Information pushing method and device, and non-transitory computer-readable storage medium

Abstract

The present disclosure relates to an information pushing method and device, and a non-transitory computer readable storage medium. The information pushing method comprises: obtaining historical display information displayed by a user terminal in a predetermined time period; determining, for each user terminal, a historical similarity between each candidate information and the historical display information of the user terminal; determining, according to the historical similarity, an efficiency value of the candidate information for the user terminal; determining a click probability of each candidate information; for each candidate information, determining a quality value of the candidate information according to the click probability of the candidate information and the efficiency value of the candidate information for the user terminal; and pushing the candidate information with the highest quality value to the user terminal for display. The present disclosure can not only display information with a higher click rate to the user, but also avoid frequently displaying similar information to the same user.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of information pushing, and in particular, to an information pushing method and device, and a non-transitory computer-readable storage medium. BACKGROUND

[0002] There are many methods for recommending information to users in the related art, which mainly include:

[0003] First, a personalized information pushing method: an information host (information provider) recommends information meeting the interests of a user to the user according to the information of the user, for example, giving cosmetic information to young women and giving automobile oil information to a user frequently browsing automobile content.

[0004] Second, a greedy method: the information host counts the click rates of all information ideas of the information host and recommends those information with the highest click rates to users; for example, the information host finds that the information of Xiaomi mobile phones always has the highest click rate, and thus the information host gives information of Xiaomi mobile phones to all users.

[0005] Third, an exploit-explore method: similar to the second method, the exploit-explore method tends to recommend the currently optimal (i.e., the information with the highest click rate) information to users, but at the same time, explores those non-optimal information and recommends them to users for trial at a certain probability. There are many choices for the probability distribution used for exploration, for example, UCB (Upper Confidence Bound) algorithm and TS (Thompson Sampling) algorithm. SUMMARY

[0006] The inventors have found through research that although the first method has the best effect and the most related research, it also has the highest requirement for data. It requires an information platform to send user information to an information host, and the information host judges which information is more in line with the interests of the user based on the user information. The detailed information of the user involves user privacy and commercial value, and many information platforms are unwilling to pass the information to the information host and only willing to pass the user identifier (such as the user equipment number) to the information host. If the information host only has the user identifier, it is obviously insufficient to make personalized information recommendations for the user.

[0007] The second method, although not as effective as the first method, does not depend on user information and is the simplest to implement. The main problem of the second method is that it is easily misled by accidental and temporary performance. For example, if a certain information S accidentally performs well, it may be recommended to the user all the time, and other information is always considered to be less effective because it does not get enough display opportunities; in fact, S may not be the most effective information.

[0008] The third method overcomes the above-mentioned shortcomings of the first and second methods, and does not require detailed user information and is not misled by accidental and temporary effects. However, it does not make full use of the important information of "user identification" provided by the information platform. Since it still tends to recommend the best information, it can repeatedly show the same information (or similar information) to the same user. The role of information is to recommend information and attract attention, and repeatedly showing the same content will only make the user feel bored. For example, the third method shows the user U five times of information about product A today; the user U has certainly paid sufficient attention to the information, and the information platform cannot attract his click by showing him more of the same information. Instead of wasting information costs on such repeated display, it is better to try other information.

[0009] In view of at least one of the above technical problems, the present disclosure provides an information pushing method and device and a non-transitory computer-readable storage medium, which can play information with higher click rates to users and avoid frequently showing similar information to the same user.

[0010] According to an aspect of the present disclosure, an information pushing method is provided, comprising:

[0011] obtaining historical display information displayed by a user terminal in a recent predetermined time period;

[0012] determining, for each user terminal, a historical similarity between each candidate information and the historical display information of the user terminal;

[0013] determining an efficiency value of the candidate information for the user terminal according to the historical similarity;

[0014] determining a click probability of each candidate information;

[0015] determining a quality value of each candidate information according to the click probability of the candidate information and the efficiency value of the candidate information for the user terminal;

[0016] pushing the candidate information with the highest quality value to the user terminal for display.

[0017] In some embodiments of the present disclosure, the information pushing method further comprises:

[0018] pre-obtaining a text vector and a non-text vector of each information, wherein the non-text vector comprises at least one of a picture vector or a video vector.

[0019] In some embodiments of the present disclosure, the determination of the historical similarity between each candidate information and the historical display information of the user terminal comprises:

[0020] The text similarity is determined based on the text vector of each candidate information and the text vector of the historical displayed information.

[0021] Non-text similarity is determined based on the non-text vector of each candidate information and the non-text vector of the historical displayed information;

[0022] The harmonic mean of text similarity and non-text similarity is used as the historical similarity.

[0023] In some embodiments of this disclosure, determining the efficiency value of each candidate piece of information for the user terminal based on its historical similarity to historically displayed information includes:

[0024] Based on the quadratic function of the similarity, the efficiency value of the candidate information for the user terminal is determined, wherein the constant term of the quadratic function is 0, the coefficient of the quadratic term is less than 0, and the coefficient of the linear term is greater than 0.

[0025] In some embodiments of this disclosure, determining the quality value of the candidate information based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal includes:

[0026] The quality value of the candidate information is determined by summing the product of the click probability of the candidate information and the fusion parameter with the efficiency value of the candidate information for the user terminal.

[0027] In some embodiments of this disclosure, the information push method further includes:

[0028] The quadratic coefficient, linear coefficient, and fusion parameter are updated by training the regression model.

[0029] In some embodiments of this disclosure, updating the quadratic coefficients, linear coefficients, and fusion parameters by training a regression model includes:

[0030] The display status of the candidate information with the highest quality value after it is pushed to the user's terminal is monitored and recorded in the display history database;

[0031] Within a predetermined time interval after the candidate information with the highest quality value is pushed to the user terminal for display, the user's clicks on the candidate information with the highest quality value are monitored, and the quadratic term coefficient, linear term coefficient, and fusion parameter are updated by training a regression model.

[0032] In some embodiments of this disclosure, monitoring user clicks on candidate information with the highest quality value and updating the quadratic term coefficients, linear term coefficients, and fusion parameters by training a regression model includes:

[0033] If, within a predetermined time interval, the user is detected to have clicked on the candidate information with the highest quality score, the quadratic term coefficient, linear term coefficient, and fusion parameter are updated by training the regression model.

[0034] If, within a predetermined time interval, the user does not click on the candidate information with the highest quality value, then the step of updating the quadratic term coefficient, linear term coefficient, and fusion parameter by training the regression model is executed.

