Information recommendation system and method based on large model implementation

Abstract

The application discloses an information recommendation system and method based on a large model, and the system comprises a data acquisition module, which acquires original related data, and obtains a processed data set by preprocessing the original related data; a feature engineering module, which performs sample screening and feature processing on the processed data set to obtain a final sample set, information features and user features; an information recommendation module, which responds to a triggering request of a front-end user, calls a recommendation engine API and then calls a recall unit, the recall unit recalls an information recommendation training model and obtains a recommendation result, the recommendation result is filtered through a filtering unit and transmitted to a sorting unit, the sorting unit sorts to obtain a sorting result, which is then transmitted to a rearrangement unit, the rearrangement unit rearranges the sorting result to form a recommendation list; and an application recommendation module. The application improves the deep representation of content and information, and thus improves the accuracy and diversity of model reasoning, so that the model can better adapt to the individual needs of different users.

Description

Technical Field

[0001] This invention relates to the field of recommendation technology, and in particular to an information recommendation system and method based on a large model. Background Technology

[0002] In the context of the rapid development of the internet, recommender systems, as a core technology for improving user experience and optimizing platform operation, have been widely used in numerous internet products. Recommender systems provide personalized recommendation services by deeply mining and analyzing user data, achieving a "one-size-fits-all" effect and accurately pushing information that meets the user's needs. Currently, most information recommendation systems are based on algorithmic modeling using basic information data and user behavior data. Information data comes in various forms, including text, images, audio, and video, which undoubtedly increases the difficulty of information understanding.

[0003] Existing recommendation technologies suffer from insufficient depth of understanding when processing diverse data, making it difficult to accurately grasp the core meaning and value of information. This is especially true in the securities or finance industry, where the amount of data is relatively small, making it impossible to provide sufficient data support for recommendation systems as in some large-scale data scenarios.

[0004] Due to limited data or misunderstandings of certain information, current securities information recommendation systems often suffer from the following shortcomings: Insufficient sample size makes it difficult to meet the ideal training requirements of the model, which undoubtedly negatively impacts the model's accuracy, reduces its generalization ability in diverse scenarios, and makes it difficult to accurately predict various user behaviors and preferences. The lack of data makes it difficult for the recommendation system to comprehensively and accurately grasp user preferences and behavioral patterns during training, resulting in relatively low accuracy. Alternatively, insufficient understanding of the information makes it difficult to accurately analyze the core elements, semantic relationships, and potential value of each modality of information, hindering the in-depth exploration of users' true needs for different modalities of information and failing to meet users' precise needs for obtaining securities information. This severely affects the accuracy of the recommendation system in providing accurate information recommendations to users.

[0005] In conclusion, the accuracy and diversity of current recommendation systems still need further improvement. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by providing an information recommendation system and method based on a large model.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] An information recommendation system based on a large model includes:

[0009] The data acquisition module acquires raw relevant data and preprocesses the raw relevant data to obtain a processed dataset. The raw relevant data includes user basic information, raw information, and user behavior data.

[0010] The feature engineering module performs sample selection and feature processing on the dataset to obtain the final sample set, information features, user features, and other relevant features.

[0011] The information recommendation module trains a pre-trained information recommendation model based on the final sample set, information features, user features, and other relevant features to obtain an information recommendation model. The information recommendation model then performs inference to obtain a recommendation list. Specifically, in response to a user's trigger request, the recommendation engine API is called, which in turn calls the recall unit. The recall unit recalls the information recommendation training model and obtains the recommendation results. These results are then filtered by the filtering unit and passed to the ranking unit. The ranking unit sorts the results and passes them to the rearrangement unit. The rearrangement unit rearranges the results to form the recommendation list.

[0012] The application recommendation module pushes a list of recommended information to the front end and displays the list to the user via the recommendation engine API.

