Commodity feature processing method and device, electronic equipment and storage medium
By acquiring contextual information from the user's browsing of product interfaces and user session sequence data, and using a multi-layer variational autoencoder to generate embedding feature vectors, the problem of lacking semantic and contextual information when representing product features using Item2vec is solved, thus achieving more accurate product recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA MOBILE INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2021-06-16
- Publication Date
- 2026-07-31
AI Technical Summary
In existing technologies, Item2vec lacks semantic and user context information when representing product features, resulting in inaccurate embedding feature representation.
By acquiring the context information of the user's browsing of the product interface and the user's session sequence data, the data is input into a variational autoencoder. The variational autoencoder with multiple encoding and decoding layers generates an embedding feature vector, which is then fused with context information and user behavior to generate a candidate product set.
It improves the accuracy of embedding feature representation, enabling more accurate identification of similar products and solving the problems of high-dimensional sparsity and difficulty in reflecting homogeneous information relationships.
Smart Images
Figure CN115482056B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a product feature processing method, apparatus, electronic device, and storage medium. Background Technology
[0002] In existing technologies, Item2vec is commonly used to represent product features. However, when using Item2vec to represent product features, it does not contain rich semantic information when processing the product access sequence for embedding (i.e., a way to convert discrete variables into continuous representations). It also loses some contextual information about the user when making the purchase, making the product embedding feature representation inaccurate. Summary of the Invention
[0003] This invention provides a product feature processing method, apparatus, electronic device, and storage medium to address the shortcomings of inaccurate embedding feature representation in the prior art and improve the accuracy of embedding feature representation.
[0004] In a first aspect, the present invention provides a product feature processing method, comprising:
[0005] Obtain the context information corresponding to the product page viewed by the user, as well as the user's session sequence data;
[0006] The context information and the user session sequence data are input into the variational autoencoder to obtain an embedding feature vector used to characterize the product features;
[0007] The context information includes the scenario of browsing a product page; the embedding feature vector is used to generate a candidate product set.
[0008] In one embodiment, inputting the context information and the user session sequence data into a variational autoencoder to obtain an embedding feature vector for characterizing product features includes:
[0009] The context information and the user session sequence data are input into the first coding layer of the variational autoencoder to obtain the initial coding latent variables;
[0010] The context information and the previous encoded latent variable are input into the second coding layer of the variational autoencoder to obtain the next encoded latent variable;
[0011] The embedding feature vector is obtained based on the last encoded latent variable.
[0012] It should be noted that latent variables are those generated during the encoding process.
[0013] In one embodiment, obtaining the embedding feature vector based on the last encoded latent variable includes:
[0014] The final encoded latent variable is input into the first decoding layer of the variational autoencoder to obtain the initial decoded latent variable;
[0015] The previous decoded latent variable is input into the second decoding layer of the variational autoencoder to obtain the next decoded latent variable;
[0016] Based on the final decoded latent variables, the embedding feature vector is obtained.
[0017] It should be noted that the latent variables in decoding are those generated during the decoding process.
[0018] In one embodiment, the initial encoded latent variable satisfies a Gaussian distribution under the context information and the user session sequence data;
[0019] The latter encoded latent variable satisfies a Gaussian distribution under the context information and the former encoded latent variable;
[0020] The initial decoded latent variable satisfies a Gaussian distribution under the context information and the final encoded latent variable;
[0021] The latter decoded latent variable satisfies a Gaussian distribution under the context information and the former decoded latent variable.
[0022] In one embodiment, the first encoding layer is used to obtain a variational lower bound of the log-likelihood function based on the context information and the user session sequence data;
[0023] The second encoding layer is used to obtain the variational lower bound of the log-likelihood function based on the context information and the previously encoded latent variables;
[0024] The variational lower bound is used to determine the accuracy of the encoded latent variables.
[0025] In one embodiment, after inputting the context information and the user session sequence data into a variational autoencoder to obtain an embedding feature vector for characterizing product features, the method further includes:
[0026] Based on the embedding feature vector, the product ID similarity matrix is obtained;
[0027] Based on the product ID similarity matrix, a candidate product set is obtained.
