accumulated in advance, and the speed of automatically adjusting the price according to market fluctuation by the DSP is improved; simultaneously, the scheme adopts a new technology, so that three modules of exposure opportunity valuation, pricing and consumption rhythm control work cooperatively together in the direction of index maximization, and the accuracy degree of the maximization index control is improved. In the following, we will explain the features and methods of the present invention with a bidding engine based on the smartphone cloud Open DSP product. The bidding engine of the Open DSP is only one embodiment of the present invention, and the technical details are only used to clearly and completely illustrate the features and methods of the present invention, but not to limit the implementation method of the present invention. All other embodiments, which can be derived by a person skilled in the art from the features and methods of the present invention, shall fall within the scope of protection of the present invention. The full scope of the invention is set forth in the appended claims.

Open DSP bidding engine

As an example of one embodiment of the present invention, FIG. 1 is an overall architecture of a version of the bidding engine of the Open DSP, where grey-marked ones do not belong to the bidding engine. 100 represents four main activity configurations entered into the bidding engine: target visitors, creative materials, activity requirements and constraints, and transaction channels. The database 125 stores a bid log from the bidding service 145 and logs from other systems, including but not limited to: click logs, web browsing logs, order logs, and the like. The bidding engine realizes the automatic optimization function based on three processes of various log feedbacks, combination and coordination, which respectively comprise: pricing optimization (115), consumption pacing control (130), and bidding service (145).

The partition valuation (105) calculates and evaluates various factors that affect the value of the exposure opportunity based on various logs, guest tags, or values of the feedback, and stores the calculated partition value (labeled seg _ v) in a partition value database (110), which then pushes seg _ v to the dynamic bid (150). An external system (such as a data management platform, DMP) defines tags for target guests, e.g., "shopping preferences → high-end handsets". The present invention allows each tag to be attached with a quantitative value, typically representing the high or low likelihood of a conversion or click.

The pricing optimization (115) periodically adjusts the pricing weights (denoted vsect _ w) based on parametric feedback from the partition value and consumption tempo control (130) of the database (110) with the goal of maximizing the campaign revenue (click-through, registration, or order), pushes into the stream value segments repository (120), and then synchronizes vsect _ w to the dynamic bids (150).

The process 140 makes real-time online statistics of the cost consumption of each activity, stores the statistics in a database (135), and provides the consumption rhythm control (130) for use. The process 130 calculates two control parameters based on the real-time consumption statistics: a flow selection probability P and a consumption proportion d. The former is used by the flow filter (155). The parameters P and d are also entered into the log database (125), and are statistically fed back to the pricing optimization (115). These statistical indicators are: 1. average flow selection probability avg _ P, 2, average consumption proportion avg _ d, 3, exposure value segment consumption proportion vsect _ c.

The bidding service (145) executes established bidding rules based on the input data and parameters. In the present invention, it is the executor of the bidding strategy and has no optimization capability. The present invention optimizes bidding strategies by adjusting the data and parameters entered into the bidding service. The bidding service includes two key processes: dynamic bidding (150) and traffic filtering (155). The process 150 parses and reads the corresponding segment value (seg _ w) and pricing weight (vsect _ w) to determine the price of the current exposure opportunity according to the bid function and the current bid request r. 155 filter the advertisement exposure opportunities according to the traffic selection probability P.

160 represents an RTB exchange. It sends an enquiry request r to the bidding service (145) which, after traffic filtering (155) and dynamic bidding (150), sends the ad creative and price to the exchange. The inquiry request r is a message sent to the DSP every time the advertisement exposure needs to be bid, the DSP must respond within a specified time, otherwise, the DSP is regarded as a disclaimer.

