Commodity click rate prediction method and commodity recommendation method
Through the multi-behavior Euler Transformer model and the comparison learning module, the problem of incomplete user preference capture in the existing technology is solved, more accurate and diverse product recommendations are achieved, and click-through rate prediction effect is improved.
Patent Information
- Application Number
- CN202411376255.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-30
- Publication Date
- 2024-12-20
AI Technical Summary
The prior art is difficult to fully capture user preferences, resulting in the recommended items being single and not diverse enough, and the single behavior sequence cannot effectively reflect the user's complex behavior patterns and changes in interests.
Using the Euler Transformer model based on multi-behavior, the intermediate features of the user behavior sequence are extracted by comparing learning and multi-behavior adaptive fusion modules, and characterizing and transforming them in complex vector space. Combining the multi-head attention mechanism to calculate attention scores, multiple Characterizes interest in this behavior, and then predicts product click-through rate.
It improves the effectiveness of product click-through rate prediction, captures user preferences more comprehensively, enhances the diversity and accuracy of recommendations, and can better model user behavior, especially in terms of long-tail distribution sequences and position differences.
Smart Images
Figure CN119168698A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of product recommendation, and in particular to a product click-through rate prediction method and a product recommendation method based on a multi-behavior Euler Transformer. Background Art
[0002] In today's era of continuous development of information technology, recommendation systems have become an important part of major Internet platforms. In many fields, whether it is e-commerce websites, social media platforms, or online video services, recommendation systems analyze user behaviors and preferences to more accurately recommend content or products that users may be interested in. The development of recommendation systems has not only improved the user experience, but also significantly increased the user stickiness and commercial value of the platform. In the recommendation system, click-through rate estimation is a crucial link and is the basis for recommendations in the fields of short videos, e-commerce, and advertising. It aims to help the system more accurately sort and display recommended content by predicting the probability of users clicking on a recommended item. Therefore, it has become a hot topic of extensive research in the industry.
[0003] In actual recommendation scenarios, user interaction data usually presents multiple behaviors and sequences. The user's historical behavior can effectively reflect their interests and preferences. For example, the products that users have clicked on, the videos they have browsed, the keywords they have searched for, and other behavioral data are all direct reflections of their potential needs and interests. However, the related technologies that use a single behavior sequence often cannot fully reflect the user's preferences, resulting in the recommendation of relatively simple items and insufficient diversity. Summary of the invention
[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a product click-through rate prediction method and a product recommendation method that have good product click-through rate prediction effect, more comprehensively capture user preferences, and better model user behavior.
[0005] The purpose of the present invention can be achieved by the following technical solutions:
[0006] According to a first aspect of the present invention, a method for predicting a click rate of a product is provided, the method comprising:
[0007] Get the user-item click dataset;
[0008] Based on the user-item click dataset, obtaining user characteristics and user behavior sequences;
[0009] The target item attributes, the user characteristics and the user behavior sequence are used as inputs, and a trained click-through rate prediction model is used to obtain the user's click rate on the target item; wherein the trained click-through rate prediction model extracts intermediate features based on the target item attributes, the user characteristics and the user behavior sequence, respectively, and obtains the click rate of the item based on the spliced features after splicing the intermediate features.
[0010] As a preferred technical solution, the click rate prediction model includes:
[0011] The embedding module is used to embed the target item attributes, the user characteristics and the user behavior sequence to obtain the item embedding representation, the user embedding representation and the user behavior embedding representation.
[0012] As a preferred technical solution, the user behavior sequence includes a like behavior sequence and a dislike behavior sequence;
[0013] The click rate prediction model includes:
[0014] The contrast learning module is used to enhance the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence.
[0015] As a preferred technical solution, the contrastive learning module is specifically used for:
[0016] At least one contrast enhancement strategy among the anti-noise strategy, the random shuffle strategy and the random loss strategy is adopted to perform feature enhancement on the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence.
[0017] As a preferred technical solution, the contrastive learning module is specifically used for:
[0018] Based on contrast loss, feature enhancement is performed on the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence;
[0019] The contrast loss is:
[0020]
[0021] Where N is the number of samples; b i + is the enhanced sample after enhancement; (b i ,b i + ) and (b i ,b j ) are positive sample pairs and negative sample pairs respectively; τ is the temperature parameter.
