A personalized shop recommendation method based on multi-dimensional space-time perception and hierarchical fusion

Abstract

This invention discloses a personalized shop recommendation method based on multi-dimensional spatiotemporal perception and hierarchical fusion, belonging to the field of intelligent product recommendation. The method includes the following steps: system initialization and multi-source data integration; real-time passenger feature extraction and spatiotemporal modeling; execution of collaborative filtering layer processing to generate a preliminary shop recommendation list; execution of geofencing filtering layer processing; execution of dynamic time warping layer processing; execution of a hierarchical fusion mechanism based on adaptive weights to dynamically integrate scores from different processing layers and generate a final personalized recommendation list; real-time monitoring of passenger behavior feedback, and dynamic optimization of the weight configuration of each processing layer through an online learning strategy to achieve self-evolution of the recommendation strategy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent product recommendation, specifically involving a personalized store recommendation method based on multi-dimensional spatiotemporal perception and hierarchical fusion. Background Technology

[0002] With the acceleration of urbanization and the rapid development of internet technology, traditional business models are gradually transforming. The rise of online e-commerce has brought enormous challenges to traditional shopping malls, and consumers' demands for shopping efficiency and personalized experiences are increasing. Traditional shopping guide models rely on human experience, resulting in problems such as low recommendation accuracy, homogenized services, and unreasonable route planning, making it difficult to meet the diversified needs of modern consumers. Furthermore, existing tourism recommendation systems also have limitations in dealing with dynamic environmental changes, struggling to integrate multi-dimensional dynamic parameters such as dense crowds, sudden weather events, and traffic control in real time, often leading to recommended routes that are out of sync with the actual carrying capacity of scenic spots. Against this backdrop, intelligent shopping guide technology has gradually become a research hotspot. However, existing intelligent shopping guide systems still have some problems, including being limited to single-dimensional purchase preference analysis, ignoring various dynamic factors, resulting in recommendations lacking timeliness and adaptability. Simultaneously, they struggle to handle high-dimensional, large-scale product data in shopping malls, leading to insufficient mining of product relationships and weak correlation in the generated recommendation lists. In addition, shopping guide route planning often only considers physical spatial distance, easily resulting in route redundancy or deviation from the actual needs of users, reducing shopping guide efficiency. Summary of the Invention

[0003] To achieve the above objectives, this invention provides a personalized shop recommendation method based on multi-dimensional spatiotemporal perception and hierarchical fusion, comprising the following steps: Step 1: System initialization and multi-source data integration to build a data infrastructure platform that supports real-time recommendation calculation; Step 2: Real-time passenger feature extraction and spatiotemporal modeling to obtain the dynamic location of passengers and construct a walking time cost model that integrates passenger flow density; Step 3: Perform collaborative filtering layer processing, mine deep preferences based on user historical behavior data, and generate a preliminary list of recommended shops; Step 4: Perform geofencing filtering, and filter shops that do not meet the geographical constraints based on the real-time location of travelers. Step 5: Perform dynamic time warping layer processing, combining the passenger's remaining waiting time with the characteristics of shop services, to evaluate and filter out recommended options with feasible times; Step 6: Execute a hierarchical fusion mechanism based on adaptive weights to dynamically integrate scores from different processing layers and generate the final personalized recommendation list; Step 7: Monitor passenger behavior feedback in real time, and dynamically optimize the weight configuration of each processing layer through online learning strategies to achieve self-evolution of the recommendation strategy.

[0004] Preferably, step 1 includes: Step 101: Build a unified data warehouse, integrating heterogeneous data sources from the airport passenger relationship management system, flight information system and passenger flow monitoring system to achieve centralized storage and management of data; Step 102: Construct a structured database of shop information, and record in detail the unique identifier, category, average order value range, and precise geographic coordinates of each shop. Step 103: Establish a high-performance geospatial database cluster and create a geospatial index for all shops in the terminal to enable rapid retrieval and related queries of shop location information.

[0005] Preferably, step 2 includes: Step 201: Capture the precise latitude and longitude location information of passengers in real time and proactively through Wi-Fi probes, Bluetooth beacon networks, or passenger-authorized mobile applications deployed in the airport environment; Step 202: Convert the obtained raw latitude and longitude coordinates into standardized GeoHash grid codes, thereby discretizing the continuous location information and mapping it to predefined grid cells; Step 203: Pre-calculate and dynamically maintain a walking time matrix between grids. This matrix combines the actual path structure of the terminal with real-time passenger flow density information to more realistically reflect the travel time cost between points.

[0006] Preferably, step 3 includes: Step 301: Construct a passenger shop interaction matrix based on passenger historical consumption and browsing records, and adaptively select a neural collaborative filtering or matrix factorization model for training and prediction for users with different data richness. Step 302: Using the selected recommendation model, learn the implicit feature vectors of users and shops, and calculate the user's preference prediction score for shops not visited in order to capture their potential interests. Step 303: Sort the calculated preference scores in descending order to generate a basic recommendation list containing the Top-N candidate shops for each traveler, which will serve as the input for subsequent processing layers.