[0035] In some embodiments of this disclosure, the information push method further includes:

[0036] Determine whether this display is the first time the user terminal has displayed the device within the most recently scheduled time period;

[0037] If this display is the first display on the user terminal within the most recent scheduled time period, then update the number of observations of clicks after the display and the number of observations of no clicks after the display;

[0038] If this display is not the first display by the user terminal in the most recent scheduled time period, then update the number of observations that clicked after the display and the number of observations that did not click after the display, and execute the steps of updating the quadratic term coefficients, linear term coefficients and fusion parameters by training the regression model.

[0039] In some embodiments of this disclosure, updating the quadratic coefficients, linear coefficients, and fusion parameters by training a regression model includes:

[0040] To obtain all display history of a piece of information, the historical similarity calculated at each display and the click identifier after each display are used as training samples. The gradient descent method is used to solve for the quadratic term coefficient, linear term coefficient and fusion parameter when the loss function is minimized.

[0041] The original quadratic coefficients, linear coefficients, and fusion parameters are updated using the quadratic coefficients, linear coefficients, and fusion parameters when the loss function is minimized.

[0042] According to another aspect of this disclosure, an information push device is provided, comprising:

[0043] The historical display information acquisition module is used to acquire the historical display information displayed by the user terminal in the most recent scheduled time period;

[0044] The historical similarity determination module is used to determine the historical similarity between each candidate piece of information and the historical information displayed by the user terminal for each user terminal.

[0045] An efficiency value determination module is used to determine the efficiency value of the candidate information for the user terminal based on the historical similarity.

[0046] The click probability determination module is used to determine the click probability of each candidate message.

[0047] The quality value determination module is used to determine the quality value of each candidate information based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal.

[0048] The information push module is used to push the candidate information with the highest quality value to the user terminal for display.

[0049] In some embodiments of this disclosure, the information push device is used to perform operations that implement the information push method as described in any of the above embodiments.

[0050] According to another aspect of this disclosure, an information push device is provided, comprising:

[0051] Memory, used to store instructions;

[0052] A processor is configured to execute the instructions, causing the information push device to perform operations implementing the information push method as described in any of the above embodiments.

[0053] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions that, when executed by a processor, implement the information push method as described in any of the above embodiments.

[0054] This disclosure can both show users information with higher click-through rates and avoid frequently showing similar information to the same user. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 These are schematic diagrams illustrating some embodiments of the information push method disclosed herein.

[0057] Figure 2 The diagram illustrates some other embodiments of the information push method disclosed herein.

[0058] Figure 3 This is a schematic diagram illustrating information vectorization in some embodiments of this disclosure.

[0059] Figure 4 This is a schematic diagram illustrating how efficiency values ​​change with repeated display of information in some embodiments of this disclosure.

[0060] Figure 5 These are schematic diagrams illustrating some embodiments of the information push device disclosed herein.

[0061] Figure 6 The diagram shows the structure of some other embodiments of the information push device disclosed herein. Detailed Implementation

[0062] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0063] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0064] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0065] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0066] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0067] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0068] The inventors discovered through research that related technologies, such as RTB (Real Time Bidding) and RTA (Real Time API (Application Programming Interface), are rapidly developing in the real-time information push sector. In the RTB scenario, information brokers send the information slots and real-time user information they collect to the DSP (Demand Side Platform, which can be considered as multiple information owners). In the RTA scenario, the information platform sends real-time user information to multiple information owners. Ultimately, each information owner decides whether to participate in the information bidding, generates a bid, and specifies the information content they want to distribute, which they then send to the information broker / information platform.

[0069] Each bidding process for real-time information can include: a user visits a website or app, and the app or website page contains an information slot, thus triggering an information request; the information broker (in RTB scenarios) or information platform (in RTA scenarios) sends the user's identifier to eligible information owners, where the user identifier is usually the user's device number, such as the mobile phone's IDFA (Identifier For Advertising) / IMEI (International Mobile Equipment Identity). Typically, more than one information owner participates in the bidding for the same information slot; based on the user identifier and other possible contextual information, the information owner determines whether to bid. If they decide to participate, they will submit an intended bid and information content, replying to the information broker / platform; the information broker / platform collects all intended bids from information owners and finds the information owner with the highest intended bid (sometimes, other factors besides bidding, such as information content quality, are considered). Then, they reply to all information owners with a message indicating whether their bid was successful; finally, the information broker / platform displays the winning information content to the user.

[0070] The information content described in the above process generally consists of text and images.

[0071] The inventors discovered through research that user appeal is crucial for information, especially for pay-per-impression information. If a bid is successful and the information is displayed to a user, but the user doesn't click or even retain any impression, it means the information owner's payment for that impression is wasted. Generally, CTR (Click Through Rate) is used to measure information quality. Information that is rich in text and images, novel and unique, and has a clear theme is obviously more likely to attract user clicks, resulting in a higher CTR. After clicking, users are directed to the information owner's page or app, and if possible, will subsequently engage in conversion behavior (user registration, order, or other high-value behaviors defined by the information owner).

[0072] It's important to note that repeatedly displaying information to the same user isn't entirely wasteful. Brand messaging, such as a brand's television advertising, aims to reinforce brand awareness through constant repetition. Therefore, brand messaging is typically found in traditional media (as opposed to internet messaging) or as packaged online advertising. The payment is usually a fixed price for the placement and duration of the ad, and we rarely evaluate its effectiveness using precise metrics like click-through rates.

[0073] The real-time information studied in this disclosure is generally performance-based information. In performance-based information systems, the information provider only needs to pay for measurable results, such as charging based on the number of times the information is displayed (the billing method for most real-time information). Performance-based information uses relatively precise revenue and cost metrics for evaluation. The real-time information discussed in this invention can be measured by the number of information clicks (click-through rate) and payment for the information. Admittedly, in real-time information represented by RTB and RTA, repeatedly showing the same information to the same user can strengthen the user's impression, but from the original intention of performance-based information, this is negligible. This is because performance-based information itself does not focus on brand information. The information creative does not emphasize the information provider's brand awareness, but rather hopes that users will notice the information, click on the information, and purchase goods on the shopping platform's app.

[0074] The real-time information represented by RTB and RTA, which are the focus of this disclosure, is generally billed based on the number of impressions. This means that information owners pay a certain cost every time they display information to users (for example, for a certain piece of information on JD.com, each display costs an average of 0.03 yuan). Therefore, information owners always hope that each information display can attract users' attention and clicks.

[0075] In the real-time information industry, information providers often struggle to obtain sufficient user information for personalized recommendations, relying solely on device IDs to differentiate users. On one hand, information providers need to show users information with higher click-through rates to achieve more clicks with the same number of impressions; on the other hand, they also need to avoid showing the same (or similar) information to the same user too frequently to prevent user boredom and wasted impressions.

