Personalized commodity recommendation engine and method
By semantic coding and deep learning semantic matching of user behavior data and product data, dynamic behavior portraits and feature portraits are constructed, which solves the problem that traditional recommendation systems cannot capture the time series characteristics of user behavior, and achieves efficient and personalized product recommendations.
Patent Information
- Application Number
- CN202411947037.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-27
- Publication Date
- 2025-04-25
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The traditional personalized product recommendation system cannot effectively capture the time series characteristics of user behavior, resulting in a lack of timeliness and targeted recommendation results, affecting user satisfaction and platform conversion rate.
By obtaining the time queue of user behavior data of the target user and the product data of the first alternative product, semantic encoding is performed to mine the timing context semantic information of the user behavior data and the context semantic information of the product, construct the user timing dynamic behavior portrait and product feature portrait, and dynamic personalized recommendation is used based on deep learning.
It enhances the relevance of recommendation results, improves user experience and platform conversion rate, and realizes smarter and more in line with actual needs.
Smart Images
Figure CN119887325A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of intelligent recommendation, and more specifically, to a personalized product recommendation engine and method. Background Art
[0002] With the booming development of e-commerce, personalized product recommendation systems have become an important tool for e-commerce platforms to improve user experience and increase sales conversion rates. Traditional recommendation systems mainly rely on collaborative filtering, content-based recommendation and other methods. Although these methods can provide personalized product recommendations to a certain extent, they begin to show certain limitations when faced with increasingly complex user behavior patterns and massive product information.
[0003] In addition, traditional recommendation algorithms usually only consider users' static preferences or historical purchase records. They collect users' behavior records, such as purchase history, browsing habits, etc., and generate a relatively fixed user preference model based on these static information, while ignoring the time series characteristics of user behavior. As a result, the user portrait construction method is relatively fixed and cannot adapt well to the dynamic changes in user interests, resulting in the lack of timeliness and pertinence in the recommendation results, which in turn affects user satisfaction and the conversion rate of the platform.
[0004] In order to overcome the above defects and meet the growing market demand, a personalized product recommendation engine and method are expected to be provided to provide a more intelligent product recommendation service that meets actual needs. Summary of the invention
[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a personalized product recommendation engine and method, which obtains the time queue of the user behavior data of the target user and the product data of the first candidate product, and performs semantic encoding on them respectively to mine the temporal contextual semantic information of the user behavior data of the target user and the contextual semantic information of the first candidate product, thereby constructing a user temporal dynamic behavior portrait and a candidate product feature portrait. Then, a semantic matching algorithm based on deep learning is further used to semantically match the user dynamic behavior characteristics of the target user and the contextual semantic information of the first candidate product, and dynamic and personalized product recommendations are achieved based on the semantic matching degree between the two. Through the personalized product recommendation method, not only the relevance of the recommendation results is enhanced, but also the user experience and platform conversion rate are improved, and a smarter and more practical product recommendation service is achieved.
[0006] According to one aspect of the present application, a personalized product recommendation engine is provided, which includes:
[0007] A user behavior data acquisition module is used to obtain a time queue of user behavior data of target users from an e-commerce platform;
[0008] A user behavior data context structured encoding module, used to perform context structured encoding on the time queue of the user behavior data to obtain a sequence of semantic encoding features of the context user behavior data;
[0009] A feature dynamic propagation coding module is used to perform feature dynamic propagation coding on the sequence of semantic coding features of the contextual user behavior data based on the private feature writing of the time series node to obtain dynamic coding features of the user portrait;
[0010] A first candidate product data acquisition module, used to acquire product data of the first candidate product;
[0011] A first candidate product information semantic encoding module, configured to input the product data of the first candidate product into a semantic encoder based on a converter structure to obtain semantic encoding features of the first candidate product information;
[0012] The recommendation result determination module is used to determine whether to recommend the shopping link of the first candidate product to the target user based on the semantic matching degree between the semantic coding feature of the first candidate product information and the dynamic coding feature of the user portrait.
[0013] According to another aspect of the present application, a personalized product recommendation method is provided, which includes:
[0014] Obtain the time queue of user behavior data of target users from the e-commerce platform;
[0015] Performing contextual structured encoding on the time queue of the user behavior data to obtain a sequence of semantic encoding features of the contextual user behavior data;
[0016] Performing feature propagation optimization based on user behavior dynamic guidance on the sequence of semantic coding features of the contextual user behavior data to obtain dynamic coding features of user portraits;
[0017] Obtain product data of the first candidate product;
[0018] Inputting the commodity data of the first candidate commodity into a semantic encoder based on a converter structure to obtain semantic encoding features of the first candidate commodity information;
[0019] Based on the semantic matching degree between the semantic coding feature of the first candidate product information and the dynamic coding feature of the user portrait, it is determined whether to recommend the shopping link of the first candidate product to the target user.
