Accurate traffic pushing method and system based on user behavior prediction, and storage medium
Through deep learning models and Bayesian networks, and combined with reinforcement learning algorithms to adjust the advertising traffic push strategy, the problem of difficult to match the dynamic changes in user behavior during RTB advertising delivery is solved, and efficient and accurate advertising traffic allocation and conversion rate improvement is achieved.
Patent Information
- Application Number
- CN202411581498.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-07
- Publication Date
- 2025-02-25
AI Technical Summary
The existing RTB advertising delivery model is difficult to adapt to the dynamic changes in user behavior, resulting in DSPs not being able to obtain traffic that accurately matches their needs, wasting resources or missing the best delivery opportunity.
By obtaining the user's multi-dimensional behavior data, preprocessing and prediction using deep learning models and Bayesian networks, user behavior prediction data is generated, and advertising traffic is sorted and priority allocation is allocated based on this, and push strategy is adjusted through reinforcement learning algorithms.
It realizes accurate prediction of user behavior and efficient allocation of traffic, significantly improves advertising conversion rate and traffic utilization rate, reduces delivery costs, and speeds up response speed.
Smart Images
Figure CN119515474A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of data processing, and in particular to a method, system and storage medium for accurately pushing traffic based on user behavior prediction. Background Art
[0002] With the rapid development of digital advertising, real-time bidding (RTB) has become the core technology in programmatic advertising. In the RTB system, the advertising exchange (ADX) sends advertising traffic to the demand-side platform (DSP) in real time through an auction mechanism, and the DSP bids for this traffic based on its own needs and bidding strategies. However, with the continuous increase in user behavior data and the increasing demand of advertisers for precise delivery, the existing RTB advertising delivery model faces some challenges.
[0003] In traditional RTB systems, traffic push often relies on rules or historical bidding records. Although this method can provide a certain reference, it is difficult to adapt to the dynamic changes in user behavior due to its limitations. DSP may waste resources or miss the best delivery opportunities because it cannot obtain traffic that accurately matches its needs.
[0004] At present, some systems try to optimize traffic delivery through data analysis and traffic filtering, but they are still insufficient in accuracy and response speed. Especially in the face of complex and changeable user behavior, traditional systems find it difficult to effectively predict user needs and respond in real time. Therefore, how to achieve accurate advertising delivery of DSP through more intelligent traffic prediction and push methods has become a technical problem that needs to be solved urgently. Summary of the invention
[0005] In order to achieve accurate delivery of DSP advertisements through a more intelligent traffic prediction and push method, the present application provides a method, system and storage medium for accurate traffic push based on user behavior prediction.
[0006] In the first aspect, the present application provides a method for accurately pushing traffic based on user behavior prediction, which adopts the following technical solution:
[0007] A method for accurately pushing traffic based on user behavior prediction includes the following steps:
[0008] Acquire multi-dimensional behavior data of the user, and pre-process the multi-dimensional behavior data to obtain a multi-dimensional feature vector;
[0009] Inputting the multi-dimensional feature vector into a deep learning model and a Bayesian network to generate user behavior prediction data;
[0010] Based on the user behavior prediction data, the advertisement traffic available for transmission is sorted by type, wherein the advertisement traffic corresponds to an advertisement type, and the advertisement type has a corresponding type matching degree with the type of the user behavior prediction data, and the type is sorted according to the type matching degree, and the advertisement traffic corresponding to the advertisement type with a preset ranking in the result of the type sorting is pushed to the DSP;
[0011] Calculate push parameters using a push algorithm based on the real-time bidding data and budget data of the DSP, calculate the similarity between the advertisement type in the pushed advertisement traffic and the push parameters according to a similarity algorithm as similarity data, and prioritize and deliver the advertisement traffic according to an allocation algorithm based on the similarity data; generate delivery feedback data after delivery;
[0012] Collecting the delivery feedback data of the DSP, and obtaining advertisement feedback information after the advertisement traffic is delivered from the delivery feedback data;
[0013] A reinforcement learning algorithm is used to adjust parameters in the similarity algorithm based on the advertisement feedback information.
[0014] By adopting the above technical solutions, artificial intelligence technology is used to predict user behavior, and combined with real-time data, advertising traffic that meets their needs is accurately pushed to the demand-side platform (DSP). The entire system includes the following main steps: real-time analysis of user behavior data through deep learning models, predicting user needs and behavior patterns in different time periods, and ensuring that the traffic pushed to the DSP is more in line with actual needs; according to the prediction results, the needs of each DSP are dynamically adjusted to push the traffic that is most likely to increase the conversion rate and reduce invalid traffic; through real-time analysis and intelligent decision-making systems, the response time of traffic push is shortened to ensure that the DSP can obtain ideal traffic in the correct time period; using AI technology to intelligently analyze DSP needs, balance the traffic acquisition of each DSP, and avoid resource waste or excess traffic; it aims to optimize traffic allocation in RTB advertising and improve advertising conversion rate and delivery efficiency.
[0015] Optionally, the deep learning model includes LSTM, wherein:
[0016] The input layer includes the historical behavior data of the input user, represented as a time series {x1, x2…xt}, where each xt represents the behavior characteristics of the user at time t;
[0017] The LSTM layer includes a forget gate, an input gate, a cell state update gate, a hidden layer state gate, and an output gate;
[0018] Forget gate: F(t) = σ(Wf×[H(t-1), Xt]+Bf), Wf is the weight matrix, Bf is the bias, and σ is the Sigmoid activation function;
[0019] Input gate: It = σ(Wi × [H(t-1), Xt] + Bi);
[0020] Cell state update gate: Ct = Ft × C(t-1) + It × tanh(Wc × [H(t-1), Xt] + Bc);
[0021] Output gate: Ot = σ(Wo × [H(t-1), Xt] + Bo);
[0022] Hidden layer state gate: Ht = Ot × tanh (Ct);
[0023] The output layer includes the prediction of the user's behavior Yt in the future, and the content of behavior Yt is the result of the LSTM layer.