[0076] This disclosure introduces the concept of the efficiency value of information for user terminals (i.e., the concept of the "freshness" of information for users), and proposes a system design for calculating the efficiency value and applying it to information display. Specifically, this disclosure uses Thompson sampling to achieve exploit-explore, displaying users with high click-through rates as much as possible, while avoiding repeated display by reducing the efficiency value of the information, thereby solving the two problems mentioned above.

[0077] The embodiments of this disclosure will be described below through specific examples.

[0078] Figure 1 This is a schematic diagram of some embodiments of the information push method of this disclosure. Preferably, this embodiment can be executed by the information push device of this disclosure. The method includes at least one of the following steps, wherein:

[0079] Step 11: Obtain the historical display information of the user terminal in the most recent scheduled time period.

[0080] In some embodiments of this disclosure, the information disclosed may be an advertisement.

[0081] Step 12: For each user terminal, determine the historical similarity between each candidate information and the historical information displayed by that user terminal.

[0082] In some embodiments of this disclosure, the information push method may further include: pre-acquiring text vectors and non-text vectors for each piece of information, wherein the non-text vectors include at least one of image vectors or video vectors.

[0083] In some embodiments of this disclosure, step 12 may include at least one of steps 121 and 123, wherein:

[0084] Step 121: Determine the text similarity based on the text vector of each candidate information and the text vector of the historical displayed information.

[0085] Step 122: Determine the non-text similarity based on the non-text vector of each candidate information and the non-text vector of the historical display information.

[0086] Step 123: The harmonic mean of text similarity and non-text similarity is taken as the historical similarity.

[0087] Step 13: Determine the efficiency value of the candidate information for the user terminal based on the historical similarity.

[0088] In some embodiments of this disclosure, step 13 may include: determining the efficiency value of the candidate information for the user terminal based on a quadratic function of the similarity, wherein the constant term of the quadratic function is 0, the coefficient of the quadratic term is less than 0, and the coefficient of the linear term is greater than 0.

[0089] Step 14: Determine the click probability of each candidate message.

[0090] Step 15: For each candidate information, determine the quality value of the candidate information based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal.

[0091] In some embodiments of this disclosure, step 15 may include: for each candidate information, determining the quality value of the candidate information based on the sum of the product of the click probability of the candidate information and the fusion parameter and the efficiency value of the candidate information for the user terminal.

[0092] Step 16: Push the candidate information with the highest quality value to the user terminal for display.

[0093] In some embodiments of this disclosure, the information push method may further include: updating the quadratic term coefficients, linear term coefficients, and fusion parameters by training a regression model.

[0094] In some embodiments of this disclosure, the step of updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model may include: monitoring the display status of the candidate information with the highest quality value after it is pushed to the user terminal and recording it in the display history database; monitoring the user's clicks on the candidate information with the highest quality value within a predetermined time interval after the candidate information with the highest quality value is pushed to the user terminal for display, and updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model.

[0095] In some embodiments of this disclosure, the step of monitoring user clicks on the candidate information with the highest quality value and updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model may include: if the user clicks on the candidate information with the highest quality value within a predetermined time interval, then updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model; if the user does not click on the candidate information with the highest quality value within a predetermined time interval, then performing the step of updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model.

[0096] In some embodiments of this disclosure, the information push method may further include: determining whether the current display is the first display of the user terminal within the most recent predetermined time period; if the current display is the first display of the user terminal within the most recent predetermined time period, then updating the number of observations of clicks after the display and the number of observations of no clicks after the display; if the current display is not the first display of the user terminal within the most recent predetermined time period, then updating the number of observations of clicks after the display and the number of observations of no clicks after the display, and performing the step of updating the quadratic term coefficient, the linear term coefficient, and the fusion parameter by training the regression model.

[0097] In some embodiments of this disclosure, the step of updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model may include: obtaining all display history of a piece of information; using the historical similarity calculated at each display and the click identifier after each display as training samples; using gradient descent to solve for the quadratic coefficient, linear coefficient, and fusion parameter when the loss function is minimized; and updating the original quadratic coefficient, linear coefficient, and fusion parameter using the quadratic coefficient, linear coefficient, and fusion parameter when the loss function is minimized.

[0098] Based on the information push method provided in the above embodiments of this disclosure, an innovative information efficiency value (freshness) model for user terminals is proposed and combined with a relatively mature click probability model in the industry (such as the Thompson sampling method). The above embodiments of this disclosure can not only show users information with higher click rates, but also avoid frequently showing similar information to the same user.

[0099] Figure 2 The diagram illustrates some other embodiments of the information push method disclosed herein. Preferably, this embodiment can be executed by the information push device of this disclosure.

[0100] like Figure 2 As shown, the architecture of this information push method mainly consists of two parts: offline information vectorization and online information service. The offline part vectorizes all information and saves these vectors in an index; while the online part is the main body of this disclosure, responsible for responding to requests, returning appropriate information to the information platform, and updating the online model.

[0101] like Figure 2 As shown, the information push method disclosed herein may include at least one of steps 100 and 200, wherein:

[0102] Step 100, Offline Part: Vectorize the information.

[0103] Figure 3 This is a schematic diagram illustrating information vectorization in some embodiments of this disclosure. For example... Figure 2 and Figure 3As shown. The information vectorization process in step 100 may include at least one of steps 101-103, wherein:

[0104] Step 101: Break down all the information and creative ideas into two parts: images and creative ideas.

[0105] In some embodiments of this disclosure, step 101 may include: splitting each piece of information into text information and non-text information, wherein the non-text information may be an image or a video.

[0106] Step 102: Vectorize the text information to generate text vectors.

[0107] In some embodiments of this disclosure, step 102 may include: text vectorization: that is, converting a sentence of information text into a d-dimensional vector. There are many mature models that can do this, including Word2vec (word to vector, a related model used to generate word vectors), GloVe (Global Vectors), ElMo (Embedding from Language Models), BERT (Bidirectional Encoder Representations from Transformers), etc. Some embodiments of this disclosure may use the word2vec model, and other models may also be used in some embodiments of this disclosure.

[0108] Step 103: Vectorize the non-text information to generate non-text vectors, wherein the non-text vectors include at least one of image vectors or video vectors.

[0109] In some embodiments of this disclosure, step 103 may include: image vectorization: converting an image into a d-dimensional vector. Similarly, many mature, pre-trained models can accomplish this task, such as AlexNet, VGG (Visual Geometry Group), and ResNet (Residual Network). Some embodiments of this disclosure may use publicly available, pre-trained VGG models.

