User behavior prediction system and method based on spiking neural network

Abstract

The application provides a user behavior prediction system and method based on a pulse neural network, and relates to the field of user behavior prediction. The method comprises the following steps: collecting user historical interaction sequence data and real-time interaction flow data to generate a user behavior pulse sequence. A double-path pulse neural network prediction model is constructed by a bionic hippocampus scene memory path and a spatial navigation path in parallel. The short-term interest mode and the long-term preference mode are fused and conflict arbitrated to obtain a comprehensive user state feature vector and generate an initial prediction list. Interaction feedback data of the user on the initial prediction list is obtained and converted into a reinforcement signal in the form of a neural pulse. The correlation strength between the user behavior features and the item features in the double-path pulse neural network prediction model is adjusted in real time through a weight self-adaptive adjustment module. The initial prediction list is reordered and finely screened to generate a prediction recommendation list. The technical problem that the long-term stable preference and the short-term accidental interest of a user cannot be effectively distinguished is solved.

Description

Technical Field

[0001] This application relates to the field of user behavior prediction, and in particular to a user behavior prediction system and method based on spiking neural networks. Background Technology

[0002] With the rapid development of the internet industry, user behavior prediction has become a core technology for improving service accuracy and optimizing business value. E-commerce recommendation systems are a prime example, urgently requiring the analysis of historical user interactions and real-time behavior to accurately predict user needs and achieve personalized recommendations. Currently, existing user behavior prediction technologies are mostly based on traditional neural networks and collaborative filtering methods. While these can achieve basic behavior prediction and recommendation functions, they cannot effectively distinguish between long-term stable preferences and short-term casual interests when faced with the dynamic and diverse characteristics of user behavior. They generally mix historical user behavior data with real-time interaction data without explicitly separating and differentiating these two types of behavioral characteristics. This easily leads to the misclassification of temporary browsing curiosity or accidental clicks in a single session as genuine purchase needs, causing interest drift and significantly increasing prediction bias. This directly results in a disconnect between recommendation results and actual user needs, not only reducing the user experience but also triggering inventory mismatch risks in e-commerce scenarios. Overstocking of products corresponding to short-term casual interests while understocking products meeting long-term stable user needs severely impacts inventory turnover efficiency and business operational effectiveness. Summary of the Invention

[0003] This application provides a user behavior prediction system and method based on spiking neural networks, which solves the technical problem that existing technologies cannot effectively distinguish between users' long-term stable preferences and short-term casual interests.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a user behavior prediction method based on spiking neural networks includes: collecting historical user interaction sequence data and real-time interaction stream data, performing structured processing and feature encoding to generate user behavior pulse sequences. A dual-pathway spiking neural network prediction model is constructed, consisting of a biomimetic hippocampal episodic memory pathway and a spatial navigation pathway operating in parallel. The biomimetic hippocampal episodic memory pathway identifies and separates short-term temporal patterns in the user behavior pulse sequences, while the spatial navigation pathway extracts and reinforces long-term stable patterns. Short-term interest patterns and long-term preference patterns are fused and conflict-arbitrated to obtain a comprehensive user state feature vector, which is then searched and matched in an item library to generate an initial prediction list. User interaction feedback data on the initial prediction list is acquired and converted into reinforcement signals in the form of neural pulses. Based on the principle of long-term synaptic enhancement and inhibition, the reinforcement signals are adjusted in real-time by a weight adaptive adjustment module to adjust the synaptic connection weights between user behavior features and item features in the dual-pathway spiking neural network prediction model. Based on the synaptic connection weights, a user-item matching score matrix is ​​calculated, and the initial prediction list is reordered and refined to generate a predicted recommendation list.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the construction process of the dual-pathway spiking neural network prediction model specifically includes: constructing a spiking neural network prediction model comprising an input layer, a shared feature extraction layer, a dual-pathway separation processing layer, and an output layer. The shared feature extraction hidden layer consists of multiple layers of convolutional spiking neurons, used for preliminary spatiotemporal feature extraction of the input pulse sequence, and outputting a shared feature pulse map. The dual-pathway separation processing layer is constructed in parallel by a physically isolated biomimetic hippocampal contextual memory pathway subnetwork and a spatial navigation pathway subnetwork. The biomimetic hippocampal contextual memory pathway subnetwork uses spiking neurons with a first preset time constant range and is configured with a short-term synaptic plasticity mechanism to extract and separate short-term interest pattern vectors from the shared feature pulse map. The spatial navigation pathway subnetwork uses spiking neurons with a second preset time constant range, the second preset time constant being greater than the first preset time constant, and is configured with a long-term synaptic plasticity foundation and a pattern extraction mechanism to extract and strengthen long-term preference pattern vectors from the shared feature pulse map.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the process by which the biomimetic hippocampal episodic memory pathway identifies and separates short-term temporal patterns in user behavior pulse sequences specifically includes: inputting a shared feature pulse map into a temporally sensitive spiking neuron layer, performing pulse temporal pattern competition and clustering operations, and activating the neuron with the highest matching degree to the pulse pattern within the current window as the winning neuron. Temporally linking the activation events of consecutive winning neurons to form interest trajectories representing different short-term interest segments. Encoding the interest trajectories into primary short-term interest pattern vectors, refining and denoising the features using a spiking autoencoder, and outputting the short-term interest pattern vector.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the process of extracting and reinforcing long-term stable patterns in user behavior pulse sequences through spatial navigation pathways specifically includes: inputting a shared feature pulse map into a spatial cell spiking neuron layer; by simulating the co-encoding mechanism of hippocampal position cells and grid cells, jointly mapping user behavior timestamps and item type features to a virtual behavior space coordinate system to generate a behavior space dwell map; inputting the behavior space dwell map into a pulse spatiotemporal integral pooling layer for nonlinear accumulation and thresholding, outputting a long-term dwell heatmap pulse map; inputting the long-term dwell heatmap pulse map into a long-term synaptic plasticity feature consolidation layer; using a learning rule based on pulse temporal dependent plasticity to generate a consolidated preference connection weight matrix; inputting the consolidated preference connection weight matrix into a preference prototype pulse self-organizing mapping layer; and clustering into several long-term preference prototype centers through competitive learning; extracting the activation intensity of the long-term preference prototype centers; and decoding and generating a long-term preference pattern vector by combining the consolidated preference connection weight matrix.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves fusing and arbitrating conflicts between short-term interest patterns and long-term preference patterns to obtain a comprehensive user state feature vector. Specifically, this includes: calculating the cosine similarity between the short-term interest pattern vector and the long-term preference pattern vector as an interest consistency index; setting a consistency threshold based on the distribution of cosine similarity between short-term and long-term vectors in historical user behavior data; when the interest consistency index is higher than the consistency threshold, performing a weighted summation of the short-term interest pattern vector and the long-term preference pattern vector to generate a comprehensive user state feature vector; and when the interest consistency index is lower than the consistency threshold, determining it as an interest drift state, assigning a higher fusion weight to the long-term preference pattern vector, and generating a comprehensive user state feature vector carrying potential incidental interest markers.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the reinforcement signal, based on the principle of long-term synaptic enhancement and inhibition, adjusts the synaptic connection weights between user behavior features and item features in the dual-path spiking neural network prediction model in real time through a weight adaptive adjustment module. Specifically, this includes: constructing a pulse-time-dependent plastic synaptic update circuit containing parallelized long-term enhancement and inhibition computation branches; receiving and parsing the reinforcement signal to obtain pulse intensity values; and reading the target synaptic weight set associated with user and item identifiers from the synaptic weight repository based on the sign and magnitude of the pulse intensity values; synchronously activating the long-term enhancement or inhibition computation branch based on the pulse intensity values; and using pipelined multiply-accumulate units to calculate in parallel the update amounts of all weight elements in the target synaptic weight set within one clock cycle; and writing the update amounts of all weight elements back to the synaptic weight repository to perform real-time updates of the synaptic connection weights between the user behavior feature dimension and the item feature dimension in the dual-path spiking neural network prediction model.

[0010] In conjunction with the first aspect mentioned above, one possible implementation involves acquiring user interaction feedback data on the initial prediction list and converting it into reinforcement signals in the form of neural impulses. Specifically, this includes: capturing user interaction events on items in the initial prediction list, including at least clicks, long presses, favorites, adding to cart, immediate purchase, ignoring, and page closing. Based on a preset event value mapping rule, a basic impulse intensity value is assigned to each type of interaction event, with positive feedback assigned a positive value and negative feedback assigned a negative value. The basic impulse intensity value is modulated by combining the occurrence time of the interaction event with the interface context information to generate a final impulse intensity. When the absolute value of the final impulse intensity is greater than a preset threshold, a pulse signal is generated at the time point corresponding to the interaction event, and the amplitude of the pulse signal is equal to the final impulse intensity. The pulse signal is associated with the corresponding user ID and item ID to form a reinforcement signal unit.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves collecting historical user interaction sequence data and real-time interaction stream data, performing structured processing and feature encoding to generate a user behavior pulse sequence. Specifically, this includes: parallel collection of historical interaction sequence data and real-time interaction stream data. The historical interaction sequence data includes the user's past purchase, favorite, and rating records across multiple periods, along with corresponding product features. The real-time interaction stream data includes page browsing, click events, and corresponding interaction features within the current session. The historical interaction sequence data is periodically segmented, and a steady-state feature vector is extracted within each period. This steady-state feature vector is a low-frequency pulse sequence and includes interest diversity entropy, average consumption level, and interaction depth score. The real-time interaction stream data is segmented using a sliding window, and a transient feature vector is extracted within each window. This transient feature vector is a high-frequency pulse sequence and includes attention shift rate, browsing stability coefficient, and interaction fineness score. The low-frequency pulse sequence and high-frequency pulse sequence are time-aligned and pulse-fused to generate a unified user behavior pulse sequence.

[0012] In conjunction with the first aspect mentioned above, one possible implementation involves reordering and refining the initial prediction list to generate a predicted recommendation list. Specifically, this includes: calculating a comprehensive recommendation score for each item in the initial prediction list based on the user-item matching score matrix. This comprehensive recommendation score includes a long-term preference score, a short-term interest score, and a time decay factor. The items in the initial prediction list are then sorted in descending order according to their comprehensive recommendation scores to form a reordered list. A dynamic threshold is set to filter out items in the reordered list whose comprehensive recommendation scores are below the dynamic threshold, retaining only those above the threshold to form the predicted recommendation list.

[0013] Secondly, a user behavior prediction system based on a spiking neural network is provided, including: a communication unit, a prediction analysis module, a fusion module, a feedback adjustment module, and a prediction generation module. The communication module collects historical user interaction sequence data and real-time interaction stream data, performs structured processing and feature encoding, and generates user behavior pulse sequences. The prediction analysis module constructs a dual-pathway spiking neural network prediction model formed by the parallel development of a biomimetic hippocampal episodic memory pathway and a spatial navigation pathway. The biomimetic hippocampal episodic memory pathway identifies and separates short-term temporal patterns in the user behavior pulse sequences, while the spatial navigation pathway extracts and reinforces long-term stable patterns. The fusion module fuses and arbitrates conflicts between short-term interest patterns and long-term preference patterns to obtain a comprehensive user state feature vector, which is then searched and matched in an item library to generate an initial prediction list. The feedback adjustment module acquires user interaction feedback data on the initial prediction list and converts it into reinforcement signals in the form of neural pulses. Based on the principle of long-term synaptic enhancement and inhibition, the reinforcement signals are adjusted in real-time by a weight adaptive adjustment module to adjust the synaptic connection weights between user behavior features and item features in the dual-pathway spiking neural network prediction model. The prediction generation module calculates the user and project matching score matrix based on synaptic connection weights, reorders and refines the initial prediction list, and generates a prediction recommendation list.