[0028] In one embodiment, after obtaining the candidate product set based on the product ID similarity matrix, the method further includes:
[0029] The candidate product set is sorted according to the target sorting algorithm to obtain the product sorting result;
[0030] The product sorting results are pushed to the target user.
[0031] In a second aspect, the present invention provides a commodity feature processing apparatus, comprising:
[0032] The data acquisition module is used to acquire the context information corresponding to the product interface viewed by the user, as well as the user session sequence data;
[0033] The feature extraction module is used to input the context information and the user session sequence data into the variational autoencoder to obtain an embedding feature vector for characterizing the product features.
[0034] The context information includes the scenario of browsing a product page; the embedding feature vector is used to generate a candidate product set.
[0035] Thirdly, the present invention provides an electronic device, including a processor and a memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of any of the above-described commodity feature processing methods.
[0036] Fourthly, the present invention provides a processor-readable storage medium storing a computer program for causing the processor to perform the steps of any of the above-described product feature processing methods.
[0037] This invention provides a product feature processing method, apparatus, electronic device, and storage medium. The product feature processing method first acquires the context information corresponding to the product interface viewed by the user, as well as the user's session sequence data. The context information includes the browsing scenario and spatiotemporal information of the product page. Then, the context information and the user session sequence data are input into a variational autoencoder to obtain an embedding feature vector representing product features. This embedding feature vector is used to generate a candidate product set. By incorporating context information and user session sequence data into the embedding feature vector, valuable scenario information is integrated. When using this embedding feature vector to generate the candidate product set, similar products can be more accurately identified as candidate products by calculating similarity. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0039] Figure 1 This is one of the flowcharts illustrating the commodity feature processing method provided by the present invention;
[0040] Figure 2 This is an overall architecture diagram of the product recommendation system provided by the present invention;
[0041] Figure 3 This is one of the schematic diagrams of the variational autoencoder provided by the present invention;
[0042] Figure 4 This is a schematic diagram of the encoding part of the variational autoencoder provided by the present invention;
[0043] Figure 5 This is a schematic diagram of the decoding part of the variational autoencoder provided by the present invention;
[0044] Figure 6 This is the second flowchart of the commodity feature processing method provided by the present invention;
[0045] Figure 7 This is a schematic diagram of the commodity feature processing device provided by the present invention;
[0046] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0048] The following is combined Figures 1-8 The present invention describes a product feature processing method, apparatus, electronic device, and storage medium.
[0049] In existing solutions for commodity feature processing, one-hot encoding (i.e., one-bit effective encoding) is generally used. However, this method has two main drawbacks:
[0050] One drawback is the high-dimensional sparsity problem: For high-dimensional sparsity problems, if there are N items, then the possible combinations of items that the user has interacted with are 2^N. N In this case, to make our one-hot encoding model more reliable, the number of samples required increases exponentially with the increase of N.
[0051] Another drawback is that the one-hot encoding model cannot reflect the relationships between objects: for homogeneous information, such as mobile phone information, or different items, such as iPhone 5 and iPhone 6, and iPhone 5 and Huawei, the distance is the same after conversion to one-hot encoding, but in reality, the distance between iPhone 5 and iPhone 6 should be closer.
[0052] For heterogeneous information, such as item ID (i.e., identification number) and store ID, the distance between them is even immeasurable. But in reality, the distance between a store selling iPhones and an iPhone should be much closer.
[0053] To address the aforementioned issues, the existing technical solution is to convert ID-type features into a low-dimensional embedding (i.e., a way to transform discrete variables into continuous representations) feature vector. This approach has achieved good results in recommendations within the e-commerce field.
[0054] Existing technologies commonly use word2vec to generate feature vectors when embedding product feature objects. While the underlying Skip-gram algorithm of word2vec can effectively extract semantic information into the embedding feature vector for text content, it doesn't capture the same rigorous semantic information as text, nor does it include contextual information about the user's purchase history. This results in the generated embedding feature vector lacking some useful information. To supplement this lost contextual information and make the embedding feature representation more accurate, this invention provides the following technical solution.