[0024] By adopting the above technical solution, the time series characteristics in the behavior are captured according to the user's historical behavior, and the user's next behavior, such as the possibility of clicking on an advertisement or conversion, is predicted.
[0025] Optionally, the Bayesian network includes:
[0026] The set of multi-dimensional feature vectors is {X1, X2, ...Xn}, and the joint probability distribution is:
[0027] Wherein Pa(Xi) is the parent node of the multi-dimensional feature vector Xi;
[0028] Each of the multi-dimensional feature vectors Xi corresponds to a conditional probability table, namely:
[0029] P(Xi=xi|Pa(Xi)=pi);
[0030] When the new multi-dimensional feature vector is input, the conditional probability is updated according to the following formula:
[0031] P(H|E)=P(E|H)×P(H) / P(E), where H is a hypothesis parameter and E is the behavior data in the multi-dimensional feature vector.
[0032] By adopting the above technical solution, the causal relationship between different user behaviors is analyzed. Based on current and past data, the system can infer the possibility of users interacting with ads in the next period of time, such as whether to click on ads, what type of ads they will be interested in, etc.
[0033] Optionally, the type sorting includes the following steps:
[0034] Calculate a first behavior value corresponding to the type of the user behavior prediction data by a first table lookup matching method;
[0035] Calculating a second behavior value corresponding to the advertisement type by a second table lookup matching method;
[0036] Calculate the difference between the first behavior value and the second behavior value as a behavior type gap value;
[0037] Sort the behavior type gap values in ascending order, thereby completing type sorting;
[0038] The similarity algorithm comprises the following steps:
[0039] Calculate a first similarity value corresponding to the advertisement type in the pushed advertisement traffic by a third table lookup matching method;
[0040] Calculate the second similarity value corresponding to the push parameter by a fourth table lookup matching method;
[0041] A weighted difference between the first similarity value and the second similarity value is calculated as similarity data.
[0042] By adopting the above technical solutions, the system sorts the available advertising traffic according to the user behavior prediction results, and prioritizes the traffic that may generate high conversions to the DSP. Combined with the real-time bidding and budget of the DSP, the system will match the most suitable traffic and allocate traffic priority to the DSP, ensuring that the DSP obtains efficient traffic that matches its needs.
[0043] Optionally, the push algorithm includes a multi-objective optimization algorithm and a Pareto algorithm;
[0044] The objective function of the multi-objective optimization algorithm is: maxF(x) = [f1(x), f2(x), ..., fk(x)], where f1(x) is the advertising conversion rate, f2(x) is the efficiency of budget consumption, and the remaining objectives include response time and traffic utilization;
[0045] The Pareto algorithm is used to calculate a set of non-dominated solutions in a set of solutions of a multi-objective optimization algorithm, that is, no solution is better than another solution in terms of the objectives.
[0046] By adopting the above technical solutions, the system not only considers the conversion rate of advertisements, but also comprehensively considers multiple goals such as DSP budget, bidding strategy, user click-through rate, etc. to ensure maximum traffic allocation efficiency.
[0047] Optionally, the content of the allocation algorithm includes a K-Means clustering algorithm;
[0048] The objective function of the K-Means clustering algorithm is:
[0049] Minimize the within-cluster squared error: Among them, Ci is the i-th cluster, μi is the center of the cluster, and x is the data point in the cluster.
[0050] By adopting the above technical solution, when pushing traffic, the system uses a clustering algorithm to group different DSP requirements and user behaviors to better allocate resources. The commonly used K-Means clustering algorithm can be used to divide user behavior data and DSP requirements into several categories.
[0051] Optionally, the reinforcement learning algorithm is modeled based on a Markov decision process, including:
[0052] State St: the state at time t, including current user behavior data and DSP demand;
[0053] Action At: Push traffic to the specified DSP;
[0054] Reward Rt: click-through rate or conversion rate of the ad;
[0055] Q-learning algorithm: Estimate the long-term reward of each state-action pair through the Q-value function Q(S, A), where α is the learning rate and γ is the discount factor.
[0056] By adopting the above technical solutions, reinforcement learning is used to dynamically adjust the traffic push strategy. By learning the effect of pushing traffic, the system can gradually optimize the decision-making process. The reinforcement learning model is mainly based on the Markov decision process (MDP) for modeling.
[0057] Optionally, the step of obtaining the multi-dimensional behavior data of the user and preprocessing the multi-dimensional behavior data to obtain a multi-dimensional feature vector further includes the following sub-steps:
[0058] Acquire historical interactive behavior data of the user, wherein the historical interactive behavior has first multi-dimensional sub-data, and the first multi-dimensional sub-data includes advertisement click data, browsing time data, and page dwell data;
[0059] Acquire real-time action behavior data of the user, wherein the real-time action behavior has second multi-dimensional sub-data, and the second multi-dimensional sub-data includes current search record data and click behavior data;
[0060] Acquire context-related data of the user, wherein the context-related data has third multi-dimensional sub-data, and the third multi-dimensional sub-data includes time data, location data, and device type data associated with the user;
[0061] Processing the first multi-dimensional sub-data by a first preprocessing algorithm to obtain first standardized data;
[0062] Processing the second multi-dimensional sub-data by a second preprocessing algorithm to obtain second standardized data;
[0063] Processing the third multi-dimensional sub-data by a third preprocessing algorithm to obtain third standardized data;
[0064] The first standardized data, the second standardized data and the third standardized data are combined into the multi-dimensional feature vector.
