Object Recommendation Method and Apparatus Based on Recommendation Model

Abstract

This invention provides an object recommendation method and apparatus based on a recommendation model. The method includes: extracting features from user data of the user to be recommended and object data of the object to be recommended using a feature extraction model to obtain corresponding features of the user to be recommended and features of the object to be recommended; extracting deep features from the combined features of the user to be recommended and the object to be recommended using a deep feature model to obtain corresponding deep features; mining frequent sequence patterns from the user to be recommended using a pattern mining model to obtain weighted features of frequent sequence patterns of the user to be recommended; fusing the deep features and weighted features using a prediction model to obtain fused features of the user to be recommended, and predicting the conversion rate based on the fused features to obtain a prediction result; and recommending the object to the user to be recommended when the prediction result indicates that the conversion rate value of the user to be recommended exceeds a probability threshold.

Description

Technical Field

[0001] This invention relates to natural language processing technology in artificial intelligence, and more particularly to an object recommendation method and apparatus based on a recommendation model. Background Technology

[0002] Natural Language Processing (NLP) is an important area of ​​artificial intelligence that enables effective communication between humans and computers using natural language. It typically includes technologies such as text processing, semantic understanding, and machine translation. Recommendation systems are a key application of NLP, automatically connecting users with recommended content. In information-overloaded environments, they help users discover information of interest and push information to other users who are also interested in it.

[0003] With the rise of online consumption and the rapid development of e-commerce platforms, a more accurate understanding of users' personalized needs has made personalized recommendations a key competitive advantage for these platforms. Therefore, establishing personalized recommendation models has become a crucial strategy for businesses to understand user consumption needs, seize market opportunities, and increase transaction volume. Furthermore, accurately identifying user behavior habits and providing personalized items based on user characteristics is also significant for applications such as benefit recommendations and advertising.

[0004] However, there is still no effective solution in terms of how to perform object recommendation efficiently and accurately. Summary of the Invention

[0005] This invention provides an object recommendation method and apparatus based on a recommendation model, which can improve the accuracy of object recommendations.

[0006] The technical solution of this invention is implemented as follows:

[0007] This invention provides an object recommendation method based on a recommendation model, wherein the recommendation model includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model, and the method includes:

[0008] Using the feature extraction model, feature extraction is performed on the user data of the user to be recommended and the object data of the object to be recommended, respectively, to obtain the corresponding features of the user to be recommended and the features of the object to be recommended.

[0009] Using the deep feature model, deep feature extraction is performed on the combined features of the user features to be recommended and the object features to be recommended to obtain the corresponding deep features;

[0010] The pattern mining model is used to mine frequent sequence patterns of the user to be recommended, and weighted features of frequent sequence patterns of the user to be recommended are obtained.

[0011] The prediction model fuses the deep features and the weighted features to obtain the fused features corresponding to the user to be recommended, and the conversion rate is predicted based on the fused features to obtain the prediction result.

[0012] When the prediction result indicates that the conversion rate of the user to be recommended exceeds the probability threshold, the object to be recommended is recommended to the user to be recommended.

[0013] This invention provides a recommendation device based on a recommendation model, wherein the recommendation model includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model; and the device includes:

[0014] The first feature extraction module is used to extract features from the user data of the user to be recommended and the object data of the object to be recommended through the feature extraction model, so as to obtain the corresponding features of the user to be recommended and the features of the object to be recommended.

[0015] The second feature extraction module is used to perform deep feature extraction on the combined features of the user features to be recommended and the object features to be recommended through the deep feature model to obtain the corresponding deep features;

[0016] The feature determination module is used to perform frequent sequence pattern mining on the features of the user to be recommended through the pattern mining model, and obtain weighted features of the frequent sequence patterns corresponding to the user to be recommended.

[0017] The feature prediction module is used to fuse the deep features and the weighted features through the prediction model to obtain the fused features corresponding to the user to be recommended, and to predict the conversion rate based on the fused features to obtain the prediction result;

[0018] The object recommendation module is used to recommend the object to the user to be recommended when the prediction result indicates that the conversion rate value of the user to be recommended exceeds a probability threshold.

[0019] In the above scheme, before performing feature extraction on the user data of the user to be recommended and the object data of the object to be recommended through the feature extraction model, the device further includes a recommendation model training module.

[0020] The recommendation model training module is used to extract features from the recommended object samples labeled with conversion rates through the feature extraction model, so as to obtain the user features and recommended object features corresponding to the recommended object samples.

[0021] Using the deep feature model, deep feature extraction is performed on the combined features of the user features and the recommended object features to obtain the corresponding deep features;

[0022] The pattern mining model is used to mine frequent sequence patterns of the user features to obtain weighted features of frequent sequence patterns corresponding to the recommended object samples.

[0023] The prediction model fuses the deep features and the weighted features to obtain the fused features corresponding to the recommended object sample, and the conversion rate is predicted based on the fused features to obtain the prediction result.

[0024] Obtain the difference between the predicted result and the labeled conversion rate, and update the model parameters of the recommendation model based on the difference.

[0025] In the above scheme, the feature extraction model includes a feature extraction layer and a feature concatenation layer for extracting features from the user data of the user to be recommended and the object data of the object to be recommended. Before performing deep feature extraction on the combined features of the user features and the object features through the deep feature model,

[0026] The first feature extraction module is further configured to concatenate the features of the user to be recommended and the features of the object to be recommended through the feature concatenation layer to obtain combined features corresponding to the user to be recommended and the object to be recommended.

[0027] In the above scheme, before performing deep feature extraction on the combined features of the user features and the object features to be recommended using the deep feature model, the device further includes a deep feature training module.

[0028] The deep feature training module is used to obtain combined feature samples labeled with target deep features. The combined feature samples include positive samples corresponding to converted users and negative samples of unconverted users.

[0029] Feature extraction is performed on the combined feature samples to obtain the user features and recommendation object features corresponding to the combined feature samples;

[0030] The combined features of the user features and the recommended object features are input into the deep feature model, and deep feature extraction is performed on the combined features of the combined feature samples to obtain the corresponding predicted deep features;

[0031] The difference between the target depth features and the predicted depth features is obtained, and the model parameters of the depth feature model are updated based on the difference.

[0032] In the above scheme, before performing deep feature extraction on the combined features of the user features and the object features to be recommended, the device further includes a preprocessing module.

[0033] The preprocessing module is configured to perform at least one of the following preprocessing operations on the combined features, so as to determine the combined features after the preprocessing operations as the combined features of the user features and the recommendation object features:

[0034] Determine the number of missing feature values ​​of the combined feature and the number of identical feature values ​​corresponding to the combined feature, and filter the combined features whose number of missing values ​​exceeds the missing value threshold or whose number exceeds the identical value threshold to obtain the filtered combined features;

[0035] Anomaly detection is performed on the feature values ​​of the combined features. When the detection result indicates that the feature value is an anomaly, the feature value is replaced with a filler feature value to obtain a combined feature containing the filler feature value.

[0036] The combined features are then subjected to derivation processing to obtain the derived combined features;

[0037] Discretize the feature values ​​corresponding to the continuous features in the combined features to obtain discrete features, and use the discrete features obtained after discretization and the discrete features in the combined features as the combined features corresponding to the combined recommendation object sample.

[0038] In the above scheme, the feature determination module is further used to perform frequent sequence pattern mining on the features of the user to be recommended, so as to obtain multiple frequent sequence patterns corresponding to the user to be recommended;

[0039] The frequency proportion of each of the frequent sequence patterns in the behavioral sequence pattern set corresponding to the user to be recommended is determined as the weight of the corresponding frequent sequence pattern. The behavioral sequence pattern set includes at least two behavioral sequence patterns.

[0040] Based on the weights, the behavioral sequence encoding features of each frequent sequence pattern are weighted to obtain the weighted features of the frequent sequence patterns of the user to be recommended.

[0041] In the above scheme, the feature determination module is further configured to determine a set of behavioral sequence patterns corresponding to the user to be recommended based on the features of the user to be recommended. The set of behavioral sequence patterns includes at least two behavioral sequence patterns, and the behavioral sequence patterns are obtained by combining multiple behavioral encoding features corresponding to the sequence samples.

[0042] In the behavior sequence pattern, behavior encoding features are selected multiple times, and the selected behavior encoding features are combined to obtain multiple different behavior sequence sub-patterns corresponding to the behavior sequence pattern.

[0043] Each selected behavioral encoding feature includes behavioral encoding features belonging to different behavioral training patterns of users to be recommended.

[0044] From the multiple different behavior sequence sub-patterns, select the behavior sequence sub-patterns whose frequency of occurrence in the behavior sequence pattern set exceeds a frequency threshold, and determine the selected behavior sequence sub-patterns as frequent sequence patterns.

[0045] In the above scheme, the feature determination module is further used to select a behavior sequence pattern prefix with a unit length of one, and determine the projection dataset corresponding to the behavior sequence pattern prefix;

[0046] When the projection dataset corresponding to a behavior sequence pattern prefix is ​​not empty, a behavior sequence pattern prefix whose frequency exceeds the frequency threshold is determined as a frequent feature encoding sequence;

[0047] Based on the aforementioned frequent behavior sequence pattern, the unit length of the selected behavior sequence pattern prefix is ​​increased sequentially to obtain n behavior sequence pattern prefixes.

[0048] Determine the projection dataset corresponding to the prefixes of the n behavioral sequence patterns;

[0049] When the projection dataset corresponding to the n behavioral sequence pattern prefixes is empty and the occurrence frequency of the n behavioral sequence pattern prefixes exceeds the frequency threshold, the n behavioral sequence pattern prefixes are determined as the frequent sequence patterns.

