Product interaction prediction method and device, equipment and storage medium

Abstract

The invention relates to a product interaction prediction method and device, computer equipment, a storage medium and a computer program product. The method can be used for scenes of cloud technology, artificial intelligence, intelligent traffic and the like. The method comprises the following steps: obtaining an interaction information sequence obtained by performing product interaction on a pushed product by a target object; determining the correlation between the product information of a to-be-pushed product and each piece of interaction information in the interaction information sequence, and extracting an original interaction mode related to the to-be-pushed product from the interaction information sequence based on the correlation; converting the original interaction mode into a standardized interaction mode; determining a target interaction mode corresponding to the to-be-pushed product, and performing mode matching on the standardized interaction mode and the target interaction mode to obtain mode matching features; and determining the probability of interaction between the target object and the to-be-pushed product based on the mode matching features. By adopting the method, the accuracy of probability prediction can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and particularly to a method, apparatus, device, and storage medium for predicting interactions of a product. Background Art

[0002] With the development of Internet technology, obtaining information through the Internet has become part of people's lives, entertainment, and work. Especially in the fields of e-commerce and online advertising, merchants and service providers increasingly rely on the Internet to promote products. Before recommending a product to a user, the interaction probability of the user with the product is usually predicted, and based on the prediction result of the interaction probability, the target product is accurately promoted to obtain a better promotion effect.

[0003] Existing interaction prediction solutions usually directly predict product interactions based on user attribute information, resulting in poor accuracy of interaction prediction. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a method, apparatus, device, and storage medium for predicting interactions of a product that can improve the accuracy of interaction prediction.

[0005] In a first aspect, this application provides a method for predicting interactions of a product. The method includes:

[0006] Obtaining an interaction information sequence obtained by a target object's product interaction with a pushed product;

[0007] Determining the correlation between the product information of the to-be-pushed product and each interaction information in the interaction information sequence, and extracting the original interaction pattern related to the to-be-pushed product from the interaction information sequence based on the correlation;

[0008] Converting the original interaction pattern into a standardized interaction pattern;

[0009] Determining the target interaction pattern corresponding to the to-be-pushed product, and performing pattern matching between the standardized interaction pattern and the target interaction pattern to obtain a pattern matching feature;

[0010] Based on the pattern matching feature, determining the probability of the target object's interaction with the to-be-pushed product.

[0011] In a second aspect, this application also provides an apparatus for predicting interactions of a product. The apparatus includes:

[0012] An interaction information sequence acquisition module, configured to obtain an interaction information sequence obtained by a target object's product interaction with a pushed product;

[0013] An original interaction mode determination module, configured to determine the correlation between the product information of the product to be pushed and each piece of interaction information in the interaction information sequence, and extract the original interaction mode related to the product to be pushed from the interaction information sequence based on the correlation;

[0014] A mode conversion module, configured to convert the original interaction mode into a standardized interaction mode;

[0015] A mode matching module, configured to determine the target interaction mode corresponding to the product to be pushed, and perform mode matching between the standardized interaction mode and the target interaction mode to obtain a mode matching feature;

[0016] A probability determination module, configured to determine the probability of interaction between the target object and the product to be pushed based on the mode matching feature.

[0017] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0018] Obtain an interaction information sequence obtained by a target object performing product interaction on a pushed product;

[0019] Determine the correlation between the product information of the product to be pushed and each piece of interaction information in the interaction information sequence, and extract the original interaction mode related to the product to be pushed from the interaction information sequence based on the correlation;

[0020] Convert the original interaction mode into a standardized interaction mode;

[0021] Determine the target interaction mode corresponding to the product to be pushed, and perform mode matching between the standardized interaction mode and the target interaction mode to obtain a mode matching feature;

[0022] Determine the probability of interaction between the target object and the product to be pushed based on the mode matching feature.

[0023] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0024] Obtain an interaction information sequence obtained by a target object performing product interaction on a pushed product;

[0025] Determine the correlation between the product information of the product to be pushed and each piece of interaction information in the interaction information sequence, and extract the original interaction mode related to the product to be pushed from the interaction information sequence based on the correlation;

[0026] Convert the original interaction mode into a standardized interaction mode;

[0027] Determine the target interaction mode corresponding to the product to be pushed, and perform pattern matching between the standardized interaction mode and the target interaction mode to obtain pattern matching features;

[0028] Based on the pattern matching features, determine the probability of interaction between the target object and the product to be pushed.

[0029] In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0030] Obtain an interaction information sequence obtained by a target object performing product interaction on a pushed product;

[0031] Determine the correlation between the product information of the product to be pushed and each interaction information in the interaction information sequence, and extract the original interaction mode related to the product to be pushed from the interaction information sequence based on the correlation;

[0032] Convert the original interaction mode into a standardized interaction mode;

[0033] Determine the target interaction mode corresponding to the product to be pushed, and perform pattern matching between the standardized interaction mode and the target interaction mode to obtain pattern matching features;

[0034] Based on the pattern matching features, determine the probability of interaction between the target object and the product to be pushed.

[0035] The interactive prediction method, device, computer device, storage medium, and computer program product for the above-mentioned product determine the relevance between the product information of the product to be pushed and each piece of interactive information in the interactive information sequence by obtaining the interactive information sequence obtained from the product interaction of the target object with the pushed product, and extract the original interactive patterns related to the product to be pushed from the interactive information sequence based on the relevance. The original interactive patterns capture the interactive patterns of the target object on products similar to the product to be pushed. Therefore, when subsequently determining the probability of the target object interacting with the product to be pushed based on the original interactive patterns, the accuracy of probability prediction can be improved; by converting the original interactive patterns into standardized interactive patterns, it is possible to remove the noise information and unimportant information in the original interactive patterns, avoiding the interference of this information on interactive prediction, thereby further improving the accuracy of probability prediction; by determining the target interactive patterns corresponding to the product to be pushed and performing pattern matching between the standardized interactive patterns and the target interactive patterns to obtain pattern matching features, the pattern matching features reflect the matching situation between the standardized interactive patterns and the target interactive patterns. Furthermore, when determining the probability of the target object interacting with the product to be pushed based on the pattern matching features, the accuracy of probability prediction can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 FIG. is an application environment diagram of the product interactive prediction method in an embodiment;

[0037] Figure 2 FIG. is a flowchart of the product interactive prediction method in an embodiment;

[0038] Figure 3 FIG. is a schematic diagram of the push information display page in an embodiment;

[0039] Figure 4 FIG. is a flowchart of the product interactive prediction method in another embodiment;

[0040] Figure 5 FIG. is a schematic diagram of the structure of the probability prediction model in an embodiment;

[0041] Figure 6 FIG. is a schematic diagram of the process of mining pattern matching features in an embodiment;

[0042] Figure 7 FIG. is a schematic diagram of the structure of the pattern refinement module in an embodiment;

[0043] Figure 8 FIG. is a schematic diagram of the structure of the pattern refinement module in another embodiment;

[0044] Figure 9 FIG. is a schematic diagram of the dataset information in an embodiment;

[0045] Figure 10 Schematic diagram of the comparison results of model performance in an embodiment;

[0046] Figure 11 Schematic diagram of the results of the ablation experiment in an embodiment;

[0047] Figure 12 Schematic diagram of the results of the compatibility analysis experiment in an embodiment;

[0048] Figure 13 Block diagram of the structure of the interaction prediction device of the product in an embodiment;

[0049] Figure 14 Block diagram of the structure of the interaction prediction device of the product in an embodiment;

[0050] Figure 15 Internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0051] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0052] The interaction prediction method of the product provided by the embodiment of the present application relates to technologies such as machine learning in artificial intelligence, where:

[0053] Artificial Intelligence (AI) is to use a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, a theory, method, technology and application system that perceives the environment, acquires knowledge and uses knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

[0054] Artificial intelligence technology is an interdisciplinary subject, involving a wide range of fields, including both hardware-level technologies and software-level technologies. The basic technologies of artificial intelligence generally include, for example, sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, pre-trained model technology, operation / interaction systems, mechatronics, etc. Among them, the pre-trained model, also known as the large model or the basic model, can be widely applied to downstream tasks in various major directions of artificial intelligence after fine-tuning. The artificial intelligence software technology mainly includes several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0055] Machine learning (ML) is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning. Pre-trained models are the latest development results of deep learning, integrating the above technologies.

