Material recall method and system and computer readable storage medium

Abstract

The embodiment of the invention provides a material recall method and system and a computer readable storage medium, and the method comprises the steps: receiving a material recommendation request of a target user, and based on the interaction behavior data of the target user, obtaining a material which has an interaction behavior of the target user within a second time range and meets a set condition, and taking the material as a trigger material; based on the periodically updated material database, similar materials corresponding to all the trigger materials are determined, and the similar materials corresponding to all the trigger materials serve as to-be-screened materials; and determining a correlation score value between each to-be-screened material and the corresponding trigger material, and determining a recalled material corresponding to the target user from all the to-be-screened materials based on the correlation score value for recall. An eges graph algorithm is adopted offline in advance, similar materials of each material are constructed, the material correlation is better, more similar materials can be recalled in the offline process, and cold materials (small in exposure quantity) can be better expressed.

Description

Technical Field

[0001] The present invention relates to the field of material recommendation, and in particular to a material recall method, system, computer-readable storage medium and computer equipment. Background Art

[0002] In Weibo's post-video recommendation business, the recommendation system architecture currently widely used in the industry is adopted. In order to provide users with personalized information flow recommendation services, the recommendation engine will go through the stages of recall (obtaining nearly 10,000 items from millions of materials) -> coarse sorting (retaining 1,000 items after model screening) -> fine sorting (scoring and sorting 1,000 items) -> rule system (selecting the 10 items with the highest ranking and in line with the strategy). In the recall stage, model recall is divided into rule recall, collaborative recall, vector recall, tree recall, etc., among which collaborative recall includes material collaborative filtering recall, which is generally achieved through matrix decomposition, graphs, deep networks and other methods.

[0003] In the prior art, the offline method for implementing collaborative filtering and recall of materials is the deepwalk graph algorithm, which constructs an interactive behavior graph between users and materials, and only uses the user's ID information and the material's ID information (without using other features of users and materials). The material embedding is obtained through skip gram training, and the faiss index is constructed to obtain the topN similar materials for each material. The similar materials and similarity scores of the materials are written into redis for online use. In response to online user requests, a list of materials that the user has interacted with in the past period of time is obtained. For each item in the trigger list, the similar material data of the material in redis is read to obtain the topN materials that are most similar to it as the result of collaborative material recall, which is used as part of the input for the subsequent link coarse sorting to recommend to users.

[0004] In the process of implementing the present invention, the applicant discovered that there are at least the following problems in the prior art:

[0005] The deepwalk graph algorithm used only has the ID information of the material node, which is missing a lot of information and has insufficient expression ability. The number of recalled materials is limited, and many similar materials will be missed. Summary of the invention

[0006] The embodiments of the present invention provide a material recall method, system, storage medium and computer device, which can solve the technical problem in the prior art that "the deepwalk graph algorithm used only has the ID information of the material node, lacks a lot of information, has insufficient expression ability, is limited in the number of recalled materials, and will miss many similar materials."

[0007] To achieve the above objectives, in a first aspect, an embodiment of the present invention provides a material recall method, comprising:

[0008] receiving a material recommendation request from a target user, and acquiring, based on the interaction behavior data of the target user, materials with which the target user has interaction behavior within a second time range and which meet set conditions as trigger materials;

[0009] Based on a periodically updated material database, similar materials corresponding to each trigger material are determined, and similar materials corresponding to all trigger materials are used as materials to be screened; each material within a first time range and at least one similar material corresponding thereto are recorded in the material database; the material database is generated by inputting the interaction behavior data of all users within the first time range into an eges graph algorithm, and training to obtain a material embedding vector corresponding to each material;

[0010] Determine the correlation score between each material to be screened and the corresponding trigger material, and determine the recall material corresponding to the target user from all materials to be screened based on the correlation score for recall; the correlation score represents the similarity between the material to be screened and the corresponding trigger material.

[0011] In a second aspect, an embodiment of the present invention provides a material recall system, including:

[0012] A trigger material acquisition unit, configured to receive a material recommendation request from a target user, and based on the interaction behavior data of the target user, acquire a material with which the target user has an interaction behavior within a second time range and which meets a set condition as a trigger material;

[0013] The first material recall unit is used to determine similar materials corresponding to each trigger material based on a periodically updated material database, and use similar materials corresponding to all trigger materials as materials to be screened; the material database records each material within the first time range and at least one similar material corresponding thereto; the material database is generated by inputting the interaction behavior data of all users within the first time range into an eges graph algorithm, and training to obtain a material embedding vector corresponding to each material;

[0014] The second material recall unit is used to determine the correlation score value between each material to be screened and the corresponding trigger material, and determine the recall material corresponding to the target user from all materials to be screened based on the correlation score value for recall; the correlation score value represents the similarity between the material to be screened and the corresponding trigger material.

[0015] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a computer device, the computer device executes the aforementioned material recall method.

[0016] The above technical solution has the following beneficial effects: the eges graph algorithm is used offline in advance to train materials with user interaction behaviors. After the material embedding vector corresponding to each material is obtained through training, similar materials for each material are constructed, and the material relevance is better. More similar materials can be recalled offline, and cold materials (with less exposure) can perform better. When there is a material recommendation request from a target user, the material with which the target user has an interactive behavior within the second time range and meets the set conditions is obtained as a trigger material, and similar materials of the trigger material are obtained. The material is recalled from all similar materials of the target user based on the relevance score value, and the recalled material has a higher similarity and will not miss any material. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 is a flow chart of a material recall method according to an embodiment of the present invention;

[0019] Figure 2 is a structural diagram of a material recall system according to an embodiment of the present invention;

[0020] Figure 3 It is the overall process framework based on collaborative filtering and recall of materials in an embodiment of the present invention;

[0021] Figure 4 It is a schematic diagram of a two-dimensional material diagram of an embodiment of the present invention;

[0022] Figure 5 It is a flow chart of offline construction of similar materials in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, in combination with an embodiment of the present invention, a material recall method is provided, comprising:

[0025] S101: receiving a material recommendation request from a target user, and acquiring, based on the interaction behavior data of the target user, materials with which the target user has interaction behavior within a second time range and which meet set conditions as trigger materials;

[0026] S102: Based on a periodically updated material database, similar materials corresponding to each trigger material are determined, and similar materials corresponding to all trigger materials are used as materials to be screened; the material database records each material within the first time range and at least one similar material corresponding thereto; the material database is generated by inputting interaction behavior data of all users within the first time range into an eges (Enhanced Graph Embedding with Side Information) graph algorithm, and training to obtain a material embedding vector corresponding to each material;

[0027] S103: Determine the correlation score between each material to be screened and the corresponding trigger material, and determine the recall material corresponding to the target user from all materials to be screened based on the correlation score; the correlation score represents the similarity between the material to be screened and the corresponding trigger material.

