Multi-objective information recommendation method, system and device based on position debiasing

Abstract

The present invention discloses a multi-objective information recommendation method, system and device based on position debiasing, the method comprising: obtaining log data of user browsing information and preprocessing and parsing to obtain user-related data sets, location-related data sets and information data sets; extracting features from the user-related data sets to obtain a user feature set, and extracting features from the information data sets to obtain an information feature set; constructing an information recommendation pre-training model; training the information recommendation pre-training model based on sample label data, location-related data and sample feature data to obtain an information recommendation model; reasoning the recommended data based on the information recommendation model to obtain click preference scores and browsing preference scores, and then obtaining information recommendation results. The present invention solves the limitations of existing information recommendation methods on information arrangement order and effective browsing, improves the accuracy and user experience of the information recommendation system, and meets the user's needs for personalized and efficient information acquisition.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation algorithms, and in particular to a multi-target information recommendation method, system and device based on position debiasing. Background Art

[0002] At present, information recommendation algorithms have been widely used in Internet products, aiming to provide users with personalized recommendation services and realize personalized experience of "one thousand faces for one thousand people". The core technology of information recommendation algorithms is to estimate the conversion rate of users to information. If the conversion rate of users to information can be accurately estimated, it means that information can be absorbed by users more efficiently. Therefore, recommending information with high conversion rate to corresponding users can achieve the purpose of improving platform interests and user stickiness, and then realize the sustainable and stable development of the platform.

[0003] Existing information recommendation algorithms can integrate multiple information such as user information and information content in terms of information and user ratings, that is, pay more attention to users, information and contextual factors, estimate the conversion rate based on complex network models, and obtain recommendation results. However, some current recommendation algorithms ignore the impact of information position on conversion rate, making it difficult for information recommendation algorithms to achieve optimal results. For example, the position of information in the information list is ignored, while the order of information in the information list can directly affect the user's attention and conversion rate. In other words, information occupying the front of the information list can attract more attention from users due to its position advantage, and therefore can achieve a higher conversion rate.

[0004] In addition, the existing information recommendation algorithm has a limitation that it focuses too much on user clicks (which can be understood as user click-through rate). This will ignore the information browsing behavior and information browsing status of users after clicking on the information and entering the details page. Therefore, it will cause "clickbait" information to have a high click-through rate, but in fact the user may not have actual information browsing behavior and information browsing status, which seriously affects the accuracy of information recommendations and the conversion rate of information.

[0005] In other words, although there are some information recommendation algorithms currently, the existing information recommendation algorithms basically limit the matching degree between the information recommendation content and the user's interests, resulting in inaccurate recommendations to users, failing to meet the user's expectations for personalized and efficient information recommendations, and may even have an adverse impact on the user experience. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a multi-target information recommendation method, system and device based on position debiasing.

[0007] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0008] A multi-objective information recommendation method based on position debiasing comprises the following steps:

[0009] Obtaining the log data of user browsing information and preprocessing and parsing it to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set;

[0010] Performing feature extraction on a user-related data set to obtain a user feature set, and performing feature extraction on an information data set to obtain an information feature set, wherein the information feature set includes information features, context features, and information sequence features; obtaining position bias data based on position-related data and then performing feature extraction to obtain a position bias feature set;

[0011] Construct an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability.

[0012] The information recommendation pre-training model is trained based on the user-related data set, the location-related data set and the information-related data set to obtain an information recommendation model;

[0013] Based on the information recommendation model, the recommended information data is inferred to obtain the user's click preference score and browsing preference score for the information, and the information recommendation results are obtained through the click preference score and browsing preference score.

[0014] As an implementable method, the information tag set includes information product tags, click behavior tags and effective browsing tags;

[0015] The user characteristics include demographic characteristics, user device characteristics, user activity characteristics and user information preference characteristics;

[0016] The information features include information attribute features and information tag features, and the information tag features include information click tags, information browsing tags and effective browsing tags;

[0017] The context features include the currently recommended information topic and time period;

[0018] The information sequence features include the information sequence clicked by the user and the information sequence browsed by the user.

[0019] As an implementable method, the effective browsing tag is obtained by the following steps:

[0020] Obtain the word length of the current information content, and calculate based on the word length and the average word length of browsing per second to obtain the information browsing time;

[0021] Get the user's browsing time, and make a judgment based on the information browsing time and the user's browsing time. When the user's browsing time is longer than the preset information browsing time, it is a valid browsing tag, which is expressed as follows:

[0022]

[0023] in, Indicates the user's browsing time. Indicates the length of words. Indicates the average length of words browsed per second. Represents a constant, Indicates a valid browsing tag.