[0020] Compared with the prior art, the present application provides a personalized product recommendation engine and method, which obtains the time queue of the target user's user behavior data and the product data of the first candidate product, and performs semantic encoding on them respectively to mine the temporal contextual semantic information of the target user's user behavior data and the contextual semantic information of the first candidate product, thereby constructing a user temporal dynamic behavior portrait and a candidate product feature portrait. Then, a semantic matching algorithm based on deep learning is further used to semantically match the user dynamic behavior characteristics of the target user and the contextual semantic information of the first candidate product, and dynamic and personalized product recommendations are achieved based on the semantic matching degree between the two. Through the personalized product recommendation method, not only the relevance of the recommendation results is enhanced, but also the user experience and platform conversion rate are improved, and a more intelligent and practical product recommendation service is achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.
[0022] Figure 1 A block diagram of a personalized product recommendation engine according to an embodiment of the present application;
[0023] Figure 2 A data flow diagram of a personalized product recommendation engine according to an embodiment of the present application;
[0024] Figure 3 It is a block diagram of a feature dynamic propagation encoding module in a personalized product recommendation engine according to an embodiment of the present application.
[0025] Figure 4 The present invention is a flowchart of a personalized product recommendation method according to an embodiment of the present application. DETAILED DESCRIPTION
[0026] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0027] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.
[0028] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.
[0029] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.
[0030] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.
[0031] Traditional recommendation algorithms usually only consider users' static preferences or historical purchase records, but ignore the time series characteristics of user behavior, which makes it impossible to capture the interest migration that changes over time, resulting in the lack of timeliness and pertinence in recommendation results, which in turn affects user satisfaction and platform conversion rate. For example, a user may have frequently browsed sports equipment in the past month because he is about to participate in a marathon, and this short-term interest may quickly weaken or change after the event.
[0032] In the technical solution of the present application, a personalized product recommendation engine is proposed. Figure 1 A block diagram of a personalized product recommendation engine according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of a personalized product recommendation engine according to an embodiment of the present application. Figure 1 and Figure 2As shown, according to the embodiment of the present application, the personalized product recommendation engine 300 includes: a user behavior data acquisition module 310, which is used to obtain a time queue of user behavior data of a target user from an e-commerce platform; a user behavior data context structured encoding module 320, which is used to perform context structured encoding on the time queue of the user behavior data to obtain a sequence of context user behavior data semantic encoding features; a feature dynamic propagation encoding module 330, which is used to perform feature dynamic propagation encoding based on the writing of private features of time series nodes on the sequence of context user behavior data semantic encoding features to obtain user portrait dynamic encoding features; a first alternative product data acquisition module 340, which is used to obtain product data of a first alternative product; a first alternative product information semantic encoding module 350, which is used to input the product data of the first alternative product into a semantic encoder based on a converter structure to obtain the first alternative product information semantic encoding features; a recommendation result determination module 360, which is used to determine whether to recommend the shopping link of the first alternative product to the target user based on the semantic matching degree between the first alternative product information semantic encoding features and the user portrait dynamic encoding features.
[0033] In particular, the user behavior data acquisition module 310 is used to obtain the time queue of the user behavior data of the target user from the e-commerce platform. It should be understood that the user behavior data of the target user has an implicit correlation characteristic in time. When the user performs operations such as searching, clicking, adding to a shopping cart, or completing a purchase, it actually reveals the user's interest migration or sudden demand during that time period. Therefore, in the technical solution of the present application, the time queue of the user behavior data of the target user is obtained from the e-commerce platform to ensure that each new behavior can be immediately included in the analysis scope, thereby realizing the dynamic update of the user portrait, taking into account both the user's long-term preferences and the changes in short-term interests, thereby providing users with more demand-oriented product recommendations.