[0055] This invention provides a product feature processing method, such as... Figure 1 As shown, the product feature processing method includes:
[0056] Step 110: Obtain the context information corresponding to the product interface viewed by the user, as well as the user session (computer term: time domain) sequence data.
[0057] It should be noted that the context information here includes the browsing scenario of the product page and spatiotemporal information, which can be the time and location information of the user browsing the page.
[0058] Session sequence data can be attribute and configuration information generated by a user during page browsing, such as user purchase history sequence data.
[0059] It is understandable that the context information and user session sequence data here include both data collected during implementation and data collected offline.
[0060] like Figure 2 As shown, the data acquisition in step 110 includes real-time data processing on the client and server sides, near-real-time data processing on the stream processing platform, and offline data processing on the big data platform. This is achieved by collecting contextual information (which may include user information, item information, and scene information).
[0061] Step 120: Input the context information and user session sequence data into the variational autoencoder to obtain the embedding feature vector used to represent the product features.
[0062] In some embodiments, step 120 may involve uniformly processing the collected context information and user session sequence data, including normalization, discretization and nonlinear transformation processing of numerical features, one-hot encoding and embedding of ID features, and finally obtaining the embedding feature vector.
[0063] The embedding feature vector is used to generate a candidate product set.
[0064] In some embodiments, context information and user session sequence data are input into a variational autoencoder to obtain an embedding feature vector for characterizing product features, including:
[0065] The context information and user session sequence data are input into the first coding layer of the variational autoencoder to obtain the initial coding latent variables;
[0066] The context information, along with the previous encoded latent variable, is input into the second coding layer of the variational autoencoder to obtain the next encoded latent variable.
[0067] Based on the final encoded latent variables, the embedding feature vector is obtained.
[0068] It should be noted that there can be multiple second coding layers here. Each coding layer is an independent variational autoencoder inference model.
[0069] In some embodiments, the embedding feature vector is obtained based on the last encoded latent variable, including:
[0070] The final encoded latent variable is input into the first decoding layer of the variational autoencoder to obtain the initial decoded latent variable;
[0071] The previous decoded latent variable is input into the second decoding layer of the variational autoencoder to obtain the next decoded latent variable;
[0072] Based on the final decoded latent variables, the embedding feature vector is obtained.
[0073] It should be noted that here, there can be multiple second decoding layers, and the number of second decoding layers can correspond to the number of second encoding layers. Each decoding layer is an independent generative model of a variational autoencoder.
[0074] Variational autoencoders have multiple coding layers and multiple decoding layers, which can transform high-bit sparse feature vectors into dense low-bit feature vectors, thus solving the high-bit sparsity problem.
[0075] It is understandable that, for the first encoding layer, the initial encoded latent variables satisfy a Gaussian distribution under contextual information and user session sequence data.
[0076] For the second coding layer, the latter encoded latent variable satisfies a Gaussian distribution under the context information and the former encoded latent variable.
[0077] For the first decoding layer, the initial decoding latent variables satisfy a Gaussian distribution under the context information and the final encoding latent variables.
[0078] For the second decoding layer, the latent variables of the latter decoding layer satisfy a Gaussian distribution under the context information and the latent variables of the former decoding layer.
[0079] In some embodiments, the first encoding layer is used to obtain the variational lower bound of the log-likelihood function based on contextual information and user session sequence data.
[0080] The second encoding layer is used to obtain the variational lower bound of the log-likelihood function based on contextual information and the previously encoded latent variables.
[0081] The variational lower bound is used to determine the accuracy of the encoded latent variables.
[0082] It is understandable that the first and second coding layers are output after the accuracy of the coding latent variables is achieved based on the variational lower bound.
[0083] In some embodiments, after inputting context information and user session sequence data into a variational autoencoder to obtain an embedding feature vector for characterizing product features, the product feature processing method further includes:
[0084] Based on the embedding feature vector, the product ID similarity matrix is obtained;
[0085] Based on the product ID similarity matrix, a candidate product set is obtained.
