Information processing method and apparatus, storage medium, and electronic device

Abstract

The application discloses an information processing method and device, a storage medium and electronic equipment. The method comprises the following steps: obtaining first interaction information between a target account and a plurality of items, and a knowledge graph corresponding to the plurality of items; based on the first interaction information and the knowledge graph, a target recommendation model is used to obtain a target recommendation result corresponding to the target account; in the model training process, the model loss comprises a first loss value determined based on an item relationship matrix and a user-item interaction matrix, and a second loss value determined based on a plurality of account prediction preference scores and actual preference scores of a plurality of items. The application solves the technical problem that in the related art, when a recommendation model is trained, a high-order neighbor cooperation signal is used to learn rich user embedding and item embedding, which causes excessive smoothing, training performance degradation, difficulty in model convergence, and inaccurate recommendation results.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to an information processing method, apparatus, storage medium, and electronic device. Background Technology

[0002] With the development of graph convolutional networks (GCNNs), related technologies have attempted to utilize GCNNs to model knowledge graphs and align their optimization objectives with those of recommendation systems, thereby improving recommendation performance. The core of GCNN-based models lies in their message-passing mechanism—aggregating neighborhood information. This is an effective method for extracting additional information from graph structures, enriching user and item representation learning, and alleviating sparsity problems. However, while using GCNNs for recommendations can leverage the collaborative signals of higher-order neighbors to learn rich user and item representations, it suffers from oversmoothing, where the representations of all nodes tend to become similar as the number of propagation layers increases, leading to performance degradation. Secondly, as graph features grow, the model struggles to converge, especially in large graphs, easily resulting in low model accuracy and poor prediction results.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides an information processing method, apparatus, storage medium, and electronic device to at least solve the technical problem in related technologies where learning rich user embeddings and item embeddings using the cooperative signals of higher-order neighbors during recommendation model training results in over-smoothing, decreased training performance, difficulty in model convergence, and consequently, inaccurate recommendation results.

[0005] According to one aspect of the present invention, an information processing method is provided, comprising: acquiring first interaction information between a target account and multiple items, and a knowledge graph corresponding to the multiple items, wherein the first interaction information is used to indicate the access status of the target account to the multiple items, and the multiple items include a first number of accessed items and a second number of unaccessed items; based on the first interaction information and the knowledge graph, employing a target recommendation model to obtain a target recommendation result corresponding to the target account, wherein the target recommendation result is used to indicate the preference score of the target account for the multiple items; wherein the target recommendation model is trained based on second interaction information between the multiple accounts and the multiple items, and the knowledge graph; the model loss during the model training process includes: a first loss value determined based on an item relationship matrix and a user-item interaction matrix, and a second loss value determined based on the predicted preference score and actual preference score of the multiple accounts for the multiple items, wherein the item relationship matrix is ​​used to indicate the association relationship between any two items in the multiple items, and the user-item interaction matrix is ​​used to indicate the initial interaction relationship between the multiple accounts and the multiple items.

[0006] According to another aspect of the present invention, an information processing apparatus is also provided, comprising: an acquisition module, configured to acquire first interaction information between a target account and multiple items, and a knowledge graph corresponding to the multiple items, wherein the first interaction information is used to indicate the access status of the target account to the multiple items, and the multiple items include a first number of accessed items and a second number of unaccessed items; a testing module, configured to obtain a target recommendation result corresponding to the target account based on the first interaction information and the knowledge graph, using a target recommendation model, wherein the target recommendation result is used to indicate the preference score of the target account for the multiple items; wherein the target recommendation model is trained based on second interaction information between the multiple accounts and the multiple items, and the knowledge graph; the model loss during the model training process includes: a first loss value determined based on an item relationship matrix and a user-item interaction matrix, and a second loss value determined based on the predicted preference score and actual preference score of the multiple accounts for the multiple items, wherein the item relationship matrix is ​​used to indicate the association relationship between any two items in the multiple items, and the user-item interaction matrix is ​​used to indicate the initial interaction relationship between the multiple accounts and the multiple items.

[0007] According to another aspect of the present invention, a non-volatile storage medium is also provided, wherein the non-volatile storage medium stores a plurality of instructions, the instructions being adapted to be loaded by a processor and executed any one of the above-described information processing methods.

[0008] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described information processing methods.