[0014] This application provides a user behavior prediction system and method based on spiking neural networks. By constructing a dual-pathway spiking neural network prediction model that combines a biomimetic hippocampal contextual memory pathway and a spatial navigation pathway, it can explicitly separate and differentiate user historical behavior data and real-time interaction data. Long-term stable preferences are encoded as low-frequency pulse sequences extracted and reinforced by the spatial navigation pathway, while short-term incidental interests are encoded as high-frequency pulse sequences identified and separated by the contextual memory pathway. This achieves accurate differentiation between long-term and short-term behavior patterns, avoiding the misjudgment of temporary browsing or accidental clicks in a single session as genuine purchase needs. It eliminates the incentive for interest drift at the mechanism level, improves the accuracy of behavior prediction, and ensures that recommendations accurately match users' true core needs. It reduces ineffective recommendations based on short-term incidental interests, allowing users to receive recommendations that better align with their long-term consumption preferences. This also prevents businesses from overstocking products corresponding to short-term incidental interests, significantly reducing inventory mismatch risk and improving overall inventory turnover efficiency. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the user behavior prediction method based on a spiking neural network provided in this application embodiment; Figure 2 A flowchart illustrating the user behavior prediction method based on a spiking neural network provided in this application embodiment; Figure 3A flowchart illustrating the user behavior prediction method based on a spiking neural network provided in this application embodiment; Figure 4 A flowchart illustrating the user behavior prediction method based on a spiking neural network provided in this application embodiment; Figure 5 A flowchart illustrating the user behavior prediction method based on a spiking neural network provided in this application embodiment; Figure 6 A flowchart illustrating the user behavior prediction method based on a spiking neural network provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of a user behavior prediction system based on a spiking neural network provided in an embodiment of this application; Figure 8 An interaction diagram of a user behavior prediction system based on a spiking neural network provided in this application embodiment. Detailed Implementation

[0016] like Figure 1 As shown in the embodiments of this application, the user behavior prediction method based on a spiking neural network includes: Step 101: Collect user historical interaction sequence data and real-time interaction stream data, and perform structured processing and feature encoding to generate user behavior pulse sequences.

[0017] Specifically, this includes: parallel acquisition of historical interaction sequence data and real-time interaction stream data; historical interaction sequence data includes user purchase, collection, and rating records from multiple past periods, as well as corresponding product characteristics; real-time interaction stream data includes page browsing, click events, and corresponding interaction features within the current session; periodically segmenting the historical interaction sequence data and extracting steady-state feature vectors within each period, the steady-state feature vectors being low-frequency pulse sequences, including interest diversity entropy, average consumption level, and interaction depth score; performing sliding window segmentation on the real-time interaction stream data and extracting transient feature vectors within each window, the transient feature vectors being high-frequency pulse sequences, including attention shift rate, browsing stability coefficient, and interaction refinement score; and time-aligning and pulse-fusion of the low-frequency pulse sequences and high-frequency pulse sequences to generate a unified user behavior pulse sequence.

[0018] In some implementations, historical interaction sequence data and real-time interaction stream data are obtained from the data source through two independent acquisition threads: A thread queries a batch of a specified user's historical interaction records over the past 12 months from a distributed user behavior database (such as a Cassandra cluster deployed on three shards). Each record contains a timestamp, behavior type (such as purchase, favorite), product ID, category tag, price, and interaction duration. Another thread subscribes to a message queue and listens in real time to the page event stream generated by the user's current session. Each event includes the event type (such as browsing, clicking), page URL, block coordinates, dwell time, and a sampled sequence of mouse trajectory points.

[0019] For historical interaction sequence data, the timeline can be divided into time periods according to natural weeks. For each time period, the frequency of occurrence of each product category is counted, the frequency distribution is calculated, and then the interest diversity entropy value is calculated using the Shannon entropy formula. Simultaneously, a weighted average of the price field of all records within the time window is calculated (the weight is the reciprocal of the interaction duration) to obtain the average consumption level. Finally, the interaction duration of each record within the time window is multiplied by an exponential decay factor (the decay base is the time difference between the record's timestamp and the end of the time window), and then summed to obtain the interaction depth score. At the end of each time window, the three-dimensional vector composed of (interest diversity entropy value, average consumption level, and interaction depth score) is passed through a fully connected layer time encoder, and a low-frequency pulse is output, which is the steady-state feature vector.

[0020] For real-time interactive stream data, the current system time is used as a baseline, and all events within the previous 60 seconds are taken as a sliding window, sliding once every 10 seconds. Within each window, the coordinate sequences of all page blocks are extracted, the vector difference between adjacent coordinates is calculated, and the attention shift rate is calculated using the Fletcher-Reeves gradient algorithm. Simultaneously, all dwell times are extracted, and the ratio of their standard deviation to the mean is calculated to obtain the browsing stability coefficient. Finally, the mouse trajectory point sequence is extracted, and the Frieze distance is calculated with a standard straight line trajectory to obtain the interaction refinement score. Whenever a new event occurs within the window (such as a click), the event-driven encoder immediately converts the current window's (attention shift rate, browsing stability coefficient, interaction refinement score) vector into a high-frequency pulse using a spiking neuron model (e.g., the Leaky Integrate-and-Fire model), which becomes the transient feature vector.

[0021] A millisecond-precise reference clock is obtained from an internal NTP server. The emission times of both low-frequency pulses (originally based on the absolute time of the end of the time window) and high-frequency pulses (originally based on the event timestamp) are calibrated to this reference clock. After calibration, the two pulses are merged in chronological order. If two pulses from low and high frequencies are found at the same millisecond on the reference clock, their amplitudes are added together, and the emission time is taken as the midpoint of that millisecond to generate a fused pulse. The generated pulse sequence is then fed into a circular buffer queue, which operates in FIFO mode, with new pulses inserted from the tail and old pulses removed from the head.

[0022] Step 102: Construct a dual-pathway spiking neural network prediction model formed by the bionic hippocampal contextual memory pathway and the spatial navigation pathway in parallel. The bionic hippocampal contextual memory pathway identifies and separates short-term temporal patterns in the user behavior pulse sequence, while the spatial navigation pathway extracts and reinforces long-term stable patterns in the user behavior pulse sequence.

[0023] The dual-pathway spiking neural network prediction model is inspired by the structure of the hippocampus, employing a spiking neural network architecture composed of parallel episodic memory and spatial navigation pathways. The biomimetic hippocampal episodic memory pathway simulates the network branch in the hippocampus that processes short-term, sequential memory, used to identify and separate short-term temporal patterns from user behavior pulse sequences. The spatial navigation pathway simulates the network branch in the hippocampus that processes spatial and long-term memory, used to extract and reinforce long-term stable patterns from the same pulse sequence. Short-term temporal patterns refer to the fluctuations in interest and behavioral sequence characteristics exhibited by users within a short period (such as a single session). Long-term stable patterns refer to the stable preferences and habitual behavioral characteristics formed by users over a longer period.

[0024] In some implementations, a prediction model with two parallel pathways is constructed based on a spiking neural network: the episodic memory pathway uses a layer of spiking neurons with fast response and short-term memory characteristics to receive encoded high-frequency spiking sequences (representing real-time interaction streams), and captures and separates the user's short-term interest shifts and behavioral sequence patterns through temporal convolution and short-term synaptic plasticity mechanisms.

[0025] The spatial navigation pathway employs a spiking neuron layer with integral firing and steady-state maintenance characteristics to receive low-frequency pulse sequences (representing historical interaction sequences). Through slow-varying weights and long-term synaptic enhancement mechanisms, it extracts and reinforces the stable preferences and consistent behavioral characteristics exhibited by users over multiple cycles.

[0026] The two pathways output short-term interest vectors and long-term preference vectors respectively, which are then integrated in the subsequent fusion module to form a complete prediction framework with the two pathways working together.

[0027] Step 103: Merge and arbitrate the short-term interest pattern and long-term preference pattern to obtain a comprehensive user state feature vector, and perform retrieval and matching in the item library to generate an initial prediction list.

[0028] The comprehensive user state feature vector is a vector representation that represents the user's current comprehensive intent, obtained after fusion and arbitration. The item library refers to a database that stores the feature vectors of all products or items to be recommended.

[0029] In some implementations, short-term interest pattern vectors and long-term preference pattern vectors output from a dual-path network are received. The cosine similarity between these two vectors is calculated to determine interest consistency, and a preset similarity threshold is applied: if the similarity is higher than the threshold, user interests are considered consistent, and the two vectors are weighted and summed (weights can be preset empirically or dynamically adjusted) to directly generate a comprehensive user state feature vector; if the similarity is lower than the threshold, it is determined to be an interest conflict or "interest drift." In this case, the system assigns a higher fusion weight to the long-term preference pattern vector and adds a "potential incidental interest" label to the generated comprehensive user state feature vector. The comprehensive vector is then compared with the embedding vectors of all items in the item library using cosine similarity calculation. The top N items are selected from high to low similarity scores to form an initial prediction list. For the list corresponding to the comprehensive vector marked as "potential incidental interest," the system also introduces a dynamic decay factor to reduce the final priority of recommended items dominated by short-term incidental interests.

[0030] Step 104: Obtain user interaction feedback data on the initial prediction list and convert it into reinforcement signals in the form of neural impulses.

[0031] Specifically, this includes: capturing user interaction events with items in the initial prediction list, including at least clicks, long presses, favorites, adding to cart, immediate purchase, ignoring, and page closing. Based on a preset event value mapping rule, a base pulse intensity value is assigned to each type of interaction event, with positive feedback assigned a positive value and negative feedback assigned a negative value. The base pulse intensity value is modulated by combining the occurrence time of the interaction event with the interface context information to generate a final pulse intensity. When the absolute value of the final pulse intensity is greater than a preset intensity threshold, a pulse signal is generated at the time point corresponding to the interaction event, and the amplitude of the pulse signal is equal to the final pulse intensity. The pulse signal is then associated with the corresponding user ID and item ID to form a reinforcement signal unit.

[0032] In some implementations, user interactions with the recommendation list can be captured in real time directly through front-end monitoring components, such as when a user clicks on a product. According to a preset event value mapping rule, the "click" action is mapped to a positive base pulse intensity value (e.g., +0.3). The timestamp of the click event and the current interface context (e.g., page type, surrounding product information) are obtained. A time decay factor is calculated based on the timestamp (e.g., earlier events decay more). A context relevance factor is calculated based on the interface context (e.g., clicks occurring in the current page's focus area have higher relevance). The base pulse intensity value is then multiplied by these two factors to obtain the final pulse intensity. If the absolute value of this intensity exceeds a preset intensity threshold (an initial threshold is calculated using a multi-objective optimization algorithm based on historical feedback signal intensity distribution, system processing capacity limitations, and signal effectiveness analysis, and an online adaptive mechanism is established to dynamically adjust the threshold based on actual processing results; e.g., 0.1), a pulse signal is generated at the time of the event (with slight delay calibration), and its amplitude is the calculated final pulse intensity. Finally, this pulse is bound to the user ID and product ID that triggered it and encapsulated into a reinforcement signal unit.