[0028] The Eges graph algorithm is used offline in advance to train materials with user interaction behaviors. After training to obtain the material embedding vector corresponding to each material, similar materials for each material are constructed. The material relevance is better, more similar materials can be recalled offline, and cold materials (with less exposure) can perform better. When there is a material recommendation request from a target user, the material with which the target user has an interactive behavior within the second time range and meets the set conditions is obtained as the trigger material, and similar materials of the trigger material are obtained. The material is recalled from all similar materials of the target user based on the relevance score value. The recalled material has a higher similarity and will not miss any material.

[0029] Preferably, the material recall method further includes:

[0030] S104: periodically obtain the interaction behavior data of all users within the first time range, and construct a bipartite graph of materials including a diversion position and a recommendation position through an eges graph algorithm; the diversion position represents the materials displayed in any primary scene, and the recommendation position represents the recommended materials in the secondary scene through a specific behavior on the diversion position;

[0031] S105: Obtain a material embedding vector corresponding to each material based on the material bipartite graph and feature information of each material, perform similarity calculation based on the material embedding vector corresponding to each material, obtain at least one similar material corresponding to each material, and write the similar materials into the material database.

[0032] The Eges graph algorithm is used offline in advance to periodically construct a bipartite graph for materials with user interaction behaviors, and obtain the material embedding vector corresponding to each material based on the bipartite graph and multiple feature information of each material. It uses multiple features of the material in addition to the node itself, has a stronger ability to characterize the material, and has better material relevance. It can recall more similar materials offline, and solve some material cold start problems, so that unpopular materials can be recalled.

[0033] Preferably, S104: periodically acquiring the interaction behavior data of all users within the first time range, and constructing a bipartite graph of materials including diversion positions and recommendation positions through an eges graph algorithm, including:

[0034] S104-1: periodically obtain the interaction behavior data of all users within a first time range, and use the eges graph algorithm to determine all diversion positions, where the diversion positions represent the materials displayed in any first-level scene;

[0035] S104-2: Determine, according to each diversion position, a recommended position corresponding to the diversion position, wherein the recommended position represents a material recommended for entering the secondary scene through a specific behavior of the diversion position after the material enters the diversion position;

[0036] S104-3: Construct each recommendation position, each diversion position, and the edge between the recommendation position and the corresponding diversion position to obtain the material bipartite graph; the wandering probability weight of the edge between the recommendation position and the corresponding diversion position is: the average viewing time of the recommendation position within the first time range.

[0037] The diversion position is the material displayed in any first-level scene. The recommended position is the material recommended in the second-level scene after entering the diversion position material through the specific behavior of the diversion position. The recommended position material is determined based on the diversion position material, and a two-dimensional material graph is constructed based on the diversion position material and the recommended position material. The two-dimensional material graph shows that the materials at both ends of the edge are related, the connection between materials is strengthened, and all materials are used. Then, when determining similar materials for each material, related materials will not be missed, and the recall rate of materials will be improved.

[0038] Preferably, in S105, obtaining a material embedding vector corresponding to each material based on the material bipartite graph and the feature information of each material includes:

[0039] S105-1: for each material in the material bipartite graph, take the material as the starting material, walk along the edges in the material bipartite graph according to the biased random walk, and sequentially construct a material sequence corresponding to the material according to the materials passed by the continuous walk;

[0040] S105-2: For each group of material sequences, training is performed using a skip-gram model based on feature information of each material in the material sequence, and a material embedding vector corresponding to each material in the material sequence is output.

[0041] If the random walk is used, it is an equal probability walk, which cannot reflect the size of the correlation between the materials at both ends of the edge. However, this application uses a biased random walk, so the probability of walking the edge of the material with high correlation is high and the number of times is high, so the materials with greater material similarity can be recalled offline, and the recall rate can also be improved.

[0042] Preferably, S105-1: for each material in the material bipartite graph, take the material as the starting material, walk on the edge in the material bipartite graph according to the biased random walk, and sequentially construct a material sequence corresponding to the material according to the materials passed by the continuous walk, specifically including:

[0043] For each material in the material bipartite graph, the material is taken as the starting material, and whether to wander to the next material connected to the previous material by an edge is determined based on the wandering probability weight of the edge;

[0044] When it is determined based on the edge walking probability weight to walk to the next material connected to the previous material by an edge, walk backward in sequence based on the edge walking probability weight, and form a material sequence corresponding to the material according to the materials passed through by continuous walking, until the number of materials in the material sequence reaches the preset material number.

[0045] For the same starting material, there may be multiple material sequences, and the same material may appear in multiple material sequences at the same time. The recall rate of similar materials will be improved by calculating the material embedding vector based on multiple feature information of the material. In particular, cold materials may also exist in multiple material sequences, so many cold materials will be started, which will also improve the recall rate of cold materials.

[0046] Preferably, S105-2: for each group of material sequences, training is performed using a skip-gram model based on feature information of each material in the material sequence, and a material embedding vector corresponding to each material in the material sequence is output, including:

[0047] For each group of material sequences, according to the set sliding window length, the starting material of the material sequence is used as the starting material of the current sliding window, the first number of materials of the sliding window length are sequentially used as material positive samples of the current sliding window, the starting materials are respectively used with other materials in the current sliding window to form positive sample pairs, material negative samples are randomly selected from the materials in all material sequences, the starting materials and the material negative samples are used to form positive and negative sample pairs, the positive sample pairs in the current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the current sliding window is output, and the current material embedding vector of the material is the vector mean of all feature information of the material, wherein the feature information includes: material node information, material author ID, primary content label corresponding to the material, secondary content label corresponding to the material, and tertiary content label corresponding to the material;

[0048] After obtaining the current material embedding vector of each material in the current sliding window through the skip-gram model, the current sliding window is changed to the previous sliding window, the previous sliding window is slid backward by one position in the material sequence to use the second material in the material sequence as the starting material of the new current sliding window, the starting material is respectively used with other materials in the current sliding window to form a positive sample pair, a material negative sample is randomly selected from the materials in all material sequences, the starting material and the material negative sample are used to form a positive and negative sample pair, the positive sample pair in the new current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the new current sliding window is output; wherein, for the material used as a positive sample pair or a positive and negative sample pair in the previous sliding window, the current material embedding vector of the corresponding output material is input into the skip-gram model as the feature information of the material;

[0049] Until the last material of the material sequence is slid into the new current sliding window, the current material embedding vector of each material in the new current sliding window is output;

[0050] The current material embedding vector of each material in the material sequence is used as the final material embedding vector of the material and outputted.

