E-commerce platform interest analysis type commodity recommendation method based on artificial intelligence

Abstract

The application relates to the technical field of artificial intelligence e-commerce, and discloses an e-commerce platform interest analysis type commodity recommendation method based on artificial intelligence. The method comprises the following steps: constructing a network topology structure describing dynamic changes of interests, i.e. an interest evolution network, based on user multi-dimensional behavior records; injecting preset virtual interaction events into the network, simulating the propagation and attenuation of the events, deducing the potential interest development track of the user, and calculating the future activity probability distribution of each interest node; combining platform real-time commodity supply flow and its environmental attribute labels, mapping commodities to the interest network, evaluating the matching strength, using the activity probability distribution for weighted correction, and generating a candidate commodity list with time weight. Through dynamic network modeling and virtual event simulation deduction, the application realizes deeper understanding and more forward grasping of user interests, and improves the accuracy, diversity and timeliness of commodity recommendation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence e-commerce technology, specifically to an interest-based product recommendation method for e-commerce platforms based on artificial intelligence. Background Technology

[0002] Existing product recommendation systems generally rely on static user profiles or temporal behavior models. Static profile methods represent user interests as a series of discrete, fixed labels or feature vectors, ignoring the inherent connections between interest points and their dynamic evolution over time. Models based on temporal behavior sequences, such as recurrent neural networks or attention models, while capturing the order in which behaviors occur, are essentially modeling behavioral sequences, making it difficult to explicitly represent and explore the complex interactions and evolutionary relationships between multiple coexisting user interests. A common drawback of these methods is that their modeling of user interests is isolated or linear, failing to effectively simulate the dynamic, networked evolution of interests as an organic whole. This results in a insufficient understanding of changes in user interests and a difficulty in capturing potential interest transfer paths.

[0003] In interest prediction, conventional techniques often employ direct extrapolation or statistical probability prediction based on historical behavioral patterns. Essentially, these are inductive and approximate predictions based on existing explicit behavioral data, falling under the category of "retrospective" prediction. Their drawback lies in their passive reliance on historical data, lacking the ability to proactively extrapolate and proactively simulate, and failing to identify and quantify potential interest development directions before users exhibit relevant behaviors. When user interests are in their nascent or transitional stages, and historical data is sparse or patterns are unclear, the predictive accuracy and foresight of these methods significantly decrease. The purpose of this invention is to address the technical problems of how to more accurately model the dynamic network structure of user interests and how to proactively simulate and extrapolate the trajectory of potential user interests. Summary of the Invention

[0004] The purpose of this invention is to provide an interest-based product recommendation method for e-commerce platforms based on artificial intelligence, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an artificial intelligence-based interest-based product recommendation method for e-commerce platforms, the method comprising:

[0006] An interest evolution network is established to describe the dynamic changes of user interests. The interest evolution network is constructed from the multi-dimensional behavior records generated by the user on the e-commerce platform within a time period. The multi-dimensional behavior records are temporally correlated and pattern abstracted to form a network topology structure containing multiple interest nodes and associated edges.

[0007] Multiple preset virtual interaction events are injected into the interest evolution network. By simulating the propagation path and decay process of the virtual interaction events in the interest evolution network, the potential interest development trajectory of the user is deduced. Based on the potential interest development trajectory, the future activity probability distribution of each interest node in the interest evolution network is calculated.

[0008] The system collects the current real-time product supply flow of the e-commerce platform. The real-time product supply flow includes a set of products and their associated environmental attribute tags. The environmental attribute tags are mapped to the interest evolution network. The matching strength between each product in the product set and the interest node is evaluated. The matching strength is then weighted and corrected according to the activity probability distribution to generate a candidate product list with time weights.

[0009] Preferably, the step of performing temporal correlation and pattern abstraction on the multi-dimensional behavior records to form a network topology containing multiple interest nodes and associated edges includes:

[0010] Independent behavioral events are extracted from the multi-dimensional behavioral records and sorted by timestamp;

[0011] Calculate the frequency and conditional probability of co-occurrence of different behavioral events within adjacent time windows;

[0012] Behavioral events that meet the co-occurrence criteria are clustered into interest clusters;

[0013] Each interest cluster is abstractly represented as an interest node;

[0014] Based on the temporal and co-occurrence relationships between the behavioral events, directional and weighted association edges are established between the nodes of interest, thereby forming the network topology.

[0015] Preferably, the deduction of the potential interest development trajectory corresponding to the user includes:

[0016] Starting from the currently active interest node in the interest evolution network;

[0017] Calculate the probability of interest state transition along the direction of the associated edge;

[0018] By combining the content attributes of the virtual interactive events, the diffusion process of interest states along different paths is simulated;

[0019] Record the new interest nodes activated during the diffusion process and their activation intensity;

[0020] The paths whose activation intensity exceeds the threshold are connected to form the potential interest development trajectory.