[0034] In one example, user behaviors include but are not limited to browsing product detail pages, adding to shopping carts, completing purchases, searching for keywords, adding products or stores to favorites, clicking on advertisements, and evaluating products. Each behavior event is accompanied by some metadata, such as user ID, product ID, behavior type, occurrence timestamp, device information, etc. The e-commerce platform can capture these user behavior events in real time through JavaScript code on the front-end page, SDK (software development kit) in mobile applications, and server-side logging. Ensuring that each behavior event has an accurate timestamp is the basis for building a user behavior data time queue. The timestamp is used to record the exact time when the event occurs, which can be the time when the server receives the request or the time when the client generates it. Based on the timestamp, the data processing layer will sort different behavior events of the same user in chronological order to form an ordered behavior sequence, namely the "time queue of user behavior data". However, in actual operation, the original behavior data may contain noise, outliers, or incomplete records. Therefore, before building the time queue, data cleaning is usually required to remove invalid or duplicate data points. At the same time, for the convenience of subsequent analysis, some fields may also be standardized, such as unifying the time format of different sources. In addition, user behavior is a dynamic process, so the time queue also needs to be updated in real time. Whenever a new user behavior occurs, the behavior will be added to the end of the corresponding user's time queue to ensure that the input of the recommendation system is always the latest user behavior snapshot.
[0035] In particular, the user behavior data context structured coding module 320 is used to perform context structured coding on the time queue of the user behavior data to obtain a sequence of context user behavior data semantic coding features. That is, in the technical solution of the present application, first, the user behavior embedding coding matrix is used to perform structured coding on each user behavior data in the time queue of the user behavior data to obtain a time queue of structured coding vectors of user behavior data. It should be understood that user behavior (such as clicks, purchases, collections, etc.) is an important way for users to interact with e-commerce platforms, which not only reflects user preferences, but also reveals users' potential needs. Therefore, in order to more accurately capture the user's dynamic behavior pattern, in the technical solution of the present application, the user behavior embedding coding matrix is used to perform structured coding on each user behavior data in the time queue of the user behavior data to obtain a time queue of structured coding vectors of user behavior data. Through structured coding, user behavior data (such as clicks, browsing, purchases, etc.) at each time point can be mapped to a vector representation in a multidimensional space. The distance between these vector representations can reflect the similarity or correlation between different user behaviors, which helps to better reveal the user's dynamic behavior association pattern. The time queue of the structured encoding vector of the user behavior data contains the multi-dimensional features of the user behavior data, which not only contains information about the changes in user interests in time series, but also the feature vector on each timestamp represents the specific behavior at that time, which lays a solid foundation for realizing personalized recommendation of commodities. In a specific example of the present application, the user behavior embedding encoding matrix is generated by Word2Vec model training. Furthermore, considering that the user's behavior not only reflects the user's immediate needs, but also implies long-term interest preferences and development trends. For example, a user's frequent browsing of sports equipment may indicate his enthusiasm for fitness, while occasional clicks on travel-related articles may imply travel plans in the near future. These behavioral characteristics are crucial to understanding the user's true intentions. However, the original behavior data is often discrete and disordered, and it is difficult to capture complex patterns and deep semantic information directly using these data. Therefore, in the technical solution of the present application, the time queue of the structured encoding vector of the user behavior data is input into the user behavior context encoder based on the converter structure to capture the dependencies between behaviors at different time points, so as to obtain a sequence of contextual user behavior data semantic encoding vectors, which helps to better understand the user's true intentions.
[0036] In particular, the feature dynamic propagation coding module 330 is used to perform feature dynamic propagation coding based on the private feature writing of the time series node on the sequence of the context user behavior data semantic coding features to obtain the dynamic coding features of the user portrait. It should be understood that since each user behavior data semantic coding feature in the sequence of the context user behavior data semantic coding features respectively contains the semantic coding features of the user behavior data at each time point, and each user behavior data semantic coding feature has a temporal correlation relationship and a temporal dynamic propagation characteristic in the time dimension, that is, the user behavior at the current time point will be affected by the semantics of the user behavior at all previous time points. Therefore, in order to be able to describe and predict the temporal propagation characteristics of the user behavior data and provide a basis for the subsequent dynamic construction of the user portrait, in the technical solution of the present application, the sequence of the context user behavior data semantic coding features is further subjected to feature dynamic propagation coding based on the private feature writing of the time series node to obtain a user portrait dynamic coding vector as a user portrait dynamic coding feature representation. In particular, in a specific example of the present application, such as Figure 3 As shown, the feature dynamic propagation coding module 330 includes: an initial feature dynamic propagation coding unit 331, which is used to perform initial feature dynamic propagation coding on the sequence of context user behavior data semantic coding vectors to obtain the user behavior data initial feature dynamic propagation coding vector; a feature dynamic propagation optimization unit 332, which is used to perform feature dynamic propagation optimization on the user behavior data initial feature dynamic propagation coding vector based on the node-private features of each context user behavior data semantic coding vector in the sequence of context user behavior data semantic coding vectors relative to the user behavior data initial feature dynamic propagation coding vector to obtain a user portrait dynamic coding vector as the user portrait dynamic coding feature.
