E-commerce product recommendation method and system based on artificial intelligence algorithm

Abstract

The invention discloses an artificial intelligence algorithm-based e-commerce product recommendation method and system, and relates to the technical field of product recommendation. Comprising the following steps: constructing a user-behavior path flow tree by adopting a multi-criterion splitting mechanism, and generating a node path set; respectively slicing and recombining the user portrait set and the commodity portrait set, and carrying out cross combination; constructing a behavior cluster atlas to which the user belongs, projecting cross commodity nodes in the node path set and the candidate cross mapping set into an anti-cross atlas structure among behavior clusters, and extracting interaction points as a candidate recommendation set; and performing weight regression calculation on each node of the path of the user-behavior path flow tree by utilizing path probability density mapping corresponding to each cross commodity node in the candidate recommendation set, and generating a commodity recommendation sequence according to the maximum path passing probability. According to the method, the recommendation process with the behavior chain criterion and the coupling map mapping is constructed, so that the user experience is improved, and the conversion efficiency and the content distribution capability of the platform are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of product recommendation, and in particular to an e-commerce product recommendation method and system based on an artificial intelligence algorithm. Background Art

[0002] With the rapid development of e-commerce platforms, personalized recommendation systems have gradually become a core technical means to improve user satisfaction and purchase conversion rates. Early recommendation methods mainly relied on strategies such as collaborative filtering, content matching, or rating matrix modeling to identify items of potential interest through similarity metrics. However, these methods have significant limitations when faced with data sparsity, cold start problems, and the complexity of user behavior. In recent years, the development of artificial intelligence, especially deep learning and graph neural networks, has provided new ideas for mining user behavior trajectories and modeling behavioral evolution paths. E-commerce recommendation technology has gradually shifted from "static preference modeling" to "dynamic behavior modeling." Under this trend, integrating multi-dimensional behavioral features, strengthening the interactive modeling of user and product profiles, and building more flexible and explainable recommendation frameworks have become important directions in the development of intelligent recommendation systems.

[0003] For example, CN118037401A is an agricultural product e-commerce recommendation system based on knowledge graphs. This system focuses on data collection and structured modeling in agricultural e-commerce scenarios. Its user portrait module is used to establish a basic interest model based on user historical behavior, while the graph construction module mainly develops knowledge modeling around agricultural product information. The system provides auxiliary recommendation functions through multiple modules such as product quality, nutritional analysis, and market monitoring. However, this technical solution mainly relies on the static representation capability of the knowledge graph and is relatively coarse-grained in modeling user behavior characteristics. It only involves user behavior aggregation and basic preference analysis, and lacks a dynamic modeling mechanism for the evolution of user behavior paths. In addition, this method does not introduce staggered cross-mapping, behavior cluster coupling, or a weighting mechanism based on path pass probability in the construction of recommendation paths between users and products. As a result, its recommendation results are difficult to fully reflect the evolution trend of user behavior. In particular, when facing multi-behavior coupling recommendation tasks, the generalization ability and timeliness are still insufficient.

[0004] CN117710064B is an e-commerce product recommendation method and device based on artificial intelligence. It generates a target value matrix through the user's purchase record, and then combines price similarity and brand similarity to perform two-stage user group matching, and generates product recommendation probabilities based on the modified model. This method improves the recommendation accuracy to a certain extent, but its algorithm core relies on purchase behavior data, ignoring the important value of user non-purchase behaviors (such as browsing, staying, jump order, etc.) in user intent identification. At the same time, its user portrait construction method is relatively simple, does not consider the behavior evolution law across time, and does not introduce path probability or graph structure to dynamically represent user behavior. In addition, this method determines the user group by static similarity, which easily ignores the coupling relationship between complex behaviors and potential temporal preferences. Summary of the Invention

[0005] In view of the problems of existing e-commerce recommendation methods in that the user behavior modeling dimension is single and there is a lack of flexible cross-mapping and behavior cluster coupling mechanism between user portraits and product portraits, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to achieve dynamic modeling and intelligent recommendation of user-product interaction relationships by constructing process paths, graph structures and cross-mapping mechanisms.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In the first aspect, the present invention provides an e-commerce product recommendation method based on an artificial intelligence algorithm, which includes extracting multi-state feature parameters based on user historical behavior data, adopting a multi-criteria splitting mechanism to construct a user-behavior path process tree, and generating a node path set with behavior evolution laws; slicing and reorganizing the user portrait set and the product portrait set under multi-dimensional features, and cross-combining them through predefined staggered mapping rules to generate a candidate cross-mapping set; constructing a user's behavior cluster graph, projecting the node path set and the cross-product nodes in the candidate cross-mapping set into an anti-interlaced graph structure between the behavior clusters, and extracting interaction points with high transfer weight and behavior coupling as a candidate recommendation set; using the path probability density mapping corresponding to each cross-product node in the candidate recommendation set, reversely searching the associated product-path probability mapping table, performing weight regression calculation on each node of the user-behavior path process tree path, and generating a product recommendation sequence based on the maximum path pass probability.