[0050] Where n is a natural number whose value gradually increases, and the range of values ​​satisfies 2≤n≤K, and K is the number of behavioral encoding features contained in the longest behavioral sequence sub-pattern in the sequence set.

[0051] In the above scheme, the feature determination module is also used to perform frequent sequence pattern mining on the features of the user to be recommended, and obtain the frequent sequence pattern corresponding to the prefix with the longest mining length.

[0052] The frequency proportion of the frequent sequence pattern in the set of behavioral sequence patterns corresponding to the user to be recommended is determined as the weight of the frequent sequence pattern. The set of behavioral sequence patterns includes at least two behavioral sequence patterns.

[0053] Based on the weights, the behavioral sequence encoding features of the frequent sequence patterns are weighted to determine the weighted features of the frequent sequence patterns corresponding to the users to be recommended.

[0054] This invention provides an electronic device, comprising:

[0055] Memory, used to store executable instructions;

[0056] A processor, when executing executable instructions stored in the memory, implements the method provided in the embodiments of the present invention.

[0057] This invention provides a computer-readable storage medium storing executable instructions for inducing a processor to execute the method provided in this invention.

[0058] The embodiments of the present invention have the following beneficial effects:

[0059] By employing deep learning to study user and recommended object features using a deep feature model, we obtain deep features that are more effective in predicting user conversion rates. Furthermore, by performing frequent sequence mining on user features, we obtain weighted features of frequent sequence patterns, which can uncover common features in the behavioral sequence patterns of user groups, thus addressing the problem of relatively coarse feature construction in related technologies. Combining deep and weighted features to predict user conversion rates improves the accuracy of user conversion rate prediction, thereby enhancing the accuracy of object recommendations based on the predicted user conversion probabilities. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of an optional architecture for an object recommendation system based on a recommendation model, provided in an embodiment of the present invention.

[0061] Figure 2 An optional structural schematic diagram of an electronic device provided in an embodiment of the present invention;

[0062] Figure 3 This is an optional flowchart illustrating an object recommendation method based on a recommendation model provided in an embodiment of the present invention.

[0063] Figure 4 A data flow diagram of the recommendation model-based recommendation method provided in this embodiment of the invention;

[0064] Figure 5 This is an optional flowchart illustrating a method for determining weighted features provided in an embodiment of the present invention;

[0065] Figure 6 This is an optional flowchart illustrating a method for determining weighted features provided in an embodiment of the present invention;

[0066] Figure 7This is an optional flowchart illustrating an object recommendation method based on a recommendation model provided in an embodiment of the present invention.

[0067] Figure 8 This is a schematic diagram of an object recommendation method provided in an embodiment of the present invention;

[0068] Figure 9 This is an optional structural diagram of an object recommendation device based on a recommendation model provided in an embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0071] In the following description, the terms “first, second, third…” are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that “first, second, third…” may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0072] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0073] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0074] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be interpreted as follows.

[0075] 1) Recommendation system: A personalized information recommendation system that recommends information and products that users are interested in to users based on their information needs and interests.

[0076] 2) Recommended items, also known as items, such as news, posts, books, shopping items, advertisements, etc.

[0077] 3) Conversion rate, which indicates the ratio of the number of conversions to the total number of clicks within a statistical period. For example, Taobao conversion rate is the ratio of all people who visit a Taobao store and make a purchase to all people who visit a Taobao store. Other examples include registered user conversion rate, which is the ratio of users attracted by news, advertisements, etc. to registered users; logged-in user conversion rate, which is the ratio of registered users to logged-in users; and paid user conversion rate, which is the ratio of registered and logged-in users who successfully pay their first fee.

[0078] 4) Sequence pattern mining refers to a set of different sequences, where each sequence consists of different elements arranged in order, and each element consists of different items. Given a minimum support threshold, sequence pattern mining aims to find all frequent sequence patterns, that is, the frequency of the subsequence in the sequence set is not lower than the minimum support threshold.

[0079] 5) Parameters of the neural network model, which are automatically updated or learned during the training process of the neural network model, including feature weights, biases, etc.

[0080] Personalized recommendation systems for business scenarios have wide applications in e-commerce platforms, advertising, and other fields. They are particularly valuable for user profiling, personalized recommendations, and targeted advertising on e-commerce platforms. Among related technologies, the mainstream methods for building recommendation systems for business scenarios include: demographic-based recommendation, content-based recommendation, and recommendation based on coordinated filtering algorithms. These specific methods will be introduced below.

[0081] 1) Recommendations based on manual statistics: This method identifies the relevance of users based on their basic information and then recommends other items liked by similar users (i.e., items to be recommended) to the current user. This method is typically used in the cold start phase of recommendation systems and does not require historical data to model user attributes.

[0082] 2) Content-based recommendation: Modeling is performed based on item attributes, and recommendations are made using the similarity of the items themselves. The accuracy and effectiveness of recommendations are improved by increasing the dimensions of item attributes.

[0083] 3) Based on coordinated filtering recommendation, coordinated filtering is divided into user-based coordinated filtering and item-based coordinated filtering. In actual implementation, coordinated filtering can be accomplished through association algorithms, clustering algorithms, classification algorithms, regression algorithms, matrix factorization, graph models, etc.

[0084] While user-based collaborative filtering also considers user similarity, it differs from demographic-based recommendation algorithms in how it calculates similarity. Demographic-based recommendations calculate similarity based on basic user information, while user-based collaborative filtering primarily relies on users' historical preference data. Similarly, item-based collaborative filtering and content-based recommendations, both based on item similarity, differ in that content-based recommendations rely on the similarity of item attributes, while item-based collaborative filtering relies on the similarity between items derived from users' historical preferences.

[0085] The above method has the following technical problems in its application:

[0086] 1) Recommendations based on manual statistics are based only on the basic user profile features. The constructed features are relatively coarse and cannot reflect the commonalities of the user group. The effect is difficult to meet the needs and is only suitable for simple recommendations. At the same time, many large-scale system products do not have the cold start problem. This algorithm is significantly different from other algorithms.

[0087] 2) Content-based recommendations only consider the similarity between the attributes of the items themselves. However, the attributes are limited, and the user's historical item data is required. It is difficult to generalize to new users, meaning that the generalization ability of content-based recommendation models is weak.

[0088] 3) Recommendations based on coordination filtering rely on users' historical preferences, regardless of the item's content attributes. New items in the cold start phase lack this historical preference information, making recommendations impossible. This results in low accuracy and generalization ability for coordination filtering-based recommendation models. Furthermore, matrix factorization requires a large number of real ratings; otherwise, the input to the factorization model becomes very sparse, hindering accurate predictions and further reducing recommendation accuracy.

[0089] In view of this, embodiments of the present invention provide an object recommendation method and apparatus based on a recommendation model to at least solve the above-mentioned problems in the related art, which will be described in detail below.

[0090] See Figure 1 , Figure 1 This is an optional architecture diagram of an object recommendation system 100 based on a recommendation model provided in an embodiment of the present invention. To support an exemplary application, a terminal 400 (terminals 400-1 and 400-2 are shown as examples) connects to a server 200 through a network 300. The network 300 can be a wide area network or a local area network, or a combination of both, and data transmission is achieved using a wireless link.

[0091] In practical applications, terminal 400 can be various types of user terminals such as smartphones, tablets, and laptops, or any combination of two or more of these data processing devices such as desktop computers, game consoles, televisions, or these data processing devices; server 200 can be a single server configured to support various services, or it can be configured as a server cluster, or it can be a cloud server, etc.

[0092] In practical applications, the terminal is equipped with a client, such as a news client or a Taobao client. For example, terminal 400-1 is located on the publishing side of the object to be recommended, and terminal 400-2 is located on the user side of the object to be recommended (i.e., the receiving side of the object to be recommended). When a user opens the client of terminal 400-1 to publish an object to be recommended, for example, a Taobao merchant publishes a product or an advertiser places an advertisement, terminal 400-1 is used to generate and send a recommendation request carrying the object data of the object to be recommended to server 200.

[0093] Server 200 is used to acquire object data of the object to be recommended and user data of the user to be recommended. Using a pre-trained feature extraction model within the recommendation model, it extracts features from the user data of the user to be recommended and the object data of the object to be recommended, respectively, to obtain the corresponding features of the user to be recommended and the object to be recommended. Using a pre-trained deep feature model within the recommendation model, it extracts deep features from the combined features of the user to be recommended and the object to be recommended, to obtain the corresponding deep features. Using a pre-trained pattern mining model within the recommendation model, it mines frequent sequence patterns from the user to be recommended, to obtain the weighted features of the frequent sequence patterns of the corresponding user to be recommended. Using a pre-trained prediction model within the recommendation model, it fuses the deep features and weighted features to obtain the fused features of the corresponding user to be recommended, and predicts the conversion rate based on the fused features, to obtain the prediction result. When the prediction result indicates that the conversion rate value of the user to be recommended exceeds a probability threshold, the object to be recommended is recommended to terminal 400-2, thus recommending the object to the user to be recommended.

[0094] The object recommendation method based on a recommendation model provided in this invention can be widely applied in e-commerce platforms, recommendation systems, and advertising. For example, on e-commerce platforms like Taobao, by mining frequent sequence patterns of user behavior based on user habits, and weighting and clustering high-discrimination features, weighted features of frequent sequence patterns are obtained. This allows for personalized recommendations of preferred items to users, which can help increase transaction volume on the platform. In advertising scenarios, targeted advertising can be performed on users with similar behavioral preferences for related items, maximizing the revenue-cost ratio. In addition, scenarios related to the construction of personalized recommendation systems are also potential application scenarios of this invention.