[0110] In some embodiments of this disclosure, the vectorized model typically uses a complex neural network such that the more similar the semantics or images, the smaller the cosine distance between their vector representations.

[0111] This disclosure describes the vectorization models for text and images described in the above embodiments, but does not involve specific methods thereof. Information providers can train models specifically for their own text and images, or they can use other pre-trained models.

[0112] Meanwhile, with the rise of short video content in recent years, much information now includes video rather than image content. While there are fewer methods for vectorizing video compared to images, some effective approaches have emerged in recent years. For example, training frame-by-frame and then merging the vectors from all frames for further training; see open-source implementations for details. Video vectorization is not the focus of this invention; this paper merely illustrates that various information formats can be decomposed into text and non-text data, and then vectorized separately.

[0113] The length d of the vectors converted from text and non-text in the embodiments of this disclosure is generally determined empirically. A larger d can describe more complex and detailed text / image features, thus achieving better results; however, this also means greater computational and storage consumption. Therefore, careful trade-offs are required during implementation. In some embodiments of this disclosure, d is set to 256.

[0114] The offline portion described in this section is often very time-consuming. This is because the aforementioned neural networks use an extremely large number of parameters (often hundreds of millions). Whether training or vectorizing existing models, it requires significant GPU (graphics processing unit), CPU (central processing unit), and memory resources. For example, if there are 100,000 pieces of information, and vectorizing each piece of information might take about 50 milliseconds, then the total time required would be approximately two hours. Therefore, the offline portion can run once every T intervals. T can be determined based on specific circumstances; in some embodiments of this disclosure, T can be set to one day.

[0115] In some embodiments of this disclosure, after the information i is vectorized, its text representation vector S is generated. i and the non-textual representation vector G i , respectively stored in the vector indices of text and non-text.

[0116] Step 200, Online Section, Online Information Optimization Service.

[0117] In some embodiments of this disclosure, the online information optimization service in step 200 may include at least one of steps 1-8, wherein:

[0118] Step 1: Information push request and competition determination.

[0119] In some embodiments of this disclosure, step 1 may include: In a real-time information scenario, when a user accesses a webpage or app embedded with information bits, the information platform immediately constructs an information request using the user identifier (usually a device number) and information bit information (such as the image size that the information bit can adapt to, and the text length limit of the information bit). This request is then sent to all eligible information owners. Information triggering is handled by the information platform, and this disclosure does not involve its implementation; it is merely illustrative.

[0120] In some embodiments of this disclosure, step 1 may further include: upon receiving the aforementioned request, the information owner immediately determines whether to participate in the bidding and generates a bid for this information push request. Many real-time information bidding methods exist in the industry, but this invention does not cover them; only the process is described. If, at this stage, the information owner determines not to participate, it will directly reply to the information platform and terminate the information service; only after deciding to participate will the information owner continue to calculate which information to return to the information platform.

[0121] In the above embodiments of this disclosure, the information provider is referred to as an information platform in the RTA scenario and as an information trader in the RTB scenario; in the context of this disclosure, the information provider, information platform, and information trader can be regarded as the same entity.

[0122] Step 2, Click Probability Model. Determine the click probability for each candidate message.

[0123] In some embodiments of this disclosure, step 2 may employ a click probability model such as the Thompson sampler to determine the click probability of each candidate piece of information.

[0124] In step 2, the request initiated by the information platform only contains the user's device ID. We need to recommend the (most likely) click-through rate information for this device ID (denoted as u). Commonly used methods in the industry include UCB (Upper Bound Confidence Algorithm) and Thompson Sampling.

[0125] In some embodiments of this disclosure, such as Figure 2 As shown, this disclosure uses the Thompson sampling method.

[0126] In some embodiments of this disclosure, step 2 may include: for information i, assuming the probability ctr of a user clicking after its display. i It follows a Bernoulli distribution, i.e., CTR i ~Bernoulli(θ), where θ is the expected value of the click probability. Note that θ is not certain and can be described by the beta distribution, the conjugate distribution of the Bernoulli distribution, with its probability density function given by formula (1), where:

[0127]

[0128] In formula (1), α and β represent the number of clicks after display and the number of observations without clicks after display, respectively. The denominator of formula (1) is only used to obtain θ in the interval 0-1. α-1 (1-θ) β-1 The integral value of the integral is given. Placing this integral value in the denominator normalizes the probability density function so that the total probability is 1, thus forming the probability density function.

[0129] In some embodiments of this disclosure, the two parameters α and β are independent of each other for different information.

[0130] Specifically, for different information, the parameters α = 0 and β = 0 are initialized respectively. For example, α is initialized for information i. i β i Subsequently, after information i is displayed, these two parameters will be updated based on whether the user clicks on it (see step 8 below for details).

[0131] After determining the participants, for each piece of information, the beta distribution described above is used for sampling to obtain θ for each piece of information. i That is, the click-through rate (CTR) of each piece of information. i The expected value of ).

[0132] In some embodiments of this disclosure, the Thompson sampling method described in step 2 is commonly used in the industry to solve the multi-armed slot machine problem; however, many other methods can be used as alternatives, such as the confidence interval upper bound method, which simply requires adjusting θ in step 2. i Simply replace it with the upper bound of the confidence interval.

[0133] Step 3, Efficiency Value Model.

[0134] In some embodiments of this disclosure, step 3 may include: for each candidate information, determining the efficiency value of the candidate information for the user terminal, that is, calculating the freshness of the requested device number u for each piece of information.

[0135] In some embodiments of this disclosure, step 3 may include: determining the efficiency value of the candidate information for the user terminal (device number u) based on a quadratic function of the similarity, wherein the constant term of the quadratic function is 0, the coefficient of the quadratic term is less than 0, and the coefficient of the linear term is greater than 0.

[0136] In some embodiments of this disclosure, step 3 may include at least one of steps 31-34, wherein:

[0137] Step 31: Obtain the information display history of u over the recent period τ. Let h be the number of pieces of information that we have displayed to u during this period. u,0 ,h u,1 ,h u,2,h u,3 ...h u,k-1 .

[0138] In some embodiments of this disclosure, τ can be determined empirically during implementation. If the value of τ is small, the efficiency of the information for the user will be relatively high, and vice versa.

[0139] In some embodiments of this disclosure, τ can be set to 1 day, but other values ​​are also possible in specific implementations.