[0009] As a preferred solution of the e-commerce product recommendation method based on artificial intelligence algorithm described in the present invention, the construction of the user-behavior path process tree includes: based on the user behavior log data recorded in the e-commerce platform, extracting the page residence time parameters, browsing frequency parameters and jump sequence codes of each behavior event, and constructing a multidimensional state vector sequence according to the time series; according to the difference change law of the parameters of each dimension in the multidimensional state vector sequence, setting a transfer threshold set, dividing the continuous state sequence into discrete state segments according to the threshold, and marking the corresponding state label set; converting the state label set into a state path chain, performing frequency statistics and co-occurrence relationship mining on the node sequence therein, and constructing a state criterion splitting condition group containing the behavior evolution relationship; based on the state label sequence and the state criterion splitting condition group, splitting the root node in sequence to generate a path process tree structure.

[0010] As a preferred solution of the e-commerce product recommendation method based on artificial intelligence algorithm described in the present invention, the slicing and reorganization includes: extracting the main feature dimensions based on the user portrait set and the product portrait set, and constructing a standardized feature matrix, wherein the standardized feature matrix includes the user feature matrix and the product feature matrix; calculating the local change rate vector on each feature dimension and performing stability check after slicing at the fluctuation point, dynamically slicing the user feature matrix and the product feature matrix respectively, and generating a feature sub-matrix set.

[0011] As a preferred solution of the e-commerce product recommendation method based on artificial intelligence algorithm described in the present invention, the generation of the candidate cross-mapping set includes: setting an offset vector sequence according to the asymmetric distribution characteristics of product portraits in the price and popularity dimensions, each offset controls the cross-combination path, and generates a staggered mapping rule group; performing a combination operation on the sliced ​​user and product sub-matrix pairs according to the staggered mapping rule group to form a candidate cross-mapping set.

[0012] As a preferred solution of the e-commerce product recommendation method based on artificial intelligence algorithm described in the present invention, when the cross-product node is projected into the deinterlaced graph structure, the associated original product ID mapping table is synchronously recorded to generate an associated product-path probability mapping table; based on the node path set, the jump frequency and time interval between user behavior nodes are counted to construct a weighted directed graph structure; based on the candidate cross-mapping set, a cross-product node set is defined, and the user behavior nodes involved therein are projected into the weighted directed graph structure to construct an index matrix consisting of behavior nodes and cross-product nodes.

[0013] As a preferred solution of the e-commerce product recommendation method based on artificial intelligence algorithm described in the present invention, the generation of the candidate recommendation set includes: for each mapping unit in the index matrix, based on the set of shortest paths between the connected behavior nodes and other nodes in the graph, calculating the coupling weight vector between the behavior paths; based on the coupling weight vector and the edge weight in the original behavior transfer graph, jointly setting the coupling degree threshold and the transfer weight threshold, screening the cross-product node pairs that meet the conditions, and forming a candidate recommendation set.

[0014] As a preferred solution of the e-commerce product recommendation method based on artificial intelligence algorithm described in the present invention, the mapping of the path probability density includes: based on the transfer frequency of each node in the user-behavior path process tree in the user behavior sequence, counting the number of historical behaviors jumping from the previous node to the current node, and smoothing all historical jump behaviors according to the distribution on the timeline; defining dynamic path entry detection rules based on the response time distribution of commodity nodes in user behavior; corresponding the above two parts in node number and path sequence to form a three-dimensional distribution structure with path node-frequency density-response time as the joint index, forming a density mapping for behavior path sorting.

[0015] In a second aspect, the present invention provides an e-commerce product recommendation system based on an artificial intelligence algorithm, which includes:

[0016] The multi-state feature extraction module extracts multi-state feature parameters based on user historical behavior data, uses a multi-criteria splitting mechanism to construct a user-behavior path process tree, and generates a node path set with behavior evolution rules;

[0017] The portrait slicing and mapping module slices and reorganizes the user portrait set and the product portrait set under multi-dimensional features, and generates a candidate cross-mapping set through predefined dislocation mapping rules;

[0018] The deinterlaced graph module constructs a graph of user behavior clusters, projects the node path set and the cross-product nodes in the candidate cross-mapping set into the deinterlaced graph structure between behavior clusters, and extracts interaction points with high transfer weight and behavior coupling as candidate recommendation sets;

[0019] The path probability regression module uses the path probability density mapping corresponding to each cross-product node in the candidate recommendation set to reversely search the associated product-path probability mapping table, perform weighted regression calculations on each node of the user-behavior path process tree, and generate a product recommendation sequence based on the maximum path pass probability.

[0020] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the e-commerce product recommendation method based on artificial intelligence algorithm as described in the first aspect of the present invention are implemented.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the e-commerce product recommendation method based on artificial intelligence algorithm as described in the first aspect of the present invention are implemented.

[0022] The beneficial effects of this invention are as follows: By introducing multi-state behavior modeling and a path process tree mechanism, it can comprehensively capture the evolving characteristics of user behavior on e-commerce platforms and effectively identify the dynamic trends of user interests. Through portrait slicing and staggered mapping strategies, it fully explores the potential matching relationships between users and products, improving the accuracy and diversity of recommendations. Combining the behavior graph structure with path probability density mapping, it can achieve quantitative evaluation and weighted regression of recommendation paths, significantly improving the relevance and click-through rate of recommendation sequences.