[0024] As an implementable method, the shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations, including the following steps:

[0025] The user features, information features and context features are converted using the look-up method to obtain the corresponding vector representations;

[0026] The information sequence features are transformed using a training method to obtain an information sequence feature vector.

[0027] As an implementable method, the main structure layer processes the vector representation to obtain the information click feature vector and the information browsing feature vector, including the following steps:

[0028] The information feature vector and information sequence feature vector are processed through the attention mechanism to obtain the cross-representation of user information;

[0029] The user information cross representation, the user feature vector and the context feature vector are concatenated to obtain a user concatenated vector;

[0030] Through at least two sets of fully connected layers and The function optimizes the user splicing vector to obtain the information click feature vector and the information browsing feature vector;

[0031] The attention mechanism is expressed as follows:

[0032]

[0033]

[0034]

[0035] Said Function, expressed as follows:

[0036]

[0037] The user splicing vector is expressed as follows:

[0038]

[0039] in, Indicates user information representation, represents the information feature vector, represents the target information vector, represents the weight parameter, Indicates information feature vector, represents the number of information feature vectors, Indicates The input of the layer network, Indicates the number of network layers, Indicates The parameters of the layer network, A, B, C represent the cross representation of user information, user feature vector and context feature vector respectively. They represent the corresponding weight coefficients respectively and their sum is 1.

[0040] As an implementable method, the position debiasing layer obtains the position bias feature set, the information click feature vector and the information browsing feature vector and processes them to obtain the information click probability and the information browsing probability, including the following steps:

[0041] The position debiasing layer includes a first input layer, an embedding layer, a fully connected layer and an output layer;

[0042] The first input layer receives and acquires the position offset feature set, the information click feature vector, and the information browsing feature vector;

[0043] The embedding layer performs end-to-end embedding on the position feature set and inputs it into the fully connected layer;

[0044] The fully connected layer analyzes the position feature set to obtain the position offset vector;

[0045] The output layer performs weighted operations on the position bias vector, the information click feature vector and the information browse feature vector to obtain a biased click vector and a biased browse vector, and inputs the biased click vector and the biased browse vector into Function is used to convert and obtain the probability of information click and information browsing;

[0046] Said The function is expressed as follows:

[0047]

[0048] in, Indicates the biased click vector or biased browsing vector, Indicates the probability of information click or information browsing.

[0049] As an implementable method, the following steps are also included:

[0050] Constructing a loss function to optimize the training process through the loss function, wherein the loss function includes an information click loss function and an information browsing loss function;

[0051] Construct an error model, and judge the accuracy of the information recommendation model based on the error model to obtain a judgment result. When the judgment result is the smallest, it means that the information recommendation model training is completed;

[0052] The information click loss function is expressed as follows:

[0053]

[0054] The information browsing loss function is expressed as follows:

[0055]

[0056] The error model is expressed as follows:

[0057]

[0058] in, represents the information click loss function, represents the information browsing loss function, Indicates clicking a label. Indicates the probability of information click. represents the probability of information browsing, Indicates the number of sample feature data, Indicates browsing tags. Indicates the actual value, represents the predicted value, Indicates the number of times, Indicates Second-rate, Indicates the error result.

[0059] As an implementable method, obtaining information recommendation results by clicking preference scores and browsing preference scores includes the following steps:

[0060] The click preference score and the browsing preference score are weighted to obtain a fusion preference score;

[0061] Sort the fusion preference scores, analyze based on the sorting results, and obtain information recommendation results;

[0062] Wherein, the fusion preference score is expressed as follows:

[0063]

[0064]

[0065]

[0066] in, represents the fusion preference score, represents the click preference score, represents the browsing preference score, represents the weight, , Indicates the average click-through rate of information. Indicates the average browsing rate of information. Indicates the number of information clicks within a fixed period of time. Indicates the amount of information exposure within a fixed period of time. Indicates the number of effective views of information within a fixed period of time.

[0067] As an implementable method, the log data of user browsing information is obtained and preprocessed and parsed to obtain a user-related data set, a location-related data set and an information data set, including the following steps:

[0068] Collect the log data of users browsing information, judge and eliminate abnormal data, and obtain the pre-processed log data set;

[0069] Extract and analyze the preprocessed log data set based on at least one or more of user behavior analysis, content preference analysis, and time series analysis to obtain a user-related data set, a location-related data set, and an information data set;

[0070] The anomaly score is calculated by the following formula, and the anomaly score is used to determine whether it is abnormal and remove abnormal data, as shown below:

[0071]

[0072] in, represents the anomaly score of the log data, Indicates the amount of data, represents the first quartile of the data, Represents the third quartile of the data. represents the interquartile range, is a constant, if Not less than 0 The value is 1 if yes, otherwise it is 0.

