An item recommendation method based on aging interest modeling

Abstract

A kind of item recommendation method based on time interest modeling, comprising the following steps: S1, collecting user identity information, user's action behavior information for the content of the article on the network and the information of the article content;S2, the time element contained in the article content is discriminated and extracted;S3, construct material-material authorized isomorphism graph and material low-dimensional vector representation based on time and interest association;S4, the real-time behavior of user is respectively matched according to time effectiveness, wherein time effectiveness behavior is directly matched with material vector, and non-time effectiveness behavior is represented as linear weighting of time effectiveness behavior by random walk method with time bias, and then matched with material vector;S5, for the user group with similar historical behavior, the new hot article content with high time effectiveness is pushed by vector matching.This application can accurately push the article content that meets the real-time preference of user, and reduce the blindness of push.

Description

Technical Field

[0001] This invention relates to data mining and recommendation technologies, and in particular to a method for recommending items based on time-sensitive interest modeling. Background Technology

[0002] With the increasing scale, level, and quality of mass consumption, the core objective and challenge of building current product recommendation systems lies in achieving real-time and accurate delivery of massive amounts of product content (such as food, clothing, and cosmetics) to a vast number of users. The existing technical approach to this is to first mine pure interest-based relationships between product content from massive user historical behavior sequences, and then match and push personalized products related to the interests of users who exhibit new behaviors. However, because this process neglects users' objective preference for the timeliness and novelty of products (such as holidays, limited-time promotions, new stores and products, and joint promotions), coupled with the inherent cold start and sparse association characteristics of highly timely content, the system's recommendation results often focus on past content with rich user behavior, making it difficult to reflect users' interest preferences for novel and timely content.

[0003] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings of the aforementioned background technology and provide a method for recommending items based on time-sensitive interest modeling.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for recommending items based on time-sensitive interest modeling includes the following steps:

[0007] S1. User behavior collection: Collect user identity information, user action behavior information on online items, and information about the items;

[0008] S2. Content Timeliness Understanding: Identify and extract the timeliness elements contained in the content of the item;

[0009] S3. Timeliness and Interest Mapping and Representation: Construct a material-material weighted identity map and a low-dimensional vector representation of materials based on timeliness and interest association;

[0010] S4. Real-time interest matching: User real-time behavior is divided into time-sensitive categories for matching. Time-sensitive behavior is directly matched with the material vector, while non-time-sensitive behavior is represented by a time-sensitive bias random walk method as a linear weighted sum of time-sensitive behavior before being matched with the material vector.

[0011] S5. New Hot Content Push: For user groups with similar historical behaviors, use vector matching to push timely new hot items and content.

[0012] Furthermore, in step S1, the collected content includes user identity ID, item content ID, item content UGC text, and user browsing or clicking time information.

[0013] Furthermore, step S1 includes collecting and preprocessing the associated data required for offline interest mining and real-time preference prediction. Specifically, for the offline part, all items that the user has browsed and clicked within a set period are collected to construct a material-material rights isomorphism graph. For the real-time part, multiple items that the user has recently clicked and their timestamps are collected and defined as the user's behavior sequence for subsequent use.

[0014] Furthermore, in step S2, the timeliness elements contained in the item content are classified and extracted using a strategy based on a large language model and a strategy based on Chinese word segmentation.

[0015] Further, in step S2, the strategy based on the large language model includes defining a parameterized Prompt and explicit examples of time-sensitive element vocabulary, and asking the large language model whether a UGC content can reflect similar time-sensitive elements; the strategy based on Chinese word segmentation includes using the jieba word segmentation tool to segment the UGC content, and using the distance of each segmented word to the predefined time-sensitive element vocabulary set as a sparse feature, and training a time-sensitive-non-time-sensitive binary classification model to perform time-sensitive discrimination.

[0016] Furthermore, in step S3, timely content and non-time-sensitive content are classified into two categories, and weighted edges are retained for two materials that are clicked more than a set number of times at the same time in the isomorphic graph. Through a random walk method with time-sensitive bias and tidal effect, only timely materials are modeled and low-dimensional vector representations are generated.

[0017] Furthermore, step S3 specifically includes the following steps:

[0018] Step S31: Based on the material timeliness labels obtained from the timeliness understanding, mark the nodes of the offline material-material map obtained by the behavior collection module to distinguish between time-sensitive and non-time-sensitive content, and perform pruning on the side.