[0021] Preferably, the step of weighting and correcting the matching strength according to the activity probability distribution to generate a candidate product list with time weights includes:

[0022] Extract the probability values ​​corresponding to specific future time periods from the activity probability distribution;

[0023] The probability value is then fused with the initial matching strength of the product and interest node;

[0024] The products in the product set are reordered based on the fusion calculation results;

[0025] Each sorted product is assigned a time-weighted coefficient that indicates the urgency of the recommendation.

[0026] The ranking results are combined with the time weight coefficient to generate the candidate product list with time weight.

[0027] Preferably, the method further includes:

[0028] An interest state comparison unit is constructed, which is used to calculate the stability index of the interest evolution network under different simulated intervention strategies in parallel. The simulated intervention strategy is formed by adding or suppressing specific commodity information flows into the network topology. The simulated intervention strategy that makes the stability index reach the optimal balance is selected as the intervention strategy to be executed.

[0029] Based on the intervention strategy to be executed and the candidate product list with time weights, a set of product information push instructions containing push timing and push sequence is generated, and the product information push instructions are input into the recommendation interface of the e-commerce platform.

[0030] Preferably, the step of selecting the simulated intervention strategy that achieves the optimal balance of the stability index as the intervention strategy to be executed includes:

[0031] The stability index includes the network structure cohesion index and the interest exploration index;

[0032] Calculate the numerical combination of the network structure cohesion index and interest exploration index under different simulated intervention strategies;

[0033] Locate the coordinate point corresponding to the numerical combination in a two-dimensional coordinate system;

[0034] Calculate the Euclidean distance from each coordinate point to the preset ideal equilibrium point;

[0035] The simulated intervention strategy associated with the coordinate point with the smallest Euclidean distance is selected as the intervention strategy to be executed.

[0036] Preferably, generating a set of product information push instructions that includes push timing and push sequence includes:

[0037] Analyze the intervention strategy to be executed and identify the set of interest nodes that need to be strengthened in the strategy;

[0038] Select the subset of products that have the highest correlation with the set of interest nodes from the list of candidate products with time weights;

[0039] Based on the time weight coefficient of each product in the product subset, the corresponding push time interval for each product is determined;

[0040] Within the push time interval, the specific push timing is calculated based on the user's historical online activity patterns;

[0041] Arrange the product subsets according to the chronological order of the push events to form the push sequence;

[0042] The push timing and push sequence are encoded into product information push instructions that can be recognized by the e-commerce platform's recommendation interface.

[0043] Preferably, the method further includes:

[0044] After executing the product information push instruction, the system monitors the secondary behavior sequence generated by the user in response to the pushed product information, and feeds the secondary behavior sequence back to the interest evolution network as a feedback signal, triggering the adjustment of parameters of the corresponding interest nodes and related edges in the interest evolution network, thus completing a complete recommendation and feedback iteration.

[0045] Preferably, the monitoring of the user's secondary behavioral sequence in response to the pushed product information includes:

[0046] Capture user behavior regarding exposure, clicks, browsing time, favorites, adding to cart, and purchases of pushed product information;

[0047] The captured behaviors are encoded into a structured sequence of behavior type and timestamp according to the chronological order in which they occur.

[0048] Calculate the time interval and behavior transition pattern between adjacent behaviors in the secondary behavior sequence;

[0049] The time interval and behavior transfer pattern are taken as the core content of the feedback signal.

[0050] Preferably, the step of feeding the secondary behavior sequence as a feedback signal back to the interest evolution network to trigger parameter adjustments to the corresponding interest nodes and associated edges in the interest evolution network includes:

[0051] Map the secondary behavior sequence back to the network topology to identify the target interest nodes and associated edges affected by the sequence;

[0052] Based on the behavior type, determine the direction and magnitude of adjusting the activity parameter of the target interest node;

[0053] Based on the behavior transfer pattern, determine the direction and magnitude of the adjustment of the weight of the associated edge;

[0054] The internal state parameters of the interest evolution network are updated by applying the adjustment direction and magnitude.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] By constructing a network topology of interest nodes and associated edges formed by temporal correlation and pattern abstraction of multi-dimensional behavioral records, user interests are transformed from static labels or linear sequences into a dynamically related network structure. This dynamic network topology can explicitly characterize the correlation strength and evolution path between different interest points, enabling the interest model to reflect the complex relationships of mutual influence, symbiosis, or transfer between interests. Compared to isolated or sequential representations, this network modeling can reveal the intrinsic structure and dynamism of user interests more refinedly and fundamentally, providing a richer and more accurate representation of interest states for subsequent recommendations. This allows product matching to delve into the association level of the interest network, improving the diversity and interpretability of recommendation results.

[0057] By injecting various pre-set virtual interaction events into a constructed interest evolution network and simulating their propagation and decay processes within the network to deduce potential interest trajectories, this method achieves proactive, mechanism-based interest prediction. The injection of virtual events is equivalent to applying controlled stimuli to the interest network, and the response and future activity probability of each node in the network are evaluated by observing their propagation effects. This approach does not rely on fitting direct historical data of the future occurrence of target interests, but rather proactively discovers potential interest points by simulating the inherent logic of interest diffusion. This allows the system to predict multiple possible branches of a user's interest development before explicit user behavior occurs, calculate a time-weighted activity probability distribution, enhance the predictability and exploratory capabilities of the recommendation system, and respond to and guide changes in user interests earlier. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based e-commerce platform interest analysis-based product recommendation method described in this invention.