[0140] Step 32: Extract the vector representations of the above k pieces of information, resulting in k vectors of the text (denoted as s). u,0 ,s u,1 ,……s u,k-1 ), and k non-text vectors (let's call them g). u,0 ,g u,1 ,……g u,k-1 ), where s u,j and g u,j Both are vectors of length d (0≤j <k)。

[0141] Step 33: Calculate the historical similarity of all information to user u in sequence. For example, the historical similarity of information i to user u is given by formula (2):

[0142]

[0143] In formula (2), S i G represents the text vector of information i. i The non-text vector representing information i, in formula (2) This can be understood as the sum of the cosine similarity between all the information displayed by user u and the text vector of information i. Similarly, the text vector similarity has been changed to image vector similarity.

[0144] The embodiments of this disclosure above calculate the sum of the cosine similarities of the text vectors and image vectors of information i and user u that have been displayed, and then take the harmonic mean of the two.

[0145] For example: If product A information has been continuously displayed to user terminal u recently, then if information i contains information about product A, r u,i It will be quite large; and if information i is product B, then r u,i The size will be relatively small.

[0146] In some embodiments of this disclosure, it is noted that r u,iThe harmonic mean of text and image similarity was used instead of the arithmetic mean because information similarity can only be considered true when both images and text are very similar. If the text is extremely similar (for example, most of the messages on a platform during Singles' Day might be very similar), but the images are less similar (even though the messages are the same, different types of products are featured), the harmonic mean will be very small. Similarly, if the images are similar but the text is quite different, the harmonic mean will also be relatively small.

[0147] In some embodiments of this disclosure, r u,i This can be used to measure how much information i user u has received in the recent time period τ.

[0148] Step 34, Efficiency Value Model. Calculate the efficiency value (freshness) of all information for user terminal u in sequence. For example, the efficiency value of information i for user terminal u is formula (3):

[0149] f u,i =-a i *r u,i 2 +b i *r u,i (3)

[0150] That is, the constant term is 0, with respect to r. u,i The quadratic function, with coefficient -a of the quadratic term i The coefficient b of the linear term is less than 0. i Greater than 0.

[0151] Figure 1 This is a schematic diagram illustrating the change in efficiency values ​​as displayed information is repeated in some embodiments of this disclosure. A quadratic function is used because the efficiency value of information for user terminal u exhibits the following characteristics: Figure 4 The trend shown.

[0152] When a user first encounters a certain type of information, the efficiency value of the information for the user's terminal u is 0, which means that the user's click rate on the information is purely the result of the calculation described in step 2.

[0153] Then, when users are initially exposed to the same type of information repeatedly, the efficiency of the information may increase until the historical similarity reaches a certain value, at which point the efficiency reaches its peak.

[0154] Finally, if users continue to be exposed to this type of information repeatedly, the efficiency of the information for users will drop rapidly; when the historical similarity reaches a certain threshold, users will feel aversion to similar information, which will not only fail to arouse users' interest, but may also generate negative emotions.

[0155] In formula (3), a i b iEach piece of information is independent of the others, meaning that different types of information have different "tolerance" for their efficiency values. For example, people browse and buy products like medicines and office supplies not out of interest, so their information tolerance is very low; even if the same user is shown the product a few times, its efficiency value may be exhausted. On the other hand, products like digital products, clothing, and shoes have relatively high tolerance.

[0156] In some embodiments of this disclosure, a i b i It is obtained by training a regression model, so that it can reflect the efficiency of information i. We set 'a' for all information. i b i Initially, all values ​​are 0, meaning that initially, the efficiency value of all information for any user is 0. As the data accumulates to a certain level, we update 'a' through model training. i b i The specific training process is detailed in step 8 below.

[0157] Step 4, Information Quality Value Model.

[0158] In some embodiments of this disclosure, step 4 may include steps 41 and 42, wherein:

[0159] Step 41: For each candidate information, determine the quality value of the candidate information based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal.

[0160] In some embodiments of this disclosure, step 41 may include: taking the expected click-through rate θ of the information i obtained in step 2. i The efficiency value of the candidate information for the user terminal obtained in step 3 (i.e., the freshness of information i for user u) is added together and normalized using the sigmoid function to obtain the quality value of information i as shown in formula (4).

[0161] V u,i =sigmoid(c i *θ i +f u,i (4)

[0162] In formula (4), the sigmoid function is a commonly used binary classification function, and its expression is:

[0163] In formula (4), V u,i This study comprehensively considers the expected click-through rate of information i and the efficiency value of information i for user u. If the information is shown to user u for the first time, then the efficiency value f is obviously high. u,i When V is 0, then V u,iActually, it's the result of sampling from a single click probability model.

[0164] c in formula (4) i It is obtained by training a regression model, and its function is to amplify the expected click-through rate obtained from click probability models such as Thompson sampling. If c i If the click-through rate is relatively high, then the quality score of the information is more closely related to its click-through rate, and the efficiency score has little impact. i The initial value is set to 1.0 (any other value greater than 0 is acceptable), which is consistent with the value of 'a' mentioned above. i b i Similar parameters are used, and the specific training process is described in step 8 below.

[0165] Step 42: Select the candidate information with the highest quality value and display it to the user terminal.

[0166] In some embodiments of this disclosure, step 42 may include: after obtaining the quality values ​​of all information pairs for user u, taking the information with the highest quality value and setting it as γ.

[0167] Step 5, Information Response.

[0168] Following step 4, this disclosure selects γ as the information to be displayed to user terminal u. The intended offer described in step 1, along with the content of information γ, is then sent back to the information platform.

[0169] Steps 1 to 5 constitute the processing procedure for the entire information request as described in this disclosure. Because the information platform triggers information very quickly, the processing procedure also needs to be very fast (typically tens of milliseconds). After the above process, the response to the entire information request is complete. The content after step 5 is an asynchronous process, meaning it is independent of the previous processes.

[0170] Step 6, Display monitoring.

[0171] After step 6 is completed, if the information platform determines that the information owner has won the bid, it will immediately display the information owner's information γ. The information owner's display monitoring system will capture this information display in real time and display information γ to user u, as well as r obtained in section 3.2.3.3. u,i Records are displayed in the historical database.

[0172] The monitoring technology itself is not within the scope of this disclosure.

[0173] In some embodiments of this disclosure, step 6 may include: monitoring the display status of the candidate information with the highest quality value after it is pushed to the user terminal, and recording it in the display history database.

[0174] Step 7, click Monitor.

[0175] In some embodiments of this disclosure, step 7 may include: if, within a predetermined time interval, the user clicks on the candidate information with the highest quality value, then the quadratic term coefficient, the linear term coefficient, and the fusion parameter are updated by training the regression model, i.e., step 8 is executed; if, within a predetermined time interval, the user does not click on the candidate information with the highest quality value, then step 8 is executed.

