Method for generating interest representation, and method and device for training interest representation model

Abstract

This application discloses a method for generating interest representations, a method for training interest representation models, and an apparatus, relating to the field of artificial intelligence technology. The method includes: acquiring a behavioral sequence of a target object; encoding the behavioral sequence of the target object to obtain a preliminary behavioral sequence representation of the target object; obtaining a discrete feature vector of the target object from a global interest table based on the preliminary behavioral sequence representation of the target object; wherein the global interest table includes multiple discrete feature vectors used to represent different interest points; and generating an interest representation of the target object based on the target discrete feature vectors, wherein the interest representation of the target object is used to represent the interests of the target object. The technical solution provided by the embodiments of this application improves the interpretability and effectiveness of interest representations by generating an interest representation of a target object based on a preliminary behavioral sequence representation and multiple discrete feature vectors used to represent different interest points.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for generating interest representations, a method for training interest representation models, and an apparatus. Background Technology

[0002] With the development of artificial intelligence technology, neural networks are widely used in recommender systems to generate interest representations, which are used to represent the interests and preferences of objects.

[0003] In related technologies, neural networks are used to represent an object's interests and preferences as a continuous and dense vector. This vector representation is then learned and optimized through supervised or self-supervised training to obtain the final representation of the object's interests.

[0004] However, continuous and dense vectors are prone to feature drift, which can lead to poor interpretability and effectiveness of interest representations. Summary of the Invention

[0005] This application provides a method for generating interest representations, a method for training interest representation models, and an apparatus, which can improve the interpretability and effectiveness of interest representations. The technical solution may include the following:

[0006] According to one aspect of the embodiments of this application, a method for generating interest representations is provided, the method comprising:

[0007] Obtain the behavior sequence of the target object;

[0008] The behavioral sequence of the target object is encoded to obtain a preliminary behavioral sequence representation of the target object;

[0009] Based on the preliminary behavioral sequence of the target object, a discrete feature vector of the target is obtained from the global interest table; wherein, the global interest table includes multiple discrete feature vectors used to represent different interest points;

[0010] Based on the target discrete feature vector, an interest representation of the target object is generated, and the interest representation of the target object is used to characterize the interest of the target object.

[0011] According to one aspect of the embodiments of this application, a training method for an interest representation model is provided, the interest representation model including an encoding network, a generator network, and a decoder network, the method comprising:

[0012] Obtain the behavior sequence of the sample object;

[0013] The behavioral sequence of the sample object is encoded by the encoding network to obtain a preliminary behavioral sequence representation of the sample object;

[0014] The generative network generates an interest representation of the sample object based on the preliminary behavioral sequence representation of the sample object, and the interest representation of the sample object is used to characterize the interest of the sample object.

[0015] The decoding network generates a behavior sequence reconstruction vector of the sample object based on the preliminary behavior sequence representation and interest representation of the sample object;

[0016] The interest representation model is trained based on the reconstructed vector of the behavior sequence and the behavior sequence of the sample object.

[0017] According to one aspect of the embodiments of this application, an apparatus for generating interest representations is provided, the apparatus comprising:

[0018] The behavior sequence acquisition module is used to acquire the behavior sequence of the target object;

[0019] The behavior sequence encoding module is used to encode the behavior sequence of the target object to obtain a preliminary behavior sequence representation of the target object;

[0020] The target vector acquisition module is used to acquire a discrete feature vector of the target from a global interest table based on the preliminary behavioral sequence representation of the target object; wherein, the global interest table includes multiple discrete feature vectors used to represent different interest points;

[0021] The interest representation generation module is used to generate an interest representation of the target object based on the target discrete feature vector, and the interest representation of the target object is used to characterize the interest of the target object.

[0022] According to one aspect of the embodiments of this application, a training apparatus for an interest representation model is provided, the interest representation model including an encoding network, a generator network, and a decoder network, the apparatus comprising:

[0023] The behavior sequence acquisition module is used to acquire the behavior sequence of the sample object;

[0024] The behavior sequence encoding module is used to encode the behavior sequence of the sample object through the encoding network to obtain a preliminary behavior sequence representation of the sample object;

[0025] An interest representation generation module is used to generate an interest representation of the sample object based on the preliminary behavioral sequence representation of the sample object through the generation network. The interest representation of the sample object is used to characterize the interest of the sample object.

[0026] The reconstruction vector generation module is used to generate a behavior sequence reconstruction vector of the sample object based on the preliminary behavior sequence representation and interest representation of the sample object through the decoding network;

[0027] The model training module is used to train the interest representation model based on the behavior sequence reconstruction vector and the behavior sequence of the sample object.

[0028] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the above-described method for generating interest representations, or to implement the above-described method for training interest representation models.

[0029] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the storage medium, the computer program being loaded and executed by a processor to implement the above-described method for generating interest representations, or to implement the above-described method for training interest representation models.

[0030] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the above-described method for generating an interest representation, or the above-described method for training an interest representation model.

[0031] The technical solutions provided in this application include at least the following beneficial effects.

[0032] By obtaining the target discrete feature vector from the global interest table based on the preliminary behavior sequence representation, and then generating the interest representation of the target object based on the target discrete feature vector, the effectiveness and interpretability of the interest representation are improved, since the target discrete feature vector is related to the behavior sequence of the target object and different target discrete feature vectors correspond to different interest points. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the implementation environment of a solution provided in one embodiment of this application;

[0035] Figure 2This is a schematic diagram of the interest representation model used in one embodiment of this application;

[0036] Figure 3 This is a schematic diagram of an interest representation model provided in one embodiment of this application;

[0037] Figure 4 This is a schematic diagram of a generative network provided in one embodiment of this application;

[0038] Figure 5 This is a flowchart of a training method for an interest representation model provided in one embodiment of this application;

[0039] Figure 6 This is a flowchart of a method for generating training-side interest representations according to an embodiment of this application;

[0040] Figure 7 This is a flowchart of a method for generating side interest representations according to an embodiment of this application;

[0041] Figure 8 This is a block diagram of an interest representation generation apparatus provided in one embodiment of this application;

[0042] Figure 9 This is a block diagram of a training apparatus for an interest representation model provided in another embodiment of this application;

[0043] Figure 10 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0045] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0046] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0047] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learn-by-doing.