[0095] The electronic device for implementing an object recommendation method based on a recommendation model, as provided in the embodiments of the present invention, will be described below. See [link to relevant documentation]. Figure 2 , Figure 2 This is an optional structural diagram of the electronic device 500 provided in an embodiment of the present invention. In practical applications, the electronic device 500 can be... Figure 1 The terminal or server in the middle, hereinafter referred to as electronic device Figure 1 Taking the server 200 shown as an example, the electronic device implementing the embodiments of the present invention will be described. Figure 2 The illustrated electronic device 500 includes at least one processor 510, a memory 550, at least one network interface 520, and a user interface 530. The various components in the service data transmission device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to implement communication between these components. In addition to a data bus, the bus system 540 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 2 The general labeled all buses as Bus System 540.

[0096] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0097] User interface 530 includes one or more output devices 531 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 530 also includes one or more input devices 532, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0098] The memory 550 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 550 may optionally include one or more storage devices physically located away from the processor 510.

[0099] The memory 550 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 550 described in this embodiment is intended to include any suitable type of memory.

[0100] In some embodiments, memory 550 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0101] Operating system 551 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0102] The network communication module 552 is used to reach other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.

[0103] Presentation module 553 is configured to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 531 (e.g., a display screen, a speaker, etc.) associated with user interface 530;

[0104] The input processing module 554 is used to detect and translate one or more user inputs or interactions from one or more input devices 532.

[0105] In some embodiments, the object recommendation device based on the recommendation model provided in this invention can be implemented in software. Figure 2 An object recommendation device 555 based on a recommendation model, stored in memory 550, is shown. It can be software in the form of programs and plug-ins, including the following software modules: a first feature extraction module 5551, a second feature extraction module 5552, a feature determination module 5553, a feature prediction module 5554, and an object recommendation module 5555. These modules are logically related and can therefore be arbitrarily combined or further split according to the functions they implement.

[0106] The functions of each module will be explained below.

[0107] In other embodiments, the object recommendation device based on the recommendation model provided in this invention can be implemented in hardware. As an example, the object recommendation device based on the recommendation model provided in this invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the object recommendation method based on the recommendation model provided in this invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0108] The object recommendation method based on the recommendation model of the present invention will be described below. In actual implementation, the object recommendation method based on the recommendation model provided in the present invention can be implemented by the server or the terminal alone, or by the server and the terminal working together. The following description takes the implementation by the server as an example.

[0109] See Figure 3-4 , Figure 3 This is an optional flowchart illustrating an object recommendation method based on a recommendation model provided in an embodiment of the present invention. Figure 4 The data flow diagram of the recommendation model-based recommendation method provided in the embodiments of the present invention is as follows: Figure 4 As shown, the recommendation model provided in this embodiment of the invention includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model, which will combine... Figure 3 and Figure 4 The steps shown are explained.

[0110] Step 101: The server uses a feature extraction model to extract features from the user data of the user to be recommended and the object data of the object to be recommended, respectively, to obtain the user features corresponding to the user to be recommended and the object features corresponding to the object to be recommended.

[0111] In practical applications, the terminal is equipped with a client, such as a news client or a Taobao client. When a user on the publishing side of the object to be recommended opens the client and publishes an object to be recommended, for example, a Taobao merchant publishes a product or an advertiser places an advertisement, the terminal generates and sends a recommendation request carrying the object data of the object to be recommended to the server.

[0112] The server receives object data of the objects to be recommended and user data of the users to be recommended. Based on the feature extraction model in the trained recommendation model, it extracts features from the object data and user data respectively to obtain corresponding object features and user features. Then, the object features and user features are concatenated to obtain the corresponding combined features.

[0113] In practical applications, before performing feature extraction on the user data of the user to be recommended and the object data of the object to be recommended using the feature extraction model, the server can train the recommendation model in the following way:

[0114] The system employs a feature extraction model to extract features from recommended object samples labeled with conversion rates, obtaining the corresponding user features and recommended object features. A deep feature model is then used to extract deep features from the combined user and recommended object features, yielding the corresponding deep features. A pattern mining model is used to mine frequent sequence patterns from the user features, obtaining the weighted features of frequent sequence patterns for the corresponding recommended object samples. A prediction model is then used to fuse the deep features and weighted features, obtaining the fused features for the corresponding recommended object samples. Conversion rates are then predicted based on these fused features, yielding the prediction results. The difference between the prediction results and the labeled conversion rates is then obtained, and the model parameters of the recommendation model are updated based on this difference.

[0115] Here, the user characteristics corresponding to the recommended object samples include basic user attributes such as age, gender, education level, and city tier; user consumption characteristics such as total number of payments, total amount, distribution of payment number within a certain time period (within 24 hours, within a week, within a month, within six months), distribution of payment amount, and average amount per payment; and user behavior characteristics such as browsing time and number of page clicks. The recommended object characteristics include basic attribute features such as recommended object category, recommended object price, recommended object brand, recommended object purchase rating, and recommended object review sentiment; and recommended object consumption characteristics such as the number of times the recommended object has been purchased, the number of times it has been clicked / viewed, the number of times it has been added to the shopping cart, and the number of times similar recommended objects have been purchased.

[0116] In practice, user data such as basic user attributes, user consumption data, and user behavior data, as well as recommendation object data such as basic attribute data and consumption data, are extracted from the user data and recommendation object data of the acquired recommendation object sample. Then, feature extraction models are used to extract features from the extracted user data and recommendation object data to obtain the corresponding user features and recommendation object features.

[0117] In some embodiments, the feature extraction model includes a feature extraction layer and a feature concatenation layer for extracting features from the recommendation object samples. The server can use the feature concatenation layer to concatenate user features and recommendation object features to obtain combined features for the corresponding recommendation object samples. Thus, combined features of <user, recommendation object> are constructed by concatenating and combining user features and recommendation object features.

[0118] In practice, the value of the loss function of the recommendation model can be determined based on the difference between the prediction results and the labeled conversion rate. When the value of the loss function reaches a preset threshold, the corresponding error signal is determined based on the value of the loss function of the recommendation model. The error signal is then backpropagated in the recommendation model, and the model parameters of each layer of the recommendation model are updated during the propagation process.

[0119] Here, we will explain backpropagation. The sample data of the recommended objects is input into the input layer of the neural network model, passes through the hidden layer, and finally reaches the output layer to output the result. This is the forward propagation process of the neural network model. Since there is an error between the output result of the neural network model and the actual result, the error between the output result and the actual value is calculated and backpropagated from the output layer to the hidden layer until it reaches the input layer. During the backpropagation process, the values ​​of the model parameters are adjusted according to the error. The above process is iterated until convergence.

[0120] Step 102: Using a deep feature model, perform deep feature extraction on the combined features of the user features to be recommended and the object features to be recommended to obtain the corresponding deep features.

[0121] In some embodiments, before performing deep feature extraction on the combined features of the user features and the recommended object features using a deep feature model, or before performing deep feature extraction on the combined features of the user features and the recommended object features using a deep feature model during the training of the recommendation model, the server can train the deep feature model in the following manner:

[0122] Obtain combined feature samples labeled with target depth features, where the combined feature samples include positive samples corresponding to converted users and negative samples corresponding to non-converted users; perform feature extraction on the combined feature samples to obtain the user features and recommended object features corresponding to the combined feature samples; input the combined features of user features and recommended object features into the deep feature model, perform deep feature extraction on the combined features of the combined feature samples to obtain the corresponding predicted depth features; obtain the difference between the target depth features and the predicted depth features, and update the model parameters of the deep feature model based on the difference.

[0123] In practice, the value of the loss function of the deep feature model can be determined based on the difference between the target deep features and the predicted deep features. When the value of the loss function reaches a preset threshold, the corresponding error signal is determined based on the value of the loss function of the deep feature model. The error signal is then backpropagated in the deep feature model, and the model parameters of each layer of the deep feature model are updated during the propagation process.

[0124] In some embodiments, the combined feature samples constructed by the server include positive samples corresponding to converted users and negative samples of non-converted users. For example, high-value converted users are used as positive samples, and churned non-converted users in the business scenario are used as negative samples. When the positive and negative samples are imbalanced, techniques such as random undersampling, random oversampling, and cluster-based oversampling can be used to solve the problem of imbalance. Of course, other techniques that can balance positive and negative samples can also be used, and this invention is not specifically limited.

[0125] In some embodiments, before performing deep feature extraction on the user features and the combined features of the recommended object samples during the training of the recommendation model, or before inputting the combined features of the user features and the combined features of the recommended object samples into the deep feature model during the training of the deep feature model, or before performing deep feature extraction on the combined features of the user features and the recommended object features during the object recommendation process, the server may also perform at least one of the following preprocessing operations on the combined features to determine the combined features after the preprocessing operations as the corresponding recommended object sample, or combined feature sample, or combined features of the user to be recommended and the recommended object:

[0126] Determine the number of missing feature values ​​of the combined features and the number of identical feature values ​​of the corresponding combined features, and filter the combined features whose number of missing values ​​exceeds the missing value threshold or whose number of identical features exceeds the threshold, to obtain the filtered combined features.

[0127] Anomaly detection is performed on the feature values ​​of the combined features. When the detection result indicates that the feature value is an anomaly, the feature value is replaced with a filler feature value to obtain a combined feature containing the filler feature value.