Before processing the inquiry request, the DSP must structure (parse and translate) the inquiry request to apply the bidding strategy to determine a reasonable price. The solicitation request may be structurally defined as:

<user_id,context_set>

the user _ ID is the ID of the guest. The Web ad is the ID stored in the Cookie and the mobile ad is the ID of the mobile device. context _ set is a set of name: value pairs. Context _ set, with a size of k >0, is defined as:

{name₁:value₁,name₂:value₂,….,name_i:value_i…,name_k,value_k}

name_iis the name of the data item in the context, such as geographic area name, media, ad slot, APP name, operating system, resolution, interface language, etc.; value_iIs a data item name_iThe value of (c).

2. Partition rules

Influencing factor

The present invention determines the price level based on the value of the exposure opportunity. The estimation of exposure opportunities is based on various classes of influencing factors. FIG. 2 is four categories of factors that affect the value of an exposure opportunity (240). These categories are explained as follows:

a. frequency (230): in addition to the number of ad exposures, other related activities such as clicks and purchases are included.

b. Medium (200): places where advertisements are shown, e.g., websites and APPs. Even to each area where an advertisement is displayed.

c. Exposure Context (210): information related to the current behavior of the visitor, such as geographic location, time, browser usage, device model, etc. In a broad sense, media is part of a Context. For convenience of description, the following describes single-column media, rather than attributing them to Context.

d. Population (220): the visitor to which the advertisement is directed.

Traffic zoning (AIS) and partitioning rule c

The traffic partition AIS (Ad expression Segment, partition for short) is a partition or classification of exposure opportunity traffic according to a certain factor. The partitioning by a certain factor allows traffic overlap to occur. The AIS division rule is not limited by the present invention, and may be determined by the implementer according to the specific situation. The use of a version of Open DSP is a common partitioning method:

a. people group

b. Media

c. Frequency of

d. Exposure Context

2. Partition estimation

One key feature of the present invention is to evaluate the value for each partition (AIS). The partition valuation (105) in FIG. 1 calculates and evaluates various factors that affect the value of the exposure opportunity based on various logs, guest tags, or values fed back, and stores the calculated partition value (labeled seg _ v) in the database (110). Although the invention is not limited to a particular process for quantifying value, it is limited

a. The quantization method must be implemented using a streaming incremental statistical method. Streaming refers to that only a log file or a data table needs to be scanned once; the incremental method is that when the value is calculated, only the newly added log is used, and the historical log does not need to be traced back.

b. The value of quantization must be normalized to a decimal number. Representing the magnitude of the increase or decrease compared to the current average level. For example, 0 means a level with the mean level, 0.5 means a 50% increase, and-0.5 means a 50% decrease.

The implementer may define revenue and consumption based on different needs. Common definitions are:

for a certain partition s, the partition estimation general rule for a certain version of Open DSP is:

recent revenue/recent consumption of s for seg _ v(s) ═ s

3. Real-time consumption statistics

The real-time expenditure statistics process (140) of fig. 1 performs real-time statistics of the cost expenditure data for each time slice of each activity, which is provided to the expenditure tempo control algorithm for use. The size of the time slice represents the budget control accuracy.

4. Consumption cadence control

The budget consumption control process (130) in fig. 1 aims to guarantee the consumption process in each budget consumption window, and the following conditions are met: a. the total consumption does not exceed the budget total; b. the consumptions are distributed as evenly as possible within the window period. The main function of budget consumption control is to calculate a traffic selection probability P, according to which the traffic filtering module of the bidding service will filter (or sample) the exposure opportunities. The time of delivery per day is divided evenly into a plurality of budget control windows, and the total budget is also distributed evenly to each budget control window. The system periodically calculates the ratio of the actual consumed budget per second to the predetermined consumed budget per second: the consumption ratio d is used for regulating P. The larger d is, the smaller the traffic selection probability P is; conversely, the greater the traffic selection probability P. The calculation process is defined as shown in fig. 3.

