Information delivery methods, devices, electronic equipment and storage media
By combining channel features and user profile features with a deep learning model, the calculation of channel weights is optimized, which solves the problem of inaccurate attribution results in existing technologies and achieves more accurate allocation of channel resources.
Patent Information
- Application Number
- CN202210194551.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-01
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2042-03-01
AI Technical Summary
Existing attribution models in the field of information delivery suffer from inaccurate attribution results, leading to low accuracy in resource allocation across different channels.
By acquiring channel feature data and user profile feature data from the training set, a prediction model is trained using deep learning methods to determine the channel feature weights. Based on these weights, the information delivery channels are adjusted. By combining an attention mechanism and a gated recurrent unit (GRU) model, the weight calculation of channel features and user profile features is optimized.
It improves the accuracy of attribution results, enhances the accuracy of resource allocation across different channels, and provides a more reliable basis for resource allocation.
Smart Images

Figure CN116739665B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet technology, and in particular to an information delivery method, apparatus, electronic device, and storage medium. Background Technology
[0002] In the field of information delivery, information is typically delivered through multiple channels. For example, in e-commerce, platforms or advertisers will distribute product information through multiple channels such as search engines, advertisements, and short videos. Users may be reached through multiple channels before purchasing a product. Therefore, it is essential to measure the conversion efficiency of different channels and adjust subsequent resource allocation based on the conversion efficiency of each channel.
[0003] In existing technologies, attribution models are used to determine the conversion rate of each channel. However, current attribution models usually only focus on channel characteristics. In other words, existing technologies have inaccurate attribution results, which leads to low accuracy in the subsequent allocation of resources to different channels. Summary of the Invention
[0004] This invention provides an information delivery method, apparatus, electronic device, and storage medium to solve the technical problem in the prior art where inaccurate attribution results lead to low accuracy in the subsequent allocation of resources across different channels.
[0005] In a first aspect, embodiments of the present invention provide an information delivery method, comprising: acquiring a training set, wherein the training set includes channel feature data, user profile feature data, and feature tag values of the channels through which target information is delivered within each time period of a first preset time period; training a prediction model to be trained based on the training set to obtain a prediction model; determining channel feature weights according to the prediction model; and adjusting the delivery channels of the target information according to the channel feature weights.
[0006] As an optional embodiment, the user profile feature data includes user behavior feature data and user category feature data.
[0007] As an optional embodiment, training the prediction model to be trained based on the training set to obtain the prediction model includes: inputting channel feature data of each time period into a gated recurrent unit (GRU) to obtain candidate hidden state sequence data, and obtaining final hidden state sequence data based on the candidate hidden state sequence data; determining the weight coefficients corresponding to the channel feature data of each time period based on an attention mechanism; performing a weighted summation on the final hidden state sequence data according to the weight coefficients to obtain aggregated feature weights; performing logarithmic transformation and normalization on the user behavior feature data to obtain processed user behavior feature data; inputting the user category feature data into an embedding layer to obtain processed user category feature data; concatenating the processed user behavior feature data and the processed user category feature data, and passing them through an activation function to obtain processed user profile feature data; and training the prediction model to be trained based on the aggregated feature weights, channel feature data, and processed user profile feature data until a preset condition is met to obtain the prediction model.
[0008] As an optional embodiment, the step of determining the weight coefficients corresponding to the channel feature data for each time period based on the attention mechanism includes: marking target users according to user behavior feature data, wherein the target users are users who are inactive in a first target time period but active in a second target time period; obtaining the channel feature data corresponding to the target information accessed by the target users in the second target time period, and determining the channel weight correction value according to the channel feature data corresponding to the target information accessed by the target users in the second target time period; determining the weight coefficients corresponding to the channel feature data for each time period according to the candidate hidden state sequence data, the channel weight correction value, and the time difference parameter, wherein the time difference parameter is the difference between the time the channel appears in a preset time period and the current time.