[0019] Step S32: Define the timeliness score of each timeliness content as 1 at the initial time t=0, and 0 for the rest;

[0020] Step S33: Starting from all nodes with non-zero timeliness scores, perform timeliness score propagation within the first-order neighborhood, following the following:

[0021]

[0022] Where v i and v j This refers to the material node, where S represents the node's timeliness score at time t. Indicates v i The first-order neighborhood node set; the process repeats multiple time steps and removes material nodes with a final timeliness score of 0;

[0023] Step S34: Starting from each time-delay node, perform multiple random walks of L steps with time-delay bias and tidal effects, where the (i+1)th step node is connected to the previous step node v. i The samples are obtained from the first-order neighborhood of the class according to a specific parameterized distribution, i.e.:

[0024]

[0025] Cat(.) generates a class distribution that satisfies the time-dependent bias and tidal effect, defined as:

[0026]

[0027] T is the maximum number of transmission rounds set. For v i and v j The strength of the interest association between them, m a m is the oversaturation coefficient for aging. i The time-dependent undersaturation coefficient is L, where L is the maximum number of steps in a random walk.

[0028] Step S35: Summarize all random walk sequences originating from each time-sensitive node, remove non-time-sensitive nodes, and construct training positive sample pairs according to the word frequency distribution of time-sensitive nodes. Construct simple negative sample pairs by randomly sampling from time-sensitive materials and construct difficult negative sample pairs by randomly sampling from hypothetical uniform word frequencies. During training, embed each material into a multi-dimensional vector, and perform gradient optimization on the co-occurrence likelihood by setting the ratio of positive samples: simple negative samples: difficult negative samples. After training multiple iterations, take the one with the largest likelihood as the final model, which is the low-dimensional embedding vector of all time-sensitive materials.

[0029] Further, in step S4, the time-sensitive behavior is directly taken from the low-dimensional embedding representation, and the top K related materials are retrieved from the Faiss library by inner product metric using existing material vectors; the non-time-sensitive behavior is constructed as a linear weighted representation of time-sensitive materials using a random walk method with time-sensitive bias before retrieval; preferably, step S4 specifically includes the following steps:

[0030] Step S41: Starting from each non-time-sensitive material node, perform multiple random walks with time-sensitivity bias. The time-sensitivity score directly uses the result of step S33, but the walk process no longer allows the tidal effect to take effect. That is, under the same conditions as the symbol definition in step S34, the category sampling follows:

[0031]

[0032] Step S42: Summarize all random walk sequences originating from each non-time-sensitive node, similarly remove non-time-sensitive nodes, and obtain the time-sensitive node representation set corresponding to the non-time-sensitive nodes. Each element of the set is a time-sensitive node and its word frequency. From the embedding vector set obtained in step S35, retrieve the embedding vectors corresponding to these time-sensitive nodes, and obtain the interpolation embedding result of each non-time-sensitive node by weighting by word frequency.

[0033] Step S43: Represent all user non-time-sensitive behaviors as embedding vectors, and use the Faiss library to retrieve the top K as relevant materials by inner product metric.

[0034] Further, in step S5, the content matched from the user behavior sequence is subjected to exponential decay weighted sorting based on the step size since the beginning of the period; preferably, step S5 specifically includes the following steps:

[0035] Step S51: Extract the most recent Q most clicked item content from a given user and invert the list as a sequence of the user's historical behavior, and perform vector matching on each content;

[0036] Step S52: Weight the matching results according to the positional order of their source index behavior, following:

[0037]

[0038] Where V i It is the ID of the weightless material obtained by matching, p(V) i Z is its normalized recommendation score. i η is the normalization coefficient within the current batch, and η is the exponential decay factor of sequence attention.

[0039] Step S53: Select factor η and execute step S51 to obtain matching material results. Generate the final recommendation list in descending order of score and push it to the user.

[0040] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the item recommendation method based on time-sensitive interest modeling.

[0041] The present invention has the following beneficial effects:

[0042] This invention provides a product recommendation method based on time-sensitive interest modeling, including steps such as user behavior collection, content timeliness understanding, timeliness interest mapping and representation, real-time interest matching, and push of new and trending content. Compared with traditional pure interest recommendation methods, this invention can effectively characterize the fluctuations in user interests caused by time-sensitive factors such as seasonal changes, new stores and products, and promotional activities. Therefore, it accurately pushes product content that matches the user's current real-time preferences, reducing the randomness of pushes and meeting the user's rapid consumption needs for new and trending items. In terms of design, this invention extracts and understands the timeliness characteristics of the product content that the user has browsed in the past, constructs a weighted mapping of materials based on timeliness × interest association, and a low-dimensional vector representation of materials. For user groups with similar historical behaviors, it pushes new and trending product content with high timeliness through vector matching.

