Information recommendation method, device, and storage medium

Abstract

The application discloses an information recommendation method and device, equipment and a storage medium, and belongs to the technical field of recommendation. The method comprises the following steps: performing information recall and rough sorting based on the main interest of a user object to obtain a first information set; performing information recall based on the long-tail interest of the user object, reserving a corresponding number of information from the recall result according to a preset proportion quota to obtain a second information set; performing fine sorting on the mixed first information set and second information set to obtain a third information set; and recommending at least one piece of information in the third information set to the user object. The scheme effectively alleviates the problem of interest solidification of the user object, can more accurately explore and meet the extensive and potential interest preferences of the user object, enhances the recommendation effect, and improves the recommendation efficiency.

Description

Technical Field

[0001] This application relates to the field of recommendation technology, and in particular to an information recommendation method, device and storage medium. Background Technology

[0002] The recommender system employs a multi-stage funnel architecture of "recall-ranking." First, it rapidly recalls a candidate set from the information base. Then, it performs initial screening through coarse ranking. Finally, a fine-ranking model scores the candidates to generate the final recommendation list, aiming to predict and satisfy the mainstream preferences of the target user group as accurately as possible. However, it suffers from problems such as fixed user interests and insufficient recommendation diversity.

[0003] Currently, to address these issues, the industry typically employs traditional diversity optimization strategies, forcibly dispersing results through rules during the post-processing stage. However, this rule-based, hard-intervention approach fails to satisfy the potential and broad interests of users, resulting in poor recommendation performance and low efficiency. Summary of the Invention

[0004] This application provides an information recommendation method, device, and storage medium, which effectively alleviates the problem of fixed user interests and can more accurately discover and satisfy the broad and potential interests and preferences of users, thereby enhancing the recommendation effect and improving recommendation efficiency. The technical solution is as follows: According to one aspect of this application, an information recommendation method is provided, the method comprising: Information retrieval and coarse sorting are performed based on the user's main interests to obtain the first information set; Information retrieval is performed based on the long-tail interests of the user objects, and a corresponding amount of information is retained from the retrieval results according to a preset ratio quota to obtain a second information set. The first information set and the second information set are mixed and then finely sorted to obtain the third information set; Recommend at least one piece of information from the third information set to the user object.

[0005] According to another aspect of this application, an information recommendation device is provided, the device comprising: The first filtering module is used to retrieve and coarsely sort information based on the user's main interests to obtain the first information set. The second filtering module is used to recall information based on the long-tail interests of the user object, and retain a corresponding number of information from the recall results according to a preset ratio quota to obtain a second information set. The third filtering module is used to mix the first information set and the second information set and then perform fine sorting to obtain the third information set. The recommendation module is used to recommend at least one piece of information from the third information set to the user object.

[0006] According to another aspect of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the information recommendation method as described above.

[0007] According to another aspect of this application, a computer-readable storage medium is provided, wherein a computer program is stored therein, the computer program being loaded and executed by a processor to implement the information recommendation method as described above.

[0008] According to another aspect of this application, a computer program product is provided, comprising a computer program executed by a processor to implement the information recommendation method provided in various alternative implementations of the above aspects.

[0009] This application provides an information recommendation scheme that establishes an independent and quota-guaranteed recall and filtering channel for long-tail interests, processing it in parallel with the main interest channel. This breaks the filtering effect of traditional recommendation processes on long-tail content, ensuring that long-tail content can stably enter the fine-ranking stage to participate in the competition. Thus, it improves the diversity of the recommendation list without overly relying on post-processing rules. This scheme effectively alleviates the problem of fixed user interests and can more accurately discover and satisfy the broad and potential interests of users, enhancing recommendation effectiveness and improving recommendation efficiency. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram illustrating the implementation environment of an information recommendation method provided in an embodiment of this application; Figure 2 This is a flowchart of an information recommendation method provided according to an embodiment of this application; Figure 3 This is a flowchart illustrating another information recommendation method provided according to an embodiment of this application; Figure 4 This is a flowchart illustrating another information recommendation method provided according to an embodiment of this application; Figure 5This is a schematic diagram of a recommended audio flow according to an embodiment of this application; Figure 6 This is a schematic diagram of the information recommendation device provided according to the embodiments of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided according to an embodiment of this application.

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0014] To enable those skilled in the art to better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0015] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0016] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the media resources involved in this application were obtained with full authorization.

[0017] The following is an explanation of the terms used in this application.

[0018] Recall refers to the process of quickly and roughly filtering out a set of candidate information (usually a few hundred to a few thousand) from the full information database based on user or scenario characteristics.

[0019] Coarse ranking refers to the process of sorting and filtering the still large candidate set generated in the recall phase to obtain a smaller (usually a few hundred) and higher-quality candidate set.

[0020] Fine-grained ranking refers to the process of using complex models and features to accurately personalize and rank the candidate set produced by coarse ranking, in order to generate a final recommendation list that is directly visible to users and has a minimal number of items (usually a few to dozens).

[0021] Figure 1 This is a schematic diagram illustrating the implementation environment of an information recommendation method provided in an embodiment of this application. See also... Figure 1 The implementation environment specifically includes: computer device 101 and server 102. Computer device 101 can be connected to server 102 via wireless network or wired network.

[0022] Computer device 101 can be at least one of the following: smartphone, smartwatch, desktop computer, laptop, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), and laptop computer. Computer device 101 can have an application installed and running. This application is compatible with the information platform and covers scenarios such as music playback, online singing, live streaming, on-demand video (both short and long), news, and information, and can be in the form of a standalone app or a mini-program. This application is associated with server 102, which provides background services to computer device 101.

[0023] Computer device 101 can refer to one of a plurality of computer devices. This embodiment uses computer device 101 as an example. Those skilled in the art will know that the number of the above-mentioned computer devices can be more or less. For example, there can be several, dozens or hundreds, or more computer devices. This application embodiment does not limit the number or type of computer devices.

[0024] Server 102 can be at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. Optionally, the number of servers can be more or less, and this embodiment does not limit this. Of course, server 102 may also include other functional servers to provide more comprehensive and diversified services. In some embodiments, server 102 undertakes the main computing work, and computer device 101 undertakes the secondary computing work; or, server 102 undertakes the secondary computing work, and computer device 101 undertakes the main computing work; or, server 102 and computer device 101 collaborate on computing using a distributed computing architecture. Server 102 can be connected to computer device 101 and other computer devices via a wireless network or a wired network. Optionally, the number of servers can be more or less, and this embodiment does not limit this.