The whole calculation process is executed once every fixed time. The execution interval is the budget control window length/m, m being a positive integer. The larger m, the finer the control. The system automatically calculates 310 the entries for storage in memory prior to invocation. P_sIs the value of P currently used by the system, P_s+1Is the P value used for the next settlement period. 320 determining whether settlement has not started, and if so, adopting default P value and d value at 340; if not, the d and P values are updated with:

d ← (B × L)/(t × B) (formula 1)

P_s+1←MIN(P_sD,1) (equation 2)

Finally, the system logs all calculated values at 360 and bases P on_s+1And updating the P value adopted by the current budget control window.

5. Flow filtration

Fig. 3 consumes the parameters calculated by the rhythm control process, which are kept in the promotional activity data. The traffic filter (155) of FIG. 1 reads these parameters, filtering out excess ad exposure, and thus controlling budget consumption. The invention does not specify a specific filtering method, but only specifies that the total flow sampling ratio meets the requirement of the P value. There are generally two approaches: 1. filtering randomly according to the P value, and not considering the partition value; 2. and filtering the bias according to the P value and the value of the advertisement flow. FIG. 4 shows a random filtering sample approach implemented by a version of the Open DSP. If 460 returns FALSE, the system abandons the subsequent bid. In step 440, the system determines whether to filter the current inquiry request based on the current P value.

6. Dynamic bidding

Dynamic pricing is based on the value of the partition (AIS), the core of which is the bid function.

Exposure value prediction

Let r be a structured exposure request, the following is the exposure cost prediction function:

1) f is the number of influencing factors and i is some kind of influencing factor.

2) And seg _ id, resolving the bidding request r to obtain the partition identification of the i-type influence factor.

3) seg _ v, the value of a certain (traffic) partition for the class i impact factor is obtained.

4) β [ i ], is the weight of the class i influencer.

The present invention is not limited to the method for determining β [ i ]. In one implementation version of Open DSP, β [ i ] anchor is 1; in another implementation version, the beta [ i ] is periodically optimized by using historical delivery data and a linear regression method.

Exposure value segmentation

The predicted exposure value is segmented according to some monotonically increasing function (denoted vsect _ index). vsect _ index maps real numbers to a continuous integer interval. One implementation version of Open DSP employs [1, N]Integer intervals, using increasing sequences of length N<v¹，v²，…，v^N>Where any i > j, then vⁱ＞v^j. The definition of vsect _ index is as follows:

later we use the function vsect _ v (i) ═ vⁱTo represent the exposure value of retrieving the ith sector according to the index i.

Bid function

The bid function fuses the influence of all factors, and the calculated value is actually the product of the weights of the factors. Let r be a structured exposure request, the following is the definition of the uniform bid function bid _ fun.

MIN{base·v^r·vsect_w[vsect_index(v^r)]Max _ b } (equation 5)

Wherein

1) base is the base price and is a predetermined constant greater than 0.

2) max _ b is the maximum price, i.e., the price cap, greater than 0;

3)v^ris the flow value predictive value of r (equation 3).

4) vsect _ index is a piecewise function (equation 4).

5) vsect _ w is a pricing weight array, calculated by the pricing optimization process. The length is the maximum value N of the piecewise function. Bidding process

Bids are calculated for each campaign based on the current bid inquiry request and based on a bid function, and the winner wins. The implementation of a version of the Open DSP bidding engine is shown in FIG. 5. Where 540 returns a bid response that is a binary of < price, creative >, if the price is null (null) indicating that the DSP is not participating in bidding. Step 530 calls QUERY _ CAMPAGIN for each activity, which calculates the price for the planned ai according to equations 3, 4, 5, and then returns with the creative. The system selects the response with high price according to the price.

7. Pricing optimization

Fig. 6A is a calculation method for automatically adjusting the pricing weight array vsect _ w. The system is executed at regular intervals, typically at intervals that are multiples of the budget consumption control algorithm execution interval m. The parameters accepted in 601 of FIG. 6A are: min _ P is the minimum average flow selection probability; min _ d minimum average consumption proportional probability. Before each execution, three items of data must be counted at 605:

1, avg _ P: obtaining the average flow selection probability of the latest period from the log;

2, avg _ d: acquiring the average consumption proportion of the latest period from the log;

vsect _ c-the consumption fraction of each exposure value segment for the last period of time, i.e. the fraction of the total cost generated by the exposures belonging to this segment, is retrieved from the log.