[0043] Compared to traditional recommendation methods, this invention can provide an accurate representation of users' real-time interests as they fluctuate with timeliness, and can deliver personalized and accurate information on new and trending items to a massive number of users, effectively meeting users' perception and consumption needs regarding the novelty and timeliness of items.

[0044] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0045] Figure 1 This is a flowchart of an item recommendation method based on time-sensitive interest modeling, according to an embodiment of the present invention.

[0046] Figure 2 This is a block diagram of the item recommendation method based on time-sensitive interest modeling according to an embodiment of the present invention. Detailed Implementation

[0047] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0048] See Figure 1 This invention provides a method for recommending items based on time-sensitive interest modeling, comprising the following steps:

[0049] S1. User behavior collection: Collect user identity information, user action behavior information on online items, and information about the items;

[0050] S2. Content Timeliness Understanding: Identify and extract the timeliness elements contained in the content of the item;

[0051] S3. Timeliness and Interest Mapping and Representation: Construct a material-material weighted identity map and a low-dimensional vector representation of materials based on timeliness and interest association;

[0052] S4. Real-time interest matching: User real-time behavior is divided into time-sensitive categories for matching. Time-sensitive behavior is directly matched with the material vector, while non-time-sensitive behavior is represented by a time-sensitive bias random walk method as a linear weighted sum of time-sensitive behavior before being matched with the material vector.

[0053] S5. New Hot Content Push: For user groups with similar historical behaviors, use vector matching to push timely new hot items and content.

[0054] The items described in this invention include, but are not limited to, various commodities such as food, clothing, cosmetics, and consumer electronics. This invention is particularly suitable for recommending food and clothing.

[0055] This invention presents a product recommendation method based on timeliness-interest modeling. It extracts and understands timeliness characteristics from items viewed historically by users, constructing a weighted composition graph of items based on timeliness and interest associations, and a low-dimensional vector representation of items. This effectively characterizes fluctuations in user interest caused by timeliness factors such as seasonal changes, new products, and promotional activities. For user groups with similar historical behaviors, vector matching is used to accurately push highly timely and popular new items that match the user's current real-time preferences. Compared to traditional recommendation methods, this invention provides an accurate representation of users' immediate interests fluctuating with timeliness, and offers personalized and precise push notifications of popular new items to a massive user base, effectively meeting users' perception and consumption needs regarding item novelty and timeliness.

[0056] The following describes specific embodiments of the present invention.

[0057] like Figure 2 As shown, an item recommendation method based on time-sensitive interest modeling is represented in a modular way, including modules such as user behavior collection, content time-sensitive understanding, time-sensitive interest mapping and representation, real-time interest matching, and push of new and trending content.

[0058] The user behavior collection module collects key information such as user identity ID, item content ID, item content UGC text, and the time when the user browses or clicks.

[0059] The content timeliness understanding module uses two strategies—large language model and Chinese word segmentation—to identify and extract timeliness elements contained in the item content.

[0060] The aforementioned timeliness interest mapping and representation module enables binary classification of timeliness content and non-timeliness content, retains weighted edges for two materials that are clicked more than 10 times simultaneously in the graph, and finally models low-dimensional vector representations only for timeliness materials through random walks with timeliness bias and tidal effects.

[0061] The real-time interest matching module enables users to classify and match real-time behaviors according to their timeliness. Timely behaviors are directly matched with the material pool vector, while non-timely behaviors are represented by a random walk with timeliness bias as a linear weighted sum of timely behaviors before matching.

[0062] The aforementioned new trending content push module performs exponential decay weighted sorting of content matched with user behavior sequences based on the time elapsed since then.

[0063] like Figure 2 As shown, the item recommendation method based on time-sensitive interest modeling in this embodiment of the invention includes modules such as user behavior collection, content time-sensitive understanding, time-sensitive interest mapping and representation, real-time interest matching, and push of new and trending content.

