Urban advertisement placement management system and method

Abstract

The application relates to the technical field of advertisement management, and discloses a city advertisement delivery management system and method. The system and method acquire a time queue of advertisement effect indexes, a time queue of external environment variables and a time queue of competitive intelligence within a scheduled time window, and then utilize a double-layer LSTM network to perform time sequence coding on the advertisement effect indexes, the external environment variables (such as air temperature, humidity and holiday flag) and the competitive intelligence (such as competitive advertisement delivery intensity within 500 meters), so as to capture periodic pattern features within a day and across days. The features are enhanced by a ternary attention network, and multi-source time sequence data is fused to generate accurate advertisement delivery prediction results. A dynamic decision module based on a Softmax function further gives an optimized advertisement delivery time period suggestion. In this way, the effect and efficiency of advertisement delivery are significantly improved, and scientific decision support is provided for brands and advertisers.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advertisement management, and more specifically, to a city advertisement delivery management system and method. BACKGROUND

[0002] In modern city advertisement delivery management, with the rapid development of the digital advertising market and the diversification of consumer behavior, how to efficiently and accurately deliver advertisements has become an important challenge for advertisers and brand managers. Traditionally, advertisement delivery strategies mainly rely on market research and historical data to determine the best delivery time and location. However, this method has obvious limitations, especially when dealing with rapidly changing market environments and competitive situations.

[0003] Firstly, the traditional advertisement delivery management method usually lacks effective consideration of dynamic external environmental factors. For example, temperature, humidity, holidays and other external variables can have a significant impact on consumer purchasing behavior, but these factors are often not fully incorporated into the advertisement delivery decision-making process. In addition, advertisement effectiveness indicators such as click-through rate and conversion rate can provide some feedback information, but they only reflect the attractiveness of the advertisement itself, ignoring the potential influence of other external factors on consumers. This single-dimensional analysis method cannot fully reflect the true effect of advertisement delivery, resulting in unreasonable allocation of advertising resources and reducing the return on investment of advertising. Secondly, the existing technology also has deficiencies in dealing with competitor advertisement delivery intensity. With the increasing market competition, competitors may also increase their advertisement delivery intensity in the same time period, which will undoubtedly affect the effectiveness of the target advertisement. However, traditional advertisement delivery management systems often cannot capture this competitive intelligence in a timely manner, let alone effectively analyze it and adjust their own advertising strategies accordingly. Therefore, even with high-quality advertising content, in a highly competitive market environment, it may lose its competitive advantage due to the inability to accurately grasp market dynamics.

[0004] Therefore, an optimized city advertisement delivery management solution is expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a city advertisement delivery management system and method, aiming to optimize advertisement delivery strategies through advanced machine learning technology. To solve the obvious deficiencies of existing technology in dealing with dynamic external environmental factors and competitor advertisement delivery intensity, resulting in inaccurate advertisement effectiveness prediction and unreasonable resource allocation.

[0006] According to an aspect of the present application, a city advertisement delivery management system is provided, comprising: inputting a time queue of advertisement effect indicators and a time queue of external environment variables within a predetermined time window, the predetermined time window being 7 days; inputting a time queue of competitive intelligence within the predetermined time window; time series encoding the time queue of advertisement effect indicators, the time queue of external environment variables and the time queue of competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector and a competitive intelligence time sequence periodic pattern feature encoding vector, wherein the first layer LSTM of the double-layer LSTM network is used to capture intra-day time sequence periodic pattern features, and the second layer LSTM of the double-layer LSTM network is used to capture cross-day time sequence periodic pattern features; inputting the advertisement effect indicator time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector into a dynamic decision module based on a Softmax function to obtain an advertisement delivery prediction result; and generating a suggestion for an advertisement delivery time period based on the advertisement delivery prediction result.

[0007] In the above city advertisement delivery management system, the advertisement effect indicators include click rate and conversion rate, the external environment variables include air temperature, humidity and holiday flag, and the competitive intelligence is the competitive product advertisement delivery intensity within 500 meters.

[0008] In the above city advertisement delivery management system, the first layer LSTM contains 64 neurons, and the second layer LSTM contains 32 neurons.

[0009] In the city advertisement management system, the time sequence encoding of the time queue of the advertisement effect index, the time queue of the external environment variable and the time queue of the competitive intelligence based on the double-layer LSTM network obtains the advertisement effect index time sequence periodicity mode feature encoding vector, the external environment variable time sequence periodicity mode feature encoding vector and the competitive intelligence time sequence periodicity mode feature encoding vector, including: inputting the time queue of the advertisement effect index, the time queue of the external environment variable and the time queue of the competitive intelligence into the first layer LSTM of the double-layer LSTM network to obtain the sequence distribution of the advertisement effect index intra-day periodicity mode feature encoding vector, the sequence distribution of the external environment variable intra-day periodicity mode feature encoding vector and the sequence distribution of the competitive intelligence intra-day periodicity mode feature encoding vector; inputting the sequence distribution of the advertisement effect index intra-day periodicity mode feature encoding vector, the sequence distribution of the external environment variable intra-day periodicity mode feature encoding vector and the sequence distribution of the competitive intelligence intra-day periodicity mode feature encoding vector into the second layer LSTM of the double-layer LSTM network respectively to obtain the advertisement effect index time sequence periodicity mode feature encoding vector, the external environment variable time sequence periodicity mode feature encoding vector and the competitive intelligence time sequence periodicity mode feature encoding vector.