[0065] By adopting the above technical solutions, the data will be preprocessed, noise will be removed, missing values will be filled, and standardized to ensure the cleanliness and uniformity of the data; the preprocessed data will form a multi-dimensional feature vector for subsequent analysis and prediction.
[0066] In the second aspect, the present application provides a traffic precision push system based on user behavior prediction, which adopts the following technical solutions:
[0067] A system for accurately pushing traffic based on user behavior prediction includes a processor, wherein the processor executes the steps of the method for accurately pushing traffic based on user behavior prediction as described in any one of the above.
[0068] In a third aspect, the present application provides a storage medium, which adopts the following technical solution:
[0069] A storage medium stores a program, which, when executed by a processor, implements the steps of any one of the above-mentioned methods for accurately pushing traffic based on user behavior prediction.
[0070] In summary, the present application includes at least one of the following beneficial technical effects: The user behavior prediction and accurate traffic push method of the present invention realizes accurate prediction of user behavior and efficient allocation of traffic by combining artificial intelligence technology, and significantly improves the overall effect of RTB advertising. The specific effects are reflected in the following aspects:
[0071] Significantly improved advertising conversion rate: Through the accurate prediction of user behavior by deep learning models and Bayesian networks, the present invention can better understand user interests and behavior patterns, and push highly matched advertising traffic to the demand-side platform (DSP). This precise traffic push strategy directly improves the click-through rate and conversion rate of advertisements;
[0072] Significantly improved traffic utilization: In traditional RTB systems, traffic distribution is often inaccurate or over-pushed, resulting in a large amount of traffic failing to be effectively converted. This invention reduces the sending of invalid traffic by predicting user behavior in real time, and pushes more high-quality traffic to DSPs in need, thereby improving the overall traffic utilization;
[0073] The response speed of advertising push is significantly accelerated: the reinforcement learning algorithm continuously optimizes traffic push decisions, achieving rapid response to DSP needs and changes in user behavior. The system can adjust the traffic allocation strategy in real time during the RTB bidding process, shortening the decision time from behavior prediction to traffic push, ensuring that advertisers can place ads within the best time window;
[0074] Reduce advertising costs: Through multi-objective optimization algorithms, the system can balance advertising conversion rates, budget consumption, and DSP needs to maximize traffic benefits. Without increasing the budget, DSP can achieve higher advertising exposure and conversion effects through more accurate traffic push, thereby reducing the unit advertising cost;
[0075] Achieve dynamic balance of traffic distribution: Through intelligent analysis of DSP demand and real-time bidding, dynamic adjustment and optimal allocation of traffic resources are achieved, avoiding waste or missed opportunities caused by excessive traffic allocation to some DSPs. Through the load balancing mechanism, the system can reasonably allocate resources and improve the advertising effect of all DSPs;
[0076] Continuous optimization of advertising delivery strategies: Through the reinforcement learning mechanism, the system can learn and improve based on the effect of each traffic push, and continuously optimize push decisions based on the constantly updated feedback data. This makes the advertising delivery strategy more and more accurate, and the traffic push effect continues to improve over time, forming an adaptive optimization closed loop;
[0077] Personalized advertising push has significant effects: By using clustering algorithms to segment user behavior patterns, personalized advertising push is achieved, which can display more attractive advertising content to different types of users. This personalized push greatly enhances the interaction between advertising and users and increases the attractiveness of advertising;
[0078] Improved system automation and reduced manual intervention: Through the introduction of artificial intelligence algorithms, the present invention reduces the need to manually adjust traffic push strategies in traditional RTB advertising systems. The system's automated push and feedback optimization mechanism makes the advertising delivery process more intelligent and reduces the complexity of manual operations. BRIEF DESCRIPTION OF THE DRAWINGS
[0079] Figure 1 This is a step-by-step diagram of the method for accurately pushing traffic based on user behavior prediction.
[0080] Figure 2 It is a step diagram of type sorting.
[0081] Figure 3 It is a step diagram of the similarity algorithm.
[0082] Figure 4 It is a sub-step diagram of obtaining multi-dimensional behavior data of users and preprocessing the multi-dimensional behavior data to obtain a multi-dimensional feature vector. DETAILED DESCRIPTION
[0083] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.
[0084] In the description of this specification, the description with reference to the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.
[0085] The embodiment of the present application discloses a method for accurately pushing traffic based on user behavior prediction.
[0086] Reference Figure 1 , a method for accurately pushing traffic based on user behavior prediction, comprising the following steps:
[0087] The multi-dimensional behavior data of the user is obtained, and the multi-dimensional behavior data is preprocessed to obtain a multi-dimensional feature vector.
[0088] The multi-dimensional feature vector is input into the deep learning model and Bayesian network to generate user behavior prediction data. The preprocessed data is input into the deep learning model (such as LSTM) and Bayesian network to predict the user's future behavior trends. Deep learning model (such as LSTM): Based on the user's historical behavior, it captures the time series characteristics in the behavior and predicts the user's next behavior, such as the possibility of clicking on an ad or converting. Bayesian network: used to analyze the causal relationship between different user behaviors. Based on current and past data, the system can infer the user's likelihood of interacting with ads in the next period of time, such as whether to click on ads, what type of ads will be of interest, etc. These prediction results will provide a basis for traffic push decisions to ensure that the pushed advertising traffic is highly matched with the user's interests.