[0073] A multi-objective information recommendation system based on position debiasing, comprising an acquisition processing module, a model building module, a model training module and a model reasoning module;

[0074] The acquisition processing module acquires the log data of the user browsing information and performs preprocessing and analysis to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set; performs feature extraction on the user-related data set to obtain a user feature set, and performs feature extraction on the information data set to obtain an information feature set, wherein the information feature set includes information features, context features and information sequence features; obtains position bias data based on the position-related data and then performs feature extraction to obtain a position bias feature set;

[0075] The model construction module constructs an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability.

[0076] The model training module trains the information recommendation pre-training model based on the user-related data set, the location-related data set and the information-related data set to obtain the information recommendation model;

[0077] The model reasoning module reasons on the recommended information data based on the information recommendation model to obtain the user's click preference score and browsing preference score for the information, and obtains the information recommendation result through the click preference score and browsing preference score.

[0078] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following method is implemented:

[0079] Obtaining the log data of user browsing information and preprocessing and parsing it to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set;

[0080] Performing feature extraction on a user-related data set to obtain a user feature set, and performing feature extraction on an information data set to obtain an information feature set, wherein the information feature set includes information features, context features, and information sequence features; obtaining position bias data based on position-related data and then performing feature extraction to obtain a position bias feature set;

[0081] Construct an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability.

[0082] The information recommendation pre-training model is trained based on the user-related data set, the location-related data set and the information-related data set to obtain an information recommendation model;

[0083] Based on the information recommendation model, the recommended information data is inferred to obtain the user's click preference score and browsing preference score for the information, and the information recommendation results are obtained through the click preference score and browsing preference score.

[0084] A multi-target information recommendation device based on position debiasing includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the following method when executing the computer program:

[0085] Obtaining the log data of user browsing information and preprocessing and parsing it to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set;

[0086] Performing feature extraction on a user-related data set to obtain a user feature set, and performing feature extraction on an information data set to obtain an information feature set, wherein the information feature set includes information features, context features, and information sequence features; obtaining position bias data based on position-related data and then performing feature extraction to obtain a position bias feature set;

[0087] Construct an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability.

[0088] The information recommendation pre-training model is trained based on the user-related data set, the location-related data set and the information-related data set to obtain an information recommendation model;

[0089] Based on the information recommendation model, the recommended information data is inferred to obtain the user's click preference score and browsing preference score for the information, and the information recommendation results are obtained through the click preference score and browsing preference score.

[0090] The present invention has significant technical effects due to the adoption of the above technical solution:

[0091] The present invention designs a multi-objective network that conforms to the actual state of information and designs a location debiasing auxiliary network to build an information recommendation pre-training model. At the same time, it obtains sample feature data based on user log data and innovatively introduces location-related data. By iteratively training the model, an information recommendation model is obtained, and data reasoning is performed based on the trained model to obtain information recommendation results. The present invention significantly improves the user's information service experience by measuring the influence of the user's deep-level interests and the location distribution of information, and realizes efficient and accurate information distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0093] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0094] Figure 2 It is an overall schematic diagram of the system of the present invention;

[0095] Figure 3 It is a schematic diagram of a specific implementation of the method of the present invention;

[0096] Figure 4is a network schematic diagram of the method of the present invention;

[0097] Figure 5 Schematic diagram of the attention mechanism of the method of the present invention;

[0098] Figure 6 It is a schematic diagram of the reasoning process of the method of the present invention. DETAILED DESCRIPTION

[0099] The present invention is further described in detail below in conjunction with embodiments. The following embodiments are for explanation of the present invention but the present invention is not limited to the following embodiments.

[0100] Embodiment 1:

[0101] A multi-objective information recommendation method based on position debiasing, such as Figure 1 As shown, the following steps are included:

[0102] S100, obtaining log data of user browsing information and performing preprocessing and parsing to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set;

[0103] S200, extracting features from a user-related data set to obtain a user feature set, and extracting features from an information data set to obtain an information feature set, wherein the information feature set includes information features, context features, and information sequence features; obtaining position bias data based on position-related data and then extracting features to obtain a position bias feature set;

[0104] S300, constructing an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability.

[0105] S400, training an information recommendation pre-training model based on a user-related data set, a location-related data set, and an information-related data set to obtain an information recommendation model;

[0106] S500: Inferring the recommended information data based on the information recommendation model to obtain the user's click preference score and browsing preference score for the information, and obtaining the information recommendation result through the click preference score and browsing preference score.

[0107] The present invention analyzes the user's operation behavior in detail, designs a multi-target network that conforms to the actual state of information browsing, obtains user-related data sets, location-related data sets and information data sets based on user log data, and iteratively trains the recommendation network based on the relevant data to obtain an information recommendation model. Through model reasoning, efficient and accurate information distribution is achieved, providing users with personalized and intelligent information service experience, significantly improving the accuracy and efficiency of information distribution, thereby improving the platform's user stickiness and providing support for the platform's sustainable development and revenue.