[0009] In this embodiment of the invention, first interaction information between a target account and multiple projects, and a knowledge graph corresponding to the multiple projects are obtained. The first interaction information indicates the target account's access to the multiple projects, which include a first number of visited projects and a second number of unvisited projects. Based on the first interaction information and the knowledge graph, a target recommendation model is used to obtain a target recommendation result corresponding to the target account. The target recommendation result indicates the target account's preference score for the multiple projects. The target recommendation model is trained based on second interaction information between the multiple accounts and the multiple projects, and the knowledge graph. The model loss during training includes a first loss value determined based on the project relationship matrix and the user-project interaction matrix. The second loss value is determined based on the predicted preference scores and actual preference scores of the aforementioned multiple accounts for multiple items. The aforementioned item relationship matrix is ​​used to indicate the association between any two items in the aforementioned multiple items, and the aforementioned user-item interaction matrix is ​​used to indicate the initial interaction relationship between the aforementioned multiple accounts and the aforementioned multiple items. This achieves the goal of learning high-order signals in the knowledge graph through the constructed item relationship matrix for model convergence judgment, effectively avoiding excessive smoothing in the model training process, and improving model performance. This achieves the technical effect of improving model training accuracy and thus improving the accuracy of recommendation results. In addition, it solves the technical problem in related technologies where learning rich user embeddings and item embeddings using high-order neighbor collaboration signals during recommendation model training results in excessive smoothing, decreased training performance, difficulty in model convergence, and inaccurate recommendation results. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0011] Figure 1 This is a schematic diagram of an information processing method according to an embodiment of the present invention;

[0012] Figure 2 This is a schematic diagram illustrating an optional user representation and item representation interaction according to an embodiment of the present invention;

[0013] Figure 3This is a schematic diagram of an optional information processing method according to an embodiment of the present invention;

[0014] Figure 4 This is a schematic diagram of an information processing device according to an embodiment of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0016] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] In recent years, recommender systems have been widely studied by researchers both domestically and internationally as an effective solution to information overload. Currently, recommender systems play a crucial role in various information systems, promoting business development and supporting decision-making processes, and are widely used in major e-commerce and multimedia websites.

[0018] A knowledge graph is a directed heterogeneous information network containing a large number of entities and relationships. It introduces more semantic relationships between items, enriches item descriptions, and facilitates the discovery of user interests at a deeper level. Secondly, knowledge graphs provide different types of relationship connections, which contributes to the diversity of recommendation results and avoids recommendations being limited to a single type. Finally, with the help of knowledge graphs, reasonable relationships between users and items can be established, improving user satisfaction with recommendation results and enhancing user trust in the recommendation system.

[0019] With the development of graph convolutional networks (GCNNs), related technologies have attempted to utilize GCNNs to model knowledge graphs and align their optimization objectives with those of recommendation systems, thereby improving recommendation performance. The core of GCNN-based models lies in their message-passing mechanism—aggregating neighborhood information. This is an effective method for extracting additional information from graph structures, enriching user and item representation learning, and alleviating sparsity problems. While using GCNNs for recommendation can leverage the collaborative signals of higher-order neighbors to learn rich user and item representations, it suffers from oversmoothing, where the representations of all nodes tend to become similar as the number of propagation layers increases, leading to performance degradation. Furthermore, as graph features grow, the model struggles to converge, especially in large graphs, easily resulting in low model accuracy and poor prediction results.

[0020] According to an embodiment of the present invention, an embodiment of an information processing method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0021] Figure 1 This is a flowchart of an information processing method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0022] Step S12: Obtain first interaction information between the target account and multiple projects, as well as the knowledge graph corresponding to the multiple projects. The first interaction information is used to indicate the access status of the target account to the multiple projects. The multiple projects include a first number of accessed projects and a second number of unaccessed projects.

[0023] Optionally, access operations may include, but are not limited to, click operations, purchase operations, rating operations, watch operations, listen operations, etc., and can be set according to specific application scenarios. For example, in a purchase scenario, the multiple items represent multiple goods, and the corresponding access operations could be click operations, purchase operations, rating operations, etc. In a movie-watching scenario, the multiple items represent multiple different movies, and the corresponding access operations could be click operations, rating operations, watch operations, etc.

[0024] It is understood that the aforementioned first interaction information is used to indicate the target account's access status to the aforementioned multiple projects, i.e., whether the user has accessed any projects. Accessed projects are those the target user has already visited (i.e., historically accessed projects), while unaccessed projects are those the target user has not yet visited. The multiple projects include both accessed and unaccessed projects, allowing for appropriate project recommendations to be made to the user based on their access history to the accessed projects. The aforementioned target account corresponds to the target user, and the aforementioned multiple accounts correspond to multiple users; that is, one account corresponds to one user.

[0025] Step S14: Based on the first interaction information and the knowledge graph, a target recommendation model is used to obtain the target recommendation result corresponding to the target account, wherein the target recommendation result is used to indicate the target account's preference score for the multiple items.

[0026] Optionally, the above-mentioned target recommendation model is trained based on the second interaction information between multiple accounts and the above-mentioned multiple items, as well as the above-mentioned knowledge graph; the model loss during the model training process includes: a first loss value determined based on the item relationship matrix and the user-item interaction matrix, and a second loss value determined based on the predicted preference scores and actual preference scores of the above-mentioned multiple accounts for the multiple items. The above-mentioned item relationship matrix is ​​used to indicate the association between any two items in the above-mentioned multiple items, and the above-mentioned user-item interaction matrix is ​​used to indicate the initial interaction relationship between the above-mentioned multiple accounts and the above-mentioned multiple items.