[0023] Compared with traditional recommendation algorithms that only rely on static features or collaborative filtering, the present invention achieves deep integration at the structural modeling and temporal response levels, and has stronger user preference recognition capabilities and real-time recommendation adaptability.

[0024] In summary, the present invention realizes the intelligent upgrade of personalized recommendation by constructing a recommendation process with behavior chain criteria and coupling graph mapping, which not only improves the user experience, but also enhances the platform's conversion efficiency and content distribution capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 A flowchart of an e-commerce product recommendation method based on artificial intelligence algorithms;

[0027] Figure 2 This is a structural diagram of the e-commerce product recommendation system based on artificial intelligence algorithms. DETAILED DESCRIPTION

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0030] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0031] As mentioned in the background technology above, early recommendation methods mainly relied on strategies such as collaborative filtering, content matching, or rating matrix modeling to identify potentially interesting products through similarity metrics. However, these methods have significant limitations when faced with data sparsity, cold start problems, and the complexity of user behavior. In recent years, the development of artificial intelligence, especially deep learning and graph neural networks, has provided new ideas for mining user behavior trajectories and modeling behavior evolution paths. E-commerce recommendation technology has gradually shifted from "static preference modeling" to "dynamic behavior modeling." Under this trend, integrating multi-dimensional behavioral features, strengthening the interactive modeling of user and product portraits, and building a more flexible and explainable recommendation framework have become important directions for the development of current intelligent recommendation systems.

[0032] Figure 1 Flowchart of the e-commerce product recommendation method based on artificial intelligence algorithm according to an embodiment of the present invention. Figure 1 As shown in the figure, the e-commerce product recommendation method based on artificial intelligence algorithm includes:

[0033] S1: Extract multi-state feature parameters based on user historical behavior data, adopt a multi-criteria splitting mechanism to construct a user-behavior path process tree, and generate a node path set with behavior evolution rules.

[0034] In an embodiment of the present invention, generating a node path set includes the following steps:

[0035] a. Based on the pre-recorded user behavior log data on the e-commerce platform, we extract the page dwell time parameters, browsing frequency parameters, and jump sequence codes of each behavior event, and construct a multidimensional state vector sequence according to the time series to represent the behavior evolution trajectory.

[0036] Specifically, each interaction a user makes while browsing products is converted into a composite vector containing multiple characteristic states. The dimensions represent browsing frequency (i.e., the number of times a user visits a product or page per unit time), page dwell time (i.e., the average time a user stays on a product page), and jump sequence code (i.e., the sequence number of paths a user takes before and after entering the current page). This process converts the original discrete behavior log into a multidimensional sequence of state vectors. Each vector sequence is arranged continuously along the timeline, thus reflecting the dynamic trajectory of the user's behavior throughout the e-commerce platform.

[0037] b. According to the different variation patterns of the parameters of each dimension in the multidimensional state vector sequence, a set of transfer thresholds is set. The continuous state sequence is divided into discrete state segments based on the thresholds, and the corresponding state label sets are marked.

[0038] In specific operations, based on the changing trends of various parameters in the multidimensional state vector, the differential features are calculated and compared with the preset transfer threshold set to determine whether the behavioral state has changed significantly.

[0039] Among them, the transfer threshold set is obtained based on the global behavior sample statistics of e-commerce platform users, which includes critical criteria for frequency mutations, significant fluctuations in stay duration, jump path changes, etc.

[0040] Once the magnitude of one or more dimensions in a continuous state vector exceeds the corresponding threshold, a behavioral state transition is considered to have occurred. Based on this, the continuous vector sequence is divided into multiple discrete state segments. Each state tag is associated with a split condition identifier, which records the threshold type that triggered the state transition and provides a matching condition for subsequent process tree splitting.

[0041] c. Convert the state label set into a state path chain, perform frequency statistics and co-occurrence relationship mining on the node sequence, and construct a state criterion splitting condition group that includes the behavior evolution relationship.

[0042] For example, first, the state labels are connected in sequence according to the time sequence of user behavior to form a path chain, and the frequency of occurrence of each label combination is counted to establish a frequency matrix to characterize the evolution probability between different states.

[0043] Meanwhile, to further enhance the ability of the path structure to express behavior patterns and to discover state pairs and path sub-segments that frequently co-occur among user groups, including: for paths with a length ≥ I steps (where I is set according to the actual situation), even if the occurrence frequency is lower than the average value, if the end node of the path is a purchase behavior, it is marked as a key conversion path; maintain an independent storage pool for long paths, and when the conventional condition group cannot cover them, extract the top 10% paths from the pool according to the conversion rate ranking and add them to the splitting condition group; accordingly, extract the state splitting condition group for process tree construction.

[0044] Adopt double-layer condition screening. Among them, the main condition group is: the transfer frequency of the transfer pair is greater than the average transfer frequency plus 3 times the standard deviation; the auxiliary condition group is: after DBSCAN clustering, select the cluster center with a silhouette coefficient > H as the structural basis for node splitting in the process tree; where H is the set silhouette coefficient threshold, which can be set according to the actual situation.

[0045] The above operations not only avoid the problem of path fragmentation caused by random behaviors, but also enhance the sensitivity and expression ability of behavior modeling to group laws.

[0046] Based on the state label sequence and the state criterion splitting condition group, split the path process tree structure is generated sequentially from the root node, and each terminal node path is classified into a set of user behavior paths, where each path represents a type of behavior evolution path.