[0056] The product interaction prediction method provided by the embodiments of this application can be applied to, for example, Figure 1 the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The product interaction prediction method can be executed independently by the terminal 102 or the server 104, or jointly executed by the terminal 102 and the server 104. In some embodiments, the product interaction prediction method is executed by the server 104. The server 104 obtains the interaction information sequence obtained by the target object's product interaction with the pushed product; determines the correlation between the product information of the product to be pushed and each interaction information in the interaction information sequence, and extracts the original interaction pattern related to the product to be pushed from the interaction information sequence based on the correlation; converts the original interaction pattern into a standardized interaction pattern; determines the target interaction pattern corresponding to the product to be pushed, and performs pattern matching between the standardized interaction pattern and the target interaction pattern to obtain pattern matching features; based on the pattern matching features, determines the probability of the target object's interaction with the product to be pushed.

[0057] Among them, the terminal can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices, portable wearable devices, and network devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The network devices can be routers, switches, firewalls, load balancers, network storage devices, network adapters, etc.

[0058] The server 104 can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected through wired or wireless communication methods, and this application does not make any restrictions in this regard.

[0059] In one embodiment, as Figure 2 shown, an interaction prediction method for a product is provided. Taking the method applied to Figure 1 the computer device (terminal or server) in

[0060] S202, obtaining an interaction information sequence obtained by a target object's product interaction with the pushed product.

[0061] The target object is an object to which products can be pushed, and the pushed product is a product that has been pushed to the target object within a historical period, which can specifically be at least one of various types of information and various types of items. Various types of information include at least one of application programs, texts, expressions, pictures, audio, video, files, or links, etc., but are not limited thereto. Various types of items can include physical items and virtual items. Physical items include various physical products, specifically, various electronic products such as mobile phones, computers, notebooks, watches, etc., and can also be clothing products such as clothes and shoes, and no excessive restrictions are made here. Virtual items include, but are not limited to, insurance products, financial products, virtual gift resources, virtual scenes, virtual characters, virtual props, etc. The virtual scene can specifically be a game scene in a game device, a virtual reality simulation scene, etc., the virtual character can specifically be various characters in a game, and the virtual prop can be various props in a game.

[0062] Product interaction refers to generating an interaction behavior with the pushed product by the target object in different product interaction scenarios. The product interaction scenario can be an online shopping scenario, an application browsing scenario, a video playing scenario, a social chat scenario, etc. Interacting with the product means generating an interaction behavior with the product, and the interaction behavior includes behaviors such as clicking, accessing, purchasing, downloading, playing, and collecting. For example, when the product interaction scenario is an online shopping scenario, the product is an item as a commodity, and the target object can purchase the product, thereby generating corresponding behavior information. The generated behavior information includes product recommendation information corresponding to the interacted product, interaction behavior, time, etc. The time can be the time when the interaction behavior occurs and / or the time difference between the time when the interaction behavior occurs and the current time.

[0063] The interaction information sequence is a sequence obtained by arranging the interaction information in the set of interaction information obtained from the product interaction between the target object and the pushed product in chronological order. The interaction information includes product information corresponding to the product being interacted with, interaction behavior, time, etc. The time can be the time when the interaction behavior occurs and / or the time difference between the time when the interaction behavior occurs and the current time.

[0064] It can be understood that, with the authorization of each object, the server can obtain and store the object information of each object and the interaction information obtained from the product interaction of each object with the pushed product, and when it is necessary to process the object information of each object and the interaction information obtained from the product interaction of each object with the pushed product, obtain the stored object information and interaction information. Among them, the collection, use, and processing of the object information of each object and the interaction information of each object need to comply with relevant laws, regulations, and standards of relevant countries and regions.

[0065] Specifically, when the product push condition is met, the computer device can obtain, from the set of interaction information of the target object stored in advance, the interaction information obtained from the product interaction of the target object with the pushed product within a specified historical period, and generate an interaction information sequence according to the time information corresponding to each interaction information.

[0066] Among them, the product push condition includes a time condition and a request condition. The time condition can specifically be a preset product push cycle. For example, if the product push cycle is 24 hours and the push time is 10 am, then when it reaches 10 am every day, the product push condition is met; the request condition can specifically be receiving a product push request, and the computer device can receive product recommendation requests from at least one.

[0067] As Figure 3 shown, the above embodiments are described by taking social circle advertisements as an example. For example, users can use a social application to browse the friend circle in daily life. The friend circle is a function of the social application. Users can share life photos, status updates, etc. in the friend circle. When the user opens the social application through the used terminal and starts to browse the friend circle, the terminal generates an advertisement push request corresponding to the trigger operation of browsing the friend circle, and sends the advertisement push request to the server corresponding to the social application. After receiving the advertisement push request, the server obtains the interaction information sequence obtained from the interaction of the user with the pushed advertisement stored in advance, and determines the target advertisement to be pushed to the user based on the interaction information sequence, and sends the determined target advertisement to the terminal. The terminal can display the target advertisement in the friend circle for the user to interact with the target advertisement.

[0068] S204. Determine the correlation between the product information of the product to be pushed and each piece of interaction information in the interaction information sequence, and extract the original interaction patterns related to the product to be pushed from the interaction information sequence based on the correlation.

[0069] The product to be pushed is a product that can be pushed to the target object, and specifically can be at least one of various types of information and various types of items. Various types of information include at least one of application programs, texts, expressions, pictures, audios, videos, files, or links, etc., but are not limited thereto. Various types of items can include physical items and virtual items. Physical items include various physical products, specifically, various electronic products such as mobile phones, computers, notebooks, watches, etc., and can also be clothing products such as clothes and shoes, and are not overly restricted herein.

[0070] The product information of the product to be pushed refers to the attribute information of the product, and specifically can include information such as product name, product number, product category, product semantics, product image, etc.

[0071] The correlation is used to quantify the degree of association between the product information and the interaction information, and specifically can be characterized by the similarity to represent the magnitude of the correlation.

[0072] The original interaction pattern refers to the initial and unprocessed interaction pattern generated when the target object interacts with the product that has been pushed. This interaction pattern reflects the natural behavior and reaction of the target object, without any screening, cleaning, or conversion, and this interaction pattern is related to the product to be pushed. The number of original interaction patterns can be at least one.

[0073] Specifically, after the computer device obtains the interaction information sequence of the target object, it can select any product to be pushed from the product pool to be pushed, obtain the product information of this product to be pushed, and use a preset correlation determination algorithm to determine the correlation between the product information of the product to be pushed and each piece of interaction information in the interaction information sequence, and based on the correlation, determine the target interaction information related to the target product from the interaction information sequence, and extract the original interaction patterns related to the product to be pushed from the interaction information sequence according to the target interaction information.

[0074] Among them, the target interaction information may include the click and browsing patterns of the user on similar products to the product to be pushed, reflecting the interest of the target object in similar products.

[0075] In one embodiment, the computer device determining the correlation between the product information of the product to be pushed and each interaction information in the interaction information sequence further includes the following steps: determining the target interaction mode of the product to be pushed according to the product information of the product to be pushed, determining the correlation between the target interaction mode and each interaction information in the interaction information sequence, and determining the correlation between the target interaction mode and each interaction information as the correlation between the product information and each interaction information in the interaction information sequence.

[0076] Among them, the target interaction mode may be an ideal or expected interaction mode for the product to be pushed. For example, for the promotion of a new mobile phone, the target interaction mode may include interaction information such as users browsing product details, clicking on the purchase link, and reading reviews.

[0077] S206, converting the original interaction mode into a standardized interaction mode.

[0078] Among them, the standardized interaction mode refers to a normalized interaction mode. The standardized interaction mode helps to more accurately analyze the behavior of the target object and understand the preferences of the target object.

[0079] Specifically, after the computer device obtains the original interaction mode related to the product to be pushed of the target object, it can use a preset standardization algorithm to process the data of the original interaction mode, thereby obtaining the standardized interaction mode.

[0080] Among them, the standardization algorithm can specifically implement at least one data processing such as data cleaning, denoising, scaling, and normalization.