[0033] The event value mapping rule extracts complete user interaction records for recommended items within a specified time period (e.g., the past 6 months) from the historical log database. For each event type (e.g., click, favorite, purchase, ignore), it calculates the proportion of events that trigger target conversion behaviors (such as placing an order or repeat purchase) within a preset time window (e.g., 7 days) after their occurrence. This proportion serves as the initial conversion tendency value for that event type. Based on the core business objectives of the current recommendation system (e.g., increasing GMV, enhancing user stickiness, and promoting new product discovery), a weighting coefficient is assigned to each business objective. For each event type, the correlation strength between it and each business objective is analyzed to calculate a business value coefficient. The business value coefficient is obtained through expert scoring or multi-objective optimization algorithms.

[0034] This process involves weighting and integrating the initial conversion tendency value with the business value coefficient, and then normalizing it to ensure that the mapping values ​​for all event types fall within a preset range (e.g., [-1, 1]), forming an initial table of event value mapping rules. The value range for positive events (e.g., purchase, favorite) is (0, 1], for negative events (e.g., ignore, bad review) it is [-1, 0), and for neutral events (e.g., browse) it is close to 0. The above steps are repeated periodically (e.g., weekly), and the rules are updated based on newly generated data to adapt to the evolution of user behavior patterns and business objectives.

[0035] Step 105: Strengthening the signal. Based on the principle of long-term synaptic enhancement and inhibition, the weights of synaptic connections between user behavior features and project features in the dual-pathway spiking neural network prediction model are adjusted in real time through the weight adaptive adjustment module.

[0036] Synaptic long-term enhancement (LTP) refers to the persistent increase in synaptic connection strength, simulating the process of positive feedback (such as clicking or purchasing) from a user to a recommended item. This is achieved by updating association weights mathematically to strengthen the positive match between user behavior features and the item. Synaptic long-term inhibition (LTD) refers to the persistent decrease in synaptic connection strength, simulating the process of negative feedback (such as ignoring or closing) or detecting conflicting interest signals. This is achieved by updating weights mathematically to suppress irrelevant or misleading feature associations. The weight adaptive adjustment module is a programmable computing unit integrated into the dual-pathway spiking neural network. It receives reinforcement signals in the form of neural pulses and, according to preset LTP / LTD mathematical rules, calculates weight adjustments in real time and updates the synaptic connection matrix in the network, enabling online and dynamic evolution of model parameters. Association strength refers to the degree of matching between user behavior feature vectors and item feature vectors, quantified by synaptic connection weights, in the dual-pathway spiking neural network prediction model. It is represented by a real value, with higher positive values ​​indicating stronger associations.

[0037] In some implementations, the weight adaptive adjustment module, upon receiving the reinforcement signal derived from real-time user feedback, directly parses the user ID, item ID, and pulse intensity value carried in the signal. The pulse intensity value is dynamically determined by the event value mapping rules, with positive feedback corresponding to a positive pulse and negative feedback to a negative pulse.

[0038] This allows for the direct triggering of different synaptic plasticity simulation processes based on the sign and magnitude of the pulse intensity: if the pulse intensity is positive and exceeds a preset activation threshold, then LTP simulation calculation is initiated, i.e., based on the pulse intensity value and the time decay factor, according to... Formula for calculating weight increment (in For learning rate, Pulse intensity, 1 represents the signal delay time. 1 is the attenuation constant), and this increment is added to the synaptic weights corresponding to the user behavior feature dimension and the project feature dimension; if the pulse intensity is negative and the absolute value exceeds the inhibition threshold, then LTD simulation calculation is started, according to... Formula for calculating weight reduction (in Negative pulse intensity As the current weight, This serves as the minimum weight lower bound to prevent excessive weight decay.

[0039] The calculated weight adjustments are applied in real time to the synaptic connection matrices of the context memory pathway and the spatial navigation pathway of the dual-path spiking neural network to update the correlation strength between user behavior features and item features.

[0040] Step 106: Based on the user-item matching score matrix, calculate the user-item matching score matrix, reorder and refine the initial prediction list, and generate a prediction recommendation list.

[0041] Specifically, this includes: receiving a user-item matching score matrix between user behavior features and item features. Each element in the user-item matching score matrix represents the synaptic connection strength between a certain feature dimension of the user behavior pulse sequence and a certain feature dimension of an item in the item library in the dual-path spiking neural network prediction model. Based on the user-item matching score matrix, a comprehensive recommendation score is calculated for each item in the initial prediction list. The comprehensive recommendation score includes a long-term preference score, a short-term interest score, and a time decay factor. The long-term preference score is obtained by normalizing the dot product of the long-term preference pattern vector output by the spatial navigation pathway and the item feature vector using the Sigmoid function. The short-term interest score is obtained by normalizing the dot product of the short-term interest pattern vector output by the bionic hippocampal episodic memory pathway and the item feature vector using the Softmax function. The time decay factor is calculated by decaying the difference between the "last interaction timestamp" in the item feature vector and the current system time using a negative exponential function. The items in the initial prediction list are sorted in descending order according to the comprehensive recommendation score to form a reordered list. A dynamic threshold is set, which is adjusted in real time based on the number and type of interaction events that have occurred in the current user session: if positive interaction events account for more than 60% in the current session, the threshold is raised to increase the strictness of the screening; if negative interaction events account for more than 40%, the threshold is lowered to expand the recall scope. Items in the re-ranked list with an overall recommendation score lower than the dynamic threshold are filtered out, and items with a score higher than the threshold are retained to form a predicted recommendation list.

[0042] In some implementations, after receiving reinforcement signals generated by real-time user feedback, the corresponding elements in the preference matching degree matrix are updated in real time based on the principle of long-term synaptic enhancement and inhibition. For example, when a user clicks on an item, the matrix element value between the corresponding user behavior feature dimension and the feature dimension of the item will increase by an increment. If the user ignores an item, the corresponding element value will decrease.

[0043] The short-term interest pattern vector and long-term preference pattern vector at the current moment are extracted from the dual-path spiking neural network prediction model, and then the dot product operation is performed with the feature vector of each item in the item library to obtain the original interest matching value. The long-term preference matching value is converted into a long-term preference score through the Sigmoid function, and the short-term interest matching value is converted into a short-term interest score through the Softmax function. At the same time, the "last interaction timestamp" is extracted from the item features, the time difference with the current time is calculated, and the time decay factor is obtained by substituting it into the decay formula. The long-term preference score, short-term interest score and time decay factor are weighted and summed according to the preset weight to obtain the comprehensive recommendation score of each item.

[0044] Based on the types of interaction events that have occurred in the current session, the proportion of positive and negative events is calculated. If the proportion of positive events exceeds 60%, the dynamic threshold is set to the upper quartile of the overall recommendation score distribution. If the proportion of negative events exceeds 40%, the dynamic threshold is set to the median. Otherwise, it is set to the lower quartile. Subsequently, the items in the initial prediction list are sorted in descending order of their overall recommendation scores, and items with scores below the dynamic threshold are removed to form the predicted recommendation list.

[0045] Based on the above technical solution, by collecting historical and real-time user interaction data in parallel, and generating a unified behavioral pulse sequence through structured processing and high- and low-frequency pulse fusion, the problem of fragmented behavioral feature extraction and insufficient consideration of both temporal sequence and real-time performance in traditional methods can be solved at the data layer, laying a comprehensive and accurate feature data foundation for behavioral prediction. Furthermore, by using a biomimetic hippocampal dual-pathway spiking neural network model to extract short-term interest and long-term preference patterns separately through two pathways, it can effectively separate and extract short- and long-term behavioral patterns during prediction generation, thus solving the technical challenge that a single model cannot simultaneously and accurately capture both fluctuating and stable user interests. This is achieved in prediction generation. In conjunction with pattern fusion and conflict arbitration, the comprehensive feature vector addresses the problem of feature fusion distortion when short- and long-term interests conflict, improving the accuracy of the initial prediction list for item library retrieval and matching. Simultaneously, user interaction feedback is converted into neural pulse reinforcement signals, and the model's association weights are adjusted in real-time based on the principle of long-term synaptic enhancement and inhibition, solving the problem of lagging model parameter updates and achieving online adaptive optimization in model management. Finally, the initial list was reordered based on the updated association strength and refined using dynamic thresholds. This solved the problem of insufficient adaptability and accuracy of the recommended list, improved the real-time performance of user behavior prediction and the relevance of the recommendation results, and made behavior prediction and recommendation more in line with the user's real needs.

[0046] In another possible implementation of the embodiments of this application, combined with Figure 1-2 As shown, the construction process of the dual-path spiking neural network prediction model can be achieved through the following steps 201 to 203, which are explained in detail below: Step 201: Construct a spiking neural network prediction model that includes an input layer, a shared feature extraction layer, a dual-pathway separation processing layer, and an output layer. The shared feature extraction hidden layer is composed of multiple convolutional spiking neurons, which are used to perform preliminary spatiotemporal feature extraction on the input spiking sequence and output a shared feature spiking map. The dual-pathway separation processing layer is composed of a biomimetic hippocampal context memory pathway subnetwork and a spatial navigation pathway subnetwork that are physically isolated.

[0047] The input layer is the external interface for the network to receive raw data. Its number of neurons matches the dimension of the input features, and it is responsible for converting external input signals (such as user behavior encoding) into an initial pulse firing pattern. The shared feature extraction layer consists of one or more hidden layers following the input layer, typically composed of convolutional spiking neurons. Its shared nature is reflected in the fact that the weights and computational patterns of this layer are reused by the subsequent parallel dual-path processing, avoiding redundant computation. A convolutional spiking neuron is a composite unit combining convolutional computation (for extracting local patterns in spatial or feature dimensions) and spiking neuron dynamics (for processing temporal information). It performs convolution and nonlinear pulse firing on the input pulse sequence, and the output can be considered a shared feature pulse map with significant features in the spatiotemporal dimensions. The dual-path separation processing layer is a parallel processing module composed of two sub-networks (i.e., the episodic memory pathway and the spatial navigation pathway) that are specialized in functional objectives, internal structure, and processing mechanisms and are physically or logically isolated. This layer receives the same feature pulse map from the shared feature extraction layer, but the two sub-networks extract and reinforce different types of behavioral patterns from it.

[0048] In some implementations, the input layer sets the number of input neurons according to the dimension of the encoded user behavior pulse sequence. The shared feature extraction layer is designed as a 1 to 3-layer convolutional spiking neural network. For example, the first layer uses a small convolutional kernel to slide along the time dimension to capture local temporal patterns, while the second layer may use a slightly larger convolutional kernel or perform convolution along the feature dimension to integrate relevant features. Each convolutional layer is followed by a spiking neuron pool, whose membrane potential integration and threshold firing mechanism convert the convolution result into a new pulse sequence, ultimately outputting a shared feature pulse map that retains key spatiotemporal features.

[0049] The dual-pathway separation processing layer is constructed as two sub-networks, both taking the feature map as input but processing it differently: the episodic memory pathway sub-network typically uses spiking neurons with small time constants and fast responses (such as LIFs), and may introduce short-term synaptic plasticity, making it sensitive to rapid changes and short-term correlations in the spiking sequence, thereby separating short-term interest patterns; while the spatial navigation pathway sub-network uses spiking neurons with larger time constants and adaptive thresholding or integral characteristics (such as AdaptiveLIFs), and focuses on extracting stable statistical features across time windows through slow integration and steady-state maintenance, thereby reinforcing long-term preference patterns. To achieve physical structural isolation, independent memory spaces and computation threads are allocated to the two sub-networks, or they are mapped to different cores or regions of the neuromorphic chip during hardware deployment.