[0048] The technical solution provided in this application relates to machine learning technology in artificial intelligence. By utilizing machine learning technology, an interest representation model is trained unsupervised based on the behavior sequence of sample objects to obtain an interest representation model that can generate interest representations based on behavior sequences.

[0049] The method provided in this application can be executed by a computer device, which refers to an electronic device with data computing, processing, and storage capabilities. This computer device can be a terminal such as a PC (Personal Computer), tablet computer, smartphone, wearable device, intelligent robot, or vehicle-mounted device; it can also be a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0050] The technical solutions provided in this application are applicable to any scenario requiring the generation of interest representations, such as recommendation scenarios, recall scenarios, shopping scenarios, entertainment scenarios, and audio / video playback scenarios. The technical solutions provided in this application can improve the effectiveness and interpretability of interest representations.

[0051] The following section will provide a detailed description of the model structure, training method, and usage method of the interest representation model provided in the embodiments of this application.

[0052] Please refer to Figure 1The diagram illustrates an implementation environment for a solution provided in one embodiment of this application. This implementation environment may include a model training device 10 and a model usage device 20.

[0053] The model training device 10 can be an electronic device such as a PC, computer, tablet computer, intelligent robot, vehicle terminal, or other electronic devices with strong computing power, or it can be a server. The model training device 10 is used to train the interest representation model 30. Optionally, the model training device 10 can be used to train the interest representation model 30 offline.

[0054] In this embodiment, the interest representation model 30 is a neural network model that can be used to generate interest representations of a target object. This interest representation can be a discrete vector, used to represent the target object's interests or characterize the target object. For example, the interest representation model 30 can generate an interest representation of the target object based on a sequence of the target object's behaviors. This sequence of behaviors can include a series of behaviors of the target object within a set time period. For instance, it can obtain a series of behaviors of the target object in the recommendation system, such as clicking, reading, watching, and purchasing, within a set time period before the current moment (including the current moment), and then sort these behaviors in chronological order to obtain the target object's behavior sequence. The target object can refer to any object, such as any user or user account in the recommendation system.

[0055] In this embodiment, the interest representation model 30 can be constructed based on the Transformer model (a model based on a multi-head attention mechanism), or it can be constructed based on any sequence-to-reconstruction sequence model. In fact, any model that satisfies the encoder-decoder structure can be used to construct the interest representation model 30, and this embodiment does not limit it.

[0056] Optionally, the model training device 10 can use machine learning to train the interest representation model 30 based on the behavioral sequences of the sample objects, so that it has better interest representation generation performance.

[0057] The trained interest representation model 30 can be deployed on the model-using device 20 to provide online interest representation generation services or behavior sequence reconstruction services. The model-using device 20 can be a terminal device such as a mobile phone, computer, smart TV, multimedia playback device, wearable device, vehicle terminal, or intelligent robot, or it can be a server; this embodiment of the application does not limit this.

[0058] In one example, a final interest representation model can be constructed based on a portion of the network in the trained interest representation model 30 for deployment on the model-using device 20, such as constructing the final interest representation model based on the encoding network 310 and the generative network 320 described below. This simplifies the model structure and saves model space.

[0059] In some embodiments, such as Figure 1 As shown, the interest representation model 30 may include an encoding network 310, a generation network 320, and a decoding network 330.

[0060] The encoding network 310 is a neural network used to encode input. The input to the encoding network 310 is an embedded sequence of behaviors, i.e., a behavior sequence vector corresponding to the behavior sequence. The output of the encoding network 310 is a preliminary behavior sequence representation, which is used to initially represent the behavior sequence of an object. For example, the preliminary behavior sequence representation may include multiple behavior sub-vectors, each representing the behavior of the object.

[0061] In one example, the encoding network 310 includes multiple sub-networks with identical structures, where the output of one sub-network serves as the input to the next. For example, refer to... Figure 3 The encoding network 310 includes N (e.g., 6) subnetworks, each of which may include a multi-head self-attention layer, a normalization layer, and a fully connected layer. For each subnetwork, the input of the subnetwork is vectorized, then passes through a multi-head self-attention layer to obtain a context-aware vector, and finally passes through several fully connected layers and normalization layers to obtain the output of the subnetwork.

[0062] Optionally, the last sub-network corresponding to the encoding network 310 is also connected to a multi-head transformation layer, which is a fully connected layer used to perform dimensionality transformation on the output of the sub-network. For example, if the output of the last sub-network is a 32*64 preliminary behavioral sequence representation, then the multi-head transformation layer can convert this preliminary behavioral sequence representation into a k*64 (i.e., multi-head structure) preliminary behavioral sequence representation, where k can be set and adjusted according to actual usage requirements. Optionally, this multi-head transformation layer can be set in the last sub-network or in the generator network 320; this application does not limit this.

[0063] Generative Network 320 is a neural network that generates interest representations. The input to Generative Network 320 is the preliminary behavioral sequence representation under the multi-head structure corresponding to the target object, and the output of Generative Network 320 is the interest representation of the target object. Generative Network 320 is connected to the aforementioned multi-head transformation layer; that is, the output of the multi-head transformation layer is the input to Generative Network 320.