[0027] Optionally, the target recommendation results are used to indicate the target account's preference scores for the above multiple items, which can reflect the probability value of recommending each item to the target account.

[0028] Optionally, based on the aforementioned first interaction information and the aforementioned knowledge graph, the target recommendation model is used to obtain the target recommendation result corresponding to the aforementioned target account, including: establishing a target user-item interaction matrix based on the first interaction information between the target account and multiple items; inputting the target user-item interaction matrix into the target recommendation model to obtain the target recommendation result corresponding to the aforementioned target account. In the target user-item interaction matrix, the value of the matrix bit corresponding to the interacted item is 1, and the value of the matrix bit corresponding to the uninteracted item is 0.

[0029] By using the above methods, during the training of the recommendation model, the model loss is determined by combining the first loss value based on the item relationship matrix and the user item interaction matrix, and the second loss value based on the predicted preference scores and actual preference scores of multiple accounts for multiple items. This effectively avoids the problem of over-smoothing during model training, resulting in a target recommendation model with higher accuracy. The recommendation results predicted by this model are more accurate and reliable, and can better meet user preferences.

[0030] Through steps S12 to S14 above, the high-order signals in the knowledge graph can be learned by constructing the item relationship matrix to judge model convergence, effectively avoiding excessive smoothing in the model training process and improving model performance. This achieves the technical effect of improving model training accuracy and thus improving the accuracy of recommendation results. It also solves the technical problem in related technologies where learning rich user embeddings and item embeddings by using the collaborative signals of high-order neighbors during recommendation model training results in excessive smoothing, decreased training performance, difficulty in model convergence, and inaccurate recommendation results.

[0031] In an optional embodiment, before obtaining the target recommendation result corresponding to the target account using a pre-trained target recommendation model based on the first interaction information and the knowledge graph, the method further includes:

[0032] Step S141: Obtain the above-mentioned user item interaction matrix and the above-mentioned knowledge graph.

[0033] Optionally, the aforementioned user-item interaction matrix is ​​used to indicate the initial interaction relationship between multiple accounts (corresponding to multiple users) and the aforementioned multiple items, and can be represented by matrix Y∈R. m×n Let u = {u1, u2, ..., u3} represent the number of users, specifically denoted as u = {u1, u2, ..., u3}. m}; n represents the number of items, that is, there are n items in the item set, specifically represented as v = {v1, v2, ..., v...} n The user project interaction matrix Y is represented as Y = {y} uv |u∈U,v∈V},Y uv =1 indicates that there is interaction between user u and project v (i.e., user u has visited project v); otherwise, there is no interaction.

[0034] It is understandable that the user-item interaction matrix and knowledge graph are the foundation for item recommendation. The user-item interaction matrix provides the recommendation model with a direct representation of user interests. Furthermore, the knowledge graph can introduce rich semantic relationships into the recommendation model, which helps to discover user interests at a deeper level. The knowledge graph is represented by the symbol G. The knowledge graph consists of a large number of entity-relation entity triples, represented as {G=(h,r,t)|h,t∈ε,r∈R}, where h and t represent the head entity and tail entity, respectively, r represents the relation, and ε and R represent the entity set and relation set, respectively.

[0035] Step S142: The initial recommendation model is trained using the aforementioned user-item interaction matrix and the aforementioned knowledge graph to obtain the trained recommendation model, the predicted preference scores of the aforementioned multiple accounts for the aforementioned multiple items, the first user representation corresponding to the aforementioned multiple accounts, and the first item representation corresponding to the aforementioned multiple items.

[0036] Optionally, the user-item interaction matrix and the knowledge graph are input into the initial recommendation model, trained for a preset number of rounds, and the trained recommendation model corresponding to the last round of training is output, along with the predicted preference scores of multiple accounts for multiple items, the first user representation corresponding to each account, and the first item representation corresponding to each item.

[0037] In an optional embodiment, the process of training the initial recommendation model using the user item interaction matrix and the knowledge graph includes:

[0038] Step S1421: Based on the knowledge graph described above, determine the first-order neighbor representations corresponding to the above multiple items.

[0039] Optionally, multiple items correspond to entities in the knowledge graph. When determining the first item representation of any one of the multiple items based on the knowledge graph, the representations of the neighbor nodes of any one item are aggregated with weights, using that item as the head entity h. To avoid over-smoothing, the representations of first-order neighbors are selected for aggregation to obtain the first-order neighbor representation of any item. The calculation method is as follows:

[0040]

[0041] Among them, e Nh Let e ​​represent the first-order neighbor representation of any item h in the knowledge graph. t Let π(h,r,t) be the entity representation of the neighboring node (i.e., the tail entity t). π(h,r,t) is the aggregation weight, calculated as follows: considering the similarity between the head entity h and the tail entity t in the entity relation r space, the higher the similarity, the greater the weight.

[0042] π(h,r,t)=(Wr e t ) T tanh(W r e h +e r )

[0043] in, W r Let e ​​be the first weight matrix to be optimized. h The entity representation of the header entity (i.e., the initial item representation of any item).