[0050] During its training, the shared feature extraction layer can be pre-trained using unsupervised learning methods (such as pulse-based Hebb learning) to learn the basic representation of user behavior data. Then, by combining labeled data (such as the user's subsequent click behavior), the entire network, including the dual-path network, can be trained end-to-end through pulse time-dependent plasticity rules or supervised pulse error backpropagation algorithms to optimize the network parameters.

[0051] Step 202: The biomimetic hippocampal episodic memory pathway subnetwork uses spiking neurons with a first preset time constant range and is configured with a short-term synaptic plasticity mechanism to extract and separate short-term interest pattern vectors from shared feature pulse maps.

[0052] The first preset time constant range refers to the membrane potential decay time constant set for spiking neurons in the episodic memory pathway subnetwork. Its value is typically small, ranging from 10 to 100 milliseconds, thus determining the neuron's rapid response to input pulses and its rapid "forgetting" characteristics, making it adept at capturing rapid changes and short-term temporal patterns in signals. The short-term interest pattern vector is a numerical vector that quantitatively and structurally represents the dynamic focus of interest and behavioral tendencies exhibited by a user within a short period (such as a single session or a few minutes).

[0053] When constructing the biomimetic hippocampal contextual memory pathway subnetwork, the spiking neuron model constituting the network is first selected and initialized. For example, the Leaky Integrate-and-Fire model is commonly used, and its membrane time constant is set to a small value within the first preset time constant range (such as 20 milliseconds to 50 milliseconds) to ensure rapid leakage of neuronal membrane potential, thereby making it more sensitive to high-frequency or recent input pulses and able to respond quickly to temporal changes.

[0054] A short-term synaptic plasticity mechanism (which can be implemented using the Tsodyks-Markram model or its simplified form) is configured on the synaptic connections of this network. By introducing variables to describe resource availability and their recovery time constants, the effectiveness of the synapse is dynamically adjusted according to the frequency and time interval of the pulse arrival: high-frequency pulse sequences cause temporary depletion of synaptic resources, resulting in short-term inhibition; while moderate intervals can trigger short-term enhancement, thereby temporarily changing the signal transmission strength to encode short-term patterns without changing the long-term weights.

[0055] Using the shared feature pulse map as input to this sub-network, during network operation, the fast-responding LIF neurons integrate and fire the input pulse map step by step. Meanwhile, the STP mechanism modulates the synaptic connectivity efficiency in real time based on the pulse activity history. These two mechanisms work together to enable the network to keenly capture and amplify feature combinations highly correlated with recent events in the input sequence, while suppressing irrelevant or background activity. After several layers of such processing (possibly including pulse convolution, pooling, etc.), the pulse firing pattern or average firing rate of the final layer of the network is encoded into a fixed-length vector, namely the short-term interest pattern vector. This vector centrally reflects the user's dynamic behavioral characteristics and interest orientation within the current short time window.

[0056] It should be noted that the specific value of the first preset time constant range needs to be determined experimentally based on the time granularity of user short-term behavior in actual application scenarios (such as the average interval of page browsing and session duration).

[0057] Step 203: The spatial navigation pathway subnetwork uses spiking neurons with a second preset time constant range, which is greater than the first preset time constant, and is configured with a long-term synaptic plasticity basis and pattern extraction mechanism to extract and enhance long-term preference pattern vectors from shared feature pulse maps.

[0058] The second preset time constant range refers to the membrane potential decay time constant set for spiking neurons in the spatial navigation pathway subnetwork. Its value is greater than the first preset time constant in the episodic memory pathway, typically set within the range of 500ms-2000ms. This increases the neuron's ability to slowly integrate and maintain potential changes over a longer period, enabling it to smooth noise, accumulate long-term statistical information, and maintain stable state representation. The long-term preference pattern vector is a numerical vector that can generally and stably represent a user's relatively persistent interests, consumption habits, or behavioral preferences over a longer period (e.g., days, weeks, or months), used to characterize the user's steady-state characteristics. The long-term synaptic plasticity basis refers to the persistent ability of synaptic connection weights in the network to change based on Heb's learning rule or its variants (e.g., pulse temporal-dependent plasticity). Its core is that when there is a strong correlation between the firing of presynaptic and postsynaptic neurons, the connection between them will be persistently strengthened or weakened. This long-term change in weights is the physical basis for the network's learning and memory of long-term stable patterns.

[0059] When constructing the space navigation pathway subnetwork, first select and initialize its spiking neuron model. For example, the Adaptive Exponential Integrate-and-Fire model or the classic LIF model with a large membrane time constant can be used. Set the time constant within the range of the second preset time constant (such as 500 ms to 2000 ms) so that the neuron can slowly integrate the input pulse, smoothing out short-term fluctuations like a low-pass filter, thereby generating a cumulative response to continuous or recurring feature signals.

[0060] The pulse-time-dependent plasticity rule is typically used as the core learning law: when the presynaptic neuron pulse occurs before the postsynaptic neuron pulse, the connection weight is enhanced; conversely, it is weakened. This time-dependent weight adjustment mechanism enables the network to learn and consolidate recurring long-term patterns with causal or statistical correlation in the input signal, thus providing a long-term synaptic plasticity basis for the network's synapses.

[0061] Furthermore, a pattern extraction mechanism was designed and implemented. For example, a pulse self-organizing map layer based on competitive learning was introduced in the last layer of the network. This layer contains a set of output neurons, each representing a potential long-term preference prototype. When an input pulse signal (from the output of the slow integral neural layer) is input, the weights of the "winning neuron" in the SOM layer that best matches the pattern and its neighboring neurons are adjusted in the direction of the input pattern according to rules such as STDP. After training with a large amount of historical data, the weight vectors of these output neurons converge to different long-term preference pattern prototypes.

[0062] During operation, a shared feature pulse map is input into the pathway, and slow integral neurons accumulate and smooth their pulse activity over time to generate potential signals that reflect long-term trends. These signals are transmitted through connections with STDP plasticity, driving competition and learning in the SOM layer. Finally, the index of the winning neuron or its weight vector (or some kind of encoding of the response pattern of the entire SOM layer) is output as a long-term preference pattern vector.

[0063] Based on the above technical solution, the constructed four-layer spiking neural network prediction model allows for preliminary spatiotemporal feature extraction of the input pulse sequence through a shared feature extraction layer composed of multiple convolutional spiking neurons, outputting a shared feature pulse map. This avoids redundant calculations in subsequent feature extraction via dual pathways, improving feature processing efficiency, and solves the problem of incomplete capture of spatiotemporal features of user behavior by traditional models, laying a unified and accurate feature foundation for subsequent pattern analysis. Furthermore, a physically isolated dual-pathway separation processing layer addresses the core problem of traditional methods' inability to explicitly separate long-term and short-term user behavior patterns. The contextual memory pathway, using small-time-constant spiking neurons combined with short-term synaptic plasticity, accurately separates short-term interest patterns from the shared feature pulse map. The spatial navigation pathway, using larger-time-constant spiking neurons combined with long-term synaptic plasticity and pattern extraction mechanisms, effectively strengthens long-term preference patterns, achieving differentiated and accurate analysis of long-term and short-term behavior patterns. Furthermore, the targeted structural design of the dual-pathway approach eliminates the incentives for interest drift at the model level, solves the technical challenge that a single model cannot simultaneously capture user interest fluctuations and stable preferences, improves the accuracy of behavior pattern extraction, and provides high-quality feature vector support for subsequent pattern fusion and user behavior prediction, ensuring that the prediction results match the user's real needs.

[0064] In another possible implementation of the embodiments of this application, combined with Figure 1-3 As shown, the process by which the biomimetic hippocampal episodic memory pathway identifies and separates short-term temporal patterns in user behavior pulse sequences can be achieved through the following steps 301 to 303, which are explained in detail below: Step 301: Input the shared feature pulse map into the time-sensitive spiking neuron layer, perform pulse temporal pattern competition and clustering operations, and activate the neuron with the highest pulse pattern matching degree in the current window as the winning neuron.

[0065] In some implementations, when inputting a time-sensitive spiking neuron layer with a shared feature pulse map, the input is segmented according to a preset sliding window length (e.g., 50 milliseconds) and step size (e.g., 25 milliseconds). Within each window, a dynamic time warping algorithm is used to align and compare the similarity of two pulse time series. The matching degree between the pulse pattern (composed of the pulse occurrence time and corresponding feature dimension) within the current window and the temporal template represented by each neuron in the layer is calculated, triggering a lateral inhibition competition mechanism: initially, all neurons responding to the current input acquire an activation potential, the magnitude of which is proportional to the matching degree; subsequently, the neuron with the highest activation potential (i.e., the winning neuron with the highest matching degree) is formally activated and fires an output pulse. Simultaneously, inhibitory signals are sent to other neurons through pre-established inhibitory connections within the layer (e.g., each neuron applies a fixed negative weight to other neurons except itself), rapidly suppressing their activation potentials and preventing multiple neurons from responding to the same pattern simultaneously. Finally, the firing of the winning neuron signifies that the dominant short-term interest pattern within the current window has been identified and selected. The entire process is repeated on each sliding window, which transforms a continuous stream of input pulses into a series of discrete sequences of winning neuron activation events representing different short-term interest segments.

[0066] Step 302: Temporally link consecutive winning neuron activation events to form interest trajectories representing different short-term interest segments.

[0067] In some implementations, winning neuron activation events are received sequentially in a streaming manner. For each newly arriving event, the system maintains an active trajectory pool containing emerging interest trajectories (each trajectory includes the neuron ID, timestamp, and a summary vector representing the trajectory pattern, such as the mean ID vector of historical events). The time difference between the new event and each trajectory in the active trajectory pool is calculated (new event timestamp minus the timestamp of the last event on the trajectory). For events with a time difference less than a time interval threshold... For each trajectory, calculate its pattern similarity. If there exists a similarity greater than or equal to the pattern similarity threshold... If a trajectory is found to be similar to another trajectory, the trajectory with the highest similarity is selected, the new event is appended to the end of that trajectory, and the summary vector of that trajectory is updated (e.g., using a moving average method). If no trajectory meets the criteria, a new interest trajectory is created starting with the new event and added to the active trajectory pool.

[0068] In another possible implementation, a trajectory timeout period is set. (For example, 2 seconds), if a certain trajectory is in If no new events are added within a certain time, the interest segment is considered to have ended, and it is removed from the active trajectory pool and output as a complete interest trajectory.

[0069] It should be noted that the time interval threshold Pattern similarity threshold Adjustments need to be made based on the typical speed and pattern stability of users' short-term interest switching in actual business scenarios.

[0070] Step 303: Encode the interest trajectory into a primary short-term interest pattern vector, refine the features and denoise the vector using an impulse autoencoder, and output the short-term interest pattern vector.

[0071] The primary short-term interest pattern vector is a preliminary vectorized representation of an interest trajectory, which usually includes the trajectory's statistical characteristics (such as duration and average impulse firing rate) and pattern characteristics (such as the distribution of winning neuron IDs in the trajectory and a certain encoding of the temporal sequence).