[0009] As an optional embodiment, determining the channel feature weights according to the prediction model includes: determining the channel feature weights according to the aggregated feature weights.
[0010] As an optional embodiment, the method further includes: obtaining a test set, the test set including channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered within each time period of a second preset time period; determining the evaluation index of the prediction model based on the test set; wherein the evaluation index includes at least one of the following: prediction accuracy AUC, channel feature weights and the stability of the prediction model.
[0011] As an optional embodiment, the stability of the prediction model includes ACU stability and / or channel feature weight stability.
[0012] Secondly, embodiments of the present invention provide an information delivery device, comprising: an acquisition module for acquiring a training set, wherein the training set includes channel feature data, user profile feature data, and feature tag values of the channels through which target information is delivered within each time period of a first preset time period; a training module for training a prediction model to be trained based on the training set to obtain a prediction model; and an adjustment module for determining channel feature weights according to the prediction model and adjusting the delivery channels of the target information according to the channel feature weights.
[0013] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor is used to implement the steps of the information delivery method described in any one of the first aspects when executing the program stored in the memory.
[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the information delivery method as described in any of the first aspects.
[0015] The information delivery method, apparatus, electronic device, and storage medium provided in this invention obtain a training set, wherein the training set includes channel feature data, user profile feature data, and feature tag values of the channels through which target information is delivered within each time period of a first preset time period; a prediction model to be trained is trained based on the training set to obtain a prediction model; channel feature weights are determined according to the prediction model, and the delivery channels of target information are adjusted according to the channel feature weights; that is, this invention simultaneously considers the temporal sequence of channel features and user profile features, thereby obtaining a more accurate prediction model, and thus obtaining more accurate channel feature weights, providing a more reliable basis for subsequent resource allocation of target information on different channels. Attached Figure Description
[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating the information delivery method provided in an embodiment of the present invention;
[0019] Figure 2 A flowchart illustrating an information delivery method provided in an embodiment of the present invention;
[0020] Figure 3 This is a flowchart illustrating the detailed implementation of step S102 in an embodiment of the present invention;
[0021] Figure 4 A training architecture diagram of a prediction model provided in an embodiment of the present invention;
[0022] Figure 5 This is a schematic diagram of the structure of an information delivery device provided in an embodiment of the present invention;
[0023] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] The era of rampant, unregulated growth for most internet products is over. The aggressive, land-grabbing approach needs to shift towards a more intensive, nurturing strategy, requiring sophisticated operations. In e-commerce, a user's journey from seeing a product to making a purchase involves a complex path. A user might initially search for a keyword, encounter a related ad, or be directed to the platform via short videos or live streaming. In short, users may be reached through multiple channels before making a purchase. Therefore, e-commerce operators need to measure the conversion efficiency of different channels or platforms and decide how much resource to invest in each.
[0026] How to allocate and adjust resources across multiple channels is typically analyzed using attribution models. Common attribution models include:
[0027] (1) Last Interaction Attribution Model, also known as the Last Click Model. The channel that makes the last interaction gets 100% credit for the transaction. This is the simplest and most direct model, and it is suitable for businesses with short conversion time intervals and simple conversion paths, such as purchases or downloads upon entry.
[0028] (2) Initial interaction attribution model, also known as the first click model. The first channel gets 100% credit. This model focuses on customer acquisition or more specific business scenarios and is suitable for the initial business expansion period.
[0029] (3) The location-based interaction attribution model has the largest weights for the first and last channels, while the weights for the middle channels are average.
[0030] (3) Linear attribution model, where all channels (platforms) along the user path have the same weight.
[0031] (4) The time decay attribution model has the highest weight for the last interaction channel (platform) and gradually decreases the weight for subsequent interactions.
[0032] (5) Custom attribution model: calculate attribution based on a certain algorithm or rule, and the audience can define an attribution model that suits them.
[0033] (6) Markov attribution model, which calculates the actual contribution of each interaction in the conversion path based on user data.