[0044] Step S1422: Based on the first-order neighbor representations corresponding to the above multiple items and the initial item representations corresponding to the above multiple items, the first item representations corresponding to the above multiple items are obtained.

[0045] In an optional embodiment, obtaining the first user representation corresponding to each of the multiple accounts based on the user item interaction matrix and the first item representation corresponding to each of the multiple items includes: determining the visited items corresponding to each of the multiple accounts and the weight value corresponding to the visited items based on the user item interaction matrix; and obtaining the first user representation corresponding to each of the multiple accounts based on the first item representation of the visited items corresponding to each of the multiple accounts and the weight value corresponding to the visited items.

[0046] It is understood that the aforementioned user item interaction matrix is ​​used to indicate the historical access information of multiple accounts. In the process of model training, based on the aforementioned user item interaction matrix, the accessed items corresponding to the aforementioned multiple accounts are determined, and the first user representation corresponding to the multiple users is generated by aggregating the user's historical records, thereby effectively learning user preferences.

[0047] Optionally, the initial item representation and the first-order neighbor representation are aggregated to jointly represent the item, resulting in the first item representation as follows:

[0048] f GCN =Leaky ReLU(W(e h +e Nh ))

[0049] Here, LeakyReLU() is the activation function, and W is the second weight matrix to be optimized.

[0050] Step S1423: Based on the above user item interaction matrix and the first item representations corresponding to the above multiple items respectively, the first user representations corresponding to the above multiple accounts are obtained respectively.

[0051] Based on the user item matrix, weighted aggregated user historical interaction items (i.e., visited items) are used to aggregate and represent users, resulting in the first user representation:

[0052]

[0053] Where N(u) represents the items that have been interacted with (i.e., visited items) by any one of the multiple user accounts u, e i β represents the initial project representation of project i. u,i This represents the weight value when aggregating item i, where d u d represents the number of historical interactions of user u, that is, how many items user u has interacted with. i β represents the number of interactions with item i, i.e., how many users have interacted with item i. u,i Its design is more in line with the idea of ​​collaborative filtering, and it can effectively learn user representations based on user history.

[0054] Step S1424: Based on the first item representations corresponding to the multiple items and the first user representations corresponding to the multiple accounts, obtain the predicted preference scores of the multiple accounts for the multiple items.

[0055] Optionally, based on the first user representations corresponding to multiple users and the first item representations corresponding to multiple items, an inner product is performed to obtain the predicted value of the interaction between user u and item i. The predicted value can be used as the predicted preference score of user u for item i, which is also equivalent to the predicted probability of recommending item i to user u.

[0056]

[0057] in, This represents the predicted preference score for user u for item i. This represents the first user representation corresponding to user u after the predetermined training rounds. This represents the first item representation corresponding to item i after a predetermined training round.

[0058] Optional, Figure 2 This is a schematic diagram illustrating an optional user representation and item representation interaction according to an embodiment of the present invention, such as... Figure 2As shown, for each item (denoted as item i) among multiple items, the first-order neighbor representations of item i (such as v1, v2, v3) are obtained. The initial item representation of item i is updated based on the first-order neighbor representations of item i to obtain the updated item representation of item i (i.e., the first item representation). Based on the user's history, the items that the user has interacted with (i.e., the items the user has visited) are determined. Based on this, the user representation is updated to obtain the updated user representation (i.e., the first user representation). Furthermore, based on the updated user representation and the item representation, the predicted probability of recommending item v to user u is determined.

[0059] Step S143: Based on the above user-item interaction matrix, the above item relationship matrix, the first user representation corresponding to the above multiple accounts, and the first item representation corresponding to the above multiple items, determine the above first loss value.

[0060] It is understandable that learning rich user embeddings and item embeddings using the collaborative signals of high-order neighbors during recommendation model training can lead to problems such as over-smoothing, decreased training performance, and difficulty in model convergence. In this embodiment of the invention, the first item representation corresponding to each of the aforementioned multiple items is determined based on the first-order neighbor representations corresponding to each of the multiple items and the initial item representations corresponding to each of the aforementioned multiple items. That is, during model training, only the first-order neighbor representations of each item are considered, skipping the multi-hop aggregation mechanism in graph convolutional networks. At the same time, during model training, the item relationship matrix is ​​introduced to learn high-order signals in the knowledge graph, thereby improving the efficiency and accuracy of model training.

[0061] Optionally, to better obtain higher-order signals from the knowledge graph, an item relationship matrix between items can be obtained from the knowledge graph, and the first user representations corresponding to the multiple users calculated above can be learned in the relationship matrix. This avoids the oversmoothing problem caused by multi-hop mechanisms and effectively learns higher-order signals from the knowledge graph, further learning user preferences.