[0072] In the model building and training phase, a spiking autoencoder network consisting of an encoder and a decoder is constructed. The encoder is composed of several layers of spiking neurons (such as LIF neurons), which maps the high-dimensional spiking patterns of the input to a low-dimensional "bottleneck layer" spiking sequence or firing rate vector. The decoder is structurally approximately a mirror image of the encoder and is responsible for reconstructing the input from the bottleneck layer representation.

[0073] When training a pulse autoencoder, primary short-term interest pattern vectors generated from a large amount of historical user behavior data and processed through temporal linking can be directly used as training samples. Each sample (i.e., the primary vector of an interest trajectory) is converted into a format suitable for network input. For example, the vector elements are converted into a constant current injected into the corresponding neuron in the encoder input layer to drive it to generate pulse firing; or the vector is directly regarded as the pulse firing rate over a period of time. The training objective is to minimize the difference between the input pulse pattern and the pulse pattern reconstructed by the decoder output. Commonly used loss functions can be the mean squared error based on the pulse firing time difference or the mean squared error based on the firing rate. At the same time, sparsity constraints on the bottleneck layer representation (such as a KL divergence term) are usually added to the loss function to encourage the learning of more effective features.

[0074] The training process employs optimization algorithms suitable for spiking neural networks, such as variants of time backpropagation or unsupervised learning rules based on STDP. After training is complete, the encoder portion of the spiking autoencoder is fixed for online refinement.

[0075] In actual operation, whenever an interest trajectory is formed and encoded into a primary short-term interest pattern vector, this vector is immediately input into the encoder part of the pre-trained pulse autoencoder. The encoder then performs a nonlinear transformation and dimensionality reduction on the input, generating a low-dimensional, dense pulse firing pattern or firing rate vector at the bottleneck layer. The output vector of this bottleneck layer is the refined and denoised short-term interest pattern vector. For example, the primary short-term interest pattern vector of an interest trajectory might be a 10-dimensional vector, containing the trajectory length (2 seconds), average pulse rate (20Hz), and intensity values ​​in eight dimensions representing different browsing modes (such as "fast swipe," "image zoom," "view comments," and "add to cart"). This vector may contain noise, such as a brief high value in the "add to cart" dimension due to an accidental touch. Alternatively, this primary vector can be directly input into the pulse autoencoder encoder, and after nonlinear transformation and compression at the bottleneck layer, the refined output vector might be a 5-dimensional vector. In this new vector, the dimensions representing the stable patterns of "rapid swiping" and "viewing reviews" are preserved and may be slightly enhanced, while the noise signal representing accidental "accidental add-to-cart" is significantly suppressed or eliminated. Simultaneously, a new comprehensive dimension may be learned to represent the higher-level interest of "deep browsing of footwear products." This allows the short-term interest pattern vector to more clearly reflect the user's true, core short-term intent during that trajectory.

[0076] Based on the above technical solution, by introducing sliding window segmentation, dynamic time warping matching, and lateral inhibition competition mechanisms into the time-sensitive spiking neuron layer, the winning neuron with the highest matching degree within the current window is accurately selected. This solves the technical problems of fuzzy short-term pulse patterns and simultaneous responses of multiple patterns in traditional pattern recognition, and enables it to dominate the accurate localization of short-term interest patterns within a single window. Furthermore, by maintaining an active trajectory pool and combining time difference and pattern similarity thresholds to temporally link consecutive winning neuron activation events, and setting a timeout mechanism to terminate complete trajectories, the problem of discrete activation events failing to represent continuous changes in user interest is solved. At the trajectory construction level, discrete events are transformed into continuous interest segment trajectories, fully capturing the dynamic transfer process of short-term interests. Finally, after encoding the interest trajectory into a primary vector, a pre-trained spiking autoencoder performs nonlinear transformation and dimensionality reduction and denoising, solving the problems of noisy and redundant primary features. At the feature optimization level, a high-purity short-term interest pattern vector is output, providing accurate and effective feature support for subsequent long-short pattern fusion. Each step progressively improves the accuracy and effectiveness of short-term behavior pattern extraction.

[0077] In another possible implementation of the embodiments of this application, combined with Figure 1-4As shown, the process of extracting and enhancing long-term stable patterns in user behavior pulse sequences through spatial navigation pathways can be achieved through the following steps 401 to 405, which are explained in detail below: Step 401: Input the shared feature pulse map into the spatial cell pulse neuron layer. By simulating the co-encoding mechanism of hippocampal position cells and grid cells, the user behavior timestamp and item type features are jointly mapped to the virtual behavior space coordinate system to generate a behavior space dwell map.

[0078] The virtual behavior space coordinate system simulates the function of hippocampal place cells and grid cells encoding physical space. It jointly maps the timestamps of user behavior and item type features to a high-dimensional abstract space, allowing behavior trajectories to form a distribution with spatial topological relationships (these relationships are automatically formed through competitive learning between place cells and grid cells in a data-driven manner). The behavior space dwell map is a two-dimensional or three-dimensional pulse density map formed by projecting user behavior onto the virtual behavior space coordinate system. The coordinate axes represent the dimensions of the behavior space, and the pixel or voxel values ​​represent the user's dwell time or access frequency in that spatial region, characterized by pulse firing frequency or accumulated membrane potential. The spatial cell spiking neuron layer is a dedicated spiking neuron layer that simulates the encoding function of hippocampal place cells and grid cells. Its neurons have specific spatial receptive fields, enabling selective responses to specific regions of input features in the virtual space. The place cell encoding mechanism is implemented by constructing radial basis function neurons with learnable center points. Each neuron responds only to inputs in the virtual space whose distance from its center point is less than a threshold. The grid cell encoding mechanism introduces a periodic hexagonal grid response pattern, performs convolution operations between input features and multi-layer grid templates of different scales and orientations, and generates a pulse output with periodic spatial modulation. The template parameters of the grid cells are initialized by simulating the spatiotemporal continuity constraints of the hippocampal development process to ensure that the virtual space can be uniformly covered in the early stage of training.

[0079] In some implementations, the user behavior timestamp and item category features associated with each pulse in the shared feature pulse map are extracted. The timestamp is compressed using a logarithmic transformation and used as a distance scale. The item category features are mapped into a dense vector through an embedding layer, thus concatenating the two to form a two-dimensional or three-dimensional coordinate vector of the behavior point in the virtual space. This coordinate vector can then be directly input into the spatial cell pulse neuron layer. This layer consists of two sets of parallel computational sublayers: the position cell sublayer contains multiple radial basis function neurons with uniformly initialized center points. Each neuron calculates the Euclidean distance between the input coordinates and its center point, converts it into a membrane potential input through a negative exponential function, and fires a pulse when the membrane potential exceeds a threshold. The neuron center points are adjusted during training through unsupervised competitive learning, allowing different neurons to specialize in responses from different behavioral regions. The grid cell sublayer constructs multiple sets of hexagonal grid templates with different frequencies and orientations. The input coordinates are convolved with each set of grid templates in two dimensions to generate a grid phase response. The weights of each set of grid templates are pre-trained using self-organizing maps.

[0080] After the pulse output from the location cells and grid cells, it can be nonlinearly accumulated in the pooling layer according to the spatial region. That is, the pulse falling into the region is integrated by leakage in each preset spatial grid. The integrated current decays with time. When the accumulated membrane potential reaches the dynamic threshold, the pulse is released, and finally a behavioral spatial residence map representing the thermal distribution of the user's long-term residence is generated.

[0081] Step 402: Input the behavior space dwell map into the pulse spatiotemporal integral pooling layer for nonlinear accumulation and threshold distribution, and output the long-term dwell thermal pulse map.

[0082] The pulse-spatiotemporal integral pooling layer is a dedicated computational layer that simulates the spatiotemporal accumulation and threshold decision-making functions of biological neurons by using a spiking neuron array with leakage integral characteristics. Each neuron corresponds to a local region in the virtual behavior space and is responsible for accumulating the pulse signals input to that region in both the temporal and spatial dimensions. Nonlinear accumulation means that the neuronal membrane potential update process follows a nonlinear dynamic equation. The input pulse is converted into an excitatory / inhibitory postsynaptic potential through synaptic connections and is nonlinearly superimposed with the current membrane potential, rather than being a simple arithmetic summation. Threshold firing means that when the neuronal membrane potential exceeds a preset firing threshold after integration, the neuron generates an output pulse, and the membrane potential is reset to the resting potential or enters a relative refractory period. The long-term residence heatmap is the spatiotemporal pulse pattern output by the pulse-spatiotemporal integral pooling layer. The firing position of each pulse in the graph corresponds to a specific region in the virtual behavior space, and the pulse firing frequency represents the intensity of the user's long-term residence in that region, forming a heatmap that reflects the spatial distribution of the user's long-term preferences.

[0083] In some implementations, the behavioral spatial dwell map is first divided into multiple local receptive field regions according to a preset spatial grid resolution, and a leaky integral firing neuron is configured in each region. The membrane potential dynamics of the neuron are mathematically simulated using the Leaky Integrate-and-Fire model, and its membrane potential update formula is as follows:

[0084] in: The membrane time constant (used to determine the speed of neuron membrane potential response to input current and leakage rate): where The larger the value, the slower the membrane potential leakage, the longer the integration time window, and the better the neuron is at accumulating long-term, repetitive signals. The smaller the value, the faster the membrane potential leaks, making it sensitive to short-term inputs but unable to maintain long-term accumulation. It is the first derivative of the membrane potential with respect to time, representing the instantaneous rate of change of the membrane potential with time; This refers to the membrane potential value of the neuron at the current moment; It refers to the resting potential (the stable membrane potential value of a neuron when it has not received any input). This refers to membrane resistance; For input synaptic current.

[0085] When a pulse in the behavioral space residency map falls into the receptive field of a neuron, the pulse is converted into a current injection of a specified intensity via synaptic connection (the amplitude of the current is proportional to the pulse intensity). The temporal superposition of multiple pulses causes the membrane potential to rise nonlinearly; each neuron is configured with an independently adjustable membrane time constant. The time constant is uniformly set to the second preset range (500ms-2000ms), thereby achieving smooth filtering of short-term fluctuations and slow integration of long-term repetitive signals.

[0086] In another possible implementation, an adaptive firing threshold mechanism is introduced, which is the neuron's base threshold. Initialized to a globally uniform value, and dynamically adjusted based on the historical firing frequency of the neuron: the threshold for high-frequency firing neurons is set according to... Slowly increase the dosage to prevent over-excitement, among which This is the threshold adjustment amount; To adaptively adjust the step size coefficient; It refers to the average actual pulse firing frequency of the neuron over the past sliding time window (usually set to 5 to 10 base time constants). This refers to the target firing frequency. The threshold of neurons firing at low frequencies is lowered accordingly to increase sensitivity; when the membrane potential V(t) exceeds the current threshold... At that moment, the neuron immediately fires an output pulse, the timing of which is precisely recorded, and the membrane potential is reset to [value missing]. It then enters an absolute refractory period of 2 ms, during which it receives no input. Once the output pulses of all neurons are labeled according to the coordinates of their spatial receptive fields, a two-dimensional spatiotemporal pulse tensor is formed, which is the long-term residence thermodynamic pulse map. The membrane time constant mentioned above... Basic threshold Hyperparameters such as the adaptive adjustment coefficient α are determined by joint optimization of grid search and impulse temporal dependence plasticity rules using historical behavioral data with dwell time labels during the offline training phase.