[0081] S208, determining the target interaction mode corresponding to the product to be pushed, and performing pattern matching between the standardized interaction mode and the target interaction mode to obtain pattern matching features.

[0082] Among them, the target interaction mode may be an ideal or expected interaction mode for the product to be pushed. For example, for the promotion of a new mobile phone, the target interaction mode may include interaction information such as users browsing product details, clicking on the purchase link, and reading reviews.

[0083] Pattern matching refers to fusing the standardized interaction mode and the target interaction mode to enhance the correlation between the two. The obtained pattern matching features are the features with enhanced correlation.

[0084] Specifically, the computer device can obtain the expected interaction mode corresponding to the product to be pushed, determine the target interaction mode of the product to be pushed according to the obtained expected interaction mode, and use a preset pattern matching algorithm to perform pattern matching between the standardized protection mode and the target interaction mode to obtain pattern matching features.

[0085] S210. Determine the probability of interaction between the target object and the product to be pushed based on the pattern matching features.

[0086] Among them, the interaction can refer to generating an interaction behavior, and the interaction behavior can specifically be behaviors such as clicking, browsing, purchasing, downloading, registering, subscribing, sharing, commenting, liking, and rating.

[0087] Specifically, after obtaining the pattern matching features, the computer device can input the pattern matching features into a pre-trained prediction network, and the prediction network performs prediction processing based on the pattern matching features to obtain the probability of interaction between the target object and the product to be pushed.

[0088] Among them, the prediction network can specifically be a deep neural network (DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), etc.

[0089] In the above product interaction prediction method, by obtaining the interaction information sequence obtained by the target object's product interaction with the pushed product, determining the correlation between the product information of the product to be pushed and each interaction information in the interaction information sequence, and extracting the original interaction pattern related to the product to be pushed from the interaction information sequence based on the correlation, the original interaction pattern captures the interaction pattern of the target object on products similar to the product to be pushed. Therefore, when subsequently determining the probability of interaction between the target object and the product to be pushed based on the original interaction pattern, the accuracy of probability prediction can be improved; by converting the original interaction pattern into a standardized interaction pattern, it is possible to remove the noise information and unimportant information in the original interaction pattern and avoid the interference of these information on interaction prediction, thereby further improving the accuracy of probability prediction; by determining the target interaction pattern corresponding to the product to be pushed and performing pattern matching between the standardized interaction pattern and the target interaction pattern to obtain the pattern matching features, the pattern matching features reflect the matching situation between the standardized interaction pattern and the target interaction pattern. Furthermore, when determining the probability of interaction between the target object and the product to be pushed based on the pattern matching features, the accuracy of probability prediction can be further improved.

[0090] In one embodiment, the above product interaction prediction method further includes the following steps: performing vectorization processing on the product information and each interaction information in the interaction information sequence to obtain a product feature vector and each interaction feature vector; performing attention processing on the product feature vector and each interaction feature vector to obtain an attention processing result; the attention processing result includes the attention weights of each interaction feature vector relative to the product feature vector; the computer device determines the correlation between the product information of the product to be pushed and each interaction information in the interaction information sequence including: determining the correlation between the product information and each interaction information based on the attention weights of each interaction feature vector relative to the product feature vector.

[0091] Among them, vectorization processing refers to converting the original data into a numerical vector so that the data can be effectively processed by subsequent algorithms.

[0092] Attention processing is used to enhance the model's ability to focus on important parts of the data. It originates from the human visual attention mechanism, that is, the tendency to concentrate attention on certain key parts of the observed scene rather than the entire scene.

[0093] Specifically, the computer device can use a preset vectorization processing algorithm to perform vectorization processing on each interaction information in the product information and the interaction information sequence respectively, obtain the product feature vector corresponding to the product information and the interaction feature vectors corresponding to each interaction information, and input the product feature vector and each interaction feature vector into a pre-trained attention network. Through the attention network, perform attention processing on the product feature vector and each interaction feature vector to obtain the attention weights of each interaction feature vector relative to the product feature vector. The attention weights reflect the importance of each interaction feature vector relative to the product feature vector, and determine the attention weights as the correlation between each interaction feature vector and the product feature vector. This correlation is the correlation between the product information and each interaction information respectively.

[0094] Among them, the preset vectorization processing algorithm can specifically be the Bag of Words model, Word Embeddings model, etc. The Word Embeddings model can specifically be Word2Vec, GloVe, etc. The attention network can specifically adopt self-attention mechanism, cross-attention mechanism or multi-head attention mechanism.

[0095] In the above embodiment, the computer device performs attention processing on the product feature vector and each of the interaction feature vectors obtained through vectorization processing, so as to effectively identify the correlation between the product information and the user interaction information, so as to accurately capture the original interaction pattern based on the correlation subsequently. Furthermore, when determining the probability of the target object interacting with the product to be pushed based on the original interaction pattern subsequently, the accuracy of probability prediction can be improved.

[0096] In one embodiment, the attention processing result further includes a fused interaction feature vector corresponding to the interaction feature vector; the process of the computer device determining the probability of the target object interacting with the product to be pushed based on the pattern matching feature further includes the following steps: determining the probability of the target object interacting with the product to be pushed based on the pattern matching feature, the fused interaction feature vector and the product feature vector.

[0097] Specifically, the computer device can use a preset vectorization processing algorithm to separately vectorize each piece of interaction information in the product information and the interaction information sequence, obtain a product feature vector corresponding to the product information and interaction feature vectors corresponding to each piece of interaction information, and input the product feature vector and each interaction feature vector into a pre-trained attention network. Through the attention network, perform attention processing on the product feature vector and each interaction feature vector to obtain the attention weights of each interaction feature vector relative to the product feature vector, and perform weighted processing on the interaction feature vectors according to the attention weights to obtain a weighted result. Then, fuse the weighted result and the product feature vector to obtain a fused interaction feature vector, and input the fused feature vector, the pattern matching feature, and the product feature vector into a pre-trained prediction network. Through the prediction network, perform prediction processing based on the fused feature vector, the pattern matching feature, and the product feature vector to obtain the probability of the target object interacting with the product to be pushed.

[0098] In one embodiment, the computer device can also obtain the object information of the target object, vectorize the object information to obtain an object feature vector, and input the object feature vector, the fused feature vector, the pattern matching feature, and the product feature vector into a pre-trained prediction network. Through the prediction network, perform prediction processing based on the object feature vector, the fused feature vector, the pattern matching feature, and the product feature vector to obtain the probability of the target object interacting with the product to be pushed.

[0099] Among them, the object information is the attribute information of the target object, such as the gender, age, height, weight, education level, occupation, region, etc. of the target object.

[0100] In the above embodiment, the computer device determines the probability of the target object interacting with the product to be pushed based on the pattern matching feature, the fused interaction feature vector, and the product feature vector, so that the probability of the target object interacting with the product to be pushed can be determined more accurately, further improving the accuracy of probability prediction.

[0101] In one embodiment, the process by which the computer device extracts the original interaction pattern related to the product to be pushed from the interaction information sequence based on relevance includes the following steps: Among the pieces of interaction information in the interaction information sequence, select the target interaction information whose relevance satisfies the relevance condition; intercept the first interaction information sequence containing the target interaction information from the interaction information sequence; and determine the first interaction information sequence as the original interaction pattern related to the product to be pushed.

[0102] Among them, the relevance condition is a standard or rule for screening interaction information related to the target product, which can specifically be a threshold condition or a ranking condition. For example, when the relevance condition is a threshold condition, among the interaction information in the interaction information sequence, the interaction information with a relevance greater than or equal to the relevance threshold can be determined as the target interaction information that meets the relevance condition; when the relevance condition is a ranking condition, the interaction information in the interaction information sequence can be sorted according to the magnitude of the relevance of each interaction information, and the top K interaction information with the greatest relevance can be determined as the target interaction information that meets the relevance condition.

[0103] The first interaction information sequence is a subsequence of the interaction information sequence, and the number of the first interaction information sequences is the same as the number of target interaction information.

[0104] Specifically, the computer device selects, from the interaction information in the interaction information sequence, the target interaction information whose relevance meets the relevance condition, and uses the position of the target interaction information in the interaction information sequence as the reference position. According to a preset truncation scheme, a first interaction information sequence containing the target interaction information is truncated from the interaction information sequence, and the first interaction information sequence is directly determined as the original interaction pattern related to the product to be pushed.