[0062] In an optional embodiment, before training the initial recommendation model using the user item interaction matrix and the knowledge graph, the method further includes: determining the association between any two items in the plurality of items based on the knowledge graph, wherein an association is determined to exist between any two items if the number of hops between any two items in the knowledge graph is less than or equal to a preset number of hops; and constructing the item relationship matrix based on the association between any two items.

[0063] Based on the knowledge graph, a project relationship matrix G is constructed. If one entity's two-hop neighbors contain another entity, then a direct relationship exists between the two entities. That is, if the hop interval between project i and project j in the knowledge graph is less than 2 hops, then project i and project j are determined to be related, which is reflected in the target relationship matrix G. ij =1.

[0064] In an optional embodiment, determining the first loss value based on the user-item interaction matrix, the item relationship matrix, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items includes: determining, based on the item relationship matrix, a first item whose similarity to an visited item is greater than a predetermined similarity threshold, and a weight value between the visited item and the first item; determining the first loss value based on the first user representations of the first items corresponding to the visited items corresponding to the multiple accounts, the weight values ​​between the visited items and the first item, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items.

[0065] Optionally, the visited items corresponding to the aforementioned multiple accounts correspond to positive (u,i) pairs in the user item interaction matrix. For each positive (u,i) pair, K first items with a similarity greater than a preset similarity threshold to item i are extracted to form a first item set S(i). Then, a weighted loss is calculated on the first items corresponding to item i to obtain the first loss value. The first loss value can be obtained, but is not limited to, through the following methods:

[0066]

[0067]

[0068] Among them, L I N represents the second loss value; + This indicates a positive (u,i) pair, corresponding to the items visited by user u; G represents the weight between project i and project j. i and g j These represent the degrees of project i and project j in the project relationship matrix G, respectively, indicating the number of projects in the project relationship matrix G that are associated with project i and project j, respectively. i,j G represents the value in the i-th row and j-th column of the project relationship matrix. This value is related to whether there is a relationship between project i and project j. If there is a relationship, then G... i,j =1, otherwise G i,j =0.

[0069] Step S144: Based on the predicted preference scores and actual preference scores of the multiple accounts for the multiple items, determine the second loss value.

[0070] Optionally, after predicting user behavior, a loss function needs to be set to enhance the learning ability of the recommendation algorithm. A negative sampling strategy is adopted (interactions present in the user's second interaction graph are considered positive samples, i.e., those with a predicted preference score greater than 0; interactions not present in the user's interaction graph are considered negative samples, i.e., those with a predicted preference score less than or equal to 0). The second loss value is calculated using the BCE loss function as follows:

[0071]

[0072] Among them, L KG This represents the second loss value. y represents the predicted preference score of user u for item i. ui This represents the actual preference score of user u for item i, where item i corresponds to the positive sample N with a predicted preference score greater than 0. + ; y represents the predicted preference score for user u on item j. uj This represents the actual preference score of user u for item j, where item j corresponds to the negative sample N with a predicted preference score less than or equal to 0. - .

[0073] Step S145: If the first loss value and the second loss value satisfy the preset loss condition, the trained recommendation model is used as the target recommendation model.

[0074] In one optional embodiment, the step of using the trained recommendation model as the target recommendation model when the first loss value and the second loss value satisfy a preset loss condition includes: determining the model loss based on the first loss value and the second loss value; and using the trained recommendation model as the target recommendation model when the model loss is greater than a preset loss threshold.

[0075] Optionally, based on the first loss value and the second loss value mentioned above, the model loss can be obtained in the following way:

[0076]

[0077] Where γ is a preset parameter used for regularization, and θ is a preset parameter set.

[0078] Optionally, if the model loss determined based on the first loss value and the second loss value meets the preset loss condition, it indicates that the model has reached the convergence condition, and the trained recommendation model can be used as the target recommendation model; if the model loss does not meet the preset loss condition, it indicates that the trained recommendation model obtained at this time has not reached convergence, and the model needs to be trained again until the trained model reaches convergence.

[0079] Based on the above embodiments and optional embodiments, the present invention proposes an optional implementation method. Figure 3 This is a flowchart of an optional information processing method according to an embodiment of the present invention, such as... Figure 3 As shown, this method includes a model training phase and a model application phase, specifically including:

[0080] During the model training phase, the user representation and the item representations of each item in the knowledge graph are first initialized to obtain the initial user representation and the initial item representations corresponding to multiple items. The first-order neighbor representations of multiple items in the knowledge graph are then obtained. Based on the first-order neighbor representations and the initial item representations, the item representations are updated to obtain the first item representations corresponding to each item. Based on the first item representations corresponding to multiple items and the historical interaction records between multiple users and multiple items, the user representation is updated to obtain the first user representations corresponding to each user. The historical interaction records indicate the interaction between users and multiple items, i.e., the user's access to multiple items. The recommendation probability values ​​for each item corresponding to each user are determined based on the first user representations and the first item representations corresponding to multiple items. The initial recommendation model is trained and optimized using the loss of multiple users on the knowledge graph relation matrix (i.e., the item relation matrix) (i.e., the first loss value) and the loss of the user's item recommendation probability (i.e., the second loss value, equivalent to the preference score loss). The trained recommendation model is then obtained. Under the condition of meeting the preset loss conditions, the model parameters are optimized, and the trained recommendation model is used as the target recommendation model. In the model application phase, for a given target user, the historical interaction information between the target user and multiple items, as well as the knowledge graph, are input into the target recommendation model. The output is the recommendation probability of recommending multiple items to the target account. The recommendation probability can be used as the target account's preference score for multiple items.