[0013] As one possible implementation, the feature engineering module includes a sample screening unit and a feature processing unit, wherein the sample screening unit includes a sample enhancement subunit and a sampling subunit;

[0014] The sample enhancement subunit mines and expands the original information through the LLM big model to generate diversified information; the LLM big model forms user preferences based on the user's past interaction history; the LLM big model generates synthetic information related to the new user's interests based on the new user's historical data, where the original information includes information title, summary and tags;

[0015] The sampling subunit samples samples from the sample set using a category weighting method to obtain the initial sampling probability of the samples, thereby forming an initial sample set. It constructs importance coefficients and sampling probabilities, and then samples the initial data sample set using the importance coefficients and sampling probabilities to obtain the sampled dataset.

[0016] The feature processing unit extracts, transforms, and optimizes data features of different data types in the sample set to obtain information features, user features, and other relevant features.

[0017] As one possible implementation method, the importance coefficient and sampling probability are determined in the following way:

[0018] Determine the first impact weight of the time difference between a user's registration time and most recent active time on the user's importance, and the corresponding first weight;

[0019] Determine the second influence weight of the time difference between a user's first click on information and the most recent click on information on the user's importance, and the corresponding second weight;

[0020] And determine the third influence weight of the difference between the user's first average browsing time and the most recent average browsing time on the user's importance, and the corresponding third weight;

[0021] Then, the importance coefficient is determined, and the sampling probability is obtained through the importance coefficient.

[0022] As one possible implementation method, the initial sampling probability is specifically:

[0023] The importance coefficient is specifically:

[0024] The sampling probability is expressed as follows:

[0025] in, These represent the user's registration time and most recent active time, respectively. These represent the time of the user's first click on the news item and the time of their most recent click on the news item, respectively. These represent the weights of the impact of time differences on user importance; This indicates the weight of the impact of duration differences on user importance; These represent the average duration of a user's first browsing session and the average duration of their most recent browsing session, respectively. These represent the corresponding weights; Represents the first sample in the initial sample set. One sample, =1... k=0 represents a negative sample, and k=1 represents a positive sample. Let represent the initial sampling probability of the k-th sample in the i-th iteration. Let k be the initial number of samples. For frequency, This represents the number of k samples during the i-th sampling.

[0026] As one possible implementation, the original relevant data includes offline data and real-time data. The offline data is used to characterize the user's offline preferences, and the real-time data is used to characterize the user's real-time preferences.

[0027] User behavior data includes browsing behavior data, click behavior data, and purchase behavior data; the original information may take the form of at least one or more of the following: text, images, audio, and video.

[0028] As one possible implementation method, information feature vectors are extracted through a large model, including:

[0029] Obtain word segmentation of the news text, and obtain feature vector representations of words in different dimensions of the news text based on the LLM large model;

[0030] Calculate the attention score for each feature vector, and then transform the score into a probability distribution using the Softmax function;

[0031] The word embeddings in the information text are summed based on attention weights and regularization terms to obtain the information feature vector, specifically:

[0032] in, Represents a regular term, Represents the information feature vector. Represents a probability distribution. This represents the eigenvector representation, where n represents the number of probability distributions and i represents the i-th probability distribution. Represents the regularity coefficient. A matrix representing the probability distributions.

[0033] As one possible implementation, it also includes extracting user feature vectors through a large model, specifically:

[0034] User behavior sequences are formed based on raw information, and user behavior sequence feature vectors are formed from information feature vectors.

[0035] The user feature vector is obtained through max pooling, average pooling, and min pooling, specifically as follows: ;

[0036] in, , Represents a feature vector of user behavior sequence. They represent the weights, This represents the user feature vector.

[0037] As one possible implementation method, the data type includes at least: user characteristics, information characteristics and other related characteristics, and other related characteristics include product characteristics, sequence characteristics, cross characteristics and contextual characteristics;

[0038] User characteristics include demographic characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user feature vectors;

[0039] Information features include information attribute features and information feature vectors;

[0040] Product characteristics include product abbreviation, product type, product risk, product company, product manager, product theme, product scale, and product vector characteristics;

[0041] Sequence features include browsing behavior, click behavior, transaction behavior, collection behavior, and forwarding behavior;

[0042] Cross features are composite features generated through feature combination and feature cross processing;

[0043] Contextual features include the currently recommended news topic and the currently recommended time period.