[0064] In the embodiments of this application, reference is made to Figure 4 The generator network 320 is configured with a global interest table 321, which includes multiple discrete feature vectors representing different interest points. For example, in a recommendation system, each category, item, behavior, and any content that can be considered an interest point can be maintained as a discrete feature vector. For instance, a sports category can be maintained as a discrete feature vector and assigned an identifier, such as a serial number or ID (identification). Similarly, a jersey within a sports category can also be maintained as a discrete feature vector. This embodiment does not limit the method of maintaining the global interest target.

[0065] The decoding network 330 is a neural network that decodes the input. The input to the decoding network 330 is a fused behavior sequence representation between the interest representation and the preliminary behavior sequence representation. The output of the decoding network 330 is a behavior sequence reconstruction vector, which can be used to reconstruct the reconstructed behavior sequence of the target object. The decoding network 330 is connected to the encoding network 310 and the generator network 320; that is, the output of the encoding network 310 and the input of the generator network 320 are the inputs to the decoding network 330.

[0066] In one example, the decoding network 330 comprises multiple subnetworks with identical structures. The output of the previous subnetwork serves as the input to the next subnetwork. For example, refer to... Figure 3 The decoding network 330 includes N (e.g., 6) sub-networks, and each sub-network may include a multi-head self-attention layer with masking operations (corresponding to...). Figure 3 The network consists of a multi-head self-attention layer 311, a normalization layer, a fully connected layer, and a multi-head self-attention layer with encoding and decoding capabilities (corresponding to multi-head self-attention layer 312 in the diagram). For each sub-network, the input is vectorized, then passes through the multi-head self-attention layer to obtain a context-aware vector, and finally passes through several fully connected layers and normalization layers to obtain the output of the sub-network. The last sub-network of the decoding network 330 outputs a behavior sequence reconstruction vector, which can be used to reconstruct the reconstructed behavior sequence of the target object.

[0067] In this context, the attention matrix corresponding to the multi-head self-attention layer 311 uses a masking operation, ensuring that when the multi-head self-attention layer 311 completes context awareness, the decoding of each behavior in the behavior sequence can only perceive the previously occurring behavior. Figure 3 The shift behavior sequence is provided. Optionally, the shift behavior sequence can be obtained by uniformly shifting the behaviors in the behavior sequence one position to the right. The multi-head self-attention layer 312 can be used to process the above fused behavior sequence representation.

[0068] In one example, reference Figure 2 During the use of the interest representation model 30, a final interest representation model can be constructed based on the encoding network 310 and the generator network 320 in the trained interest representation model 30. This final interest representation model is then deployed in the model-using device 20 to generate interest representations for target objects online. Optionally, the interest representation can be used as a feature of the object and applied to the recall, coarse ranking, and fine ranking modules of the downstream recommendation system. This application embodiment does not limit the application scenarios of the interest representation model; it can be used in recommendation processes for various scenarios such as text, images, videos, items, and news.

[0069] Please refer to Figure 5 The diagram illustrates a flowchart of a training method for an interest representation model according to an embodiment of this application. The execution entity for each step of this method can be the model training device described above. The method may include the following steps (501-505).

[0070] Step 501: Obtain the behavior sequence of the sample object.

[0071] A sample object refers to the object collected in the training data; it can refer to any object. For example, a sample object can refer to any user or user account in the recommendation system. The behavioral sequence of the sample object serves as the training sample for the interest representation model. The behavioral sequence of the sample object can include a series of behaviors of the sample object within a set time period. For example, obtaining a series of behaviors of the sample object in the recommendation system, such as clicking, reading, watching, and purchasing, within a set time period before the current moment (including the current moment), and then sorting these behaviors in chronological order, yields the behavioral sequence of the sample object.

[0072] The interest representation model on the training side is the same as described in the above embodiments. For content not described in the embodiments of this application, please refer to the above embodiments, and it will not be repeated here.

[0073] Step 502: Encode the behavior sequence of the sample object through an encoding network to obtain a preliminary behavior sequence representation of the sample object.

[0074] A preliminary behavior sequence representation is a sequence of behaviors used to initially represent an object. For example, a preliminary behavior sequence representation of a sample object is a sequence of behaviors used to initially represent the sample object. This preliminary behavior sequence representation may include multiple behavior sub-vectors, each representing a behavior.

[0075] In one example, reference Figure 3 The behavior sequence of the sample object can be embedded and encoded first to obtain the behavior sequence vector of the sample object. Then, the behavior sequence vector of the sample object can be encoded by the encoding network 310 to obtain the preliminary behavior sequence representation of the sample object.

[0076] Step 503: Generate an interest representation of the sample object based on the preliminary behavioral sequence representation of the sample object through a generative network. The interest representation of the sample object is used to characterize the interest of the sample object.

[0077] In this embodiment, the interest representation can be a discrete vector used to characterize the interest of an object.

[0078] In one example, reference Figure 6 Step 503 may also include the following sub-steps.

[0079] Step 503a: The generator network determines the weight parameters corresponding to multiple discrete feature vectors based on the preliminary behavioral sequence representation of the sample objects and the multiple discrete feature vectors contained in the global interest table; wherein, different discrete feature vectors are used to represent different interest points.

[0080] A global interest table can be used to represent all points of interest across the entire network, system, or client. For example, in a recommendation system, all points of interest that can be considered objects can be maintained as discrete feature vectors, and each discrete feature vector can be numbered to obtain the global interest table corresponding to the recommendation system. The dimension of the discrete feature vectors can be the same as the dimension of the initial behavior sequence representation (or the behavior sub-vectors in the initial behavior sequence representation).

[0081] For example, the process of obtaining the weight parameters can be as follows:

[0082] 1. Perform dimensional transformation on the preliminary behavioral sequence representation of the sample object to obtain the preliminary behavioral sequence representation under a multi-head structure. This preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object.