[0081] It is understood that the embodiments of the present invention provide a simplified model training method for training recommendation models. Compared with existing methods for obtaining recommendation models, the embodiments of the present invention have at least the following technical advantages: when obtaining item representations from the knowledge graph, multi-layer aggregation mechanisms are skipped, which can effectively reduce over-smoothing; the user-item interaction matrix is ​​fully utilized, and user representations are generated by aggregating users' historical access records, effectively learning user preferences; finally, an item relationship matrix is ​​constructed based on the knowledge graph, which effectively learns high-order relationships between items by removing the message passing mechanism.

[0082] This embodiment also provides an information processing apparatus for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the terms "module" and "apparatus" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0083] According to an embodiment of the present invention, an apparatus embodiment for implementing the above-described information processing method is also provided. Figure 4 This is a schematic diagram of the structure of an information processing device according to an embodiment of the present invention, such as... Figure 4 As shown, the above-mentioned information processing device includes: an acquisition module 400 and a testing module 402, wherein:

[0084] The aforementioned acquisition module 400 is used to acquire first interaction information between the target account and multiple projects, as well as the knowledge graph corresponding to the multiple projects. The first interaction information is used to indicate the access status of the target account to the multiple projects. The multiple projects include a first number of accessed projects and a second number of unaccessed projects.

[0085] The aforementioned test module 402, connected to the aforementioned acquisition module 400, is used to obtain the target recommendation result corresponding to the target account based on the aforementioned first interaction information and the aforementioned knowledge graph, using a target recommendation model. The target recommendation result indicates the target account's preference score for the multiple items. The target recommendation model is trained based on the second interaction information between the multiple accounts and the multiple items, and the aforementioned knowledge graph. The model loss during training includes: a first loss value determined based on the item relationship matrix and the user-item interaction matrix, and a second loss value determined based on the predicted and actual preference scores of the multiple accounts for the multiple items. The item relationship matrix indicates the association between any two items among the multiple items, and the user-item interaction matrix indicates the initial interaction relationship between the multiple accounts and the multiple items.

[0086] In this embodiment of the invention, the acquisition module 400 is configured to acquire first interaction information between a target account and multiple projects, and a knowledge graph corresponding to the multiple projects. The first interaction information indicates the target account's access to the multiple projects, and the multiple projects include a first number of visited projects and a second number of unvisited projects. The testing module 402, connected to the acquisition module 400, is configured to use a target recommendation model based on the first interaction information and the knowledge graph to obtain a target recommendation result corresponding to the target account. The target recommendation result indicates the target account's preference score for the multiple projects. The target recommendation model is trained based on second interaction information between the multiple accounts and the multiple projects, and the knowledge graph. The model loss during training includes: based on project relationships... The first loss value is determined by the matrix and the user-item interaction matrix, and the second loss value is determined based on the predicted preference scores and actual preference scores of the multiple accounts for the multiple items. The item relationship matrix is ​​used to indicate the association between any two items in the multiple items, and the user-item interaction matrix is ​​used to indicate the initial interaction between the multiple accounts and the multiple items. This achieves the goal of learning high-order signals in the knowledge graph through the constructed item relationship matrix for model convergence judgment, effectively avoiding excessive smoothing in the model training process, and improving model performance. This achieves the technical effect of improving model training accuracy and thus improving the accuracy of recommendation results. It also solves the technical problem in related technologies where learning rich user embeddings and item embeddings using high-order neighbor collaboration signals during recommendation model training results in excessive smoothing, decreased training performance, difficulty in model convergence, and inaccurate recommendation results.

[0087] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0088] It should be noted that the acquisition module 400 and the testing module 402 mentioned above correspond to steps S12 to S14 in the embodiments. The instances and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules, as part of the device, can run on a computer terminal.

[0089] It should be noted that the optional or preferred implementation methods of this embodiment can be found in the relevant descriptions in the embodiments, and will not be repeated here.

[0090] The aforementioned information processing device may further include a processor and a memory. The aforementioned acquisition module 400, test module 402, etc., are all stored in the memory as program modules, and the processor executes the aforementioned program modules stored in the memory to realize the corresponding functions.

[0091] The processor contains a core that retrieves the corresponding program modules from memory. One or more cores may be configured. Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip.

[0092] According to an embodiment of this application, an embodiment of a non-volatile storage medium is also provided. Optionally, in this embodiment, the non-volatile storage medium includes a stored program, wherein, when the program is running, it controls the device where the non-volatile storage medium is located to execute any of the aforementioned information processing methods.