[0059] Figure 2 A flowchart for constructing a network topology;

[0060] Figure 3 A flowchart for generating a candidate product list with time weights;

[0061] Figure 4 A comparison chart showing the matching scores between products and interest nodes related to fitness training and healthy eating;

[0062] Figure 5 A stacked graph showing the distribution of user secondary behavior types. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1 This invention provides an AI-based interest-based product recommendation method for e-commerce platforms. The method includes: establishing an interest evolution network to describe the dynamic changes in user interests. This network is constructed from multi-dimensional behavioral records generated by users on the e-commerce platform within a time period. The multi-dimensional behavioral records are temporally correlated and pattern abstracted to form a network topology containing multiple interest nodes and associated edges. Multiple preset virtual interaction events are injected into the interest evolution network. By simulating the propagation path and decay process of these virtual interaction events in the network, the potential interest development trajectory of the user is deduced. Based on the potential interest development trajectory, the future activity probability distribution of each interest node in the network is calculated. The current real-time product supply flow of the e-commerce platform is collected. This real-time product supply flow includes a set of products and their associated environmental attribute tags. The environmental attribute tags are mapped to the interest evolution network. The matching strength between each product in the product set and the interest nodes is evaluated. The matching strength is weighted and adjusted according to the activity probability distribution to generate a candidate product list with time weights.

[0065] Example 1: See Figure 2Independent behavioral events are extracted from multi-dimensional behavioral records and sorted by timestamp. The frequency and conditional probability of co-occurrence of different behavioral events within adjacent time windows are calculated. Behavioral events that meet the co-occurrence condition are clustered into interest clusters, and each interest cluster is abstractly represented as an interest node. Based on the temporal and co-occurrence relationships between behavioral events, directional and weighted association edges are established between interest nodes to form a network topology. Starting from the currently active interest node in the interest evolution network, the probability of interest state transition is calculated along the direction of the association edge. Combining the content attributes of virtual interactive events, the diffusion process of interest states along different paths is simulated. The newly activated interest nodes and their activation intensity are recorded during the diffusion process. Paths with activation intensity exceeding a threshold are connected to form potential interest development trajectories.

[0066] In practice, for a user with behavioral records on an e-commerce platform, the process of establishing an interest evolution network begins with the analysis of multi-dimensional behavioral records. These records include event types such as browsing, searching, clicking, adding to favorites, adding to cart, and purchasing. Each behavioral event is accompanied by a precise timestamp, product identifier, and contextual attributes at the time of the action. Independent behavioral events are extracted from these records and sorted by timestamps to form a chronological sequence of original behavioral events. The frequency and conditional probability of co-occurrence of different behavioral events within adjacent time windows are calculated. The length of the time window can be configured according to platform characteristics. Co-occurrence of behavioral events refers to two or more behavioral events being triggered by the same user within the same time window. Behavioral events that meet the co-occurrence criteria are clustered into interest clusters. The co-occurrence criteria can be a co-occurrence frequency or conditional probability exceeding a preset threshold. The clustering process is based on the semantic similarity and co-occurrence strength between behavioral events.

[0067] In some embodiments, the abstract representation process of interest clusters maps each interest cluster to an interest node. Interest nodes are represented in the interest evolution network using a high-dimensional vector, which is a weighted aggregation of the product attribute vectors associated with all behavioral events within the cluster and the behavioral intensity. Based on the temporal and co-occurrence relationships between behavioral events, directional and weighted association edges are established between interest nodes. The direction of the association edges is determined by the chronological order of the behavioral events, and the initial weights of the association edges are jointly determined by the transition frequency and conditional probability between behavioral events, thus forming the network topology. It can be understood that interest nodes in the network topology represent user interests at different granularities, while the association edges characterize the possibility and intensity of evolution between these interest points.

[0068] In practice, the process of extrapolating the development trajectory of potential interests begins with the currently active interest nodes in the interest evolution network. These nodes are determined by analyzing the user's behavior records over the most recent period. Along the directions of the associated edges, the probability of interest state transitions is calculated, based on the weights of the associated edges and the Markov property of historical transitions. Combining the content attributes of virtual interactive events, the diffusion process of interest states along different paths is simulated. Virtual interactive events are modeled as external stimulus signals, and their content attributes are encoded as vectors and interact with the representation vectors of interest nodes to calculate the initial activation energy of each interest node. New interest nodes activated during the diffusion process and their activation strengths are recorded. The activation strength depends on the initial activation energy, the weights of the associated edges on the path, and the decay coefficient during the simulation. Paths with activation strengths exceeding a threshold are connected to form the potential interest development trajectory.