[0083] Optionally, the preliminary behavioral sequence representation of the sample object can be dimensionally transformed using the aforementioned multi-head transformation layer to obtain a preliminary behavioral sequence representation under a multi-head structure. The head vector of the sample object is used to represent the prediction interest of the sample object.

[0084] For example, suppose the initial behavior sequence representation under a multi-head structure includes 4 head vectors, and the initial behavior sequence representation of a sample object includes 32 behavior sub-vectors, each with 64 dimensions. Then, the initial behavior sequence representation can be converted into 4 head vectors, each with 64 dimensions, through a multi-head transformation layer.

[0085] 2. For the first discrete feature vector among multiple discrete feature vectors, the first similarity between multiple head vectors and the first discrete feature vector is calculated by generating a network.

[0086] The first discrete eigenvector can refer to any one of multiple discrete eigenvectors.

[0087] For example, refer to Figure 4 The first similarity between the four head vectors and the first discrete feature vector in the global interest table 321 can be calculated by generating network 320.

[0088] The first similarity can be calculated using algorithms such as cosine similarity, Euclidean distance, Manhattan distance, and Chebyshev distance.

[0089] 3. Sum the first similarity to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector.

[0090] For example, the first similarity between the multiple head vectors and the first discrete feature vector is summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector.

[0091] For example, based on the above embodiments, the first similarity between the four head vectors and the first discrete feature vector is summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector.

[0092] 4. Determine the weight parameters corresponding to the first discrete feature vector based on the second similarity.

[0093] Optionally, the second similarity between the initial behavioral sequence representation under the multi-head structure and the first discrete feature vector can be directly determined as the weight parameter corresponding to the first discrete feature vector.

[0094] Using the same method as for the first discrete feature vector, the weight parameters corresponding to the multiple discrete feature vectors are obtained respectively.

[0095] Step 503b: Based on the weight parameters, perform a weighted summation of multiple discrete feature vectors to generate an interest representation of the sample object.

[0096] The interest representation of the sample object can be generated by weighting and summing multiple discrete feature vectors based on the weight parameters corresponding to each discrete feature vector.

[0097] Step 504: Generate a behavior sequence reconstruction vector of the sample object by decoding the network based on the preliminary behavior sequence representation and interest representation of the sample object.

[0098] The behavior sequence reconstruction vector represents the reconstructed behavior sequence. Further transformation of the sequence reconstruction vector yields the reconstructed behavior sequence.

[0099] In one example, the process of obtaining the behavior sequence reconstruction vector can be as follows: the preliminary behavior sequence representation and interest representation of the sample object are added together to obtain the fused behavior sequence representation of the sample object; the fused behavior sequence representation of the sample object is decoded by a decoding network to generate the behavior sequence reconstruction vector of the sample object.

[0100] For example, refer to Figure 3 The preliminary behavior sequence representation output by the encoding network 310 and the interest representation output by the generator network 320 are added together to obtain the fused behavior sequence representation of the sample object. Then, the fused behavior sequence representation is decoded by the decoding network 330 to obtain the behavior sequence reconstruction vector of the sample object, and thus the reconstructed behavior sequence of the sample object.

[0101] Step 505: Based on the behavior sequence, reconstruct the vector and the behavior sequence of the sample object, and train the interest representation model.

[0102] Optionally, the reconstruction loss of the interest representation model can be obtained based on the reconstructed vector of the behavior sequence and the corresponding behavior sequence vector of the sample object. Then, the interest representation model is iteratively trained based on the reconstruction loss to obtain the trained interest representation model. The reconstruction loss can be obtained using cross-entropy loss functions, mean squared error loss functions, etc. The stochastic gradient descent method can be used for iterative training of the interest representation model.

[0103] In one example, the final interest representation model can be constructed based on a portion of the network in the trained interest representation model. For instance, referring to Figure 2, the final interest representation model can be constructed based on the encoding network 310 and the generator network 320 in the trained interest representation model 30 for deployment in the model-using device 20.

[0104] In summary, the technical solution provided in this application generates an interest representation based on a preliminary behavior sequence through a generative network. Then, it combines the interest representation and the preliminary behavior sequence representation to obtain a behavior sequence reconstruction vector. Finally, it trains the interest representation model based on the behavior sequence reconstruction vector and the behavior sequence, thereby achieving unsupervised training of the interest representation model and reducing the training cost of the interest representation model.

[0105] Furthermore, by combining interest representation and preliminary behavior sequence representation, a more accurate behavior sequence reconstruction vector can be obtained, thereby improving the generation accuracy of the interest representation model.

[0106] Furthermore, by generating interest representations based on the global interest table, the interest representations generated by the interest representation model become more effective (e.g., the meaning of the vectors has substantial significance) and interpretable (e.g., the meaning of the vectors can be understood by humans), thereby further improving the generation accuracy of the interest representation model.

[0107] Please refer to Figure 7 The diagram illustrates a flowchart of a method for generating an interest representation according to an embodiment of this application. The entity executing each step of this method can be the model-using device described above. The method may include the following steps (701-704).

[0108] Step 701: Obtain the behavior sequence of the target object.

[0109] In this application, the target object is not limited. For example, in a recommendation system, the target object can refer to any object in the recommendation system. The behavior sequence of the target object can include a series of behaviors of the target object within a set time period. For example, in a recommendation system, the behavior sequence of the target object can include a series of behaviors such as clicking, reading, watching, and purchasing by the target object within the recommendation system and the set time period.

[0110] Step 702: Encode the behavior sequence of the target object to obtain a preliminary behavior sequence representation of the target object.

[0111] A preliminary behavior sequence represents a sequence of behaviors used to initially represent an object. For example, a preliminary behavior sequence for a target object represents a sequence of behaviors used to initially represent the target object.

[0112] In one example, the interest representation in this application embodiment is generated by an interest representation model, which includes an encoding network and a generating network. The interest representation model used is the same as that described in the above embodiment, and will not be repeated here.