[0176] In some embodiments of this disclosure, step 7 may include: if, within M minutes after the information γ displayed to user u in step 6 is detected, user u clicks on information γ, then the model update in step 8 is performed; otherwise, after M minutes, it is assumed that user u has not clicked on information γ, and the model update in step 8 is performed in the same manner.

[0177] In some embodiments of this disclosure, the value of M can be determined according to specific circumstances during implementation. For example, in the case of mobile information displayed on a mobile phone, users generally either click on the information quickly or skip it quickly, so M is generally smaller in the mobile information scenario. In the case of PC (Personal Computer) information, users may have opened other windows to browse and only return to the information page and click on the information after a long time, so M is relatively larger in the PC information scenario.

[0178] In some embodiments of this disclosure, M can be set to 10.

[0179] Step 8, model update.

[0180] This step is triggered by step 7. The identifier ζ is used to indicate whether a click has occurred (1 for a click, 0 for no click). There are two cases:

[0181] In some embodiments of this disclosure, step 7 may include at least one of steps 71-73, wherein:

[0182] Step 71: Determine whether this display is the first display of the user terminal in the most recent scheduled time period.

[0183] Step 72: If this display is the first display by the user terminal in the most recent scheduled time period, then update the number of observations of clicks after the display and the number of observations of no clicks after the display.

[0184] In some embodiments of this disclosure, step 72 may include: if this display is the first display for user u within the most recent τ time period, then update the click probability model such as the Thompson sampler: if ζ = 1, then α γ =α γ +1; otherwise β γ =β γ +1.

[0185] Step 73, if the current display is not the first display of the user terminal in the most recent predetermined time period, update the observed number of clicks after the display and the observed number of non-clicks after the display, and perform the step of updating the quadratic coefficient, the linear coefficient, and the fusion parameter by training the regression model, that is, for the parameter a γ , b γ , c γ for update.

[0186] In some embodiments of the present disclosure, step 73 may include at least one of steps 731-step 733, where:

[0187] Step 731, retrieve all the display histories of the information γ (recorded in step 6). Assume there are n displays in total. Let the historical similarity calculated at the t-th display be r t , and whether it is finally clicked is denoted as y t (y t ∈[0, 1]).

[0188] Step 732, according to formula (4) in step 4, V u,γ = sigmoid(c γ *θ γ + f u,i ) = sigmoid(c γ *θ γ - a γ *r t 2 + b γ *r t ).

[0189] In the above formula, θ γ is the expected click-through rate of the information γ ( the expectation of the beta distribution), which is currently known. Therefore, V u,γ becomes a function of r t , denoted as

[0190] Step 733, according to the above formula, train the logistic regression model with parameters a γ , b γ , c γ . Specifically, take (r t , y t ) as the training samples (0 ≤ t < n), and use the gradient descent method to solve for a γ , b γ , c γ parameters when the loss function in formula (5) is minimized:

[0191]

[0192] After solving formula (5), update a of the information γγ b γ c γ parameter.

[0193] Thus, this disclosure completes the parameter updates for the click probability model, efficiency value model, and information quality value model.

[0194] In some embodiments of this disclosure, the information quality value model method in step 4 uses logistic regression for expression and training. There are also implementation methods based on neural networks, factorization machines, etc. in the industry. Although alternative solutions may have a slight advantage in effect, logistic regression is more practical in terms of overall performance.

[0195] The embodiments of this disclosure employ a method for calculating the similarity between a user's historical displayed information and current information by pre-vectorizing the text and non-textual aspects of the information.

[0196] The above embodiments of this disclosure describe a method for measuring the efficiency of current information to the user by comparing the similarity between the user's historical displayed information and the current information.

[0197] The above embodiments of this disclosure use logistic regression to create an information quality value model that combines the efficiency value model with the click probability model.

[0198] The above embodiments of this disclosure update the above efficiency value model and information quality value model by monitoring the display and clicks of information.

[0199] The methods described in the above embodiments and the corresponding technical architecture improve information efficiency and ensure system performance, and should be protected.

[0200] In the embodiments of this disclosure, the click probability model is only updated when user u sees the information for the most recent first time. Therefore, the click probability model reflects only the click-through rate of the information itself, completely disregarding the psychological effects of repeated exposure. Thus, the design of the embodiments of this disclosure effectively separates the actual effect of the information from the effect loss caused by repeated exposure (the user's psychological effect), avoiding inaccuracies in the model due to the overlap of the two.

[0201] The online information selection service described in the above embodiments of this disclosure selects information with better click-through rates using a click probability model (e.g., Thompson sampler), measures the psychological impact of historically displayed information on users using an efficiency value model, and combines the two using an information quality value model. In this way, this disclosure solves two major technical problems in related technologies.

[0202] Compared to the third method in related technologies, this disclosure effectively achieves a balance between centralization and decentralization without adding any additional data. Centralization refers to the disclosure's aim to ensure that the information with the best results and highest click-through rate receives the most display opportunities; decentralization refers to the disclosure's aim to avoid displaying similar information too frequently to each user, so as not to cause users to lose interest or even become annoyed. Specifically, this disclosure brings the following beneficial effects:

[0203] The above-described embodiments of this disclosure innovatively propose an efficiency value model for information to user terminals and combine it with click probability models such as the Thompson sampling method, which are relatively mature in the industry. As a result, this invention can not only show users information with higher click rates, but also avoid frequently showing similar information to the same user.

[0204] The embodiments disclosed above do not require the use of an information platform to provide user information or information context information, and there is no risk to user privacy. They can be widely used in commercial systems.

[0205] As described in step 5 above, in real-time internet information scenarios, each information request must be responded to within tens of milliseconds. The methods used in the above embodiments of this disclosure, whether Thompson sampling or logistic regression, are very mature and computationally simple. The above embodiments of this disclosure require very few computing resources to complete the calculation and respond to information requests within tens of milliseconds.

[0206] Figure 5 These are schematic diagrams illustrating some embodiments of the information push device disclosed herein. For example... Figure 5 As shown, the information push device disclosed herein may include a historical display information acquisition module 51, a historical similarity determination module 52, an efficiency value determination module 53, a click probability determination module 54, a quality value determination module 55, and an information push module 56, wherein:

[0207] The historical display information acquisition module 51 is used to acquire the historical display information displayed by the user terminal in the most recent scheduled time period.

[0208] The historical similarity determination module 52 is used to determine the historical similarity between each candidate piece of information and the historical display information of the user terminal for each user terminal.