[0108] The present invention adopts a more comprehensive strategy in the process of information recommendation prediction, such as Figure 3 As shown in the figure, we design multiple objectives based on the user's browsing status of information, obtain sample feature data and sample label data through user log data, creatively introduce location-related data of information information, and explore a multi-objective analysis method that can effectively integrate information click data information and user browsing behavior in the details page, build an information recommendation pre-training model, and iteratively train the information recommendation pre-training model based on sample label data, sample feature data and location-related data by designing and optimizing the objective function to obtain the information recommendation model, and finally perform reasoning analysis based on the information recommendation model to obtain the information recommendation result. We achieve efficient and accurate information recommendation, provide users with a more personalized and intelligent information service experience, and significantly improve the efficiency and accuracy of information recommendation.

[0109] In the process of constructing sample data, the present invention designs two core goals: information clicks and effective browsing. Among them, information clicks, as the primary link of information browsing, aim to attract users and conduct preliminary interactions; effective browsing is used to measure users' deep interest and participation in information. Specifically, when designing an information recommendation model, it is necessary to first consider whether the user is interested in the information module, and also to consider the browsing status of the user after clicking into the information details page, that is, whether the user has achieved effective browsing of the information information, which can better express the user's real interest and realize the transition from clicks to effective browsing. In the process of effective browsing by users, the user's behavior constitutes the user behavior link: browsing the information list -> clicking on information -> browsing information.

[0110] After acquiring the data, the data is preprocessed and analyzed to obtain user-related data sets, location-related data sets and information data sets, wherein the information data sets include information tag sets and information-related data sets. The information tag sets include information product tags, click behavior tags and effective browsing tags.

[0111] In one embodiment, it is necessary to make a detailed distinction between information product labels, click behavior labels and effective browsing labels. By simulating user browsing behavior, the effective browsing label indicates whether the user has effectively browsed. Effective browsing is not just a simple up and down slide, but is determined based on the viewing time. The time required to browse the full text of the information is calculated by the length of the information, and the actual length of time the user stays on the current information details page is counted. When the actual stay time is a preset multiple of the full text browsing time, it is considered that the user has effectively browsed the current information and an effective browsing label is obtained. The effective browsing label indicates that the user is interested in the current information rather than clicking by mistake. Among them, the effective browsing label is expressed as follows:

[0112]

[0113] in, Indicates the user's browsing time. Indicates the length of words. Indicates the average length of words browsed per second. Represents a constant, Indicates a valid browsing tag.

[0114] Of course, not all acquired data can be used, and the data needs to be screened. The acquisition of the log data of user browsing information and preprocessing and parsing to obtain user-related data sets, location-related data sets and information data sets includes the following steps:

[0115] Collect the log data of users browsing information, judge and eliminate abnormal data, and obtain the pre-processed log data set;

[0116] Extract and analyze the preprocessed log data set based on at least one or more of user behavior analysis, content preference analysis, and time series analysis to obtain a user-related data set, a location-related data set, and an information data set;

[0117] The anomaly score is calculated by the following formula, and the anomaly score is used to determine whether it is abnormal and remove abnormal data, as shown below:

[0118]

[0119] in, represents the anomaly score of the log data, Indicates the amount of data, represents the first quartile of the data, Represents the third quartile of the data. represents the interquartile range, is a constant, if Not less than 0 The value is 1 if yes, otherwise it is 0.

[0120] By extracting features from user log data, the features required for sample data are calculated, including user feature data, information feature data, context feature data, and information sequence feature data. In this embodiment, user feature data includes demographic attribute features (age, gender, and region, etc.), user device features (device operating system and device operator, etc.), user activity features (platform activity, scene activity, and user account level, etc.), and user information preference features (calculated through user clicks, browsing information, etc.); information feature data includes information attribute features (labels, release time and source, etc.), information pre-training features (trained through information co-occurrence data); context feature data includes the currently recommended information topic and the currently recommended time period, etc.; information sequence feature data includes the user information sequence composed of clicks or browsing by users over a period of time.

[0121] The impact of information location on information browsing. This embodiment takes the information recommendation module on the homepage of an app as an example. When the user enters the app, the user first sees the top information. The information behind requires the user to slide to be effectively exposed. At this time, the position advantage of the information is obvious. The information in the front has an exposure advantage and has more chances of being clicked and browsed. Therefore, if direct training is performed based on a conventional algorithm model, the resulting model will naturally be biased towards information located in the front row, thereby causing a position deviation problem and affecting the accuracy of information recommendation. The present invention introduces position-related data to weaken the click-through rate difference caused by information location information for information browsing, thereby improving the accuracy of the information recommendation model.