[0037] Specifically, the initial feature dynamic propagation coding unit 331 is used to perform initial feature dynamic propagation coding on the sequence of context user behavior data semantic coding vectors to obtain the initial feature dynamic propagation coding vector of user behavior data. Specifically, the sequence of context user behavior data semantic coding vectors is input into a sequence reasoning module based on a forward LSTM model to obtain the initial feature dynamic propagation coding vector of user behavior data. That is, the forward LSTM model is used to capture the initial propagation reasoning information of user behavior data in the sequence of context user behavior data semantic coding vectors to obtain the initial feature dynamic propagation coding vector of user behavior data. More specifically, the sequence of context user behavior data semantic coding vectors is subjected to initial feature dynamic propagation coding using the following coding formula to obtain the initial feature dynamic propagation coding vector of user behavior data; wherein, the coding formula is:
[0038] X={x1,x2,…,x i , …, x n}
[0039] v i =LSTM({x1, x2, ..., x i , …, x n})
[0040] Where X represents the sequence of semantic encoding vectors of the contextual user behavior data, x1, x2, x i 、x n denote the first, second, i-th and n-th contextual user behavior data semantic encoding vectors in the sequence of contextual user behavior data semantic encoding vectors, respectively. LSTM(·) is the forward LSTM encoding. v i Represents the dynamic propagation coding vector of the initial features of the user behavior data.
[0041] Specifically, the feature dynamic propagation optimization unit 332 is used to perform feature dynamic propagation optimization on the initial feature dynamic propagation coding vector of user behavior data based on the node-private features of each contextual user behavior data semantic coding vector in the sequence of contextual user behavior data semantic coding vectors relative to the initial feature dynamic propagation coding vector of user behavior data to obtain a user portrait dynamic coding vector as the user portrait dynamic coding feature. In an embodiment of the present application, first, the node-private features of each contextual user behavior data semantic coding vector in the sequence of contextual user behavior data semantic coding vectors relative to the initial feature dynamic propagation coding vector of user behavior data are calculated to obtain a sequence of user behavior data node-private feature coding vectors. That is, the node-private features of each contextual user behavior data semantic coding vector in the sequence of the contextual user behavior data semantic coding vector relative to the dynamic propagation coding vector of the initial feature of the user behavior data are calculated respectively, and the user behavior information that is not focused on or even ignored in the propagation process of the sequence of user behavior data semantic coding vectors is captured by comparing each node feature with the dynamic propagation coding feature of the initial feature of the user behavior data, thereby generating a series of node-private feature coding vectors that emphasize the individual differences of user behavior nodes, so as to obtain the sequence of user behavior data node-private feature coding vectors. In a specific example, the specific steps of calculating the node-private features of each contextual user behavior data semantic coding vector in the sequence of contextual user behavior data semantic coding vectors relative to the user behavior data initial feature dynamic propagation coding vector include: respectively calculating the user behavior data node difference representation vectors between each contextual user behavior data semantic coding vector in the sequence of contextual user behavior data semantic coding vectors and the user behavior data initial feature dynamic propagation coding vector to obtain a sequence of user behavior data node difference representation vectors; based on the sequence of user behavior data node difference representation vectors, performing node-private feature modulation on the sequence of contextual user behavior data semantic coding vectors to obtain a sequence of user behavior data node-private feature coding vectors. More specifically, the node-private features of each contextual user behavior data semantic coding vector in the sequence of contextual user behavior data semantic coding vectors relative to the user behavior data initial feature dynamic propagation coding vector are calculated using the following node-private feature calculation formula to obtain a sequence of user behavior data node-private feature coding vectors; wherein the node-private feature calculation formula is:
[0042]
[0043] in, is positional subtraction, |·| means taking the absolute value, conv 1×1 (·) represents point convolutional coding, W dis the difference weight matrix, Sigmoid(·) is the Sigmoid function, g(x i , v i ) represents the calculation of x i and v i The node private features between i For x i The corresponding user behavior data node difference representation vector, xb i Represents the i-th user behavior data node private feature coding vector in the sequence of user behavior data node private feature coding vectors.