[0209] In some embodiments of this disclosure, the information push device may also be used to pre-acquire text vectors and non-text vectors for each piece of information, wherein the non-text vectors include at least one of image vectors or video vectors.

[0210] In some embodiments of this disclosure, the historical similarity determination module 52 can be used to determine text similarity based on the text vector of each candidate information and the text vector of the historical display information; determine non-text similarity based on the non-text vector of each candidate information and the non-text vector of the historical display information; and take the harmonic mean of the text similarity and non-text similarity as the historical similarity.

[0211] The efficiency value determination module 53 is used to determine the efficiency value of the candidate information for the user terminal based on the historical similarity.

[0212] In some embodiments of this disclosure, the efficiency value determination module 53 can be used to determine the efficiency value of the candidate information for the user terminal based on the quadratic function of the similarity, wherein the constant term of the quadratic function is 0, the coefficient of the quadratic term is less than 0, and the coefficient of the linear term is greater than 0.

[0213] Click probability determination module 54 is used to determine the click probability of each candidate information.

[0214] The quality value determination module 55 is used to determine the quality value of each candidate information based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal.

[0215] In some embodiments of this disclosure, the quality value determination module 55 can be used to determine the quality value of the candidate information based on the sum of the product of the click probability of the candidate information and the fusion parameter and the efficiency value of the candidate information for the user terminal.

[0216] The information push module 56 is used to push the candidate information with the highest quality value to the user terminal for display.

[0217] In some embodiments of this disclosure, the information push device can also be used to update the quadratic coefficients, linear coefficients, and fusion parameters by training a regression model.

[0218] In some embodiments of this disclosure, the information push device can monitor the display of the candidate information with the highest quality value after it is pushed to the user terminal and record it in the display history database, after updating the quadratic coefficient, linear coefficient and fusion parameter by training a regression model. Within a predetermined time interval after the candidate information with the highest quality value is detected and pushed to the user terminal for display, the device can monitor the user's click on the candidate information with the highest quality value and update the quadratic coefficient, linear coefficient and fusion parameter by training a regression model.

[0219] In some embodiments of this disclosure, the information push device monitors user clicks on candidate information with the highest quality value. When updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model, it can be used to update the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model if the user clicks on the candidate information with the highest quality value within a predetermined time interval; and to perform the step of updating the quadratic coefficient, linear coefficient, and fusion parameter by training a regression model if the user does not click on the candidate information with the highest quality value within the predetermined time interval.

[0220] In some embodiments of this disclosure, the information push device can also be used to determine whether the current display is the first display of the user terminal in the most recent predetermined time period; if the current display is the first display of the user terminal in the most recent predetermined time period, then update the number of observations of clicks after the display and the number of observations of no clicks after the display; if the current display is not the first display of the user terminal in the most recent predetermined time period, then update the number of observations of clicks after the display and the number of observations of no clicks after the display, and perform the step of updating the quadratic term coefficient, linear term coefficient and fusion parameter by training the regression model.

[0221] In some embodiments of this disclosure, the information push device can be used to obtain all display history of a piece of information by updating the quadratic coefficient, linear coefficient, and fusion parameter through training a regression model. The historical similarity calculated at each display and the click identifier after each display are used as training samples. The gradient descent method is used to solve for the quadratic coefficient, linear coefficient, and fusion parameter when the loss function is minimized. The original quadratic coefficient, linear coefficient, and fusion parameter are updated using the quadratic coefficient, linear coefficient, and fusion parameter when the loss function is minimized.

[0222] In some embodiments of this disclosure, the information push device is used to perform any of the embodiments described above (e.g., Figures 1-4 The operation of the information push method described in any embodiment.

[0223] Based on the information push device provided in the above embodiments of this disclosure, an innovative information efficiency value (freshness) model for user terminals is proposed and combined with a relatively mature click probability model in the industry (such as the Thompson sampling method). The above embodiments of this disclosure can not only show users information with higher click rates, but also avoid frequently showing similar information to the same user.

[0224] Figure 6 The diagram shows the structure of some other embodiments of the information push device disclosed herein. For example... Figure 6 As shown, the information push device includes a memory 61 and a processor 62.

[0225] Memory 61 is used to store instructions, and processor 62 is coupled to memory 61. Processor 62 is configured to execute instructions stored in memory to implement any of the above embodiments (e.g., Figures 1-4 The information push method described in any embodiment.

[0226] like Figure 6 As shown, the information push device also includes a communication interface 63 for information interaction with other devices. Simultaneously, the information push device also includes a bus 64, through which the processor 62, communication interface 63, and memory 61 communicate with each other.

[0227] Memory 61 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive. Memory 61 may also be a memory array. Memory 61 may also be divided into blocks, and the blocks may be combined into virtual volumes according to certain rules.

[0228] Furthermore, processor 62 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present disclosure.

[0229] In the embodiments of this disclosure, the click probability model is only updated when user u sees the information for the most recent first time. Therefore, the click probability model reflects only the click-through rate of the information itself, completely disregarding the psychological effects of repeated exposure. Thus, the design of the embodiments of this disclosure effectively separates the actual effect of the information from the effect loss caused by repeated exposure (the user's psychological effect), avoiding inaccuracies in the model due to the overlap of the two.

[0230] The online information selection service described in the above embodiments of this disclosure selects information with better click-through rates using a click probability model (e.g., Thompson sampler), measures the psychological impact of historically displayed information on users using an efficiency value model, and combines the two using an information quality value model. In this way, this disclosure solves two major technical problems in related technologies.

[0231] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions that, when executed by a processor, implement any of the embodiments described above (e.g., Figures 1-4 The information push method described in any embodiment.

[0232] Compared to the third method in related technologies, this disclosure effectively achieves a balance between centralization and decentralization without adding any additional data. Centralization refers to the disclosure's aim to ensure that the information with the best results and highest click-through rate receives the most display opportunities; decentralization refers to the disclosure's aim to avoid displaying similar information too frequently to each user, so as not to cause users to lose interest or even become annoyed. Specifically, this disclosure brings the following beneficial effects:

[0233] The above-described embodiments of this disclosure innovatively propose an efficiency value model for information to user terminals and combine it with click probability models such as the Thompson sampling method, which are relatively mature in the industry. As a result, this invention can not only show users information with higher click rates, but also avoid frequently showing similar information to the same user.

[0234] The embodiments disclosed above do not require the use of an information platform to provide user information or information context information, and there is no risk to user privacy. They can be widely used in commercial systems.