[0105] The preset truncation scheme includes the truncation length, the truncation direction, etc. For example, taking the reference position as the benchmark, M interaction information are truncated forward and N interaction information are truncated backward in the interaction information sequence to obtain a first interaction information sequence with a length of L, where L = M + 1 + N.

[0106] In the above embodiment, the computer device selects, from the interaction information in the interaction information sequence, the target interaction information whose relevance meets the relevance condition, truncates a first interaction information sequence containing the target interaction information from the interaction information sequence, and determines the first interaction information sequence as the original interaction pattern related to the product to be pushed. Thus, the original interaction pattern related to the product to be pushed can be accurately extracted from the interaction information sequence based on the relevance, and when subsequently determining the probability of the target object interacting with the product to be pushed based on the original interaction pattern, the accuracy of probability prediction can be improved.

[0107] In one embodiment, the process of the computer device converting the original interaction pattern into a standardized interaction pattern includes the following steps: performing vectorization processing on the first interaction information corresponding to the original interaction pattern to obtain each first interaction feature vector; the first interaction information is the interaction information in the original interaction pattern; performing attention processing on each first interaction feature vector respectively to obtain each enhanced first interaction feature vector; performing mapping processing on each enhanced first interaction feature vector to obtain the first standardized probability corresponding to each first interaction information; and denoising the original interaction pattern based on the first standardized probability corresponding to each first interaction information to obtain the standardized interaction pattern.

[0108] Among them, the first interaction information corresponding to the original interaction mode refers to the interaction information in the original interaction mode, that is, the interaction information in the first interaction information sequence. Vectorization processing refers to converting the original data into a numerical vector so that the data can be effectively processed by subsequent algorithms.

[0109] Specifically, the computer device can perform vectorization processing on the first interaction information corresponding to the original interaction mode by using a preset vectorization processing algorithm to obtain each first interaction feature vector, and input each first interaction feature vector into a pre-trained neural network. Through the attention sub-network of the neural network, attention processing is performed on each first interaction feature vector to obtain an enhanced first interaction feature vector corresponding to each first interaction feature vector, and each enhanced first interaction feature vector is input into the multi-layer perceptron sub-network of the neural network. Through the multi-layer perceptron sub-network, mapping processing is performed on each enhanced first interaction feature vector to obtain a first normalized probability corresponding to each first interaction information, and based on the first normalized probability, noise information in the original interaction mode is determined, and the noise information is removed from the original interaction mode to obtain a normalized interaction mode.

[0110] Among them, the attention sub-network can specifically adopt an encoder (Transformer) structure. The Transformer encoder usually consists of multiple identical encoding layers, and each encoding layer contains a self-attention mechanism and a feed-forward neural network. In the embodiments of the present application, the Transformer encoder can specifically adopt a double-layer structure.

[0111] The multi-layer perceptron (MLP, Multi-Layer Perceptron) sub-network is used to extract deeper patterns and relationships from the input data, including at least one network layer, and each network layer contains a group of neurons, and these neurons can learn the non-linear representation of the input data.

[0112] In the above embodiments, the computer device performs vectorization processing on the first interaction information corresponding to the original interaction mode to obtain each first interaction feature vector, and the first interaction information is the interaction information in the original interaction mode; attention processing is respectively performed on each first interaction feature vector to obtain each enhanced first interaction feature vector, mapping processing is performed on each enhanced first interaction feature vector to obtain a first normalized probability corresponding to each first interaction information, and the original interaction mode is denoised based on the first normalized probability corresponding to each first interaction information to obtain a normalized interaction mode, which can accurately remove the noise information in the original interaction mode and avoid the interference of this information on the interaction prediction, thereby further improving the accuracy of probability prediction.

[0113] In one embodiment, the enhanced first interaction feature vector and the first normalized probability are obtained by processing through a neural network; the interaction prediction method of the above product further includes the process of training the neural network, and this process specifically includes the following steps: adding noise to the sample interaction information sequence to obtain a second interaction information sequence; vectorizing each second interaction information in the second interaction information sequence to obtain each second interaction feature vector; performing attention processing on the second interaction feature vectors through the neural network to be trained to obtain each enhanced second interaction feature vector, and performing mapping processing on each second interaction feature vector to obtain the second normalized probability corresponding to each second interaction information; optimizing the neural network to be trained based on the second normalized probability corresponding to each second interaction information to obtain the neural network.

[0114] Among them, the sample interaction information sequence is the training data for training the neural network, and can also be called the sample interaction pattern. Specifically, it can be the sample normalized interaction pattern, and the sample normalized interaction pattern can be obtained after artificially normalizing the original interaction pattern.

[0115] Adding noise means adding noise information to the sample interaction information sequence to achieve data augmentation for training the neural network with the augmented data. For example, if the sample interaction information sequence is , where S represents the number of sample interaction information in the sample interaction information sequence P, then L - S interaction information can be randomly mixed into this sample interaction information sequence to obtain the second interaction information sequence , and the second interaction information sequence is also the enhanced interaction pattern obtained after data augmentation of the sample interaction pattern.

[0116] Specifically, after the computer device adds noise to the sample interaction information sequence to obtain the second interaction information sequence, a preset vectorization processing algorithm can be used to vectorize each second interaction information in the second interaction information sequence to obtain each second interaction feature vector, and input each second interaction feature vector into the neural network to be trained. Through the attention sub-network of the neural network to be trained, attention processing is performed on each second interaction feature vector to obtain the enhanced second interaction feature vector corresponding to each second interaction feature vector, and input each enhanced second interaction feature vector into the multi-layer perceptron sub-network of the neural network to be trained. Through the multi-layer perceptron sub-network, mapping processing is performed on each enhanced second interaction feature vector to obtain the second normalized probability corresponding to each second interaction information, and the training loss value is determined based on the second normalized probability. The neural network to be trained is optimized based on the training loss value to obtain the neural network.

[0117] Among them, the attention sub-network can specifically adopt an encoder (Transformer) structure. The Transformer encoder usually consists of multiple identical encoding layers, and each encoding layer contains a self-attention mechanism and a feed-forward neural network. In the embodiments of the present application, the Transformer encoder can specifically adopt a double-layer structure.

[0118] The multi-layer perceptron (MLP, Multi-Layer Perceptron) sub-network is used to extract deeper patterns and relationships from the input data, including at least one network layer, and each network layer contains a group of neurons, and these neurons can learn the non-linear representation of the input data.

[0119] In one embodiment, the attention sub-network can specifically adopt an encoder (Transformer) structure. The attention sub-network includes M encoding layers, and the data processing process of the attention sub-network can be represented by the following formula:

[0120]

[0121]

[0122]

[0123] Among them, represents the input data of the first encoding layer of the attention sub-network, represents the vector representation of the i-th second interaction information in the second interaction information sequence, that is, the i-th second interaction feature vector, represents the position encoding of the i-th second interaction feature vector; represents the output of the attention mechanism of the m-th encoding layer, represents self-attention processing of the output of the (m - 1)-th encoding layer through the attention network of the m-th encoding layer, represents adding the result of the self-attention processing to the original input. This process can also be called a residual connection. LayerNorm represents performing layer normalization on the result to ensure network stability; represents the output of the feed-forward neural network of the m-th encoding layer, that is, the output of the m-th encoding layer, represents processing the output of the attention mechanism of the m-th encoding layer through the feed-forward neural network of the m-th encoding layer, represents adding the processing result of the feed-forward neural network to the original input. This process can also be called a residual connection. LayerNorm represents performing layer normalization on the result.

[0124] In one embodiment, the output of the multi-layer perceptron can be represented as a class score vector (logits vector) o of length L, and the processing process of the multi-layer perceptron can be represented as follows:

[0125]

[0126] where o is the class score vector (logits vector) o, and MLP represents the processing of the multi-layer perceptron. represents the output result of the attention sub-network, that is, the output result of the M-th encoding layer of the attention sub-network.