[0034] (7) The Shapley Value model calculates the contribution value of each interaction based on marginal effects. It was developed in cooperative game theory.
[0035] However, both the final interaction attribution model and the initial interaction attribution model clearly ignore the impact of other channels on attribution; location-based interaction attribution models do not consider the effectiveness of post-conversion touchpoint marketing; linear attribution models do not differentiate the importance of channels, making it easy to miss key points, especially in complex scenarios; time decay attribution models are not favorable to e-commerce operators, as channels at the top of the conversion funnel do not receive high weight; Markov attribution models do not consider different user habits and only consider the previous state; and the Shapley value model completely ignores user characteristics. In summary, current attribution models suffer from technical problems of inaccurate attribution.
[0036] To address the aforementioned technical problems, the technical concept of this invention is to propose a deep learning attribution method that combines user profile feature data. This method not only considers the temporal characteristics of each channel but also the influence of different user profile features, thereby improving the accuracy of the attribution results and further enhancing the accuracy of resource allocation for different channels.
[0037] In a first aspect, embodiments of the present invention provide an information delivery method. Figure 1 This is a flowchart illustrating the information delivery method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the information delivery method includes:
[0038] Step S101: Obtain the training set, wherein the training set includes channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered within each time period of the first preset time period.
[0039] Specifically, the target information can be product information or product links placed by e-commerce operators on different channels (such as search, recommendations, advertisements, SMS, etc.). When a user sees product information or a product link that interests them on a certain channel, they can click on the target information to enter the e-commerce platform where the product is located and complete the purchase. The first preset time period can be any historical time before the current time. The first preset time period can be divided into multiple time periods to obtain channel feature data within each time period, as well as user profile feature data and corresponding feature tag values.
[0040] For example, when an e-commerce operator needs to evaluate the resource allocation ratios for a certain mobile phone category across different channels from September 1, 2021 to September 7, 2021, these resources include, but are not limited to, search, recommendations, advertising, SMS, coupons, red envelopes, and even consumption on and off-site channels. In this step, characteristic data for all channels can be collected for each day of the 14 days prior to September 1, 2021 (i.e., the first preset time period is 14 days, with one day as a time cycle). This includes numerical data such as the number of visits, browsing duration, and clicks for each channel. If a channel is inactive on a particular day, it is supplemented with a characteristic value of 0.
[0041] As an optional embodiment, the user profile feature data includes user behavior feature data and user category feature data. Specifically, user behavior feature data can be channel feature data of accessing target information in different historical time periods, such as various numerical features of users in search, recommendation or advertising channels in the last 30 days, 15 days, 7 days and 3 days; user category features can be user characteristics such as gender, age, occupation, etc.
[0042] The training set is shown in Table 1, including channel feature data F_r1, F_r2, ..., F_r 14 The user profile feature data includes user behavior feature data UD_1, UD_2...UD_m and user category feature data UCat_1, UCat_2...UCat_n; as well as feature label values, where label=1 represents that the user has placed an order and label=0 represents that the user has not placed an order or converted on the last day of the first preset time period.
[0043] Table 1
[0044]
[0045] Optionally, before step S101, the process further includes: processing the raw data to obtain a training set. Specifically, the raw data collected on e-commerce platforms is relatively messy and scattered, so it is necessary to process the raw data first to organize it into three categories: channel characteristics, user behavior characteristics, and user category characteristics.
[0046] Step S102: Train the prediction model to be trained based on the training set to obtain the prediction model.
[0047] Specifically, the prediction model can be a multilayer perceptron neural network model. The prediction model is trained using the aforementioned training set. The input to the trained prediction model is the channel feature data and user profile feature data (including user behavior feature data and user category feature data) of the sample to be predicted. The output is whether the sample to be predicted has completed the order conversion or other target actions specified by the e-commerce operator.
[0048] Step S103: Determine the channel feature weights according to the prediction model, and determine the target information delivery channels according to the channel feature weights.