[0044] Next, the sequence of private feature coding vectors of the user behavior data nodes is gated and written to obtain a sequence of private feature coding vectors of the user behavior data gated write nodes. In the technical solution of the present application, the sequence of private feature coding vectors of the user behavior data nodes is input into a gated write unit based on the Softmax function to obtain a sequence of private feature coding vectors of the user behavior data gated write nodes. In this process, the Softmax function is used to convert the private feature coding vectors of the user behavior data nodes into a probability distribution, so as to give different importance scores to different user behavior features. Next, a gating mechanism is used to determine whether each private feature of the user behavior node needs to be written into the initial propagation inference information of the feature sequence based on the importance score, so as to retain the node information with the most representative and relevant user behavior features. More specifically, the sequence of private feature coding vectors of the user behavior data nodes is gated and written using the following gated write formula to obtain a sequence of private feature coding vectors of the user behavior data gated write nodes; wherein the gated write formula is:
[0045]
[0046] Y={xb1·q1,xb2·q2,…,xb i ·q i , …, xb n ·q n}
[0047] Among them, W i and b i are the modulation weight matrix and the modulation bias vector respectively, Indicates xb i The node position in the sequence of the private feature encoding vector of the user behavior data node, softmax(·) is the softmax function, a i It is xb i The corresponding probabilistic user behavior data node private feature encoding vector, τ is the predetermined threshold, mask(·) is the mask operation, xb1, xb2, xbi and xb n Represent x1, x2, x i 、x n The corresponding user behavior data node private feature encoding vector, q1, q2, q i and q n Respectively represent xb1, xb2, xb i and xb n The corresponding user behavior data node private feature gating write weight vector, Y represents the sequence of the user behavior data gating write node private feature encoding vector.
[0048] Furthermore, the sequence of the private feature coding vectors of the user behavior data gated write nodes and the dynamic propagation coding vectors of the initial features of the user behavior data are fused to obtain the dynamic coding vector of the user portrait. In order to obtain a comprehensive feature representation that reflects the dynamic interaction characteristics between user behavior feature nodes and the uniqueness of user behavior feature nodes, in the technical solution of the present application, the sequence of the private feature coding vectors of the user behavior data gated write nodes and the dynamic propagation coding vectors of the initial features of the user behavior data are fused to obtain the dynamic coding vector of the user portrait. That is, the gated private features of the user behavior are orderly and selectively written into the initial propagation inference information of the user behavior feature sequence to obtain the optimized dynamic propagation coding features of the user behavior features, so that the final dynamic coding vector of the user portrait can reflect the dynamic interaction mode between nodes and accurately capture the unique attributes of each node. In a specific example, the positional sum of each user behavior data gated write node private feature encoding vector in the sequence of the user behavior data gated write node private feature encoding vector is first calculated to obtain the user behavior data gated write node private feature global representation vector; then, the user behavior data gated write node private feature global representation vector and the user behavior data initial feature dynamic propagation encoding vector are cascaded to obtain the user portrait dynamic encoding vector. More specifically, the sequence of the user behavior data gated write node private feature encoding vector and the user behavior data initial feature dynamic propagation encoding vector are fused with the following fusion formula to obtain the user portrait dynamic encoding vector; wherein, the fusion formula is:
[0049] v f =Concat{v h ; Y}
[0050] Among them, Concat{·;·} is a cascade operation, v f Represents the user portrait dynamic encoding vector.
[0051] In particular, the first candidate product data acquisition module 340 is used to acquire product data of the first candidate product. The product data of the first candidate product includes multi-dimensional information such as detailed description, attributes, classification, price, user evaluation, etc. of the product. By acquiring the product data of the first candidate product, the system can grasp the detailed information of the product, so as to more accurately screen out the product that best meets the user's current interests based on the user portrait dynamic coding vector.
[0052] In one example, commodity data is usually stored in one or more databases, which contain basic information of commodities (such as name, description, category), price information, inventory status, sales records, etc. In addition, with the development of e-commerce, commodity data may also come from third-party service providers, supplier API interfaces, or even public data captured by web crawlers. For some large e-commerce platforms, commodity data may also come from internal management systems, such as ERP (Enterprise Resource Planning) systems. After determining the source of commodity data, the next step is to extract data. This usually requires writing specific query statements or calling APIs to obtain the required commodity data. In order to ensure the consistency and accuracy of the data, the data may be initially cleaned and verified during the extraction process, such as removing duplicate entries, repairing data in incorrect formats, filling missing values, etc. At the same time, in order to improve performance and efficiency, batch loading can also be used to extract a large amount of commodity data at one time. The key to this stage is to ensure the quality and integrity of the data and lay a solid foundation for subsequent processing.