[0122] In this embodiment, the information recommendation pre-training model constructed is composed of a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. Figure 4 As shown, the data input layer is used to receive sample feature data; the shared feature layer converts the sample feature data received by the data input layer into vectors, that is, the user features, information features and context features are converted using the look-up method to obtain the corresponding vector representation; the information sequence features are converted using the training method to obtain the information sequence feature vector. The specific method is as follows:

[0123] Step 1: The data input layer receives user feature data , Information feature data , context feature data and information sequence feature data ;

[0124] Step 2: Transform the user feature data, information feature data, and context feature data into vectors using the look-up method to obtain the user feature vector , information feature vector and context feature vector ;

[0125] Step 3: Convert the information sequence feature data into vectors through end-to-end training methods or pre-training methods to obtain the information sequence feature vector .

[0126] The main structure layer processes the vectorized information feature vector and information sequence feature vector with attention mechanism. The structure of the attention mechanism is as follows: Figure 5 As shown, the attention mechanism is an important concept in deep learning. It enables the model to focus on the important parts of the input data. When processing sequence data, it imitates the working mode of human attention and is used to effectively cross users and information. It can well learn the complex cross-information between users and information. At the same time, this method ignores irrelevant or less important parts while focusing on the key parts of the input data. In this embodiment, the user is characterized by adopting the attention mechanism, and the information feature vector is weightedly summed to obtain the user portrait feature to obtain the user information representation. The user information representation processed by the attention mechanism is compared with the user feature vector. and context feature vector Perform data splicing to obtain the user splicing vector , please follow the steps below for specific operations:

[0127] The information feature vector and the information sequence feature vector are processed through the attention mechanism to obtain the user information cross representation; the user information cross representation, the user feature vector and the context feature vector are concatenated to obtain the user concatenated vector; at least two groups of fully connected layers and The function optimizes the user splicing vector to obtain the information click feature vector and the information browsing feature vector;

[0128] The attention mechanism is expressed as follows:

[0129]

[0130]

[0131]

[0132] Said Function, expressed as follows:

[0133]

[0134] The user splicing vector is expressed as follows:

[0135]

[0136] in, Indicates user information representation, represents the information feature vector, represents the target information vector, represents the weight parameter, Indicates information feature vector, represents the number of information feature vectors, Indicates The input of the layer network, Indicates the number of network layers, Indicates The parameters of the layer network, A, B, and C represent the cross-representation of user information, the user feature vector, and the context feature vector, respectively. They represent the corresponding weight coefficients respectively and their sum is 1.

[0137] During the user's splicing vector confirmation process, They represent the corresponding weight coefficients respectively. These weight coefficients can be assigned based on experience, for example, which type of features can better represent the user's actual preferences or portraits can account for a larger proportion. Of course, some existing algorithms can also be used to calculate and adjust to obtain the corresponding weight coefficients. The calculation or assignment method of the weight coefficients will not be described here.

[0138] In this embodiment, in order to achieve a better simulation of information clicks and information browsing, network optimization is performed through two three-layer fully connected layers to obtain information click feature vectors and information browsing feature vectors. In the fully connected layer, each layer is connected through a nonlinear function. as an activation function.

[0139] The position debiasing layer selects an independent auxiliary network to process the position-related data. Specifically in this embodiment, the position debiasing layer includes a first input layer, an embedding layer, a fully connected layer and an output layer; the first input layer receives and obtains the position bias feature set, the information click feature vector and the information browsing feature vector; the embedding layer performs end-to-end embedding processing on the position feature set and inputs it to the fully connected layer; the fully connected layer analyzes the position feature set to obtain the position bias vector; the output layer performs weighted operations on the position bias vector with the information click feature vector and the information browsing feature vector to obtain the bias click vector and the bias browsing vector, and inputs the bias click vector and the bias browsing vector into Function is used to convert and obtain the probability of information click and information browsing;

[0140] Said The function is expressed as follows:

[0141]

[0142] in, Indicates the biased click vector or biased browsing vector, Indicates the probability of information click or information browsing.

[0143] The position-related data of information can be understood as position information, and the position bias ID can be obtained based on the position information. The position bias ID is used in the Transformer model to capture the relationship between different positions in the sequence, which can be understood as the relationship between different information positions. The position bias data is obtained from the position-related data. The position bias data can be realized by adding a position-related sine and cosine function encoding (Sinusoidal Positional Encoding). The calculation process of the position bias data can be realized in the following ways:

[0144] Initialize the bias table; build an index system to find the value in the bias table based on the relative position between the information. It can be understood that this index is built by calculating the relative row and column positions between each pair of information; calculate the relative position index: get the relative position index by calculating the difference between the coordinates. For example: if there are two coordinates, the position bias data is obtained. Perform feature extraction and other operations on the position bias data to obtain the position bias feature.