[0049] Specifically, after obtaining the prediction model, the channel feature weight of each channel will be determined based on the weights related to the channel feature data in the prediction model, and the resource allocation or marketing strategy of the target information on different channels will be adjusted according to the channel feature weight.
[0050] Figure 2 This is a flowchart of an information delivery method provided in an embodiment of the present invention, as shown below. Figure 2 As shown, firstly, the acquired raw data is processed into the required data, including channel feature data, user behavior feature data, and user category feature data. Then, the channel feature data, user behavior feature data, and user category feature data are used as training sets to train the prediction model to be trained, obtaining the trained prediction model and channel feature weights. Finally, the bidding strategy and marketing strategy of the target information on different channels are adjusted according to the channel feature weights.
[0051] The information delivery method provided in this embodiment of the invention obtains a training set, wherein the training set includes channel feature data, user profile feature data, and feature tag values of the channels through which target information is delivered within each time period of a first preset time period; a prediction model to be trained is trained based on the training set to obtain a prediction model; channel feature weights are determined according to the prediction model, and the delivery channels of target information are adjusted according to the channel feature weights; that is, this embodiment of the invention simultaneously considers the temporal sequence of channel features and user profile features, thereby obtaining a more accurate prediction model, and thus obtaining more accurate channel feature weights, providing a more reliable basis for subsequent resource allocation of target information on different channels.
[0052] Based on the above embodiments, Figure 3 This is a detailed flowchart illustrating the implementation of step S102 in an embodiment of the present invention, as shown below. Figure 3 As shown, step S102 includes:
[0053] Step S1021: Input the channel feature data of each time period into the gated loop unit (GRU) to obtain candidate hidden state sequence data, and obtain the final hidden state sequence data based on the candidate hidden state sequence data.
[0054] Taking the training set shown in Table 1 as an example, the channel feature data F_r1, F_r2, ... F_r for each time period are... 14 The data is fed into a GRU to obtain candidate hidden state sequence data. As shown in formula (1); and based on the candidate hidden state sequence data Obtain the final hidden state sequence data F_d1, F_d2...F_d 14 Its formula is shown in (2).
[0055]
[0056]
[0057] in, This represents the candidate hidden state at the current time step t (t = 1, 2, ..., 14). This represents the candidate hidden state at the previous time step t-1, where w1 is the initialization parameter; F_d t F_d represents the final hidden state at the current time step t. t-1 This represents the final hidden state at the previous time step t-1. Z represents the gating signal, which ranges from 0 to 1. The closer it is to 1, the more data is "remembered"; while the closer it is to 0, the more data is "forgotten". w1 is the initialization parameter.
[0058] Step S1022: Determine the weight coefficients corresponding to the channel feature data for each time period based on the attention mechanism.
[0059] Specifically, F_r1, F_r2, ..., F_r are determined based on the attention mechanism for each time period. 14 The corresponding weight coefficients F_a1, F_a2, ..., F_a are respectively. 14 .
[0060] Step S1023: Perform a weighted summation on the final hidden state sequence data according to the weight coefficients to obtain the aggregated feature weights.
[0061] Specifically, according to F_a1, F_a2...F_a 14 For the final hidden state sequence data F_d1, F_d2...F_d 14 Perform a weighted summation to obtain the aggregated feature weights F. weight As shown in formula (3):
[0062]
[0063] Step S1024: Perform logarithmic transformation and normalization on the user behavior feature data to obtain the processed user behavior feature data.
[0064] Specifically, the user behavior action features UD_1, UD_2...UD_m in the training set shown in Table 1 are first transformed by log and then normalized to obtain the processed user behavior feature data UDN_1, UDN_2...UDN_m.
[0065] Among them, normalization methods include linear function (min-max) normalization and zero mean (Z-score) normalization. The min-max normalization is shown in formula (4):
[0066] UDN=(log(UD)-min(log(UD)) / (max(log(UD))-min(log(UD)))) (4)
[0067] Step S1025: Input the user category feature data into the embedding layer to obtain the processed user category feature data.