[0053] In particular, the first candidate product information semantic encoding module 350 is used to input the product data of the first candidate product into a semantic encoder based on a converter structure to obtain the semantic encoding features of the first candidate product information. By semantically encoding the product data of the first candidate product, it is possible to identify the subtle differences between the products, rather than just the superficial feature similarities, so that the system can deeply understand the unique attributes and potential value of each product, laying the foundation for the accuracy of the semantic matching of subsequent product features and user profile features.
[0054] In the technical solution of the present application, the user portrait dynamic coding vector represents the temporal semantic aggregation coding features of the user behavior data, and the first candidate product information semantic coding vector represents the semantic coding features of the product data of the first candidate product. Therefore, in the process of inputting the first candidate product information semantic coding vector and the user portrait dynamic coding vector into the product-user semantic matching module to obtain the matching result, the inconsistent feature distribution semantic density of the first candidate product information semantic coding vector relative to the user portrait dynamic coding vector will cause the first candidate product information semantic coding vector to have a feature distribution spatial structure difference relative to the user portrait dynamic coding vector, affecting the accuracy of the matching result.
[0055] Therefore, in one example, the semantic coding vector of the first candidate product information is optimized, and the optimization process includes:
[0056] The absolute value sum and the square root of the square sum of all eigenvalues of the first candidate product information semantic encoding vector are calculated to obtain the first candidate product information semantic encoding space structure value and the second candidate product information semantic encoding space structure value, that is:
[0057] w1=∑ i |v i |
[0058]
[0059] Among them, v i represents the i-th eigenvalue of the semantic coding vector of the first candidate product information, w1 and w2 represent the first and second candidate product information semantic coding space structure values respectively;
[0060] Determine the number of eigenvalues of all eigenvalues of the first candidate product information semantic encoding vector, that is, the length L of the first candidate product information semantic encoding vector;
[0061] For each eigenvalue of the first candidate product information semantic encoding vector, the first candidate product information semantic encoding space structure value minus the product of the eigenvalue and the number of eigenvalues is calculated to obtain the first candidate product information semantic encoding long-range dependency value x i =w1-v i ×L;
[0062] Calculate the product of the square root of the number of eigenvalues multiplied by the eigenvalue minus the second first candidate product information semantic encoding spatial structure value to obtain the second first candidate product information semantic encoding long-range dependency value
[0063] The index value calculated by taking the first first candidate product information semantic encoding long-range dependency value as the exponent of the natural constant and the inverse of the second first candidate product information semantic encoding long-range dependency value are weighted summed to obtain the optimized eigenvalue corresponding to each eigenvalue
[0064] The optimized feature values are combined into an optimized semantic encoding vector of the first candidate product information.
[0065] In this way, a long-distance feature dependency relationship is established based on the spatial structure representation of the first candidate product information semantic coding vector relative to the overall feature scale of the first candidate product information semantic coding vector, so as to establish the feature local connectivity of the first candidate product information semantic coding vector, and capture the spatial ambiguous information of the object feature value through the unstructured feature value point prediction of the first candidate product information semantic coding vector, thereby improving the spatial inductive bias perception ability of the feature set of the first candidate product information semantic coding vector, thereby improving the feature distribution consistency of the first candidate product information semantic coding vector relative to the user portrait dynamic coding vector, so as to improve the analysis accuracy of the matching result.
[0066] In particular, the recommendation result determination module 360 is used to determine whether to recommend the shopping link of the first candidate product to the target user based on the semantic matching degree between the semantic encoding feature of the first candidate product information and the dynamic encoding feature of the user portrait. That is, in the technical solution of the present application, first, the semantic encoding vector of the first candidate product information and the dynamic encoding vector of the user portrait are input into the product-user semantic matching module to obtain a matching result, and the matching result is used to indicate whether the semantic matching degree between the first candidate product and the target user exceeds a preset threshold. Among them, the semantic matching module determines the degree of association between the two by calculating the similarity between the product information and the user portrait, so as to deeply explore the potential relationship between the product and the user, capture deeper points of interest and needs, and thus achieve more accurate and practical product recommendations. In one example, the similarity between the product information and the user portrait can be measured by calculating the cosine similarity between the semantic encoding vector of the first candidate product information and the dynamic encoding vector of the user portrait. Further, in response to the matching result that the semantic matching degree between the first candidate product and the target user exceeds the preset threshold, the shopping link of the first candidate product is recommended to the target user. That is, by setting a preset threshold to screen out products that are highly relevant to user interests, it is ensured that the recommendation results have high relevance and accuracy. In one example, first, by analyzing a large amount of existing user behavior data and recommendation records, the average matching degree between the products that are finally successfully converted and the users is extracted, and the initial threshold is set based on this. Then, according to the pre-set threshold, it is determined whether the calculated matching degree exceeds the threshold. If the obtained semantic matching degree exceeds the preset threshold, it is considered that there is a high correlation between the product and the user. Finally, the products in the recommendation list are pushed to the target users in an appropriate form (such as notifications, advertising banners, recommendation pages, etc.). Through personalized product recommendations, user stickiness can be enhanced and the overall service quality of the platform can be improved.