[0010] In the city advertisement management system, the time sequence encoding of the time queue of the advertisement effect index, the time queue of the external environment variable and the time queue of the competitive intelligence based on the double-layer LSTM network obtains the advertisement effect index time sequence periodicity mode feature encoding vector, the external environment variable time sequence periodicity mode feature encoding vector and the competitive intelligence time sequence periodicity mode feature encoding vector, including: inputting the time queue of the advertisement effect index, the time queue of the external environment variable and the time queue of the competitive intelligence into the first layer LSTM of the double-layer LSTM network to obtain the sequence distribution of the advertisement effect index intra-day periodicity mode feature encoding vector, the sequence distribution of the external environment variable intra-day periodicity mode feature encoding vector and the sequence distribution of the competitive intelligence intra-day periodicity mode feature encoding vector; inputting the sequence distribution of the advertisement effect index intra-day periodicity mode feature encoding vector, the sequence distribution of the external environment variable intra-day periodicity mode feature encoding vector and the sequence distribution of the competitive intelligence intra-day periodicity mode feature encoding vector into the second layer LSTM of the double-layer LSTM network respectively to obtain the advertisement effect index time sequence periodicity mode feature encoding vector, the external environment variable time sequence periodicity mode feature encoding vector and the competitive intelligence time sequence periodicity mode feature encoding vector.

[0011] The advertisement effect index time sequence periodicity mode feature encoding vector, the external environment variable time sequence periodicity mode feature encoding vector and the competitive intelligence time sequence periodicity mode feature encoding vector are inputted into the dynamic decision module based on the Softmax function to obtain the advertisement prediction result.

[0012] In the city advertisement management system, the feature encoding vector of the time series periodic pattern of the advertisement effect index, the feature encoding vector of the time series periodic pattern of the external environment variable, and the feature encoding vector of the time series periodic pattern of the competitive intelligence are input into a ternary attention network to obtain an enhanced feature encoding vector of the time series periodic pattern of the advertisement effect index, including: calculating a feature fine-grained association matrix of the feature encoding vector of the time series periodic pattern of the external environment variable and the feature encoding vector of the time series periodic pattern of the competitive intelligence to obtain an external environment-competitive intelligence time series collaborative attention matrix; performing feature activation based on a Sigmoid function on the external environment-competitive intelligence time series collaborative attention matrix to obtain an external environment-competitive intelligence time series collaborative attention mask matrix; mapping the feature encoding vector of the time series periodic pattern of the advertisement effect index to the external environment-competitive intelligence time series collaborative attention mask matrix to obtain an attention modulated feature encoding vector of the time series periodic pattern of the advertisement effect index; inputting the attention modulated feature encoding vector of the time series periodic pattern of the advertisement effect index and the feature encoding vector of the time series periodic pattern of the advertisement effect index into a residual unit to obtain the enhanced feature encoding vector of the time series periodic pattern of the advertisement effect index.

[0013] In the city advertisement management system, the feature fine-grained association matrix of the feature encoding vector of the time series periodic pattern of the external environment variable and the feature encoding vector of the time series periodic pattern of the competitive intelligence is calculated to obtain an external environment-competitive intelligence time series collaborative attention matrix, including: calculating the feature fine-grained association matrix of the feature encoding vector of the time series periodic pattern of the external environment variable and the feature encoding vector of the time series periodic pattern of the competitive intelligence to obtain the external environment-competitive intelligence time series collaborative attention matrix by the following formula; wherein the formula is: ; wherein, the feature encoding vector of the time series periodic pattern of the external environment variable is denoted as the feature encoding vector of the time series periodic pattern of the competitive intelligence is denoted as matrix multiplication is denoted as transposition operation is denoted as the feature length of the feature encoding vector of the time series periodic pattern of the external environment variable and the feature encoding vector of the time series periodic pattern of the competitive intelligence is denoted as the external environment-competitive intelligence time series collaborative attention matrix is denoted as

[0014] In the city advertisement delivery management system, the feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector is calculated to obtain an external environment-competition intelligence time series collaborative attention matrix, which includes: performing distribution stability probability based on interaction global mean on the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector to obtain an external environment variable time series periodic pattern feature stability probability vector and a competition intelligence time series periodic pattern feature stability probability vector; based on the feature mean of the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector, the interaction step and the generalized interaction coefficient are calculated; based on the calculated interaction step and generalized interaction coefficient, the external environment variable time series periodic pattern feature stability probability vector and the competition intelligence time series periodic pattern feature stability probability vector are subjected to truncated approximation based on historical values to obtain an external environment variable time series periodic pattern feature truncated approximation vector and a competition intelligence time series periodic pattern feature truncated approximation vector; the feature fine-grained association matrix of the external environment variable time series periodic pattern feature truncated approximation vector and the competition intelligence time series periodic pattern feature truncated approximation vector is calculated to obtain an external environment-competition intelligence time series collaborative attention matrix.

[0015] In the city advertisement delivery management system, the advertisement delivery prediction result is a probability value of each time period as the best delivery opportunity.