[0047] In the process tree construction link, the present invention adopts an iterative splitting strategy based on the state label sequence and the state splitting condition group, and generates a complete user behavior path process tree structure from top to bottom starting from the root node.

[0048] Specifically, with all state path chains as the initial input, perform the first state splitting operation on the root node, and judge whether the sub-path division conditions are met according to the aforementioned state splitting conditions; if so, generate two or more sub-nodes, and each sub-node represents a state evolution direction; thereafter, repeat the above splitting process for each sub-node respectively, and the termination conditions are satisfied by any of the following: the current node depth ≥ I, the number of paths contained in the node < max(0.1%, co-occurrence frequency threshold of the same type of users / 10), and no gain is generated after 3 consecutive splits, and the construction ends.

[0049] In the final process tree, each terminal node represents an independent user behavior evolution path, and they are uniformly classified into a set of user behavior paths.

[0050] It's worth emphasizing that, compared to traditional user clustering and behavioral induction recommendation modeling approaches, the proposed path process tree construction method can more accurately depict the behavioral evolution trajectory of individual users on e-commerce platforms. It is particularly well-suited for nonlinear behavioral paths such as shallow visits, deep stays, and skip browsing, followed by target clicks. By introducing multidimensional state features and split condition groups, this method not only improves the discriminability of path classification but also enhances the adaptability of behavioral modeling to temporal heterogeneity and multimodal features, thereby constructing more targeted and dynamically evolving user behavior profiles.

[0051] S2: Slice and reorganize the user portrait set and the product portrait set under multi-dimensional features, and cross-combine them through predefined dislocation mapping rules to generate a candidate cross-mapping set.

[0052] S2.1: Based on the user portrait set U and the product portrait set G, extract the main feature dimensions and construct a standardized feature matrix, where the standardized feature matrix includes the user feature matrix and the product feature matrix.

[0053] In the technical solution of the present invention, the selection of main feature dimensions includes but is not limited to interest categories, consumption levels, active cycles (for user portraits) as well as price ranges, product popularity, category labels (for product portraits), etc.

[0054] In the actual construction process, the Z-score normalization method is used to normalize the original features, unify the dimensions and eliminate the feature offset problem caused by inconsistent dimensions. After the construction is completed, two standardized feature matrices are formed: user feature matrix and product feature matrix Here, m and n represent the number of users and items, respectively, and d represents the uniform number of feature dimensions. This standardized approach ensures comparability and mathematical consistency in subsequent slicing and offset mapping operations.

[0055] S2.2: Calculate the local change rate vector on each feature dimension and perform stability check after slicing the fluctuation point. Dynamically slice the user feature matrix and the product feature matrix respectively to generate a set of feature sub-matrices.

[0056] Among them, the local change rate vector is obtained by performing sliding window difference on the eigenvalue in the corresponding dimension, and the local fluctuation point is calibrated by the peak value of the absolute value of the first-order derivative.

[0057] In the specific operation of the present invention, stability checks are performed after slicing the fluctuation points. These checks include merging slices if the Jaccard similarity between adjacent slices exceeds a similarity threshold, and ensuring that the offset mapping offset is ≤ the length of the merged slice. This process yields a set of user and product sub-feature matrices. This improves the ability to characterize implicit local connections between users and products, avoiding the problem of detail obscuration caused by overall vector operations.

[0058] S2.3: Based on the asymmetric distribution characteristics of product portraits in the price and popularity dimensions, set the offset vector sequence. Each offset controls the cross-combination path to generate a set of dislocation mapping rules.

[0059] It should be noted that each offset δ i Indicates the moving step of the product sub-matrix along the feature dimension axis when the user sub-matrix is ​​cross-combined with the product sub-matrix. For example, when δ i = +1, it means that the user sub-matrix U j and the product submatrix G j+1 Combinations are made instead of corresponding combinations.

[0060] The theoretical basis of this offset combination design is that there is a time dislocation or trend deviation between user interest hotspots and product popularity high points in the feature space. By constructing an offset relationship, potential cross-interest matching units can be effectively identified, thereby expanding the potential interaction space between users and products.

[0061] Then, according to the offset vector sequence set above, a set of offset mapping rules R is generated, where each rule r i By quintuple (U j , δ i , mapping type, path fingerprint), where path fingerprint = Hash (user sub-matrix features + product sub-matrix features).

[0062] The combination method of user sub-matrix and product sub-matrix and their offset relationship are clearly stipulated.

[0063] Guided by this rule group, all combinable paths are traversed in the entire quantum space to achieve multi-path, multi-step, and multi-dimensional cross-combination operations, thereby constructing a set of cross-matching units with differences and structural jump characteristics.

[0064] S2.4: Perform a combination operation on the sliced ​​user and item sub-matrices according to the staggered mapping rule group to form a candidate cross-mapping set.

[0065] Under the guidance of the above-mentioned staggered mapping rule group, the user sub-matrix and the product sub-matrix are combined pair by pair to form a candidate cross-mapping set C, where each candidate mapping C i For one During the combination process, the two matrices are fused into a unified expression structure by using feature splicing and linkage encoding, while their offset labels and path metadata are recorded.