[0089] The system allocates advertising traffic based on predicted user behavior and the needs of each DSP. During this process, the system will perform traffic sorting and traffic allocation operations based on the bidding strategy, budget and target audience of each DSP. Specifically:
[0090] Based on the user behavior prediction data, the types of advertising traffic available for sending are sorted. Among them, advertising traffic corresponds to advertising types, and the advertising types have corresponding type matching degrees with the types of user behavior prediction data; type sorting is performed according to the type matching degree, and the advertising traffic corresponding to the advertising types with the top set rankings in the type sorting results is pushed to the DSP. The system sorts the advertising traffic available for sending based on the user's behavior prediction results, and prioritizes the traffic that may generate high conversions to the DSP.
[0091] Based on the real-time bidding data and budget data of DSP, push parameters are calculated using push algorithm. The similarity between the ad type and push parameters in the pushed ad traffic is calculated as similarity data according to the similarity algorithm. Ad traffic is prioritized and delivered according to the allocation algorithm based on the similarity data. Delivery feedback data is generated after delivery. Combined with the real-time bidding and budget of DSP, the system will match the most suitable traffic and allocate traffic priority to DSP. Ensure that DSP obtains efficient traffic that matches its needs.
[0092] When traffic is pushed to DSP, the system will monitor its performance in real time, especially key indicators such as ad click-through rate and conversion rate.
[0093] Collect DSP's delivery feedback data, and obtain advertising feedback information after the advertising traffic is delivered from the delivery feedback data. Feedback collection: Collect delivery feedback data from each DSP and record the advertising performance after the traffic is sent. Dynamic adjustment: The system will use reinforcement learning (such as Q-learning algorithm) based on feedback information to adjust the next round of traffic allocation decisions. Specifically, the system will dynamically adjust the traffic push strategy according to the delivery effect of different traffic, and gradually optimize the traffic allocation.
[0094] The parameters in the similarity algorithm are adjusted using a reinforcement learning algorithm based on advertising feedback information.
[0095] The entire process forms a complete closed loop, from data collection, user behavior prediction, traffic matching, push, to feedback learning and optimization. The decision-making and execution of each step will be continuously optimized to ensure the accuracy and efficiency of traffic distribution. Through the continuous learning and improvement of AI algorithms, the system can achieve the following goals:
[0096] Provide more accurate advertising traffic for each DSP.
[0097] Improve ad click-through rate and conversion rate and reduce invalid traffic.
[0098] Dynamically adjust traffic push strategies to optimize advertising effectiveness.
[0099] This closed-loop optimization system can effectively improve the efficiency of RTB advertising and help advertisers and DSPs obtain higher returns amid fierce competition.
[0100] Among them, the deep learning model includes LSTM (Long Short-Term Memory Network), which is a deep learning model suitable for processing and predicting time series data, and is particularly good at capturing the temporal changes in user behavior. Through LSTM, the system can analyze users' browsing, clicking and interaction records, identify their behavior patterns and predict future advertising interactions.
[0101] Among them: the input layer includes the historical behavior data of the input user, represented as a time series {x1, x2...xt}, where each xt represents the behavior characteristics of the user at time t; behavior characteristics include browsing time, number of clicks, advertising display, etc.
[0102] The LSTM layer includes a forget gate, an input gate, a cell state update gate, a hidden layer state gate, and an output gate;
[0103] Forget gate: F(t)=σ(Wf×[H(t-1),Xt]+Bf), the forget gate determines how much past information the current state needs to "forget", Wf is the weight matrix, Bf is the bias, and σ is the Sigmoid activation function.
[0104] Input gate: It = σ(Wi × [H(t-1), Xt] + Bi); The input gate determines how the current input affects the state of the LSTM.
[0105] Cell state update gate: Ct = Ft×C(t-1)+It×tanh(Wc×[H(t-1),Xt]+Bc); The cell state update gate updates the cell state Ct, taking into account the combined effects of forgetting and new input.
[0106] Output gate: Ot = σ(Wo × [H(t-1), Xt] + Bo); the output gate determines the output state at the current moment.
[0107] Hidden layer state gate: Ht = Ot × tanh (Ct).
[0108] Output layer: predicts the user's behavior Yt at the future moment. The content of behavior Yt is the result of the LSTM layer. The output of the output layer is the result of the classification or regression task, expressed as traffic demand or click probability.
[0109] Based on the user's historical behavior, the time series characteristics in the behavior are captured to predict the user's next behavior, such as the possibility of clicking on an ad or converting.
[0110] Among them, the Bayesian network is a graphical model used to represent the conditional dependencies between random variables, which can effectively infer future behaviors in user behavior prediction.
[0111] The Bayesian network represents the dependency relationship between variables through a directed acyclic graph (DAG). Assuming that the set of multi-dimensional feature vectors is {X1, X2, ... Xn}, the joint probability distribution is:
[0112] Among them, Pa(Xi) is the parent node of the multi-dimensional feature vector Xi, that is, the other multi-dimensional feature vectors it depends on.
[0113] Each multi-dimensional feature vector Xi corresponds to a conditional probability table (CPT), namely:
[0114] P(Xi=xi|Pa(Xi)=pi); represents the probability distribution of each multi-dimensional feature vector under the condition of a given parent node. The Bayesian network calculates the conditional probability of user behavior through CPT and predicts whether the user will be interested in the advertisement under certain conditions.
[0115] When a new multi-dimensional feature vector is input, the conditional probability is updated according to the following formula:
[0116] P(H|E)=P(E|H)×P(H) / P(E), where H is the hypothesis parameter (such as the probability of a user clicking on an ad); E is the behavior data in the multi-dimensional feature vector, that is, the observed behavior data. Through Bayesian updating, the system can continuously correct its prediction of user behavior.
[0117] Used to analyze the causal relationship between different user behaviors. Based on current and past data, the system can infer the user's likelihood of interacting with ads in the next period of time, such as whether to click on ads, what type of ads will interest the user, etc.