[0051] The skip-gram model considers that the materials in the current sliding window are similar, namely, positive samples, and negative samples are randomly drawn from global materials to form positive sample pairs and negative sample pairs. The skip-gram model training will make the distance between positive samples closer and the distance between positive and negative samples farther, so that the distance between the material embedding vectors of more similar materials will be closer, thus achieving rapid aggregation of similar materials.

[0052] Preferably, in S105, the similarity calculation is performed based on the material embedding vector corresponding to each material to obtain at least one similar material corresponding to each material, including:

[0053] S105-3: For each material in the material bipartite graph, use faiss to calculate the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials, sort the distances from small to large, take a preset number of other materials before sorting as similar materials of the material, and construct an index of similar materials for each material, the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials is equal to the inner product of the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials.

[0054] Since the number of candidate materials in the recall scenario is large, the Faiss method is used to quickly calculate the distance between vectors to meet the requirements of a large number of recalled materials and a large number of calculation vectors.

[0055] Preferably, in S103, determining the correlation score between each to-be-screened material and the corresponding trigger material includes:

[0056] S103-1: for each material to be screened, determining a content correlation coefficient according to a content distance between the material to be screened and a corresponding trigger material;

[0057] S103-2: determining a time decay factor according to the time difference between the release time of the material to be screened and the corresponding trigger material, wherein the time decay factor is negatively correlated with the time difference between the release times;

[0058] S103-3: Determine the Wilson correlation coefficient based on the relationship between the estimated viewing time of the material to be screened and the duration of the material to be screened, wherein the Wilson correlation coefficient is positively correlated with the ratio of the estimated viewing time to the duration of the material to be screened;

[0059] S103-4: determining an interaction coefficient according to the interaction frequency between the user and the material to be screened, wherein the interaction coefficient is positively correlated with the interaction frequency;

[0060] S103-5: The product of the content correlation coefficient, the time decay factor, the Wilson correlation coefficient and the interaction coefficient is used as the correlation score value between the material to be screened and the corresponding trigger material.

[0061] The correlation score value is designed according to the business characteristics, using the formula that maximizes the indicator benefit. The correlation coefficient is a score calculated offline, which is the inner product of the two material embedding vectors. The score ranges from 0 to 1. The more similar the two materials are, the closer they are to 1. The time decay factor is that the closer the material release time is to the current time, the higher the factor score. The Wilson correlation coefficient is positively correlated with the ratio between the estimated viewing time and the time of the material to be screened. The closer the viewing time is to the time of the material to be screened, the higher the score. The more likes, comments, and reposts of the material are scored, the higher the interaction score coefficient is.

[0062] like Figure 2 As shown, in combination with an embodiment of the present invention, a material recall system is provided, including:

[0063] The trigger material acquisition unit 21 is used to receive a material recommendation request from a target user, and based on the interaction behavior data of the target user, acquire a material with which the target user has an interaction behavior within a second time range and which meets a set condition as a trigger material;

[0064] The first material recall unit 22 is used to determine similar materials corresponding to each trigger material based on a periodically updated material database, and use similar materials corresponding to all trigger materials as materials to be screened; the material database records each material within the first time range and at least one similar material corresponding thereto; the material database is generated by inputting the interaction behavior data of all users within the first time range into an eges graph algorithm, and training to obtain a material embedding vector corresponding to each material;

[0065] The second material recall unit 23 is used to determine the correlation score value between each material to be screened and the corresponding trigger material, and determine the recall material corresponding to the target user from all materials to be screened based on the correlation score value for recall; the correlation score value represents the similarity between the material to be screened and the corresponding trigger material.

[0066] The Eges graph algorithm is used offline in advance to train materials with user interaction behaviors. After training to obtain the material embedding vector corresponding to each material, similar materials for each material are constructed. The material relevance is better, more similar materials can be recalled offline, and cold materials (with less exposure) can perform better. When there is a material recommendation request from a target user, the material with which the target user has an interactive behavior within the second time range and meets the set conditions is obtained as the trigger material, and similar materials of the trigger material are obtained. The material is recalled from all similar materials of the target user based on the relevance score value. The recalled material has a higher similarity and will not miss any material.

[0067] Preferably, the material recall system further includes a similar material offline construction unit, and the similar material offline construction unit includes:

[0068] The material bipartite graph construction subunit is used to periodically obtain the interactive behavior data of all users within the first time range, and construct a material bipartite graph including a diversion position and a recommendation position through the eges graph algorithm; the diversion position represents the material displayed in any primary scene, and the recommendation position represents the recommended material in the secondary scene through a specific behavior on the diversion position;

[0069] The similar material offline determination subunit is used to obtain the material embedding vector corresponding to each material based on the material bipartite graph and the characteristic information of each material, perform similarity calculation based on the material embedding vector corresponding to each material, obtain at least one similar material corresponding to each material, and write it into the material database.

[0070] The Eges graph algorithm is used offline in advance to periodically construct a bipartite graph for materials with user interaction behaviors, and obtain the material embedding vector corresponding to each material based on the bipartite graph and multiple feature information of each material. It uses multiple features of the material in addition to the node itself, has a stronger ability to characterize the material, and has better material relevance. It can recall more similar materials offline, and solve some material cold start problems, so that unpopular materials can be recalled.