[0066] S3: Construct a user behavior cluster graph, project the node path set and the cross-product nodes in the candidate cross-mapping set into the anti-interlaced graph structure between the behavior clusters, and extract interaction points with high transfer weight and behavior coupling as the candidate recommendation set.

[0067] When projecting the cross-product node into the deinterlaced graph structure, the associated original product ID mapping table is recorded synchronously to generate an associated product-path probability mapping table to ensure a one-to-many association between the cross node and the product entity.

[0068] It should be noted that the cross-product node is a composite node generated by the user feature sub-matrix and the product feature sub-matrix through the staggered mapping rule; the product node is an original entity node containing only the product ID and inherent attributes.

[0069] First, based on the node path set, the jump frequency and time interval between user behavior nodes are counted to construct a weighted directed graph structure, in which the edge weight is defined by the transfer frequency vector and the time delay factor.

[0070] Specifically, based on the state sequence and path frequency data extracted from the node path set, a weighted directed graph structure is constructed to represent the jump relationship between user behavior nodes.

[0071] During graph construction, each state label with significant behavioral characteristics is treated as a node element in the graph. A directed edge is constructed for each pair of states with a continuous jump relationship, reflecting the migration direction of the behavioral path. Edge weights are defined using a dual weighting element consisting of a transition frequency vector and a time delay factor.

[0072] Among them, the transfer frequency vector is used to quantify the relative frequency of the behavior jump in the sample users, reflecting the dominance of the behavior transfer; the time delay factor measures the closeness and response rate of the jump by calculating the average dwell time between adjacent state labels. The combination of the two through weighted calculation can make the graph not only express the structural relationship of the behavior, but also reflect the execution rhythm and response rhythm of the user behavior, which is helpful to identify the main transfer flow in the behavior path. In traditional methods, only the transfer frequency vector or the time delay factor is used for path construction, which cannot accurately characterize the actual importance difference between high-frequency but slow behavior and low-frequency but fast behavior. The present invention forms a weighted structure that is more sensitive to actual path selection through fusion modeling.

[0073] Secondly, a cross-product node set is defined based on the candidate cross-mapping set, and the user behavior nodes involved are projected into a weighted directed graph structure. An index matrix consisting of behavior nodes and cross-product nodes is constructed. This matrix is ​​used to represent the anti-interlaced position relationship of the product mapping points in the behavior graph.

[0074] During the mapping process, we first search whether the behavior node already exists in the currently constructed graph structure based on the unique identifier of each cross-product or behavior node in the candidate cross-product node set; if it does, a corresponding relationship is directly established; if it does not exist, it is inserted into the graph structure according to the new behavior node insertion judgment rule to establish an edge connection between it and the high-frequency jump node, and the initial frequency and time factor of the connecting edge are calculated to improve the boundary information of the graph structure.

[0075] Among them, the judgment rules for inserting new behavior nodes include: there is at least one indirect association (within two hops) in the current user's historical behavior; the co-occurrence frequency of similar users is greater than 3 times the average level of the group standard deviation. If the end of the path is a purchase behavior, even if the co-occurrence frequency is low, the weight is increased by half, and the S1 termination condition is exempted (the path terminated by S1 is temporarily stored. If S3 triggers the insertion rule and there is a matching path in the buffer, it will be reused first); the initial value of the new edge weight = min (the current graph edge weight average, 0.1).

[0076] Subsequently, an index matrix between behavior nodes and cross-product nodes is constructed. This matrix expresses the anti-interlaced mapping relationship between nodes in the behavior graph in a two-dimensional form, that is, the behavior anchor points corresponding to each cross-product node in different path sequences, their occurrence frequency, and structural position.

[0077] Finally, for each mapping unit in the index matrix, the coupling weight vector between the behavior paths is calculated based on the set of shortest paths connecting the behavior node with other behavior nodes in the graph. The weight is obtained by weighting the path length, jump frequency, and common product density. According to the coupling weight vector and the edge weight in the original behavior transfer graph, the coupling degree threshold and transfer weight threshold are jointly set to screen the cross-product node pairs that meet the conditions to form a candidate recommendation set, and its path sequence and coupling background parameters are recorded.

[0078] In the embodiment of the present invention, the coupling analysis is based on a set of all shortest paths connecting cross-product nodes and original behavior nodes in the graph, and a coupling weight vector between behavior paths is comprehensively calculated.

[0079] This weight vector incorporates three core factors into its weighting mechanism: First, path length. Shorter paths generally indicate a higher likelihood of direct behavioral transitions, thus assigning a higher coupling tendency. Second, jump frequency, representing the number of times the shortest path has actually been triggered in historical user behavior, reflects the path's reliability. Third, common product density, representing the degree of overlap in products or pages visited by behavioral nodes on the path within a user group; a higher density indicates stronger behavioral coupling. When constructing this weighted coupling model, weight parameters are set for different factors, ensuring that the sum of these weight parameters is 1. This allows for flexible control over the contribution of different factors to the final coupling weight in implementation.

[0080] After the coupling weight vector is calculated, the transfer weight information of the edges in the original graph is combined with a dual-threshold judgment mechanism to screen cross-product node pairs and construct a recommendation candidate set.