[0128] The combined features are then subjected to derivation processing to obtain the derived combined features;

[0129] Discretize the feature values ​​corresponding to the continuous features in the combined features to obtain discrete features. Then, perform one-hot encoding on the discrete features obtained after discretization and the discrete features in the combined features, and use the resulting encoded features as the corresponding combined features.

[0130] Here, the missing value threshold is a = b × G, G ∈ [0,1], where b is the sample data size and G is the missing value filtering coefficient. The smaller the missing value filtering coefficient, the higher the filtering threshold, the fewer the number of features after filtering, and the more complete the feature values ​​corresponding to the filtered features. The same quantity threshold is c = d × M, M ∈ [0,1], where d is the sample data size and M is the same quantity filtering coefficient. If 90% of the values ​​(i.e., feature values) of a certain feature are the same, or all values ​​are the same, then the feature is called a single-value feature and needs to be filtered. Because the larger the number of identical feature values ​​corresponding to user features, the less useful it is for subsequent screening of strongly correlated features, so it needs to be filtered.

[0131] Here, the feature values ​​of combined features with outliers are discarded and replaced. That is, outliers corresponding to combined features are discarded. For example, based on the feature distribution, outliers with excessively large feature values ​​or outliers ranked in the top 0.0001 are discarded. Feature values ​​are filled for the discarded features. For example, the missing feature values ​​in continuous features are filled with the mean, and the missing feature values ​​in discrete features are filled with a constant.

[0132] As an example, discrete processing can be binning operations, including equidistant binning and equal-frequency binning. For instance, age is a continuous feature; after equal-frequency binning of the feature values ​​corresponding to age, the discrete age feature becomes {18-25, 26-30, 31-35, 36-40, 41-45, 46-50, 51-60, 60 and above}. The distribution of payment transactions within a certain time period (within 24 hours, one week, one month, or six months) is a continuous feature; after equidistant binning of the feature values ​​corresponding to the distribution of payment transactions within a certain time period (within 24 hours, one week, one month, or six months), the discrete distribution of payment transactions within that time period (within 24 hours, one week, one month, or six months) becomes {0-10, 11-20, 21-30, 31-40, 41-50, 51-60, 61-70, 70 and above}. Gender is a discrete feature, therefore, discrete processing of gender is unnecessary.

[0133] In practice, the process of deriving combined features includes at least one of the following: performing a basic transformation on a single variable in multiple combined features; or combining multiple variables in multiple combined features, such as feature squaring, feature addition and subtraction, etc.

[0134] It should be noted that the above preprocessing operations on combined features can be used to preprocess the combined features of combined training samples for training deep feature models, as well as to preprocess the combined features of combined features of recommended object samples used to train recommendation models, and also to preprocess the combined features of the object to be recommended and the user to be recommended when making object recommendations.

[0135] In this embodiment of the invention, filtering and filling combined features can better construct strongly correlated features, which can improve the accuracy of deep feature models trained based on strongly correlated combined features in predicting user conversion rates. By performing derivation processing on combined features, not only can the number of combined features be increased, but the cohesion between features can also be improved, thereby avoiding underfitting and overfitting problems in the trained deep feature models and recommendation models. Discretizing the feature values ​​corresponding to continuous features before encoding can reduce the time and space overhead of deep feature models and recommendation models, and can improve the classification and clustering ability and noise resistance of recommendation models. Furthermore, discrete features are easier to understand than continuous features and are closer to knowledge-level expressions, thereby overcoming the hidden defects in user data and making the object recommendation results based on the recommendation model more stable.

[0136] Step 103: Using a pattern mining model, perform frequent sequence pattern mining on the features of the users to be recommended to obtain the weighted features of the frequent sequence patterns of the corresponding users to be recommended.

[0137] In some embodiments, see Figure 5 , Figure 5 This is an optional flowchart illustrating a method for determining weighted features provided in an embodiment of the present invention. Figure 3 Step 103 shown can be achieved as follows: Figure 5 Steps 1031-1033 shown are implemented as follows:

[0138] Step 1031: Perform frequent sequence pattern mining on the features of the users to be recommended to obtain multiple frequent sequence patterns corresponding to the users to be recommended.

[0139] In some embodiments, the server can perform frequent sequence pattern mining on user features to obtain multiple frequent sequence patterns corresponding to the user to be recommended, in the following manner:

[0140] Based on user characteristics, a set of behavioral sequence patterns for the corresponding user to be recommended is determined. The set of behavioral sequence patterns includes at least two behavioral sequence patterns, which are obtained by combining multiple behavioral encoding features corresponding to the sequence samples. In the behavioral sequence patterns, behavioral encoding features are selected multiple times, and the selected behavioral encoding features are combined each time to obtain multiple different behavioral sequence sub-patterns for the corresponding behavioral sequence pattern. Each selected behavioral encoding feature includes behavioral encoding features belonging to the behavioral training patterns of different users to be recommended. From the multiple different behavioral sequence sub-patterns, behavioral sequence sub-patterns that appear more frequently than a frequency threshold in the set of behavioral sequence patterns are selected and determined as frequent sequence patterns.

[0141] In some embodiments, the server may select, from a set of sequences formed by multiple different behavior sequence sub-patterns, behavior sequence sub-patterns whose frequency of occurrence exceeds a frequency threshold as frequent sequence patterns in the following manner:

[0142] In the sequence set, a behavior sequence pattern prefix with a unit length of one is selected, and the projection dataset corresponding to the behavior sequence pattern prefix is ​​determined. When the projection dataset corresponding to the behavior sequence pattern prefix is ​​not empty, the behavior sequence pattern prefix whose frequency exceeds the frequency threshold is determined as a frequent one-feature encoding sequence. Based on the frequent one-feature behavior sequence pattern, the unit length of the selected behavior sequence pattern prefix is ​​increased sequentially to obtain n behavior sequence pattern prefixes. The projection dataset corresponding to the n behavior sequence pattern prefixes is determined. When the projection dataset corresponding to the n behavior sequence pattern prefixes is empty, and the frequency of the n behavior sequence pattern prefixes exceeds the frequency threshold, the n behavior sequence pattern prefixes are determined as the frequent sequence pattern. Here, n is a natural number that gradually increases in value, and the value range satisfies 2≤n≤K, and K is the number of behavior encoding features contained in the behavior sequence sub-pattern with the longest length in the sequence set.

[0143] Here, the frequency threshold, also known as the minimum support threshold, is min_sup = e × f, where e is the number of high-conversion-rate user samples (i.e., positive samples), and f is the minimum support rate. The minimum support rate f can be adjusted according to the number of high-conversion-rate user sample sets. The larger the number of sample sets, the larger the minimum support rate.

[0144] The following sections will illustrate the behavioral sequence patterns formed by the sequence information left by users clicking and browsing on the product, and the behavioral sequence patterns formed by a series of behavioral trajectories that lead to conversions from other channels.

[0145] For behavioral sequence patterns formed by users clicking and browsing on a product, for example, user Xiaoming clicks button 'a' on page A to enter page B, browses for a while, and then clicks button 'b' to enter page C; user Xiaoli clicks button 'a' on page A to enter page B, browses for a while, and then clicks button 'c' to return to page A. Then user Xiaoming's browsing sequence can be labeled as: AaBbC, and user Xiaoli's browsing sequence can be labeled as: AaBcA. The sequence information in the behavioral sequence patterns is sequential, therefore, frequent sequence patterns can be extracted based on sequence pattern mining algorithms. These frequent sequence patterns represent shared behavioral habits among different users.

[0146] Below, using a minimum support threshold (i.e., frequency threshold) of 0.5 as an example, we illustrate the specific process of mining frequent sequence patterns that meet the frequency threshold. First, we start mining behavioral sequence patterns from prefixes of length 1, searching the corresponding projection database to obtain the projection dataset (i.e., the corresponding suffixes) corresponding to prefixes of length 1, as shown in Table 1. For each prefix A, prefix a, and prefix B that meets the minimum support threshold, the corresponding suffixes are all non-empty. Therefore, we can conclude that the frequent sequence pattern of one item (i.e., length 1) in Table 1 is... 、 and .

[0147] Table 1

[0148]

[0149] Then, based on a single prefix, the frequent sequence patterns corresponding to prefixes of length 2 are recursively mined, and so on, until no longer prefixes can be mined. The binomial prefixes and corresponding suffixes that satisfy the minimum support threshold are shown in Table 2. The frequent sequence patterns of binomial prefixes (i.e., those of length 2) in Table 2 are... <aa>and <ab>.

[0150] Table 2

[0151]

[0152] Similarly, the three prefixes and corresponding suffixes that satisfy the minimum support threshold are shown in Table 3. The frequent sequence patterns of the three terms (i.e., those of length 3) in Table 3 are... <aab>.

[0153] Table 3

[0154]

[0155] At this point, the frequent sequence patterns of users Xiaoming and Xiaoli obtained are as follows:< / aab> < / ab> < / aa> 、 、 、 <aa> 、 <ab>and <aab>That is, users Xiaoming and Xiaoli share the aforementioned behavioral habits.

[0156] For a series of behavioral trajectories formed by users achieving conversion from other channels, the behavioral trajectory consists of sequence information composed of behavioral tags. The behavioral tag classification includes, but is not limited to, the tags shown in Table 4, and can be further defined and marked according to the actual application scenario and behavioral category.

[0157] Table 4

[0158] Behavioral tags Behavioral coding Purchasing behavior h Add to cart behavior g collecting behavior f Commenting behavior e Search behavior d Login behavior c Registration behavior b Browsing behavior a

[0159] In practice, the server extracts user behavior features, which consist of multiple user behavior label features, from the user features of the user to be recommended; based on the mapping relationship between behavior label features and behavior coding features, it determines the behavior coding features corresponding to each behavior label feature; and combines each behavior coding feature to obtain the behavior sequence pattern of the corresponding user to be recommended.