[0025] Figure 2 This is a flowchart of an information recommendation method provided according to an embodiment of this application, such as... Figure 2 As shown, this method is executed by a computer device, which can be... Figure 1 The method, implemented via a client application in the computer device 101, includes the following steps: Step 201: Based on the user's main interests, perform information retrieval and coarse sorting to obtain the first information set.

[0026] In this embodiment, the primary interest of a user object refers to the preferences identified by analyzing the user object's long-term, stable, and high-frequency historical behavioral data (such as playback, favorites, searches, etc.). Primary interests can be represented by tags or categories, such as "Chinese pop music," "rock music," or "technology news." The user object can be a user account, which represents the user.

[0027] Optionally, the recommendation system uses an offline computing platform to analyze the user's long-term historical behavioral data, such as playback and search records from the past 90 days, to calculate the user's preference intensity scores across various interest dimensions (such as music genres, video categories, and product categories), forming a stable long-term interest profile. Based on this long-term interest profile, the user's primary interests can be identified. For example, the top N interest tags in each category with scores exceeding a certain threshold can be selected as the primary interests.

[0028] For example, the long-term interest profile is {"Chinese pop": 95, "rock music": 85, "electronic music": 15}. If N=2 and the threshold=80, then the identified primary interests are "Chinese pop" and "rock music".

[0029] Information can represent different content in different recommendation scenarios. For example, in an audio / music recommendation scenario, information may represent songs, podcasts, audiobooks, audio programs, radio channels, etc. In a video recommendation scenario, information may represent short videos, long videos, movies, TV series, live streams, etc. In an e-commerce recommendation scenario, information may represent products, stores, services, coupons, etc. In a content and information recommendation scenario, information may represent news, articles, posts, pictures, columns, etc. In an advertising recommendation scenario, information may represent image and text ads, video ads, product promotion links, or promotional activities, etc. In an application distribution and game recommendation scenario, information may represent mobile applications, games, software tools, or game skins, etc. This application's embodiments do not impose any limitations on this.

[0030] Information retrieval refers to filtering information from a database based on the user's primary interests, selecting information relevant to those interests, and using this candidate information set as input for the coarse-sorting stage.

[0031] Coarse ranking refers to using a relatively lightweight and efficient model to initially score and rank the candidate information set obtained in the recall stage, and then selecting hundreds of the highest quality information to significantly reduce the size of the candidate set.

[0032] The first information set refers to the candidate information set that represents the mainstream preferences of the user group after the path processing of recall and coarse ranking.

[0033] Optionally, when a user initiates a request, the recommendation system immediately retrieves all information entries with or highly related to the user's primary interest tag (such as "Chinese pop music") from the entire database, forming an initial selection set of several thousand items. Then, the recommendation system inputs this initial selection set into a lightweight ranking model. This ranking model can comprehensively consider user characteristics, information popularity, contextual features, etc., to quickly score and rank these several thousand pieces of information, and extract the top few hundred. These selected few hundred pieces of information constitute the "first information set" representing the user's mainstream interests.

[0034] Step 202: Based on the long-tail interests of the user objects, information is recalled, and a corresponding amount of information is retained from the recall results according to a preset ratio quota to obtain the second information set.

[0035] In this embodiment of the application, the long-tail interests of the user object refer to niche, personalized, and sparse interests other than the main interests.

[0036] Optionally, the recommendation system obtains a long-tail interest list of the user group. This list is dynamically updated and may include recently appeared interests with low scores.

[0037] Information retrieval refers to the operation of quickly retrieving relevant content from the entire information database based on long-tail interests. It should be noted that, since long-tail interests themselves are sparse data, the retrieval strategy can differ from that of the main interest, such as relying on content tags, knowledge graphs, or embedding similarity.

[0038] The preset quota ratio is a rule set in advance to implement a quantity preservation strategy. The purpose is to ensure that a certain share of information belonging to long-tail interests is reserved in the subsequent screening results.

[0039] Optionally, the recommendation system uses tags representing long-tail interests as search keys to retrieve an initial selection set (e.g., thousands of related audio files) from the full database. The system then calls a preset "proportional quota" (e.g., long-tail content should account for 5% of the final ranked candidates) to calculate a specific retention quantity N. Finally, it performs a quick filter or sort (e.g., lightweight quality filtering or relevance ranking) and strictly adheres to the quantity N, retaining the top N pieces of information. The final output set of information is the set composed of these retained N pieces of information.

[0040] Step 203: Mix the first information set and the second information set and then perform fine sorting to obtain the third information set.

[0041] In this embodiment, the first information set is a candidate set representing the mainstream preferences of the user group. The second information set is a candidate set representing the long-tail, emerging, or potential preferences of the user group. By simply merging these two candidate sets from different sources and with different characteristics into a larger mixed candidate set, the mixed candidate set is input into the fine-ranking model. This model does not distinguish the source of the information but scores each piece of information uniformly based on all available information for the current user group. Optionally, some information from the first information set may receive higher scores due to its high relevance to the current scenario. Some information from the second information set may receive higher scores than expected due to its content quality, novelty, or simply matching the user group's momentary interests, thus ranking higher. The fine-ranking model sorts all items in descending order based on their scores and typically selects the top few dozen (e.g., Top 50) to form the final ordered third information set.

[0042] Step 204: Recommend at least one piece of information from the third information set to the user.

[0043] In this embodiment of the application, the information contained in the third information set obtained by the above steps is recommended to the user.

[0044] This application provides an information recommendation scheme that establishes an independent and quota-guaranteed recall and filtering channel for long-tail interests, processing it in parallel with the main interest channel. This breaks the filtering effect of traditional recommendation processes on long-tail content, ensuring that long-tail content can stably enter the fine-ranking stage to participate in the competition. Thus, it improves the diversity of the recommendation list without overly relying on post-processing rules. This scheme effectively alleviates the problem of fixed user interests and can more accurately discover and satisfy the broad and potential interests of users, enhancing recommendation effectiveness and improving recommendation efficiency.