[0113] The encoding network encodes the behavioral sequence of the target object to obtain a preliminary behavioral sequence representation of the target object; the generator network generates an interest representation of the target object based on the preliminary behavioral sequence representation and the global interest table.

[0114] For example, refer to Figure 2 The interest representation model includes an encoding network 310 and a generator network 320. By encoding the behavior sequence of the target object through the encoding network 310, a preliminary behavior sequence representation of the target object can be obtained.

[0115] Step 703: Obtain the target discrete feature vector from the global interest table based on the preliminary behavioral sequence representation of the target object; wherein, the global interest table includes multiple discrete feature vectors used to represent different interest points.

[0116] A global interest table can be used to represent all points of interest across the entire network, system, or client. A target discrete feature vector can be a discrete feature vector in the global interest table associated with the initial behavioral sequence representation of the target object. For example, a target discrete feature vector can be matched with the initial behavioral sequence representation of the target object, or it can be matched with the head vector in the initial behavioral sequence representation of the target object. The head vector is used to represent the predicted interest of the target object.

[0117] Optionally, the preliminary behavioral sequence representation of the target object can be first transformed to obtain a preliminary behavioral sequence representation under a multi-head structure. This preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the prediction interest of the target object. Based on the preliminary behavioral sequence representation under the multi-head structure and multiple discrete feature vectors, the target discrete feature vector is obtained.

[0118] In one example, the process of obtaining the target discrete feature vector can be as follows: obtain the first similarity between each head vector and each discrete feature vector; obtain n target discrete feature vectors corresponding to multiple head vectors in descending order of the first similarity, where n is a positive integer.

[0119] For example, suppose the initial behavioral sequence representation E under a multi-head structure includes: E1, E2, ..., Ei (the i-th head vector), and the global interest table D includes: D1, D2, ..., Dj (the j-th discrete feature vector). Taking E1 as an example, the first similarity between E1 and the j discrete feature vectors can be obtained respectively. The first similarity can be calculated using algorithms such as cosine similarity, Euclidean distance, Manhattan distance, and Chebyshev distance. Based on the order of the first similarity from largest to smallest, n target discrete feature vectors corresponding to E1 are determined from the j discrete feature vectors, where n is less than or equal to j.

[0120] In another example, the process of obtaining the target discrete feature vector can also be as follows: obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the multiple discrete feature vectors respectively; obtain m target discrete feature vectors in descending order of the second similarity, where m is a positive integer.

[0121] Specifically, for the first discrete feature vector among multiple discrete feature vectors, the first similarity between the first discrete feature vector and each of the multiple head vectors is calculated; the first similarity is then summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector. The first discrete feature vector can refer to any discrete feature vector in the global interest table.

[0122] For example, based on the above embodiments, taking D1 as an example, the first similarity between D1 and i head vectors is obtained respectively, and then the i first similarities are summed to obtain the preliminary behavioral sequence representation under the multi-head structure and the second similarity between D1 and D1. The same method is used to obtain the second similarity corresponding to D2 to Dj respectively. Finally, according to the second similarity from large to small, m target discrete feature vectors are obtained from the global interest table, where m is less than j.

[0123] Step 704: Generate the interest representation of the target object based on the target discrete feature vector. The interest representation of the target object is used to characterize the interest of the target object.

[0124] Optionally, a discrete vector-type interest representation can be generated based on the identifier of the target discrete feature vector.

[0125] For example, the identifier of the target discrete feature vector can be the ID of the target discrete feature vector. The IDs of the i*n target discrete feature vectors are concatenated to obtain the discrete vector-type interest representation corresponding to the target object. Alternatively, the IDs of the m target discrete feature vectors are concatenated to obtain the discrete vector-type interest representation corresponding to the target object. Compared to continuous vector-type interest representations, discrete vector-type interest representations are more stable and occupy less space. Optionally, steps 703 and 704 can be performed using a generative network. For example, refer to... Figure 2 By generating network 320, the target discrete feature vector is obtained from the global interest table based on the preliminary behavioral sequence representation of the target object, and then the interest representation of the target object is generated based on the target discrete feature vector.

[0126] Optionally, after generating the interest representation of the target object, a set of recommended content for the target object can be obtained based on the interest representation of the target object; wherein, the recommended content in the set of recommended content refers to content that matches the interest representation of the target object.

[0127] For example, consider a recommendation system. Content retrieval can be performed based on the target object's interest representation to obtain candidate content. This content can include images and text, videos, audio, news, items, knowledge, etc. Then, combined with the interest representation, the candidate content undergoes coarse-sorting and fine-sorting processes to obtain the recommended content corresponding to the target object, thereby generating a set of recommended content for the target object, which is then recommended to the target object.

[0128] In some embodiments, in actual business scenarios, comparative experiments were conducted on content domain interests and advertising domain interests to determine whether the interest representation obtained through the technical solution provided in the embodiments of this application can effectively distinguish between objects and the interests of the represented objects.

[0129] Given an object to be evaluated, objects with similar interest representations are found according to the Jaccard distance of interest representations. The overlap between the two objects in content domain behavior and advertising domain behavior is calculated. Objects are randomly selected for comparison to observe whether objects with similar interests also have similar interest representations, so as to prove the effectiveness of the technical solution provided in the embodiments of this application.

[0130] The degree of behavioral overlap between the object to be evaluated and randomly selected objects can be set to 1 to compare relative values. The comparison results of content domain interests are shown in Table 1 below.

[0131] Table 1

[0132]

[0133] The comparison results of advertising domain interests are shown in Table 2 below.