[0016] According to another aspect of the present application, a city advertisement delivery management system is provided, comprising: a time data input module configured to input a time queue of advertisement effect indicators and a time queue of external environment variables within a predetermined time window, the predetermined time window being 7 days; a competitive intelligence input module configured to input a time queue of competitive intelligence within the predetermined time window; a periodic pattern feature extraction module configured to perform time series encoding on the time queue of advertisement effect indicators, the time queue of external environment variables and the time queue of competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector and a competitive intelligence time sequence periodic pattern feature encoding vector, wherein a first layer LSTM of the double-layer LSTM network is configured to capture intra-day time sequence periodic pattern features, and a second layer LSTM of the double-layer LSTM network is configured to capture cross-day time sequence periodic pattern features; an advertisement delivery prediction result determination module configured to input the advertisement effect indicator time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector into a dynamic decision module based on a Softmax function to obtain an advertisement delivery prediction result; and an advertisement delivery suggestion generation module configured to generate a suggestion of an advertisement delivery time period based on the advertisement delivery prediction result.

[0017] Compared with the prior art, the city advertisement delivery management system and method provided by the present application acquires a time queue of advertisement effect indicators, a time queue of external environment variables and a time queue of competitive intelligence within a predetermined time window, and then performs time series encoding on the advertisement effect indicators, the external environment variables (such as temperature, humidity, holiday flag) and the competitive intelligence (such as competitive advertisement delivery intensity within 500 meters) by using a double-layer LSTM network to capture intra-day and cross-day periodic pattern features. The feature representation is enhanced by a ternary attention network, and multi-source time sequence data is fused to generate a precise advertisement delivery prediction result. A dynamic decision module based on a Softmax function further gives a suggestion of an optimized advertisement delivery time period. In this way, the effect and efficiency of advertisement delivery are significantly improved, and scientific decision support is provided for brands and advertisers. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which like reference characters refer to like parts throughout the figures. The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of this specification, illustrate embodiments of the present application, and together with the description serve to explain the present application. In the drawings, like reference characters refer to like parts throughout the figures.

[0019] Figure 1Fig. 1 shows a flow diagram of a city advertisement delivery management method according to an embodiment of the present application.

[0020] Figure 2 Fig. 3 shows a flow diagram of step S3 in the city advertisement delivery management method according to an embodiment of the present application.

[0021] Figure 3 Fig. 4 shows a flow diagram of step S4 in the city advertisement delivery management method according to an embodiment of the present application.

[0022] Figure 4 Fig. 5 shows a flow diagram of step S41 in the city advertisement delivery management method according to an embodiment of the present application.

[0023] Figure 5 Fig. 6 shows a structural diagram of a city advertisement delivery management system according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part but not all of the embodiments of the present application, and the present application can be implemented in many different ways. Therefore, the contents described should not be considered as limitation of the present application.

[0025] Figure 1 Fig. 1 shows a flow diagram of a city advertisement delivery management method according to an embodiment of the present application. As shown in Fig. 1, the city advertisement delivery management method comprises: S1, inputting a time queue of advertisement effect indicators and a time queue of external environment variables within a predetermined time window, the predetermined time window being 7 days; S2, inputting a time queue of competitive intelligence within the predetermined time window; S3, performing time series encoding on the time queue of advertisement effect indicators, the time queue of external environment variables and the time queue of competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector and a competitive intelligence time sequence periodic pattern feature encoding vector, wherein a first layer LSTM of the double-layer LSTM network is used to capture intra-day time sequence periodic pattern features, and a second layer LSTM of the double-layer LSTM network is used to capture cross-day time sequence periodic pattern features; S4, inputting the advertisement effect indicator time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector into a dynamic decision module based on a Softmax function to obtain an advertisement delivery prediction result; and S5, generating a suggestion of an advertisement delivery time period based on the advertisement delivery prediction result. Figure 1

[0026] ​Specifically, in step S1, a time queue of advertising effectiveness indicators and a time queue of external environment variables within a predetermined time window of 7 days are input. In one example, the advertising effectiveness indicators include click-through rate and conversion rate, and the external environment variables include air temperature, humidity, and holiday flag. Specifically, the time queue of advertising effectiveness indicators is considered first. The advertising effectiveness indicators referred to herein generally include click-through rate (CTR), conversion rate (CVR), and other key performance indicators. The click-through rate (CTR) and conversion rate (CVR) are aggregated by time period and form a seven-day time series. In one embodiment, the time series data is obtained in hourly dimensions, and of course, the time period can also be selected according to the type of advertisement. In this way, the trend of changes in advertising effectiveness over time can be clearly observed, such as whether there is a significant increase or decrease on weekends or during specific holidays. For example, at the same time, the time queue of external environment variables is also crucial. Such variables can include air temperature, humidity, holiday flag, and other factors. Therefore, the air temperature, humidity, and other meteorological data of each time period, as well as information such as whether it is a holiday, need to be recorded. These data are also organized according to a seven-day time window to form a time series of external environment variables. This approach helps to identify how external conditions indirectly affect the performance of the advertisement.