[0069] In some embodiments, a specific formula for calculating conditional probabilities to quantify co-occurrence relationships can be expressed as:

[0070]

[0071] Where: characters Indicating in behavioral events Behavioral events occurring within the same time window The conditional probability of occurrence, character Represents behavioral events and behavioral events The total number of times a character co-occurs within the same time window in the historical record. Represents behavioral events The total number of occurrences in historical records. This conditional probability value will be used to evaluate the correlation strength between behavioral events and as one of the bases for determining whether to cluster them into the same interest cluster and for subsequently establishing the weights of the association edges. It can be understood that the implementation of the entire embodiment strictly follows the logical chain from the original behavioral sequence to network structure construction, and then to interest diffusion simulation based on the network, where the flow and calculation of data constitute the core of the method.

[0072] Example 2: See Figure 3The system extracts probability values ​​from the activity probability distribution corresponding to specific future time periods. These probability values ​​are then fused with the initial matching strength of products and interest nodes. Based on the fusion result, the products in the product set are reordered, and each sorted product is assigned a time weight coefficient representing the urgency of the recommendation. The sorting result and the time weight coefficient are then integrated to generate a candidate product list with time weights. In practice, the activity probability distribution is a vector, where each dimension corresponds to an interest node, and its value represents the probability that the interest node will be activated or in a highly active state within a specific future time period. The system extracts probability values ​​from the activity probability distribution corresponding to specific future time periods, which could be an upcoming workday, a weekend, or a promotional period. The system selects the corresponding time period index based on the recommendation task's objective and extracts the corresponding probability value vector from the multi-dimensional activity probability distribution matrix.

[0073] In some embodiments, the probability value is fused with the initial matching strength of the product and the interest node. The initial matching strength is obtained by calculating the cosine similarity between the environmental attribute label vector of the product and the representation vector of the interest node. The purpose of the fusion operation is to inject future interest trends and preferences into the current static matching evaluation. A specific fusion operation formula is expressed as follows:

[0074]

[0075] Where: characters This represents the final overall matching score of the product after weighted adjustments, and the character count is... Represents the total number of interest nodes in the interest evolution network, character This indicates that the product is related to the first The initial matching strength of each interest node, character Represents the extraction of the first value from the activity probability distribution. The probability value of the activity of an interest node in a specific time period in the future, character It is a base coefficient greater than 1, used to amplify the non-linear impact of the high activity probability value on the final result. Based on the fusion calculation results, the products in the product set are re-sorted according to the final comprehensive matching score of each product. Proceed from high to low.

[0076] It is understandable that a time-weighted coefficient representing the urgency of the recommendation is attached to each sorted product. The calculation of the time-weighted coefficient depends on the final comprehensive matching score generated for that product. The time period used is a specific future time interval. If the specific future time interval is close to the current time, a higher time weight coefficient is assigned, indicating that the recommendation is urgent; if the specific future time interval is far away, a lower time weight coefficient is assigned. The ranking results and time weight coefficients are integrated to generate a candidate product list with time weights. Each entry in this list includes a product identifier, a final comprehensive matching score, and a time weight coefficient.

[0077] In a practical implementation, an example scenario involves a user's interest evolution network comprising three interest nodes: "outdoor camping," "digital photography," and "home organization." Their activity probabilities for the "future weekend" timeframe are 0.8, 0.3, and 0.1, respectively. An e-commerce platform's real-time product supply stream includes a product called "ultralight portable tent," whose initial matching strengths with the aforementioned three interest nodes are 0.9, 0.1, and 0.0, respectively. Calculations are performed using a fusion calculation formula, with a base coefficient set. If the score is 2, then the final overall matching score for the "ultralight portable tent" product is 2. Meanwhile, another product, "Professional Camera Lens Cleaning Kit," had initial match strengths of 0.2, 0.9, and 0.0 with the three interest nodes, respectively, and its final comprehensive match score... The calculation results show that although the initial matching strength between the "Professional Camera Lens Cleaning Kit" and the "Digital Photography" interest node is high, its final comprehensive matching score may be lower than that of the "Ultra-light Portable Tent" product because the "Digital Photography" interest node has a low probability of activity during the target time period.

[0078] In some embodiments, the time weighting coefficient is calculated based on the time difference between the midpoint of a specific future time period and the current time; the smaller the time difference, the larger the time weighting coefficient. It can be understood that the candidate product list with time weights is not only sorted according to the matching degree between the products and the user's dynamic interests, but also uses the time weighting coefficient to mark the timeliness priority of the recommendation action.

[0079] Example 3: Construct an interest state comparison unit. This unit is used to calculate the stability index of the interest evolution network under different simulated intervention strategies in parallel. The simulated intervention strategies are formed by adding or suppressing product information flows in the network topology. The simulated intervention strategy that achieves the optimal balance of stability index is selected as the intervention strategy to be executed. The stability index includes the network structure cohesion index and the interest exploration index. The numerical combination of the network structure cohesion index and the interest exploration index under different simulated intervention strategies is calculated. The coordinate points corresponding to the numerical combinations are located in a two-dimensional coordinate system. The Euclidean distance from each coordinate point to the preset ideal equilibrium point is calculated. The simulated intervention strategy associated with the coordinate point with the smallest corresponding Euclidean distance is selected as the intervention strategy to be executed.