[0145] During the model training process, a loss function is constructed, and the training process is optimized through the loss function. The loss function includes an information click loss function and an information browsing loss function. An error model is constructed, and the accuracy of the information recommendation model is judged based on the error model to obtain a judgment result. When the judgment result is the smallest, it means that the information recommendation model training is completed. In this embodiment, the cross entropy loss function is used as the objective function to be optimized, and the information click loss function is expressed as follows:

[0146]

[0147] The information browsing loss function is expressed as follows:

[0148]

[0149] The error model is expressed as follows:

[0150]

[0151] in, represents the information click loss function, represents the information browsing loss function, Indicates clicking a label. Indicates the probability of information click. represents the probability of information browsing, Indicates the number of sample feature data, Indicates browsing tags. Indicates the actual value or actual label, which can be understood as the actual click situation. represents the predicted value, Indicates the number of times, Indicates Second-rate, Indicates the error result.

[0152] Based on the loss function, sample label data, location-related data and sample feature data, the information recommendation pre-training model is trained and optimized. During the training process, the training samples are divided into batches for batch training, and the entire training process is a continuous iterative process. In each iteration, the forward propagation algorithm calculates the information prediction value of the current batch of training data based on the value of the current parameter, and then the back propagation algorithm calculates the gradient of the parameter based on the loss function and updates the parameter. In this embodiment, AdamOptimizer is used as the optimizer for the optimization of the target network. The optimizer can adaptively adjust the learning rate and can effectively adjust the pace of parameter update during the training process. It is also suitable for processing large data sets and high-dimensional data spaces. The prediction value and loss function obtained by forward propagation are calculated, and then the AdamOptimizer optimizer is used for back propagation to achieve model parameter update, thereby continuously iterating and optimizing the performance of the information recommendation pre-training model. The information recommendation pre-training model is continuously iterated and optimized until the expected convergence state or preset convergence condition is reached to obtain the information recommendation model. The information recommendation model obtained by this training strategy can maintain high efficiency and accuracy in a complex data environment.

[0153] Get the data to be recommended and load the trained information recommendation model file, and obtain the click preference score and browsing preference score through inference. The inference process is as follows: Figure 6 As shown, during the inference process, the position bias layer is not calculated, so it will not affect the final inference result. Taking into account the difference between the daily click-through rate and the effective browsing rate, in order to avoid too large a difference when weighting the click preference score and the browsing preference score, the weight is set to the ratio of the average click rate of the information to the average browsing rate of the information, so as to achieve the purpose of smoothing the difference between the two predicted scores, so that the fused preference score is more accurate and closer to the user's actual conversion rate. Based on the obtained weight, the click preference score and the browsing preference score are weighted to obtain the fused preference score, and the fused preference score is reversed. This embodiment sets a recommendation threshold, and the information information corresponding to the recommended data that meets the recommendation threshold is the information recommendation result. Among them, the fused preference score is expressed as follows:

[0154]

[0155]

[0156]

[0157] in, , represents the fusion preference score, represents the click preference score, represents the browsing preference score, represents the weight, Indicates the average click-through rate of information. Indicates the average browsing rate of information. Indicates the number of information clicks within a fixed period of time. Indicates the amount of information exposure within a fixed period of time. Indicates the number of effective views of information within a fixed period of time.

[0158] Embodiment 2:

[0159] A multi-objective information recommendation system based on position debiasing, such as Figure 2 As shown, it includes an acquisition processing module 100, a model building module 200, a model training module 300 and a model reasoning module 400;

[0160] The acquisition processing module 100 acquires the log data of the user browsing information and performs preprocessing and analysis to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set; performs feature extraction on the user-related data set to obtain a user feature set, and performs feature extraction on the information data set to obtain an information feature set, wherein the information feature set includes information features, context features and information sequence features; obtains location bias data based on location-related data and then performs feature extraction to obtain a location bias feature set;

[0161] The model construction module 200 constructs an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability.

[0162] The model training module 300 trains the information recommendation pre-training model based on the user-related data set, the location-related data set and the information-related data set to obtain the information recommendation model;

[0163] The model inference module 400 infers the recommended information data based on the information recommendation model to obtain the user's click preference score and browsing preference score for the information, and obtains the information recommendation result through the click preference score and browsing preference score.

[0164] Various changes and modifications can be made without departing from the spirit and scope of the present invention, and all equivalent technical solutions also belong to the scope of the present invention.

[0165] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0166] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0167] The present invention is described with reference to the flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0168] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0169] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0170] It should be noted that:

[0171] The "one embodiment" or "embodiment" mentioned in the specification means that the specific features, structures or characteristics described in conjunction with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrases "one embodiment" or "embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment.