[0160] Taking a shopping platform product as an example, user Xiaoming enters the platform through a channel, then registers and logs in. After browsing the page for a while, he clicks to enter the item details page, then clicks the favorite button to favorite the item, and then clicks to add the item to the shopping cart to purchase it. Then, the user's behavior sequence pattern is: bcafgh.

[0161] User Xiao Li accessed the platform through a channel, then registered and logged in. After browsing the page for a period of time, he clicked to search for a specific product, added it to his shopping cart, and made a purchase. After the purchase, he added it to his favorites. The user's behavioral sequence pattern is: bcdaghf.

[0162] Taking a minimum support threshold (i.e., frequency threshold) of 0.5 as an example, the specific process of mining frequent sequence patterns that meet the frequency threshold is explained. First, behavioral sequence patterns are mined starting from prefixes of length 1, resulting in a prefix and its corresponding suffix that meet the minimum support threshold, as shown in Table 5. It can be seen that one frequent sequence pattern (i.e., length 1) in Table 5 is... 、 <c> 、< / c> < / aab> < / ab> < / aa> 、 <g>and <h>.

[0163] Table 5

[0164]

[0165] Then, based on a single prefix, the frequent sequence patterns corresponding to prefixes of length 2 are recursively mined, and so on, until no longer prefixes can be mined. The binomial prefixes and corresponding suffixes that satisfy the minimum support threshold are shown in Table 6. The frequent sequence patterns of binomial prefixes (i.e., those of length 2) in Table 6 are... <bc> 、 <ca> 、 <ag>and <gh>.

[0166] Table 6

[0167]

[0168] Similarly, the three prefixes and corresponding suffixes that satisfy the minimum support threshold are shown in Table 7. The frequent sequence patterns of the three terms (i.e., length 3) in Table 7 are... <bca> 、 <cag>and <agh>.

[0169] Table 7

[0170]

[0171]

[0172] Similarly, the four prefixes and corresponding suffixes that satisfy the minimum support threshold are shown in Table 8. The frequent sequence patterns of the four terms (i.e., length 4) in Table 8 are... <bcag>and <cagh>.

[0173] Table 8

[0174]

[0175] Similarly, the five prefixes and corresponding suffixes that satisfy the minimum support threshold are shown in Table 9. The frequent sequence patterns of the five items (i.e., those of length 5) in Table 9 are... <bcagh>.

[0176] Table 9

[0177] Five prefixes Corresponding suffix bcagh f

[0178] At this point, the frequent sequence patterns of users Xiaoming and Xiaoli obtained are as follows: 、 <c> 、< / c> < / bcagh> < / cagh> < / bcag> < / agh> < / cag> < / bca> < / gh> < / ag> < / ca> < / bc> < / h> < / g> 、 <g> 、 <h> 、 <bc> 、 <ca> 、 <ag> 、 <gh> 、 <bca> 、 <cag> 、 <agh> 、 <bcag> 、 <cagh>and <bcagh>That is, users Xiaoming and Xiaoli share the aforementioned behavioral habits.

[0179] Step 1032: Determine the weight of each frequent sequence pattern as the frequency proportion of each frequent sequence pattern in the set of behavioral sequence patterns corresponding to the user to be recommended.

[0180] The behavior sequence pattern set includes at least two behavior sequence patterns.

[0181] In some embodiments, after mining and transforming frequent sequence patterns (i.e., common user behavior patterns) from user behavior sequences and user browsing sequences, the modeling method focuses more on the behavioral encoding features of these frequent sequence patterns. These features are specially processed, such as repeatedly constructing these features to make the model learning more inclined towards these features, or the numerical range of these features is expanded to ensure that it is larger than the numerical range of other features, or the feature weights are initialized to weight the behavioral sequence features of frequent sequence patterns to obtain weighted features.

[0182] Here, the probability value of each frequent training pattern appearing in the behavioral sequence pattern set is determined as the weight of the corresponding frequent training pattern. The larger the weight, the more representative the corresponding frequent sequence pattern is of the user's common habits.

[0183] Step 1033: Based on the weights, the behavioral sequence encoding features of each frequent sequence pattern are weighted to obtain the weighted features of the frequent sequence patterns of the corresponding user to be recommended.

[0184] In this way, by using weighted features as strong features of common user habits, the model's learning is more inclined to these features, which is conducive to improving the accuracy of the recommendation model based on training in predicting user conversion rates.

[0185] In some embodiments, see< / bcagh> < / cagh> < / bcag> < / agh> < / cag> < / bca> < / gh> < / ag> < / ca> < / bc> < / h> < / g> Figure 6 , Figure 6 This is an optional flowchart illustrating a method for determining weighted features provided in an embodiment of the present invention. Figure 3 Step 103 shown can be achieved as follows: Figure 6 Steps 1034-1036 shown are implemented as follows:

[0186] Step 1034: Perform frequent sequence pattern mining on user features to obtain the frequent sequence pattern corresponding to the longest mined prefix;

[0187] Step 1035: Determine the weight of the frequent sequence pattern by the frequency percentage of the frequent sequence pattern in the set of behavioral sequence patterns corresponding to the user to be recommended.

[0188] The behavior sequence pattern set includes at least two behavior sequence patterns;

[0189] Step 1036: Based on the weights, the behavioral sequence encoding features of frequent sequence patterns are weighted to determine the weighted features of the frequent sequence patterns of the corresponding users to be recommended.

[0190] Since the mining of frequent sequence patterns is carried out on the premise of meeting the set minimum support threshold, the behavioral encoding features in the longest frequent sequence pattern also cover the behavioral encoding features in the shorter frequent sequence patterns, such as a five-item frequent sequence pattern. <bcagh>The behavioral encoding features in the data cover four frequent sequence patterns. <bcag>Therefore, the longest frequent sequence pattern best represents user behavior habits. Thus, the frequency proportion of the longest frequent sequence pattern within the set of behavioral sequence patterns corresponding to the user to be recommended can be determined as the weight of the frequent sequence pattern for that user. This weight is then used to weight the behavioral sequence encoding feature of the longest frequent sequence pattern, resulting in a weighted feature for the frequent sequence pattern of the user to be recommended. This approach uncovers common characteristics in the behavioral sequence patterns of a user group, solving the problem of relatively coarse feature construction in related technologies.

[0191] Step 104: Using a prediction model, the deep features and weighted features are fused to obtain the fused features of the corresponding user to be recommended, and the conversion rate is predicted based on the fused features to obtain the prediction result.

[0192] In some embodiments, the prediction model includes a feature fusion layer and a feature prediction layer. In actual implementation, the feature fusion layer fuses deep features and weighted features to obtain fused features corresponding to the user to be recommended. The feature prediction layer then uses the obtained fused features to predict the conversion rate and obtain the prediction result.

[0193] In some embodiments, the prediction model is a fully connected neural network, including an input layer, a hidden layer, and an output layer. In actual implementation, the server inputs the fused features of the acquired deep features and weighted features into the input layer, so that they can be passed to the hidden layer through the input layer. Through the hidden layer, the activation function is called to obtain the hidden features of the corresponding fused features. Through the output layer, the conversion rate of the obtained hidden features is predicted to obtain the prediction result.

[0194] Step 105: When the prediction result indicates that the conversion rate of the user to be recommended exceeds the probability threshold, the object to be recommended is recommended to the user to be recommended.

[0195] Here, if the conversion rate of the user to be recommended exceeds the probability threshold, it indicates that the user is highly likely to actually convert to the recommended item. Recommending the item to this user satisfies their needs and improves the accuracy of the recommendation. For example, on an e-commerce shopping platform, if the probability threshold is set to 80%, and user A's conversion rate for product B is 85%, it means that user A is highly likely to actually purchase product B. Recommending product B to user A improves the actual conversion rate of the product.

[0196] By employing the methods described above, a deep feature model trained on user features and recommended object features is used for deep learning. This yields deep features that are more conducive to predicting user conversion rates. Frequent sequence mining of user features yields weighted features of frequent sequence patterns, enabling the discovery of common features in user behavior sequence patterns and addressing the issue of coarse feature construction in related technologies. Combining deep and weighted features to predict user conversion rates improves the accuracy of conversion rate prediction, thereby enhancing the accuracy of object recommendations based on the predicted user conversion probabilities and improving the generalization ability of the recommendation model.

[0197] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0198] With the rise of online consumption and the rapid development of e-commerce platforms, a more accurate understanding of users' personalized needs has made personalized recommendations a key competitive advantage for these platforms. Therefore, establishing personalized recommendation models has become a crucial strategy for businesses to understand user consumption needs, seize market opportunities, and increase transaction volume. Furthermore, accurately identifying user behavior habits and providing personalized items based on user characteristics is also significant for applications such as benefit recommendations and advertising.

[0199] Personalized recommendation systems for business scenarios have wide applications in e-commerce platforms, advertising, and other fields, demonstrating high application value in areas such as user profiling, personalized recommendations, and targeted advertising on e-commerce platforms. Among related technologies, mainstream methods for building recommendation systems for business scenarios include: demographic-based recommendations, content-based recommendations, and recommendations based on coordinated filtering algorithms. However, these methods face the following technical challenges in application:

[0200] 1) Recommendations based on manual statistics are based only on the basic user profile features. The constructed features are relatively coarse and cannot reflect the commonalities of the user group. The effect is difficult to meet the needs and is only suitable for simple recommendations. At the same time, many large-scale system products do not have the cold start problem. This algorithm is significantly different from other algorithms.