[0087] Step 403: Input the long-term resident thermal pulse map into the long-term synaptic plasticity feature consolidation layer, and generate the consolidation preference connection weight matrix by adopting the learning rule based on pulse time-dependent plasticity.

[0088] Among them, the consolidated preference connection weight matrix is ​​a stable synaptic connection matrix generated by the layer after iterative updates of the impulse temporal dependence plasticity learning rule. Each element in the matrix represents the long-term consolidated association strength between the impulse pattern of a certain local region in the virtual behavior space and the preference prototype neuron, which is used to support the decoding and output of long-term preference patterns.

[0089] In some implementations, a long-term resident thermal pulse map is input to a long-term synaptic plasticity feature consolidation layer. This layer contains a set of plastic synaptic connections, each of which connects a spatial neuron output from a pulse spatiotemporal integral pooling layer to an input neuron in a subsequent preferred prototype pulse self-organizing map layer. Simultaneously, a basic weight value is initialized for each synaptic connection (the weight range is preset to [0,1], and the initial value is either randomly initialized using a uniform distribution or assigned according to the pre-training results of Heblin learning).

[0090] At each moment, the precise firing time of the pulse is recorded as it is transmitted through the presynaptic neuron in the long-term resident thermogram. When a postsynaptic neuron fires a pulse due to the accumulation of sufficient postsynaptic potential, the system records the firing time. Thus, it is possible to pass Calculate the time difference .

[0091] Based on the impulse-time-dependent plasticity learning rule, a double-exponential plasticity window is adopted: like And if it is less than the preset window length (usually 20ms-40ms), then pass. ,in The positive learning rate is set to 0.01-0.05. It is a positive time constant (default is 10ms-20ms).

[0092] like And if the absolute value is less than the preset window length, then ,in A negative learning rate, It is a negative time constant (default is 20ms-40ms).

[0093] The calculated Δw1 is added to the current weight w, and the result is truncated to the [0,1] interval to complete one weight update. When long-term user behavior data is repeatedly input, presynaptic and postsynaptic pulse pairs with stable causal relationships to the user's stable preference patterns appear frequently, causing the corresponding synaptic weights to continuously increase and eventually converge to near 1. Synapses corresponding to random noise or accidental interests, due to unstable pulse timing relationships, maintain low weights or gradually decay to 0. After multiple rounds of offline pre-training and online incremental updates, all synaptic weights tend to stabilize, and the output synaptic connection matrix at this point is the consolidated preference connection weight matrix.

[0094] Step 404: Input the consolidated preference connection weight matrix into the preference prototype impulse self-organizing map layer, and cluster it into several long-term preference prototype centers through competitive learning.

[0095] The preference prototype spiking self-organizing mapping layer is an output plane composed of spiking neurons arranged in a two-dimensional grid. It is a dedicated spiking computation layer that simulates the topological organization and competitive neural cluster encoding mechanisms in the biological cerebral cortex. Competitive learning is an unsupervised learning paradigm where all neurons within the layer compete to respond to the same input pattern. Only the neuron that best matches the input pattern and its topological neighborhood neurons are eligible for weight updates. The winning neuron is the neuron whose weight vector is closest to the input feature vector, has the highest firing rate, or the strongest response under the current input pattern, and is uniquely determined through lateral inhibition. The long-term preference prototype center is the convergent weight vector of the winning neuron and its neighboring neurons after repeated competitive learning. This vector forms several cluster centers in the virtual behavior feature space, each center representing a stable and distinguishable long-term user interest category.

[0096] In some implementations, this layer pre-constructs a two-dimensional planar grid consisting of N×N spiking neurons. The number of neurons N is preset to be between 10 and 20 according to business needs. The grid topology adopts hexagonal or rectangular neighborhood relationships. Each neuron maintains a weight vector with the same dimension as the input. The initial values ​​are initialized by random sampling from the historical user behavior feature space or by principal component analysis, thereby ensuring coverage of a wide range of potential preference regions.

[0097] In each training iteration, the current input vector is broadcast to all neurons in the layer. Each neuron calculates the Euclidean distance between its weight vector and the input vector as a matching index. The smaller the distance, the higher the matching degree. A winner-takes-all competition mechanism is implemented through pre-established global inhibitory connections in the layer: all neurons send their own matching degree as an inhibitory signal to other neurons. The neuron with the lowest matching degree (smallest distance) receives the strongest activation potential and sends high-intensity inhibitory pulses to the surrounding neurons to suppress the firing activity of the other neurons. Finally, the winning neuron in this input pattern is uniquely determined.

[0098] Once the winning neuron is determined, its topological neighborhood range can be calculated based on the preset neighborhood radius function (the neighborhood radius decreases exponentially with the number of training rounds, initially covering half of the grid plane, and eventually shrinking to include only the winning neuron itself), allowing all neurons within the neighborhood to receive different weight update magnitudes according to their topological distance from the winning neuron.

[0099] The weight update rule adopts an unsupervised variant of impulse-time-dependent plasticity, through... Update weight vector update amount .in The learning rate decays with each training epoch, with an initial value of 0.1 and a decay factor of 0.99. The Gaussian decay function is the neighborhood function centered on the winning neuron. This is the current input vector; This is the current weight vector of the neuron.

[0100] The entire competitive learning process takes place during the offline training phase, iterating repeatedly using massive amounts of historical user behavior data, typically requiring 500 to 2000 rounds, until the neuron weight vectors tend to stabilize, forming several discrete long-term preference prototype centers. Then, during the online execution phase, this layer stops updating weights and only performs the forward competition process, outputting the index of the winning neuron and its corresponding prototype center vector.

[0101] Step 405: Extract the activation intensity of the long-term preference prototype center, combine it with the consolidated preference connection weight matrix, and decode to generate the long-term preference pattern vector.

[0102] Activation intensity refers to the pulse firing frequency or membrane potential amplitude generated by each neuron in the preferred prototype pulse self-organizing map layer under the current input mode due to the matching of the weight vector and the input feature vector. It is usually represented by a normalized real value. The higher the value, the stronger the fit between the prototype center and the user's current long-term behavior pattern.

[0103] In some implementations, the firing rates of all neurons under the current input are read from the preferred prototype pulse self-organizing map layer, and the firing rate of each neuron is compressed to the [0,1] interval through minimum and maximum normalization to obtain the activation intensity value of that neuron. At the same time, the consolidated preferred connection weight matrix is ​​extracted from the long-term synaptic plasticity feature consolidation layer. At this time, each row of the matrix corresponds to the long-term consolidated connection weights between a preferred prototype neuron and each region of the virtual behavior space.

[0104] A preset activation neuron selection threshold, typically 0.3 to 0.5, is used to retain only neurons with activation intensities higher than this threshold for subsequent decoding, thus suppressing interference from noisy prototypes. For each retained activated neuron, its weight vector is extracted as the baseline representation vector for that prototype, and its activation intensity is used as the weight coefficient. A weighted summation operation is then performed: the baseline representation vectors of all retained neurons are linearly combined according to their corresponding activation intensities to generate an initial aggregate vector. The row vectors corresponding to high-activation-intensity neurons in the consolidated preference connection weight matrix are then weighted according to activation intensity and concatenated to the initial aggregate vector to form an expanded-dimensional original long-term preference representation. Finally, this original representation is input into a decoder network consisting of three fully connected spiking neurons. This decoder is pre-trained using a pulse-time backpropagation algorithm (the training objective is to minimize the mean squared error between the decoded output vector and the embedding vectors of positive sample items in the user's historical behavior sequence; the decoder weights are fixed after pre-training). This allows the decoder to non-linearly map the original representation to an embedding space consistent with the dimension of the item feature library, outputting a final 256-dimensional or 512-dimensional long-term preference pattern vector.

[0105] Based on the above technical solution, by utilizing the hippocampal co-encoding mechanism of the spatial cell spiking neuron layer, user behavior timestamps and item category features are mapped to a virtual spatial dwell map. This solves the problems of insufficient fusion of long-term behavioral feature temporal and category dimensions and unstructured representation, constructing a spatial feature foundation for long-term pattern extraction and enabling dispersed long-term behavioral data to form a quantifiable dwell distribution. Simultaneously, through nonlinear accumulation and thresholding of the spiking spatiotemporal integral pooling layer, noise caused by short-term behavioral fluctuations is smoothed, generating a long-term dwell heatmap, solving the problem of long-term behavioral signals being easily interfered with by accidental interests, and effectively strengthening the feature signals of real long-term behavior. Furthermore, through the spiking temporal dependence plasticity rules of the long-term synaptic plasticity feature consolidation layer, a stable preference connection weight matrix is ​​generated, solving the problem of long-term preference features lacking stable synaptic-level association support, and realizing feature consolidation of long-term behavioral patterns. Finally, through competitive learning clustering of preference prototype centers by the preference prototype spiking self-organizing mapping layer, the problems of ambiguous long-term preference pattern division and lack of clear clustering representation are solved, forming discrete and distinguishable long-term preference categories. Finally, the activation intensity is extracted and combined with the weight matrix to decode and generate a long-term preference pattern vector. This solves the problem that long-term preference features cannot be directly output as structured and usable vectors, and provides accurate and effective long-term preference feature support for subsequent fusion of long-term and short-term patterns and prediction of user behavior. Each step is progressive and achieves efficient extraction and representation of long-term stable patterns.

[0106] In another possible implementation of the embodiments of this application, combined with Figure 1-5 As shown, the process of fusing and arbitrating conflicts between short-term interest patterns and long-term preference patterns to obtain a comprehensive user state feature vector can be achieved through the following steps 501 to 503, which are explained in detail below: Step 501: Calculate the cosine similarity between the short-term interest pattern vector and the long-term preference pattern vector as an interest consistency index, and set a consistency threshold based on the distribution of the cosine similarity between short-term and long-term vectors in historical user behavior data.

[0107] In some implementations, the short-term interest pattern vector output by the context memory pathway and the long-term preference pattern vector output by the spatial navigation pathway are obtained from the dual-path spiking neural network prediction model, and cosine similarity parallel calculation is immediately initiated. The two vectors are then dot-producted and divided by the product of their respective magnitudes to obtain an interest consistency index between -1 and 1.

[0108] Step 502: When the interest consistency index is higher than the consistency threshold, the short-term interest pattern vector and the long-term preference pattern vector are weighted and summed to generate a comprehensive user state feature vector.

[0109] In some implementations, when the interest consistency index exceeds the consistency threshold, the fusion and conflict arbitration module immediately reads the short-term interest pattern vector and long-term preference pattern vector for the current prediction period from the shared cache. The weighted summation unit within the module loads the preset fusion weights stored in the system configuration center, multiplying each dimension of the short-term interest pattern vector by the short-term weight and each dimension of the long-term preference pattern vector by the long-term weight. The two weighted vectors are then element-wise summed by the vector addition unit to generate a comprehensive user state feature vector.

[0110] The preset fusion weights stored in the configuration center are determined by performing a grid search on a validation set containing three months of historical user behavior logs during the offline phase. The search range is set to short-term weights of 0.3 to 0.5 and long-term weights of 0.5 to 0.7, with a step size of 0.05. The optimization objective is to optimize the cumulative gain of normalized loss. Finally, short-term weights of 0.4 and long-term weights of 0.6 are selected as global initial values.