[0071] Preferably, the material bipartite graph constructs a subunit, specifically for:

[0072] Periodically obtain the interactive behavior data of all users within the first time range, and use the eges graph algorithm to determine all diversion positions, where the diversion positions represent the materials displayed in any first-level scene;

[0073] According to each of the diversion positions, a recommended position corresponding to the diversion position is determined, wherein the recommended position represents the recommended material in the secondary scene after the material enters the diversion position through a specific behavior of the diversion position;

[0074] Each recommendation position, each diversion position, and the edge between the recommendation position and the corresponding diversion position are constructed to obtain the material bipartite graph; the wandering probability weight of the edge between the recommendation position and the corresponding diversion position is: the average viewing time of the recommendation position within the first time range.

[0075] The diversion position is the material displayed in any first-level scene. The recommended position is the material recommended in the second-level scene after entering the diversion position material through the specific behavior of the diversion position. The recommended position material is determined based on the diversion position material, and a two-dimensional material graph is constructed based on the diversion position material and the recommended position material. The two-dimensional material graph shows that the materials at both ends of the edge are related, the connection between materials is strengthened, and all materials are used. Then, when determining similar materials for each material, related materials will not be missed, and the recall rate of materials will be improved.

[0076] Preferably, the similar material offline determination subunit is specifically used for:

[0077] A material sequence construction module is used for taking each material in the material bipartite graph as a starting material, walking along the edges in the material bipartite graph according to a biased random walk, and sequentially constructing a material sequence corresponding to the material according to the materials passed by the continuous walk;

[0078] The material embedding vector output module is used to train the skip-gram model based on the feature information of each material in each material sequence for each group of material sequences, and output the material embedding vector corresponding to each material in the material sequence.

[0079] If the random walk is used, it is an equal probability walk, which cannot reflect the size of the correlation between the materials at both ends of the edge. However, this application uses a biased random walk, so the probability of walking the edge of the material with high correlation is high and the number of times is high, so the materials with greater material similarity can be recalled offline, and the recall rate can also be improved.

[0080] Preferably, the material sequence building module is specifically used for:

[0081] For each material in the material bipartite graph, the material is taken as the starting material, and whether to wander to the next material connected to the previous material by an edge is determined based on the wandering probability weight of the edge;

[0082] When it is determined based on the edge walking probability weight to walk to the next material connected to the previous material by an edge, walk backward in sequence based on the edge walking probability weight, and form a material sequence corresponding to the material according to the materials passed through by continuous walking, until the number of materials in the material sequence reaches the preset material number.

[0083] For the same starting material, there may be multiple material sequences, and the same material may appear in multiple material sequences at the same time. The recall rate of similar materials will be improved by calculating the material embedding vector based on multiple feature information of the material. In particular, cold materials may also exist in multiple material sequences, so many cold materials will be started, which will also improve the recall rate of cold materials.

[0084] Preferably, the material embedding vector output module is specifically used for:

[0085] For each group of material sequences, according to the set sliding window length, the starting material of the material sequence is used as the starting material of the current sliding window, the first number of materials of the sliding window length are sequentially used as material positive samples of the current sliding window, the starting materials are respectively used with other materials in the current sliding window to form positive sample pairs, material negative samples are randomly selected from the materials in all material sequences, the starting materials and the material negative samples are used to form positive and negative sample pairs, the positive sample pairs in the current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the current sliding window is output, and the current material embedding vector of the material is the vector mean of all feature information of the material, wherein the feature information includes: material node information, material author ID, primary content label corresponding to the material, secondary content label corresponding to the material, and tertiary content label corresponding to the material;

[0086] After obtaining the current material embedding vector of each material in the current sliding window through the skip-gram model, the current sliding window is changed to the previous sliding window, the previous sliding window is slid backward by one position in the material sequence to use the second material in the material sequence as the starting material of the new current sliding window, the starting material is respectively used with other materials in the current sliding window to form a positive sample pair, a material negative sample is randomly selected from the materials in all material sequences, the starting material and the material negative sample are used to form a positive and negative sample pair, the positive sample pair in the new current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the new current sliding window is output; wherein, for the material used as a positive sample pair or a positive and negative sample pair in the previous sliding window, the current material embedding vector of the corresponding output material is input into the skip-gram model as the feature information of the material;

[0087] Until the last material of the material sequence is slid into the new current sliding window, the current material embedding vector of each material in the new current sliding window is output;

[0088] The current material embedding vector of each material in the material sequence is used as the final material embedding vector of the material and outputted.

[0089] The skip-gram model considers that the materials in the current sliding window are similar, namely, positive samples, and negative samples are randomly drawn from global materials to form positive sample pairs and negative sample pairs. The skip-gram model training will make the distance between positive samples closer and the distance between positive and negative samples farther, so that the distance between the material embedding vectors of more similar materials will be closer, thus achieving rapid aggregation of similar materials.

[0090] Preferably, the similar material offline determination subunit is specifically used for:

[0091] For each material in the material bipartite graph, faiss is used to calculate the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials, the distances are sorted from small to large, a preset number of other materials before sorting are taken as similar materials of the material, and an index of similar materials of each material is constructed, and the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials is equal to the inner product of the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials.

[0092] Since the number of candidate materials in the recall scenario is large, the Faiss method is used to quickly calculate the distance between vectors to meet the requirements of a large number of recalled materials and a large number of calculation vectors.

[0093] Preferably, the second material recall unit 23 is specifically used for:

[0094] For each material to be screened, determining a content correlation coefficient according to a content distance between the material to be screened and a corresponding trigger material;

[0095] Determine a time decay factor according to the time difference between the release time of the material to be screened and the corresponding trigger material, wherein the time decay factor is negatively correlated with the time difference between the release times;

[0096] Determine the Wilson correlation coefficient according to the relationship between the estimated viewing time of the material to be screened and the duration of the material to be screened, wherein the Wilson correlation coefficient is positively correlated with the ratio between the estimated viewing time and the duration of the material to be screened;

[0097] Determine an interaction coefficient according to the interaction frequency between the user and the material to be screened, wherein the interaction coefficient is positively correlated with the interaction frequency;

[0098] The product of the content correlation coefficient, the time decay factor, the Wilson correlation coefficient and the interaction coefficient is used as the correlation score value between the material to be screened and the corresponding trigger material.