If the condition 610 is met, it indicates that the system has selected most of the flow, but the average consumption rate is still low, requiring an increase in price level. If the condition of 610 is not met, then the condition of 615 needs to be detected. If the condition of 615 is met, it indicates that the average consumption proportion has reached a high point, but too much traffic has been filtered out, the overall bid is high, and the price level needs to be lowered. 620 calculating the magnitude of the boost, using the formula

g ← (min _ d-avg _ d)/min _ d (formula 6)

The result of the formula calculation is between 0 and 1. 625 calculating the magnitude of the drop, using the formula

g ← (min _ P-avg _ P)/min _ P (equation 7)

The result of the formula calculation is between 0 and 1. After the adjustment amplitude calculation is completed, the adjust _ vsect _ w process of fig. 6B is invoked. In step 630, the adjustment direction is to increase the weight from N to 1, and in step 635, the weight is decreased from 1 to N.

At 640 of fig. 6B, an adjustment magnitude g and an adjustment direction sp are received. A negative magnitude, according to the decision 645, represents a decreasing weight and a positive increasing weight. 650 sets the weights to adjust vsect _ w elements from low to high prediction weights, and 655 sets the weights to decrease vsect _ w elements from high to low.

660 of FIG. 6B attempts to search for an element from the beginning, with the current accumulated weight adjustment magnitude being exactly close to g, based on the set adjustment direction. Each search updates the weight values of the elements from the starting point until the current element K. If the weight is increased, the formula for adjusting the weight of the ith segment is as follows:

b (i) ← MIN [ vsect _ v (i)/vsect _ v (K) × z (i), L ] (equation 8)

W (i) ← w (i) × 1+ b (i) ] (formula 9)

Wherein L is the maximum amplitude; if the weight is decreased, the formula for adjusting the weight of the ith segment is as follows:

b (i) ← MIN [ vsect _ v (K)/vsect _ v (i) × z (i),1] (equation 10)

W (i) ← w (i) × [1-b (i) ] (formula 11)

Wherein before each search, W is a copy of vsect _ W, but after the computation is finished W is copied to vsect _ W.

Claims

1. a fully automatic bidding optimization method based on streaming computing, applied to the bidding of demand side platform, is characterized in that, comprises the steps:

a) The possible values of various factors that affect the value of advertising exposure opportunities are merged into multiple partitions, and each partition is attached with one or more value indicators; various factors include crowd tags, visitor exposure frequency, media, browsers , region and time period;

b) Regularly read the newly added bidding logs and feedback logs from other systems in the log database, and use the streaming method to count and update the value of each partition;

c) Based on the partition value of various factors that affect the advertisement exposure opportunity, predict the exposure value of the advertisement exposure opportunity, and segment the exposure value according to a monotonically increasing function;

d) Regularly adjust and calculate the traffic selection probability and consumption ratio required by the subsequent method according to the latest budget consumption situation and budget configuration, and save them to the log database;

e) use the traffic selection probability and consumption ratio logs saved in the log database in step d, and regularly adjust the pricing weight of the corresponding segment of the exposure value generated in step c;

f) According to the traffic selection probability calculated in step d, for each advertisement exposure opportunity, determine whether to participate in the bidding;

g) According to the pricing weight of the partition value and exposure value in steps c and e, calculate the price of the advertisement exposure opportunity online, and send the price and the creative to the RTB exchange as a bid reply; wherein, the RTB bid transaction feedback includes bid and exposure The bidding log is saved to the log database; the log database receives various feedback logs collected by the demand-side platform from other systems; among which, the cost consumption data of each time slice of each activity is counted in real time, recorded, and fed back to the step d.