[0022] As a preferred technical solution, the user behavior sequence includes a like behavior sequence, a dislike behavior sequence, a click behavior sequence and a browse behavior sequence;
[0023] The click rate prediction model also includes:
[0024] The Euler Transformer module is used to use the item embedding representation as the Transformer query Q, use the user behavior sequence embedding representation as the index K and content V respectively, use the Euler transformation to convert the embedding representation to the complex vector space, and use the multi-head attention mechanism based on vector rotation operation to calculate the attention score to obtain multiple behavior interest representations.
[0025] As a preferred technical solution, the click rate prediction model also includes:
[0026] The multi-behavior adaptive fusion module is used to extract the weights of the multiple behavior interest representations using a trained neural network, and perform weighted processing on the multiple behavior interest representations based on the weights to obtain a weighted representation.
[0027] As a preferred technical solution, the click rate prediction model also includes:
[0028] A multi-layer perceptron is used to predict the click rate of the item based on the spliced features of the item embedding representation, the user embedding representation and the weighted representations of the multiple behavioral interest representations.
[0029] As a preferred technical solution, the method further includes:
[0030] The click-through rate prediction model is trained using the Adam optimizer, and the model parameters of the click-through rate prediction model are adjusted until the model performance index reaches the optimal level, thereby obtaining the trained click-through rate prediction model.
[0031] According to a second aspect of the present invention, a commodity recommendation method is provided, the method comprising:
[0032] Based on the click-through rate of the target items, the top K target items with greater probability are selected as the products to be recommended; wherein the click-through rate of the target items is obtained based on the product click-through rate prediction method provided by the first aspect or any optional implementation manner of the first aspect;
[0033] Push the recommended products to the user.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] 1. Good product click-through rate prediction effect: The product click-through rate prediction method provided by the present invention adopts multi-behavior Euler Transformer to better model user behavior, enhances item representation through comparative learning and multi-behavior adaptive fusion, and improves the effect of product click-through rate estimation.
[0036] 2. More comprehensive capture of user preferences: The product click-through rate prediction method provided by the present invention introduces multi-behavior sequences and uses the designed Euler Transformer. At the same time, it designs comparative learning tasks for sequences with long-tail distribution, supplements semi-supervised signals, and more comprehensively captures user preferences.
[0037] 3. Better modeling of user behavior: The commodity click rate prediction method provided by the present invention describes the semantic differences and position differences between sequence items in a unified form in the complex vector space, thereby better modeling user behavior. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 A schematic diagram of a flow chart of a method for predicting a commodity click rate provided by an embodiment of the present invention;
[0039] Figure 2 A schematic diagram of a model architecture of a click rate prediction model provided in an embodiment of the present invention;
[0040] Figure 3 A schematic diagram of the architecture of a comparative learning module provided in an embodiment of the present invention;
[0041] Figure 4 A schematic diagram of the architecture of the Euler Transformer module provided in an embodiment of the present invention;
[0042] Figure 5 A schematic diagram of the structure of a multi-behavior adaptive fusion module provided in an embodiment of the present invention;
[0043] Figure 6 A schematic diagram of a flow chart of a product recommendation method provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0045] "One embodiment" or "embodiment" referred to herein refers to a particular feature, structure, or characteristic that may be included in at least one implementation of the present application.
[0046] The click rate prediction model follows the multi-layer perceptron & embedding paradigm, and the input usually contains several fields, such as user behavior sequence, user attributes, target items, target item attributes, etc. In behavior sequence modeling, deep learning methods have shown significant results due to their powerful feature extraction and expression capabilities. Compared with traditional methods, deep learning can better capture complex behavior patterns and changes in interests over long time spans. Among these behavior sequence modeling methods, some works use methods based on recurrent neural networks, long short-term memory networks, Transformer, etc. to convert behavior sequences of different lengths into fixed-dimensional embeddings, capturing a single interest from a single user-item interaction behavior.