[0067] As described above, the personalized product recommendation engine 300 according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a personalized product recommendation algorithm. In a possible implementation, the personalized product recommendation engine 300 according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the personalized product recommendation engine 300 can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the personalized product recommendation engine 300 can also be one of the many hardware modules of the wireless terminal.
[0068] Alternatively, in another example, the personalized product recommendation engine 300 and the wireless terminal may also be separate devices, and the personalized product recommendation engine 300 may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.
[0069] Furthermore, a personalized product recommendation method is also provided.
[0070] Figure 4 Flowchart of a personalized product recommendation method according to an embodiment of the present application. Figure 4 As shown, according to the personalized product recommendation method of the embodiment of the present application, the steps include: S1, obtaining a time queue of user behavior data of a target user from an e-commerce platform; S2, performing context-structured encoding on the time queue of the user behavior data to obtain a sequence of semantic coding features of context user behavior data; S3, performing feature propagation optimization based on user behavior dynamic guidance on the sequence of semantic coding features of context user behavior data to obtain dynamic coding features of user portraits; S4, obtaining product data of a first alternative product; S5, inputting the product data of the first alternative product into a semantic encoder based on a converter structure to obtain semantic coding features of first alternative product information; S6, determining whether to recommend the shopping link of the first alternative product to the target user based on the semantic matching degree between the semantic coding features of the first alternative product information and the dynamic coding features of the user portrait.
[0071] In summary, according to the embodiment of the present application, the personalized product recommendation method is explained, which obtains the time queue of the user behavior data of the target user and the product data of the first candidate product, and semantically encodes them respectively to mine the temporal contextual semantic information of the user behavior data of the target user and the contextual semantic information of the first candidate product, thereby constructing a user temporal dynamic behavior portrait and a candidate product feature portrait. Then, a semantic matching algorithm based on deep learning is further used to semantically match the user dynamic behavior characteristics of the target user and the contextual semantic information of the first candidate product, and dynamic and personalized product recommendations are achieved based on the semantic matching degree between the two. Through the personalized product recommendation method, not only the relevance of the recommendation results is enhanced, but also the user experience and platform conversion rate are improved, and a smarter and more practical product recommendation service is achieved.
[0072] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.
Claims
1. A personalized product recommendation engine, characterized in that: include: A user behavior data acquisition module is used to obtain a time queue of user behavior data of target users from an e-commerce platform; A user behavior data context structured encoding module, used to perform context structured encoding on the time queue of the user behavior data to obtain a sequence of semantic encoding features of the context user behavior data; A feature dynamic propagation coding module is used to perform feature dynamic propagation coding on the sequence of semantic coding features of the contextual user behavior data based on the private feature writing of the time series node to obtain dynamic coding features of the user portrait; A first candidate product data acquisition module, used to acquire product data of the first candidate product; A first candidate product information semantic encoding module, configured to input the product data of the first candidate product into a semantic encoder based on a converter structure to obtain semantic encoding features of the first candidate product information; The recommendation result determination module is used to determine whether to recommend the shopping link of the first candidate product to the target user based on the semantic matching degree between the semantic coding feature of the first candidate product information and the dynamic coding feature of the user portrait.
2. The personalized product recommendation engine according to claim 1, characterized in that: The user behavior data context structured encoding module includes: A structured coding unit, configured to perform structured coding on each user behavior data in the time queue of the user behavior data using a user behavior embedding coding matrix to obtain a time queue of structured coding vectors of the user behavior data; A context semantic encoding unit is used to input the time queue of the structured encoding vector of the user behavior data into a user behavior context encoder based on a converter structure to obtain a sequence of context user behavior data semantic encoding vectors as a sequence of context user behavior data semantic encoding features.