[0044] As one possible implementation method, the recall unit includes at least one or more of collaborative filtering, preference recall, vector recall, popularity recall, and latest recall;

[0045] The filtering unit includes at least exposure filtering and rule filtering;

[0046] Ranking units include logistic regression, XGBoost, machine learning, and deep learning;

[0047] Rearrangement units include data insertion, data fragmentation, content strategy, and business strategy.

[0048] An information recommendation method based on a large model includes the following steps:

[0049] Acquire relevant data, preprocess the dataset to obtain a processed dataset, wherein the relevant data includes user-related data and information-related data;

[0050] The dataset is subjected to sample selection and feature processing to obtain the final sample set, information features, and user features.

[0051] The information recommendation pre-training model is trained based on the final sample set, information features, and user features to obtain the information recommendation model. The information recommendation model is then used for inference to obtain the recommendation list. The inference process is as follows: in response to the trigger request from the front-end user, the recommendation engine API is called, which in turn calls the recall unit. The recall unit recalls the information recommendation training model and obtains the recommendation results. The results are then filtered by the filtering unit and passed to the ranking unit. The ranking unit sorts the results and passes them to the rearrangement unit. The rearrangement unit rearranges the results to form the recommendation list.

[0052] The recommendation engine API is used to push a list of recommended news items to the front end and display the list to the user.

[0053] This invention, by adopting the above technical solutions, has significant technical effects:

[0054] This invention utilizes large-scale model technology for feature engineering, effectively addressing the problem of insufficient content and user representation in recommendation systems. It introduces the semantic parsing capabilities of large-scale models to deeply extract more content features, enabling deeper content and user understanding. This improves the deep representation of content and information, thereby enhancing the accuracy and diversity of model reasoning, allowing the model to better adapt to the personalized needs of different users.

[0055] This invention introduces a method for sample expansion based on the large model generation capability. By enriching the sample data, it effectively solves the recommendation accuracy problem caused by model samples and improves the accuracy of information recommendation systems. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of the overall structure of the system of the present invention;

[0058] Figure 2 This is a schematic diagram of the technical architecture of the present invention;

[0059] Figure 3 This is a schematic diagram of data acquisition according to the present invention;

[0060] Figure 4 This is a schematic diagram of sample expansion according to the present invention;

[0061] Figure 5 This is a schematic diagram of the enhancement process of the present invention;

[0062] Figures 6-8 This is a schematic diagram of the feature processing of the present invention;

[0063] Figure 9 This is a schematic diagram of the recommendation system process of this invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to the embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.

[0065] Example 1:

[0066] An information recommendation system based on a large model, such as Figure 1 As shown, it includes:

[0067] The data acquisition module 100 acquires raw relevant data and preprocesses the raw relevant data to obtain a processed dataset. The raw relevant data includes user basic information, raw information, and user behavior data.

[0068] The feature engineering module 200 performs sample selection and feature processing on the processed dataset to obtain the final sample set, information features, user features and other relevant features;

[0069] The information recommendation module 300 trains a pre-trained information recommendation model based on the final sample set, information features, user features, and other relevant features to obtain an information recommendation model. The information recommendation model then performs inference to obtain a recommendation list. Specifically, in response to a trigger request from a front-end user, the recommendation engine API is called, which in turn calls the recall unit. The recall unit recalls the information recommendation training model and obtains the recommendation results. The results are then filtered by the filtering unit and passed to the sorting unit. The sorting unit sorts the results and passes them to the rearrangement unit. The rearrangement unit rearranges the sorted results to form the recommendation list.

[0070] The application recommendation module 400 pushes a list of recommended information to the front end and displays the list to the user via the recommendation engine API.

[0071] Figure 2 This is a schematic diagram of the technical architecture for implementing a recommendation system.

[0072] In one embodiment, the raw relevant data includes offline data and real-time data. Offline data is used to characterize users' offline preferences, and real-time data is used to characterize users' real-time preferences. User behavior data includes browsing behavior data, click behavior data, and purchase behavior data. The raw information is in at least one or more of the following formats: text, images, audio, and video. See the appendix for a detailed data collection diagram. Figure 3 As shown.