[0045] The above Figure 2 The diagram shows the main flow of an information recommendation method provided in this application. The process of obtaining the second information set based on long-tail interests in this information recommendation scheme will be further explained below. Figure 3 This is a flowchart illustrating another information recommendation method provided according to an embodiment of this application. The method is executed by a computer device, or by the computer device through a client of an installed application, such as... Figure 3 As shown, the method includes: Step 301: Based on the user's main interests, perform information retrieval and coarse sorting to obtain the first information set.

[0046] In this embodiment of the application, this step is the same as step 201 above. Please refer to step 201 above, and it will not be repeated here.

[0047] Step 302: Retrieve information based on the long-tail interests of the user objects to obtain a long-tail information set.

[0048] In this embodiment, the recommendation system can acquire real-time behavior sequences of users outside the recommendation scenario and quantify and score them based on positive and negative behaviors within these sequences. If the interest score of a certain interest category changes from 0 to a positive number for the first time, that interest category is activated as a long-tail interest. After activation, information under that interest category can participate in the recall and ranking process, thus having the opportunity to be recommended to users.

[0049] In some embodiments, the long-tail interest is determined as follows: First, the user's interaction behavior sequence outside the recommendation scenario is obtained. The interaction behavior sequence includes positive and negative interaction behaviors, and different types of interaction behaviors correspond to different quantitative scores. Then, for any interest type, starting from the moment when the user first interacts with the interest type, the interest score of the interest type is calculated based on the interaction behaviors between the user and the interest type in the interaction behavior sequence. Finally, if the interest score of the interest type is greater than zero, the interest type is determined as a long-tail interest.

[0050] In this context, "outside of the recommendation scenario" refers to user behavior in product modules outside the main information flow (such as the "Recommendations" page), such as actions in the "Favorites" section, editing in the "Playlist" section, and active searches. These behaviors have more explicit intentions.

[0051] The interaction sequence is a log of user behavior arranged chronologically. Each behavior is tagged with its type (e.g., favorite, play, skip) and object (e.g., song ID). Different behavior types are assigned weights reflecting the strength of the user's preference. For example: favorite (+10 points) > play (+8 points) > skip (-3 points). The scoring design reflects the information content of the behavior. Favorites and play are positive behaviors, while skipping is a negative behavior.

[0052] For any interest type, the recommendation system finds the moment T when a user first interacts with that type, that is, the moment of the first interaction with that type within a certain time range. Starting from moment T, the system accumulates the quantitative scores corresponding to all behaviors belonging to that interest type that occur after moment T, obtaining the interest score S for that interest type. When S first changes from 0 to a positive number greater than 0 (for example, when a user listens to a song completely for the first time and receives +8 points), the system determines that the interest type is an activated long-tail interest. Once the interest type is activated, information retrieval is based on information about the user's actual positive behaviors under that interest (such as adding to favorites or completing playback). For example, the SASRec time-series model can be used, taking the user's behavior sequence (such as recent behavior sequence) as input to predict "the next song they might like." Alternatively, the DSSM deep semantic matching model can be used, matching the vector of information activating the interest type (the semantic representation obtained through learning) with the information vector of the entire database to find similar information. Songs of that interest type will then be recommended to the user.

[0053] It should be noted that, in addition to recommendation scenarios, user data from other scenarios can be integrated, such as social interactions, geolocation information, device models (obtained through authorization), and interest tags imported through cross-platform authorization, to construct a more comprehensive user interest profile. This profile allows for the identification of more diverse long-tail interests. Accordingly, user interest profiles are constructed based on multi-source data outside of recommendation scenarios. This multi-source data includes at least one of the following: social interaction data, geolocation data, device information data, and interest tag data imported through cross-platform authorization. Based on these user interest profiles, the user's primary and long-tail interests are determined.

[0054] Social interaction data can be obtained by extracting user behavior, constructing relationship graphs between users and information, or using graph embedding techniques to obtain user social vectors. Geographic location data includes precise coordinates provided by GPS or fuzzy locations inferred from IP addresses or base station information. It should be noted that geographic location data needs to be anonymized or aggregated; raw, precise coordinates that pinpoint an individual are not directly stored or used. Device information data can be obtained through standard interfaces provided by the operating system or browser, acquiring a set of non-resettable, non-personally identifiable device parameters, including device brand, model, operating system, and version. Cross-platform authorized data import involves obtaining limited data explicitly agreed to be shared by the user after they actively click and confirm, including publicly available nicknames, avatars, and preference lists on authorized social media platforms.

[0055] Optionally, unsupervised clustering algorithms or knowledge graph embedding techniques can be used to identify the main interests and long-tail interests of user subjects from the interest profile. That is, the identification of long-tail interests can use unsupervised clustering (such as K-Means, GMM) or knowledge graph embedding techniques to automatically discover and define the main interests and sparse interests (long-tail interests) of user subjects, thereby achieving more dynamic and fine-grained interest segmentation.

[0056] First, principal component analysis or model-based feature importance assessment is used to select the most effective feature subset for interest discrimination from the profile, in order to remove noise, reduce computational complexity, and improve subsequent clustering results.

[0057] Then, the feature vectors of the user profiles (or the features of all user objects) are input into algorithms such as K-Means and GMM for clustering. Each cluster center can be regarded as an "interest prototype", and objects within the cluster share similar interests. Alternatively, user objects, information, and entities (such as singers and locations) can be embedded into a knowledge graph, and potential interest association clusters of user objects can be discovered through graph reasoning and path mining.

[0058] Then, for each cluster of interests derived from clustering or graph inference, calculate the intensity score of the user's membership in that cluster (e.g., the inverse of the distance to the cluster center, the sum of graph relation weights). This yields a list of quantified scores for all user interests.

[0059] Finally, the top-N interests with the highest ratings exceeding the absolute threshold are selected as primary interests. All other interests with ratings greater than zero but not selected as primary interests, or interests with low ratings but belonging to newly emerging clusters, are classified as long-tail interests.