[0081] This mechanism sets a coupling threshold and a transfer weight threshold, and periodically calculates the correlation coefficient between the coupling and transfer weights. If a strong positive correlation is found, the coupling threshold is lowered, for example, to 20%, and the transfer weight threshold is raised, for example, to 15%. These values ​​are for reference only and should be adjusted in practice. If a negative correlation is found, the transfer weight threshold is frozen, and only the coupling threshold is used for screening.

[0082] This joint screening structure can effectively avoid the bias problem of a single indicator dominating the recommendation results, and improve the behavioral correlation rationality and structural stability of the recommendation nodes.

[0083] Through the above steps, the present invention achieves in-depth mining of potential recommendation nodes within user behavior paths. In particular, in terms of graph construction and cross-item node linkage, by combining temporal weighting with structural co-occurrence, recommendation results take into account both the true semantics of behavior and the stability of group behavior.

[0084] S4: Using the path probability density mapping corresponding to each cross-product node in the candidate recommendation set, reversely search the associated product-path probability mapping table, perform weighted regression calculation on each node of the user-behavior path process tree, and generate a product recommendation sequence based on the maximum path pass probability.

[0085] First, we use path probability density mapping rather than a simple node scoring mechanism because user behavior within complex process paths exhibits significant temporal, dynamic, and contextual dependencies. A single scoring model cannot effectively reflect the transition patterns and temporal response characteristics between nodes. Therefore, this paper proposes a probability density mapping based on the behavioral path process tree, aiming to quantify the fluctuations in user behavior probability along a specific path and its impact on recommendation effectiveness.

[0086] S4.1: For each cross-product node in the candidate recommendation set, based on the historical travel frequency of the behavior node in the associated path set within the process tree path, analyze its associated original product set and construct a path probability density map. This path probability density map is obtained by jointly fitting the node transition frequency kernel density estimate and the product behavior response time distribution. The specific steps are as follows:

[0087] 1. Based on the transfer frequency of each node in the user behavior sequence in the user-behavior path process tree, the number of historical behaviors jumping from the previous node to the current node is counted, and all historical jump behaviors are smoothed according to their distribution on the timeline.

[0088] It should be noted that the present invention introduces time axis density estimation to express the activity differences of jump behaviors in different time periods, thereby improving the ability to sensitively perceive changes in behavior frequency.

[0089] In order to achieve the continuous expression of the above-mentioned node transfer relationship, the present invention adopts a kernel density estimation method to smooth the node transfer frequency.

[0090] 2. Define dynamic path entry detection rules based on the response time distribution of product nodes in user behavior.

[0091] Specifically, when the number of entry paths to a node (i.e., the number of arrows from other nodes pointing to it) exceeds twice the number of exit paths (i.e., the number of arrows from it pointing to other nodes), it is automatically marked as a candidate entry node; the response time calculation starts from the most recently passed candidate entry node; if there are multiple candidate entry nodes in the path, the entry closest to the current node is selected.

[0092] 3. The above two parts are matched in terms of node number and path sequence to form a three-dimensional distribution structure with "path node-frequency density-response time" as the joint index. Specifically, it is path node-frequency density-response time, where the path node is the numbered node in the process tree, the frequency density reflects its transfer activity, and the response time dimension reflects the time characteristics of the behavior reach, forming a density map that can be used for behavior path sorting.

[0093] S4.2: Taking each node on each path as input, the node attribution weight vector is calculated through weighted regression based on its matching degree in the probability density mapping of different paths. This vector reflects the contribution of each node in the corresponding product path of the candidate recommendation set.

[0094] Compared with the traditional simple weighting method based on click rate, dwell time or TF-IDF calculation, the present invention can capture the coupling relationship between structural context and behavioral dynamics through density matching of nodes within the path, and has a higher behavioral interpretation ability.

[0095] Specifically, the present invention sequentially inputs the joint feature value of each node in each path in the path probability density map, that is, the kernel density frequency value and response time feature vector of the current node in a certain product path, and calculates the matching score in the product path.

[0096] The calculation of the matching score includes: first, determining the probability density value of the node appearing in the target product path; second, calculating the actual time interval between the node and its previous node, and at the same time, finding the average residence time and standard deviation of the behavior jump on the target product path from the map. If the difference between the actual time interval and the average residence time is less than the standard deviation, it is considered that the response time is consistent and is assigned a value of 1; if the difference is between the standard deviation and 2 times the standard deviation, it is assigned a value of 0.5; if it exceeds 2 times the standard deviation, it is assigned a value of 0; the above two scores are weighted summed to obtain the result.

[0097] The score reflects the actual influence of the node in the current path on the triggering of the product. To ensure the stability and comparability of the calculation results, the present invention normalizes the matching scores under each path through a standardized method.

[0098] Furthermore, in order to realize the path attribution judgment at the node level, the present invention takes the matching score of each node under multiple paths as the independent variable and the historical product trigger label as the dependent variable, and obtains the attribution weight vector of each node through least squares regression. The vector represents the attribution degree of the node under different product paths, that is, whether the node belongs to the core behavior facilitation node of a certain recommended path.

[0099] It can be seen that, unlike the traditional label classification method, this method can reflect the cross-contribution of a node in multiple recommended paths and has a stronger path structure representation ability.