[0047] User behavior data is often represented as a time-ordered interaction sequence. In Transformer, position encoding is an important component for sequential data. The existing position encoding is usually an adaptive and learnable absolute position encoding (i.e., a unique embedding is assigned to each position), which is superimposed with feature embedding to supplement the position semantics of context representation.
[0048] It is worth noting that a single behavior sequence often cannot fully reflect the user's preferences, resulting in a relatively simple and insufficiently diverse recommended item. At the same time, the absolute position encoding in the Transformer can only encode the observed position and cannot be generalized to the unobserved position in the long sequence. Moreover, the absolute position encoding cannot capture the relative position relationship between items in the sequence, which is very important for understanding the user's behavior paradigm and sequence dependency. The items in the recommendation system are distributed in a long tail, and some items cannot be fully represented due to the small amount of data. The above defects seriously restrict the application of existing click-through rate prediction methods in actual recommendation systems.
[0049] Based on this, the present application embodiment provides a method for predicting the click rate of a product. Figure 1 As shown, Figure 1 This is a flowchart of a method for predicting the click rate of a product provided in the embodiments of the present application. This application provides method operation steps as described in the embodiments or flowcharts, but based on conventional or non-creative labor, more or fewer operation steps may be included. The order of steps listed in the embodiments is only one of many step execution orders and does not represent the only execution order. The method should be implementable by software and / or hardware. Please refer to Figure 1 , the above-mentioned commodity click rate prediction method may include:
[0050] Step S110: Obtaining a user-item click data set;
[0051] Step S120: obtaining user features and user behavior sequences based on the user-item click data set;
[0052] Step S130: Using the target item attributes, user characteristics and user behavior sequence as input, using the trained click-through rate prediction model to obtain the user's click rate on the target item; wherein the trained click-through rate prediction model extracts intermediate features based on the target item attributes, user characteristics and user behavior sequence, respectively, and obtains the item's click rate based on the spliced features after splicing the intermediate features.
[0053] The above steps S110 to S130 are described in detail below:
[0054] The user-item click data set in the above step S110 includes user characteristics such as user age, gender, region, occupation, and user behavior logs, wherein the user behavior logs record the time when the user behavior occurs.
[0055] It is understandable that the data contained in the above user-item click data set can include numerical data and categorical data according to the type. For numerical features, they can be standardized so that they are between 0 and 1. For categorical features, different categories can be encoded separately, and the default value can be filled with 0.
[0056] The method for obtaining the user behavior sequence in step S120 may be: generating a user behavior sequence according to the user behavior, behavior type and behavior occurrence time recorded in the user behavior log. Taking the KuaiRand dataset as an example, the previous behavior can be traced back according to the time when the current data occurred, and four behavior sequences of click behavior, browsing behavior, like behavior and dislike behavior with a length of 50 are constructed for each data. If the sequence is less than 50, it can be filled with 0.
[0057] The click rate prediction model in step S130 is introduced below:
[0058] like Figure 2 As shown, the above click rate prediction model may include:
[0059] An embedding module is used to embed the target item attributes, user characteristics and user behavior sequence to obtain the item embedding representation, user embedding representation and user behavior embedding representation;
[0060] A contrastive learning module, used for enhancing the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence;
[0061] The Euler Transformer module is used to use the item embedding representation as the Transformer query Q, use the user behavior sequence embedding representation as the index K and content V respectively, use the Euler transformation to convert the embedding representation to the complex vector space, and use the multi-head attention mechanism based on vector rotation operation to calculate the attention score to obtain multiple behavior interest representations;
[0062] A multi-behavior adaptive fusion module is used to extract the weights of multiple behavior interest representations using a trained neural network, and perform weighted processing on the multiple behavior interest representations based on the weights to obtain a weighted representation;
[0063] A multi-layer perceptron is used to predict the click rate of items based on the concatenated features of the item embedding representation, the user embedding representation, and the weighted representation of multiple behavioral interest representations.
[0064] The following is an introduction to each of the above modules:
[0065] (1) Embedded module:
[0066] The sparse one-hot encoded categorical data can be converted into low-dimensional dense vectors, and the numerical features can be concatenated with the dense vectors to obtain user embedding representations and item embedding representations;
[0067] The user behavior sequence can be represented as an embedding matrix with a size of sequence length × embedding dimension.