[0086] It is understandable that the product ID similarity matrix is a matrix composed of the IDs of similar products.
[0087] In some embodiments, after obtaining the candidate product set based on the product ID similarity matrix, the product feature processing method further includes:
[0088] The candidate product set is sorted according to the target sorting algorithm to obtain the product sorting result;
[0089] Push the product sorting results to the target users.
[0090] In some embodiments, the candidate product set is sorted according to the target sorting algorithm to obtain the product sorting result, including:
[0091] The candidate product set is input into the recall layer to obtain the product recall results;
[0092] The product recall results are input into the sorting layer to obtain the product sorting results.
[0093] The recall layer aims to quickly and accurately filter out relevant target items and narrow down the candidate set, while the ranking layer aims to improve the recommendation effect by ranking the recommendation list accurately. The ranking layer is equipped with a target ranking algorithm.
[0094] You can continue to refer to this. Figure 2 In the model section, the recommendation model mainly consists of two parts: the recall layer and the ranking layer. The main relevant recommendation algorithms are also concentrated in this part. The recall layer includes multi-path recall such as popular item recall, category-related recall, related item recall, and social relationship recall. The recommendation engine filters the results of multi-path recall, and the remaining items serve as the candidate set for the ranking layer.
[0095] The ranking layer includes methods such as LR (logistic regression), FM (factor factorization machine), MLR (a non-linear model equivalent to clustering plus LR) and combined models. The ranking layer performs fine ranking on the candidate set selected by the recall layer, and takes the top n products from the ranking results as a list to recommend to the user.
[0096] The method provided by this invention improves the embedding method for one of the recall layers in a recommendation model. It fuses contextual information (including time, location, and page context) with user session sequence data and inputs it into a variational autoencoder. This variational autoencoder, with multiple encoding and decoding layers, learns a representation of product ID features. Finally, it uses the embedding feature vector of the product ID to filter candidate products from a massive database, serving as the candidate set for one of the recall layers. Figure 3 As shown, latent variables are generated under the combined influence of sample data and contextual information. The process of reconstructing input data is also generated under the combined influence of latent variables and contextual information.
[0097] For a variational autoencoder that incorporates contextual information, reconstructing the input data requires maximizing the learning model parameters of the joint log-likelihood function of the sample data x and the contextual information t. The joint log-likelihood function is:
[0098]
[0099] Where, p θ This represents the probability distribution of the corresponding events during the decoding process:
[0100] p θ (x|t) represents the probability of x occurring given that t occurs during the decoding process.
[0101] p θ (z|x,t) represents the probability of z occurring given that x and t have occurred during the decoding process.
[0102] p θ (x,z|t) represents the probability of x and z occurring during the decoding process, given that t occurs.
[0103] q φ This represents the probability distribution of the corresponding event during the encoding process:
[0104] q φ (z|x) represents the probability of z occurring given that x occurs during the encoding process.
[0105] q φ (z|x,t) represents the probability of z occurring given that x and t occur during the encoding process.
[0106] Understandably, KL represents divergence, and E represents expected value. That is, q φ The expected value of (z|x,t).
[0107] This is called cross-entropy, which is the difference between the predicted value and the actual value.
[0108] The variational lower bound of the joint log-likelihood function is used as the surrogate objective function for optimization, and its variational lower bound l T (x,t;θ,φ) can be written as:
[0109]
[0110] Here, θ and φ are the neural network parameters of the variational autoencoder, namely the weights and biases. θ and φ can be randomly initialized, for example, to a normal distribution or a uniform distribution, and then adjusted through learning.
[0111] However, shallow models with only one layer of random hidden variables limit the flexibility of the hidden representation. To improve this, this invention proposes a deep variational autoencoder model that integrates contextual information and user session sequences (i.e., a model of a variational autoencoder with stacked multi-layer coding and decoding layers). Compared to single-layer variational autoencoder models, deep variational autoencoder models are better at fitting complex functions.
