Advertising data analysis system based on multi-resolution exponential weighting

Abstract

The present invention provides an advertising data analysis system based on multi-resolution exponential weighting, which performs binning operations on the estimated achievement data caused by advertising feature data, and intelligently adjusts the width of the bin interval during the historical evolution process, thereby automatically and intelligently adapting to the historical evolution of advertising and conforming to the current advertising situation, and on this basis, combining the cost data to calculate the probability of successful delivery of any advertisement, and effectively solves the technical difficulty of unpredictable advertising data that cannot be avoided in the field of advertising technology with the idea of ​​probability.

Description

Technical Field

[0001] The present invention relates to the fields of advertising and big data artificial intelligence, and in particular to an advertising data analysis system based on multi-resolution index weighting. Background Art

[0002] Currently, achieving precise advertising delivery is extremely difficult in the advertising sector. Technically, the key difficulty lies in the highly unpredictable nature of advertising data. It's virtually impossible to predict future advertising data based on existing data using specific functional patterns.

[0003] In other words, the current technical limitations of advertising data analysis have greatly reduced the accuracy of advertising delivery.

[0004] To this end, the most primitive solution is to simply pile up manpower, use the advantage of many people to constantly observe the changes in past advertising data, then summarize these changes and summarize the patterns of changes in relevant data based on personal experience, and then make manual predictions.

[0005] Obviously, this approach is too clumsy and time-consuming, and since it relies solely on human experience to estimate and predict, there is a high possibility that the predicted results will deviate significantly from the actual situation.

[0006] Therefore, in the field of advertising, there is an urgent need for technological breakthroughs and the emergence of a large data system that can automatically and intelligently predict advertising data. Summary of the Invention

[0007] The present invention provides an advertising data intelligent analysis system based on multi-resolution index weighting, which effectively solves the above-mentioned problems in the prior art.

[0008] Specifically, the present invention provides an advertising data analysis system based on multi-resolution exponential weighting, which includes a historical database, an operation module, an update module, and a probability prediction module, wherein the historical database obtains the periodic advertising feature data of K advertisements at time T0, K is an integer and K≥2, the operation module obtains the maximum estimated result V based on the periodic advertising data and sets the estimated result interval (0, V], and sets the initial number of boxes M0, M0 is an integer and M0≥2, the estimated result interval (0, V] is initially evenly divided into M0 boxes, and a count value is set for each box, any advertisement forms actual conversion result data based on the periodic advertising feature data, if the actual conversion result data falls within any box, the count value of the box is accumulated 1. From time T0 to time T1 after a predetermined time step T, if the accumulated count value of any box is greater than the maximum detection threshold, the box is split; if the accumulated count value of any box is less than the minimum detection threshold, the box is merged with the adjacent box. Due to the splitting and merging of boxes, the number of boxes is updated from M0 to M1, where M1 is an integer and M1≥1. The update module sets a historical change variable for each box. The historical change variable is accumulated based on the staged advertising feature data and forms a final accumulated value at time T1. The probability prediction module assigns a weight to the final accumulated value of the historical change variable, and performs a weighted sum operation on the M1 boxes and the K advertisements, thereby calculating the probability of successful delivery of each advertisement.

[0009] Preferably, in the calculation module, the calculation formula for the advertising estimated result pValue is: ,in, Indicates the estimated advertising results of each advertisement. Indicates the average conversion result of each advertisement. Indicates the estimated click-through rate of each ad. It represents the estimated conversion rate of each advertisement. The maximum value of the pValue values ​​of all K advertisements is the maximum estimated result V.

[0010] Preferably, in the calculation module, the estimated result interval (0, V] is initially divided into M0 bins, and an integer i is set to represent the bin number, i is sequentially set to 0, 1, 2, .... M0-1, and the i-th bin is i The lower bound of v is expressed as: low =i*V / M0, its upper bound is expressed as: v high =(i+1)*V / M0, the i-th bin i The expression is the interval (v low , v high ].