[0172] In addition, it should be noted that the shapes and names of the parts and components of the specific embodiments described in this specification may be different. Any equivalent or simple changes made based on the structure, features and principles described in the patent concept of the present invention are included in the protection scope of the patent of the present invention. The technicians in the technical field of the present invention can make various modifications or supplements to the specific embodiments described or replace them in a similar manner, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A multi-objective information recommendation method based on position debiasing, characterized in that: The following steps are involved: Obtaining the log data of user browsing information and preprocessing and parsing it to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set; Performing feature extraction on a user-related data set to obtain a user feature set, and performing feature extraction on an information data set to obtain an information feature set, wherein the information feature set includes information features, context features, and information sequence features; obtaining position bias data based on position-related data and then performing feature extraction to obtain a position bias feature set; Construct an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability. The information feature vector and information sequence feature vector are processed through the attention mechanism to obtain the user information cross representation; The user information cross representation, the user feature vector and the context feature vector are concatenated to obtain a user concatenated vector; the user concatenated vector is optimized by at least two sets of fully connected layers and functions to obtain an information click feature vector and an information browsing feature vector; The attention mechanism is expressed as follows: Said Function, expressed as follows: The user splicing vector is expressed as follows: in, Indicates user information representation, represents the information feature vector, represents the target information vector, represents the weight parameter, Indicates information feature vector, represents the number of information feature vectors, Indicates The input of the layer network, Indicates the number of network layers, Indicates The parameters of the layer network, A, B, C represent the cross representation of user information, user feature vector and context feature vector respectively. They represent the corresponding weight coefficients respectively and the sum is 1; The position debiasing layer includes a first input layer, an embedding layer, a fully connected layer and an output layer; The first input layer receives and acquires the position offset feature set, the information click feature vector, and the information browsing feature vector; The embedding layer performs end-to-end embedding processing on the position bias feature set and inputs it into the fully connected layer; The fully connected layer analyzes the position offset feature set to obtain the position offset vector; The output layer performs weighted operations on the position bias vector, the information click feature vector and the information browse feature vector to obtain a biased click vector and a biased browse vector, and inputs the biased click vector and the biased browse vector into Function is used to convert and obtain the probability of information click and information browsing; Said The function is expressed as follows: in, Indicates the biased click vector or biased browse vector, Indicates the probability of information click or information browsing; The information recommendation pre-training model is trained based on the user-related data set, the location-related data set and the information-related data set to obtain an information recommendation model; Based on the information recommendation model, the recommended information data is inferred to obtain the user's click preference score and browsing preference score for the information, and the information recommendation results are obtained through the click preference score and browsing preference score.

2. The multi-objective information recommendation method based on position debiasing according to claim 1, characterized in that: The information tag set includes information product tags, click behavior tags and effective browsing tags; The user characteristics include demographic characteristics, user device characteristics, user activity characteristics and user information preference characteristics; The information features include information attribute features and information tag features, and the information tag features include information click tags, information browsing tags and effective browsing tags; The context features include the currently recommended information topic and time period; The information sequence features include the information sequence clicked by the user and the information sequence browsed by the user.

3. The multi-objective information recommendation method based on position debiasing according to claim 1, characterized in that: The effective browsing tag is obtained by following the steps below: Obtain the word length of the current information content, and calculate based on the word length and the average word length of browsing per second to obtain the information browsing time; Get the user's browsing time, and make a judgment based on the information browsing time and the user's browsing time. When the user's browsing time is longer than the preset information browsing time, it is a valid browsing tag, which is expressed as follows: in, Indicates the user's browsing time. Indicates the length of words. Indicates the average length of words browsed per second. Represents a constant, Indicates a valid browsing tag.

4. The multi-objective information recommendation method based on position debiasing according to claim 1, characterized in that: The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations, including the following steps: The user features, information features and context features are converted using the look-up method to obtain the corresponding vector representations; The information sequence features are transformed using a training method to obtain an information sequence feature vector.

5. The multi-objective information recommendation method based on position debiasing according to claim 1, characterized in that: The following steps are also included: Constructing a loss function to optimize the training process through the loss function, wherein the loss function includes an information click loss function and an information browsing loss function; Construct an error model, and judge the accuracy of the information recommendation model based on the error model to obtain a judgment result. When the judgment result is the smallest, it means that the information recommendation model training is completed; The information click loss function is expressed as follows: The information browsing loss function is expressed as follows: The error model is expressed as follows: in, represents the information click loss function, represents the information browsing loss function, Indicates clicking a label. Indicates the probability of information click. represents the probability of information browsing, Indicates the number of sample feature data, Indicates browsing tags. Indicates the actual value, represents the predicted value, Indicates the number of times, Indicates Second-rate, Indicates the error result.