3. The personalized product recommendation engine according to claim 2, characterized in that: The feature dynamic propagation encoding module comprises: An initial feature dynamic propagation coding unit, used for performing initial feature dynamic propagation coding on the sequence of context user behavior data semantic coding vectors to obtain an initial feature dynamic propagation coding vector of user behavior data; A feature dynamic propagation optimization unit is used to perform feature dynamic propagation optimization on the initial feature dynamic propagation coding vector of user behavior data based on the node-private features of each contextual user behavior data semantic coding vector in the sequence of contextual user behavior data semantic coding vectors relative to the initial feature dynamic propagation coding vector of user behavior data, so as to obtain a user portrait dynamic coding vector as the user portrait dynamic coding feature.
4. The personalized product recommendation engine according to claim 3, characterized in that: The initial feature dynamic propagation encoding unit is used to: The sequence of the contextual user behavior data semantic encoding vectors is input into a sequence reasoning module based on a forward LSTM model to obtain the initial feature dynamic propagation encoding vector of the user behavior data.
5. The personalized product recommendation engine according to claim 4, characterized in that: The feature dynamic propagation optimization unit comprises: A node private feature calculation subunit, used to calculate the node private feature of each context user behavior data semantic coding vector in the sequence of context user behavior data semantic coding vectors relative to the user behavior data initial feature dynamic propagation coding vector to obtain a sequence of user behavior data node private feature coding vectors; A gated writing subunit, used for gated writing the sequence of private feature coding vectors of the user behavior data nodes to obtain a sequence of private feature coding vectors of the gated writing nodes of the user behavior data; A fusion subunit is used to fuse the sequence of the user behavior data gated write node private feature coding vector and the user behavior data initial feature dynamic propagation coding vector to obtain the user portrait dynamic coding vector.
6. The personalized product recommendation engine according to claim 5, characterized in that: The node private feature calculation subunit is used to: Respectively calculating the user behavior data node difference representation vectors between each context user behavior data semantic encoding vector in the sequence of context user behavior data semantic encoding vectors and the user behavior data initial feature dynamic propagation encoding vector to obtain a sequence of user behavior data node difference representation vectors; Based on the sequence of the user behavior data node difference representation vectors, the sequence of the context user behavior data semantic encoding vectors is modulated by node private features to obtain the sequence of the user behavior data node private feature encoding vectors.
7. The personalized product recommendation engine according to claim 6, characterized in that: The gated writing subunit is used for: Each user behavior data node private feature coding vector in the sequence of user behavior data node private feature coding vectors is input into a gated writing unit based on a Softmax function to obtain a sequence of user behavior data gated writing node private feature coding vectors.
8. The personalized product recommendation engine according to claim 7, characterized in that: The fusion subunit is used for: Calculating the positional sum of each user behavior data gated write node private feature encoding vector in the sequence of the user behavior data gated write node private feature encoding vector to obtain a user behavior data gated write node private feature global representation vector; The user behavior data gated writing node private feature global representation vector and the user behavior data initial feature dynamic propagation coding vector are cascaded to obtain the user portrait dynamic coding vector.
9. The personalized product recommendation engine according to claim 8, characterized in that: The recommendation result determination module is used to: Inputting the semantic coding vector of the first candidate product information and the dynamic coding vector of the user portrait into a product-user semantic matching module to obtain a matching result, wherein the matching result is used to indicate whether the semantic matching degree between the first candidate product and the target user exceeds a preset threshold; In response to the matching result that the semantic matching degree between the first candidate product and the target user exceeds a preset threshold, a shopping link of the first candidate product is recommended to the target user.
10. A personalized product recommendation method, characterized in that: include: Obtain the time queue of user behavior data of target users from the e-commerce platform; Performing contextual structured encoding on the time queue of the user behavior data to obtain a sequence of semantic encoding features of the contextual user behavior data; Performing feature propagation optimization based on user behavior dynamic guidance on the sequence of semantic coding features of the contextual user behavior data to obtain dynamic coding features of user portraits; Get the product data of the first candidate product; Inputting the commodity data of the first candidate commodity into a semantic encoder based on a converter structure to obtain semantic encoding features of the first candidate commodity information; Based on the semantic matching degree between the semantic coding feature of the first candidate product information and the dynamic coding feature of the user portrait, it is determined whether to recommend the shopping link of the first candidate product to the target user.