[0112] Figure 4 The diagram shows the inference model of a deep variational autoencoder, also known as the coding part or coding layer. Figure 5 The image shows the generative model of the deep variational autoencoder, also known as the decoding part or decoding layer.
[0113] The stacked form of variational autoencoders is similar to a stacked autoencoder. For the (i-1)th inference model (i.e., the (i-1)th layer of the deep variational autoencoder model), zi The output z of the previous inference model i-1 The context information t is used as the input vector for the current layer, while z... i This is the output of the current layer.
[0114] This stacking method eventually forms a deep network with L layers. In this deep network, there are L latent variables z. i , where z i =1, 2, ..., L, a hierarchical inference model allows hidden variables to be highly correlated with the coding layer, while still maintaining the computational efficiency of a fully decomposed model.
[0115] In the model, the hidden variable z is divided into L layers. The parameter learning of the inference model in the deep variational autoencoder model is a bottom-up process. Each inference layer is defined as a Gaussian distribution given the inference model below it (i.e., the previous encoding layer) and the topic vectors (i.e., the hidden variables). (Each inference layer uses a Gaussian distribution for definition.)
[0116]
[0117]
[0118] The random variables in each layer of the generative model (i.e., the latent variables output by each decoding layer) are considered to be Gaussian distributions resulting from the combined effects of the random variables in the previous layer and contextual information.
[0119]
[0120]
[0121] The above formula can be optimized using stochastic gradient descent to obtain the parameters θ and φ of the generative and inference models, respectively. μ and σ represent the expected value and variance, respectively.
[0122] like Figure 5 The diagram shows the learning process of a deep variational autoencoder model for the feature representation of product IDs in an online marketplace.
[0123] Figure 6 In this model, 't' represents the user's context information, and 'x' represents the user's session sequence data of browsing products as sample data. This user session sequence data, incorporating context information, is used to train a deep variational autoencoder model to learn the feature representation of product IDs. The similarity relationships between products are calculated using the learned embedding feature vectors representing product IDs. A candidate product set is generated as one recall candidate result, which is then refined through a ranking layer. Finally, the results from multiple recall paths are combined and recommended to the user.
[0124] Online testing was conducted on the product feature processing method provided by this invention. Users were divided into three groups: A, B, and C, corresponding to the product feature processing method provided by this invention, as well as the collaborative filtering model and the GBDT (Gradient Boosting Decision Tree) + LR (Logistic Regression) model in existing technical solutions. After one week of observation, the test results are as follows:
[0125] The accuracy rates for Group A were 11.60%, 5.02%, 11.39%, 18.36%, 16.98%, and 14.49%, respectively, with an average accuracy rate of 12.97%.
[0126] The accuracy rates for Group B were 7.79%, 5.59%, 11.51%, 16.79%, 16.90%, and 13.79%, respectively, with an average accuracy rate of 12.06%.
[0127] The accuracy rates for Group C were 9.80%, 8.93%, 14.29%, 16.67%, 11.11%, and 14.29%, respectively, with an average accuracy rate of 12.52%.
[0128] The average accuracy of the method provided by this invention is higher than that of the other two models, indicating that the method provided by this invention is effective.
[0129] In summary, the product feature processing method provided by this invention first obtains the context information corresponding to the product interface browsed by the user, as well as the user's session sequence data. The context information includes the browsing scenario and spatiotemporal information of the product page. Then, the context information and the user's session sequence data are input into a variational autoencoder to obtain an embedding feature vector used to characterize product features. The embedding feature vector is used to generate a candidate product set. The embedding feature vector incorporates context information and user session sequence data, that is, it incorporates valuable scenario information. When using the embedding feature vector to generate a candidate product set, similar products can be more accurately found as candidate products by calculating similarity.