[0118] Reference Figure 2 Type sorting is mainly based on the behavior type gap value between user behavior prediction data and advertisement types. Type sorting includes the following steps:
[0119] Calculate the first behavior value corresponding to the type of user behavior prediction data through a first table lookup matching method; input user behavior prediction data, such as user click, browse, purchase and other behavior data; match the type of user behavior prediction data through a table lookup, such as a predefined rule table or model output; output a first behavior value used to represent a quantitative value of the user behavior type.
[0120] The second behavior value corresponding to the advertisement type is calculated by the second table lookup matching method; the advertisement type is input, such as video advertisement, graphic advertisement, etc.; the advertisement type is matched by table lookup, such as a predefined rule table or model output; and the second behavior value used to represent the quantitative value of the advertisement type is output.
[0121] Calculate the difference between the first behavior value and the second behavior value as the behavior type gap value; input the first behavior value and the second behavior value; calculate the difference between the two behavior values; output the behavior type gap value; the behavior type gap value represents the degree of mismatch between the user behavior type and the advertisement type.
[0122] Sort the behavior type gap values in ascending order to complete the type sorting. Input the behavior type gap value; sort the behavior type gap values of all advertisements in ascending order; output the sorted advertisement list, which is sorted according to the matching degree between the user behavior type and the advertisement type.
[0123] In addition, the type gap value is adjusted by adaptive gain or loss according to the amount of flow, the steps are as follows:
[0124] Monitor system traffic in real time or regularly, including user visits, ad impressions, clicks, etc.
[0125] Based on system performance, business requirements, or historical data, set one or more traffic thresholds to determine whether the current traffic is high, medium, or low.
[0126] According to the traffic threshold, the adjustment strategy for the behavior type gap value at different traffic levels is determined.
[0127] During high traffic: We want to increase diversity and reduce recommendations that are overly dependent on user behavior. Therefore, we can deduct the behavior type gap value to reduce its impact.
[0128] Medium traffic: Maintain balance, no need to significantly adjust the behavior type gap value.
[0129] When the traffic volume is low: You want to improve the accuracy of the recommendation, so you can adjust the behavior type gap value by gain, that is, increase its influence.
[0130] According to the determined adjustment strategy, an adjustment factor (gain or loss) is applied to the behavior type gap value. This factor can be a fixed value, a percentage, or a function calculated dynamically according to the traffic flow.
[0131] The adjusted behavior type gap value is used in subsequent sorting, recommendation or analysis processes.
[0132] For example, suppose there is an ad recommendation system, the first behavior value is the user's click rate for a certain type of ad (between 0 and 1), and the second behavior value is the ad type score (also between 0 and 1). The calculated behavior type gap value may be the difference between the two (or some transformed difference).
[0133] During high traffic, you want to increase the diversity of ads, so you can apply a reduction factor, such as 0.8, to the behavior type gap value to reduce its impact.
[0134] At medium flow, leave it as is and do not apply any adjustment factor.
[0135] During low traffic, we want to improve the accuracy of the recommendations, so we can apply a gain factor, such as 1.2, to the behavior type gap value to increase its impact.
[0136] Reference Figure 3 , the similarity algorithm includes the following steps:
[0137] Calculate the first similarity value corresponding to the advertisement type in the pushed advertisement traffic by the third table lookup matching method; input the advertisement type in the pushed advertisement traffic; match the advertisement type by looking up a table, such as a correspondence table between advertisement type and similarity value; output the first similarity value; the first similarity value indicates the similarity between the advertisement type and the push target.
[0138] The second similarity value corresponding to the push parameter is calculated by the fourth table lookup matching method; the push parameter is input, such as the target user group, delivery time, delivery location, etc.; the push parameter is matched by table lookup, such as a correspondence table between push parameters and similarity values; the second similarity value is output, and the second similarity value represents the similarity between the push parameter and the advertisement content.
[0139] Calculate the weighted difference between the first similarity value and the second similarity value as similarity data. Input the first similarity value and the second similarity value; calculate the weighted difference between the two similarity values according to a predetermined weight; output similarity data, the similarity data representing the overall similarity between the advertisement and the push target. Similarity data = (first similarity value × first weight) - (second similarity value × second weight); or, similarity data = | (first similarity value × first weight) - (second similarity value × second weight) |.
[0140] The system sorts the available advertising traffic based on the user behavior prediction results, and prioritizes the traffic that may generate high conversions to the DSP. Combined with the DSP's real-time bidding and budget, the system will match the most suitable traffic and prioritize the traffic to the DSP, ensuring that the DSP obtains efficient traffic that matches its needs.
[0141] In the process of traffic push, the system must not only optimize the conversion rate, but also consider the constraints of DSP budget, bidding, etc. Multi-objective optimization algorithms are used to solve such multi-dimensional optimization problems.
[0142] In the embodiment of the present application, the content of the push algorithm includes a multi-objective optimization algorithm and a Pareto algorithm;
[0143] The objective function of the multi-objective optimization algorithm is: maxF(x) = [f1(x), f2(x), ..., fk(x)], where f1(x) is the advertising conversion rate, f2(x) is the efficiency of budget consumption, and the remaining objectives include response time and traffic utilization.
[0144] The Pareto algorithm is used to calculate a set of non-dominated solutions in a set of solutions of a multi-objective optimization algorithm, that is, no solution is better than another solution in terms of the objectives.
[0145] When pushing traffic, the system uses a clustering algorithm to group different DSP requirements and user behaviors to better allocate resources. The commonly used K-Means clustering algorithm can be used to divide user behavior data and DSP requirements into several categories.