[0201] 2) Content-based recommendations only consider the similarity between the attributes of the items themselves. However, the attributes are limited, and the user's historical item data is required. It is difficult to generalize to new users, meaning that the generalization ability of content-based recommendation models is weak.

[0202] 3) Recommendations based on coordination filtering rely on users' historical preferences, regardless of the item's content attributes. New items in the cold start phase lack this historical preference information, making recommendations impossible. This results in low accuracy and generalization ability for coordination filtering-based recommendation models. Furthermore, matrix factorization requires a large number of real ratings; otherwise, the input to the factorization model becomes very sparse, hindering accurate predictions and further reducing recommendation accuracy.

[0203] In view of the problems existing in the above methods, the embodiments of the present invention provide an object recommendation method and apparatus based on a recommendation model. The embodiments of the present invention have at least the following advantages compared with related technologies:

[0204] 1) By constructing a sample library of actual converted users, this invention mines user attributes and behavioral sequence patterns based on the Prefixspan algorithm, discovers the common characteristics of the converted user group in behavioral sequence patterns, can mine deeper common attributes of converted users and focus on these features, and better construct the behavioral features of converted users.

[0205] 2) This invention integrates and processes strong features (i.e., weighted features of frequent sequence patterns) and transforms the range of feature values, enabling strong features to play a greater role in applications. This solves the problem that related technologies do not adequately assess the importance of user features and perform strong feature analysis. By fusing the deep features and strong features output by the deep feature model, the accuracy of the recommendation model can be improved. At the same time, by constructing a deep feature model to learn the deep features of user items (i.e., recommended objects), the recommendation model can make decisions from a broader perspective. By fusing these two types of features to recommend users, the recommendation effect of the recommendation model can be improved.

[0206] 3) The user recommendation system construction method provided in this embodiment of the invention, which performs strong feature processing based on the user behavior sequence pattern and combines deep model features for feature fusion classification, improves the generalization ability of the recommendation model.

[0207] See Figure 7 , Figure 7 This is an optional flowchart illustrating an object recommendation method based on a recommendation model provided in an embodiment of the present invention. The recommendation model provided in this embodiment includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model, which will be combined with... Figure 7 The steps shown are explained.

[0208] Step 201: Extract user features and item features using a feature extraction model, and construct a sample of recommended objects.

[0209] Here, the user characteristics corresponding to the recommended object samples include basic user attributes such as age, gender, education level, and city tier; user consumption characteristics such as total number of payments, total amount, distribution of payment number within a certain time period (within 24 hours, within a week, within a month, within six months), distribution of payment amount, and average amount per payment; and user behavior characteristics such as browsing time and number of page clicks. Item characteristics include basic item attributes such as item category, item price, item brand, item purchase rating, and item review sentiment; and item consumption characteristics such as the number of times an item has been purchased, the number of times it has been clicked / viewed, the number of times it has been added to the shopping cart, and the number of times similar items have been purchased.

[0210] The <user, item> combined feature is constructed by concatenating and combining two parts of features, and data preprocessing is performed. The specific processing steps include:

[0211] 1) Discard features with too many missing values: Set the missing value filtering threshold = sample data size * G, G∈[0,1], G can be set to 0.4. When the number of missing values ​​of a certain feature data (i.e. the feature value or feature value mentioned above) exceeds the threshold, filter the feature and delete the single-value feature.

[0212] 2) Handle outliers (i.e., the outlier features mentioned above): Based on the feature distribution, discard outliers that are too large or ranked in the top 0.0001 (i.e., one in ten thousand).

[0213] 3) Missing value handling: Missing values ​​of continuous features are filled with the mean, and missing values ​​of discrete features are filled with a constant as a separate category.

[0214] 4) Feature derivation: Feature combination and derivation are performed through feature transformation, feature squaring, and feature addition and subtraction.

[0215] 5) Feature processing: Continuous features are binned and discretized (the binning method can be based on the distribution ratio of user feature fields in each interval, i.e., equal frequency binning operation), and the binned and discretized continuous features and discrete features are type-coded.

[0216] Users who actually convert within the business scenario are considered high-value users. High-value users are defined as those who have purchased memberships, have high points and have redeemed points, and have large historical transaction amounts. These users are used as positive samples to build a high-value user sample library. At the same time, churned users in the business scenario are used as negative samples. In this way, the constructed recommendation object sample is labeled with conversion probability, including positive samples of users who convert and negative samples of users who do not convert.

[0217] Step 202: Learn the deep features of user items through a deep feature model.

[0218] Here, based on step 201, user features and item features are constructed and processed to obtain the combined features of each user and item pair. A deep learning model (i.e., a deep feature model) using a convolutional neural network (CNN) or a long short-term memory network (LSTM) is built to learn the deep features corresponding to the combined features of user and item.

[0219] Here, using the aforementioned constructed sample of recommended objects, a sampling algorithm such as undersampling is employed to balance the number of positive and negative samples, and the deep feature model is trained to obtain a well-trained deep feature model. In practical applications, the combined features of the user features and the item features to be recommended are input into the trained deep feature model to obtain the deep features corresponding to the combined features.

[0220] Step 203: Using a pattern mining model, perform frequent sequence pattern mining on user features to obtain the user's frequent sequence patterns.

[0221] Here, the Prefixspan algorithm is used to mine user behavior sequence patterns and discover user groups with common habits / browsing habits from initial contact to conversion. The Prefixspan algorithm is also used to mine frequent sequence patterns of various lengths that satisfy the minimum support threshold in user behavior trajectories. Simultaneously, a multiple minimum support strategy is employed, and the minimum support is calculated as shown in formula (1).

[0222] min_sup=a×n(1)

[0223] Where n is the number of recommended object samples, and a is the minimum support rate, which is adjusted according to the number of recommended object samples. Furthermore, this embodiment of the invention employs a "snowball" approach, where each round of mining sets a high support level to ensure the accuracy of behavioral sequence pattern mining, and multiple rounds of iterative mining improve the recall rate of sequence pattern mining.

[0224] The specific process of the sequence pattern mining algorithm is as follows:

[0225] (1) Find the prefix and corresponding projection dataset of user behavior sequence with a unit length of 1;

[0226] (2) Count the frequency of occurrence of prefixes in user behavior sequences and add prefixes with support higher than the minimum support threshold to the dataset to obtain frequent item set behavior sequence patterns (i.e., frequent sequence patterns of an item).

[0227] (3) Recursively mine all prefixes of length i that meet the minimum support requirement:

[0228] 1) Mine the projected dataset of the prefix; if the projected data is an empty set, return recursively.

[0229] 2) Calculate the minimum support of each item in the corresponding projected dataset, merge each item that meets the minimum support threshold with the current prefix to obtain a new prefix, and recursively return if the minimum support threshold is not met.

[0230] 3) Let i = i + 1, and let the prefixes be the new prefixes after merging individual terms. Recursively execute step (3) respectively;

[0231] (4) Return all frequent sequence patterns in the set of user behavior sequences.

[0232] Using the methods described above, common behavioral patterns (i.e., frequent sequence patterns) of users can be mined from user behavior sequences and user browsing sequences.

[0233] Step 204: Using a pattern mining model, the frequent sequence pattern features of users are weighted by support to obtain the weighted features of frequent sequence patterns.

[0234] The frequent sequence patterns of users obtained through step 203 reflect their common habits. When building the recommendation model, more attention can be paid to these weighted features, and these features should be specially processed. The steps for this special processing are as follows:

[0235] 1) Special transformations for strong features, including reconstructing important features repeatedly, so that the recommendation model learns to favor these features;

[0236] 2) Perform numerical transformations on important features to expand their numerical range and ensure that their numerical range is larger than that of other features;

[0237] 3) Initialize feature weights by weighting the features. The weights are set to the frequency percentage of each field type. For example, if the minimum support threshold is set to 0.5, and the frequency percentage of all types of values ​​for a certain field is less than the minimum support, then that field is removed. If the frequency percentage of field type "collection behavior f" is 0.7, then the weight of that field type is 0.7; if the frequency percentage of field type "browsing sequence AaBcA" is 0.56, then the weight of that field type is 0.56. Based on this weight, the sequence pattern features are weighted. The original sequence pattern features are the user behavior sequence encoding features mined above, which are then numerically represented and weighted.

[0238] Step 205: Use the prediction model to fuse deep features and weighted features for classification and scene recommendation.

[0239] In practical applications, the prediction model may include a feature fusion layer and a feature prediction layer. The feature fusion layer fuses the deep features and weighted features obtained above to obtain fused features. In practical applications, the feature prediction layer includes a fully connected layer and a feature classification layer (i.e., a classifier). In actual implementation, the server inputs the fused features of the deep features and weighted features into the fully connected layer. The fully connected layer calls the activation function to obtain the hidden features of the corresponding fused features. The feature classification layer then performs conversion rate prediction on the obtained hidden features to obtain the prediction result.

[0240] Here, the difference between the predicted results and the labeled conversion rate can be obtained, and the model parameters of the recommendation model can be updated based on the difference to complete the object recommendation based on the recommendation model.

[0241] When using a trained recommendation model for item recommendation, see [link to relevant documentation]. Figure 8 , Figure 8 This is a schematic diagram of the object recommendation method provided in an embodiment of the present invention, such as... Figure 8 As shown, the item data of the item to be recommended and the user data of the user to be recommended are input into the trained recommendation model to obtain user-item combination features and sequence pattern weighted features. The user-item combination features are then processed by a deep feature model of CNN or LTSM to obtain corresponding deep features. The deep features and weighted features are then fused to obtain corresponding fused features. The fused features are then processed through a fully connected layer to obtain the hidden features of the fused features. A classifier is used to predict the conversion rate of the obtained hidden features to obtain the prediction results. When the prediction result indicates that the conversion rate value of the user to be recommended exceeds the probability threshold, the item to be recommended is recommended to the user. That is, the user with a higher conversion probability is more worthy of recommendation.