[0060] In some embodiments, when the user object is a cold-start object, a deep interest network (DIN or DIEN) based on the behavior of highly active user objects can be pre-trained. Then, the feature extraction layer of this model is transferred to the cold-start user object model, fine-tuned with only a small amount of new data, to quickly generate a high-quality user object representation. The user object's primary and long-tail interests are then determined using this representation. Specifically, the primary and long-tail interests are determined as follows: limited behavioral data of the cold-start object is input into the feature extraction layer of the pre-trained interest network model to obtain an initial representation of the object. The pre-trained interest network model is trained based on the behavioral data of highly active objects. Based on this initial representation, the primary and long-tail interests of the cold-start object are determined.

[0061] In training the interest network model, a large amount of long-term, rich, and diverse historical behavioral data (such as click, favorite, and view sequences) of highly active objects is first acquired. Then, the model learns, through extensive training data, how to map the original object behavior sequences (a string of item IDs and timestamps) into a high-quality "object representation vector" that summarizes its interest preferences. During training, the model's feature extraction layer learns how to extract general patterns useful for judging interests from behavior. Next, limited behavioral data of cold-start objects is acquired. This limited behavioral data is input into the trained model, using only the model's feature extraction layer. Since the feature extraction layer is already a mature feature understander, it can transform the limited and sparse behavioral features of these new objects into a high-information-density initial object representation that resides in the same semantic space as the representation of highly active objects. Finally, based on this initial object representation, the potential primary and long-tail interests of the cold-start object can be predicted.

[0062] In some embodiments, during the recall phase, a Variational Autoencoder (VAE) or Generative Adversarial Network (GAN) can be used to generate "virtual" information feature vectors that match the user's potential preferences based on a small number of existing interest points. These "virtual" information vectors are then used to retrieve real information from a vectorized database, thereby discovering unexpected information. Correspondingly, information recall based on the user's long-tail interests to obtain a long-tail information set includes: inputting the user's long-tail interests into a pre-trained generative model to obtain at least one virtual information feature vector; the pre-trained generative model being a Variational Autoencoder or a Generative Adversarial Network; calculating the similarity between at least one virtual information feature vector and the feature vectors of information entities in the information database; and recalling the corresponding information entities from the information database based on the calculated similarity to obtain the long-tail information set.

[0063] Variational Autoencoders (VAEs) can learn to compress information feature vectors into a low-dimensional, continuous latent space and decode (generate) a reasonable information feature vector from any point in this space. Their latent space has good interpolation properties, and smoothly transitioning points correspond to reasonable information. Generative Adversarial Networks (GANs) are trained adversarially between a generator and a discriminator. The generator learns to generate information feature vectors from random noise sufficient to "deceive" the discriminator, thus learning to capture the distribution of real information feature vectors. Optionally, the generative model uses the user's long-tail interest vector and random noise vector as common inputs to generate diverse virtual information feature vectors. The long-tail interest representation of the current user object (which can be an embedding vector of an interest label or a fusion of a few related information feature vectors) is input into the pre-trained generative model. In VAEs, the interest representation can be used as the mean, with a small amount of random noise added to form a latent variable, and then the decoder generates a vector. In GANs, the interest representation can be concatenated with random noise and input into the generator to generate a vector.

[0064] Step 303: Determine the first quantity based on the preset ratio quota.

[0065] In this embodiment, a pre-defined quantitative rule is used to ensure the exposure of information belonging to long-tail interests. The preset proportion quota is a "proportion" or "share" that defines the portion of the output that long-tail content should occupy. For example, the first quantity in this step is the specific number (N) of information items retained from the "long-tail information set".

[0066] Optionally, the preset proportional quota is any one of the following: A fixed percentage means that the long tail set is truncated at a fixed ratio, regardless of its size.

[0067] The ratio is dynamically adjusted based on the total number of long-tail information sets, that is, the ratio is intelligently adjusted based on the size of the long-tail information sets to avoid insufficient truncation when the set is too small.

[0068] The proportion is dynamically adjusted based on the total number of items in the first information set, so that the long-tail information and the mainstream information maintain a relatively stable structural ratio.

[0069] The proportion is dynamically adjusted based on the user object state, which includes at least one of real-time behavior sequence and interest distribution. That is, the user object state can describe the characteristics of the user's current real-time context, including real-time behavior sequence (recent clicks, what was searched) and real-time interest distribution (the current activity level of each interest).

[0070] Step 304: Extract the first amount of information from the long-tail information set to obtain the second information set.

[0071] In this embodiment, firstly, all information in the long-tail information set is quickly scored and sorted. Then, according to a first quantity N, the top N pieces of information are extracted from the sorted list, starting from the highest score. This is the Top-N selection strategy. The extracted N pieces of information constitute the second information set. The information in this second information set is the most relevant batch of information after sorting from the long-tail information set.

[0072] Alternatively, the criteria for quick scoring and sorting can be any of the following: Sort by relevance score, that is, score and sort according to the degree of matching between the information and the interest tags that trigger the recall.

[0073] Sort by coarse ranking score, which is based on the estimated scores (such as CTR estimate) given by a simplified ranking model.

[0074] The overall quality score is a score calculated by combining factors such as the freshness, popularity, and production quality of the information, and the information is ranked based on this score.

[0075] Sort by model score, that is, by matching score calculated using a lightweight DSSM dual-tower model specifically optimized for long-tail information.

[0076] This application does not impose any limitations on this.

[0077] Step 305: Mix the first information set and the second information set and then perform fine sorting to obtain the third information set.

[0078] In this embodiment of the application, this step is the same as step 203 above. Please refer to step 203 above, and it will not be repeated here.

[0079] Step 306: Recommend at least one piece of information from the third information set to the user.

[0080] In this embodiment of the application, this step is the same as step 204 above. Please refer to step 204 above, and it will not be repeated here.

[0081] It should be noted that this application embodiment illustrates the use of a volume-preserving strategy during the coarse-ranking stage as an example. Optionally, the volume-preserving intervention can be extended from the single coarse-ranking stage to the entire recommendation chain. For example, after recall, a fixed number of slots can be reserved for long-tail interests to directly advance to the coarse-ranking stage, or after fine-ranking, information that ranks lower but belongs to long-tail interests can be weighted and ranked higher. This application embodiment does not impose any limitations on this.