[0080] In practical implementation, based on the generated interest evolution network and a time-weighted candidate product list, the core lies in constructing an interest state comparison unit. This unit is an independent computational module used to compute the stability indices of the interest evolution network under different simulated intervention strategies in parallel. Simulated intervention strategies are formed by adding or suppressing product information flows in the network topology. Adding product information flows corresponds to introducing new product exposures strongly associated with specific interest nodes into the network, while suppressing product information flows corresponds to temporarily blocking product information related to certain interest nodes in the simulated environment. The stability indices include network structure cohesion and interest exploration. The network structure cohesion quantifies the tightness and robustness of connections between interest nodes in the interest evolution network, while the interest exploration quantifies the extent to which simulated intervention strategies may guide user interests to migrate to previously weak or underdeveloped areas of the network.

[0081] In some embodiments, numerical combinations of network structure cohesion and interest exploration indices under different simulated intervention strategies are calculated. For each simulated intervention strategy, the interest state comparison unit applies the strategy in an isolated simulation environment and observes the feedback of virtual user interests to calculate a set. The coordinate points corresponding to the numerical combinations are located in a two-dimensional coordinate system. The horizontal axis of this coordinate system represents the network structure cohesion index, and the vertical axis represents the interest exploration index. The evaluation result of each simulated intervention strategy is represented as a specific point in this coordinate system. The Euclidean distance from each coordinate point to a preset ideal equilibrium point is calculated. The coordinates of the ideal equilibrium point are predefined, representing the optimal balance between network structure cohesion and interest exploration. The simulated intervention strategy associated with the coordinate point with the smallest corresponding Euclidean distance is selected as the intervention strategy to be executed.

[0082] It is understandable that, through the above quantitative calculation and comparison process, the interest state comparison unit can automatically select a simulated intervention strategy from a variety of possible intervention paths that can both maintain the inherent structure of the user's interest network and appropriately guide interest exploration. In some embodiments, the coordinates of the ideal equilibrium point are not fixed and can be dynamically adjusted according to the user's life cycle stage or the platform's overall operational goals, thereby making the selection of the intervention strategy to be executed more flexible and targeted.

[0083] Example 4: Analyze the intervention strategy to be executed, identify the set of interest nodes that need to be strengthened in the strategy, select the subset of products with the highest correlation with the set of interest nodes from the candidate product list with time weight, determine the push time interval for each product based on the time weight coefficient of each product in the product subset, calculate the specific push timing based on the user's historical online activity pattern within the push time interval, arrange the product subsets according to the time sequence of the push timing to form a push sequence, and encode the push timing and push sequence into a product information push instruction that can be recognized by the e-commerce platform recommendation interface.

[0084] In practice, the intervention strategy to be executed is a structured data object that explicitly includes the strategy intent, a list of target interest nodes, and the expected intervention intensity coefficient. Parsing the intervention strategy to be executed and identifying the set of interest nodes that need to be strengthened is accomplished by reading the target interest node identifier list field from the intervention strategy data object. The identified interest nodes constitute the set of interest nodes that need to be strengthened.

[0085] In some embodiments, a subset of products with the highest relevance to the set of interest nodes is selected from a time-weighted candidate product list. The relevance is calculated based on the final comprehensive matching score between the product and each node in the set of interest nodes. One approach is to sum or take the maximum of the final comprehensive matching scores between the product and all interest nodes in the set, and then rank the candidate products according to the calculated total relevance score, selecting the top-ranked products to form a product subset. Refer to Table 1, which shows a fragment of a candidate product list and its relevance calculation.

[0086] Table 1: Calculation Table of Correlation Between Candidate Product List and Interest Node Set

[0087]

[0088] Suppose that the set of interest nodes to be reinforced by the intervention strategy to be implemented includes "fitness training" and "healthy eating," and the total relevance score is the sum of the matching scores with these two nodes. According to Table 1, products P001 and P003 have the highest total relevance scores, and therefore are preferentially selected into the product subset.

[0089] In practice, the push time interval for each product is determined based on its time weight coefficient within the product subset. A higher time weight coefficient corresponds to a more urgent push time interval and a shorter time window. One method for determining the starting point of the push time interval is based on calculations using the time weight coefficient, for example, using the formula:

[0090]

[0091] Where: characters Indicates the start time of the push time interval, character Indicates the current system time, character Represents a basic time delay constant, character This represents the time weighting coefficient of the product. For product P003, with a time weighting coefficient of 0.90, its... Product P004, which has a time weighting coefficient of 0.70, is included. Within the push time interval, the specific push timing is calculated by combining the user's historical online activity patterns. The historical online activity patterns can be extracted from the user's past login and browsing records to show the activity probability distribution at different times of the day. The specific push timing is determined as one or more time points within the push time interval where the user's activity probability is the highest.