[0068] Specifically, the user category features UCat_1, UCat_2...UCat_n in the training set shown in Table 1 are input into the embedding layer to obtain the dimensionality-reduced user category feature data.
[0069] Step S1026: Concatenate the processed user behavior feature data and the processed user category feature data, and obtain the processed user profile feature data through an activation function.
[0070] Specifically, the processed user behavior feature data UDN_1, UDN_2...UDN_m and the dimensionality-reduced user category feature data are concatenated together and then processed through a three-layer ReLU activation function to obtain the final user profile feature data (User Feature, or UF for short).
[0071] Step S1027: Train the prediction model to be trained according to the aggregated feature weights, channel feature data and processed user profile feature data until the preset conditions are met, and obtain the prediction model.
[0072] Specifically, based on the aggregated feature weight F weight The channel feature data is obtained after processing. The processed channel feature data and the processed user profile feature data are combined to train the prediction model to be trained (such as CVR, GMV, CTR, etc.) until the preset conditions are met and the prediction model is obtained.
[0073] Optionally, the preset condition can be a preset optimization function, as shown in formula (5):
[0074]
[0075] Where m is the total number of samples in the training set, y i Indicates whether the i-th sample has placed an order for conversion, h(UF) i ) represents whether the predicted sample will be converted to an order, and J(Θ) represents the optimization function.
[0076] As an optional embodiment, step S1022 includes: marking target users according to user behavior feature data, wherein the target users are users who are inactive during a first target time period but active during a second target time period; obtaining channel feature data corresponding to the target information accessed by the target users during the second target time period, and determining channel weight correction values based on the channel feature data corresponding to the target information accessed by the target users during the second target time period; determining the weight coefficients corresponding to the channel feature data of each time period based on the candidate hidden state sequence data, the channel weight correction values, and the time difference parameter, wherein the time difference parameter is the difference between the time the channel appears in a preset time period and the current time.
[0077] Specifically, users who have been inactive for more than 30 days but have been active in the last 14 days can be identified as target users based on user behavior data. The system then obtains channel feature data showing these target users' access to target information within the last 14 days and learns a channel weight correction value w2 based on this data. This approach optimizes attribution weights by using long-inactive users as cold starters, avoiding the problem of large discrepancies in attribution results from the same attribution algorithm across different periods. Then, based on candidate hidden state sequence data... Channel weight correction value w2 and time difference parameter T t Determine the weighting coefficients F_a1, F_a2, ..., F_a 14 Its formula is shown in (6):
[0078]
[0079] Where λ is a parameter greater than 0, and Tt is the difference between the time the channel appears and the time the channel ends on a daily basis. The larger Tt is, the greater the penalty for similarity distance. It is believed that the further away the number of days the channel is active is from the present, the further away the contribution will be. Conversely, if the channel is active and the conversion is on the same day, then the penalty is 0. Then, substitute formula (6) into formula (3) to obtain the aggregated feature weights.
[0080] As an optional embodiment, determining the channel feature weights according to the prediction model includes: determining the channel feature weights based on the aggregated feature weights. Specifically, the channel feature weights corresponding to each channel can be obtained by averaging the aggregated feature weights or through other processing.
[0081] Figure 4 This is a training architecture diagram of a prediction model provided in an embodiment of the present invention. The input features of the prediction model are shown in Table 1, including channel feature data F_r1, F_r2, ..., F_r14, user profile feature data including user behavior feature data UD_1, UD_2, ..., UD_m, and user category features UCat_1, UCat_2, ..., UCat_n.