[0073] This process allows for in-depth exploration of the potential value of data from different modalities, laying a solid foundation for building a recommendation system that supports multimodal information and meeting users' personalized recommendation needs in diverse information consumption scenarios.

[0074] The relevant data used in this invention includes user-related data and information-related data. User-related data includes user data and behavioral data. Behavioral data includes user stock behavior data and user information behavior data. Information-related data includes information data and sequence data. The sequence data includes information sequence data and individual stock sequence data. The information sequence data is information click records and / or information browsing records within a certain time period. The individual stock sequence data is the records of held stocks and / or selected stocks within a certain time period. User features are obtained based on the relevant data, information features and information text tags are obtained based on the information-related data, and a cross-feature set is formed based on the user features and information text tags.

[0075] The raw relevant data includes basic user information, various information data, and user behavior data such as browsing, clicking, and purchasing on the platform. The raw relevant data includes offline data and real-time data. Offline data can deeply analyze users' long-term behavioral patterns to depict users' offline preferences. Real-time data can capture users' immediate dynamics in a timely manner to accurately depict users' real-time preferences.

[0076] The collected information data has multimodal features, covering various forms such as text, images, audio, and video. In actual operation, non-text information data is first converted into text features through a large model. For example, images, audio, and video are converted into text and then text features are extracted. Then, the information features are uniformly processed based on the text features to form information features.

[0077] Considering the differences in information representation and feature extraction among different modal data, we will leverage the understanding and processing capabilities of large language models to conduct targeted feature processing for various modal data.

[0078] In another embodiment, the feature engineering module includes a sample screening unit and a feature processing unit, wherein the sample screening unit includes a sample enhancement subunit and a sampling subunit;

[0079] The sample enhancement subunit mines and expands the original information through the LLM big model to generate diversified information; the LLM big model forms user preferences based on the user's past interaction history; the LLM big model generates synthetic information related to the new user's interests based on the new user's historical data, where the original information includes information title, summary and tags;

[0080] The sampling subunit samples samples from the sample set using a category weighting method to obtain the initial sampling probability of the samples, thereby forming an initial sample set. It constructs importance coefficients and sampling probabilities, and then samples the initial data sample set using the importance coefficients and sampling probabilities to obtain the sampled dataset.

[0081] The feature processing unit extracts, transforms, and optimizes data features of different data types in the sample set to obtain information features, user features, and other relevant features.

[0082] In other words, the entire feature engineering module includes sample selection and feature processing;

[0083] The sample screening module selects representative samples based on screening criteria to ensure the high quality and validity of the data.

[0084] The feature processing module uses professional algorithms and technologies to extract, transform, and optimize data features for different types of data in order to uncover the potential value behind the data.

[0085] These two modules work closely together, providing a solid data foundation and feature support for the subsequent modeling work of the recommendation system, which is an important prerequisite for achieving accurate recommendations.

[0086] In building an information recommendation system, sufficient data is the cornerstone of model training, ensuring its effectiveness and stability. However, in reality, most brokerage apps have relatively limited user behavior data, making it difficult to meet the ideal model training requirements. This undoubtedly negatively impacts the model's accuracy, reduces its generalization ability across different scenarios, and hinders its ability to accurately predict various user behaviors and preferences.

[0087] In the training data, the appropriate ratio of positive to negative samples is crucial to the model's performance. However, in some brokerage apps, the ratio of positive to negative samples is often unreasonable. This can lead to model bias during training, causing an overemphasis on the class with the larger proportion of samples. Consequently, the model's ability to identify and predict minority classes decreases significantly, making it unable to learn data features comprehensively and accurately, ultimately affecting the overall performance of the recommendation system.