[0082] This application provides an information recommendation scheme that establishes an independent and quota-guaranteed recall and filtering channel for long-tail interests, processing it in parallel with the main interest channel. This breaks the filtering effect of traditional recommendation processes on long-tail content, ensuring that long-tail content can stably enter the fine-ranking stage to participate in the competition. Thus, it improves the diversity of the recommendation list without overly relying on post-processing rules. This scheme effectively alleviates the problem of fixed user interests and can more accurately discover and satisfy the broad and potential interests of users, enhancing recommendation effectiveness and improving recommendation efficiency.

[0083] The above Figure 2 The diagram shows the main flow of an information recommendation method provided in this application. Before recommending information in a third information set, the third information set can be rearranged to include information on long-tail interests. The process of rearranging the third information set in this information recommendation scheme is explained below. Figure 4 This is a flowchart illustrating another information recommendation method provided according to an embodiment of this application. The method is executed by a computer device, such as... Figure 4 As shown, the method includes: Step 401: Based on the user's main interests, perform information retrieval and coarse sorting to obtain the first information set.

[0084] In this embodiment of the application, this step is the same as step 201 above. Please refer to step 201 above, and it will not be repeated here.

[0085] Step 402: Based on the long-tail interests of the user objects, information is recalled, and a corresponding amount of information is retained from the recall results according to a preset ratio quota to obtain the second information set.

[0086] In this embodiment of the application, this step is the same as step 202 above. Please refer to step 202 above, and it will not be repeated here.

[0087] Step 403: Mix the first information set and the second information set and then perform fine sorting to obtain the third information set.

[0088] In this embodiment of the application, this step is the same as step 203 above. Please refer to step 203 above, and it will not be repeated here.

[0089] Step 404: Obtain the interest scores of each long-tail interest of the user object.

[0090] In this embodiment of the application, for any long-tail interest, the interest score of the long-tail interest is obtained based on the user object's interaction behavior with information belonging to the long-tail interest.

[0091] Interest score is a numerical metric used to measure the real-time strength or activity of a user's preference for a specific long-tail interest. A higher score generally indicates a greater current interest in that interest. Interest score is related to interactive behaviors; positive interactions increase the interest score, while negative interactions decrease it. Examples include: saving, playing in full, liking, sharing, searching, clicking, skipping, and blocking.

[0092] Optionally, the recommendation system continuously acquires all user interactions. When an action (such as adding a song to favorites) occurs, the system attributes that action to its corresponding interest category. For each interest category, the system maintains a first action timestamp T_first in the background. When the system first detects that a user has generated any interaction with that category, it records that moment as T_first. For that interest category, the system only accumulates the quantitative scores corresponding to all user interactions belonging to that category after the T_first boundary point. For example, adding to favorites (+10 points), playing the entire song (+8 points), searching (+5 points), skipping (-2 points). For each identified long-tail interest, the system outputs a dynamically updated interest score calculated according to the above rules.

[0093] Step 405: Determine the rearrangement probability based on the sum of interest scores for each long-tail interest. The rearrangement probability is positively correlated with the sum.

[0094] In this embodiment, this step provides a data-driven decision-making mechanism to determine whether the ranking results need to be adjusted to explore long-tail interests. That is, while ensuring the recommendation results are relevant to the main interest, information matching the long-tail interests of the user is introduced.

[0095] To achieve the above objectives, a quantitative aggregation operation is first performed, summing the interest scores of all long-tail interests of the currently identified user to obtain a total value A. The total value A is an aggregate indicator; its value directly reflects the user's recent exploration activity and preferences beyond mainstream interests. A higher value indicates more frequent interaction with various long-tail interests and stronger interest signals, meaning the user has recently shown greater interest in long-tail interests.

[0096] Optionally, the recommender system does not directly compare the total value A with a fixed threshold. Instead, it inputs the total value A into a probability function, mapping it to a continuous probability value between 0 and 1, which is the reorder probability. That is, the reorder probability can be obtained by mapping the total value A through a Sigmoid function. The Sigmoid function has the following characteristics: the larger the value of A, the closer the output probability is to 1; the smaller the value of A, the closer the output probability is to 0. In other words, when the input total value A changes within a specific range, the output probability P exhibits a smooth S-shaped curve change. Specifically, when the total interest score A is low, the function output probability P approaches 0, indicating that the recommender system should maintain the stability of the ranking results; as the value of A increases, the value of P begins to grow rapidly and non-linearly; when the value of A is high enough, the value of P will approach 1 infinitely, indicating that a reorder exploration is recommended. Parameters a (steepness) and b (center point) are used to control the sensitivity and trigger threshold of the function. See formula (1) below.

[0097] (1).

[0098] Here, A represents the total interest score. 'a' controls the steepness of the curve, i.e., the sensitivity of probability P to changes in the total score A; a larger 'a' value results in a sharper decision boundary. 'b' defines the total interest score threshold corresponding to a probability of 0.5, i.e., sets the baseline for triggering exploration. By adjusting these two parameters, the recommendation system can flexibly adapt to the requirements of exploration aggressiveness and stability in different business scenarios.

[0099] Step 406: If the third information set is rearranged based on the rearrangement probability, then the target long-tail interest is selected based on the interest score of each long-tail interest. The probability of each long-tail interest being selected as the target long-tail interest is positively correlated with the interest score of each long-tail interest.

[0100] In this embodiment of the application, after deciding on the rearrangement, a long-tail interest needs to be selected as the target for exploration. Here, exploration refers to the attempt to recommend information about this long-tail interest to the target user.

[0101] Optionally, the recommendation system calculates a probability distribution using the Softmax function based on the scores of each long-tail interest. Interests with higher scores have a greater probability of being selected. See formula (2) below.

[0102] (2).

[0103] in, It is the interest score of the i-th long-tail interest. It is a temperature parameter. The larger the value, the smoother the probability distribution, and the more likely long-tail interests with low scores will be selected; the smaller the value, the sharper the probability distribution, and the system will almost certainly select the interest with the highest score. It is the probability that the i-th long-tail interest is selected.

[0104] The selection is based on the probability of each long-tail interest being selected, and the selected ones are taken as the target long-tail interests.

[0105] Step 407: Insert information belonging to the target long-tail interest into the third information set to update the third information set.

[0106] In this embodiment, the recommendation system can find one or more pieces of information that belong to the target long-tail interest and are of the highest quality (e.g., have the highest score in the ranking) from the candidate pool (such as the second information set). This information is then inserted into the third information set to update the third information set.