[0027] The reason for using the above two types of time series data together is that they can jointly reveal the complex dynamic relationship behind advertising effectiveness. The choice of 7 days as a predetermined time window is not arbitrary. On the one hand, 7 days is long enough to capture intra-day and cross-day pattern changes; on the other hand, it is not too long to lose timeliness. For example, in a rapidly changing market environment, a long time window can lead to lag effects, making decisions based on old data no longer applicable. In contrast, a 7-day time window can reflect both short-term fluctuations and capture periodic trends.

[0028] Specifically, in step S2, a time queue of competitive intelligence within the predetermined time window is input. In one example, the competitive intelligence is the intensity of competitor advertising within 500 meters. It should be understood that market competition is a dynamic process, and the advertising activities of competitors can significantly affect advertising effectiveness. For example, if it is found that the number of competitor ad displays has increased significantly on a certain day, even if the ad content is attractive, it may reduce the click-through rate and conversion rate due to the distraction of consumer attention. By continuously monitoring these changes, advertising placement strategies can be adjusted in a timely manner to avoid waste of resources.

[0029] In one specific implementation example, an effective data collection mechanism is established to obtain these competitive intelligence. This can be achieved in various ways, such as using the data interface of a third-party advertising service provider to regularly crawl the display data of competitors' online advertisements; or by recording the advertising arrangement of surrounding stores through on-site research. These information is recorded every day and organized into a seven-day time sequence. For example, in each time period of each day from Monday to Sunday, the total number of displays or distribution density of all competitor advertisements within a 500-meter range on that day is counted. In addition, organizing these competitive intelligence according to a 7-day time window also helps to identify periodic patterns. For example, some competitors may increase their advertising investment during the weekend to attract more people who go shopping during the weekend. If this can be predicted in advance, and the advertising investment plan is adjusted accordingly, the company can take the initiative in the fierce market competition.

[0030] Specifically, in step S3, the time sequence encoding is performed on the time queue of the advertising effect indicator, the time queue of the external environment variable, and the time queue of the competitive intelligence based on a double-layer LSTM network to obtain an advertising effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector, and a competitive intelligence time sequence periodic pattern feature encoding vector. The first layer LSTM of the double-layer LSTM network is used to capture intra-day time sequence periodic pattern features, and the second layer LSTM of the double-layer LSTM network is used to capture cross-day time sequence periodic pattern features. In one example, the first layer LSTM contains 64 neurons, and the second layer LSTM contains 32 neurons. Specifically, the reason for using a double-layer LSTM network for time sequence encoding is that this method can effectively capture complex pattern features at different time scales. First, the changes in advertising effect, external environment variable, and competitive intelligence often have multi-level dynamic characteristics. For example, the effect of an advertisement may fluctuate significantly within a day, but there is also a trend change across days. The first layer LSTM captures intra-day patterns through a larger number of neurons (64), which can reflect these short-term fluctuations in detail; while the second layer LSTM captures cross-day patterns through a smaller number of neurons (32), which helps to identify long-term trends. In addition, the design of the double-layer LSTM network also considers the balance between computational efficiency and model complexity. The first layer LSTM captures intra-day patterns through a larger number of neurons (64), ensuring that short-term fluctuations can be reflected in detail; while the second layer LSTM captures cross-day patterns through a smaller number of neurons (32), which ensures the generalization ability of the model and avoids the risk of overfitting. This design enables high accuracy while maintaining good real-time performance and scalability.

[0031] In one example, as shown in FIG. 6, the time sequence encoding is performed on the time queue of the advertising effect indicator, the time queue of the external environment variable, and the time queue of the competitive intelligence based on a double-layer LSTM network to obtain an advertising effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector, and a competitive intelligence time sequence periodic pattern feature encoding vector. Figure 2As shown, in step S3, the time series encoding of the time queue of the advertising effect index, the time queue of the external environment variable and the time queue of the competitive intelligence based on the double-layer LSTM network to obtain the advertising effect index time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector, comprising: S31, inputting the time queue of the advertising effect index, the time queue of the external environment variable and the time queue of the competitive intelligence into the first layer LSTM of the double-layer LSTM network to obtain the sequence distribution of the advertising effect index intra-day periodic pattern feature encoding vector, the sequence distribution of the external environment variable intra-day periodic pattern feature encoding vector and the sequence distribution of the competitive intelligence intra-day periodic pattern feature encoding vector; S32, inputting the sequence distribution of the advertising effect index intra-day periodic pattern feature encoding vector, the sequence distribution of the external environment variable intra-day periodic pattern feature encoding vector and the sequence distribution of the competitive intelligence intra-day periodic pattern feature encoding vector into the second layer LSTM of the double-layer LSTM network respectively to obtain the advertising effect index time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector.

[0032] Specifically, the time series of the advertising effect index, the time series of the external environment variable and the time series of the competitive intelligence are sequentially input into the first layer LSTM. The first layer LSTM processes these data day by day and extracts the intra-day pattern features in each time series. For example, it may find that the click rate of certain advertisements is particularly high around 8 pm, or the conversion rate on weekends is significantly higher than on weekdays. These intra-day pattern features are encoded into a series of vectors, forming the sequence distribution of the advertising effect index, the external environment variable and the competitive intelligence intra-day periodic pattern feature encoding vector. Then, these sequence distributions of the intra-day periodic pattern feature encoding vector are input into the second layer LSTM. The role of the second layer LSTM is to capture the cross-day pattern features, that is, to identify the trend changes between different dates within a week. For example, it may find that the advertising effect is better in the first half of the week, but the effect decreases in the second half due to the increased advertising investment by competitors. In this way, the second layer LSTM can generate more comprehensive advertising effect index time sequence periodic pattern feature encoding vector, external environment variable time sequence periodic pattern feature encoding vector and competitive intelligence time sequence periodic pattern feature encoding vector.