[0088] When dealing with insufficient sample size, a large-scale model-based sample augmentation technique is employed. Leveraging the powerful generative capabilities of large models, the limited original samples are reasonably expanded. For users with sparse historical behavior data, similar pseudo-samples are generated based on a small number of behavior samples to expand user feature input, effectively enriching the total sample size and providing sufficient data support for model training, thereby improving the model's accuracy and generalization ability. Specifically, the tagged data of some information that the user has viewed is input into the large language model. The large language model can then generate some "pseudo-information" that the user hasn't seen but might be interested in, based on the information they've viewed. This "pseudo" interaction data is then used as part of the training set. Based on the generated pseudo-samples, user understanding is enhanced, and then more accurate pseudo-samples are generated. The entire process is as follows: Figure 5 As shown.

[0089] The data augmentation of the large model mainly involves information enhancement, generating user interests, and synthesizing new data, as detailed below:

[0090] Information Enhancement: Based on the provided information titles, summaries, and tags, the large model can perform in-depth mining and expansion of the original information, thereby generating information with richer content and more diverse layers.

[0091] Generate user interests: By leveraging users' past interaction history, large language models can accurately extract user preferences;

[0092] Information Synthesis: Due to a lack of interaction data with new users, models struggle to learn effectively. By leveraging a large model and using extremely limited historical data as "raw material," information tailored to the interests of new users is generated, filling the data gap.

[0093] To address the issue of an unreasonable ratio of positive to negative samples, this application employs a sample sampling method.

[0094] By reducing the relative number of negative samples in the training data through sampling, the ratio of positive to negative samples is brought closer to a reasonable level. This avoids the model overfitting to negative samples during training, enhances the model's ability to learn from minority class positive samples, and ensures that the model can comprehensively and accurately capture data features, thereby significantly optimizing the overall performance of the recommendation system.

[0095] In one embodiment, the importance coefficient and sampling probability are determined in the following manner:

[0096] Determine the first impact weight of the time difference between a user's registration time and most recent active time on the user's importance, and the corresponding first weight;

[0097] Determine the second influence weight of the time difference between a user's first click on information and the most recent click on information on the user's importance, and the corresponding second weight;

[0098] And determine the third influence weight of the difference between the user's first average browsing time and the most recent average browsing time on the user's importance, and the corresponding third weight;

[0099] Then, the importance coefficient is determined, and the sampling probability is obtained through the importance coefficient.

[0100] The initial sampling probability is specifically as follows:

[0101]

[0102] The importance coefficient is specifically:

[0103]

[0104] The sampling probability is expressed as follows:

[0105] in, These represent the user's registration time and most recent active time, respectively. These represent the time of the user's first click on the news item and the time of their most recent click on the news item, respectively. These represent the weights of the impact of time differences on user importance; This indicates the weight of the impact of duration differences on user importance; These represent the average duration of a user's first browsing session and the average duration of their most recent browsing session, respectively. These represent the corresponding weights; Represents the first sample in the initial sample set. One sample, =1... k=0 represents a negative sample, and k=1 represents a positive sample. Let represent the initial sampling probability of the k-th sample in the i-th iteration. Let k be the initial number of samples. For frequency, This represents the number of k samples during the i-th sampling.

[0106] This sampling method can solve the problem of imbalance between positive and negative samples, thus making the subsequent modeling results more accurate.

[0107] In the feature processing module, large language modeling (LLM) is used to perform user and information understanding, constructing user profiles and information profiles, and completing real-time and offline feature processing tasks. Simultaneously, in-depth mining and analysis are conducted on sequence features and cross-features. The data types involved in the entire information recommendation system include at least: user features, information features, and other related features. Other related features include product features, sequence features, cross-features, and contextual features. User features include demographic features, user device features, user activity features, user information preference features, user stock preference features, and user feature vectors. Information features include information attribute features and information feature vectors. Product features include product abbreviation, product type, product risk, product company, product manager, product theme, product size, product net value, product price fluctuation data for different time periods, and product vector features. Sequence features include browsing behavior, click behavior, transaction behavior, collection behavior, and forwarding behavior. Cross-features are composite features generated through feature combination and cross-feature processing. Contextual features include the currently recommended information theme and the currently recommended time period.

[0108] The following characteristics are obtained through the above series of operations, as shown in Table 1:

[0109] Table 1 shows examples of features.