[0111] Step 503: When the interest consistency index is lower than the consistency threshold, it is determined to be an interest drift state. A higher fusion weight is assigned to the long-term preference pattern vector, and a comprehensive user state feature vector carrying potential accidental interest tags is generated.

[0112] In some implementations, when the fusion and conflict arbitration module completes the cosine similarity calculation and determines that the current interest consistency index is below the consistency threshold, it immediately marks the current prediction period as an interest drift state. The arbitration logic loads the pre-set exclusive fusion weights for the drift state from the configuration center, multiplying each dimension of the long-term preference pattern vector by 0.85 and each dimension of the short-term interest pattern vector by 0.15, and then performing element-wise accumulation to generate the basic fusion vector. Simultaneously, a tag generator writes an identifier into the protocol buffer metadata field of this vector, encapsulating the specific value of the interest consistency index. When the comprehensive user state feature vector carrying this tag is sent to the product matching module, the retrieval logic still performs the complete cosine similarity calculation to recall all relevant products. However, in the subsequent re-ranking stage, the ranking engine identifies this tag and dynamically calculates a time decay factor in the range of 0.5 to 0.8 multiplied by the comprehensive recommendation score for the top K products in the recall project with the short-term interest pattern vector similarity, based on the degree of deviation of the interest consistency index. The greater the deviation, the stronger the decay, thus achieving effective filtering for incidental interests without completely cutting off the recall path.

[0113] Based on the above technical solution, a cosine similarity metric is used to quantify the consistency index of long-term and short-term interests, solving the technical problem of the lack of objective quantitative judgment standards for the fit between long-term and short-term interests and avoiding feature representation bias caused by subjective fusion. When the index is higher than a preset threshold, a weighted summation of long-term and short-term interest pattern vectors is performed to generate a comprehensive feature vector. This solves the problems of redundancy and representation distortion caused by simple splicing of long-term and short-term features when there is no interest drift. From the generation level, a structured and effective fusion of long-term and short-term interest features is achieved, accurately representing the comprehensive behavioral intention of users in a drift-free state. When the indicator falls below the threshold, it is determined to be interest drift, and a higher fusion weight is assigned to long-term preferences. At the same time, potential incidental interest tags are added to the comprehensive vector. This solves the technical problem that short-term incidental interests dominate feature fusion and cause subsequent prediction and recommendation deviations when interest drift occurs. Thus, it not only anchors the user's core long-term needs with high weights to eliminate the cause of interest drift from a mechanism perspective, but also allows subsequent steps to perform targeted attenuation filtering on items related to short-term incidental interests through tagging, while retaining a reasonable recall path. Each step improves the accuracy of the comprehensive user state feature vector, providing high-quality feature support for subsequent item library retrieval and matching, and ensuring that user behavior prediction and recommendation results are consistent with real needs.

[0114] In another possible implementation of the embodiments of this application, combined with Figure 1-6 As shown, the process by which the reinforcement signal, based on the principle of long-term synaptic enhancement and inhibition, adjusts the synaptic connection weights between user behavior features and item features in the dual-pathway spiking neural network prediction model in real time through a weight adaptive adjustment module, can be achieved through the following steps 601 to 604, which are explained in detail below: Step 601: Construct a pulse timing-dependent plastic synaptic update circuit that includes parallelized long-term enhancement computation branches and long-term suppression computation branches.

[0115] The pulse-time-dependent plastic synaptic update circuit refers to mapping the plasticity rule of synaptic connection strength in biological neural systems, which varies with the time difference between presynaptic and postsynaptic neuronal pulse firing, into combinational and sequential logic circuits within a field-programmable gate array (FPGA) using a hardware description language. This circuit implements pulse time difference detection, exponential decay function calculation, and weight increment determination using pure hardware logic gates, without the need for software instructions. The long-term enhancement calculation branch is a physically independent arithmetic logic unit within the pulse-time-dependent plastic synaptic update circuit, dedicated to weight enhancement calculation when the presynaptic pulse fires before the postsynaptic pulse. It integrates a positive time difference sampling module, a positive exponential decay function generator, and a positive multiplier-accumulator. The long-term suppression calculation branch is another physically independent arithmetic logic unit in the pulse-time-dependent plastic synaptic update circuit, completely parallel to the long-term enhancement calculation branch. It is dedicated to weight suppression calculation when the presynaptic pulse fires after the postsynaptic pulse. It integrates a negative time difference sampling module, a negative exponential decay function generator, and a negative multiplier-accumulator.

[0116] Step 602: Receive and parse the reinforcement signal to obtain the pulse intensity value. Based on the sign and magnitude of the pulse intensity value, read the target synaptic weight set associated with the user identifier and the project identifier from the synaptic weight repository.

[0117] In some implementations, the enhancement signal unit is fed into the input pin of the field-gate array neuromorphic coprocessor as a high-speed serial data stream. The enhancement signal unit parser can then be used to first lock the start flag field of the data frame, and then sequentially extract the fixed-length user identifier field and item identifier field according to the protocol. Within the same clock cycle, the encapsulated 16-bit fixed-point pulse strength value is extracted from the end of the frame, simultaneously triggering the frame reception completion flag. Upon receiving the user identifier and item identifier, the hash address mapper immediately starts two independent CRC32 hash engines, compressing the identifier bitstream into a 16-bit hash value. The user hash value is used as the high 16 bits, and the item hash value as the low 16 bits, concatenating them to form a 32-bit physical base address. This base address precisely points to the storage unit in the synaptic weight repository that stores the first element of the weight vector associated with the user and item. Simultaneously, after obtaining a valid base address, the synaptic weight repository read controller generates N consecutive read addresses starting from the base address and with an address step size of 1, based on the preset feature dimension N in the system configuration register (this value is determined by the output layer dimension of the dual-path spiking neural network prediction model and is fixed to the coprocessor configuration space during the model compilation stage). At the same time, it enables N independent read ports of the block random access memory (or N burst read operations of a single read port), and completes the parallel or pipelined readout of all N-dimensional weights within one clock cycle. The readout raw weight data stream is aligned with the sign bit and 15-bit amplitude output from the pulse intensity sign and amplitude separator in the pipeline register. The sign bit is directly sent to the branch enable terminal of the subsequent pulse timing-dependent plastic synaptic update circuit, while the amplitude is sent to the multiply-accumulate unit as the pulse intensity factor A to participate in the weight increment calculation.

[0118] Step 603: Based on the pulse intensity value, synchronously activate the long-term enhancement calculation branch or the long-term suppression calculation branch, and use the pipelined multiply-accumulate operation unit to calculate in parallel the update amount of all weight elements in the target synaptic weight set within one clock cycle.

[0119] The pipelined multiply-accumulate unit refers to the decomposition of fixed-point multiplication and addition operations involved in the calculation of weight update into five independent pipelined stages: fetch, decode, multiply, accumulate, and write back. Each stage is isolated by a dedicated register and the processing progress of different data is advanced in parallel under the drive of the clock edge.

[0120] In some implementations, the pipelined register treats the latched pulse intensity amplitude A, pulse intensity sign bit S, time difference Δt, and N-dimensional weights read from the synaptic weight repository as initial weight values ​​w[i] and simultaneously pushes them to the input port of the pulse timing-dependent plasticity synaptic update circuit. At this point, the branch-enabled arbiter first captures the sign bit S: If S is 0 (positive pulse), then an enable signal is sent to the long-term enhancement branch. At the same time, a disabling signal is sent to the long-term suppression branch; Conversely, if S is 1 (negative pulse), the opposite operation is performed. This enable signal remains stable throughout the entire clock cycle to ensure computational determinism.

[0121] The enabled computation branch then initiates its internal 5-stage pipeline operation: In the first instruction fetch stage, the Δt value is latched into the time difference synchronization register group, and the index i of the currently processed feature dimension is sent into the address counter. In the second-stage decoding phase, the high 4 bits of the Δt value are used as an index to access the exponential function lookup table, followed by long-term enhanced branch readout. approximation Long-term suppressed branch readout approximation The lookup table output of branches that are not enabled is forced to be set to zero; The third multiplication stage combines the pulse intensity amplitude A and the lookup table output value. The current weight value w[i] is simultaneously fed into an 18×18-bit fixed-point multiplier to calculate A× The product of ×w[i] and P[i]; In the fourth accumulation stage, the product P[i] and the preset learning rate parameter η (long-term enhancement branch fixed value 0.01, long-term suppression branch fixed value 0.008) are input into the weight increment accumulation tree to generate the weight update amount Δw[i]=η×P[i] for this dimension. At the same time, the accumulation tree merges the update amounts of all dimensions in this period in parallel. In the fifth write-back stage, Δw[i] and the original weight w[i] are input into the adder to generate a new weight w'[i] = w[i] + Δw[i] (long-term enhancement branch) or w'[i] = w[i] - Δw[i] (long-term suppression branch), and the result is latched into the output register to wait for the write-back bus.

[0122] The entire 5-stage pipeline processes weight data for N dimensions simultaneously in each clock cycle. While the first stage is processing dimension i, the second stage is already processing dimension i-1, and so on. After a pipeline setup delay of 5 clock cycles, each clock cycle can output the updated weight value of a complete dimension. The N dimensions consume a total of (N+4) clock cycles to complete all calculations.

[0123] Step 604: Write the updated values ​​of all weight elements back to the synaptic weight repository to perform real-time updates of the synaptic connection weights between the user behavior feature dimension and the project feature dimension in the dual-path spiking neural network prediction model.

[0124] In some implementations, the updated weight value w'[i], its corresponding feature dimension index i, and the 32-bit physical base address BASE_ADDR generated by the hash address mapper are synchronously pushed to the write-back buffer queue entry. At this time, the write-back buffer queue manages the storage space in a circular pointer manner. When the queue is not full, the current data pair is written, and a queue validity flag trigger address and event indicate a bus write transaction request are generated.

[0125] The write address remapper reads the base address BASE_ADDR and dimension index i from the head of the queue. Based on the storage organization structure of the synaptic weight repository (the N-dimensional weights associated with each user and item are regarded as a vector stored continuously, and each weight element occupies an independent column address), it calculates the target physical address TARGET_ADDR=BASE_ADDR+(i×WEIGHT_WIDTH / 8), where WEIGHT_WIDTH is a 16-bit fixed-point number width.

[0126] After receiving the remapped target physical address and the updated weight value w'[i], the atomic write operation controller first queries the write lock register status of the address. If it is not locked, it immediately drives the address line, data line and write enable signal to the write data bus of the block random access memory, and completes the weight value writing within a single clock cycle.

[0127] After the write operation is completed, the weight update submission acknowledgment initiates a readback verification to the same target physical address in the next clock cycle. It performs a bitwise XOR comparison between the readback data w''[i] and w'[i]. If the XOR result is zero, it generates an ACK flag and removes the entry from the writeback buffer queue. If the comparison fails, it retryes the write operation. If the retry exceeds three times, it triggers an interrupt and reports to the central processing unit.

[0128] Upon successful writing, the association strength version stamp register associated with the target physical address is automatically incremented by 1. This version stamp is returned with all subsequent read requests to that address, allowing other coprocessors in the distributed system to determine the validity of the cached data. Once the weight updates for all N dimensions in the write-back buffer queue have been completed and acknowledged, the address-event bus controller releases the bus channel resources occupied by the strengthening signal unit and sets the completion flag register. At this point, the entire process of real-time updating of the association strength of the user and project pair is complete.