[0146] In the embodiment of the present application, the content of the allocation algorithm includes a K-Means clustering algorithm;
[0147] The objective function of the K-Means clustering algorithm is:
[0148] Minimize the within-cluster squared error: Among them, Ci is the i-th cluster, μi is the center of the cluster, and x is the data point in the cluster. Through continuous iteration, the K-Means algorithm finds the optimal center of each cluster, thus completing clustering.
[0149] Through reinforcement learning and multi-objective optimization, the system can continuously optimize the push strategy based on the feedback from the DSP. This process involves learning the best traffic delivery plan from the feedback data, combining the Q learning and optimization algorithms mentioned above, dynamically adjusting the push decision, and gradually improving the delivery effect.
[0150] Reinforcement learning is used to dynamically adjust traffic push strategies. By learning the effect of pushing traffic, the system can gradually optimize the decision-making process. The reinforcement learning model is mainly based on the Markov decision process (MDP).
[0151] State St: The state of the system at time t, including current user behavior data and DSP demand.
[0152] Action At: Push traffic to the specified DSP; the push decision selected by the system, such as pushing a certain type of traffic to the specified DSP.
[0153] Reward Rt: the click-through rate or conversion rate of the ad; the reward brought by the action, usually measured by the click-through rate or conversion rate of the ad.
[0154] Q-learning algorithm: Estimate the long-term reward of each state-action pair through the Q-value function Q(S, A): Q Among them, α is the learning rate and γ is the discount factor, which indicates the importance of future returns. The system gradually learns the optimal push strategy by continuously trying different traffic push strategies and updating the Q value function.
[0155] Reinforcement learning is used to dynamically adjust traffic push strategies. By learning the effect of pushing traffic, the system can gradually optimize the decision-making process. The reinforcement learning model is mainly based on the Markov decision process (MDP).
[0156] Reference Figure 4 , the step of obtaining the multi-dimensional behavior data of the user and preprocessing the multi-dimensional behavior data to obtain a multi-dimensional feature vector also includes the following sub-steps:
[0157] The user's historical interactive behavior data is obtained, where the historical interactive behavior has first multi-dimensional sub-data, and the first multi-dimensional sub-data includes advertisement click data, browsing time data, and page dwell data.
[0158] The real-time action behavior data of the user is obtained, and the real-time action behavior has second multi-dimensional sub-data, and the second multi-dimensional sub-data includes current search record data and click behavior data.
[0159] The context-related data of the user is obtained, the context-related data having third multi-dimensional sub-data, and the third multi-dimensional sub-data including time data, location data and device type data associated with the user.
[0160] The first multi-dimensional sub-data is processed by a first preprocessing algorithm to obtain first standardized data; the first preprocessing algorithm includes a noise elimination algorithm and a standard data format conversion and unification.
[0161] The second multi-dimensional sub-data is processed by a second preprocessing algorithm to obtain second standardized data; the second preprocessing algorithm includes filling missing values + standard data format conversion and unification.
[0162] The third multi-dimensional sub-data is processed by a third preprocessing algorithm to obtain third standardized data; the third preprocessing algorithm includes noise removal algorithm + missing value filling + standard data format conversion and unification.
[0163] The first standardized data, the second standardized data and the third standardized data are combined into a multi-dimensional feature vector. The combining step is stacking of standard data formats, similar to matrix stacking; the steps are as follows:
[0164] Step 1: Confirm data dimensions
[0165] Check the dimensions of each normalized dataset:
[0166] First, standardize the data (historical interactive behavior data): confirm the number of features it contains, such as the number of ad clicks, average browsing time, page dwell time, etc.
[0167] Second standardized data (real-time action behavior data): confirm the number of features it contains, such as the number of current search keywords, number of clicks, etc.
[0168] Third, standardized data (contextual data): confirm the number of features it contains, such as timestamp (such as converted into a time period of the day), geographic location (such as converted into longitude and latitude or area code), device type (such as mobile phone, tablet, desktop computer), etc.
[0169] Step 2: Data alignment
[0170] Make sure the timestamps or user IDs of each dataset are aligned:
[0171] If the data is organized by timestamp or user ID, make sure they are aligned before merging. That is, the records in each dataset correspond to the same user or the same point in time.
[0172] Step 3: Merge data
[0173] Combine datasets by rows or columns:
[0174] Merge by rows: If each dataset represents user behavior at a different point in time, you can stack them up in chronological order to form a long-format dataset.
[0175] Merge by columns: If each dataset represents different dimensional features of the user, they can be concatenated by columns to form a wide-format dataset, i.e., a multi-dimensional feature vector.
[0176] Handling duplicate features:
[0177] If you find duplicate features during the merge process (for example, the feature “browsing time” exists in both datasets), you need to decide whether to keep one or perform some form of aggregation (such as sum, average, etc.).
[0178] Fill missing values:
[0179] If missing values still exist in the merged data set, choose to fill in default values, use interpolation methods, or delete records containing missing values, depending on the specific situation.
[0180] Step 4: Verification and cleaning
[0181] Verify data integrity:
[0182] Check that the merged dataset contains all expected features and that no important information is lost.
[0183] Cleaning data:
[0184] If necessary, further clean the data, such as removing outliers and processing data type mismatches.
[0185] Example:
[0186] Assume there are three standardized datasets:
[0187] First standardized data (historical interactive behavior data): [user ID, number of ad clicks, average browsing time, page dwell time];
[0188] Second standardized data (real-time action data): [user ID, number of current search keywords, number of clicks];
[0189] Third standardized data (contextual data): [user ID, time period, geographic location, device type];
[0190] The merged multi-dimensional feature vector is as follows:
[0191] [User ID, number of ad clicks, average browsing time, length of time on page, number of current search keywords, number of clicks, time period, geographic location, device type].
[0192] The merging here is based on user ID, and it is assumed that each user ID is unique in each dataset.