[0064] The user behavior collection module collects key information such as user identity ID, item content ID, UGC text of item content, and user browsing or clicking times. Specifically, it includes collecting and preprocessing the associated data required for offline interest mining and real-time preference prediction. Offline, it collects all item content (materials) viewed and clicked by the user within 31 days to construct an isomorphic graph of material-material pairs. Real-time, it collects the 100 most recently clicked items and their timestamps, defining them as the user's behavior sequence for subsequent use.

[0065] The content timeliness understanding module employs two strategies—a large language model and Chinese word segmentation—to classify and extract timeliness elements contained in item content. The large language model strategy defines a parameterized Prompt and several examples of words explicitly representing timeliness, then asks the language model whether a UGC piece of content can embody similar timeliness elements. The Chinese word segmentation strategy uses jieba to segment the UGC content, and uses the distance of each segmented word to a predefined set of timeliness element words as a sparse feature to train a timeliness-non-timeliness binary classification model for timeliness discrimination.

[0066] The aforementioned timeliness-based interest mapping and representation module performs binary classification of timeliness-based and non-timeliness-based content. It retains weighted edges between two items that are clicked more than 10 times simultaneously in the graph, and models only timeliness-based items using a low-dimensional vector representation through a random walk with timeliness bias and tidal effects. The specific steps are as follows:

[0067] Step S31: Based on the material timeliness labels obtained from the timeliness understanding, perform node annotation on the offline material-material map obtained by the behavior collection module (i.e., distinguish between time-sensitive content and non-time-sensitive content), and prune the edges of materials that users click on simultaneously no more than 10 times to save performance.

[0068] Step S32: Define the timeliness score of each timeliness content as 1 at the initial time t=0, and 0 for the rest.

[0069] Step S33: Starting from all nodes with non-zero timeliness scores, perform timeliness score propagation within the first-order neighborhood, following the following:

[0070] Where v i and v j This refers to the material node, where S represents the node's timeliness score at time t. Indicates v i The first-order neighborhood node set. The process is repeated for 15 time steps, and material nodes with a final timeliness score of 0 are removed.

[0071] Step S34: Starting from each time-delay node, perform 30 random walks of L steps with time-delay bias and tidal effects, where the (i+1)th step node is derived from the previous step node v. i The samples are obtained from the first-order neighborhood of the class according to a specific parameterized distribution, i.e.:

[0072] Cat(.) generates a class distribution that satisfies the time-dependent bias and tidal effect, defined as:

[0073]

[0074] In the above parameters, T=15 is the maximum number of transmission rounds set. For v i and v j The strength of the interest association between them, m a =0.2 is the oversaturation coefficient for aging, m i =0.4 is the undersaturation coefficient, and L=60 is the maximum number of steps in a random walk.

[0075] Step S35: Summarize all random walk sequences originating from each time-sensitive node, remove non-time-sensitive nodes, and construct training positive sample pairs according to the word frequency distribution of time-sensitive nodes. Construct simple negative sample pairs by randomly sampling from time-sensitive materials and construct difficult negative sample pairs by randomly sampling from hypothetical uniform word frequencies. During training, first embed each material into a 16-dimensional vector, and perform gradient optimization on the co-occurrence likelihood according to the ratio of positive samples: simple negative samples: difficult negative samples = 1:100:0.1. After training for 500 iterations, take the one with the largest likelihood as the final model, that is, the low-dimensional embedding vector of all time-sensitive materials.

[0076] The aforementioned real-time interest matching module divides and matches user real-time behavior sequences according to their timeliness. Specifically, time-sensitive behaviors are directly taken from the low-dimensional embedding representation obtained in step S35, and therefore can be directly compared with existing material vectors using the Faiss library to retrieve the top K as relevant materials based on the inner product metric. Non-time-sensitive behaviors are retrieved by constructing a linearly weighted representation of time-sensitive materials through a random walk with a timeliness bias, as detailed below:

[0077] Step S41: Starting from each non-time-sensitive material node, perform 20 random walks with a time-sensitive bias. The time-sensitive score directly uses the final result of step S33, but the walk process no longer allows the tidal effect to take effect. That is, under the same conditions as the symbol definition in step S34, the category sampling follows:

[0078]

[0079] Step S42: Summarize all random walk sequences originating from each non-time-sensitive node, and similarly remove non-time-sensitive nodes to obtain the time-sensitive node representation set corresponding to the non-time-sensitive nodes. Each element in the set is a time-sensitive node and its word frequency. From the embedding vector set obtained in Step S35, retrieve the embedding vectors corresponding to these time-sensitive nodes, and weight them according to word frequency to obtain the interpolation embedding result for each non-time-sensitive node.