[0082] First, the channel feature data is processed through a gated recurrent unit (GRU) to retain temporal information. Specifically, candidate hidden state sequence data is obtained according to formula (1), and final hidden state sequence data F_d1, F_d2, ..., F_d are obtained according to formula (2). 14 Then F_d1, F_d2...F_d 14The input is fed into a fully connected layer (Dense Layer) and then into an attention layer (Attention Layer). The attention layer also includes a Softmax function for normalization. The attention layer will introduce channel feature data of users who have been inactive for more than 30 days and active in the last 14 days. The weight coefficients F_a1, F_a2...F_a are obtained according to formula (6). 14 Then, according to formula (3), the aggregated feature weights are obtained, and then the processed channel feature data is obtained according to the aggregated feature weights and channel feature coefficients.
[0083] User behavior feature data UD_1, UD_2...UD_m are processed through a log trans and normalization layer to obtain processed user behavior feature data UDN_1, UDN_2...UDN_m; user category feature data UCat_1, UCat_2...UCat_n are input into an embedding layer to obtain a dimensionality-reduced user category feature vector; UDN_1, UDN_2...UDN_m and the dimensionality-reduced user category feature vector are concatenated and input into a three-layer ReLU layer to obtain UF.
[0084] Then, the processed channel feature data and UF are combined for feature processing, and the prediction model to be trained is trained to obtain the prediction model.
[0085] As an optional embodiment, the method further includes: obtaining a test set, the test set including channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered within each time period of a second preset time period; determining the evaluation index of the prediction model based on the test set; wherein the evaluation index includes at least one of the following: prediction accuracy AUC, channel feature weights and the stability of the prediction model.
[0086] Specifically, after training the prediction model, a test set is used to evaluate its performance. In this embodiment, three evaluation metrics are used to evaluate the prediction model. The first evaluation metric is the prediction accuracy (Area UnderCurve, or AUC), which represents the correctness of the prediction model itself, as shown in formula (7). The second evaluation metric is the magnitude of the output channel feature weights and the downstream task that makes decisions based on the magnitude of the channel feature weights. The third evaluation metric is the stability of the prediction model. Optionally, the stability of the prediction model includes the stability of AUC and / or the stability of the channel feature weights. The stability of AUC is shown in formula (8), and the stability of the channel feature weights is shown in formula (9). That is, the daily fluctuation of the prediction model's AUC should not be too large, and the daily fluctuation of the channel feature weights should not be too large either, because the output of the prediction model needs to be used for downstream decision-making. Therefore, it is necessary to ensure that the changes in AUC_Range and Weight_Range are minimal within a prediction period, otherwise it will affect the performance and decision-making.
[0087]
[0088] Among them, pred pos To predict the probability of a positive sample, pred neg To predict the probability of negative samples, N pos N represents the number of positive samples. neg This represents the number of negative samples.
[0089]
[0090]
[0091] The information delivery method provided in this invention involves inputting channel feature data from each time period into a gated recurrent unit (GRU) to obtain candidate hidden state sequence data, and then obtaining final hidden state sequence data based on the candidate hidden state sequence data. Weight coefficients corresponding to the channel feature data from each time period are determined based on an attention mechanism. The final hidden state sequence data is then weighted and summed according to the weight coefficients to obtain aggregated feature weights. Logarithmic transformation and normalization are performed on the user behavior feature data to obtain processed user behavior feature data. The user category feature data is input into an embedding layer to obtain processed user category feature data. The processed user behavior feature data and the processed user category feature data are concatenated and activated using an activation function to obtain processed user profile feature data. The prediction model to be trained is then trained based on the aggregated feature weights, channel feature data, and processed user profile feature data until preset conditions are met to obtain the prediction model. In other words, this invention simultaneously considers the temporal nature of channel features and user profile features, resulting in more accurate predictions. The predictive model obtains more accurate channel feature weights, providing a more reliable basis for subsequent resource allocation of target information on different channels. Furthermore, target users are marked based on user behavior feature data, where the target users are inactive during a first target time period but active during a second target time period. Channel feature data corresponding to the target users' access to target information during the second target time period is obtained, and channel weight correction values are determined based on this data. Weight coefficients corresponding to channel feature data for each time period are determined based on the candidate hidden state sequence data, channel weight correction values, and time difference parameters, where the time difference parameter is the difference between the time a channel appears within a preset time period and the current time. In other words, this embodiment initializes the weights of each channel by expanding the natural population data, while penalizing channels far from conversion rates. It considers population characteristics, time-series characteristics, and the weight coefficients of natural population data, achieving significant effects in terms of model interpretability and operability.