[0110]

[0111] In the process of feature processing, such as Figures 7-8 As shown, information feature vectors are extracted through a large model, including:

[0112] Obtain word segmentation of the news text, and obtain feature vector representations of words in different dimensions of the news text based on the LLM large model;

[0113] Calculate the attention score for each feature vector, and then transform the score into a probability distribution using the Softmax function;

[0114] The word embeddings in the information text are summed based on attention weights and regularization terms to obtain the information feature vector.

[0115] The detailed calculation process is as follows:

[0116] The information text is segmented and used as input. The input is:

[0117]

[0118] This represents words in the information text, which are then semantically analyzed by a large model. The embedding vector is represented as follows:

[0119]

[0120] This indicates the dimension of the embedding vector, which defaults to 100 here.

[0121] Next, calculate the attention score for each word, as shown below:

[0122]

[0123] in, , These are trainable parameters;

[0124] The attention score for each word is transformed into a probability distribution using the Softmax function, as shown below:

[0125]

[0126] The word embeddings are weighted according to attention weights and summed with regularization terms to obtain the final information embedding:

[0127]

[0128] In other words, n represents the number of probability distributions, and i represents the i-th probability distribution. Represents a regular term, Represents the regularity coefficient. The matrix representing the probability distribution is used to prevent overfitting when calculating information embedding, thereby enhancing the information features.

[0129] In addition, extracting information feature vectors through large models also includes:

[0130] User behavior sequences are formed based on raw information, and user behavior sequence feature vectors are formed from information feature vectors.

[0131] The user feature vector is obtained through max pooling, average pooling, and min pooling.

[0132]

[0133] in, , , , Represents a feature vector of user behavior sequence. They represent the weights, This represents the user feature vector.

[0134] The detailed calculation process is as follows:

[0135] The input is represented as:

[0136] in, Indicates the first Based on the previously calculated information embedding, we can obtain the following information:

[0137]

[0138] This invention employs a hybrid pooling approach to calculate user embedding features, combining max pooling, average pooling, and min pooling to balance salient features with detail preservation, as shown below:

[0139]

[0140] in, , , , Represents a feature vector of user behavior sequence. They represent the weights, This represents a user feature vector. Based on these weighting formulas, if there are significantly high activation regions in the vector (i.e., ... If the value is relatively large, then enhance max pooling (i.e., let the value be larger). Increase); if the feature distribution is uniform (i.e. If the pooling is dominant, then average pooling is enhanced, which allows for better calculation of user characteristics.

[0141] Finally, as Figure 9 As shown, in the entire information recommendation system, the recall unit includes at least one or more of the following: collaborative filtering, preference recall, vector recall, popularity recall, and latest recall; the filtering unit includes at least exposure filtering and rule filtering; the ranking unit includes logistic regression, XGBoost, machine learning, and deep learning; and the re-ranking unit includes data imputation, data shuffling, content strategy, and business strategy.

[0142] This invention establishes an information recommendation system, see [link to relevant documentation]. Figure 9 As shown, the system integrates multi-path recall, filtering, ranking, and reordering algorithms, enabling users to independently configure various algorithm strategies based on the actual needs of different business scenarios. Simultaneously, the platform also integrates real-time event tracking data and supports A / B testing. By comparing the recommendation effects under different strategies, it accurately selects the most suitable solution, ultimately achieving real-time personalized recommendation management. With this plug-and-play design, the platform can not only meet diverse business needs but also continuously optimize recommendation strategies based on market feedback, providing users with more tailored personalized recommendation services. In summary, this information recommendation system has broad applicability, supporting multiple business touchpoints, including apps, WeChat Work, mini-programs, PCs, and official accounts, facilitating accurate information delivery across various business scenarios.

[0143] Example 2:

[0144] An information recommendation method based on a large model includes the following steps:

[0145] Acquire relevant data, preprocess the dataset to obtain a processed dataset, wherein the relevant data includes user-related data and information-related data;

[0146] The dataset is subjected to sample selection and feature processing to obtain the final sample set, information features, and user features.