[0099] The correlation score value is designed according to the business characteristics, using the formula that maximizes the indicator benefit. The correlation coefficient is a score calculated offline, which is the inner product of the two material embedding vectors. The score ranges from 0 to 1. The more similar the two materials are, the closer they are to 1. The time decay factor is that the closer the material release time is to the current time, the higher the factor score. The Wilson correlation coefficient is positively correlated with the ratio between the estimated viewing time and the time of the material to be screened. The closer the viewing time is to the time of the material to be screened, the higher the score. The more likes, comments, and reposts of the material are scored, the higher the interaction score coefficient is.

[0100] In combination with an embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a computer device, the computer device executes any one of the aforementioned material recall methods.

[0101] The above technical solution of the embodiment of the present invention is described in detail below in conjunction with specific application examples. For technical details not introduced during the implementation process, please refer to the relevant description in the previous text.

[0102] A material recall method and system, which uses the collaborative filtering recall method of materials based on the Eges graph algorithm to construct a bipartite graph of diversion materials and recommendation materials, and introduces more feature information of the materials, and then performs training to obtain the material embedding vectors corresponding to the materials. This method has a stronger ability to express materials, better material relevance, improved offline recall rate evaluation indicators, and better performance of cold materials (less exposure), and significantly improved online indicators, solving the problems of previous graph models.

[0103] 1. Periodically (hourly) obtain the interactive behavior data of all users within the first time range, and construct a bipartite graph of materials including diversion positions and recommendation positions through the eges graph algorithm; the diversion position represents the materials displayed in any primary scene, and the recommendation position represents the recommended materials that enter the secondary scene through specific behaviors of the diversion position; based on the bipartite graph of materials and the characteristic information of each material, the material embedding vector corresponding to each material is obtained, and the similarity is calculated based on the material embedding vector corresponding to each material to obtain at least one similar material corresponding to each material, and write it into the material database. Specifically: In the post-video recommendation system, the overall process framework based on collaborative filtering and recall of materials is as follows: Figure 3As shown in the figure, in the offline construction process of similar materials, the interactive behavior data of users on the entire site within the first time range, such as 24 hours (referring to materials with user interactive behavior) is used to construct a two-dimensional material graph as input, and the material embedding vectors of these materials are obtained by training the eges graph algorithm based on the two-dimensional material graph. The index is built by faiss to obtain the topN similar materials of each material, and the results are written to redis. The entire offline process is updated hourly. Because the number of candidate materials in the recall scenario is large, the faiss method must be used to quickly calculate the distance between vectors. If it is not used, it will take a lot of time online and cannot meet the requirements of the real-time recommendation system. Faiss has a fast construction speed and can meet the requirements of a large number of recalled materials and a large number of calculation vectors.

[0104] (I) Offline construction of similar materials using Eges graph algorithm

[0105] Periodically obtain the interactive behavior data of all users within the first time range, and use the eges graph algorithm to determine all diversion positions, which represent the materials displayed in any first-level scene; determine the recommendation position corresponding to each diversion position according to each diversion position, and the recommendation position represents the recommended materials that enter the second-level scene through specific behaviors of the diversion position after entering the diversion position material; construct the bipartite graph of the material by combining each recommendation position, each diversion position, and the edges between the recommendation position and the corresponding diversion position; the walking probability weight of the edge between the recommendation position and the corresponding diversion position is: the average viewing time of the recommendation position within the first time range. Specific examples are:

[0106] like Figure 4 As shown in the figure, all video materials covering the first time range, for example, within 24 hours, are obtained for the entire site (Weibo). The interactive behavior data of the entire site users within 24 hours (referring to materials with user interactive behavior) are used to screen materials. The material screening conditions are: the viewing time of the material is greater than 12 seconds, or likes, reposts, and comments are given. A bipartite graph of the diversion material lmid and the recommendation material mid is constructed. lmid is the diversion material, mid is the recommendation material, and the connecting line is represented as follows: Figure 4 The connecting line between lmid1 and mid5 in the figure shows that after watching the diversion material lmid1, the video mid5 is recommended after scrolling down the page, that is, lmid1 and mid5 are connected. Diversion materials refer to the display materials of any first-level scene of Weibo app. After clicking on the material, scrolling down the video page enters the second-level scene after the video is recommended. The video materials recommended in this second-level scene are all recommended materials, that is, the recommended materials are recommended after scrolling down through the diversion material. Constructing the bipartite graph of materials can obtain the material embedding vector embedding of the diversion material and improve the coverage of trigger mid.

[0107] Construct a two-part graph of traffic diversion and recommendation position co-occurrence (samples of 6 main scenarios: hot search, relationship flow, popular flow, hot flow, text page, and personal page), and map the traffic diversion material features and recommendation position features to a unified feature space, with a data volume of millions. Both the traffic diversion material features and recommendation position features include: material node information, material author id, first-level content label corresponding to the blog post, (whole site) second-level content label corresponding to the blog post, (whole site) third-level content label corresponding to the blog post. That is, in addition to using the id information feature of the material node itself, add side info supplementary features: author id, (whole site) first-level content label corresponding to the blog post, (whole site) second-level content label corresponding to the blog post, (whole site) third-level content label corresponding to the blog post. The weights of the 4 features are set to the same (this application believes that the contributions of the 4 features are the same). Remove noise during random walk of the graph: filter out the traffic diversion material and recommendation material pairs published 8 months ago from the current time.

[0108] After the material bipartite graph is constructed, for each material in the material bipartite graph, the material is taken as the starting material, and the material is walked on the edges in the material bipartite graph according to biased random walks, and a material sequence corresponding to the material is sequentially constructed according to the materials passed through by the continuous walk; specifically including: for each material in the material bipartite graph, the material is taken as the starting material, and it is determined based on the edge walking probability weight whether to walk to the next material connected to the previous material by an edge; when it is determined based on the edge walking probability weight that the material has walked to the next material connected to the previous material by an edge, the material sequence corresponding to the material is constructed based on the edge walking probability weight, until the number of materials in the material sequence reaches the preset number of materials.