[0107] In some embodiments, inserting information belonging to the target long-tail interest into the third information set includes: obtaining first information from the ranking results, wherein the first information is the information belonging to the target long-tail interest and ranked highest in the ranking results; and inserting the first information at a preset position in the third information set.

[0108] The preset position can be a fixed position, such as always inserting it in the 2nd or 3rd position of the third information set. The insertion method can be direct insertion, which adds the first information to the third information set. Alternatively, it can be replacement insertion, which replaces the information in the corresponding position in the third information set while keeping the total number of information in the third information set unchanged.

[0109] Step 408: If it is determined based on the rearrangement probability that the third information set will not be rearranged, then recommend at least one piece of information from the third information set to the user.

[0110] In this embodiment of the application, if the information is not rearranged, the information in the third information set can be directly recommended to the user.

[0111] It should be noted that even if the recommendation system has decided to start the reordering exploration, if the condition that no further insertion is needed is met, the recommendation system will stop the insertion operation and directly use the fine-ranking result.

[0112] Optionally, if the third information set is rearranged based on the rearrangement probability, and the third information set contains information belonging to the target long-tail interest, then the step of recommending at least one piece of information from the third information set to the user object is performed.

[0113] In this scenario, when the recommendation system detects that the third information set already contains information belonging to the target long-tail interest, it indicates that the ranking model has recognized the value of that interest. Inserting homogeneous content at this point would not only fail to increase diversity but could also disrupt the list's coherence and alienate the target user. Therefore, this insertion is abandoned, and the original ranking judgment is retained.

[0114] Optionally, if the third information set is rearranged based on the rearrangement probability, and the second information set does not include information on the target long-tail interests, then the step of recommending at least one piece of information from the third information set to the user object is performed.

[0115] If no information in the second information set (i.e., the candidate information set after quantitative intervention of long-tail interest information) belongs to the target long-tail interest, then the recommendation system lacks information for insertion. Therefore, no insertion operation needs to be performed.

[0116] Optionally, if the third information set is rearranged based on the rearrangement probability, and the proportion of information belonging to long-tail interests in the third information set reaches a preset proportion, then the step of recommending at least one piece of information from the third information set to the user is executed.

[0117] When the proportion of long-tail interest information in the third information set reaches or exceeds a preset ratio, it means that the diversity of the current list has been fully satisfied, or the quota allocated to long-tail interests has been exhausted. Continuing to insert more information at this point would bring minimal benefits and would even reduce user satisfaction; therefore, insertion is abandoned.

[0118] It should be noted that the embodiments of this application provide two formulas for deciding whether to explore (i.e., recommend) long-tail interests and which long-tail interests to explore. However, it is not limited to this.

[0119] In some embodiments, the decision-making mechanism can abandon fixed formula calculations and adopt a reinforcement learning model (such as Contextual Bandit) as the "exploration decision-making brain." This reinforcement learning model uses the user's state (real-time behavior sequence, interest distribution) as context to dynamically decide whether to explore, which interest category to explore, and what content to insert, aiming to maximize long-term user satisfaction (such as dwell time, diversity rating) through online learning. The decision-making is performed by the reinforcement learning model trained on it.

[0120] This application provides an information recommendation scheme that establishes an independent and quota-guaranteed recall and filtering channel for long-tail interests, processing it in parallel with the main interest channel. This breaks the filtering effect of traditional recommendation processes on long-tail content, ensuring that long-tail content can stably enter the fine-ranking stage to participate in the competition. Thus, it improves the diversity of the recommendation list without overly relying on post-processing rules. This scheme effectively alleviates the problem of fixed user interests and can more accurately discover and satisfy the broad and potential interests of users, enhancing recommendation effectiveness and improving recommendation efficiency.

[0121] To make the information recommendation scheme provided in this application easier to understand, an example of audio information will be used for illustration. See [link to relevant documentation]. Figure 5 As shown, Figure 5 This is a schematic diagram illustrating a recommended audio flow according to an embodiment of this application. For example... Figure 5 As shown, the process includes the following steps: 501. Obtain the offline profile of the user object. This refers to the long-term interest profile determined by the offline computing platform in step 201. 502. Obtain the real-time behavior sequence of the user object, such as liking, favorited, and skipping songs. 503. Obtain multiple long-tail interests and calculate the interest score for each long-tail interest. 504. Perform information retrieval based on the user object's main interests to obtain retrieval results. 505. Perform coarse ranking on the above retrieval results to obtain a first information set. 506. Perform information retrieval based on the user object's long-tail interests and preserve the quantity of retrieval results to obtain a second information set. 507. Perform fine ranking on the first and second information sets to obtain a third information set. 508. Re-rank the third information set. Step 508 includes the following five sub-steps: 5081. Determine whether to re-rank based on the sum of the interest scores of each long-tail interest. 5082. Determine the target long-tail interest based on the interest scores of each long-tail interest. 5083. Determine if the third information set contains information about the target long-tail interest. 5084. If not, insert at the preset position. 5085. If yes, cancel the insertion.

[0122] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user data. These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their data is being collected. This ensures that the application only begins the steps for collecting user data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without user confirmation), the steps for collecting user data end, meaning no user data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0123] It should be noted that the order of the method steps provided in the embodiments of this application can be appropriately adjusted, and the steps can also be added or removed as appropriate. Any method variations that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application, and therefore will not be elaborated further.

[0124] Figure 6This is a schematic diagram of the information recommendation device provided according to an embodiment of this application. For example... Figure 6 As shown, the device includes: a first screening module 601, a second screening module 602, a third screening module 603, and a recommendation module 604.

[0125] The first filtering module 601 is used to perform information retrieval and coarse sorting based on the user's main interests to obtain a first information set; The second filtering module 602 is used to recall information based on the long-tail interests of the user object, and retain a corresponding number of information from the recall results according to a preset ratio quota to obtain a second information set. The third filtering module 603 is used to mix the first information set and the second information set and then perform fine sorting to obtain the third information set. Recommendation module 604 is used to recommend at least one piece of information from a third information set to the user.