[0033] Specifically, in step S4, the advertising effect index time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector are input into the dynamic decision module based on the Softmax function to obtain the advertising investment prediction result.

[0034] In one example, such as Figure 3 As shown, step S4, which involves inputting the advertising effectiveness index time-series periodic pattern feature encoding vector, the external environmental variable time-series periodic pattern feature encoding vector, and the competitive intelligence time-series periodic pattern feature encoding vector into a dynamic decision-making module based on the Softmax function to obtain the advertising placement prediction result, includes: S41, inputting the advertising effectiveness index time-series periodic pattern feature encoding vector, the external environmental variable time-series periodic pattern feature encoding vector, and the competitive intelligence time-series periodic pattern feature encoding vector into a three-element attention network to obtain an enhanced advertising effectiveness index time-series periodic pattern feature encoding vector; S42, fusing the enhanced advertising effectiveness index time-series periodic pattern feature encoding vector, the external environmental variable time-series periodic pattern feature encoding vector, and the competitive intelligence time-series periodic pattern feature encoding vector to obtain an advertising placement multi-source time-series fusion encoding vector; and S43, inputting the advertising placement multi-source time-series fusion encoding vector into the dynamic decision-making module based on the Softmax function to obtain the advertising placement prediction result.

[0035] Specifically, in step S41, when predicting the effectiveness of ad placement, the influence mechanisms of external environmental factors and competitor ad placement intensity on ad placement effectiveness are different, and there are also implicit pattern associations between external environmental factors and competitor ad placement intensity. Therefore, to optimize the accuracy of ad placement prediction, the ternary attention network is constructed. Specifically, the design of the ternary attention network aims to enhance the encoding vector of the temporal periodic pattern features of ad performance indicators. Ad performance itself is a complex concept; it depends not only on the quality of ad content but also on the influence of the external environment and competitors. By introducing an attention mechanism, it is possible to automatically identify which external factors have the greatest impact on ad performance and adjust the ad placement strategy accordingly.

[0036] In one example, such as Figure 4As shown, in step S41, the advertisement effect index time series periodic pattern feature encoding vector, the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector are input into a ternary attention network to obtain an enhanced advertisement effect index time series periodic pattern feature encoding vector, including: S411, calculating a feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector to obtain an external environment-competitive intelligence time series collaborative attention matrix; S412, performing feature activation based on a Sigmoid function on the external environment-competitive intelligence time series collaborative attention matrix to obtain an external environment-competitive intelligence time series collaborative attention mask matrix; S413, mapping the advertisement effect index time series periodic pattern feature encoding vector to the external environment-competitive intelligence time series collaborative attention mask matrix to obtain an attention-modulated advertisement effect index time series periodic pattern feature encoding vector; S414, inputting the attention-modulated advertisement effect index time series periodic pattern feature encoding vector and the advertisement effect index time series periodic pattern feature encoding vector into a residual unit to obtain the enhanced advertisement effect index time series periodic pattern feature encoding vector.

[0037] In one example, calculating a feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector to obtain an external environment-competitive intelligence time series collaborative attention matrix, includes: calculating a feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector to obtain the external environment-competitive intelligence time series collaborative attention matrix in the following formula; wherein the formula is: ; wherein, denotes the external environment variable time series periodic pattern feature encoding vector, denotes the competitive intelligence time series periodic pattern feature encoding vector, denotes matrix multiplication, denotes the transpose operation, is the feature length of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector, denotes an external environment-competitive intelligence time series collaborative attention matrix.

[0038] Specifically, first, a feature fine-grained correlation matrix between the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector needs to be calculated. Then, the external environment-competitive intelligence time series collaborative attention matrix is activated based on the Sigmoid function to obtain the external environment-competitive intelligence time series collaborative attention mask matrix. The role of the Sigmoid function is to map the numerical value to between 0 and 1, so that some features are more prominent, while other features are suppressed. In this way, the implicit relationship between the external environment variables and the competitive intelligence can be better captured. Map the advertisement effect index time series periodic pattern feature encoding vector onto the above attention mask matrix to obtain the attention modulated advertisement effect index time series periodic pattern feature encoding vector. The purpose of this step is to enable the advertisement effect index to better reflect the influence of the external environment and the competitive situation. For example, if the temperature is lower on a certain day and the competitor's advertising intensity is higher, the advertisement effect index on that day may be greatly affected, and attention modulation can help identify this influence. The attention modulated advertisement effect index time series periodic pattern feature encoding vector and the original advertisement effect index time series periodic pattern feature encoding vector are input into the residual unit. The role of the residual unit is to retain the original information while introducing new attention modulated information, thereby obtaining an enhanced advertisement effect index time series periodic pattern feature encoding vector.