[0093] Optionally, in this embodiment, the non-volatile storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals, and the non-volatile storage medium includes stored programs.

[0094] Optionally, during program execution, the device containing the non-volatile storage medium performs the following functions: acquiring first interaction information between the target account and multiple items, and a knowledge graph corresponding to the multiple items, wherein the first interaction information is used to indicate the target account's access to the multiple items, and the multiple items include a first number of accessed items and a second number of unaccessed items; based on the first interaction information and the knowledge graph, using a target recommendation model to obtain a target recommendation result corresponding to the target account, wherein the target recommendation result is used to indicate the target account's preference score for the multiple items; wherein the target recommendation model is trained based on second interaction information between the multiple accounts and the multiple items, and the knowledge graph; the model loss during model training includes: a first loss value determined based on the item relationship matrix and the user-item interaction matrix, and a second loss value determined based on the predicted preference score and actual preference score of the multiple accounts for the multiple items, wherein the item relationship matrix is ​​used to indicate the association between any two items in the multiple items, and the user-item interaction matrix is ​​used to indicate the initial interaction relationship between the multiple accounts and the multiple items.

[0095] According to an embodiment of this application, an embodiment of a processor is also provided. Optionally, in this embodiment, the processor is used to run a program, wherein the program executes any of the above-described information processing methods during runtime.

[0096] According to an embodiment of this application, an embodiment of a computer program product is also provided, which, when executed on a data processing device, is suitable for executing a program that initializes any of the above-described information processing method steps.

[0097] Optionally, when the aforementioned computer program product is executed on a data processing device, it is suitable to execute an initialization program with the following method steps: obtaining first interaction information between a target account and multiple items, and a knowledge graph corresponding to the multiple items, wherein the first interaction information is used to indicate the target account's access to the multiple items, and the multiple items include a first number of accessed items and a second number of unaccessed items; based on the first interaction information and the knowledge graph, using a target recommendation model to obtain a target recommendation result corresponding to the target account, wherein the target recommendation result is used to indicate the target account's preference score for the multiple items; wherein the target recommendation model is trained based on second interaction information between the multiple accounts and the multiple items, and the knowledge graph; the model loss during model training includes: a first loss value determined based on an item relationship matrix and a user-item interaction matrix, and a second loss value determined based on the predicted preference score and actual preference score of the multiple accounts for the multiple items, wherein the item relationship matrix is ​​used to indicate the association between any two items in the multiple items, and the user-item interaction matrix is ​​used to indicate the initial interaction relationship between the multiple accounts and the multiple items.

[0098] This invention provides an electronic device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: acquiring first interaction information between a target account and multiple items, and a knowledge graph corresponding to the multiple items; wherein the first interaction information indicates the target account's access to the multiple items, and the multiple items include a first number of accessed items and a second number of unaccessed items; based on the first interaction information and the knowledge graph, using a target recommendation model to obtain a target recommendation result corresponding to the target account; wherein the target recommendation result indicates the target account's preference score for the multiple items; wherein the target recommendation model is trained based on second interaction information between the multiple accounts and the multiple items, and the knowledge graph; the model loss during training includes a first loss value determined based on an item relationship matrix and a user-item interaction matrix, and a second loss value determined based on the predicted preference score and actual preference score of the multiple accounts for the multiple items; the item relationship matrix indicates the association between any two items in the multiple items, and the user-item interaction matrix indicates the initial interaction relationship between the multiple accounts and the multiple items.

[0099] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0100] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of modules described above can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between modules, and may be electrical or other forms.

[0102] The modules described above as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0103] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0104] If the aforementioned integrated modules are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable non-volatile storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a non-volatile 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 steps of the methods of the various embodiments of this invention. The aforementioned non-volatile storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0105] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An information processing method, characterized in that, include: Obtain first interaction information between a target account and multiple projects, as well as a knowledge graph corresponding to the multiple projects. The first interaction information is used to indicate the target account's access status to the multiple projects. The multiple projects include a first number of accessed projects and a second number of unaccessed projects. Based on the first interaction information and the knowledge graph, a target recommendation model is used to obtain the target recommendation result corresponding to the target account, wherein the target recommendation result is used to indicate the target account's preference score for the multiple items; The target recommendation model is trained based on the second interaction information between multiple accounts and the multiple items, as well as the knowledge graph. The model loss during training includes a first loss value determined based on the item relationship matrix and the user-item interaction matrix, and a second loss value determined based on the predicted preference scores and actual preference scores of the multiple accounts for the multiple items. The item relationship matrix is ​​used to indicate the association between any two items in the multiple items, and the user-item interaction matrix is ​​used to indicate the initial interaction relationship between the multiple accounts and the multiple items. Before obtaining the target recommendation result corresponding to the target account using a pre-trained target recommendation model based on the first interaction information and the knowledge graph, the method further includes: acquiring the user-item interaction matrix and the knowledge graph; training an initial recommendation model using the user-item interaction matrix and the knowledge graph to obtain a trained recommendation model, the predicted preference scores of the multiple accounts for the multiple items, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items, wherein the first item representations corresponding to the multiple items are obtained based on the first-order neighbor representations corresponding to the multiple items and the initial item representations corresponding to the multiple items; determining a first loss value based on the user-item interaction matrix, the item relationship matrix, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items; determining a second loss value based on the predicted preference scores and actual preference scores of the multiple accounts for the multiple items; and using the trained recommendation model as the target recommendation model when the first loss value and the second loss value satisfy a preset loss condition.