[0242] The following describes an exemplary structure of the software module. In some embodiments, see [link to documentation]. Figure 9 , Figure 9 This is an optional structural diagram of an object recommendation device 555 based on a recommendation model provided in an embodiment of the present invention. The recommendation model includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model. The device 555 includes:

[0243] The first feature extraction module 5551 is used to extract features from the user data of the user to be recommended and the object data of the object to be recommended through the feature extraction model, so as to obtain the user features corresponding to the user to be recommended and the object features corresponding to the object to be recommended.

[0244] The second feature extraction module 5552 is used to perform deep feature extraction on the combined features of the user features and the features of the object to be recommended through the deep feature model to obtain the corresponding deep features;

[0245] The feature determination module 5553 is used to perform frequent sequence pattern mining on the user features through the pattern mining model to obtain weighted features of frequent sequence patterns corresponding to the user to be recommended.

[0246] The feature prediction module 5554 is used to fuse the deep features and the weighted features through the prediction model to obtain the fused features corresponding to the user to be recommended, and to predict the conversion rate based on the fused features to obtain the prediction result.

[0247] The object recommendation module 5555 is used to recommend the object to the user to be recommended when the prediction result indicates that the conversion rate value of the user to be recommended exceeds a probability threshold.

[0248] In some embodiments, before performing feature extraction on the user data of the user to be recommended and the object data of the object to be recommended using the feature extraction model, the device further includes a recommendation model training module.

[0249] The recommendation model training module is used to extract features from the recommended object samples labeled with conversion rates through the feature extraction model, so as to obtain the user features and recommended object features corresponding to the recommended object samples.

[0250] Using the deep feature model, deep feature extraction is performed on the combined features of the user features and the recommended object features to obtain the corresponding deep features;

[0251] The pattern mining model is used to mine frequent sequence patterns of the user features to obtain weighted features of frequent sequence patterns corresponding to the recommended object samples.

[0252] The prediction model fuses the deep features and the weighted features to obtain the fused features corresponding to the recommended object sample, and the conversion rate is predicted based on the fused features to obtain the prediction result.

[0253] Obtain the difference between the predicted result and the labeled conversion rate, and update the model parameters of the recommendation model based on the difference.

[0254] In some embodiments, the feature extraction model includes a feature extraction layer and a feature concatenation layer for extracting features from the user data of the user to be recommended and the object data of the object to be recommended. This occurs before the deep feature model performs deep feature extraction on the combined features of the user features and the object features to be recommended.

[0255] The first feature extraction module is further configured to concatenate the features of the user to be recommended and the features of the object to be recommended through the feature concatenation layer to obtain combined features corresponding to the user to be recommended and the object to be recommended.

[0256] In some embodiments, before performing deep feature extraction on the combined features of the user features and the object features to be recommended using the deep feature model, the device further includes a deep feature training module.

[0257] The deep feature training module is used to obtain combined feature samples labeled with target deep features. The combined feature samples include positive samples corresponding to converted users and negative samples of unconverted users.

[0258] Feature extraction is performed on the combined feature samples to obtain the user features and recommendation object features corresponding to the combined feature samples;

[0259] The combined features of the user features and the recommended object features are input into the deep feature model, and deep feature extraction is performed on the combined features of the combined feature samples to obtain the corresponding predicted deep features;

[0260] The difference between the target depth features and the predicted depth features is obtained, and the model parameters of the depth feature model are updated based on the difference.

[0261] In some embodiments, before performing deep feature extraction on the combined features of the user features and the object features to be recommended, the apparatus further includes a preprocessing module.

[0262] The preprocessing module is configured to perform at least one of the following preprocessing operations on the combined features, so as to determine the combined features after the preprocessing operations as the combined features of the user features and the recommendation object features:

[0263] Determine the number of missing feature values ​​of the combined feature and the number of identical feature values ​​corresponding to the combined feature, and filter the combined features whose number of missing values ​​exceeds the missing value threshold or whose number exceeds the identical value threshold to obtain the filtered combined features;

[0264] Anomaly detection is performed on the feature values ​​of the combined features. When the detection result indicates that the feature value is an anomaly, the feature value is replaced with a filler feature value to obtain a combined feature containing the filler feature value.

[0265] The combined features are then subjected to derivation processing to obtain the derived combined features;

[0266] Discretize the feature values ​​corresponding to the continuous features in the combined features to obtain discrete features, and use the discrete features obtained after discretization and the discrete features in the combined features as the combined features corresponding to the combined recommendation object sample.

[0267] In some embodiments, the feature determination module is further configured to perform frequent sequence pattern mining on the features of the user to be recommended, and obtain multiple frequent sequence patterns corresponding to the user to be recommended;

[0268] The frequency proportion of each of the frequent sequence patterns in the behavioral sequence pattern set corresponding to the user to be recommended is determined as the weight of the corresponding frequent sequence pattern. The behavioral sequence pattern set includes at least two behavioral sequence patterns.

[0269] Based on the weights, the behavioral sequence encoding features of each frequent sequence pattern are weighted to obtain the weighted features of the frequent sequence patterns of the user to be recommended.

[0270] In some embodiments, the feature determination module is further configured to determine a set of behavioral sequence patterns corresponding to the user to be recommended based on the features of the user to be recommended. The set of behavioral sequence patterns includes at least two behavioral sequence patterns, and the behavioral sequence patterns are obtained by combining multiple behavioral encoding features corresponding to the sequence samples.

[0271] In the behavior sequence pattern, behavior encoding features are selected multiple times, and the selected behavior encoding features are combined to obtain multiple different behavior sequence sub-patterns corresponding to the behavior sequence pattern.

[0272] Each selected behavioral encoding feature includes behavioral encoding features belonging to different behavioral training patterns of users to be recommended.

[0273] From the multiple different behavior sequence sub-patterns, select the behavior sequence sub-patterns whose frequency of occurrence in the behavior sequence pattern set exceeds a frequency threshold, and determine the selected behavior sequence sub-patterns as frequent sequence patterns.

[0274] In some embodiments, the feature determination module is further configured to select a behavior sequence pattern prefix with a unit length of one, and determine the projection dataset corresponding to the behavior sequence pattern prefix;

[0275] When the projection dataset corresponding to a behavior sequence pattern prefix is ​​not empty, a behavior sequence pattern prefix whose frequency exceeds the frequency threshold is determined as a frequent feature encoding sequence;

[0276] Based on the aforementioned frequent behavior sequence pattern, the unit length of the selected behavior sequence pattern prefix is ​​increased sequentially to obtain n behavior sequence pattern prefixes.

[0277] Determine the projection dataset corresponding to the prefixes of the n behavioral sequence patterns;

[0278] When the projection dataset corresponding to the n behavioral sequence pattern prefixes is empty and the occurrence frequency of the n behavioral sequence pattern prefixes exceeds the frequency threshold, the n behavioral sequence pattern prefixes are determined as the frequent sequence patterns.

[0279] Where n is a natural number whose value gradually increases, and the range of values ​​satisfies 2≤n≤K, and K is the number of behavioral encoding features contained in the longest behavioral sequence sub-pattern in the sequence set.

[0280] In some embodiments, the feature determination module is further configured to perform frequent sequence pattern mining on the features of the user to be recommended, and obtain the frequent sequence pattern corresponding to the prefix with the longest mining length;

[0281] The frequency proportion of the frequent sequence pattern in the set of behavioral sequence patterns corresponding to the user to be recommended is determined as the weight of the frequent sequence pattern. The set of behavioral sequence patterns includes at least two behavioral sequence patterns.

[0282] Based on the weights, the behavioral sequence encoding features of the frequent sequence patterns are weighted to determine the weighted features of the frequent sequence patterns corresponding to the users to be recommended.

[0283] This invention provides an electronic device, comprising:

[0284] Memory, used to store executable instructions;

[0285] A processor, when executing executable instructions stored in the memory, implements the method provided in the embodiments of the present invention.

[0286] This invention provides a computer-readable storage medium storing executable instructions for inducing a processor to execute the method provided in this invention.

[0287] In some embodiments, the storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0288] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0289] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., files that store one or more modules, subroutines, or code sections).

[0290] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0291] In summary, the embodiments of the present invention have the following beneficial effects:

[0292] 1) By performing frequent sequence mining on user features, the weighted features of users' frequent sequence patterns are obtained, which can uncover the common features of user groups in behavioral sequence patterns and solve the problem that the features constructed by related technologies are relatively coarse.

[0293] 2) By using a deep feature model to perform deep learning on user features and recommended object features, we obtain deep features that are more conducive to predicting user conversion rates. By combining deep features and weighted features, we can predict user conversion rates, which improves the accuracy of user conversion rate prediction. Then, based on the predicted user conversion probability, we can make object recommendations, which not only improves the accuracy of recommending objects based on the recommendation model, but also improves the generalization ability of the recommendation model.

[0294] 3) Filtering and filling combined features can better construct strongly correlated features, which can improve the accuracy of deep feature models trained on strong correlated combined features in predicting user conversion rates.

[0295] 4) By performing derivation processing on the combined features, not only can the number of combined features be increased, but also the cohesion between features can be improved, thereby avoiding the problems of underfitting and overfitting in the deep feature models and recommendation models obtained through training.