[0109] like Figure 4 The example shows a material sequence of length 10: lmid3, mid10, lmid2, mid1, ..., lmid7, mid8. The probability weight of the edge between material nodes is further explained as follows: the log value of the average viewing time of the recommended material node within 24 hours is the weight of the edge, the average viewing time refers to the total viewing time of all people within 24 hours divided by the number of people, and the edge refers to the edge between the recommended material and the previous diversion material. Among them, biased walking means that the weights of the edges are different, and the greater the weight, the greater the probability of walking this edge.

[0110] like Figure 5 As shown in the figure, after obtaining the video material sequence, for each group of material sequences, the skip-gram model is trained based on the feature information of each material in the material sequence, and the material embedding vector corresponding to each material in the material sequence is output:

[0111] For each group of material sequences, according to the set sliding window length, the starting material of the material sequence is used as the starting material of the current sliding window, the first number of materials of the sliding window length are sequentially used as material positive samples of the current sliding window, the starting materials are respectively used with other materials in the current sliding window to form positive sample pairs, material negative samples are randomly selected from the materials in all material sequences, the starting materials and the material negative samples are used to form positive and negative sample pairs, the positive sample pairs in the current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the current sliding window is output, and the current material embedding vector of the material is the vector mean of all feature information of the material, wherein the feature information includes: material node information, material author ID, primary content label corresponding to the material, secondary content label corresponding to the material, and tertiary content label corresponding to the material;

[0112] After obtaining the current material embedding vector of each material in the current sliding window through the skip-gram model, the current sliding window is changed to the previous sliding window, the previous sliding window is slid backward by one position in the material sequence to use the second material in the material sequence as the starting material of the new current sliding window, the starting material is respectively used with other materials in the current sliding window to form a positive sample pair, a material negative sample is randomly selected from the materials in all material sequences, the starting material and the material negative sample are used to form a positive and negative sample pair, the positive sample pair in the new current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the new current sliding window is output; wherein, for the material used as a positive sample pair or a positive and negative sample pair in the previous sliding window, the current material embedding vector of the corresponding output material is input into the skip-gram model as the feature information of the material;

[0113] Until the last material of the material sequence is slid into the new current sliding window, the current material embedding vector of each material in the new current sliding window is output;

[0114] The current material embedding vector of each material in the material sequence is used as the final material embedding vector of the material and outputted.

[0115] Assume the window length is 4, such as Figure 4The material sequence is lmid3, mid10, and lmid2, which are the materials in the current sliding window. After the current sliding window is slid back by 1 position, it becomes mid10, lmid2, and mid1, which slide in sequence. The skip-gram model considers that the materials in the current sliding window are similar, that is, positive samples. Negative samples are randomly drawn from global materials to form positive sample pairs and negative sample pairs. The skip-gram model makes the distance between positive samples closer and the distance between positive and negative samples farther, so that the distance between the material embedding vectors of similar materials will be closer. When the skip-gram model is trained, the side info feature embedding and id embedding of the material are constantly updated and changed. The final material embedding of each material is the mean of the side info feature embedding and id embedding. For example, a material has 2 side infonce features, which means that the material has three embeddings in total. The sum of the three vectors divided by 3 is the mean. The side info feature embedding and id embedding correspond to the side info feature and the material node feature respectively. The elimination time of the material id and sideinfo feature vectors is the length of the third time period, such as one week, and they are continuously updated and iterated with incremental training.

[0116] (ii) For each material in the material bipartite graph, use faiss to calculate the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials, sort the distances from small to large, take a preset number of other materials before sorting as similar materials of the material, and construct an index of similar materials for each material, the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials is equal to the inner product of the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials.

[0117] 2. If Figure 3 As shown, online, a material recommendation request from a target user is received, and based on the interactive behavior data of the target user, materials with which the target user has interactive behavior and that meet set conditions within a second time range are obtained as trigger materials; based on a periodically updated material database, similar materials corresponding to each trigger material are determined, and similar materials corresponding to all trigger materials are used as materials to be screened. The correlation score value between each material to be screened and the corresponding trigger material is determined, and the recall material corresponding to the target user is determined from all materials to be screened based on the correlation score value for recall; the correlation score value represents the similarity between the material to be screened and the corresponding trigger material.

[0118] For example, according to the material recommendation request of the target user, by reading the characteristics of the target user's behavior history (specifically, the video sequence materials that the user is interested in in the past week), the video materials that the user is interested in interacting with in the past week (watching time greater than 12s or completion rate greater than 90%) are obtained, which are used as trigger materials trigger mids, and the offline results are read from the database redis and stored in the double cache. The topn similar materials of trigger mids (greater than the offline topN) are recalled from the offline results through i2i (I2i is item2item) as the materials to be screened. After deduplication and filtering, the correlation score value is calculated. Among them, the correlation score value is designed according to the business characteristics, and the formula with the maximum indicator benefit is adopted. For each material to be screened, the content correlation coefficient is determined according to the content distance between the material to be screened and the corresponding trigger material; the correlation coefficient is a score calculated offline, which is the inner product of the two material embedding vectors, and the score range is between 0 and 1. The more similar the two materials are, the closer they are to 1. The time decay factor is determined according to the time difference between the release time of the material to be screened and the corresponding trigger material, and the time decay factor is negatively correlated with the time difference between the release times; the time decay factor is that the closer the material release time is to the current time, the higher the factor score is. The Wilson correlation coefficient is determined according to the relationship between the estimated viewing time of the material to be screened and the duration of the material to be screened, and the Wilson correlation coefficient is positively correlated with the ratio between the estimated viewing time and the duration of the material to be screened; the closer the viewing time is to the duration of the material to be screened, the higher the score is. The interaction coefficient is determined according to the interaction frequency between the user and the material to be screened, and the interaction coefficient is positively correlated with the interaction frequency; the number of likes, comments, and reposts of the material is scored, and the more reposts, comments, and likes, the higher the interaction score coefficient.

[0119] The product of the content correlation coefficient, the time decay factor, the Wilson correlation coefficient and the interaction coefficient is used as the correlation score between the material to be screened and the corresponding trigger material, and the recalled materials are returned after sorting.

[0120] The method of the embodiment of the present invention is easy to implement, has low engineering implementation cost, has good personalized recall capability, has good algorithm interpretability, and has a very short online time consumption of only about 5ms.