[0129] Based on the above technical solution, by constructing a hardware-based synaptic update circuit with parallelized long-term enhancement and suppression computation branches, the hardware-level separate execution of synaptic plasticity rules can be achieved. This solves the technical problems of poor timing and low execution efficiency in traditional software-based synaptic update implementations, while providing an efficient hardware computing foundation for real-time updates of association strength. Furthermore, by combining the parsing of enhancement signals and the precise location of user and project association weight addresses through hash mapping, along with the parallel reading of the target weight set, the problems of slow weight retrieval and insufficient targeting are effectively solved, enabling high-speed and accurate retrieval of associated weights and making weight updates more targeted. Simultaneously, by synchronously activating the corresponding computation branch based on the positive and negative pulse intensity, and relying on pipelined multiply-accumulate units to calculate the weight update amount in parallel, the problems of serial processing and time-consuming multi-element operations in weight update calculations are solved, allowing multiple weight elements to be processed synchronously in parallel, improving the efficiency of update calculations. Finally, the weight update is written back to the synaptic weight repository and verified, which solves the problems of error-prone weight writing and version inconsistency. It achieves reliable real-time updates of association strength at the model parameter level, and progressively completes the dynamic adaptive adjustment of the association strength between user and project features in the dual-path model. This ensures the real-time and accuracy of model parameter updates, allowing the model to quickly adapt to user behavior feedback and continuously improve the accuracy of behavior prediction and recommendation.

Claims

1. A user behavior prediction method based on spiking neural networks, characterized in that, include: Collect historical user interaction sequence data and real-time interaction stream data, perform structured processing and feature encoding, and generate user behavior pulse sequences; A dual-path spiking neural network prediction model is constructed, which is formed by the bionic hippocampal contextual memory pathway and the spatial navigation pathway in parallel. The bionic hippocampal contextual memory pathway identifies and separates short-term temporal patterns in the user behavior pulse sequence, and the spatial navigation pathway extracts and reinforces long-term stable patterns in the user behavior pulse sequence. The short-term interest pattern and long-term preference pattern are fused and conflict arbitration is performed to obtain a comprehensive user state feature vector, which is then searched and matched in the item library to generate an initial prediction list. Acquire user interaction feedback data on the initial prediction list and convert it into reinforcement signals in the form of neural impulses; The enhancement signal, based on the principle of long-term synaptic enhancement and inhibition, adjusts the synaptic connection weights between user behavior features and item features in the dual-pathway spiking neural network prediction model in real time through a weight adaptive adjustment module. Based on the synaptic connection weights, a user-item matching score matrix is ​​calculated, and the initial prediction list is reordered and refined to generate a prediction recommendation list.

2. The user behavior prediction method based on a spiking neural network according to claim 1, characterized in that, The construction process of the dual-path spiking neural network prediction model specifically includes: A spiking neural network prediction model is constructed, comprising an input layer, a shared feature extraction layer, a dual-pathway separation processing layer, and an output layer. The shared feature extraction hidden layer is composed of multiple convolutional spiking neurons, which is used to perform preliminary spatiotemporal feature extraction on the input spiking sequence and output a shared feature spiking map. The dual-pathway separation processing layer is composed of a biomimetic hippocampal context memory pathway subnetwork and a spatial navigation pathway subnetwork that are physically isolated. The biomimetic hippocampal contextual memory pathway subnetwork uses spiking neurons with a first preset time constant range and is configured with a short-term synaptic plasticity mechanism to extract and separate short-term interest pattern vectors from the shared feature pulse map. The spatial navigation pathway subnetwork employs spiking neurons with a second preset time constant range, which is greater than the first preset time constant. It is also configured with a long-term synaptic plasticity basis and a pattern extraction mechanism to extract and enhance long-term preference pattern vectors from the shared feature pulse map.

3. The user behavior prediction method based on a spiking neural network according to claim 2, characterized in that, The process by which the biomimetic hippocampal episodic memory pathway identifies and separates short-term temporal patterns in the user behavior pulse sequence specifically includes: The shared feature pulse map is input into the time-sensitive spiking neuron layer, and pulse temporal pattern competition and clustering operations are performed. The neuron with the highest matching degree with the pulse pattern in the current window is activated as the winning neuron. The consecutive activation events of the winning neurons are sequentially linked to form interest trajectories representing different short-term interest segments; The interest trajectory is encoded into a primary short-term interest pattern vector, and then refined and denoised using a pulse autoencoder to output the short-term interest pattern vector.

4. The user behavior prediction method based on a spiking neural network according to claim 3, characterized in that, The process of extracting and enhancing long-term stable patterns in the user behavior pulse sequence through the spatial navigation path specifically includes: The shared feature pulse map is input into the spatial cell pulse neuron layer. By simulating the co-encoding mechanism of hippocampal position cells and grid cells, the user behavior timestamp and item type features are jointly mapped to the virtual behavior space coordinate system to generate a behavior space dwell map. The behavior space dwell map is input into the pulse spatiotemporal integral pooling layer for nonlinear accumulation and threshold distribution, and a long-term dwell thermal pulse map is output. The long-term resident thermopulse map is input into the long-term synaptic plasticity feature consolidation layer, and a consolidation-preferred connection weight matrix is ​​generated by adopting a learning rule based on pulse time-dependent plasticity. The consolidated preference connection weight matrix is ​​input into the preference prototype impulse self-organizing map layer, and clustered into several long-term preference prototype centers through competitive learning. The activation intensity of the long-term preference prototype center is extracted and combined with the consolidated preference connection weight matrix to decode and generate the long-term preference pattern vector.

5. The user behavior prediction method based on a spiking neural network according to claim 4, characterized in that, The process of fusing and arbitrating conflicts between the short-term interest patterns and long-term preference patterns to obtain a comprehensive user state feature vector specifically includes: The cosine similarity between the short-term interest pattern vector and the long-term preference pattern vector is calculated as an interest consistency index, and a consistency threshold is set according to the distribution of the cosine similarity between short-term and long-term vectors in historical user behavior data. When the interest consistency index is higher than the consistency threshold, the short-term interest pattern vector and the long-term preference pattern vector are weighted and summed to generate a comprehensive user state feature vector. When the interest consistency index is lower than the consistency threshold, it is determined to be an interest drift state. A higher fusion weight is assigned to the long-term preference pattern vector, and a comprehensive user state feature vector carrying potential accidental interest tags is generated.

6. The user behavior prediction method based on a spiking neural network according to claim 5, characterized in that, The enhancement signal, based on the principle of long-term synaptic enhancement and inhibition, adjusts the synaptic connection weights between user behavior features and item features in the dual-pathway spiking neural network prediction model in real time through a weight adaptive adjustment module. Specifically, this includes: Construct a pulse timing-dependent plastic synaptic update circuit that includes parallelized long-term enhancement computation branches and long-term suppression computation branches; Receive and parse the enhancement signal to obtain the pulse intensity value, and read the target synaptic weight set associated with the user identifier and the project identifier from the synaptic weight repository according to the sign and magnitude of the pulse intensity value; Based on the pulse intensity value, the long-term enhancement calculation branch or the long-term suppression calculation branch is activated synchronously, and the update amount of all weight elements in the target synaptic weight set within one clock cycle is calculated in parallel using a pipelined multiply-accumulate operation unit. The updated values ​​of all weight elements are written back to the synaptic weight repository to perform real-time updates of the synaptic connection weights between the user behavior feature dimension and the project feature dimension in the dual-path spiking neural network prediction model.

7. The user behavior prediction method based on a spiking neural network according to claim 6, characterized in that, The process of acquiring user interaction feedback data on the initial prediction list and converting it into reinforcement signals in the form of neural impulses specifically includes: Capture user interaction events on items in the initial prediction list, including at least click, long press, favorite, add to cart, buy now, ignore, and close the page; According to the preset event value mapping rules, a basic pulse intensity value is assigned to each type of interactive event, wherein positive feedback is assigned a positive value and negative feedback is assigned a negative value. By combining the occurrence time of the interactive event and the interface context information, the basic pulse intensity value is modulated to generate the final pulse intensity; When the absolute value of the final pulse intensity is greater than a preset intensity threshold, a pulse signal is generated at the time point corresponding to the interactive event, and the amplitude of the pulse signal is equal to the final pulse intensity. The pulse signal is associated with the corresponding user ID and project ID to form an enhanced signal unit.

8. The user behavior prediction method based on a spiking neural network according to claim 7, characterized in that, The process of collecting historical user interaction sequence data and real-time interaction stream data, performing structured processing and feature encoding, and generating user behavior pulse sequences specifically includes: The system collects historical interaction sequence data and real-time interaction stream data in parallel. The historical interaction sequence data includes the user's purchase, favorite, and rating records over multiple periods, as well as the corresponding product characteristics. The real-time interaction stream data includes page browsing, click events, and corresponding interaction characteristics within the current session. The historical interaction sequence data is periodically segmented, and a steady-state feature vector is extracted in each period. The steady-state feature vector is a low-frequency pulse sequence and includes interest diversity entropy, consumption level, and interaction depth score. The real-time interactive stream data is segmented by a sliding window, and transient feature vectors are extracted in each window. The transient feature vectors are high-frequency pulse sequences and include attention shift rate, browsing stability coefficient, and interaction fineness score. The low-frequency pulse sequence and the high-frequency pulse sequence are time-aligned and pulse-fused to generate a unified user behavior pulse sequence.

9. The user behavior prediction method based on a spiking neural network according to claim 8, characterized in that, The process of reordering and refining the initial prediction list to generate a prediction recommendation list specifically includes: Based on the user-item matching score matrix, the comprehensive recommendation score for each item in the initial prediction list is calculated. The comprehensive recommendation score includes long-term preference score, short-term interest score, and time decay factor. Based on the comprehensive recommendation score, the items in the initial prediction list are sorted in descending order to form a reordered list; A dynamic threshold is set, and items in the re-ranking list whose overall recommendation score is lower than the dynamic threshold are filtered out, while items with a score higher than the threshold are retained to form a predicted recommendation list.

10. A user behavior prediction system based on a spiking neural network, characterized in that, The user behavior prediction method based on a spiking neural network according to any one of claims 1-9, the prediction system specifically includes: a communication unit, a prediction analysis module, a fusion module, a feedback adjustment module, and a prediction generation module; The communication module collects historical user interaction sequence data and real-time interaction stream data, performs structured processing and feature encoding, and generates user behavior pulse sequences. The predictive analysis module constructs a dual-path spiking neural network prediction model formed by the bionic hippocampal contextual memory pathway and the spatial navigation pathway in parallel. The bionic hippocampal contextual memory pathway identifies and separates short-term temporal patterns in the user behavior pulse sequence, and the spatial navigation pathway extracts and reinforces long-term stable patterns in the user behavior pulse sequence. The fusion module fuses and arbitrates conflicts between the short-term interest pattern and the long-term preference pattern to obtain a comprehensive user state feature vector, and then performs retrieval and matching in the item library to generate an initial prediction list. The feedback adjustment module acquires the user's interactive feedback data on the initial prediction list and converts it into reinforcement signals in the form of neural impulses; The enhancement signal, based on the principle of long-term synaptic enhancement and inhibition, adjusts the synaptic connection weights between user behavior features and item features in the dual-pathway spiking neural network prediction model in real time through a weight adaptive adjustment module. The prediction generation module calculates the user-item matching score matrix based on the synaptic connection weights, reorders and refines the initial prediction list, and generates a prediction recommendation list.

Images