[0130] Furthermore, existing technologies often use one-hot encoding for feature representation, which leads to high-dimensional sparsity and fails to reflect relationships between objects with homogeneous information. The product feature processing method provided in this invention utilizes a multi-layer variational autoencoder to process contextual information and user session sequence data. This transforms the contextual information and user session sequence data from high-dimensional sparse feature vectors into dense low-dimensional feature vectors, and the incorporation of contextual information effectively solves the homogeneity problem.
[0131] The commodity feature processing apparatus provided by the present invention is described below. The commodity feature processing apparatus described below and the commodity feature processing method described above can be referred to in correspondence.
[0132] like Figure 7 As shown, the product feature processing device 700 includes a data acquisition module 710 and a feature extraction module 720.
[0133] The data acquisition module 710 is used to acquire the context information corresponding to the product interface viewed by the user, as well as the user session sequence data.
[0134] The feature extraction module 720 is used to input context information and user session sequence data into the variational autoencoder to obtain an embedding feature vector that represents the product features;
[0135] The context information includes the browsing scenario of the product page and spatiotemporal information; the embedding feature vector is used to generate a candidate product set.
[0136] In some embodiments, the feature extraction module 720 includes a first encoding unit, a second encoding unit, and a feature extraction unit.
[0137] The first coding unit is used to input context information and user session sequence data into the first coding layer of the variational autoencoder to obtain the initial coding latent variables.
[0138] The second coding unit is used to input the context information and the previous encoded latent variable into the second coding layer of the variational autoencoder to obtain the next encoded latent variable.
[0139] The feature extraction unit is used to obtain the embedding feature vector based on the last encoded latent variables.
[0140] In some embodiments, the feature extraction unit includes a first decoding unit, a second decoding unit, and a feature generation unit.
[0141] The first decoding unit is used to input the last encoded latent variable into the first decoding layer of the variational autoencoder to obtain the initial decoded latent variable.
[0142] The second decoding unit is used to input the previous decoded latent variable into the second decoding layer of the variational autoencoder to obtain the next decoded latent variable.
[0143] The feature generation unit is used to obtain the embedding feature vector based on the last decoded latent variables.
[0144] The first encoding layer, the second encoding layer, the first decoding layer, and the second decoding layer all satisfy a Gaussian distribution of the output data under the input data.
[0145] In some embodiments, the first coding layer is used to obtain the variational lower bound of the log-likelihood function based on contextual information and user session sequence data.
[0146] The second coding layer is used to obtain the variational lower bound of the log-likelihood function based on contextual information and the previously encoded latent variables.
[0147] The variational lower bound is used to determine the accuracy of the encoded latent variables.
[0148] In some embodiments, the product feature processing device 700 further includes a matrix generation module and a product set generation module.
[0149] The matrix generation module is used to obtain a product ID similarity matrix based on the embedding feature vector after inputting context information and user session sequence data into the variational autoencoder to obtain the embedding feature vector that represents the product features.
[0150] The product set generation module is used to obtain candidate product sets based on the product ID similarity matrix.
[0151] In some embodiments, the product feature processing device 700 further includes a product sorting module and a push module.
[0152] The product sorting module is used to sort the candidate product set according to the target sorting algorithm after obtaining the candidate product set based on the product ID similarity matrix, so as to obtain the product sorting result.
[0153] The push module is used to push product sorting results to target users.
[0154] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a product feature processing method, which includes:
[0155] Step 110: Obtain the context information corresponding to the product interface viewed by the user, as well as the user session sequence data;
[0156] Step 120: Input the context information and user session sequence data into the variational autoencoder to obtain the embedding feature vector used to represent the product features;
[0157] The context information includes the browsing scenario of the product page and spatiotemporal information; the embedding feature vector is used to generate a candidate product set.
[0158] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0159] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the commodity feature processing method provided by the above methods, the method comprising:
[0160] Step 110: Obtain the context information corresponding to the product interface viewed by the user, as well as the user session sequence data;
[0161] Step 120: Input the context information and user session sequence data into the variational autoencoder to obtain the embedding feature vector used to represent the product features;
[0162] The context information includes the browsing scenario of the product page and spatiotemporal information; the embedding feature vector is used to generate a candidate product set.