[0039] In a preferred example, the feature fine-grained correlation matrix of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector is calculated to obtain an external environment-competitive intelligence time series collaborative attention matrix, comprising: performing distribution stabilization probability based on interaction global mean on the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector to obtain an external environment variable time series periodic pattern feature stabilized probability vector and a competitive intelligence time series periodic pattern feature stabilized probability vector; based on the feature mean of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector, calculating an interaction step length and a generalized interaction coefficient; based on the calculated interaction step length and generalized interaction coefficient, performing truncated approximation based on historical values on the external environment variable time series periodic pattern feature stabilized probability vector and the competitive intelligence time series periodic pattern feature stabilized probability vector to obtain an external environment variable time series periodic pattern feature truncated approximation vector and a competitive intelligence time series periodic pattern feature truncated approximation vector; calculate the feature fine-grained correlation matrix of the external environment variable time series periodic pattern feature truncated approximation vector and the competitive intelligence time series periodic pattern feature truncated approximation vector to obtain an external environment-competitive intelligence time series collaborative attention matrix.

[0040] Specifically, first, the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector are subjected to distribution stabilization probability based on interaction global mean, that is: ; wherein, and are the feature mean of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector , represents vector addition, represents Hadamard product, represents calculating the reciprocal of each feature value in the vector, is the external environment variable time series periodic pattern feature stabilization probability vector, is the competitive intelligence time series periodic pattern feature stabilization probability vector.

[0041] Then, the interaction step length of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector is obtained by , and further, the generalized interaction coefficient is calculated: ; wherein, represents rounding up, represents interaction step length, represents the factorial of , is the generalized interaction coefficient for the external environment variable time series periodic pattern feature encoding vector, is the generalized interaction coefficient for the competitive intelligence time series periodic pattern feature encoding vector; thus, the interaction long-tail distribution is flexibly inferred on the basis of historical interaction non-locality.

[0042] Finally, the truncation approximation based on historical values is performed in a gradual recursive manner: ; wherein, represents the external environment variable time series periodic pattern feature truncation approximation vector, represents the competitive intelligence time series periodic pattern feature truncation approximation vector.

[0043] That is, the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector The stable global historical interaction is taken as a boundary condition, and the non-local long-tail distribution is inferred from the global by a truncated approximation, so as to realize the explicit full-history discretization interaction and compensate for the possible global historical explicit inference deficiency of the feature fine-grained correlation.

[0044] Specifically, in step S42, the enhanced advertising effect index time series periodic pattern feature encoding vector, the external environment variable time series periodic pattern feature encoding vector, and the competitive intelligence time series periodic pattern feature encoding vector are fused to obtain an advertising launch multi-source time series fusion encoding vector. Specifically, the three encoding vectors are weighted and summed or spliced, etc., to form a multi-dimensional vector representation. This vector not only contains the features of the advertising effect itself, but also fuses the information of the external environment and the competitive situation, thereby providing a more comprehensive perspective. That is, the generation of the multi-source time series fusion encoding vector is to integrate multiple types of data and provide a more comprehensive feature representation. There may be complex interaction between different data sources, and it is difficult to fully understand the influencing factors of advertising effect by relying on a certain type of data. By fusing the time series encoding vectors of advertising effect indicators, external environment variables, and competitive intelligence, more rich feature representations can be extracted, thereby generating more accurate advertising launch prediction results.

[0045] Specifically, in step S43, the dynamic decision module based on the Softmax function is used to generate an advertising launch prediction result according to the fused encoding vector. The advantage of the Softmax function is that it can clearly present the probability distribution of multiple candidate schemes, which is convenient for subsequent decision making. For example, when facing multiple different advertising launch time period options, the Softmax function can give the probability value of each time period as the best launch time, helping advertisers choose the optimal launch time. In addition, the dynamic decision module can also dynamically adjust according to real-time data and historical data to ensure that the advertising launch strategy can quickly respond to market changes. This flexibility enables high accuracy while maintaining good real-time performance and scalability.

[0046] The specific steps of the dynamic decision module based on the Softmax function for generating an advertising launch prediction result according to the fused encoding vector are as follows: the Softmax function maps the multi-source time series fusion encoding vector to the probability distribution of each possible advertising launch time period. The probability of each time period reflects the possibility of that time period as the best launch time. For example, if there are multiple different advertising launch time period options, the Softmax function will give the probability value of each time period as the best launch time.

[0047] Here, the design of the ternary attention network and the dynamic decision module based on the Softmax function also takes into account the balance between computational efficiency and model complexity. The ternary attention network can significantly improve the quality of feature representation while keeping the model complexity relatively low by introducing an attention mechanism. The dynamic decision module based on the Softmax function provides a clear and intuitive basis for decision-making through the use of probability distribution. This design enables high accuracy while maintaining good real-time performance and scalability.

[0048] Specifically, in step S5, based on the advertising prediction results, a suggestion for the advertising time period is generated. It should be understood that the dynamic decision module based on the Softmax function will give a probability value for each time period as the best delivery opportunity. Based on the advertising prediction results, a specific advertising time period suggestion can be generated. Specifically, time periods with higher probability values are recommended as priority delivery periods. In addition, dynamic adjustments can be made according to market changes and real-time data. For example, if it is found that the probability value of a certain time period suddenly decreases (possibly due to increased advertising efforts by competitors or changes in weather conditions), the suggestion can be adjusted in time to avoid wasting resources. For example, if the probability value of a certain time period decreases from 55% to 45% throughout the day, it may be recommended to reduce the advertising frequency in this time period and instead increase the delivery in other high-success rate time periods.