[0126] In some embodiments, the second filtering module 602 is used to perform information retrieval based on the long-tail interests of user objects to obtain a long-tail information set; determine a first quantity based on a preset ratio quota; and extract the first quantity of information from the long-tail information set to obtain a second information set.

[0127] In some embodiments, the second filtering module 602 is used to input the user's long-tail interests into a pre-trained generative model to obtain at least one virtual information feature vector. The pre-trained generative model is a variational autoencoder or a generative adversarial network. The similarity between the at least one virtual information feature vector and the feature vector of the information entity in the information database is calculated respectively. Based on the calculated similarity, the corresponding information entity is recalled from the information database to obtain a long-tail information set.

[0128] In some embodiments, the preset proportional quota is any one of the following: Fixed percentage; The proportion is dynamically adjusted based on the total number of long-tail information sets. The proportion is dynamically adjusted based on the total quantity of the first information set; The proportion is dynamically adjusted based on the user object state, which includes at least one of real-time behavior sequence and interest distribution.

[0129] In some embodiments, the recommendation module 604 is further configured to obtain the interest scores of each long-tail interest of the user object, wherein, for any long-tail interest, the interest score of the long-tail interest is obtained based on the user object's interaction behavior with information belonging to the long-tail interest; based on the sum of the interest scores of each long-tail interest, a reordering probability is determined, and the reordering probability is positively correlated with the sum; if it is determined to reorder the third information set based on the reordering probability, then based on the interest scores of each long-tail interest, a target long-tail interest is selected, and the probability of each long-tail interest being selected as the target long-tail interest is positively correlated with the interest score of each long-tail interest; and information belonging to the target long-tail interest is inserted into the third information set to update the third information set.

[0130] In some embodiments, the recommendation module 604 is further configured to obtain first information from the ranking results, wherein the first information is the information in the ranking results that belongs to the target long-tail interest and has the highest ranking; and insert the first information at a preset position in the third information set.

[0131] In some embodiments, the recommendation module 604 is further configured to implement at least one of the following: If it is determined based on the rearrangement probability that the third information set will not be rearranged, then the step of recommending at least one piece of information from the third information set to the user object is executed. If the third information set is rearranged based on the rearrangement probability, and the third information set contains information belonging to the target long-tail interest, then the step of recommending at least one piece of information from the third information set to the user object is executed. If the third information set is rearranged based on the rearrangement probability, and the second information set does not include information on the target long-tail interests, then the step of recommending at least one piece of information from the third information set to the user object is executed. If the third information set is rearranged based on the rearrangement probability, and the proportion of information belonging to long-tail interests in the third information set reaches a preset ratio, then the step of recommending at least one piece of information from the third information set to the user is executed.

[0132] In some embodiments, the second filtering module 602 is further configured to obtain the user object's interaction behavior sequence outside the recommendation scenario, the interaction behavior sequence including positive interaction behavior and negative interaction behavior, and different types of interaction behavior correspond to different quantitative scores; For any interest type, starting from the moment when the user object first interacts with the interest type, the interest score of the interest type is calculated based on the interaction behavior between the user object and the interest type in the interaction behavior sequence. If the interest score of an interest type is greater than zero, then the interest type is identified as a long-tail interest.

[0133] In some embodiments, the second filtering module 602 is further configured to construct an object interest profile based on multi-source data of the user object outside the recommendation scenario, wherein the multi-source data includes at least one of social interaction data, geographic location data, device information data, and interest tag data imported through cross-platform authorization; and determine the user object's main interests and long-tail interests based on the object interest profile.

[0134] In some embodiments, the second filtering module 602 is further configured to identify the main interests and long-tail interests of user objects from the interest profile using unsupervised clustering algorithms or knowledge graph embedding techniques.

[0135] In some embodiments, the user object is a cold-start object with sparse behavioral data; the second filtering module 602 is further configured to input the limited behavioral data of the cold-start object into the feature extraction layer of a pre-trained interest network model to obtain an initial representation of the object, the pre-trained interest network model being trained based on the behavioral data of highly active objects; and to determine the main interest and long-tail interest of the cold-start object based on the initial representation of the object.

[0136] This application provides an information recommendation device that establishes an independent and quota-guaranteed recall and filtering channel for long-tail interests, processing it in parallel with the main interest channel. This breaks the filtering effect of traditional recommendation processes on long-tail content, ensuring that long-tail content can stably enter the fine-ranking stage to participate in the competition. Thus, it improves the diversity of the recommendation list without overly relying on post-processing rules. This solution effectively alleviates the problem of fixed user interests and can more accurately discover and satisfy the broad and potential interests of users, enhancing recommendation effectiveness and improving recommendation efficiency.

[0137] It should be noted that the information recommendation device provided in the above embodiments is only an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the information recommendation device and the information recommendation method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0138] Embodiments of this application also provide a computer device including a processor and a memory, wherein the memory stores a computer program that is loaded and executed by the processor to implement the information recommendation method provided in the above-described method embodiments.

[0139] Figure 7 This is a schematic diagram of the structure of a computer device provided according to an embodiment of this application.

[0140] Computer device 700 includes a central processing unit (CPU) 701, a system memory 704 including random access memory (RAM) 702 and read-only memory (ROM) 703, and a system bus 705 connecting the system memory 704 and the CPU 701. Computer device 700 also includes a basic input / output system (I / O system) 706 that facilitates information transfer between various devices within the computer device, and a mass storage device 707 for storing the operating system 713, application programs 714, and other program modules 715.

[0141] The basic input / output system 706 includes a display 708 for displaying information and an input device 709 for user input, such as a mouse or keyboard. Both the display 708 and the input device 709 are connected to the central processing unit 701 via an input / output controller 710 connected to the system bus 705. The basic input / output system 706 may also include the input / output controller 710 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 710 also provides output to a display screen, printer, or other types of output devices.

[0142] Mass storage device 707 is connected to central processing unit 701 via a mass storage controller (not shown) connected to system bus 705. Mass storage device 707 and its associated computer-readable storage media provide non-volatile storage for computer device 700. That is, mass storage device 707 may include computer-readable storage media (not shown) such as hard disk or compact disc read-only memory (CD-ROM) drive.

[0143] Without loss of generality, computer-readable storage media can include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technique for storing information such as computer-readable storage instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically-erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage devices, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 704 and mass storage device 707 described above can be collectively referred to as memory.