[0092] Secondly, embodiments of the present invention provide an information delivery device. Figure 5 This is a schematic diagram of the structure of an information delivery device provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the information delivery device includes:
[0093] The acquisition module 10 is used to acquire a training set, wherein the training set includes channel feature data, user profile feature data and feature label values of the channels through which the target information is delivered within each time period of the first preset time period; the training module 20 is used to train the prediction model to be trained based on the training set to obtain the prediction model; the adjustment module 30 is used to determine the channel feature weights according to the prediction model and adjust the delivery channels of the target information according to the channel feature weights.
[0094] As an optional embodiment of the present invention, the user profile feature data includes user behavior feature data and user category feature data.
[0095] As an optional embodiment of the present invention, the training module 20 is specifically used for: inputting channel feature data of each time period into a gated recurrent unit (GRU) to obtain candidate hidden state sequence data, and obtaining final hidden state sequence data based on the candidate hidden state sequence data; determining the weight coefficients corresponding to the channel feature data of each time period based on an attention mechanism; performing a weighted summation on the final hidden state sequence data according to the weight coefficients to obtain aggregated feature weights; performing logarithmic transformation and normalization on the user behavior feature data to obtain processed user behavior feature data; inputting the user category feature data into an embedding layer to obtain processed user category feature data; concatenating the processed user behavior feature data and the processed user category feature data, and obtaining processed user profile feature data through an activation function; and training the prediction model to be trained according to the aggregated feature weights, channel feature data, and processed user profile feature data until a preset condition is met to obtain the prediction model.
[0096] As an optional embodiment of the present invention, the training module 20 is specifically used for: marking target users according to user behavior feature data, wherein the target users are users who are inactive in a first target time period but active in a second target time period; acquiring channel feature data corresponding to the target information accessed by the target users in the second target time period, and determining channel weight correction values based on the channel feature data corresponding to the target information accessed by the target users in the second target time period; determining the weight coefficients corresponding to the channel feature data of each time period based on the candidate hidden state sequence data, the channel weight correction values, and the time difference parameter, wherein the time difference parameter is the difference between the time the channel appears in a preset time period and the current time.
[0097] As an optional embodiment of the present invention, the adjustment module 30 is specifically used to: determine the channel feature weights based on the aggregated feature weights.
[0098] As an optional embodiment of the present invention, the device further includes an evaluation module 40; the evaluation module 40 is specifically used for: acquiring a test set, the test set including channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered in each time period of a second preset time period; determining the evaluation index of the prediction model based on the test set; wherein the evaluation index includes at least one of the following: prediction accuracy AUC, channel feature weight and prediction model stability.
[0099] As an optional embodiment of the present invention, the stability of the prediction model includes ACU stability and / or channel feature weight stability.
[0100] The information delivery device provided in this embodiment of the invention has a similar implementation principle and technical effect to the above embodiments, and will not be described again here.
[0101] like Figure 6 As shown, this embodiment of the invention provides an electronic device, including a processor 111, a communication interface 112, a memory 113, and a communication bus 114, wherein the processor 111, the communication interface 112, and the memory 113 communicate with each other through the communication bus 114.
[0102] Memory 113 is used to store computer programs;
[0103] In one embodiment of the present invention, when the processor 111 executes the program stored in the memory 113, it implements the steps of the information delivery method provided in any of the foregoing method embodiments.
[0104] The electronic device provided in this embodiment of the invention has a similar implementation principle and technical effect to the above embodiments, and will not be described again here.