[0193] These data will be preprocessed to remove noise, fill missing values, and standardize to ensure the cleanliness and uniformity of the data; the preprocessed data will form a multi-dimensional feature vector for subsequent analysis and prediction.
[0194] The core of the entire system to achieve accurate push and efficient traffic distribution lies in:
[0195] 1. Real-time prediction and dynamic adjustment of user behavior
[0196] Technological innovation: By combining deep learning models (such as LSTM) with Bayesian networks, the system can conduct in-depth analysis of users' historical behaviors and predict their future behaviors based on the latest data. This combination enables highly accurate predictions of user behaviors, ensuring that the DSP is pushed the advertising traffic that best matches the user's current needs during the RTB bidding process.
[0197] Key point: Compared with traditional fixed rules or push solutions based on historical data, the present invention dynamically adjusts traffic push strategies through real-time behavior prediction, greatly improving the click-through rate and conversion rate of advertisements.
[0198] 2. Traffic Accurate Matching Mechanism
[0199] Technological innovation: The system uses AI technology to model user behavior and pushes traffic that best meets the needs of each DSP based on the needs of the DSP (budget, target user group, bidding strategy, etc.). In this process, the system will provide DSP with higher-quality traffic based on the user's current interaction and conversion tendency, reducing unnecessary waste of resources.
[0200] Key point: Through dual analysis of user behavior and DSP needs, the system achieves accurate matching of traffic, improves traffic utilization, and avoids the sending of invalid traffic.
[0201] 3. Traffic distribution optimization driven by reinforcement learning
[0202] Technological innovation: Through reinforcement learning, the system can continuously learn and optimize in the process of pushing traffic. The reinforcement learning model records the performance of each push strategy, such as click-through rate, conversion rate, etc., and updates the push decision based on feedback data. Through continuous self-learning and improvement, the system has achieved dynamic adjustment in push strategy and gradually improved the efficiency of traffic distribution.
[0203] Key point: Unlike the fixed or manually adjusted traffic push strategies in existing systems, the reinforcement learning algorithm enables the system to have adaptive capabilities, adjust traffic push according to actual conditions, reduce manual intervention and improve push effects.
[0204] 4. Multi-objective optimization algorithm
[0205] Technological innovation: This invention uses a multi-objective optimization algorithm, which not only considers the conversion rate of advertisements during traffic push, but also integrates multiple factors such as DSP budget, bidding, response speed, etc. The system can find the optimal solution to balance multiple objectives and ensure the maximum utilization of advertising traffic.
[0206] Key point: Through multi-objective optimization, the system can find the most suitable push solution for each DSP in a complex bidding environment, thereby improving the effectiveness and efficiency of overall advertising delivery.
[0207] 5. Real-time feedback and adaptive optimization of traffic transmission
[0208] Technological innovation: The system can monitor and provide feedback on the traffic effect in real time, and obtain information such as DSP bidding results, ad click-through rate and conversion rate in a timely manner. Based on these data, the system will automatically adjust the subsequent traffic push strategy to ensure that the traffic pushed each time is more accurate. The feedback mechanism is combined with reinforcement learning to form a closed loop to improve traffic utilization efficiency.
[0209] Key point: Unlike traditional systems that can only be optimized over a long period of time, the present invention relies on real-time feedback mechanisms and reinforcement learning models to achieve rapid response and optimization, greatly improving the flexibility and accuracy of push notifications.
[0210] 6. Dynamic DSP demand analysis and load balancing
[0211] Technological innovation: The system can analyze the needs of each DSP in real time and dynamically adjust traffic allocation based on its needs and advertising performance. This mechanism ensures balanced allocation of resources, preventing some DSPs from obtaining a large amount of invalid traffic while other DSPs miss out on delivery opportunities due to insufficient traffic.
[0212] Key point: This dynamic adjustment mechanism not only improves the efficiency of traffic usage, but also helps DSP better achieve its advertising goals and optimizes resource allocation issues in the entire RTB ecosystem.
[0213] 7. Refined traffic push based on user behavior
[0214] Technological innovation: This invention achieves in-depth mining of user needs by finely processing user behavior data. For example, clustering algorithms are used to classify different user behavior patterns, helping the system to achieve "different for each person" when pushing traffic and realizing personalized advertising push strategies.
[0215] Key point: This personalized push strategy significantly improves the click-through rate and conversion rate of ads, enables ads to accurately reach the target audience, and reduces traffic waste.
[0216] In summary, the present invention significantly improves the advertising conversion rate and traffic utilization rate, reduces the delivery cost, and speeds up the response speed through real-time prediction and accurate push of user behavior. Through the continuous optimization of AI technology, the system not only realizes the automation and precision of traffic push, but also improves the delivery efficiency of advertisers and DSPs.
[0217] An embodiment of the present application also discloses a system for accurately pushing traffic based on user behavior prediction, including a processor, wherein the processor executes the steps of the method for accurately pushing traffic based on user behavior prediction as described in any one of the above.
[0218] An embodiment of the present application further discloses a storage medium, in which a program is stored. When the program is executed by a processor, the steps of any one of the above-mentioned methods for accurately pushing traffic based on user behavior prediction are implemented.
[0219] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.