[0011] Preferably, the calculation module sets a count value count for each box.i , for M0 bins, M0 bins bin0, bin1, ..., bin M0-1 The count values ​​are count0, count1, ..., count M0-1 , the count value of each box is count i , are initially set to 0. If the actual conversion result data falls in the i-th bin i In the interval, the count value count i Accumulate 1.

[0012] Preferably, in the update module, each bin i The cost data of the corresponding advertisement that has evolved over time is recorded as price i The historical change variables set for each box are counted as history i , historical change variable history i Initialized to 0, in the historical evolution process, if each bin i Price data price i If the winning data is not greater than the preset mp, the historical change variable history i Remain unchanged; if the price data price i If it is greater than the preset winning data mp, the historical change variable history i Accumulated value pValue-price i .

[0013] Preferably, in the calculation module, the maximum detection threshold N1 and the minimum detection threshold N2 are set, wherein N1>N2, if the i-th bin i The count value count i Satisfy count i >N1, then bin i i Split into two bins evenly i’ and bin i’’ , after splitting, the box bin i’ The interval expression is (v low ,(v low + v high ) / 2], box bin i’’ The interval expression is ( (v low + v high ) / 2,v high ], the two bins after splitting i’ and bin i’’ The count value is set to count i / 2.

[0014] Preferably, in the operation module, if the count value i count i of the i-th bin i satisfies count i < N2, then the i-th bin and the bin with the smallest count value among the adjacent bins are merged, and the sum of the count values of the two merged bins becomes the count value of the merged bin.

[0015] Preferably, in the probability prediction module, there are K items of advertisements in total. The corresponding advertisement feature data of any j-th advertisement falls into each bin i , where j is an integer and 1 ≤ j ≤ K. The calculation formula for the advertisement placement success probability p j of any j-th advertisement is as follows:

[0016] ,

[0017] where history ij represents the historical change variable history i of any j-th advertisement in any i-th bin i , represents the weighting coefficient for any i-th bin i in the summation formula.

[0018] Generally speaking, the present invention provides an intelligent advertising data analysis system based on multi-resolution exponential weighting, which performs binning operations on the estimated result data caused by advertisement feature data, and intelligently adjusts the width of the bin interval during the historical evolution process, thereby automatically and intelligently adapting to the historical evolution of advertisements, conforming to the current advertising situation, and calculating the advertisement placement success probability of any item of advertisement in combination with cost data on this basis, effectively solving the technical difficulty of unpredictable advertisement data that cannot be avoided in the field of advertising technology with the idea of probability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will discuss the drawings required for use in the description of the embodiments or the prior art. Obviously, the technical solutions described in conjunction with the drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other embodiments and their drawings can also be obtained based on these embodiments shown in the drawings.

[0020] Figure 1 shows the specific operation flow of the intelligent advertising data analysis system based on multi-resolution exponential weighting according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will clearly and completely describe the technical solutions of various embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments described in 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.

[0022] In summary, this invention provides an intelligent advertising data analysis system based on multi-resolution exponential weighting. By integrating the realities of the advertising industry and the technical characteristics of advertising data, a relatively comprehensive data analysis model system has been constructed from a technical perspective. This enables relatively accurate technical estimation of advertising data, which is typically highly unpredictable. The following details the specific operation of this intelligent advertising data analysis system based on multi-resolution exponential weighting.

[0023] Figure 1 The specific operation process of the advertising data intelligent analysis system based on multi-resolution index weighting according to the present invention is shown.

[0024] like Figure 1 As shown, the advertising data intelligent analysis system based on multi-resolution exponential weighting provided by the present invention first establishes a historical database. The historical database obtains the phased advertising feature data of a specific advertisement at a specific time in the past (for example, 24 hours, 30 days, or a quarter) at time T0 as a training data set, and counts the actual total conversion number and the actual total conversion results of the phased advertising feature data, and then divides the actual total conversion results by the actual total conversion number to obtain the average conversion result.