[0049] The present application also provides a city advertising management system for executing the city advertising management method described above. Figure 5 Fig. 1 illustrates a structural schematic diagram of a city advertising management system according to an embodiment of the present application. As shown in Fig. 1, the city advertising management system comprises a data collection module 101, a feature extraction module 102, a ternary attention network 103, a dynamic decision module 104, and a suggestion generation module 105. Figure 5As shown, the urban advertisement delivery management system 500 includes: a time data input module 510 configured to input a time queue of advertisement effect indicators and a time queue of external environment variables within a predetermined time window, the predetermined time window being 7 days; a competitive intelligence input module 520 configured to input a time queue of competitive intelligence within the predetermined time window; a periodic pattern feature extraction module 530 configured to perform time series encoding on the time queue of advertisement effect indicators, the time queue of external environment variables, and the time queue of competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector, and a competitive intelligence time sequence periodic pattern feature encoding vector, wherein a first layer LSTM of the double-layer LSTM network is configured to capture intra-day time sequence periodic pattern features, and a second layer LSTM of the double-layer LSTM network is configured to capture cross-day time sequence periodic pattern features; an advertisement delivery prediction result determination module 540 configured to input the advertisement effect indicator time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector, and the competitive intelligence time sequence periodic pattern feature encoding vector into a dynamic decision module based on a Softmax function to obtain an advertisement delivery prediction result; and an advertisement delivery suggestion generation module 550 configured to generate an advertisement delivery time period suggestion based on the advertisement delivery prediction result.

[0050] Here, those skilled in the art can understand that the specific operations of each module in the above urban advertisement delivery management system have been described in detail above with reference to the description of the urban advertisement delivery management method Figures 1 to 4 of the above embodiments, and thus repeated descriptions thereof will be omitted.

[0051] The embodiments of the present application further provide a computer readable storage medium having computer program codes stored therein, which, when executed on a computer, cause the computer to perform the above related method steps to implement the urban advertisement delivery management method provided by the above embodiments.

[0052] The embodiments of the present application further provide a computer program product, which, when executed on a computer, causes the computer to perform the above related steps to implement the urban advertisement delivery management method provided by the above embodiments.

[0053] Among them, the system, computer readable storage medium or computer program product provided by the embodiments of the present application are all used to execute the corresponding method provided above, so the beneficial effects they can achieve can refer to the beneficial effects in the corresponding method provided above, which will not be described here.

[0054] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments.

[0055] The processes depicted in the figures do not require the particular order shown, or sequential order, to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous. The various embodiments described herein are described with reference to a particular sequential order, and the interrelationship between various elements depicted in the drawings are not necessarily to scale, and certain implementations can be practiced with other sequential orders, and other element interrelationships.

Claims

1. A city advertisement delivery management method characterized by, The method comprises the following steps: inputting a time queue of advertisement effect indicators and a time queue of external environment variables within a predetermined time window, the predetermined time window being 7 days; inputting a time queue of competitive intelligence within the predetermined time window; performing time series coding on the time queue of advertisement effect indicators, the time queue of external environment variables and the time queue of competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector and a competitive intelligence time sequence periodic pattern feature encoding vector, wherein a first layer LSTM of the double-layer LSTM network is used to capture intra-day time sequence periodic pattern features, and a second layer LSTM of the double-layer LSTM network is used to capture cross-day time sequence periodic pattern features; inputting the advertisement effect indicator time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector into a dynamic decision module based on a Softmax function to obtain an advertisement launching prediction result; and generating a suggestion for an advertisement launching time period based on the advertisement launching prediction result.

2. The urban advertisement delivery management method according to claim 1, characterized by, The advertisement effect indicators include click rates and conversion rates, the external environment variables include air temperatures, humidities and holiday flag bits, and the competitive intelligence is competitive advertisement launching intensity within 500 meters.

3. The urban advertisement delivery management method according to claim 2, characterized by, The first layer LSTM contains 64 neurons, and the second layer LSTM contains 32 neurons.

4. The urban advertisement delivery management method according to claim 3, characterized by, The method of performing time series coding on the time queue of advertisement effect indicators, the time queue of external environment variables and the time queue of competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect indicator time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector and a competitive intelligence time sequence periodic pattern feature encoding vector comprises the following steps: inputting the time queue of advertisement effect indicators, the time queue of external environment variables and the time queue of competitive intelligence into a first layer LSTM of the double-layer LSTM network to obtain a sequence distribution of advertisement effect indicator intra-day periodic pattern feature encoding vectors, a sequence distribution of external environment variable intra-day periodic pattern feature encoding vectors and a sequence distribution of competitive intelligence intra-day periodic pattern feature encoding vectors; and inputting the sequence distribution of advertisement effect indicator intra-day periodic pattern feature encoding vectors, the sequence distribution of external environment variable intra-day periodic pattern feature encoding vectors and the sequence distribution of competitive intelligence intra-day periodic pattern feature encoding vectors into a second layer LSTM of the double-layer LSTM network respectively to obtain the advertisement effect indicator time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector.