[0163] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the aforementioned product feature processing methods, the method comprising:
[0164] Step 110: Obtain the context information corresponding to the product interface viewed by the user, as well as the user session sequence data;
[0165] Step 120: Input the context information and user session sequence data into the variational autoencoder to obtain the embedding feature vector used to represent the product features;
[0166] The context information includes the browsing scenario of the product page and spatiotemporal information; the embedding feature vector is used to generate a candidate product set.
[0167] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0168] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A commodity feature processing method characterized by comprising: include: Obtain the context information corresponding to the product page viewed by the user, as well as the user's session sequence data; The context information and the user session sequence data are input into the variational autoencoder to obtain an embedding feature vector used to characterize the product features; The context information includes the scenario of browsing a product page; the embedding feature vector is used to generate a candidate product set. The step of inputting the context information and the user session sequence data into a variational autoencoder to obtain an embedding feature vector for characterizing product features includes: The context information and the user session sequence data are input into the first coding layer of the variational autoencoder to obtain the initial coding latent variables; The context information and the previous encoded latent variable are input into the second coding layer of the variational autoencoder to obtain the next encoded latent variable; The embedding feature vector is obtained based on the last encoded latent variable.
2. The commodity feature processing method according to claim 1, characterized in that, The process of obtaining the embedding feature vector based on the last encoded latent variable includes: The final encoded latent variable is input into the first decoding layer of the variational autoencoder to obtain the initial decoded latent variable; The previous decoded latent variable is input into the second decoding layer of the variational autoencoder to obtain the next decoded latent variable; Based on the final decoded latent variables, the embedding feature vector is obtained.
3. The commodity feature processing method according to claim 2, characterized in that, The initial encoded latent variable satisfies a Gaussian distribution under the context information and the user session sequence data; The latter encoded latent variable satisfies a Gaussian distribution under the context information and the former encoded latent variable; The initial decoded latent variable satisfies a Gaussian distribution under the context information and the final encoded latent variable; The latter decoded latent variable satisfies a Gaussian distribution under the context information and the former decoded latent variable.
4. The commodity feature processing method according to claim 1, characterized in that, The first encoding layer is used to obtain the variational lower bound of the log-likelihood function based on the context information and the user session sequence data; The second encoding layer is used to obtain the variational lower bound of the log-likelihood function based on the context information and the previously encoded latent variables; The variational lower bound is used to determine the accuracy of the encoded latent variables.
5. The commodity feature processing method according to any one of claims 1-4, characterized in that, After inputting the context information and the user session sequence data into the variational autoencoder to obtain the embedding feature vector used to represent product features, the method further includes: Based on the embedding feature vector, the product ID similarity matrix is obtained; Based on the product ID similarity matrix, a candidate product set is obtained.
6. The commodity feature processing method according to claim 5, characterized in that, After obtaining the candidate product set based on the product ID similarity matrix, the process further includes: The candidate product set is sorted according to the target sorting algorithm to obtain the product sorting result; The product sorting results are pushed to the target user.
7. A product feature processing device, characterized in that, include: The data acquisition module is used to acquire the context information corresponding to the product interface viewed by the user, as well as the user session sequence data; The feature extraction module is used to input the context information and the user session sequence data into the variational autoencoder to obtain an embedding feature vector for characterizing the product features. The context information includes the scenario of browsing a product page; the embedding feature vector is used to generate a candidate product set. The step of inputting the context information and the user session sequence data into a variational autoencoder to obtain an embedding feature vector for characterizing product features includes: The context information and the user session sequence data are input into the first coding layer of the variational autoencoder to obtain the initial coding latent variables; The context information and the previous encoded latent variable are input into the second coding layer of the variational autoencoder to obtain the next encoded latent variable; The embedding feature vector is obtained based on the last encoded latent variable.
8. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the commodity feature processing method according to any one of claims 1 to 6.
9. A processor-readable storage medium, characterized in that, The processor-readable storage medium stores a computer program for causing the processor to perform the steps of the commodity feature processing method according to any one of claims 1 to 6.