[0121] 3. Experimental Data

[0122] The eges graph algorithm is used to construct similar materials offline, and the recall rate of materials is 2% higher than the recall@100 of the existing technology, and the recall rate is significantly improved; and because materials with interactive behaviors are used as similar materials for offline construction and the material sequence is constructed using the walk probability of the edge between the recommendation position and the corresponding diversion position, and the corresponding material embedding vector is constructed using multiple features of the material, many cold start materials (with small exposure) can be recalled. According to the model online experiment, it can be seen that the online indicators are also significantly improved. Table 1, Table 2 and Table 3 are the experimental indicator data.

[0123] Table 1

[0124] Scale data Effective playback volume of recommended positions Effective playback time of recommended position income +0.4％ +0.38％

[0125] Table 2

[0126] Efficiency data ROI_3S VV ROI_3S duration Recommended reading number after ROI Average number of reads per person after recommendation income +0.53％ +0.52％ +0.53％ +0.39％

[0127] Table 3

[0128] Single channel data Exposure ratio 3S_VV Recommended UV income 19.33％(+6.8％) 3368w(+7.2％) 166w(+5.6％)

[0129] The beneficial technical effects achieved by the embodiments of the present invention are as follows:

[0130] 1. Compared with the graph algorithm using deepwalk, the embodiment of the present invention adopts the eges graph algorithm to construct similar materials offline, and uses multiple features of the material in addition to the node itself, so as to have a stronger ability to characterize the material. At the same time, it can solve some material cold start problems, and unpopular materials can be recalled.

[0131] 2. According to the particularity of the recommendation scenario after a Weibo video, the embodiment of the present invention constructs a bipartite graph of materials for the diversion position and materials for the recommendation position, which improves the material similarity effect to a certain extent and significantly improves the recall rate.

[0132] 3. Use faiss to build an index to obtain the top N similar materials for each material. Faiss has a fast construction speed and can meet the requirements of recalling a large number of materials and calculating a large number of vectors.

[0133] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of protection of the present disclosure. The attached method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0134] In the above detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are clearly stated in each claim. On the contrary, as reflected in the appended claims, the invention is in a state of having less than all the features of the disclosed individual embodiments. Therefore, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0135] The disclosed embodiments are described above to enable any person skilled in the art to implement or use the present invention. Various modifications of these embodiments are obvious to those skilled in the art, and the general principles defined herein may also be applied to other embodiments without departing from the spirit and scope of the present disclosure. Therefore, the present disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0136] The above description includes examples of one or more embodiments. Of course, it is impossible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but it should be recognized by those skilled in the art that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to cover all such changes, modifications and variations that fall within the scope of protection of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, the word is covered in a manner similar to the term "including", just as "including," is explained as a transitional word in the claims. In addition, any term "or" used in the specification of the claims is intended to mean "non-exclusive or".

[0137] Those skilled in the art may also understand that the various illustrative logical blocks, units, and steps listed in the embodiments of the present invention may be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly demonstrate the interchangeability of hardware and software, the various illustrative components, units, and steps described above have generally described their functions. Whether such functions are implemented by hardware or software depends on the specific application and the design requirements of the entire system. Those skilled in the art may use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the scope of protection of the embodiments of the present invention.

[0138] The various illustrative logic blocks or units described in the embodiments of the present invention can be implemented or operated by a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field programmable gate array or other programmable logic device, a discrete gate or transistor logic, a discrete hardware component, or any combination of the above. The general-purpose processor can be a microprocessor, and optionally, the general-purpose processor can also be any conventional processor, controller, microcontroller or state machine. The processor can also be implemented by a combination of computing devices, such as a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a digital signal processor core, or any other similar configuration.

[0139] The steps of the method or algorithm described in the embodiments of the present invention can be directly embedded in hardware, a software module executed by a processor, or a combination of the two. The software module can be stored in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or other storage media of any form in the art. Exemplarily, the storage medium can be connected to the processor so that the processor can read information from the storage medium and can write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. The processor and the storage medium can be arranged in an ASIC, and the ASIC can be arranged in a user terminal. Optionally, the processor and the storage medium can also be arranged in different components in the user terminal.

[0140] In one or more exemplary designs, the above functions described in the embodiments of the present invention can be implemented in hardware, software, firmware or any combination of the three. If implemented in software, these functions can be stored on a computer-readable medium, or transmitted in the form of one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media that facilitate the transfer of computer programs from one place to another. The storage medium can be any available medium that can be accessed by any general or special computer. For example, such computer-readable media can include but are not limited to RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store program codes in the form of instructions or data structures and other forms that can be read by general or special computers, or general or special processors. In addition, any connection can be appropriately defined as a computer-readable medium, for example, if the software is transmitted from a website site, server or other remote resource through a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL) or wirelessly, such as infrared, wireless and microwave, it is also included in the defined computer-readable medium. The disk and disc include compact disk, laser disk, optical disk, DVD, floppy disk and blue-ray disk. Disks usually copy data magnetically, while discs usually copy data optically with lasers. The above combination can also be included in computer readable media.

[0141] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A material recall method, characterized in that: include: Receive a material recommendation request from a target user, and based on the interaction behavior data of the target user, obtain a material with which the target user has an interaction behavior within a second time range and which meets a set condition as a trigger material; Based on a periodically updated material database, similar materials corresponding to each trigger material are determined, and similar materials corresponding to all trigger materials are used as materials to be screened; each material within a first time range and at least one similar material corresponding thereto are recorded in the material database; the material database is generated by inputting the interaction behavior data of all users within the first time range into an eges graph algorithm, and training to obtain a material embedding vector corresponding to each material; Determine the correlation score between each material to be screened and the corresponding trigger material, and determine the recall material corresponding to the target user from all materials to be screened based on the correlation score for recall; the correlation score represents the similarity between the material to be screened and the corresponding trigger material.

2. The material recall method according to claim 1, characterized in that: Also includes: Periodically obtain the interactive behavior data of all users within the first time range, and construct a bipartite graph of materials including diversion positions and recommendation positions through the Eges graph algorithm; the diversion position represents the materials displayed in any primary scene, and the recommendation position represents the recommended materials in the secondary scene through specific behaviors on the diversion position; A material embedding vector corresponding to each material is obtained based on the material bipartite graph and the characteristic information of each material, and a similarity calculation is performed based on the material embedding vector corresponding to each material to obtain at least one similar material corresponding to each material, and the similar material is written into the material database.