[0134] Table 2

[0135]

[0136] As can be seen, the solution provided in this application embodiment can effectively represent the interests of objects in both the content domain and the advertising domain, for use in downstream recommendation systems. Taking advertising recommendation business as an example, in both business scenarios, the offline model evaluation Recall@10 and HR@100 of the solution provided in this application embodiment show an improvement of more than 1%.

[0137] The solution provided in this application also features relatively stable interest representation. For example, Table 3 below shows the similarity between the interest representations of the same object on two consecutive days in different scenarios.

[0138] Table 3

[0139] Scene Similarity Scenario A 0.764 Scene B 0.803

[0140] In summary, the technical solution provided in this application obtains the target discrete feature vector from the global interest table based on the preliminary behavior sequence representation, and then generates the interest representation of the target object based on the target discrete feature vector. Since the target discrete feature vector is related to the behavior sequence of the target object, and different target discrete feature vectors correspond to different interest points, the effectiveness and interpretability of the interest representation are improved.

[0141] Furthermore, based on the identifier of the target discrete feature vector, a discrete vector-based interest representation is constructed, avoiding the problems of large space occupation, easy feature drift, and poor interpretability of continuous vector-based interest representations. This reduces the space occupied by the interest representation and improves its stability and interpretability.

[0142] Furthermore, compared to related technologies that require manually pre-defining an interest system and setting up an additional understanding system for item attributes, resulting in high generation costs and susceptibility to human error, the technical solution provided in this application improves the efficiency and accuracy of interest representation generation and reduces the cost of interest representation generation because it eliminates the need for manually pre-defining an interest system and setting up an additional understanding system for item attributes.

[0143] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0144] Please refer to Figure 8 This diagram illustrates a block diagram of an interest representation generation apparatus according to an embodiment of this application. The apparatus can be used to implement the above-described interest representation generation method. The apparatus 800 may include: a behavior sequence acquisition module 801, a behavior sequence encoding module 802, a target vector acquisition module 803, and an interest representation generation module 804.

[0145] The behavior sequence acquisition module 801 is used to acquire the behavior sequence of the target object.

[0146] The behavior sequence encoding module 802 is used to encode the behavior sequence of the target object to obtain a preliminary behavior sequence representation of the target object.

[0147] The target vector acquisition module 803 is used to acquire the target discrete feature vector from the global interest table based on the preliminary behavioral sequence representation of the target object; wherein, the global interest table includes multiple discrete feature vectors used to represent different interest points.

[0148] The interest representation generation module 804 is used to generate an interest representation of the target object based on the target discrete feature vector, and the interest representation of the target object is used to characterize the interest of the target object.

[0149] In some embodiments, the target vector acquisition module 803 is configured to:

[0150] The preliminary behavioral sequence representation of the target object is subjected to dimensionality transformation to obtain a preliminary behavioral sequence representation under a multi-head structure. The preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object.

[0151] Based on the preliminary behavioral sequence representation under the multi-head structure and the multiple discrete feature vectors, the target discrete feature vector is obtained.

[0152] In some embodiments, the target vector acquisition module 803 is further configured to:

[0153] The first similarity between each head vector and each discrete feature vector is obtained respectively;

[0154] Based on the first similarity in descending order, obtain n discrete feature vectors corresponding to the multiple head vectors, where n is a positive integer.

[0155] In some embodiments, the target vector acquisition module 803 is further configured to:

[0156] The second similarity between the preliminary behavioral sequence representation under the multi-head structure and the multiple discrete feature vectors is obtained respectively;

[0157] Based on the second similarity in descending order, obtain m discrete feature vectors of the target, where m is a positive integer.

[0158] In some embodiments, the target vector acquisition module 803 is further configured to:

[0159] For the first discrete feature vector among the plurality of discrete feature vectors, calculate the first similarity between the plurality of head vectors and the first discrete feature vector;

[0160] The first similarity is summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector.

[0161] In some embodiments, the interest representation generation module 804 is used to generate a discrete vector type interest representation based on the identifier of the target discrete feature vector.

[0162] In some embodiments, the interest representation is generated by an interest representation model, which includes an encoding network and a generative network;

[0163] The encoding network is used to encode the behavioral sequence of the target object to obtain a preliminary behavioral sequence representation of the target object;

[0164] The generative network is used to generate an interest representation of the target object based on the preliminary behavioral sequence representation of the target object and the global interest table.

[0165] In some embodiments, the device 800 further includes: a recommended content acquisition module ( Figure 8 (Not shown in the image).

[0166] The recommended content acquisition module is used to acquire a set of recommended content for the target object based on the target object's interest representation; wherein, the recommended content in the recommended content set refers to content that matches the target object's interest representation.

[0167] In summary, the technical solution provided in this application obtains the target discrete feature vector from the global interest table based on the preliminary behavior sequence representation, and then generates the interest representation of the target object based on the target discrete feature vector. Since the target discrete feature vector is related to the behavior sequence of the target object, and different target discrete feature vectors correspond to different interest points, the effectiveness and interpretability of the interest representation are improved.

[0168] Please refer to Figure 9 This diagram illustrates a block diagram of a training apparatus for an interest representation model according to an embodiment of this application. The apparatus 900 can be used to implement the training method for the aforementioned interest representation model. The apparatus 900 may include: a behavior sequence acquisition module 901, a behavior sequence encoding module 902, an interest representation generation module 903, a reconstruction vector generation module 904, and a representation model training module 905.

[0169] The behavior sequence acquisition module 901 is used to acquire the behavior sequence of the sample object.

[0170] The behavior sequence encoding module 902 is used to encode the behavior sequence of the sample object through the encoding network to obtain a preliminary behavior sequence representation of the sample object.

[0171] The interest representation generation module 903 is used to generate an interest representation of the sample object based on the preliminary behavioral sequence representation of the sample object through the generation network. The interest representation of the sample object is used to characterize the interest of the sample object.