2. The fully automatic bidding optimization method based on streaming computing according to claim 1, wherein the partition is to merge the possible values of each type of factors into a limited number of divisions or categories, and the merging rules are predefined and do not require Advertiser participation.

3 . The fully automatic bidding optimization method based on streaming computing according to claim 1 , wherein the value of the value index of the partition represents the range of improvement or reduction compared with the average value. 4 .

4. The fully automatic bidding optimization method based on stream computing according to claim 1, wherein the newly added log data is read regularly, and one or more statistical values related to the value index are calculated for each partition, Then update the value indicator.

5. the fully automatic bidding optimization method based on streaming computing according to claim 1, is characterized in that, predicts the value of advertisement exposure opportunity according to following formula:

where r is the advertisement inquiry request, F is the number of impact factors, and i is the i-th type of impact factor; the seg_id function analyzes r to obtain the partition ID of the i-th type of impact factor; seg_v obtains the partition value specified by the seg_id of the i-type impact factor; β [i] is the partition i influence weight.

6. The fully automatic bidding optimization method based on streaming computing according to claim 1, characterized in that, what is adopted for the segmentation of exposure value is a monotonically increasing function, and the exposure value of the real number type is mapped to an integer interval, and in the integer interval The integer is the index of the exposure value segment.

7 . The fully automatic bidding optimization method based on streaming computing according to claim 1 , wherein the daily delivery time is evenly divided into a plurality of budget control windows, and the total budget is also equally distributed to each budget control window. 8 .

8. The fully automatic bidding optimization method based on streaming computing according to claim 1, wherein the consumption ratio is the ratio of the current actual average consumption per second to the configured average per second consumption budget, and the traffic selection probability is the current The traffic selection probability used divided by the consumption ratio; if the calculation result is greater than 1, take 1.

9. The fully automatic bidding optimization method based on streaming computing according to claim 1, wherein each exposure value segment is equipped with a pricing weight and a consumption ratio, and the consumption ratio is a certain exposure value score in a recent period of time. The segment's ad impression cost as a percentage of the total ad impression cost.

10. The fully automatic bidding optimization method based on streaming computing according to claim 1, characterized in that, it is necessary to regularly count the average traffic selection probability and average consumption ratio in a recent period of time, and then adjust the pricing of exposure value segments according to the following rules Weights:

a) If avg_d is lower than min_d, but avg_P is not lower than min_P, increase the pricing weight of some exposure value segments. The total increase is: (min_d-avg_d)/min_d;

b) If avg_d is not lower than min_d, but avg_P is lower than min_P, reduce the pricing weight of some exposure value segments, and the total reduction is: (min_P-avg_P)/min_P;

Where avg_P is the average traffic selection probability obtained from the log in the recent period, min_P is the minimum threshold; avg_d is the average consumption ratio obtained from the log in the recent period, where min_d is the minimum threshold.

11. The fully automatic bidding optimization method based on streaming computing according to claim 1, characterized in that, according to the value of the exposure value segment from high to low, try to allocate the increased range until the total range defined in claim step a is allocated. ; The steps for assigning pricing weight to each exposure value segment are as follows:

a)b(i)=MIN[vsect_v(i)/vsect_v(K)*z(i),L];

b)W(i)←W(i)*[1+b(i)]

c)z(i-1)←z(i)-vsect_c(i)*b(i);

where z(i) is the increase in the pricing weight to be assigned to segment i, vsect_c is the consumption ratio of segments with exposure value; W(i) is the pricing weight of segment i with exposure value; vsect_v is the i-th segment The exposure value of ; L is the maximum increase, K is the current termination segment, and z(i) satisfies that there is no segment with an index less than K greater than 0.