[0127] In the above embodiment, the computer device obtains a second interaction information sequence by adding noise to the sample interaction information sequence; performs vectorization processing on each second interaction information in the second interaction information sequence to obtain each second interaction feature vector; performs attention processing on the second interaction feature vectors through the neural network to be trained to obtain each enhanced second interaction feature vector, and performs mapping processing on each second interaction feature vector to obtain the second normalized probability corresponding to each second interaction information; optimizes the neural network to be trained based on the second normalized probability corresponding to each second interaction information, so that a neural network capable of accurately predicting the normalized probability corresponding to the interaction information can be obtained. Subsequently, when predicting the first normalized probability corresponding to each first interaction information based on the neural network, the accuracy of the normalized probability prediction can be improved.

[0128] In one embodiment, the process of training the neural network further includes the following steps: determining a loss coefficient corresponding to each second interaction information based on the matching relationship between the interaction information in the sample interaction information sequence and each second interaction information in the second interaction information sequence; the process by which the computer device optimizes the neural network to be trained based on the second normalized probability corresponding to each second interaction information to obtain the neural network includes the following steps: determining a training loss value based on the loss coefficient corresponding to each second interaction information and the second normalized probability corresponding to each second interaction information; adjusting the parameters of the neural network to be trained based on the training loss value until the training stop condition is reached to obtain the neural network.

[0129] Among them, the training loss value is used to measure the performance of the model on the training data to guide the optimization process of the model parameters. The training stop condition is the condition used to determine when to end the model training, which can prevent overfitting and ensure that the model reaches sufficient performance. Specifically, it can be a loss threshold condition, a performance threshold condition, an iteration number condition, or a custom condition. The loss threshold condition can specifically be setting a specific loss function threshold. When the loss value of the model drops below this threshold, it is considered that the model has been sufficiently optimized and the training can be stopped. The performance threshold condition can specifically be based on the performance of the model on the validation set, such as accuracy, recall, etc. If the model performance no longer improves significantly or reaches a predetermined performance level, the training is stopped. The iteration number condition can specifically be setting the maximum number of iterations for training. Regardless of the model performance, the training is stopped after reaching this iteration number. The custom condition can specifically be customizing a set of rules for stopping training according to specific tasks or requirements. This set of rules can be a combination of the above conditions or conditions based on other specific metrics.

[0130] Specifically, for each second interaction information in the second information sequence, it is determined whether it belongs to the noise information outside each interaction information in the sample interaction information sequence. If so, the loss coefficient of the second interaction information is determined to be the first coefficient value. If not, the loss coefficient of the second interaction information is determined to be the second coefficient value. For example, the first coefficient value is 0 and the second coefficient value is 1. After obtaining the second prediction probability values corresponding to each second interaction information, numerical processing is performed on the second prediction probability value to obtain a numerical processing result, and weighted summation is performed according to the numerical processing results corresponding to each second interaction information based on the loss coefficient to obtain the training loss value. The backpropagation algorithm is used to adjust the parameters of the neural network to be trained based on the training loss value. Through multiple iterations, the training loss value is gradually reduced until the training stop condition is reached, and a trained neural network is obtained.

[0131] In one embodiment, the following relationship is satisfied among the training loss value, the loss coefficient, and the second prediction probability:

[0132]

[0133] Among them, represents the training loss value, represents the loss coefficient corresponding to the j-th second interaction information. When takes the value of 1, When takes the value of 0, represents the second prediction probability corresponding to the j-th second interaction information, represents the second information sequence.

[0134] In the above embodiments, the computer device determines the loss coefficient corresponding to each second interaction information based on the matching relationship between the interaction information in the sample interaction information sequence and each second interaction information in the second interaction information sequence, determines the training loss value based on the loss coefficient corresponding to each second interaction information and the second normalization probability corresponding to each second interaction information. The training loss value reflects the gap between the model prediction and the actual situation. The parameters of the neural network to be trained are adjusted based on the training loss value until the training stop condition is reached, and a neural network is obtained. The training process ensures the sufficiency and stability of the model training, so that a neural network that can accurately predict the normalization probability corresponding to the interaction information can be obtained. Furthermore, when predicting the first normalization probability corresponding to each first interaction information based on the neural network later, the accuracy of the normalization probability prediction can be improved.

[0135] In one embodiment, the process by which the computer device denoises the original interaction pattern based on the first normalization probability corresponding to each first interaction information to obtain the normalized interaction pattern includes the following steps: In the original interaction pattern, the target first interaction information whose first normalization probability does not meet the probability condition is determined as noise information; the noise information is deleted from the original interaction pattern to obtain the normalized interaction pattern.

[0136] Among them, the probability condition can be a threshold condition or a sorting condition. For example, when the probability condition is a threshold condition, the first interaction information in which the first normalization probability is less than the probability threshold among each first interaction information can be determined as the target first interaction information; when the probability condition is a sorting condition, each first interaction information can be sorted according to the magnitude of the first normalization probability, and the largest S first interaction information with the first normalization probability can be determined as the target first interaction information.

[0137] For example, for the first interaction information sequence with a length of L, the largest L - S target first interaction information with the first normalization probability can be determined, and the L - S target first interaction information is deleted from the original interaction pattern to obtain a normalized interaction pattern with a length of S.

[0138] In the above embodiments, the computer device can accurately remove the noise information in the original interaction pattern by determining the target first interaction information whose first normalization probability does not meet the probability condition as noise information in the original interaction pattern and deleting the noise information from the original interaction pattern to obtain the normalized interaction pattern, avoiding the interference of this information on the interaction prediction, thereby further improving the accuracy of the probability prediction.

[0139] In one embodiment, the process by which the computer device determines the target interaction mode corresponding to the product to be pushed includes the following steps: extracting a third interaction information sequence from the interaction information sequence; determining the expected interaction information sequence of the product to be pushed according to the product information; splicing the third interaction information sequence and the expected interaction information sequence to obtain the target interaction mode.

[0140] Among them, the third interaction information sequence includes at least one interaction information, and the at least one interaction information is the at least one interaction information closest to the current time in the interaction information sequence of the target object. For example, if the interaction information sequence of the target object includes 10 interaction information, 4 interaction information closest to the current time can be selected, and the 4 interaction information form the third interaction information sequence.

[0141] The expected interaction information sequence refers to the interaction mode between the target object and the product to be pushed under ideal circumstances. For example, in an e-commerce platform, it includes interaction information such as searching for products, viewing product details, adding to the shopping cart, and making a purchase.

[0142] Specifically, the computer device can extract the third interaction information sequence from the interaction information sequence according to a preset selection method, and according to the product information of the product to be recommended, the key features of the product to be pushed, such as functions, uses, market positioning, etc., generate the expected interaction information sequence according to the key features, and connect the end of the third interaction information sequence and the expected interaction information sequence to obtain the target interaction mode.

[0143] In the above embodiment, the computer device extracts the third interaction information sequence from the interaction information sequence; determines the expected interaction information sequence of the product to be pushed according to the product information; splices the third interaction information sequence and the expected interaction information sequence to obtain the target interaction mode, so that the standardized interaction mode can be further enhanced based on the target interaction mode in the subsequent process to obtain the pattern matching feature. The pattern matching feature reflects the matching situation between the standardized interaction mode and the target interaction mode. Furthermore, when determining the probability of the interaction between the target object and the product to be pushed based on the pattern matching feature, the accuracy of probability prediction can be further improved.

[0144] In one embodiment, the process by which the computer device performs pattern matching between the standardized interaction mode and the target interaction mode to obtain the pattern matching feature includes the following steps: performing corresponding fusion on the interaction information in the standardized interaction mode and the interaction information in the target interaction mode to obtain a fusion result; performing pooling processing on the fusion result to obtain the pattern matching feature.

[0145] Specifically, after obtaining the standardized interaction pattern and the target interaction pattern, the computer device can perform Hadamard multiplication on the standardized interaction pattern and the target interaction pattern to obtain the Hadamard product between the standardized interaction pattern and the target interaction pattern. This Hadamard product is the fusion result, and the computer device can perform sum pooling on the fusion result to obtain the pattern matching feature.

[0146] In one embodiment, the following relationship holds among the pattern matching feature, the standardized interaction pattern, and the target interaction pattern:

[0147]

[0148] Wherein, represents the pattern matching feature, represents the standardized interaction pattern, represents the target interaction pattern, and ⨂ represents Hadamard multiplication.