6. The multi-objective information recommendation method based on position debiasing according to claim 1, characterized in that: The method of obtaining information recommendation results by clicking preference scores and browsing preference scores includes the following steps: The click preference score and the browsing preference score are weighted to obtain a fusion preference score; Sort the fusion preference scores, analyze based on the sorting results, and obtain information recommendation results; Wherein, the fusion preference score is expressed as follows: in, represents the fusion preference score, represents the click preference score, represents the browsing preference score, represents the weight, , Indicates the average click-through rate of information. Indicates the average browsing rate of information. Indicates the number of information clicks within a fixed period of time. Indicates the amount of information exposure within a fixed period of time. Indicates the number of effective views of information within a fixed period of time.

7. The multi-objective information recommendation method based on position debiasing according to claim 1, characterized in that: The method of obtaining the log data of user browsing information and preprocessing and parsing it to obtain a user-related data set, a location-related data set and an information data set includes the following steps: Collect the log data of users browsing information, judge and eliminate abnormal data, and obtain the pre-processed log data set; Extract and analyze the preprocessed log data set based on at least one or more of user behavior analysis, content preference analysis, and time series analysis to obtain a user-related data set, a location-related data set, and an information data set; The anomaly score is calculated by the following formula, and the anomaly score is used to determine whether it is abnormal and remove abnormal data, as shown below: in, represents the anomaly score of the log data, Indicates the amount of data, represents the first quartile of the data, Represents the third quartile of the data. represents the interquartile range, is a constant, if Not less than 0 The value is 1 if yes, otherwise it is 0.

8. A multi-objective information recommendation system based on position debiasing, characterized in that: It includes acquisition processing module, model building module, model training module and model reasoning module; The acquisition processing module acquires the log data of the user browsing information and performs preprocessing and analysis to obtain a user-related data set, a location-related data set and an information data set, wherein the information data set includes an information tag set and an information-related data set; performs feature extraction on the user-related data set to obtain a user feature set, and performs feature extraction on the information data set to obtain an information feature set, wherein the information feature set includes information features, context features and information sequence features; obtains position bias data based on the position-related data and then performs feature extraction to obtain a position bias feature set; The model construction module constructs an information recommendation pre-training model, which includes a data input layer, a shared feature layer, a main structure layer and a position debiasing layer. The data input layer receives a user feature set, an information feature set, an information sequence feature set and a context feature set. The shared feature layer processes the user feature set, the information feature set, the information sequence feature set and the context feature set to obtain corresponding vector representations. The main structure layer processes the vector representations to obtain an information click feature vector and an information browsing feature vector. The position debiasing layer obtains a position bias feature set, an information click feature vector and an information browsing feature vector and processes them to obtain an information click probability and an information browsing probability. The information feature vector and information sequence feature vector are processed through the attention mechanism to obtain the user information cross representation; The user information cross representation, the user feature vector and the context feature vector are concatenated to obtain a user concatenated vector; the user concatenated vector is optimized by at least two sets of fully connected layers and functions to obtain an information click feature vector and an information browsing feature vector; The attention mechanism is expressed as follows: Said Function, expressed as follows: The user splicing vector is expressed as follows: in, Indicates user information representation, represents the information feature vector, represents the target information vector, represents the weight parameter, Indicates information feature vector, represents the number of information feature vectors, Indicates The input of the layer network, Indicates the number of network layers, Indicates The parameters of the layer network, A, B, C represent the cross representation of user information, user feature vector and context feature vector respectively. They represent the corresponding weight coefficients respectively and the sum is 1; The position debiasing layer includes a first input layer, an embedding layer, a fully connected layer and an output layer; The first input layer receives and acquires the position offset feature set, the information click feature vector, and the information browsing feature vector; The embedding layer performs end-to-end embedding processing on the position bias feature set and inputs it into the fully connected layer; The fully connected layer analyzes the position offset feature set to obtain the position offset vector; The output layer performs weighted operations on the position bias vector, the information click feature vector and the information browse feature vector to obtain a biased click vector and a biased browse vector, and inputs the biased click vector and the biased browse vector into Function is used to convert and obtain the probability of information click and information browsing; Said The function is expressed as follows: in, Indicates the biased click vector or biased browse vector, Indicates the probability of information click or information browsing; The information recommendation pre-training model is trained based on the user-related data set, the location-related data set and the information-related data set to obtain an information recommendation model; The model training module trains the information recommendation pre-training model based on the user-related data set, the location-related data set and the information-related data set to obtain the information recommendation model; The model reasoning module reasons on the recommended information data based on the information recommendation model to obtain the user's click preference score and browsing preference score for the information, and obtains the information recommendation result through the click preference score and browsing preference score.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A multi-objective information recommendation device based on position debiasing, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images