[0025] It's important to note that after entering the historical database, ad feature data is processed, analyzed, and judged based on its inherent attributes within the overall big data system. This system doesn't consider its commercial attributes, but rather its technical analysis and processing. Therefore, this disclosure focuses on the technical analysis and processing within the overall big data system.

[0026] Next, regarding the historical database, we've mentioned the actual conversions triggered by the periodic advertising feature data stored there. In reality, these periodic ads, upon entry into the historical database, have already been estimated by the calculation module, generating a maximum estimated outcome V based on the periodic advertising data. This estimated ad outcome data can be derived from the advertising feature data using a specific estimation function. For example, the specific estimation function can be a simple, straightforward linear function. In this case, the advertising feature data is simply multiplied by a coefficient to generate the estimated outcome data.

[0027] Of course, more specific solutions can be provided. For example, advertising feature data is not a single piece of data. The estimated advertising results can examine the click-through rate, conversion results, and specific conversion rate of the advertisement. The estimated results of the advertisement can be calculated based on the average conversion results and the predicted click-through rate and predicted conversion rate in the data. , its calculation formula can be expressed as: ,

[0028] in, Indicates the estimated results of the advertisement. represents the average conversion result, Indicates the estimated click-through rate. Represents the estimated conversion rate. The maximum pValue value among all ad samples is the maximum estimated outcome V.

[0029] In the historical database, the calculation module sets the initial number of boxes M0 at the original time T0 and sets the estimated result interval (0, V]. It should be noted that the square brackets "[", "]" at both ends of this interval indicate that the interval is closed and includes the number; if the interval is enclosed by parentheses "(", ")", it means that the interval is open and does not include the number.

[0030] The estimated result interval (0, V] is initially divided into M0 bins evenly. The integer i is set to represent the bin number. The values ​​of i are 0, 1, 2, .... M0-1 in sequence. In other words, when i is 0, it represents the 0th bin, when i is 1, it represents the 1st bin, and so on. Finally, when i is M0-1, it represents the M0-1th bin.

[0031] The i-th bin i The lower bound of v is expressed as: low =i*V / M0, its upper bound is expressed as: v high =(i+1)*V / M0, so the i-th bin i The expression is the interval (v low , v high ]. As mentioned above, the right side of the interval is a square bracket “]”, indicating that the number is included, while the left side is a small bracket “(”, indicating that the number is not included.

[0032] Thus, the integer i takes the values ​​0, 1, 2, ....M0-1 in sequence, forming M0 bins bin0, bin1, ..., bin M0-1 .

[0033] In the calculation module, a count value is set for each box. i In other words, for M0 bins, M0 bins bin0, bin1, ..., bin M0-1The count values ​​are count0, count1, ..., count M0-1 The count value of each box is count i , are initially set to 0.

[0034] As mentioned in the title of this invention, "multi-resolution" is the embodiment of "resolution." The smaller the bin, the greater the resolution. As will be explained in detail below, this "resolution" will also change in real time as the system progresses.

[0035] As time goes by, the advertisement will form actual conversion result data based on the staged advertisement data, and the actual conversion result data will also be input into the historical database. If the actual conversion result data falls into the i-th bin, i In the interval, the count value count i Accumulate 1. Thus, as time goes by, the actual conversion result data obtained by the advertisements with the above-mentioned phased advertising data will be successively entered into the historical database and fall within the intervals of different boxes, resulting in changes in the count values ​​of different boxes.

[0036] As mentioned above, each bin i In fact, it is a range of estimated results. Therefore, for advertising, this range of estimated results corresponds to a price data of advertising, which is calculated as price. i This cost data will, of course, change over time. As mentioned above, cost data reflects the unpredictability that is a defining characteristic of advertising data. However, this system overcomes this unpredictability by inputting this cost data in real time and identifying patterns within this data model. This task is handled by the update module provided by this system.