[0080] Step S43: Using steps S41 and S42, represent all user non-time-sensitive behaviors as embedding vectors, and similarly use the Faiss library to retrieve the top K as relevant materials by inner product metric.

[0081] The aforementioned new trending content push module performs exponential decay weighted sorting of content matched with user behavior sequences, based on a step size elapsed since the beginning of the year. Specific steps include:

[0082] Step S51: Extract the most recent Q = 100 most clicked item content from the given user and sort it in reverse order as the user's historical behavior sequence. According to step S42, take K = 5 and perform vector matching on each content.

[0083] Step S52: Weight the matching results according to the positional order of their source index behavior, following: Where V i It is the ID of the weightless material obtained by matching, p(V) i Z is its normalized recommendation score. i η is the normalization coefficient within the current batch, and η is the exponential decay factor of sequence attention.

[0084] Step S53: Select 3 factors η∈{0.99,0.9,0.75} and execute step S51 respectively to obtain a maximum of 15×100 matching material results. Generate a final recommendation list in descending order of scores and push it to the user.

[0085] This invention extracts and understands the timeliness characteristics of items viewed in a user's history, constructs a weighted isomorphic graph of items based on timeliness and interest associations, and a low-dimensional vector representation of items. For user groups with similar historical behaviors, vector matching is used to push highly timely and popular new items. Compared to traditional pure interest recommendation systems, this invention effectively characterizes fluctuations in user interests caused by timeliness factors such as seasonal changes, new stores and products, and promotional activities. Therefore, it can accurately push items that match the user's current real-time preferences, reducing the randomness of push notifications and meeting users' rapid consumption needs for new and popular items.

[0086] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0087] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0088] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0089] The storage medium can be implemented by any type of volatile or non-volatile storage device, or a combination thereof. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM). The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0090] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0091] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0092] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0093] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0095] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0096] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0097] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0098] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A method for recommending items based on time-sensitive interest modeling, characterized in that, Includes the following steps: S1. User behavior collection: Collect user identity information, user action behavior information on online items, and information about the items; S2. Content Timeliness Understanding: Identify and extract the timeliness elements contained in the content of the item; S3. Timeliness and Interest Mapping and Representation: Construct a weighted isomorphic graph of materials based on timeliness and interest association and a low-dimensional vector representation of materials; wherein, time-sensitive content and non-time-sensitive content are classified into two categories, and weighted edges are retained for two materials that are clicked more than a set number of times at the same time in the isomorphic graph. Through a random walk method with timeliness bias and tidal effect, only time-sensitive materials are modeled and a low-dimensional vector representation is generated. S4. Real-time Interest Matching: User real-time behaviors are divided into time-sensitive categories for matching. Time-sensitive behaviors are directly matched with material vectors, while non-time-sensitive behaviors are represented as linear weights of time-sensitive behaviors using a time-sensitive biased random walk method before being matched with material vectors. Specifically, time-sensitive behaviors are directly taken from low-dimensional embedding representations and retrieved as relevant materials from the Faiss library using inner product metric, while non-time-sensitive behaviors are constructed as linear weighted representations of time-sensitive materials using a time-sensitive biased random walk method before retrieval. S5. New Hot Content Push: For user groups with similar historical behaviors, use vector matching to push timely new hot items and content.

2. The item recommendation method based on time-sensitive interest modeling as described in claim 1, characterized in that, In step S1, the collected information includes user identity ID, item content ID, item content UGC text, and user browsing or clicking time information.

3. The item recommendation method based on time-sensitive interest modeling as described in claim 2, characterized in that, Step S1 includes collecting and preprocessing the associated data required for offline interest mining and real-time preference prediction. For the offline part, all items that the user has browsed and clicked within a set period are collected to construct a material-material rights isomorphism graph. For the real-time part, multiple items that the user has recently clicked and their timestamps are collected and defined as the user's behavior sequence for subsequent use.

4. The item recommendation method based on time-sensitive interest modeling as described in any one of claims 1 to 3, characterized in that, In step S2, the timeliness elements contained in the item content are identified and extracted using a strategy based on a large language model and a strategy based on Chinese word segmentation.