[0092] The product subsets are arranged in chronological order of push timing to form a push sequence. The push timing and push sequence are encoded into product information push instructions recognizable by the e-commerce platform's recommendation interface. These instructions typically use a structured data format, such as JSON or Protocol Buffers, and include fields such as: instruction type, target user identifier, a sequentially ordered list of product identifiers, and a pre-defined push timestamp for each product identifier. In some embodiments, in an example scenario, the system's current time... The base delay is 10:00 AM. Set to 12 hours. For product P003 (low-calorie chicken breast, time weighting coefficient 0.90), calculate the start time of its push notification time interval. Based on the user's historical activity patterns, it was found that this user was most active between 12:00 PM and 1:00 PM. Therefore, the specific push notification time for product P003 was determined to be 12:30 PM, which falls within the calculated push time interval. For product P001 (whey protein powder, time weighting coefficient 0.85), its... The specific push timing, based on activity patterns, may be set at 13:00. The resulting push sequence would be [Product P003, Product P001].

[0093] It is understandable that the generation process of product information push instructions strictly follows the intent of the intervention strategy to be executed. From product selection to push timing arrangement, it is subject to the dual constraints of the relevance of the interest node set and the time weight coefficient, ensuring that the final execution action is aligned with the previously calculated interest evolution trend and timeliness requirements. In some embodiments, if the specific push timings calculated for multiple products are the same or very close, these products are then sorted a second time according to their total relevance score to determine the display order under the same push timing, thereby ensuring the logical consistency of the push sequence.

[0094] See Figure 4In the interest node matching analysis during the product screening stage, the distribution of matching scores between different products and the two interest nodes "fitness training" and "healthy eating" is presented intuitively. Specifically, whey protein powder, yoga mats, sports water bottles, elastic bands, and quick-drying sportswear have significantly higher matching scores with the "fitness training" node (all above 0.75), reflecting a strong correlation between these products and the interest in "fitness training." Meanwhile, low-calorie chicken breast, whole-wheat bread, and sugar-free yogurt have more prominent matching scores with the "healthy eating" node (all above 0.85), corresponding to the matching advantage of the "healthy eating" interest. The difference in score distribution is highly consistent with the attribute characteristics of the products: fitness equipment and sportswear are naturally suited to the "fitness training" interest, while low-calorie food products are more aligned with the "healthy eating" interest. Furthermore, the score data in the figure can also serve as a basis for calculating the correlation between products and interest nodes. For example, by summing the scores or taking the maximum value, the correlation strength between products and the set of interest nodes can be quantified, providing data support for the subsequent screening of product subsets.

[0095] Example 5: After executing the product information push instruction, the system monitors the secondary behavior sequences generated by the user in response to the pushed product information. These secondary behavior sequences are then fed back to the interest evolution network as feedback signals, triggering parameter adjustments to the corresponding interest nodes and associated edges in the network, completing a full recommendation and feedback iteration. The system captures user exposure, clicks, browsing duration, favorites, adding to cart, and purchase behaviors related to the pushed product information. These behaviors are encoded into structured sequences of behavior types and timestamps according to their chronological order. The time intervals and behavior transition patterns between adjacent behaviors in the secondary behavior sequences are calculated, and these time intervals and behavior transition patterns are used as the core content of the feedback signal. The secondary behavior sequences are mapped back to the network topology, identifying the target interest nodes and associated edges affected by the sequences. Based on the behavior type, the direction and magnitude of adjusting the activity parameters of the target interest nodes are determined. Based on the behavior transition pattern, the direction and magnitude of adjusting the weights of the associated edges are determined. The internal state parameters of the interest evolution network are updated by applying these adjustments.

[0096] In practice, the system is activated immediately after the product information push command is executed via the e-commerce platform's recommendation interface and pushed to the user. Following the execution of the push command, it monitors the user's secondary behavior sequence in response to the pushed product information. This secondary behavior sequence refers to the sequential set of all traceable interactive behaviors a user performs on one or more product information items pushed by the system within a complete recommendation exposure cycle. The system captures user exposure, clicks, browsing duration, favorites, adding to cart, and purchase behaviors related to the pushed product information. The e-commerce platform's front-end tracking and back-end logging system collaboratively record the user's behavior type, associated product identifier, and timestamp accurate to milliseconds when interacting with the pushed information. Based on the chronological order of the behaviors, the captured behaviors are encoded into a structured sequence of behavior type and timestamp.

[0097] In some embodiments, the time interval and behavior transition pattern between adjacent behaviors in the secondary behavior sequence are calculated. The time interval is the difference in timestamps between two consecutive behaviors in the sequence, and the behavior transition pattern is a combination of the preceding and following behavior types. For example, "click -> browse" and "browse -> favorite" constitute specific behavior transition patterns. The time interval and behavior transition pattern are used as the core content of the feedback signal. The feedback signal is a structured data object that, in addition to containing the core content, is associated with the original push instruction identifier that triggered the feedback and the corresponding user identifier. In a specific implementation, the secondary behavior sequence is fed back to the interest evolution network as a feedback signal, triggering parameter adjustments to the corresponding interest nodes and associated edges in the interest evolution network. This process is executed by the network update module. The secondary behavior sequence is mapped back to the network topology, identifying the target interest nodes and associated edges affected by the sequence. The mapping is based on the product identifier associated with each behavior in the secondary behavior sequence. By querying the matching relationship database between products and interest nodes, one or more interest nodes mainly associated with the product are determined. These nodes are the target interest nodes, and the edges connecting these target interest nodes are the target associated edges. Based on the behavior type, the direction and magnitude of adjustments to the activity parameters of the target interest nodes are determined. Different behavior types are assigned different feedback weights. For example, purchasing behavior usually brings a larger positive adjustment than clicking behavior, while long-term browsing may bring a more significant positive adjustment than short-term exposure. Based on the behavior transfer pattern, the direction and magnitude of adjustments to the weights of related edges are determined. Positive behavior transfer patterns will strengthen the weights of the corresponding related edges, while negative or ineffective transfers may maintain or slightly weaken the weights.