2. The method according to claim 1, characterized in that, The process of training the initial recommendation model using the user item interaction matrix and the knowledge graph includes: Based on the knowledge graph, the first-order neighbor representations corresponding to the multiple items are determined respectively; Based on the first-order neighbor representations corresponding to the multiple items and the initial item representations corresponding to the multiple items, the first item representations corresponding to the multiple items are obtained. Based on the user item interaction matrix and the first item representations corresponding to the multiple items, the first user representations corresponding to the multiple accounts are obtained. Based on the first item representation corresponding to each of the multiple items and the first user representation corresponding to each of the multiple accounts, the predicted preference scores of the multiple accounts for the multiple items are obtained.

3. The method according to claim 1, characterized in that, Before training the initial recommendation model using the user item interaction matrix and the knowledge graph, the method further includes: Based on the knowledge graph, the association between any two projects in the plurality of projects is determined, wherein the association between any two projects is determined when the number of hops between any two projects in the knowledge graph is less than or equal to a preset number of hops. Construct the project relationship matrix based on the association between any two projects.

4. The method according to claim 2, characterized in that, The step of obtaining the first user representation corresponding to each of the multiple accounts based on the user item interaction matrix and the first item representation corresponding to each of the multiple items includes: Based on the user item interaction matrix, determine the accessed items corresponding to the multiple accounts, and the weight value corresponding to the accessed items. Based on the first item representation of the visited items corresponding to the multiple accounts, and the weight value corresponding to the visited items, the first user representation corresponding to the multiple accounts is obtained.

5. The method according to claim 4, characterized in that, The step of determining the first loss value based on the user-item interaction matrix, the item relationship matrix, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items includes: Based on the project relationship matrix, a first project with a similarity greater than a predetermined similarity threshold with the visited project is determined, and a weight value is assigned between the visited project and the first project. The first loss value is determined based on the first user representation of the first project corresponding to the visited projects of the multiple accounts, the weight value between the visited projects and the first project, the first user representation of the multiple accounts, and the first project representation of the multiple projects.

6. The method according to any one of claims 1 to 5, characterized in that, The step of using the trained recommendation model as the target recommendation model when the first loss value and the second loss value satisfy a preset loss condition includes: Based on the first loss value and the second loss value, determine the model loss; If the model loss is greater than a preset loss threshold, the trained recommendation model is used as the target recommendation model.

7. An information processing device, characterized in that, include: The acquisition module is used to acquire first interaction information between a target account and multiple projects, as well as a knowledge graph corresponding to the multiple projects. The first interaction information is used to indicate the access status of the target account to the multiple projects. The multiple projects include a first number of accessed projects and a second number of unaccessed projects. The testing module is used to obtain a target recommendation result corresponding to the target account based on the first interaction information and the knowledge graph, using a target recommendation model. The target recommendation result indicates the target account's preference score for the multiple items. The target recommendation model is trained based on second interaction information between the multiple accounts and the multiple items, and the knowledge graph. The model loss during training includes a first loss value determined based on the item relationship matrix and the user-item interaction matrix, and a second loss value determined based on the predicted and actual preference scores of the multiple accounts for the multiple items. The item relationship matrix indicates the association between any two items in the multiple items, and the user-item interaction matrix indicates the initial interaction relationship between the multiple accounts and the multiple items. The device is further configured to, before obtaining the target recommendation result corresponding to the target account using a pre-trained target recommendation model based on the first interaction information and the knowledge graph, acquire the user-item interaction matrix and the knowledge graph; train an initial recommendation model using the user-item interaction matrix and the knowledge graph to obtain a trained recommendation model, the predicted preference scores of the multiple accounts for the multiple items, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items, wherein the first item representations corresponding to the multiple items are obtained based on the first-order neighbor representations corresponding to the multiple items and the initial item representations corresponding to the multiple items; determine a first loss value based on the user-item interaction matrix, the item relationship matrix, the first user representations corresponding to the multiple accounts, and the first item representations corresponding to the multiple items; determine a second loss value based on the predicted preference scores and actual preference scores of the multiple accounts for the multiple items; and, if the first loss value and the second loss value satisfy a preset loss condition, use the trained recommendation model as the target recommendation model.

8. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions, which are adapted to be loaded by a processor and executed by the information processing method according to any one of claims 1 to 6.

9. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the information processing method according to any one of claims 1 to 6.

Images