[0100] S4.3: Based on the regressed path node weight vector and the node sequence relationship of the path structure in the process tree, the pass probability of each path is calculated, and the path set is sorted based on the pass probability. The top N paths are extracted as the recommended path candidate set, where N is a constant.

[0101] Specifically, the calculation process of the passing probability is:

[0102] The historical number of transfers between every two adjacent nodes is recorded in the order of the path, and the transfer probability of each step is calculated (the number of occurrences of the next node / the total number of transfers to the current node); for the nodes that must be passed (such as the login node), a minimum transfer probability is set (such as not less than 20%) to avoid blocking the critical path due to accidental low weight; the transfer probabilities of all adjacent node pairs in the path are multiplied in sequence to obtain the pass probability value of the complete path.

[0103] After completing the calculation of the pass probability, the present invention sorts all candidate path sets according to the pass probability, and selects the top N paths with the largest pass probabilities as the recommended path candidate set.

[0104] This operation realizes the effective transition from weight modeling at the node granularity to traffic determination at the path granularity, and is one of the key steps in realizing multi-path normalization selection in the present invention.

[0105] S4.4: Aggregate the cross-product nodes associated with each path in the recommended path candidate set according to their corresponding path pass probability, parse them into the original product set through the product-path probability mapping table, sort them by aggregation weight, and form the final product recommendation sequence. Simultaneously retain the associated path index and regression weight parameters for subsequent iterative optimization.

[0106] Specifically, the cosine similarity between paths is calculated, and paths with a similarity < 0.3 are considered conflicting paths (0.3 is the set threshold here, which needs to be set according to the actual situation during operation); if it is a non-conflicting path, the maximum pass probability value is taken; if it is a conflicting path, the Borda counting method is used to sort it, and finally the overall recommendation probability index of the product is formed.

[0107] Unlike existing techniques that rank products solely based on click counts or average ratings, our recommendation mechanism focuses on paths, integrating structural, behavioral, and temporal features to generate a recommendation sequence that is more informed by behavioral evolutionary logic. After ranking, the final product recommendation sequence is output to the user, while the path index and node regression weight vector are retained to serve as foundational data for the next round of recommendation training and iterative model optimization.

[0108] It can be seen that the present invention carries out a series of regression analysis and sorting and screening operations around the probability density characteristics of path nodes, aiming to realize a dynamic product recommendation and sorting process driven by historical behavior trajectories. This not only improves the personalized accuracy of the recommendation results, but also enhances the adaptability to the process path topology structure.

[0109] Further, such as Figure 2 As shown, this embodiment also provides an e-commerce product recommendation system based on artificial intelligence algorithms, including:

[0110] The multi-state feature extraction module 100 extracts multi-state feature parameters based on user historical behavior data, constructs a user-behavior path process tree using a multi-criteria splitting mechanism, and generates a node path set with behavior evolution rules;

[0111] The portrait slicing and mapping module 200 slices and reorganizes the user portrait set and the product portrait set under multi-dimensional features, and generates a candidate cross-mapping set through predefined dislocation mapping rules;

[0112] The deinterlaced graph module 300 constructs a graph of user behavior clusters, projects the node path set and the cross-product nodes in the candidate cross-mapping set into the deinterlaced graph structure between the behavior clusters, and extracts interaction points with high transfer weight and behavior coupling as candidate recommendation sets;

[0113] The path probability regression module 400 uses the path probability density map corresponding to each cross-product node in the candidate recommendation set to perform weighted regression calculation on each node of the user-behavior path process tree path, and generates a product recommendation sequence based on the maximum path pass probability.

[0114] This embodiment also provides a computer device suitable for the e-commerce product recommendation method based on artificial intelligence algorithm, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the e-commerce product recommendation method based on artificial intelligence algorithm proposed in the above embodiment.

[0115] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0116] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the e-commerce product recommendation method based on the artificial intelligence algorithm proposed in the above embodiment is implemented.

[0117] In summary, the present invention, by introducing multi-state behavior modeling and path process tree mechanism, can comprehensively capture the behavioral evolution characteristics of users on the e-commerce platform and effectively identify the dynamic change trend of user interests; through portrait slicing and staggered mapping strategies, it can fully explore the potential matching relationship between users and commodities, and improve the accuracy and diversity of recommendations. Combining the behavioral graph structure with the path probability density mapping, it can achieve quantitative evaluation and weight regression of the recommendation path, thereby significantly improving the relevance and click-through rate of the recommendation sequence. Compared with the traditional recommendation algorithm that only relies on static features or collaborative filtering methods, the present invention achieves deep integration at the structural modeling and temporal response levels, and has stronger user preference recognition capabilities and real-time recommendation adaptability. In summary, the present invention realizes the intelligent upgrade of personalized recommendations by constructing a recommendation process with behavioral chain criteria and coupled graph mapping, which not only improves the user experience, but also enhances the platform's conversion efficiency and content distribution capabilities.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An e-commerce product recommendation method based on artificial intelligence algorithm, characterized by: include: Based on the user's historical behavior data, multi-state feature parameters are extracted, and a multi-criteria splitting mechanism is used to construct a user-behavior path process tree, generating a node path set with behavior evolution rules; The user portrait set and the product portrait set are sliced ​​and reorganized under multi-dimensional features, and cross-combined using predefined dislocation mapping rules to generate a candidate cross-mapping set; Construct a user behavior cluster graph, project the node path set and the cross-product nodes in the candidate cross-mapping set into the deinterlaced graph structure between the behavior clusters, and extract the interaction points with high transfer weight and behavior coupling as the candidate recommendation set; Using the path probability density mapping corresponding to each cross-product node in the candidate recommendation set, reversely search the associated product-path probability mapping table, perform weighted regression calculation on each node of the user-behavior path process tree, and generate a product recommendation sequence based on the maximum path pass probability.