[0098] The direction and magnitude of the adjustment are used to update the internal state parameters of the interest evolution network. A specific parameter adjustment can be achieved using the formula:

[0099]

[0100] Where: characters This represents the adjustment amount applied to the target interest node activity parameter or the target associated edge weight parameter. (Character) This represents the number of behavioral events in a secondary behavioral sequence that are related to a specific target interest node or associated edge. Indicates the first The behavior type weight coefficient corresponding to each relevant behavioral event is a predefined coefficient representing the influence strength of different behavior types. Indicates the first The duration or normalized intensity of each relevant behavioral event, for instantaneous behaviors such as clicks, A fixed value can be set for browsing behavior. This refers to the browsing duration, measured in seconds. Adjustment amount. This will be added to the current value of the target parameter, thereby completing the real-time update of the internal state parameters of the interest evolution network.

[0101] Understandably, in an example scenario, a user receives push notifications for products P003 and P001, generating a secondary behavioral sequence including clicking on P003, browsing for 180 seconds, adding it to favorites, and clicking on P001, browsing for 15 seconds, and then leaving. The system identifies P003 as associated with the "healthy eating" interest node, and P001 as associated with the "fitness training" interest node. For the "healthy eating" node, its activity parameter will receive a significant positive adjustment based on the favorites behavior and the long browsing duration. Furthermore, if there is a correlation edge between the "healthy eating" node and the "fitness training" node, the weight of this correlation edge may receive a positive adjustment based on the behavior shift pattern, since the user behavior sequence shows a shift from focusing on fitness supplements to focusing on health foods. Conversely, for the "Fitness Training" node, because the browsing time for P001 is very short, the adjustment amount obtained is minimal. These adjustments may be very small or even slightly negative. After all calculations are complete, these adjustments are applied to update the corresponding network parameters, thus enabling the interest evolution network to more accurately reflect the latest dynamics of user interests.

[0102] In some embodiments, completing parameter adjustments marks the end of a full recommendation and feedback iteration. The updated interest evolution network serves as the starting point for the next recommendation cycle calculation, thereby enabling continuous learning and adaptive optimization of the interest model. It can be understood that the monitoring and feedback mechanism of secondary behavior sequences ensures that the recommendation system forms a closed loop from perception, decision-making, execution to learning, enabling the recommendation strategy to dynamically respond to real user feedback rather than relying on static historical profiles.

[0103] See Figure 5The graph uses stacked bar charts to visually represent the frequency distribution of secondary behavior types (add to cart, favorite, exposure, browsing, clicking, and purchasing) for different users (U001 to U005). Specifically, each bar corresponds to a single user, and different colored blocks within the bar represent the frequency percentage of each behavior type. The height of the block directly indicates the number of times the corresponding behavior occurred. Data analysis shows that U001's total behavior frequency is significantly higher than other users, with a prominent proportion of browsing and clicking behaviors. Other users (such as U002 and U005) have relatively lower total behavior frequencies, and their behavior type distributions also exhibit differentiated characteristics. The core value of this graph lies in quantifying the secondary behavior combination patterns of different users within the recommendation feedback cycle, providing direct behavioral data support for subsequently mapping behavior sequences to the interest evolution network and adjusting the activity of interest nodes and the weights of associated edges.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for interest-based product recommendation on e-commerce platforms based on artificial intelligence, characterized in that, The method includes: An interest evolution network is established to describe the dynamic changes of user interests. The interest evolution network is constructed from the multi-dimensional behavior records generated by the user on the e-commerce platform within a time period. The multi-dimensional behavior records are temporally correlated and pattern abstracted to form a network topology structure containing multiple interest nodes and associated edges. Multiple preset virtual interaction events are injected into the interest evolution network. By simulating the propagation path and decay process of the virtual interaction events in the interest evolution network, the potential interest development trajectory of the user is deduced. Based on the potential interest development trajectory, the future activity probability distribution of each interest node in the interest evolution network is calculated. The real-time product supply flow of the e-commerce platform is collected. The real-time product supply flow includes a set of products and their associated environmental attribute tags. The environmental attribute tags are mapped to the interest evolution network. The matching strength between each product in the product set and the interest node is evaluated. The matching strength is weighted and corrected according to the activity probability distribution to generate a candidate product list with time weight. The step of weighting and correcting the matching strength based on the activity probability distribution to generate a candidate product list with time weights includes: Extract the probability values ​​corresponding to specific future time periods from the activity probability distribution; The probability value is then fused with the initial matching strength of the product and interest node; The products in the product set are reordered based on the fusion calculation results; Each sorted product is assigned a time-weighted coefficient that indicates the urgency of the recommendation. The ranking results are integrated with the time weight coefficient to generate the candidate product list with time weight; Also includes: An interest state comparison unit is constructed, which is used to calculate the stability index of the interest evolution network under different simulated intervention strategies in parallel. The simulated intervention strategy is formed by adding or suppressing specific commodity information flows into the network topology. The simulated intervention strategy that makes the stability index reach the optimal balance is selected as the intervention strategy to be executed. Based on the intervention strategy to be executed and the candidate product list with time weights, a set of product information push instructions containing push timing and push sequence is generated, and the product information push instructions are input into the recommendation interface of the e-commerce platform; The selection of the simulated intervention strategy that achieves the optimal balance of the stability index as the intervention strategy to be executed includes: The stability index includes the network structure cohesion index and the interest exploration index; Calculate the numerical combination of the network structure cohesion index and interest exploration index under different simulated intervention strategies; Locate the coordinate point corresponding to the numerical combination in a two-dimensional coordinate system; Calculate the Euclidean distance from each coordinate point to the preset ideal equilibrium point; The simulated intervention strategy associated with the coordinate point with the smallest Euclidean distance is selected as the intervention strategy to be executed.