5. The item recommendation method based on time-sensitive interest modeling as described in claim 4, characterized in that, In step S2, the strategy based on the large language model includes defining a parameterized Prompt and explicit examples of time-sensitive element vocabulary, and asking the large language model whether a UGC content can reflect similar time-sensitive elements; the strategy based on Chinese word segmentation includes using the jieba word segmentation tool to segment the UGC content, and using the distance of each segmented word to the predefined time-sensitive element vocabulary set as a sparse feature. The time-sensitive and non-time-sensitive binary classification model is used for time-sensitive discrimination.

6. The item recommendation method based on time-sensitive interest modeling as described in any one of claims 1 to 3, characterized in that, Step S3 specifically includes the following steps: Step S31: Based on the material timeliness labels obtained from the timeliness understanding, mark the nodes of the offline material-material map obtained by the behavior collection module to distinguish between time-sensitive and non-time-sensitive content, and perform pruning on the side. Step S32: Define the timeliness score of each timeliness content at the initial time t=0 as 1, and the rest as 0; Step S33: Starting from all nodes with non-zero timeliness scores, perform timeliness score propagation within the first-order neighborhood, following the following: ； Where v i and v j This refers to the material node, where S represents the node's timeliness score at time t. Indicates v i The first-order neighborhood node set; the above process in step S33 is repeated for multiple time steps, and material nodes with a final timeliness score of 0 are removed; Step S34: Starting from each time-delay node, perform multiple random walks of L steps with time-delay bias and tidal effects, where the (i+1)th step node is connected to the previous step node v. i The samples are obtained from the first-order neighborhood of the class according to a specific parameterized distribution, i.e.: ； Cat(.) generates a class distribution that satisfies the time-dependent bias and tidal effect, defined as: ； T is the maximum number of transmission rounds set. For v i and v j The strength of the interest association between them, m a m is the oversaturation coefficient for aging. in The time-dependent undersaturation coefficient is L, where L is the maximum number of steps in a random walk. Step S35: Summarize all random walk sequences originating from each time-sensitive node, remove non-time-sensitive nodes, and construct training positive sample pairs according to the word frequency distribution of time-sensitive nodes. Construct simple negative sample pairs by randomly sampling from time-sensitive materials and construct difficult negative sample pairs by randomly sampling from hypothetical uniform word frequencies. During training, embed each material into a multi-dimensional vector, and perform gradient optimization on the co-occurrence likelihood by setting the ratio of positive samples: simple negative samples: difficult negative samples. After training multiple iterations, take the one with the largest likelihood as the final model, which is the low-dimensional embedding vector of all time-sensitive materials.

7. The item recommendation method based on time-sensitive interest modeling as described in claim 6, characterized in that, Step S4 specifically includes the following steps: Step S41: Starting from each non-time-sensitive material node, perform multiple random walks with time-sensitivity bias. The time-sensitivity score directly uses the result of step S33, but the walk process no longer allows the tidal effect to take effect. That is, under the same conditions as the symbol definition in step S34, the category sampling follows: ； Step S42: Summarize all random walk sequences originating from each non-time-sensitive node, similarly remove non-time-sensitive nodes, and obtain the time-sensitive node representation set corresponding to the non-time-sensitive nodes. Each element of the set is a time-sensitive node and its word frequency. From the embedding vector set obtained in step S35, retrieve the embedding vectors corresponding to these time-sensitive nodes, and obtain the interpolation embedding result of each non-time-sensitive node by weighting by word frequency. Step S43: Represent all user non-time-sensitive behaviors as embedding vectors, and use the Faiss library to retrieve the top K as relevant materials by inner product metric.

8. The item recommendation method based on time-sensitive interest modeling as described in claim 7, characterized in that, In step S5, the content matched by the user behavior sequence is sorted by exponential decay weighted mixing according to the step size since then.

9. The item recommendation method based on time-sensitive interest modeling as described in claim 8, characterized in that, Step S5 specifically includes the following steps: Step S51: Extract the most recent Q most clicked item content from a given user and invert the list as a sequence of the user's historical behavior, and perform vector matching on each content; Step S52: Weight the matching results according to the positional order of their source index behavior, following: ; Where V i It is the ID of the weightless material obtained by matching, p(V) i Z is its normalized recommendation score. i It is the normalization coefficient within the current batch. It is the exponential decay factor of sequence attention; Step S53: Selecting factors Step S51 is executed to obtain the matching material results. The final recommendation list is generated in descending order of score and pushed to the user.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the item recommendation method based on time-sensitive interest modeling as described in any one of claims 1 to 9.

Images