2. The e-commerce product recommendation method based on an artificial intelligence algorithm according to claim 1, characterized in that: The construction of the user-behavior path process tree includes: Based on the user behavior log data recorded in the e-commerce platform, the page dwell time parameters, browsing frequency parameters and jump sequence codes of each behavior event are extracted, and a multi-dimensional state vector sequence is constructed according to the time series; According to the difference change law of each dimension parameter in the multidimensional state vector sequence, a transfer threshold set is set, and the continuous state sequence is divided into discrete state segments according to the threshold, and the corresponding state label set is marked; Convert the state label set into a state path chain, perform frequency statistics and co-occurrence relationship mining on the node sequence, and construct a state criterion splitting condition group containing the behavior evolution relationship; According to the state label sequence and the state criterion splitting condition group, the path process tree structure is generated by splitting in sequence from the root node.

3. The e-commerce product recommendation method based on an artificial intelligence algorithm according to claim 1, characterized in that: The slice reorganization includes: Based on the user portrait set and the product portrait set, the main feature dimensions are extracted and a standardized feature matrix is ​​constructed. The standardized feature matrix includes the user feature matrix and the product feature matrix. The local change rate vector on each feature dimension is calculated and stability check is performed after slicing the fluctuation points. The user feature matrix and the product feature matrix are dynamically sliced ​​to generate a set of feature sub-matrices.

4. The e-commerce product recommendation method based on an artificial intelligence algorithm according to claim 3, characterized in that: Generating the candidate cross-mapping set includes: Based on the asymmetric distribution characteristics of product images in terms of price and popularity, an offset vector sequence is set. Each offset controls the cross-combination path to generate a set of dislocation mapping rules. The sliced ​​user and item sub-matrices are combined according to the staggered mapping rule group to form a candidate cross-mapping set.

5. The e-commerce product recommendation method based on artificial intelligence algorithm according to claim 1, characterized in that: When projecting the cross-product node into the deinterlaced graph structure, the associated original product ID mapping table is recorded simultaneously to generate the associated product-path probability mapping table; Based on the node path set, the jump frequency and time interval between user behavior nodes are counted to construct a weighted directed graph structure; A cross-product node set is defined based on the candidate cross-mapping set, and the user behavior nodes involved are projected into a weighted directed graph structure to construct an index matrix consisting of behavior nodes and cross-product nodes.

6. The e-commerce product recommendation method based on artificial intelligence algorithm according to claim 5, characterized in that: The generation of the candidate recommendation set includes: For each mapping unit in the index matrix, the coupling weight vector between the behavior paths is calculated based on the set of shortest paths connecting the behavior node with other nodes in the graph. Based on the coupling weight vector and the edge weights in the original behavior transfer graph, the coupling degree threshold and transfer weight threshold are jointly set to screen the cross-product node pairs that meet the conditions to form a candidate recommendation set.

7. The e-commerce product recommendation method based on artificial intelligence algorithm according to claim 1, characterized in that: The mapping of the path probability density includes: Based on the frequency of each node in the user behavior sequence in the user-behavior path process tree, the number of historical behaviors jumping from the previous node to the current node is counted, and all historical jump behaviors are smoothed according to their distribution on the timeline; Define dynamic path entry detection rules based on the response time distribution of cross-product nodes in user behavior; The above two parts are matched in terms of node number and path order to form a three-dimensional distribution structure with path node-frequency density-response time as the joint index, which constitutes a density map for behavior path sorting.

8. An e-commerce product recommendation system based on an artificial intelligence algorithm, based on the e-commerce product recommendation method based on an artificial intelligence algorithm according to any one of claims 1 to 7, characterized in that: Also includes: The multi-state feature extraction module extracts multi-state feature parameters based on user historical behavior data, uses a multi-criteria splitting mechanism to construct a user-behavior path process tree, and generates a node path set with behavior evolution rules; The portrait slicing and mapping module slices and reorganizes the user portrait set and the product portrait set under multi-dimensional features, and generates a candidate cross-mapping set through predefined dislocation mapping rules; The deinterlaced graph module constructs a graph of user behavior clusters, projects the node path set and the cross-product nodes in the candidate cross-mapping set into the deinterlaced graph structure between behavior clusters, and extracts interaction points with high transfer weight and behavior coupling as candidate recommendation sets; The path probability regression module uses the path probability density mapping corresponding to each cross-product node in the candidate recommendation set to reversely search the associated product-path probability mapping table, perform weighted regression calculations on each node of the user-behavior path process tree, and generate a product recommendation sequence based on the maximum path pass probability.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the e-commerce product recommendation method based on artificial intelligence algorithm described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the e-commerce product recommendation method based on artificial intelligence algorithm described in any one of claims 1 to 7 are implemented.

Images