2. The AI-based interest-based product recommendation method for e-commerce platforms as described in claim 1, characterized in that, The process of performing temporal correlation and pattern abstraction on the multi-dimensional behavior records to form a network topology structure containing multiple interest nodes and associated edges includes: Independent behavioral events are extracted from the multi-dimensional behavioral records and sorted by timestamp; Calculate the frequency and conditional probability of co-occurrence of different behavioral events within adjacent time windows; Behavioral events that meet the co-occurrence criteria are clustered into interest clusters; Each interest cluster is abstractly represented as an interest node; Based on the temporal and co-occurrence relationships between the behavioral events, directional and weighted association edges are established between the nodes of interest, thereby forming the network topology.

3. The AI-based interest-based product recommendation method for e-commerce platforms as described in claim 2, characterized in that... The deduction of the potential interest development trajectory corresponding to the user includes: Starting from the currently active interest node in the interest evolution network; Calculate the probability of interest state transition along the direction of the associated edge; By combining the content attributes of the virtual interactive events, the diffusion process of interest states along different paths is simulated; Record the new interest nodes activated during the diffusion process and their activation intensity; The paths whose activation intensity exceeds the threshold are connected to form the potential interest development trajectory.

4. The AI-based interest-based product recommendation method for e-commerce platforms as described in claim 3, characterized in that... The process of generating a set of product information push instructions, including push timing and push sequence, includes: Analyze the intervention strategy to be executed and identify the set of interest nodes that need to be strengthened in the strategy; Select the subset of products that have the highest correlation with the set of interest nodes from the list of candidate products with time weights; Based on the time weight coefficient of each product in the product subset, the corresponding push time interval for each product is determined; Within the push time interval, the specific push timing is calculated based on the user's historical online activity patterns; Arrange the product subsets according to the chronological order of the push events to form the push sequence; The push timing and push sequence are encoded into product information push instructions that can be recognized by the e-commerce platform's recommendation interface.

5. The AI-based interest-based product recommendation method for e-commerce platforms as described in claim 4, characterized in that... The method further includes: After executing the product information push instruction, the system monitors the secondary behavior sequence generated by the user in response to the pushed product information, and feeds the secondary behavior sequence back to the interest evolution network as a feedback signal, triggering the adjustment of parameters of the corresponding interest nodes and related edges in the interest evolution network, thus completing a complete recommendation and feedback iteration.

6. The AI-based interest-based product recommendation method for e-commerce platforms as described in claim 5, characterized in that... The monitoring of secondary behavioral sequences generated by the user in response to the pushed product information includes: Capture user behavior regarding exposure, clicks, browsing time, favorites, adding to cart, and purchases of pushed product information; The captured behaviors are encoded into a structured sequence of behavior type and timestamp according to the chronological order in which they occur. Calculate the time interval and behavior transition pattern between adjacent behaviors in the secondary behavior sequence; The time interval and behavior transfer pattern are taken as the core content of the feedback signal.

7. The AI-based interest-based product recommendation method for e-commerce platforms as described in claim 6, characterized in that... The step of feeding the secondary behavior sequence back to the interest evolution network as a feedback signal to trigger parameter adjustments to the corresponding interest nodes and associated edges in the interest evolution network includes: Map the secondary behavior sequence back to the network topology to identify the target interest nodes and associated edges affected by the sequence; Based on the behavior type, determine the direction and magnitude of adjusting the activity parameter of the target interest node; Based on the behavior transfer pattern, determine the direction and magnitude of the adjustment of the weight of the associated edge; The internal state parameters of the interest evolution network are updated by applying the adjustment direction and magnitude.

Images