[0147] The information recommendation pre-training model is trained based on the final sample set, information features, and user features to obtain the information recommendation model. The information recommendation model is then used for inference to obtain the recommendation list. The inference process is as follows: in response to the trigger request from the front-end user, the recommendation engine API is called, which in turn calls the recall unit. The recall unit recalls the information recommendation training model and obtains the recommendation results. The results are then filtered by the filtering unit and passed to the ranking unit. The ranking unit sorts the results and passes them to the rearrangement unit. The rearrangement unit rearranges the results to form the recommendation list.

[0148] The recommendation engine API is used to push a list of recommended news items to the front end and display the list to the user.

[0149] Various changes and modifications made without departing from the spirit and scope of this invention, and all equivalent technical solutions, also fall within the scope of this invention.

[0150] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0151] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0155] It should be noted that:

[0156] The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0157] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.

Claims

1. An information recommendation system based on a large model implementation, characterized in that, include: The data acquisition module acquires raw relevant data and preprocesses the raw relevant data to obtain a processed dataset. The raw relevant data includes user basic information, raw information, and user behavior data. The feature engineering module performs sample selection and feature processing on the dataset to obtain the final sample set, information features, user features, and other relevant features. The information recommendation module trains a pre-trained information recommendation model based on the final sample set, information features, user features, and other relevant features to obtain an information recommendation model. The information recommendation model then performs inference to obtain a recommendation list. Specifically, in response to a user's trigger request, the recommendation engine API is called, which in turn calls the recall unit. The recall unit recalls the information recommendation training model and obtains the recommendation results. These results are then filtered by the filtering unit and passed to the ranking unit. The ranking unit sorts the results and passes them to the rearrangement unit. The rearrangement unit rearranges the results to form the recommendation list. The application recommendation module pushes a list of recommended information to the front end and displays the list to the user via the recommendation engine API. The feature engineering module includes a sample screening unit and a feature processing unit. The sample screening unit includes a sample enhancement subunit and a sampling subunit. The sampling subunit samples samples from the sample set using a category weighting method to obtain the initial sampling probability of the samples, thereby forming an initial sample set. It constructs importance coefficients and sampling probabilities, and then samples the initial data sample set using the importance coefficients and sampling probabilities to obtain the sampled dataset. The importance coefficient and sampling probability are determined in the following way: Determine the first impact weight of the time difference between a user's registration time and most recent active time on the user's importance, and the corresponding first weight; Determine the second influence weight of the time difference between a user's first click on information and the most recent click on information on the user's importance, and the corresponding second weight; And determine the third influence weight of the difference between the user's first average browsing time and the most recent average browsing time on the user's importance, and the corresponding third weight; Then, the importance coefficient is determined, and the sampling probability is obtained through the importance coefficient.

2. The large model-based information recommendation system of claim 1, wherein, The sample enhancement subunit mines and expands the original information through the LLM big model to generate diversified information; the LLM big model forms user preferences based on the user's past interaction history; the LLM big model generates synthetic information related to the new user's interests based on the new user's historical data, wherein the original information includes information title, summary and tags; The feature processing unit extracts, transforms, and optimizes data features of different data types in the sample set to obtain information features, user features, and other relevant features.

3. The information recommendation system based on a large model according to claim 1, characterized in that, The initial sampling probability, in particular is: The importance coefficient, in particular is: The sampling probability is expressed as follows: wherein, respectively represent the registration time and the last active time of the user, respectively represent the first click information time and the last click information time of the user; respectively represent the influence weight of the time difference on the importance of the user; represent the influence weight of the time length difference on the importance of the user; respectively represent the first average browsing time length and the last average browsing time length of the user, respectively represent the corresponding weights; represent the initial sample set, the k-th sample, =1... , k=0 represents a negative sample, and k=1 represents a positive sample, represent the initial sampling probability of the k-th sample, is the initial number of k samples, is the frequency.

4. The large model implementation-based information recommendation system according to claim 1, characterized in that, The original relevant data includes offline data and real-time data. Offline data is used to characterize users' offline preferences, and real-time data is used to characterize users' real-time preferences. User behavior data includes browsing behavior data, click behavior data, and purchase behavior data; the original information may take the form of at least one or more of the following: text, images, audio, and video.