[0235] As described in step 5 above, in real-time internet information scenarios, each information request must be responded to within tens of milliseconds. The methods used in the above embodiments of this disclosure, whether Thompson sampling or logistic regression, are very mature and computationally simple. The above embodiments of this disclosure require very few computing resources to complete the calculation and respond to information requests within tens of milliseconds.

[0236] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, apparatus, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable non-transitory storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0237] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0238] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0239] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0240] The information push device described above can be implemented as a general-purpose processor, programmable logic controller (PLC), digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component or any suitable combination thereof for performing the functions described in this application.

[0241] This concludes the detailed description of the present disclosure. To avoid obscuring the concept of the disclosure, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions disclosed herein based on the above description.

[0242] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing the relevant hardware to implement them. The program can be stored in a non-transitory computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0243] The description in this disclosure is provided for illustrative and descriptive purposes only and is not intended to be exhaustive or to limit the disclosure to its forms. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of this disclosure and to enable those skilled in the art to understand this disclosure and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. An information push method, characterized in that, include: Obtain historical display information of the user terminal during the most recently scheduled time period; For each user terminal, the historical similarity between each candidate information and the historical display information of that user terminal is determined, wherein the historical similarity is used to measure the number of related historical display information of the candidate information that the user terminal has received in the most recent predetermined time. Based on the historical similarity, the efficiency value of the candidate information for the user terminal is determined, wherein when the historical similarity is less than a predetermined value, the efficiency value increases with the increase of the historical similarity, and when the historical similarity is greater than the predetermined value, the efficiency value decreases with the increase of the historical similarity. Determine the click probability for each candidate message; For each candidate information, a quality value is determined based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal, wherein the quality value decreases as the efficiency value decreases; The candidate information with the highest quality score is pushed to the user's terminal for display.

2. The information push method according to claim 1, characterized in that, Also includes: The text vector and non-text vector of each piece of information are obtained in advance, wherein the non-text vector includes at least one of image vector or video vector; Determining the historical similarity between each candidate piece of information and the historical display information of the user terminal includes: The text similarity is determined based on the text vector of each candidate information and the text vector of the historical displayed information. Non-text similarity is determined based on the non-text vector of each candidate information and the non-text vector of the historical displayed information; The harmonic mean of text similarity and non-text similarity is used as the historical similarity.

3. The information push method according to claim 1 or 2, characterized in that, The step of determining the efficiency value of each candidate information for the user terminal based on its historical similarity to historically displayed information includes: Based on the quadratic function of the similarity, the efficiency value of the candidate information for the user terminal is determined, wherein the constant term of the quadratic function is 0, the coefficient of the quadratic term is less than 0, and the coefficient of the linear term is greater than 0.

4. The information push method according to claim 3, characterized in that, The process of determining the quality value of the candidate information based on its click probability and efficiency value for the user terminal includes: The quality value of the candidate information is determined by summing the product of the click probability of the candidate information and the fusion parameter with the efficiency value of the candidate information for the user terminal.

5. The information push method according to claim 4, characterized in that, Also includes: The quadratic coefficient, linear coefficient, and fusion parameter are updated by training the regression model.

6. The information push method according to claim 5, characterized in that, The step of updating the quadratic term coefficient, linear term coefficient, and fusion parameter by training the regression model includes: The display status of the candidate information with the highest quality value after it is pushed to the user's terminal is monitored and recorded in the display history database; Within a predetermined time interval after the candidate information with the highest quality value is pushed to the user terminal for display, the user's clicks on the candidate information with the highest quality value are monitored, and the quadratic term coefficient, linear term coefficient, and fusion parameter are updated by training a regression model.

7. The information push method according to claim 6, characterized in that, The monitoring of user clicks on candidate information with the highest quality value, and the updating of quadratic term coefficients, linear term coefficients, and fusion parameters through training a regression model, includes: If, within a predetermined time interval, the user is detected to have clicked on the candidate information with the highest quality score, the quadratic term coefficient, linear term coefficient, and fusion parameter are updated by training the regression model. If, within a predetermined time interval, the user does not click on the candidate information with the highest quality value, then the step of updating the quadratic term coefficient, linear term coefficient, and fusion parameter by training the regression model is executed.

8. The information push method according to claim 7, characterized in that, Also includes: Determine whether this display is the first time the user terminal has displayed the device within the most recently scheduled time period; If this display is the first display on the user terminal within the most recent scheduled time period, then update the number of observations of clicks after the display and the number of observations of no clicks after the display; If this display is not the first display by the user terminal in the most recent scheduled time period, then update the number of observations that clicked after the display and the number of observations that did not click after the display, and execute the steps of updating the quadratic term coefficients, linear term coefficients and fusion parameters by training the regression model.

9. The information push method according to claim 8, characterized in that, The step of updating the quadratic term coefficient, linear term coefficient, and fusion parameter by training the regression model includes: To obtain all display history of a piece of information, the historical similarity calculated at each display and the click identifier after each display are used as training samples. The gradient descent method is used to solve for the quadratic term coefficient, linear term coefficient and fusion parameter when the loss function is minimized. The original quadratic coefficients, linear coefficients, and fusion parameters are updated using the quadratic coefficients, linear coefficients, and fusion parameters when the loss function is minimized.

10. An information push device, characterized in that, include: The historical display information acquisition module is used to acquire the historical display information displayed by the user terminal in the most recent scheduled time period; The historical similarity determination module is used to determine the historical similarity between each candidate information and the historical display information of the user terminal for each user terminal, wherein the historical similarity is used to measure the number of related historical display information of the candidate information that the user terminal has received in the most recent predetermined time. An efficiency value determination module is used to determine the efficiency value of the candidate information for the user terminal based on the historical similarity, wherein when the historical similarity is less than a predetermined value, the efficiency value increases as the historical similarity increases, and when the historical similarity is greater than the predetermined value, the efficiency value decreases as the historical similarity increases. The click probability determination module is used to determine the click probability of each candidate message. A quality value determination module is used to determine the quality value of each candidate information based on the click probability of the candidate information and the efficiency value of the candidate information for the user terminal, wherein the quality value decreases as the efficiency value decreases; The information push module is used to push the candidate information with the highest quality value to the user terminal for display.

11. The information push device according to claim 10, characterized in that, The information push device is used to perform operations that implement the information push method as described in any one of claims 2-9.

12. An information push device, characterized in that, include: Memory, used to store instructions; A processor is configured to execute the instructions, causing the information push device to perform operations that implement the information push method as described in any one of claims 1-9.

13. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that, when executed by a processor, implement the information push method as described in any one of claims 1-9.

Images