[0144] The memory stores one or more programs, which are configured to be executed by one or more central processing units 701. The one or more programs contain instructions for implementing the above method embodiments, and the central processing unit 701 executes the one or more programs to implement the methods provided by the various method embodiments described above.

[0145] According to various embodiments of this application, the computer device 700 can also be connected to a remote computer device on a network, such as the Internet, for operation. That is, the computer device 700 can be connected to a network 712 via a network interface unit 711 connected to the system bus 705, or the network interface unit 711 can be used to connect to other types of networks or remote computer device systems (not shown).

[0146] The memory also includes one or more programs stored in the memory, and the one or more programs include steps performed by a computer device in the methods provided in the embodiments of this application.

[0147] This application also provides a computer-readable storage medium storing a computer program that is loaded and executed by a processor to implement the information recommendation method provided in the above-described method embodiments.

[0148] This application also provides a computer program product, which includes a computer program executed by a processor to implement the information recommendation method provided in the above-described method embodiments.

[0149] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0150] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent switching, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An information recommendation method, characterized in that, The method includes: Information retrieval and coarse sorting are performed based on the user's main interests to obtain the first information set; Information retrieval is performed based on the long-tail interests of the user objects, and a corresponding amount of information is retained from the retrieval results according to a preset ratio quota to obtain a second information set. The first information set and the second information set are mixed and then finely sorted to obtain the third information set; Recommend at least one piece of information from the third information set to the user object.

2. The method according to claim 1, characterized in that, The information retrieval based on the long-tail interests of the user object, and the retention of a corresponding amount of information from the retrieval results according to a preset proportion quota, to obtain a second information set, including: Information retrieval is performed based on the long-tail interests of the user objects to obtain a long-tail information set; Based on the preset ratio quota, a first quantity is determined; The first amount of information is extracted from the long-tail information set to obtain the second information set.

3. The method according to claim 1 or 2, characterized in that, The information retrieval based on the long-tail interests of the user object yields a long-tail information set, including: The user's long-tail interests are input into a pre-trained generative model to obtain at least one virtual information feature vector. The pre-trained generative model is a variational autoencoder or a generative adversarial network. Calculate the similarity between the feature vector of the at least one virtual information and the feature vector of the information entity in the information database; Based on the calculated similarity, the corresponding information entities are retrieved from the information database to obtain the long-tail information set.

4. The method according to any one of claims 1-3, characterized in that, The preset ratio quota is any one of the following: Fixed percentage; The proportion is dynamically adjusted based on the total number of the long-tail information set. The ratio is dynamically adjusted based on the total number of items in the first information set; The proportion is dynamically adjusted according to the user object state, which includes at least one of real-time behavior sequence and interest distribution.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain the interest scores of each long-tail interest of the user object, wherein, for any long-tail interest, the interest score of the long-tail interest is obtained based on the user object's interaction behavior with information belonging to the long-tail interest; The rearrangement probability is determined based on the sum of the interest scores of each long-tail interest, and the rearrangement probability is positively correlated with the sum. If the third information set is rearranged based on the rearrangement probability, then a target long-tail interest is selected based on the interest scores of each long-tail interest. The probability that each long-tail interest is selected as the target long-tail interest is positively correlated with the interest scores of each long-tail interest. Information belonging to the target long-tail interest is inserted into the third information set to update the third information set.

6. The method according to claim 5, characterized in that, The insertion of information belonging to the target long-tail interest into the third information set includes: Obtain first information from the ranking results, wherein the first information is the information in the ranking results that belongs to the target long-tail interest and has the highest ranking; The first information is inserted at a preset position in the third information set.

7. The method according to claim 5 or 6, characterized in that, The method further includes at least one of the following: If it is determined based on the rearrangement probability that the third information set will not be rearranged, then the step of recommending at least one piece of information from the third information set to the user object is executed; If the third information set is rearranged based on the rearrangement probability, and the third information set contains information belonging to the target long-tail interest, then the step of recommending at least one piece of information from the third information set to the user object is executed. If it is determined based on the rearrangement probability that the third information set is rearranged, and the second information set does not include information about the target long-tail interest, then the step of recommending at least one piece of information from the third information set to the user object is executed. If the third information set is rearranged based on the rearrangement probability, and the proportion of information belonging to long-tail interests in the third information set reaches a preset proportion, then the step of recommending at least one piece of information from the third information set to the user object is executed.

8. The method according to any one of claims 1 to 7, characterized in that, The method further includes: Obtain the user's interaction behavior sequence outside the recommendation scenario. The interaction behavior sequence includes positive interaction behavior and negative interaction behavior. Different types of interaction behavior correspond to different quantitative scores. For any interest type, starting from the moment when the user object first interacts with the interest type, the interest score of the interest type is calculated based on the interaction behavior between the user object and the interest type in the interaction behavior sequence. If the interest score of the interest type is greater than zero, then the interest type is identified as a long-tail interest.

9. The method according to any one of claims 1 to 8, characterized in that, The method further includes: Based on the user object's multi-source data outside of the recommendation scenario, construct the object's interest profile. The multi-source data includes at least one of social interaction data, geographic location data, device information data, and cross-platform authorized import interest tag data. Based on the object interest profile, the user object's primary interest and long-tail interest are determined.

10. The method according to claim 9, characterized in that, The step of determining the user object's primary interest and long-tail interest based on the object's interest profile includes: The main interests and long-tail interests of the user object are identified from the interest profile using unsupervised clustering algorithms or knowledge graph embedding techniques.

11. The method according to any one of claims 1 to 10, characterized in that, The user object is a cold-start object with sparse behavioral data; The method further includes: The limited behavioral data of the cold-start object is input into the feature extraction layer of the pre-trained interest network model to obtain the initial representation of the object. The pre-trained interest network model is trained based on the behavioral data of highly active objects. Based on the initial characterization of the object, the primary interest and long-tail interest of the cold-start object are determined.

12. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the information recommendation method as described in any one of claims 1 to 11.

13. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program, which is loaded and executed by a processor to implement the information recommendation method as described in any one of claims 1 to 11.

14. A computer program product, characterized in that, The computer program product includes a computer program executed by a processor to implement the information recommendation method as described in any one of claims 1 to 11.

Images