[0172] The reconstruction vector generation module 904 is used to generate a behavior sequence reconstruction vector of the sample object through the decoding network based on the preliminary behavior sequence representation and interest representation of the sample object.

[0173] The model training module 905 is used to train the interest representation model based on the behavior sequence reconstruction vector and the behavior sequence of the sample object.

[0174] In some embodiments, the interest representation generation module 903 is configured to:

[0175] The generative network determines the weight parameters corresponding to the multiple discrete feature vectors based on the preliminary behavioral sequence representation of the sample object and the multiple discrete feature vectors contained in the global interest table; wherein, different discrete feature vectors are used to represent different interest points.

[0176] Based on the weight parameters, the multiple discrete feature vectors are weighted and summed to generate the interest representation of the sample object.

[0177] In some embodiments, the interest representation generation module 903 is further configured to:

[0178] The preliminary behavioral sequence representation of the sample object is subjected to dimensionality transformation to obtain a preliminary behavioral sequence representation under a multi-head structure. The preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object.

[0179] For the first discrete feature vector among the plurality of discrete feature vectors, the first similarity between the plurality of head vectors and the first discrete feature vector is calculated by the generator network;

[0180] The first similarity is summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector;

[0181] Based on the second similarity, the weight parameters corresponding to the first discrete feature vector are determined.

[0182] In some embodiments, the reconstructed vector generation module 904 is configured to:

[0183] The preliminary behavioral sequence representation and interest representation of the sample object are added together to obtain the fused behavioral sequence representation of the sample object;

[0184] The fusion behavior sequence representation of the sample object is decoded by the decoding network to generate the behavior sequence reconstruction vector of the sample object.

[0185] In summary, the technical solution provided in this application generates an interest representation based on a preliminary behavior sequence through a generative network. Then, it combines the interest representation and the preliminary behavior sequence representation to obtain a behavior sequence reconstruction vector. Finally, it trains the interest representation model based on the behavior sequence reconstruction vector and the behavior sequence, thereby achieving unsupervised training of the interest representation model and reducing the training cost of the interest representation model.

[0186] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0187] Please refer to Figure 10This illustration shows a schematic diagram of a computer device according to an embodiment of this application. The computer device can be any electronic device with data computing, processing, and storage functions, and can be implemented as... Figure 1 The model training device 10 and / or model usage device 20 in the implementation environment of the scheme shown. Specifically, it may include the following:

[0188] The computer device 1000 includes a central processing unit (such as a CPU, GPU, or FPGA) 1001, a system memory 1004 including RAM (Random-Access Memory) 1002 and ROM (Read-Only Memory) 1003, and a system bus 1005 connecting the system memory 1004 and the central processing unit 1001. The computer device 1000 also includes a basic input / output system (I / O system) 1006 to facilitate information transfer between various devices within the server, and a large-capacity storage device 1007 for storing the operating system 1013, application programs 1014, and other program modules 1015.

[0189] In some embodiments, the basic input / output system 1006 includes a display 1008 for displaying information and an input device 1009 for user input, such as a mouse or keyboard. Both the display 1008 and the input device 1009 are connected to the central processing unit 1001 via an input / output controller 1010 connected to the system bus 1005. The basic input / output system 1006 may also include the input / output controller 1010 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1010 also provides output to a display screen, printer, or other types of output devices.

[0190] The mass storage device 1007 is connected to the central processing unit 1001 via a mass storage controller (not shown) connected to the system bus 1005. The mass storage device 1007 and its associated computer-readable media provide non-volatile storage for the computer device 1000. That is, the mass storage device 1007 may include computer-readable media (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0191] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage medium is not limited to the above-mentioned types. The system memory 1004 and mass storage device 1007 described above can be collectively referred to as memory.

[0192] According to an embodiment of this application, the computer device 1000 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1000 can be connected to the network 1012 via the network interface unit 1011 connected to the system bus 1005, or the network interface unit 1011 can be used to connect to other types of networks or remote computer systems (not shown).

[0193] The memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the above-described method for generating interest representations or the above-described method for training interest representation models.

[0194] In some embodiments, a computer-readable storage medium is also provided, wherein a computer program is stored therein, which, when executed by a processor, implements the above-described method for generating interest representations or the above-described method for training interest representation models.

[0195] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random-Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0196] In some embodiments, a computer program product or computer program is also provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the above-described method for generating interest representations, or the above-described method for training interest representation models.

[0197] It should be noted that all information (including but not limited to object device information, object personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the object or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the behavioral sequences and global interest tables involved in this application were obtained with full authorization.

[0198] It should be understood that "multiple" as used herein refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

[0199] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating interest representations, characterized in that, The method includes: Obtain the behavior sequence of the target object; The behavioral sequence of the target object is encoded to obtain a preliminary behavioral sequence representation of the target object; The preliminary behavioral sequence representation is subjected to dimensionality transformation to obtain a preliminary behavioral sequence representation under a multi-head structure. The preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object. The first similarity between each head vector and each discrete feature vector in the global interest table is obtained. The global interest table includes multiple discrete feature vectors used to represent different interest points. The interest point is at least one of the following: category, item, behavior. Based on the first similarity in descending order, obtain n discrete feature vectors corresponding to the multiple head vectors, where n is a positive integer; The identifiers of the n discrete feature vectors corresponding to the multiple head vectors are concatenated to generate the interest representation of the target object. The interest representation of the target object is used to characterize the interest of the target object, and the interest representation is a discrete vector.

2. The method according to claim 1, characterized in that, The method further includes: The second similarity between the preliminary behavioral sequence representation under the multi-head structure and the multiple discrete feature vectors is obtained respectively; Based on the second similarity in descending order, obtain m discrete feature vectors of the target, where m is a positive integer.