12. The fully automatic bidding optimization method based on streaming computing according to claim 10, characterized in that, according to the value of the exposure value segment from low to high, try to allocate the reduced range until the total range defined in claim step b is allocated. , the pricing weight allocation steps for each exposure value segment are as follows:

a)b(i)=MIN[vsect_v(K)/vsect_v(i)*z(i),1];

b)W(i)←W(i)*[1-b(i)]

c)z(i-1)←z(i)-vsect_c(i)*b(i);

where z(i) is the increase in the pricing weight to be assigned to segment i; vsect_c is the consumption ratio of segments with exposure value; W(i) is the pricing weight of segment i with exposure value; vsect_v is the i-th segment The exposure value of , K is the current termination segment, satisfies that there is no segment with index greater than K whose z(i) is greater than 0.

13. The fully automatic bidding optimization method based on streaming computing according to claim 1, wherein the bidding price of each advertisement exposure opportunity is calculated by the following formula

MIN{base v ^r vsect_w[vsect_index(v ^r )], max_b}

calculation, where

1) base is the base price;

2) max_b is the maximum price;

3) v ^r is the exposure value of the inquiry request r calculated according to the method of claim 5;

4) vsect_index is a piecewise function of exposure value;

5) vsect_w is an array of pricing weights.

14. A system of a fully automatic bidding optimization method based on streaming computing, applied in a bidding system of a demand-side platform, characterized in that it comprises the following concurrent processes:

a) The log database receives various feedback logs collected by the demand-side platform from other systems; the system automatically divides the possible values of various factors that affect the value of advertising exposure opportunities according to predefined merging rules, and attaches one or Multiple value indicators; regularly read various new bidding logs and feedback logs from other systems in the log database, and use streaming methods to count and update the value of each partition; various factors include crowd tags, visitor exposure frequency, media , browser, region and time of day;

b) The system calculates the consumption ratio and the traffic selection probability according to the budget of average consumption per second and the historical data of cost consumption in the recent period of time, and records them in the log database;

c) Receive an inquiry request for advertising exposure opportunities from the advertising exchange center, select the advertising exposure opportunities according to the traffic selection probability, and then predict the exposure value of each advertising exposure opportunity according to the partition value of various factors that affect the advertising exposure opportunities; According to the value of each partition in the inquiry request and the pricing weight of the exposure value, calculate the price and participate in the bidding, send the price and idea to the RTB trading platform as a bid reply, and save the bidding log including the bid and exposure of the RTB transaction feedback. Go to the log database, and count the cost consumption data of each time slice of each activity, record it, and feed it back to process b;

d) The system automatically segments the predicted exposure value, reads the log of the traffic selection probability and consumption ratio generated by process b from the log database, counts the average traffic selection probability and average consumption ratio in the recent period, and adjusts the exposure value. Pricing weight for each segment;

Among them, c is the real-time calculation process, and the processes a, b and d are executed regularly at small intervals.

15. The system of the fully automatic bidding optimization method based on streaming computing according to claim 14, wherein the consumption ratio is equal to the ratio of the current actual average consumption per second budget to the predetermined average per second consumption budget, and the traffic selection probability is The currently used traffic selection probability divided by the consumption ratio.

16 . The system of the automatic bidding optimization method based on streaming computing according to claim 14 , wherein the exposure value of the advertisement exposure opportunity is calculated. 17 .

17. The system of automatic bidding optimization method based on streaming computing according to claim 14, characterized in that, after finding the pricing weight of the segment to which it belongs, the price is calculated by "basic price x pricing weight x exposure value".

18 . The system for the automatic bidding optimization method based on streaming computing according to claim 14 , wherein the overall increase or decrease range of the pricing weight is determined according to the average traffic selection probability and average consumption ratio in a recent period of time. 19 .

19. The system of the fully automatic bidding optimization method based on streaming computing according to claim 14, characterized in that, if the pricing weight needs to be increased, the range of increase is allocated from high to low according to the value of the exposure value segment, until the total range is reached. The allocation is completed; otherwise, the value of the exposure value segment will try to allocate the reduction range from low to high.