[0149] In the above embodiment, the computer device corresponds and fuses the interaction information in the standardized interaction pattern with the interaction information in the target interaction pattern to obtain the fusion result. By performing pooling on the fusion result, the computer device can further enhance the alignment interaction pattern to obtain the pattern matching feature. The pattern matching feature reflects the matching between the standardized interaction pattern and the target interaction pattern. Therefore, when determining the probability of interaction between the target object and the product to be pushed based on the pattern matching feature, the accuracy of probability prediction can be further improved.

[0150] In one embodiment, the probability of interaction between the target object and the product to be pushed is the first probability. After determining the probability of interaction between the target object and the product to be pushed based on the pattern matching feature, the computer device can further: obtain the second probability of interaction between the target object and other products to be pushed; sort the product to be pushed and other products to be pushed based on the first probability and the second probability to obtain a sorting result; select the target product to be pushed that meets the push condition from the product to be pushed and other products to be pushed according to the sorting result; and push the target product to be pushed to the target object.

[0151] Wherein, the second probability is the probability of interaction between the target object and other products to be pushed. Specifically, the push condition can be a ranking condition. For example, the product ranked first among the products to be pushed and other products to be pushed is determined as the target product to be pushed.

[0152] Specifically, the computer device can also obtain the probability of interaction between the target object and other products to be pushed through the execution of steps S202 to S210. This probability is the second probability of interaction between the target object and other products to be pushed. Then, the products to be pushed and other products to be pushed are sorted according to the probability values of the first probability and the second probability to obtain a sorting result. The target product to be pushed that meets the push condition is selected from the products to be pushed and other products to be pushed, and the product information of the target product to be pushed is obtained. The product information is sent to the terminal used by the target object, so that the terminal can display the product information of the target product to be pushed received, so that the target object can interact with the target product to be pushed.

[0153] In one embodiment, after determining the target product to be pushed, the computer device can also obtain the product display configuration information of the terminal corresponding to the target object, and adjust the product information of the target product to be pushed according to the product display configuration information to obtain the adjusted product information of the target product to be pushed, and send the adjusted product information of the target product to the terminal corresponding to the target object, so that the terminal corresponding to the target object can display the target product based on the received adjusted product information.

[0154] For example, the terminal product display configuration information includes the size of the product picture display window. After the server obtains the size of the product picture display window of the terminal, it can adjust the size of the product image of the target product according to this size to adapt to the size of the product picture display window of the terminal, and send the product image of the target product with the adjusted size to the terminal corresponding to the target object, so that the terminal corresponding to the target object can display the received product image of the target product.

[0155] In the above embodiment, the computer device obtains the second probability of interaction between the target object and other products to be pushed; sorts the products to be pushed and other products to be pushed based on the first probability and the second probability to obtain a sorting result; selects the target product to be pushed that meets the push condition from the products to be pushed and other products to be pushed according to the sorting result; and pushes the target product to be pushed to the target object, thereby improving the accuracy of product recommendation.

[0156] In one embodiment, as Figure 4 shown, there is also provided an interaction prediction method for products. Taking the computer device (terminal or server) in Figure 1 as an example, the method includes the following steps:

[0157] S402. Obtain the interaction information sequence obtained by the target object's interaction with the pushed product; perform vectorization processing on the product information and each interaction information in the interaction information sequence respectively to obtain a product feature vector and each interaction feature vector; perform attention processing on the product feature vector and each interaction feature vector to obtain an attention processing result; the attention processing result includes the attention weight of each interaction feature vector relative to the product feature vector and the fused interaction feature vector corresponding to the interaction feature vector.

[0158] S404. Based on the attention weight of each interaction feature vector relative to the product feature vector, determine the correlation between the product information and each interaction information respectively; among the interaction information in the interaction information sequence, select the target interaction information whose correlation satisfies the correlation condition; intercept the first interaction information sequence containing the target interaction information from the interaction information sequence; determine the first interaction information sequence as the original interaction pattern related to the product to be pushed.

[0159] S406. Perform vectorization processing on the first interaction information corresponding to the original interaction pattern to obtain each first interaction feature vector; the first interaction information is the interaction information in the original interaction pattern.

[0160] S408. Perform attention processing on each first interaction feature vector respectively through a pre-trained neural network to obtain each enhanced first interaction feature vector; perform mapping processing on each enhanced first interaction feature vector to obtain the first normalized probability corresponding to each first interaction information.

[0161] S410. In the original interaction pattern, determine the target first interaction information whose first normalized probability does not satisfy the probability condition as noise information; delete the noise information from the original interaction pattern to obtain a normalized interaction pattern.

[0162] S412. Extract a third interaction information sequence from the interaction information sequence; determine the expected interaction information sequence of the product to be pushed according to the product information; splice the third interaction information sequence and the expected interaction information sequence to obtain a target interaction pattern.

[0163] S414. Perform corresponding fusion on the interaction information in the normalized interaction pattern and the interaction information in the target interaction pattern to obtain a fusion result; perform pooling processing on the fusion result to obtain a pattern matching feature.

[0164] S416. Based on the pattern matching feature, the fused interaction feature vector and the product feature vector, determine the probability of the target object's interaction with the product to be pushed.

[0165] This application also provides an application scenario, which implements the above product interaction prediction method through a probability prediction model. Refer to Figure 5Schematic diagram of the structure of the probability prediction model shown. The probability prediction model includes a Deep Pattern Network (DPN) and a basic network. The Deep Pattern Network (DPN) includes a Target-aware Pattern Retrieval Module (TPRM), a Self-supervised Pattern Refinement Module (SPRM), and a Pattern Matching Module (PMM). The basic network includes a word embedding layer, a target attention network, a feature fusion network, an activation function layer (PReLU), and a probability layer (softmax). The interactive prediction method of this product can specifically include the following steps:

[0166] Step Y1, data acquisition.

[0167] Obtain the product information of the product to be pushed, the interaction information sequence obtained by the target object's interaction with the pushed products, and other relevant information. Among them, the other relevant information can specifically be the object information of the target object.

[0168] Step Y2, vectorization processing.

[0169] Perform vectorization on the product information of the product to be pushed, the interaction information sequence obtained by the target object's interaction with the pushed products, and other relevant information respectively to obtain a product feature vector, each interaction feature vector, and other relevant feature vectors.

[0170] Step Y3, target attention processing.

[0171] Perform attention processing on the product feature vector and each interaction feature vector through the target attention network to obtain an attention processing result. The attention processing result includes the attention weights (which can also be called attention scores) of each interaction feature vector relative to the product feature vector and the fused interaction feature vectors corresponding to the interaction feature vectors.

[0172] Step Y4, pattern retrieval.

[0173] Through the target-aware pattern retrieval module, according to the attention scores of each interaction feature vector relative to the product feature vector, retrieve the target interaction information corresponding to the top-K highest attention scores in the interaction information sequence, and intercept K first interaction information sequences of length L with the target interaction information as the end point from the interaction information sequence. The first interaction information sequence is the original interaction pattern retrieved and related to the product to be pushed.

[0174] Step Y5, Pattern Refinement.

[0175] The original interaction pattern is refined by a self-supervised pattern refinement module to obtain a standardized pattern related to the product to be pushed.

[0176] The following uses an actual example to illustrate the mining process of pattern matching features in the above scenario. Refer to Figure 6 the schematic diagram of the pattern matching feature mining process shown. The vector representation of the expected interaction information corresponding to the target product is "1591862", and the vector representations of the two most recent interaction information of the target object are "3371523" and "2367945". Based on this, the target interaction pattern of the target product is determined as "3371523, 2367945, 1591862". The Top-1 original interaction pattern related to the target interaction pattern, "4196768, 3371523, 2367945, 4196768, 1591862", is retrieved from the interaction information sequence of the target object through TPRM. The original interaction pattern is further refined by the SPRM module to obtain the standardized interaction pattern "3371523, 2367945, 1591862". Finally, the standardized interaction pattern "3371523, 2367945, 1591862" is pattern-matched with the target interaction pattern "3371523, 2367945, 1591862" through the PMM module to obtain the pattern matching feature.