[0037] In the update module, each bin i Set a history change variable history i , indicating the historical transformation results within the box, and the historical change variable history i Initialized to 0. According to the historical evolution of the historical database, each bin i Corresponding price data price i It is also changing.

[0038] During the time evolution, if each bin i Price data price i If the winning data is not greater than the preset mp, the historical change variable history of the box i Remain unchanged. If the price data price iIf it is greater than the preset winning data mp, the historical change variable history of the specific advertisement i Accumulate the value pValue - price i .

[0039] Return to the operation module again. At time T1 after a predetermined time step T, the system detects the count values of all bins, and sets a maximum detection threshold N1 and a minimum detection threshold N2, where N1 > N2.

[0040] If the count value count i of the i-th bin i satisfies count i > N1, it indicates that the samples in this bin are overly dense. Therefore, split the i-th bin i into two bins i’ and bin i’’ on average. The original interval of the i-th bin i is [v low , v high )]. After splitting, the interval expression of bin i’ is (v low , (v low + v high ) / 2], and the interval expression of bin i’’ is ((v low + v high ) / 2, v high ). Each such split will add one bin. Correspondingly, the count values of these two split bins are both half of the original, that is, count i / 2.

[0041] If the count value count i of the i-th bin i satisfies count i < N2, it indicates that the samples in this bin are overly sparse. Therefore, merge the i-th bin i with the bin with the smallest count value among its adjacent bins. The sum of the count values of the two merged bins becomes the count value of the merged bin. It can be imagined that each such merge will reduce one bin. Of course, it should be noted that in all bins, generally each bin has two adjacent bins on the left and right. Therefore, when merging, select the one with the smaller count value between the left and right adjacent bins for merging. For the two outermost bins, there is only one adjacent bin, so there is no need to compare, and the adjacent bin is the one with the "smallest count value", and just merge with it.

[0042] After the above splitting and merging, the number of boxes changes from the initial number M0 to M1. It's important to note that because the splitting process increases the number of boxes, while the merging process decreases the number of boxes, M1 may be greater than, equal to, or less than M0. However, it is certain that the current number of boxes, M1, is closer to the current reality.

[0043] When the above box i 、price i 、history i All have been updated. At time T1, the price data of each box can be calculated based on the change in the number of boxes from M0 to M1. i Changes to the history variable history i Assigning weighted coefficients and calculating the success probability of each advertisement is a task undertaken by the probability prediction module provided by this system.

[0044] Specifically, in the probability prediction module, it is assumed that there are K advertisements in total, where K is an integer and K≥2, and the corresponding advertisement feature data of any j-th advertisement falls into each bin. i , where j is an integer and 1≤j≤K. Therefore, the following weighted probability analysis involves both any advertisement (i.e., the jth advertisement) and any bin i The specific weighted success probability p for any j-th advertisement j The calculation formula is as follows:

[0045] Among them, history ij Indicates that any j-th advertisement is in any i-th bin i Historical changes in the variable history i , Indicates the summation formula for any i-th bin i The weighting coefficient of .

[0046] At this point, the system has calculated the weighted success probability of any advertisement by inputting advertisement-related data. The success probability of all related advertisements can be seen at a glance, thus realizing intelligent advertisement delivery.

[0047] This concludes the basic introduction to the present invention. In summary, the present invention provides an intelligent advertising data analysis system based on multi-resolution exponential weighting. This system bins the estimated outcome data generated by advertising feature data and intelligently adjusts the width of the bins over time. This system automatically and intelligently adapts to the historical evolution of advertising and the current state of advertising. Furthermore, it combines cost data to calculate the probability of successful placement of any advertisement, effectively resolving the unavoidable technical difficulty of unpredictable advertising data in the advertising technology field using probabilistic thinking.