[0105] The aforementioned memory 113 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 113 has storage space for program code used to perform any of the method steps described above. For example, the storage space for program code may include individual program codes for implementing the various steps in the methods described above. This program code can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, optical discs (CDs), memory cards, or floppy disks. Such computer program products are typically portable or fixed storage units. The storage unit may have storage segments or storage spaces arranged similarly to memory 113 in the aforementioned electronic device. The program code may be compressed, for example, in a suitable form. Typically, the storage unit includes programs for performing the method steps according to embodiments of this disclosure, i.e., code that can be read by a processor such as 111, which, when run by the electronic device, causes the electronic device to perform the various steps in the methods described above.
[0106] Embodiments of the present invention also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the information delivery method described above.
[0107] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.
[0108] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0109] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0110] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.
Claims
1. An information delivery method, characterized by, include: Obtain a training set, wherein the training set includes channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered within each time period of the first preset time period; The prediction model to be trained is trained based on the training set to obtain the prediction model; The channel feature weights are determined based on the prediction model, and the channels for distributing target information are adjusted according to the channel feature weights. The user profile feature data includes user behavior feature data and user category feature data; The step of training the prediction model to be trained based on the training set to obtain the prediction model includes: Channel feature data for each time period is input into the gated recurrent unit (GRU) to obtain candidate hidden state sequence data, and the final hidden state sequence data is obtained based on the candidate hidden state sequence data. The weight coefficients corresponding to the channel feature data of each time period are determined based on the attention mechanism. The final hidden state sequence data is weighted and summed according to the weight coefficients to obtain the aggregated feature weights. The user behavior feature data is subjected to logarithmic transformation and normalization to obtain processed user behavior feature data. The user category feature data is input into the embedding layer to obtain processed user category feature data; The processed user behavior feature data and the processed user category feature data are concatenated, and then processed user profile feature data is obtained by using an activation function. The prediction model to be trained is trained based on the aggregated feature weights, channel feature data, and processed user profile feature data until the preset conditions are met, thereby obtaining the prediction model.
2. The method of claim 1, wherein, The weighting coefficients for determining the channel feature data for each time period based on the attention mechanism include: Target users are identified based on user behavior characteristic data, wherein the target users are those who are inactive during a first target time period but active during a second target time period; Obtain channel feature data corresponding to the target user's access to target information within the second target time period, and determine the channel weight correction value based on the channel feature data corresponding to the target user's access to target information within the second target time period; The weight coefficients corresponding to the channel feature data for each time period are determined based on the candidate hidden state sequence data, channel weight correction value, and time difference parameter, wherein the time difference parameter is the difference between the time when the channel appears within a preset time period and the current time.
3. The method according to claim 1 or 2, characterized in that, Determining the channel feature weights based on the prediction model includes: The channel feature weights are determined based on the aggregated feature weights.
4. The method according to claim 1 or 2, characterized in that, The method further includes: Obtain a test set, which includes channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered within each time period of the second preset time period; The evaluation metrics for the prediction model are determined based on the test set. The evaluation metrics include at least one of the following: prediction accuracy AUC, channel feature weights, and prediction model stability.
5. The method of claim 4, wherein, The stability of the prediction model includes AUC stability and / or channel feature weight stability.
6. An information delivery apparatus, characterized by comprising: The apparatus for the information delivery method as described in any one of claims 1-5 includes: The acquisition module is used to acquire a training set, wherein the training set includes channel feature data, user profile feature data and feature tag values of the channels through which the target information is delivered within each time period of a first preset time period; The training module is used to train the prediction model to be trained based on the training set to obtain the prediction model; The adjustment module is used to determine the channel feature weights based on the prediction model, and adjust the distribution channels of the target information based on the channel feature weights.
7. An electronic device, comprising: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in a memory, it implements the steps of the information delivery method according to any one of claims 1-5.
8. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the information delivery method as described in any one of claims 1-5.
Citation Information
Patent Citations
Internet information putting channel optimizing system based on distributed computing
CN104133837A
Prediction model generation method and device, equipment and readable medium
CN113283948A