[0177] Step Y6, Pattern Matching.

[0178] The standardized interaction pattern and the target interaction pattern are fused by the pattern matching module using feature crossing to obtain a fusion result, and the fusion result is processed by sum pooling to obtain the pattern matching feature.

[0179] Among them, the target interaction pattern is generated based on the most recent interaction information of the target object and the expected interaction information of the product to be pushed.

[0180] Step 7, Probability Prediction.

[0181] The pattern matching feature, product feature vector, fused interaction feature vector, and other relevant feature vectors are fused by a feature fusion network to obtain a fused feature; the fused feature is processed by an activation function layer and a probability layer in sequence to obtain the probability of the interaction between the target object and the product to be pushed.

[0182] Refer to Figure 7Schematic structural diagram of the pattern refinement module shown. The pattern refinement module includes an attention sub-network and a multi-layer perceptron network. The attention sub-network adopts an encoder (Transformer) structure. The Transformer encoder includes 2 encoding layers, and each encoding layer contains a self-attention mechanism and a feed-forward neural network. The process of the pattern refinement module for refining the original interaction pattern is as follows: After obtaining the first interaction feature vectors of each first interaction information in the first interaction information sequence, each first interaction feature vector is respectively subjected to attention processing through the attention sub-network to obtain each enhanced first interaction feature vector; each enhanced first interaction feature vector is subjected to mapping processing through a multi-layer perceptron to obtain the first normalized probability corresponding to each first interaction information. In the original interaction pattern, the target first interaction information whose first normalized probability does not meet the probability condition is determined as noise information; the noise information is deleted from the original interaction pattern to obtain a normalized interaction pattern.

[0183] In addition, the pattern refinement module is obtained by pre-training with self-supervision. Refer to Figure 8 Schematic structural diagram of the pattern refinement module based on self-supervision shown. The training process includes:

[0184] Step X1, data acquisition.

[0185] Obtain a sample interaction information sequence, and add noise to the sample interaction information sequence to obtain a second interaction information sequence. For example, the sample interaction information sequence is , S represents the number of sample interaction information in the sample interaction information sequence P. Then L - S interaction information can be randomly mixed into the sample interaction information sequence to obtain the second interaction information sequence . The second interaction information sequence is also the enhanced interaction pattern obtained after data augmentation of the sample interaction pattern.

[0186] In addition, based on the matching relationship between the interaction information in the sample interaction information sequence and each second interaction information in the second interaction information sequence, the loss coefficient corresponding to each second interaction information can be determined.

[0187] Step X2, vectorization processing.

[0188] Perform vectorization processing on each second interaction information in the second interaction information sequence to obtain each second interaction feature vector.

[0189] Step X3, refinement training.

[0190] Performing attention processing on the second interaction feature vector through the to-be-trained refinement network to obtain enhanced second interaction feature vectors, and performing mapping processing on each second interaction feature vector to obtain second normalized probabilities corresponding to each second interaction information; determining a training loss value based on the loss coefficients corresponding to each second interaction information and the second normalized probabilities corresponding to each second interaction information; adjusting the parameters of the to-be-trained refinement network based on the training loss value until a training stop condition is reached, thereby obtaining a refinement network.

[0191] In addition, the performance of the probability prediction model provided by the present application was also verified on three different data sets. The specific situations of the three data sets are as Figure 9 shown. Figure 10 The performance of different probability prediction models on these three data sets is shown. The performance metric used is AUC. AUC is the abbreviation of "Area Under the Curve", which is a commonly used metric for evaluating the performance of classification models. Specifically, AUC measures the trade-off between the true positive rate (also known as sensitivity or recall) and the false positive rate of a binary classification model (usually a classifier) at different thresholds. It can be seen from the figure that the probability prediction model provided by the present application that includes DPN has better performance.

[0192] In addition, ablation experiments were also conducted on each module of the DPN of the probability prediction model provided by the present application. Figure 11 The results of the ablation experiments are shown. It can be seen from the figure that removing any one of the three modules TPRM, SPRM, and PMM of DPN will lead to a decrease in the performance of the probability prediction model, demonstrating the effectiveness and necessity of each module.

[0193] It should be noted that the DPN network provided by the present application can be applied not only to Figure 5 the probability prediction model shown, but also to probability prediction models with other structures. It can be understood that the base network in probability prediction models with other structures is different from Figure 5 the base network in the probability prediction model shown. Based on this, a compatibility analysis experiment was also conducted on the DPN network provided by the present application. The DPN network provided by the present application was respectively applied to each probability prediction model of DNN, DIN, and DIEN. From Figure 12 the results shown, it can be seen that when the DPN network is applied in the probability prediction model, the performance of the probability prediction model will be significantly improved.

[0194] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0195] Based on the same inventive concept, an embodiment of the present application further provides an interactive prediction device for a product for implementing the interactive prediction method of the product involved above. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the interactive prediction device for the product provided below can refer to the limitations on the interactive prediction method of the product in the above text, and will not be repeated here.

[0196] In one embodiment, as Figure 13 shown, an interactive prediction device for a product is provided, including: an interactive information sequence acquisition module 1302, an original interaction mode determination module 1304, a mode conversion module 1306, a mode matching module 1308, and a probability determination module 1310, where:

[0197] The interactive information sequence acquisition module 1302 is configured to acquire an interactive information sequence obtained by a target object performing product interaction on a pushed product.

[0198] The original interaction mode determination module 1304 is configured to determine the correlation between the product information of the product to be pushed and each piece of interactive information in the interactive information sequence, and extract the original interaction mode related to the product to be pushed from the interactive information sequence based on the correlation.

[0199] The mode conversion module 1306 is configured to convert the original interaction mode into a standardized interaction mode.

[0200] The mode matching module 1308 is configured to determine the target interaction mode corresponding to the product to be pushed, and perform mode matching between the standardized interaction mode and the target interaction mode to obtain mode matching features.

[0201] The probability determination module 1310 is configured to determine the probability of the target object interacting with the product to be pushed based on the mode matching features.

[0202] In one embodiment, as Figure 14As shown, the device further includes: an attention processing module 1312, configured to respectively perform vectorization processing on each piece of interaction information in the product information and the interaction information sequence to obtain a product feature vector and each interaction feature vector; perform attention processing on the product feature vector and each interaction feature vector to obtain an attention processing result; the attention processing result includes the attention weights of each interaction feature vector relative to the product feature vector; the original interaction mode determination module 1304 is further configured to: determine the correlation between the product information and each piece of interaction information based on the attention weights of each interaction feature vector relative to the product feature vector.

[0203] In one embodiment, the attention processing result further includes a fused interaction feature vector corresponding to the interaction feature vector; the probability determination module 1310 is further configured to: determine the probability that the target object interacts with the product to be pushed based on the pattern matching feature, the fused interaction feature vector, and the product feature vector.

[0204] In one embodiment, the original interaction mode determination module 1304 is further configured to: select, from each piece of interaction information in the interaction information sequence, a target interaction information whose correlation satisfies the correlation condition; intercept a first interaction information sequence including the target interaction information from the interaction information sequence; and determine the first interaction information sequence as the original interaction mode related to the product to be pushed.

[0205] In one embodiment, the mode conversion module 1306 is further configured to: perform vectorization processing on the first interaction information corresponding to the original interaction mode to obtain each first interaction feature vector; the first interaction information is the interaction information in the original interaction mode; perform attention processing on each first interaction feature vector to obtain each enhanced first interaction feature vector; perform mapping processing on each enhanced first interaction feature vector to obtain the first normalized probability corresponding to each first interaction information; and denoise the original interaction mode based on the first normalized probability corresponding to each first interaction information to obtain a normalized interaction mode.

[0206] In one embodiment, the enhanced first interaction feature vector and the first normalized probability are obtained by processing through a neural network; as Figure 14 shown, the device further includes a model training module 1314, configured to: add noise to the sample interaction information sequence to obtain a second interaction information sequence; perform vectorization processing on each piece of second interaction information in the second interaction information sequence to obtain each second interaction feature vector; perform attention processing on the second interaction feature vectors through the neural network to be trained to obtain each enhanced second interaction feature vector, and perform mapping processing on each second interaction feature vector to obtain the second normalized probability corresponding to each second interaction information; and optimize the neural network to be trained based on the second normalized probability corresponding to each second interaction information to obtain a neural network.