[0296] 5) Discretizing the feature values ​​corresponding to continuous features before encoding can reduce the time and space overhead of deep feature models and recommendation models, and improve the classification and clustering ability and noise resistance of recommendation models. Furthermore, discrete features are easier to understand than continuous features and are closer to the expression of knowledge, thereby overcoming the hidden defects in user data and making the object recommendation results based on the recommendation model more stable.

[0297] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.< / bcag> < / bcagh>

Claims

1. An object recommendation method based on a recommendation model, characterized in that, The recommendation model includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model; the method includes: Using the feature extraction model, feature extraction is performed on the user data of the user to be recommended and the object data of the object to be recommended, respectively, to obtain the corresponding features of the user to be recommended and the features of the object to be recommended. Using the deep feature model, deep feature extraction is performed on the combined features of the user features to be recommended and the object features to be recommended to obtain the corresponding deep features; Based on the features of the user to be recommended, the pattern mining model determines a set of behavioral sequence patterns corresponding to the user to be recommended. The set of behavioral sequence patterns includes at least two behavioral sequence patterns, which are obtained by combining multiple behavioral encoding features of the corresponding sequence samples. In the behavior sequence pattern, behavior encoding features are selected multiple times, and the selected behavior encoding features are combined to obtain multiple different behavior sequence sub-patterns corresponding to the behavior sequence pattern. Each selected behavioral encoding feature includes behavioral encoding features belonging to different behavioral training modes of users to be recommended. From the multiple different behavior sequence sub-patterns, select behavior sequence sub-patterns whose frequency of occurrence in the behavior sequence pattern set exceeds a frequency threshold, and determine the selected behavior sequence sub-patterns as frequent sequence patterns. Then, determine the weighted features of the frequent sequence patterns. The weighted features are obtained by weighting the behavior sequence encoding features of each frequent sequence pattern based on weights. The prediction model fuses the deep features and the weighted features to obtain the fused features corresponding to the user to be recommended, and the conversion rate is predicted based on the fused features to obtain the prediction result. When the prediction result indicates that the conversion rate of the user to be recommended exceeds the probability threshold, the object to be recommended is recommended to the user to be recommended.

2. The method as described in claim 1, characterized in that, Before performing feature extraction on the user data of the user to be recommended and the object data of the object to be recommended using the feature extraction model, the method further includes: Using the feature extraction model, feature extraction is performed on the recommended object samples labeled with conversion rates to obtain the user features and recommended object features corresponding to the recommended object samples. Using the deep feature model, deep feature extraction is performed on the combined features of the user features and the recommended object features to obtain the corresponding deep features; The pattern mining model is used to mine frequent sequence patterns of the user features to obtain weighted features of frequent sequence patterns corresponding to the recommended object samples. The prediction model fuses the deep features and the weighted features to obtain the fused features corresponding to the recommended object sample, and the conversion rate is predicted based on the fused features to obtain the prediction result. Obtain the difference between the predicted result and the labeled conversion rate, and update the model parameters of the recommendation model based on the difference.

3. The method as described in claim 1, characterized in that, The feature extraction model includes a feature extraction layer and a feature concatenation layer for extracting features from the user data of the user to be recommended and the object data of the object to be recommended. Before performing deep feature extraction on the combined features of the user features and the object features to be recommended using the deep feature model, the method further includes: The feature splicing layer is used to splice the features of the user to be recommended and the features of the object to be recommended to obtain the combined features corresponding to the user to be recommended and the object to be recommended.

4. The method as described in claim 1, characterized in that, Before performing deep feature extraction on the combined features of the user features and the object features to be recommended using the deep feature model, the method further includes: Obtain combined feature samples labeled with target depth features, wherein the combined feature samples include positive samples corresponding to converted users and negative samples of unconverted users; Feature extraction is performed on the combined feature samples to obtain the user features and recommendation object features corresponding to the combined feature samples; The combined features of the user features and the recommended object features are input into the deep feature model, and deep feature extraction is performed on the combined features of the combined feature samples to obtain the corresponding predicted deep features; The difference between the target depth features and the predicted depth features is obtained, and the model parameters of the depth feature model are updated based on the difference.

5. The method as described in claim 1, characterized in that, Before performing deep feature extraction on the combined features of the user features and the object features to be recommended, the method further includes: Perform at least one of the following preprocessing operations on the combined features to determine the combined features after the preprocessing operations as the combined features of the user features to be recommended and the object features to be recommended: Determine the number of missing feature values ​​of the combined feature and the number of identical feature values ​​corresponding to the combined feature, and filter the combined features whose number of missing values ​​exceeds the missing value threshold or whose number exceeds the identical value threshold to obtain the filtered combined features; Anomaly detection is performed on the feature values ​​of the combined features. When the detection result indicates that the feature value is an anomaly, the feature value is replaced with a filler feature value to obtain a combined feature containing the filler feature value. The combined features are subjected to derivation processing to obtain derived combined features, wherein the process of derivation processing of the combined features includes at least one of the following: performing a basic transformation on a single variable in multiple combined features; or combining multiple variables in multiple combined features. Discretize the feature values ​​corresponding to the continuous features in the derived combined features to obtain discrete features, and use the discrete features obtained after discretization and the discrete features in the combined features as the corresponding combined features.

6. The method as described in claim 1, characterized in that, The determination of the weighted features of the frequent sequence pattern includes: The frequency proportion of each of the frequent sequence patterns in the behavioral sequence pattern set corresponding to the user to be recommended is determined as the weight of the corresponding frequent sequence pattern. The behavioral sequence pattern set includes at least two behavioral sequence patterns. Based on the weights, the behavioral sequence encoding features of each frequent sequence pattern are weighted to obtain the weighted features of the frequent sequence patterns of the user to be recommended.

7. The method as described in claim 1, characterized in that, The step of selecting a sub-pattern of a behavior sequence that occurs more frequently than a frequency threshold in the behavior sequence pattern set, and determining the selected sub-pattern of the behavior sequence as a frequent sequence pattern, includes: Select a behavior sequence pattern prefix with a length of one element, and determine the suffix corresponding to the behavior sequence pattern prefix; When the suffix corresponding to a behavior sequence pattern prefix is ​​not empty, a behavior sequence pattern prefix whose frequency exceeds the frequency threshold is determined as a frequent feature encoding sequence. Based on the frequent behavior sequence pattern, the length of the selected behavior sequence pattern prefix is ​​increased sequentially to obtain n behavior sequence pattern prefixes; Determine the suffix corresponding to the prefix of the n behavioral sequence patterns; When the suffix corresponding to the prefix of the n behavioral sequence patterns is empty and the frequency of the occurrence of the prefix of the n behavioral sequence patterns exceeds the frequency threshold, the prefix of the n behavioral sequence patterns is determined as the frequent sequence pattern. Where n is a natural number whose value gradually increases, and the range of values ​​satisfies 2≤n≤K, and K is the number of behavioral encoding features contained in the longest behavioral sequence sub-pattern in the sequence set.

8. A recommendation device based on a recommendation model, characterized in that, The recommendation model includes: a feature extraction model, a deep feature model, a pattern mining model, and a prediction model; the device includes: The first feature extraction module is used to extract features from the user data of the user to be recommended and the object data of the object to be recommended through the feature extraction model, so as to obtain the corresponding features of the user to be recommended and the features of the object to be recommended. The second feature extraction module is used to perform deep feature extraction on the combined features of the user features to be recommended and the object features to be recommended through the deep feature model to obtain the corresponding deep features; The feature determination module is used to determine a set of behavioral sequence patterns corresponding to the user to be recommended based on the features of the user to be recommended through the pattern mining model. The set of behavioral sequence patterns includes at least two behavioral sequence patterns. The behavioral sequence patterns are obtained by combining multiple behavioral encoding features of the corresponding sequence samples. In the behavioral sequence patterns, behavioral encoding features are selected multiple times and the selected behavioral encoding features are combined each time to obtain multiple different behavioral sequence sub-patterns corresponding to the behavioral sequence patterns. Each selected behavior encoding feature includes behavior encoding features belonging to different behavior training patterns of users to be recommended; from the multiple different behavior sequence sub-patterns, behavior sequence sub-patterns that appear more frequently than a frequency threshold in the behavior sequence pattern set are selected, and the selected behavior sequence sub-patterns are determined as frequent sequence patterns, and the weighted features of the frequent sequence patterns are determined. The weighted features are obtained by weighting the behavior sequence encoding features of each frequent sequence pattern based on weights. The feature prediction module is used to fuse the deep features and the weighted features through the prediction model to obtain the fused features corresponding to the user to be recommended, and to predict the conversion rate based on the fused features to obtain the prediction result; The recommendation module is used to recommend the object to the user to be recommended when the prediction result indicates that the conversion rate value of the user to be recommended exceeds a probability threshold.

9. The apparatus according to claim 8, characterized in that, The feature extraction model includes a feature extraction layer and a feature concatenation layer for extracting features from user data of the user to be recommended and object data of the object to be recommended. The first feature extraction module is further configured to concatenate the features of the user to be recommended and the features of the object to be recommended through the feature concatenation layer before performing deep feature extraction on the combined features of the user to be recommended and the object to be recommended through the deep feature model, so as to obtain the combined features corresponding to the user to be recommended and the object to be recommended.

10. An electronic device, characterized in that, The electronic device includes: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the object recommendation method based on a recommendation model as described in any one of claims 1 to 7.

11. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, they implement the object recommendation method based on a recommendation model as described in any one of claims 1 to 7.

Images