[0068] (2) Comparative learning module:
[0069] It is understandable that some behaviors, such as likes and dislikes, have strong emotional connotations, which leads to the long-tail distribution problem of such behavior sequences. Figure 3 As shown, the contrastive learning module can adopt at least one contrast enhancement strategy among the anti-noise strategy, the random shuffle strategy and the random loss strategy to enhance the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence. And based on the contrast loss, the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence are enhanced, and the contrast loss is:
[0070]
[0071] Where N is the number of samples; b i + is the enhanced sample after enhancement; (b i ,b i + ) and (b i ,b j ) are positive sample pairs and negative sample pairs respectively; τ is the temperature parameter.
[0072] (3)Eulerian Transformer module:
[0073] like Figure 4 As shown in the figure, the item embedding is used as the query of the Transformer, and the embedding corresponding to the behavior sequence is indexed and the content is converted to the complex vector space using Euler transformation. The vector dot product operation in the multi-head attention is converted into a vector rotation operation, and the attention score is calculated to obtain the representation of various behavioral interests.
[0074] The Euler transformation is fully described by the following formula:
[0075]
[0076] Among them, e j is the representation of the object, r j =e j [1:d / 2],s j =e j [d / 2+1:d] represent e j After the division, the first half acts as the real part and the second half acts as the imaginary part. yes The polar coordinate representation of .
[0077] The calculation of the attention score evolves into a rotation operation, which is fully described by the following formula:
[0078]
[0079] Among them, q i ,k j Represents the query and key index of item i and item j respectively, β i ,β j Represent the fixed rotation angles of the i-th position and the j-th position, β j =j·10000 -2j / d , d is the embedding dimension.
[0080] (4) Multi-behavior adaptive fusion module:
[0081] like Figure 5 As shown in the figure, the Euler Transformer is used to obtain the representation of the interests of different behaviors. In order to distinguish the importance of different behaviors, the above multiple representations are concatenated and input into the average pooling layer and the maximum pooling layer respectively to obtain the pooled feature vector, and the pooled vector is input into the neural network containing the linear layer and the Gelu activation function. The above transformed feature vectors are added and passed through the Sigmoid activation function to obtain the final weighted vector. The different behavior representations are weighted respectively to obtain the weighted final feature vector.
[0082] (5) Multilayer Perceptron:
[0083] For the representations after the embedding module, Euler Transformer module, multi-behavior adaptive fusion module, and contrastive learning module, they are connected to obtain the final feature y. After full connection, the click probability of the user-item pair is obtained through the sigmoid function. The mathematical formula is as follows:
[0084] p = sigmoid(y)
[0085] Where p represents the click probability of a user-item pair.
[0086] The above click-through rate prediction model can use the Adam optimizer to train the click-through rate prediction model during training, and adjust the model parameters of the click-through rate prediction model until the model performance index reaches the optimal level, thereby obtaining a trained click-through rate prediction model.
[0087] In addition, the training data set can use the KuaiRand data set as a basis. In order to ensure the rationality of the recommendation and take the time sequence into consideration, the click behavior data that occurred before the set time node is used as the training set, and the click behavior data that occurred after the set time node is used as the validation set of the model.
[0088] The following is a specific application case to introduce the application process of the above product click rate prediction method:
[0089] The present invention is explained by taking the KuaiRand dataset as an example. The KuaiRand dataset is the first recommendation dataset that inserts random products into the real sequence recommendation scenario in the industry. It records all the user's interaction records within four weeks, as well as their source scenarios (15 in total) and includes 12 types of user feedback (whether to click, viewing time, whether to like, etc.). Specifically, it includes the following steps:
[0090] Mainly include:
[0091] Step 1. In order to ensure the rationality of the recommendation, the time sequence is taken into consideration. 20220417 is used as the time node, the data before it is used as the training set, and the data after it is used as the validation set. 900,000 pieces of data are used as the training set for training, and 250,000 pieces of data are used to verify the performance of this method. Generate behavior sequences: According to the time when the current data occurs, trace back the previous behaviors, and construct four behavior sequences of 50 click behaviors, browsing behaviors, like behaviors, and dislike behaviors for each piece of data. Sequences less than 50 are filled with 0; Different category features are encoded separately, and missing values are filled with 0.