5. The urban advertisement delivery management method according to claim 4, characterized in that, The advertisement effect index time series periodic pattern feature encoding vector, the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector are input into a dynamic decision module based on a Softmax function to obtain an advertisement launching prediction result, including: inputting the advertisement effect index time series periodic pattern feature encoding vector, the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector into a ternary attention network to obtain an enhanced advertisement effect index time series periodic pattern feature encoding vector; fusing the enhanced advertisement effect index time series periodic pattern feature encoding vector, the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector to obtain an advertisement launching multi-source time series fusion encoding vector; inputting the advertisement launching multi-source time series fusion encoding vector into the dynamic decision module based on the Softmax function to obtain the advertisement launching prediction result.

6. The urban advertisement delivery management method according to claim 5, characterized by, The advertisement effect index time series periodic pattern feature encoding vector, the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector are input into a ternary attention network to obtain an enhanced advertisement effect index time series periodic pattern feature encoding vector, including: calculating a feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competition intelligence time series periodic pattern feature encoding vector to obtain an external environment-competition intelligence time series collaborative attention matrix; performing feature activation based on a Sigmoid function on the external environment-competition intelligence time series collaborative attention matrix to obtain an external environment-competition intelligence time series collaborative attention mask matrix; mapping the advertisement effect index time series periodic pattern feature encoding vector to the external environment-competition intelligence time series collaborative attention mask matrix to obtain an attention modulated advertisement effect index time series periodic pattern feature encoding vector; inputting the attention modulated advertisement effect index time series periodic pattern feature encoding vector and the advertisement effect index time series periodic pattern feature encoding vector into a residual unit to obtain the enhanced advertisement effect index time series periodic pattern feature encoding vector.

7. The urban advertisement delivery management method according to claim 6, characterized by, The feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector is calculated to obtain an external environment-competitive intelligence time series collaborative attention matrix, including: the feature fine-grained association matrix of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector is calculated to obtain the external environment-competitive intelligence time series collaborative attention matrix in the following formula; wherein the formula is: ; wherein, represents the external environment variable time series periodic pattern feature encoding vector, represents the competitive intelligence time series periodic pattern feature encoding vector, represents matrix multiplication, represents a transposition operation, is the feature length of the external environment variable time series periodic pattern feature encoding vector and the competitive intelligence time series periodic pattern feature encoding vector, represents an external environment-competitive intelligence time series collaborative attention matrix.

8. The urban advertisement delivery management method according to claim 7, characterized by, The feature fine-grained association matrix of the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector is calculated to obtain an external environment-competitive intelligence time sequence collaborative attention matrix, including: performing distribution stabilization probability based on interaction global mean on the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector to obtain an external environment variable time sequence periodic pattern feature stabilization probability vector and a competitive intelligence time sequence periodic pattern feature stabilization probability vector; calculating an interaction step and a generalized interaction coefficient based on the feature mean of the external environment variable time sequence periodic pattern feature encoding vector and the competitive intelligence time sequence periodic pattern feature encoding vector; performing truncated approximation based on historical values on the external environment variable time sequence periodic pattern feature stabilization probability vector and the competitive intelligence time sequence periodic pattern feature stabilization probability vector to obtain an external environment variable time sequence periodic pattern feature truncated approximation vector and a competitive intelligence time sequence periodic pattern feature truncated approximation vector based on the calculated interaction step and the generalized interaction coefficient; and calculating the feature fine-grained association matrix of the external environment variable time sequence periodic pattern feature truncated approximation vector and the competitive intelligence time sequence periodic pattern feature truncated approximation vector to obtain an external environment-competitive intelligence time sequence collaborative attention matrix.

9. The urban advertisement delivery management method according to claim 8, characterized by, The advertisement delivery prediction result is a probability value of each time period as the best delivery time.

10. A city advertisement delivery management system for executing the city advertisement delivery management method according to claims 1 to 9, characterized by The method comprises: a time data input module configured to input a time queue of an advertisement effect index and a time queue of an external environment variable within a predetermined time window, the predetermined time window being 7 days; a competitive intelligence input module configured to input a time queue of competitive intelligence within the predetermined time window; a periodic pattern feature extraction module configured to perform time series encoding on the time queue of the advertisement effect index, the time queue of the external environment variable, and the time queue of the competitive intelligence based on a double-layer LSTM network to obtain an advertisement effect index time sequence periodic pattern feature encoding vector, an external environment variable time sequence periodic pattern feature encoding vector, and a competitive intelligence time sequence periodic pattern feature encoding vector, wherein a first layer LSTM of the double-layer LSTM network is configured to capture intra-day time sequence periodic pattern features, and a second layer LSTM of the double-layer LSTM network is configured to capture cross-day time sequence periodic pattern features; an advertisement delivery prediction result determination module configured to input the advertisement effect index time sequence periodic pattern feature encoding vector, the external environment variable time sequence periodic pattern feature encoding vector, and the competitive intelligence time sequence periodic pattern feature encoding vector into a dynamic decision module based on a Softmax function to obtain an advertisement delivery prediction result; and an advertisement delivery suggestion generation module configured to generate a suggestion of an advertisement delivery time period based on the advertisement delivery prediction result.

Images