3. The material recall method according to claim 2, characterized in that: The periodic acquisition of the interaction behavior data of all users within the first time range and the construction of a bipartite graph of materials including diversion positions and recommendation positions by using an eges graph algorithm include: Periodically obtain the interactive behavior data of all users within the first time range, and use the eges graph algorithm to determine all diversion positions, where the diversion positions represent the materials displayed in any first-level scene; According to each of the diversion positions, a recommended position corresponding to the diversion position is determined, wherein the recommended position represents the recommended material in the secondary scene after the material enters the diversion position through a specific behavior of the diversion position; Each recommendation position, each diversion position, and the edge between the recommendation position and the corresponding diversion position are constructed to obtain the material bipartite graph; the wandering probability weight of the edge between the recommendation position and the corresponding diversion position is: the average viewing time of the recommendation position within the first time range.

4. The material recall method according to claim 3, characterized in that: The obtaining a material embedding vector corresponding to each material based on the material bipartite graph and the feature information of each material includes: For each material in the material bipartite graph, the material is taken as the starting material, and the biased random walk is used to walk along the edges in the material bipartite graph, and a material sequence corresponding to the material is sequentially constructed according to the materials passed by the continuous walk; For each group of material sequences, a skip-gram model is used for training based on the feature information of each material in the material sequence, and a material embedding vector corresponding to each material in the material sequence is output.

5. The material recall method according to claim 3, characterized in that: For each material in the material bipartite graph, the material is taken as the starting material, and according to the biased random walk, the material sequence corresponding to the material is sequentially constructed according to the materials passed by the continuous walk. Specifically, it includes: For each material in the material bipartite graph, the material is taken as the starting material, and whether to wander to the next material connected to the previous material by an edge is determined based on the wandering probability weight of the edge; When it is determined based on the edge walking probability weight to walk to the next material connected to the previous material by an edge, walk backward in sequence based on the edge walking probability weight, and form a material sequence corresponding to the material according to the materials passed through by continuous walking, until the number of materials in the material sequence reaches the preset material number.

6. The material recall method according to claim 4, characterized in that: For each group of material sequences, training is performed through a skip-gram model based on the feature information of each material in the material sequence, and a material embedding vector corresponding to each material in the material sequence is output, including: For each group of material sequences, according to the set sliding window length, the starting material of the material sequence is used as the starting material of the current sliding window, the first number of materials of the sliding window length are sequentially used as material positive samples of the current sliding window, the starting materials are respectively used with other materials in the current sliding window to form positive sample pairs, material negative samples are randomly selected from the materials in all material sequences, the starting materials and the material negative samples are used to form positive and negative sample pairs, the positive sample pairs in the current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the current sliding window is output, and the current material embedding vector of the material is the vector mean of all feature information of the material, wherein the feature information includes: material node information, material author ID, primary content label corresponding to the material, secondary content label corresponding to the material, and tertiary content label corresponding to the material; After obtaining the current material embedding vector of each material in the current sliding window through the skip-gram model, the current sliding window is changed to the previous sliding window, the previous sliding window is slid backward by one position in the material sequence to use the second material in the material sequence as the starting material of the new current sliding window, the starting material is respectively used with other materials in the current sliding window to form a positive sample pair, a material negative sample is randomly selected from the materials in all material sequences, the starting material and the material negative sample are used to form a positive and negative sample pair, the positive sample pair in the new current sliding window and the feature information corresponding to each material in the positive and negative sample pairs are input into the skip-gram model, and the current material embedding vector of each material in the new current sliding window is output; wherein, for the material used as a positive sample pair or a positive and negative sample pair in the previous sliding window, the current material embedding vector of the corresponding output material is input into the skip-gram model as the feature information of the material; Until the last material of the material sequence is slid into the new current sliding window, the current material embedding vector of each material in the new current sliding window is output; The current material embedding vector of each material in the material sequence is used as the final material embedding vector of the material and outputted.

7. The material recall method according to claim 2, characterized in that: The similarity calculation is performed based on the material embedding vector corresponding to each material to obtain at least one similar material corresponding to each material, including: For each material in the material bipartite graph, faiss is used to calculate the distance between the material embedding vector corresponding to the material and the material embedding vectors corresponding to other materials, the distances are sorted from small to large, a preset number of other materials before the sorting are taken as similar materials of the material, and an index of similar materials of each material is constructed.

8. The material recall method according to claim 1, characterized in that: Determining the correlation score between each to-be-screened material and the corresponding trigger material includes: For each material to be screened, determining a content correlation coefficient according to a content distance between the material to be screened and a corresponding trigger material; Determine a time decay factor according to the time difference between the release time of the material to be screened and the corresponding trigger material, wherein the time decay factor is negatively correlated with the time difference between the release times; Determine the Wilson correlation coefficient according to the relationship between the estimated viewing time of the material to be screened and the duration of the material to be screened, wherein the Wilson correlation coefficient is positively correlated with the ratio between the estimated viewing time and the duration of the material to be screened; Determine an interaction coefficient according to the interaction frequency between the user and the material to be screened, wherein the interaction coefficient is positively correlated with the interaction frequency; The product of the content correlation coefficient, the time decay factor, the Wilson correlation coefficient and the interaction coefficient is used as the correlation score value between the material to be screened and the corresponding trigger material.

9. A material recall system, characterized in that: include: A trigger material acquisition unit, configured to receive a material recommendation request from a target user, and based on the interaction behavior data of the target user, acquire a material with which the target user has an interaction behavior within a second time range and which meets a set condition as a trigger material; The first material recall unit is used to determine similar materials corresponding to each trigger material based on a periodically updated material database, and use similar materials corresponding to all trigger materials as materials to be screened; the material database records each material within the first time range and at least one similar material corresponding thereto; the material database is generated by inputting the interaction behavior data of all users within the first time range into an eges graph algorithm, and training to obtain a material embedding vector corresponding to each material; The second material recall unit is used to determine the correlation score value between each material to be screened and the corresponding trigger material, and determine the recall material corresponding to the target user from all materials to be screened based on the correlation score value for recall; the correlation score value represents the similarity between the material to be screened and the corresponding trigger material.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a computer device, the computer device executes the material recall method according to any one of claims 1 to 8.

Images