[0207] In one embodiment, the model training module 1314 is further configured to: determine a loss coefficient corresponding to each second interaction information based on a matching relationship between the interaction information in the sample interaction information sequence and each second interaction information in the second interaction information sequence; determine a training loss value based on the loss coefficient corresponding to each second interaction information and the second normalized probability corresponding to each second interaction information; and adjust the parameters of the neural network to be trained based on the training loss value until a training stop condition is reached, thereby obtaining a neural network.

[0208] In one embodiment, the mode conversion module 1306 is further configured to: in the original interaction mode, determine target first interaction information whose first normalized probability does not meet the probability condition as noise information; and delete the noise information from the original interaction mode to obtain a normalized interaction mode.

[0209] In one embodiment, the mode matching module 1308 is further configured to: extract a third interaction information sequence from the interaction information sequence; determine an expected interaction information sequence of the product to be pushed based on the product information; and splice the third interaction information sequence and the expected interaction information sequence to obtain a target interaction mode.

[0210] In one embodiment, the mode matching module 1308 is further configured to: perform corresponding fusion on the interaction information in the normalized interaction mode and the interaction information in the target interaction mode to obtain a fusion result; and perform pooling processing on the fusion result to obtain a mode matching feature.

[0211] In one embodiment, the probability is a first probability; as Figure 14 shown, the apparatus further includes a product pushing module 1316, configured to: obtain a second probability of the target object interacting with other products to be pushed; sort the product to be pushed and other products to be pushed based on the first probability and the second probability to obtain a sorting result; select a target product to be pushed that meets the pushing condition from the products to be pushed and other products to be pushed according to the sorting result; and push the target product to be pushed to the target object.

[0212] Each module in the above product interaction prediction apparatus can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above respective modules.

[0213] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as Figure 15As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store at least one of interaction information and product information of products to be pushed. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements an interaction prediction method for a product.

[0214] Those skilled in the art can understand that Figure 15 the structure shown in the figure is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0215] In one embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the steps in the above method embodiments are implemented.

[0216] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0217] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0218] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. And the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards of relevant countries and regions.

[0219] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0220] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0221] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. An interaction prediction method for a product, characterized in that, The method includes: Obtaining an interaction information sequence obtained by a target object's product interaction with the already pushed products; Determining the relevance between the product information of the product to be pushed and each interaction information in the interaction information sequence, and extracting the original interaction patterns related to the product to be pushed from the interaction information sequence based on the relevance; Converting the original interaction patterns into standardized interaction patterns; Determining the target interaction pattern corresponding to the product to be pushed, and performing pattern matching between the standardized interaction pattern and the target interaction pattern to obtain pattern matching features; Based on the pattern matching features, determining the probability of the target object interacting with the product to be pushed.

2. The method according to claim 1, wherein The method further includes: Performing vectorization processing on the product information and each interaction information in the interaction information sequence to obtain a product feature vector and each interaction feature vector; Performing attention processing on the product feature vector and each of the interaction feature vectors to obtain an attention processing result; the attention processing result includes the attention weights of each of the interaction feature vectors relative to the product feature vector; The determining the relevance between the product information of the product to be pushed and each interaction information in the interaction information sequence includes: Based on the attention weights of each of the interaction feature vectors relative to the product feature vector, determining the relevance between the product information and each of the interaction information.

3. The method according to claim 2, wherein The attention processing result further includes the fused interaction feature vector corresponding to the interaction feature vector; the determining the probability of the target object interacting with the product to be pushed based on the pattern matching features includes: Based on the pattern matching features, the fused interaction feature vector, and the product feature vector, determining the probability of the target object interacting with the product to be pushed.

4. The method according to claim 1, characterized in that The extracting the original interaction patterns related to the product to be pushed from the interaction information sequence based on the relevance includes: Among each interaction information in the interaction information sequence, selecting the target interaction information whose relevance satisfies the relevance condition; Intercepting a first interaction information sequence including the target interaction information from the interaction information sequence; Determining the first interaction information sequence as the original interaction pattern related to the product to be pushed.

5. The method according to claim 1, wherein The converting the original interaction patterns into standardized interaction patterns includes: Performing vectorization processing on the first interaction information corresponding to the original interaction pattern to obtain each first interaction feature vector; the first interaction information is the interaction information in the original interaction pattern; Performing attention processing on each of the first interaction feature vectors to obtain each enhanced first interaction feature vector; Performing mapping processing on each of the enhanced first interaction feature vectors to obtain the first standardized probability corresponding to each of the first interaction information; Based on the first standardized probability corresponding to each of the first interaction information, denoising the original interaction pattern to obtain a standardized interaction pattern.

6. The method according to claim 5, wherein The enhanced first interaction feature vector and the first standardized probability are obtained by processing through a neural network; the method further includes: Adding noise to the sample interaction information sequence to obtain a second interaction information sequence; Vectorize each second interaction information in the second interaction information sequence to obtain each second interaction feature vector; Perform attention processing on the second interaction feature vectors through the neural network to be trained to obtain each enhanced second interaction feature vector, and perform mapping processing on each of the second interaction feature vectors to obtain the second normalized probability corresponding to each of the second interaction information; Optimize the neural network to be trained based on the second normalized probability corresponding to each of the second interaction information to obtain the neural network.

7. The method according to claim 6, wherein The method further includes: Determine the loss coefficient corresponding to each of the second interaction information based on the matching relationship between the interaction information in the sample interaction information sequence and each of the second interaction information in the second interaction information sequence; The optimizing the neural network to be trained based on the second normalized probability corresponding to each of the second interaction information to obtain the neural network includes: Determine a training loss value based on the loss coefficient corresponding to each of the second interaction information and the second normalized probability corresponding to each of the second interaction information; Adjust the parameters of the neural network to be trained based on the training loss value until a training stop condition is reached to obtain the neural network.

8. The method according to claim 5, characterized in that, The denoising the original interaction pattern based on the first normalized probability corresponding to each of the first interaction information to obtain a normalized interaction pattern includes: In the original interaction pattern, determine the target first interaction information whose first normalized probability does not satisfy the probability condition as noise information; Delete the noise information from the original interaction pattern to obtain a normalized interaction pattern.

9. The method according to claim 1, wherein The determining the target interaction pattern corresponding to the product to be pushed includes: Extract a third interaction information sequence from the interaction information sequence; Determine the expected interaction information sequence of the product to be pushed according to the product information; Concatenate the third interaction information sequence and the expected interaction information sequence to obtain a target interaction pattern.

10. The method according to claim 1, wherein The performing pattern matching between the normalized interaction pattern and the target interaction pattern to obtain a pattern matching feature includes: Perform corresponding fusion on the interaction information in the normalized interaction pattern and the interaction information in the target interaction pattern to obtain a fusion result; Perform pooling processing on the fusion result to obtain a pattern matching feature.

11. The method according to any one of claims 1 to 10, characterized in that, The probability is the first probability; after determining the probability that the target object interacts with the product to be pushed based on the pattern matching feature, the method further includes: Obtain a second probability that the target object interacts with other products to be pushed; Rank the product to be pushed and the other products to be pushed based on the first probability and the second probability to obtain a ranking result; Select a target product to be pushed that meets the push condition from the product to be pushed and the other products to be pushed according to the ranking result; Push the target product to be pushed to the target object.

12. An interaction prediction device for a product, characterized in that, The device includes: An interaction information sequence acquisition module, configured to acquire an interaction information sequence obtained by a target object performing product interaction on a product that has been pushed; An original interaction mode determination module, configured to determine the relevance between the product information of the product to be pushed and each piece of interaction information in the interaction information sequence, and extract the original interaction mode related to the product to be pushed from the interaction information sequence based on the relevance; A mode conversion module, configured to convert the original interaction mode into a standardized interaction mode; A mode matching module, configured to determine the target interaction mode corresponding to the product to be pushed, and perform mode matching between the standardized interaction mode and the target interaction mode to obtain mode matching features; A probability determination module, configured to determine the probability of interaction between the target object and the product to be pushed based on the mode matching features.

13. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 11 are implemented.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images