[0092] Step 2: The processed data is first converted into a dense vector through an embedding layer, and the embedding dimension is set to 4. The like and dislike behavior sequence data is enhanced using a contrastive learning module because of its long-tail distribution;
[0093] Auxiliary contrast loss is introduced to enhance the representation of long-tail distribution items. The contrast loss is as follows:
[0094]
[0095] Where N is the number of samples, b i + represents the enhanced sample after adopting the enhancement strategy, (b i ,b i + ),(b i ,b j ) represent positive sample pairs and negative samples respectively, and τ is the temperature parameter.
[0096] Euler Transformer is used to process four kinds of behavior sequences. Taking the click sequence as an example, a detailed introduction is given. The sequence length is 50. The target item serves as the query, and the click sequence serves as the index and content. That is, the query sequence Q = [q1, q2, ..., q 50 ], index sequence K = [k1, k2, ..., k 50 ] and the content sequence V = V = [v1, v2, ..., v 50 ], perform Euler transformation on the query sequence and index sequence and calculate the attention score as follows:
[0097]
[0098] Among them, q i ,k j Represents the query and index of item i and item j respectively, β i ,β j Represent the fixed rotation angles of the i-th position and the j-th position, β j =j·10000 -2jd , d is the embedding dimension.
[0099] After the above transformation, the output of the behavior sequence after Euler Transformer is as follows:
[0100]
[0101] Here, d is the dimension of the embedding.
[0102] The four behavior sequences are processed by Euler Transformer to obtain the characteristic behavior interest representation, which is recorded as hist click,hist pv ,hist like ,hist dislike , the four interest representations are input into the multi-behavior adaptive fusion module, the above multiple representations are concatenated and input into the average pooling layer and the maximum pooling layer respectively to obtain the pooled feature vector, the pooled vector is input into the neural network containing the linear layer and the Gelu activation function, the above transformed feature vectors are added together and passed through the Sigmoid activation function to obtain the final weighted vector, and the different behavior representations are weighted respectively to obtain the weighted final feature vector.
[0103] The processed embeddings are connected to get the final feature y. After full connection, the click probability of the user-item pair is obtained through the sigmoid function, which is fully described by the following formula:
[0104] p = sigmoid(y)
[0105] Where p represents the click probability of a user-item pair.
[0106] Step 3: Input the segmented data set into the model and train the model using the Adam optimizer. The loss function is as follows:
[0107] L overall =L ctr +γL cl
[0108]
[0109] Among them, y i ,p i are the label value and the predicted value respectively, and γ is the weight of the auxiliary loss function introduced by the contrastive learning task.
[0110] Adjust the parameters until the model indicator AUC reaches the optimal value and save the weights.
[0111] In this embodiment, the number of Transformer layers is selected to be 3, the number of multi-head attention heads is selected to be 2, and the auxiliary loss function weight γ=0.01.
[0112] The comparison with other models is carried out in the same way. The comparison results are shown in Table 1. The AUC of the method of the present invention reaches 0.7767, which is better than other click rate prediction methods.
[0113] Table 1 Comparison of model effects
[0114]
[0115] Step 4: Input the user-item pair data into the click rate prediction model, and use the weights obtained in the above steps to obtain the click probability of the user-item pair.
[0116] Based on the same inventive concept, the embodiment of the present application also provides a product recommendation method, which includes:
[0117] Based on the click-through rate of the target items, the top K target items with the highest probability are selected as the products to be recommended; wherein the click-through rate of the target items is obtained based on any of the above-mentioned product click-through rate prediction methods;
[0118] Push the recommended products to the user.
[0119] It is understandable that if Figure 6 As shown, the user-item pairs can be output in the form of an item recall table, grouped by user number, and based on the probability size, for each user, the top K items with higher probability are taken as the recommendation results.