3. The method according to claim 2, characterized in that, The step of obtaining the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the multiple discrete feature vectors includes: For the first discrete feature vector among the plurality of discrete feature vectors, calculate the first similarity between the plurality of head vectors and the first discrete feature vector; The first similarity is summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector.

4. The method according to any one of claims 1 to 3, characterized in that, The interest representation is generated by an interest representation model, which includes an encoding network and a generative network. The encoding network is used to encode the behavioral sequence of the target object to obtain a preliminary behavioral sequence representation of the target object; The generative network is used to generate an interest representation of the target object based on the preliminary behavioral sequence representation of the target object and the global interest table.

5. The method according to any one of claims 1 to 3, characterized in that, After concatenating the identifiers of the n discrete feature vectors corresponding to the plurality of head vectors to generate the interest representation of the target object, the method further includes: Based on the target object's interest representation, a set of recommended content for the target object is obtained; wherein, the recommended content in the set of recommended content refers to content that matches the target object's interest representation.

6. A training method for an interest representation model, characterized in that, The interest representation model includes an encoding network, a generating network, and a decoding network; the method includes: Obtain the behavior sequence of the sample object; The behavioral sequence of the sample object is encoded by the encoding network to obtain a preliminary behavioral sequence representation of the sample object; The preliminary behavioral sequence representation of the sample object is subjected to dimensionality transformation to obtain a preliminary behavioral sequence representation under a multi-head structure. The preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object. For the first discrete feature vector among multiple discrete feature vectors in the global interest table, the first similarity between the multiple head vectors and the first discrete feature vector is calculated by the generator network; wherein, different discrete feature vectors are used to represent different interest points, and the interest points are at least one of the following: category, item, behavior; Based on the first similarity, determine the weight parameters corresponding to the first discrete feature vector; Based on the weight parameters, the plurality of discrete feature vectors are weighted and summed to generate the interest representation of the sample object, which is used to characterize the interest of the sample object; The decoding network generates a behavior sequence reconstruction vector of the sample object based on the preliminary behavior sequence representation and interest representation of the sample object; The interest representation model is trained based on the reconstructed vector of the behavior sequence and the behavior sequence of the sample object.

7. The method according to claim 6, characterized in that, The step of determining the weight parameters corresponding to the first discrete feature vector based on the first similarity includes: The first similarity is summed to obtain the second similarity between the preliminary behavioral sequence representation under the multi-head structure and the first discrete feature vector; Based on the second similarity, the weight parameters corresponding to the first discrete feature vector are determined.

8. The method according to claim 6, characterized in that, The step of generating a behavior sequence reconstruction vector for the sample object through the decoding network based on the preliminary behavior sequence representation and interest representation of the sample object includes: The preliminary behavioral sequence representation and interest representation of the sample object are added together to obtain the fused behavioral sequence representation of the sample object; The fusion behavior sequence representation of the sample object is decoded by the decoding network to generate the behavior sequence reconstruction vector of the sample object.

9. An apparatus for generating an interest representation, characterized in that, The device includes: The behavior sequence acquisition module is used to acquire the behavior sequence of the target object; The behavior sequence encoding module is used to encode the behavior sequence of the target object to obtain a preliminary behavior sequence representation of the target object; The target vector acquisition module is used to perform dimensionality transformation on the preliminary behavior sequence representation to obtain a preliminary behavior sequence representation under a multi-head structure. The preliminary behavior sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object. The module then acquires a first similarity between each head vector and each discrete feature vector in the global interest table. The global interest table includes multiple discrete feature vectors used to represent different interest points, where the interest point is at least one of the following: category, item, or behavior. Finally, the module acquires n target discrete feature vectors corresponding to each of the multiple head vectors, where n is a positive integer, in descending order of the first similarity. The interest representation generation module is used to concatenate the identifiers of the n discrete feature vectors corresponding to the multiple head vectors to generate the interest representation of the target object. The interest representation of the target object is used to characterize the interest of the target object, and the interest representation is a discrete vector.

10. A training device for an interest representation model, characterized in that, The interest representation model includes an encoding network, a generating network, and a decoding network, and the apparatus includes: The behavior sequence acquisition module is used to acquire the behavior sequence of the sample object; The behavior sequence encoding module is used to encode the behavior sequence of the sample object through the encoding network to obtain a preliminary behavior sequence representation of the sample object; An interest representation generation module is used to perform dimensionality transformation on the preliminary behavioral sequence representation of the sample object to obtain a preliminary behavioral sequence representation under a multi-head structure. The preliminary behavioral sequence representation under the multi-head structure includes multiple head vectors, which are used to represent the predicted interest of the object. For a first discrete feature vector among multiple discrete feature vectors in the global interest table, the generation network calculates a first similarity between the multiple head vectors and the first discrete feature vector. Different discrete feature vectors are used to represent different interest points, which are at least one of the following: category, item, or behavior. Based on the first similarity, a weight parameter corresponding to the first discrete feature vector is determined. Based on the weight parameter, the multiple discrete feature vectors are weighted and summed to generate an interest representation of the sample object, which is used to represent the interest of the sample object. The reconstruction vector generation module is used to generate a behavior sequence reconstruction vector of the sample object based on the preliminary behavior sequence representation and interest representation of the sample object through the decoding network; The model training module is used to train the interest representation model based on the behavior sequence reconstruction vector and the behavior sequence of the sample object.

11. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the method for generating interest representations as described in any one of claims 1 to 5, or to implement the method for training an interest representation model as described in any one of claims 6 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the method for generating interest representations as described in any one of claims 1 to 5, or to implement the method for training an interest representation model as described in any one of claims 6 to 8.

13. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium, and a processor reads and executes the computer instructions from the computer-readable storage medium to implement the method for generating an interest representation as described in any one of claims 1 to 5, or to implement the method for training an interest representation model as described in any one of claims 6 to 8.

Images