Claims
1. A method for accurately pushing traffic based on user behavior prediction, characterized in that: The steps include: Acquire multi-dimensional behavior data of the user, and pre-process the multi-dimensional behavior data to obtain a multi-dimensional feature vector; Inputting the multi-dimensional feature vector into a deep learning model and a Bayesian network to generate user behavior prediction data; Based on the user behavior prediction data, the advertisement traffic available for transmission is sorted by type, wherein the advertisement traffic corresponds to an advertisement type, and the advertisement type has a corresponding type matching degree with the type of the user behavior prediction data, and the type is sorted according to the type matching degree, and the advertisement traffic corresponding to the advertisement type with a preset ranking in the result of the type sorting is pushed to the DSP; Calculate push parameters using a push algorithm based on the real-time bidding data and budget data of the DSP, calculate the similarity between the advertisement type in the pushed advertisement traffic and the push parameters according to a similarity algorithm as similarity data, and prioritize and deliver the advertisement traffic according to an allocation algorithm based on the similarity data; generate delivery feedback data after delivery; Collecting the delivery feedback data of the DSP, and obtaining advertisement feedback information after the advertisement traffic is delivered from the delivery feedback data; A reinforcement learning algorithm is used to adjust parameters in the similarity algorithm based on the advertisement feedback information.
2. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The deep learning model includes LSTM, where: The input layer includes the historical behavior data of the input user, represented as a time series {x1, x2…xt}, where each xt represents the behavior characteristics of the user at time t; The LSTM layer includes a forget gate, an input gate, a cell state update gate, a hidden layer state gate, and an output gate; Forget gate: F(t) = σ(Wf×[H(t-1), Xt]+Bf), Wf is the weight matrix, Bf is the bias, and σ is the Sigmoid activation function; Input gate: It = σ(Wi × [H(t-1), Xt] + Bi); Cell state update gate: Ct = Ft × C(t-1) + It × tanh(Wc × [H(t-1), Xt] + Bc); Output gate: Ot = σ(Wo × [H(t-1), Xt] + Bo); Hidden layer state gate: Ht = Ot × tanh (Ct); The output layer includes the prediction of the user's behavior Yt in the future, and the content of behavior Yt is the result of the LSTM layer.
3. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The Bayesian network includes: The set of multi-dimensional feature vectors is {X1, X2, ...Xn}, and the joint probability distribution is: Wherein Pa(Xi) is the parent node of the multi-dimensional feature vector Xi; Each of the multi-dimensional feature vectors Xi corresponds to a conditional probability table, namely: P(Xi=xi|Pa(Xi)=pi); When the new multi-dimensional feature vector is input, the conditional probability is updated according to the following formula: P(H|E)=P(E|H)×P(H) / P(E), where H is a hypothesis parameter and E is the behavior data in the multi-dimensional feature vector.
4. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The content of the type sorting includes the following steps: Calculate a first behavior value corresponding to the type of the user behavior prediction data by a first table lookup matching method; Calculating a second behavior value corresponding to the advertisement type by a second table lookup matching method; Calculate the difference between the first behavior value and the second behavior value as a behavior type gap value; Sort the behavior type gap values in ascending order, thereby completing type sorting; The similarity algorithm comprises the following steps: Calculate a first similarity value corresponding to the advertisement type in the pushed advertisement traffic by a third table lookup matching method; Calculate the second similarity value corresponding to the push parameter by a fourth table lookup matching method; A weighted difference between the first similarity value and the second similarity value is calculated as similarity data.
5. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The push algorithm includes a multi-objective optimization algorithm and a Pareto algorithm; The objective function of the multi-objective optimization algorithm is: maxF(x) = [f1(x), f2(x), ..., fk(x)], where f1(x) is the advertising conversion rate, f2(x) is the efficiency of budget consumption, and the remaining objectives include response time and traffic utilization; The Pareto algorithm is used to calculate a set of non-dominated solutions in a set of solutions of a multi-objective optimization algorithm, that is, no solution is better than another solution in terms of the objectives.
6. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The content of the allocation algorithm includes K-Means clustering algorithm; The objective function of the K-Means clustering algorithm is: Minimize the within-cluster squared error: Among them, Ci is the i-th cluster, μi is the center of the cluster, and x is the data point in the cluster.
7. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The reinforcement learning algorithm is modeled based on the Markov decision process, including: State St: the state at time t, including current user behavior data and DSP demand; Action At: Push traffic to the specified DSP; Reward Rt: click-through rate or conversion rate of the ad; Q-learning algorithm: Estimate the long-term reward of each state-action pair through the Q-value function Q(S, A), where α is the learning rate and γ is the discount factor.
8. The method for accurately pushing traffic based on user behavior prediction according to claim 1 is characterized in that: The step of obtaining the multi-dimensional behavior data of the user and pre-processing the multi-dimensional behavior data to obtain a multi-dimensional feature vector further includes the following sub-steps: Acquire historical interactive behavior data of the user, wherein the historical interactive behavior has first multi-dimensional sub-data, and the first multi-dimensional sub-data includes advertisement click data, browsing time data, and page dwell data; Acquire real-time action behavior data of the user, wherein the real-time action behavior has second multi-dimensional sub-data, and the second multi-dimensional sub-data includes current search record data and click behavior data; Acquire context-related data of the user, wherein the context-related data has third multi-dimensional sub-data, and the third multi-dimensional sub-data includes time data, location data, and device type data associated with the user; Processing the first multi-dimensional sub-data by a first preprocessing algorithm to obtain first standardized data; Processing the second multi-dimensional sub-data by a second preprocessing algorithm to obtain second standardized data; Processing the third multi-dimensional sub-data by a third preprocessing algorithm to obtain third standardized data; The first standardized data, the second standardized data and the third standardized data are combined into the multi-dimensional feature vector.
9. A traffic precision push system based on user behavior prediction, characterized in that: It includes a processor, in which the steps of the method for accurately pushing traffic based on user behavior prediction as described in any one of claims 1-8 are executed.
10. A storage medium, characterized in that: The medium stores a program, and when the program is executed by the processor, the steps of the method for accurately pushing traffic based on user behavior prediction described in any one of claims 1 to 8 are implemented.