[0048] The above description is merely an exemplary embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An advertising data analysis system based on multi-resolution exponential weighting, characterized in that: The system includes a historical database, a calculation module, an update module and a probability prediction module, among which: The historical database obtains the periodic advertising feature data of K advertisements at time T0, where K is an integer and K≥2. The calculation module obtains the maximum estimated result V in the advertising estimated results based on the staged advertising feature data and sets the estimated result interval (0, V], and sets the initial number of boxes M0, M0 is an integer and M0 ≥ 2, and performs a binning operation on the estimated result interval (0, V], initially dividing it into M0 boxes on average, setting an integer i to represent the box number, i takes values ​​of 0, 1, 2, .... M0-1 in turn, then the i-th box is expressed as box bin i , set a count value for each box, any advertisement forms actual conversion result data based on the staged advertisement feature data, if the actual conversion result data falls within any box, the count value of the box is accumulated by 1, wherein, in the binning operation, the smaller the box, the greater the resolution, in this calculation module, the calculation formula of the advertisement estimated result pValue is: pValue = mCom*pCtr*pCvr, wherein pValue represents the advertisement estimated result of each advertisement, mCom represents the average conversion result of each advertisement, pCtr represents the estimated click rate of each advertisement, pCvr represents the estimated conversion rate of each advertisement, the maximum value of the pValue values ​​of all K advertisements is the maximum estimated result V, At time T1 after a predetermined time step T from time T0, if the accumulated count value of any box is greater than the maximum detection threshold, the box is split; if the accumulated count value of any box is less than the minimum detection threshold, the box is merged with its adjacent boxes. Due to the splitting and merging of boxes, the number of boxes is updated from M0 to M1, where M1 is an integer and M1≥1. The update module sets a historical change variable for each box. The historical change variable is accumulated based on the periodic advertising feature data and forms the final accumulated value at time T1. In the update module, each bin i The cost data of the corresponding advertisement that evolves over time is recorded as price i , the historical change variables set for each box are counted as history i , historical change variable history i Initialized to 0, during the historical evolution, if each bin i Price data price i If it is not greater than the preset winning data mp, the historical change variable history i Remain unchanged; if the price data price i If it is greater than the preset winning data mp, the historical change variable history i Accumulated value pValue-price i , The probability prediction module assigns weights to the final accumulated values ​​of the historical change variables, and performs weighted sum operations on the M1 boxes and the K advertisements, thereby calculating the probability of successful delivery of each advertisement. In the probability prediction module, there are K items of advertisements in total, and the corresponding advertisement feature data of any j-th advertisement falls into each bin i , where j is an integer and 1≤j≤K, the probability of successful delivery of any j-th advertisement is p j The calculation formula is as follows: Among them, history ij Indicates that any j-th advertisement is in any i-th bin i The historical change variable in μ i Indicates the summation formula for any i-th bin i The weighting coefficient of .

2. The system according to claim 1, characterized in that In the calculation module, the i-th bin i The lower bound of is expressed as: v low =i*V / M0, its upper limit is expressed as: v high =(i+1)*V / M0, the i-th bin i The expression of the interval (v low , v high ].

3. The system according to claim 2, characterized in that The calculation module sets a count value for each box. i , for M0 bins, M0 bins bin0, bin1, ..., bin M0-1 The count values ​​are count0, count1, ..., count M0-1 , the count value of each box is count i , are initially set to 0. If the actual conversion result data falls in the i-th bin i If the count value is within the interval of i Accumulate 1.

4. The system according to claim 1, characterized in that In the calculation module, the maximum detection threshold N1 and the minimum detection threshold N2 are set, where N1>N2. If the i-th bin i The count value count i Satisfy count i > N1, then the i-th bin i Split into two bins evenly i’ and bin i” , after splitting, the box bin i’ The interval expression of (v low ,(v low +v high ) / 2], box bin i” The interval expression of ((v low +v high ) / 2,v high ], the two boxes after splitting i’ and bin i” The count value of i / 2.

5. The system according to claim 4, characterized in that In the operation module, if the count value i count i of the i-th bin i satisfies count i < N2, then the i-th bin and the bin with the minimum count value among its adjacent bins are merged, and the sum of the count values of the two merged bins becomes the count value of the merged bin.

Images