[0120] The embodiment of the present application provides a method for predicting the click-through rate of goods based on a multi-behavior Euler Transformer, which has the following outstanding innovations compared to the methods in the prior art: First, the present invention captures the user's specific behavioral interests more comprehensively by introducing multi-behavior sequences, enhances the representation of long-tail distribution sequences by using a contrastive learning module, and clarifies the importance of different behavioral interests through a multi-behavior adaptive fusion module, thereby improving the diversity and comprehensiveness of recommendations. Secondly, the present invention designs a new Euler Transformer, which models semantic differences and positional differences in a unified form in complex vector space, so that user behavioral interests can be better modeled. Finally, compared with other click-through rate prediction models, the present invention has a significant improvement and has a certain improvement in the click-through rate prediction effect.
[0121] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. A method for predicting a product click rate, characterized in that: The method comprises: Get the user-item click dataset; Based on the user-item click dataset, obtaining user characteristics and user behavior sequences; The target item attributes, the user characteristics and the user behavior sequence are used as inputs, and a trained click-through rate prediction model is used to obtain the user's click rate on the target item; wherein the trained click-through rate prediction model extracts intermediate features based on the target item attributes, the user characteristics and the user behavior sequence, respectively, and obtains the click rate of the item based on the spliced features after splicing the intermediate features.
2. The method for predicting the click rate of a product according to claim 1, characterized in that: The click rate prediction model includes: The embedding module is used to embed the target item attributes, the user characteristics and the user behavior sequence to obtain the item embedding representation, the user embedding representation and the user behavior embedding representation.
3. The method for predicting the click rate of a commodity according to claim 2, characterized in that: The user behavior sequence includes a like behavior sequence and a dislike behavior sequence; The click rate prediction model includes: A contrastive learning module is used to enhance the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence.
4. The method for predicting the click rate of a commodity according to claim 3, characterized in that: The contrastive learning module is specifically used for: At least one contrast enhancement strategy among the anti-noise strategy, the random shuffle strategy and the random loss strategy is adopted to perform feature enhancement on the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence.
5. The method for predicting the click rate of a commodity according to claim 3, characterized in that: The contrastive learning module is specifically used for: Based on contrast loss, feature enhancement is performed on the intermediate features extracted based on the like behavior sequence and the intermediate features extracted based on the dislike behavior sequence; The contrast loss is: Where N is the number of samples; b i + is the enhanced sample after enhancement; (b i ,b i + ) and (b i ,b j ) are positive sample pairs and negative sample pairs respectively; τ is the temperature parameter.
6. The method for predicting the click rate of a commodity according to claim 2, characterized in that: The user behavior sequence includes a like behavior sequence, a dislike behavior sequence, a click behavior sequence and a browse behavior sequence; The click rate prediction model also includes: The Euler Transformer module is used to use the item embedding representation as the Transformer query Q, use the user behavior sequence embedding representation as the index K and content V respectively, use the Euler transformation to convert the embedding representation to the complex vector space, and use the multi-head attention mechanism based on vector rotation operation to calculate the attention score to obtain multiple behavior interest representations.
7. The method for predicting the click rate of a commodity according to claim 6, characterized in that: The click rate prediction model also includes: The multi-behavior adaptive fusion module is used to extract the weights of the multiple behavior interest representations using a trained neural network, and perform weighted processing on the multiple behavior interest representations based on the weights to obtain a weighted representation.
8. The method for predicting the click rate of a commodity according to claim 7, characterized in that: The click rate prediction model also includes: A multi-layer perceptron is used to predict the click rate of the item based on the spliced features of the item embedding representation, the user embedding representation and the weighted representations of the multiple behavioral interest representations.
9. The method for predicting the click rate of a commodity according to any one of claims 1 to 8, characterized in that: The method further comprises: The click-through rate prediction model is trained using the Adam optimizer, and the model parameters of the click-through rate prediction model are adjusted until the model performance index reaches the optimal level, thereby obtaining the trained click-through rate prediction model.
10. A product recommendation method, characterized in that: The method comprises: Based on the click-through rate of the target items, the top K target items with greater probability are selected as the products to be recommended; wherein the click-through rate of the target items is obtained based on the product click-through rate prediction method according to any one of claims 1 to 9; Push the recommended products to the user.