5. The information recommendation system based on a large model according to claim 1, characterized in that, Information feature vectors are extracted using large models, including: Obtain word segmentation of the news text, and obtain feature vector representations of words in different dimensions of the news text based on the LLM large model; Calculate the attention score for each feature vector, and then transform the score into a probability distribution using the Softmax function; The word embeddings in the information text are summed based on attention weights and regularization terms to obtain the information feature vector, specifically: in, Represents a regular term, Represents the information feature vector. Represents a probability distribution. This represents the eigenvector representation, where n represents the number of probability distributions and i represents the i-th probability distribution. Represents the regularity coefficient. A matrix representing the probability distributions.

6. The information recommendation system based on a large model according to claim 5, characterized in that, This also includes extracting user feature vectors through large models, specifically: User behavior sequences are formed based on raw information, and user behavior sequence feature vectors are formed from information feature vectors. The user feature vector is obtained through max pooling, average pooling, and min pooling, specifically as follows: ; in, , Represents a feature vector of user behavior sequence. They represent the weights, This represents the user feature vector.

7. The information recommendation system based on a large model according to claim 1, characterized in that, Data types should include at least: user characteristics, information characteristics, and other relevant characteristics. Other relevant characteristics include product characteristics, sequence characteristics, cross characteristics, and contextual characteristics. User characteristics include demographic characteristics, user device characteristics, user activity characteristics, user information preference characteristics, and user feature vectors; Information features include information attribute features and information feature vectors; Product characteristics include product abbreviation, product type, product risk, product company, product manager, product theme, product scale, and product vector characteristics; Sequence features include browsing behavior, click behavior, transaction behavior, collection behavior, and forwarding behavior; Cross features are composite features generated through feature combination and feature cross processing; Contextual features include the currently recommended news topic and the currently recommended time period.

8. The information recommendation system based on a large model according to claim 1, characterized in that, The recall unit includes at least one or more of the following: collaborative filtering, preference recall, vector recall, popularity recall, and latest recall. The filtering unit includes at least exposure filtering and rule filtering; Ranking units include logistic regression, XGBoost, machine learning, and deep learning; Rearrangement units include data insertion, data fragmentation, content strategy, and business strategy.

9. An information recommendation method based on a large model, characterized in that, Includes the following steps: Acquire relevant data, preprocess the dataset to obtain a processed dataset, wherein the relevant data includes user-related data and information-related data; The dataset is subjected to sample selection and feature processing to obtain the final sample set, information features, and user features. The information recommendation pre-training model is trained based on the final sample set, information features, and user features to obtain the information recommendation model. The information recommendation model is then used for inference to obtain the recommendation list. The inference process is as follows: in response to the trigger request from the front-end user, the recommendation engine API is called, which in turn calls the recall unit. The recall unit recalls the information recommendation training model and obtains the recommendation results. The results are then filtered by the filtering unit and passed to the ranking unit. The ranking unit sorts the results and passes them to the rearrangement unit. The rearrangement unit rearranges the results to form the recommendation list. The recommendation engine API is used to push the information recommendation list to the front end and display the recommendation list to the user. The sample screening and feature processing include a sample screening unit and a feature processing unit. The sample screening unit includes a sample enhancement subunit and a sampling subunit. The sampling subunit samples samples from the sample set using a category weighting method to obtain the initial sampling probability of the samples, thereby forming an initial sample set. It constructs importance coefficients and sampling probabilities, and then samples the initial data sample set using the importance coefficients and sampling probabilities to obtain the sampled dataset. The importance coefficient and sampling probability are determined in the following way: Determine the first impact weight of the time difference between a user's registration time and most recent active time on the user's importance, and the corresponding first weight; Determine the second influence weight of the time difference between a user's first click on information and the most recent click on information on the user's importance, and the corresponding second weight; And determine the third influence weight of the difference between the user's first average browsing time and the most recent average browsing time on the user's importance, and the corresponding third weight; Then, the importance coefficient is determined, and the sampling probability is obtained through the importance coefficient.

Images