[0007] Preferably, step 4 includes: Step 401: Using the established geospatial index, with the real-time grid code of the traveler as the center, perform a high-efficiency geofence range query to quickly obtain all potential accessible shops within a radius of R meters. Step 402: For each candidate shop within the fence, calculate its precise spherical distance from the traveler's current location, and generate a quantified spatial fit score based on this distance; the closer the distance, the higher the score. Step 403: Based on the preset hard distance threshold and the estimated time obtained from the walking time matrix, perform a second screening of the shops within the fence, and remove options that are too far away or are not expected to be reached within a reasonable time.

[0008] Preferably, step 5 includes: Step 501: Obtain the passenger's remaining waiting time from the flight information and query the specific one-way travel time required to reach each candidate shop from the walking time matrix; Step 502: Design a time warping function that comprehensively considers the remaining time, walking time, and the service type of the shop, and calculates a time-dimensional suitability score for each shop to quantify its recommendation priority under time constraints. Step 503: Based on the time suitability score, re-sort the shops that have passed the previous steps and directly filter out those that are completely infeasible in terms of time, such as shops where the round trip plus consumption time exceeds the remaining time.

[0009] Preferably, step 6 includes: Step 601: Initialize a set of dynamically adjustable weight parameters for the three processing layers representing user preference, spatial proximity, and temporal feasibility, respectively, whose sum is always 1; Step 602: For the final candidate store set that has passed both geographical and time filtering, the scores calculated at each layer are linearly weighted according to the current weight to calculate the final comprehensive recommendation score for each store. Step 603: Sort all candidate shops globally based on the final comprehensive score, select the Top-N shops with the highest scores, and generate and output the final personalized recommendation list for the traveler.

[0010] Preferably, step 7 includes: Step 701: Set a fixed monitoring time window and continuously collect data on actual interaction behavior of travelers with the recommendation results, including key feedback signals such as clicks, in-store navigation, and consumption conversion. Step 702: Based on the aggregated feedback data within a time window, analyze the contribution of each recommendation layer to generating positive feedback, and calculate the adjustment amount and direction of the weight of each layer accordingly. Step 703: Based on the calculated adjustment amount, update the weights of each layer in a smooth manner to ensure that the system can adaptively adjust the emphasis on different dimensions of preference, distance, and time according to changes in the overall recommendation effect, thereby achieving continuous self-optimization.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves comprehensive analysis of multi-dimensional information through a parallel processing mode of collaborative filtering, geofencing, and dynamic time warping, overcoming the problems of traditional shopping guide models that rely on human experience and have low recommendation accuracy, and significantly improving the accuracy and personalization of shop recommendations; 2. This invention constructs a complete user profile system, which achieves accurate understanding of user preferences by collecting and analyzing multi-dimensional user data in real time, effectively solving the shortcomings of existing systems that ignore multiple dynamic factors; 3. By setting a dynamic time warping layer and combining it with real-time calculation of the walking time matrix, the rationality of the recommended route is ensured, avoiding route redundancy or deviation from the user's actual needs. 4. This invention employs an adaptive weighted fusion module based on online learning. By monitoring passenger behavior feedback in real time, it can dynamically adjust the weights of each layer, enabling the system to continuously optimize itself and improving the adaptability and accuracy of the recommendation strategy. 5. This invention overcomes the limitations of existing systems in data collection and processing by effectively integrating and analyzing multi-source heterogeneous data, improving the accuracy and completeness of the data, and providing data support for more accurate store recommendations. Attached Figure Description

[0012] Figure 1 This is a flowchart of the recommended method described in this invention. Detailed Implementation

[0013] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Example

[0014] This invention provides a personalized shop recommendation method based on multi-dimensional spatiotemporal perception and hierarchical fusion. The method employs a parallel processing mode of collaborative filtering, geofencing, and dynamic time warping in its system architecture. It utilizes an adaptive weighted fusion module based on online learning for decision-making, ultimately forming an intelligent recommendation system capable of perceiving user preferences, real-time location, and time constraints, and capable of self-optimization. The specific implementation steps are as follows: Step 1: System Initialization and Multi-Source Data Integration Step 101: Build a data warehouse, integrating data from the airport passenger relationship management system, flight information system, and passenger flow monitoring system. An offline data warehouse based on Hadoop Hive is used to store and process historical data. Real-time access is provided via message queues for flight numbers, boarding gates, scheduled / estimated departure times, historical consumption records, passenger profiles based on membership levels, and real-time passenger flow heatmaps published by the passenger flow monitoring system. A unified data model is established to clean, correlate, and dimensionally model data from different systems.

[0015] Step 102: Construct a structured store database, entering the store's unique identifier, category, average order value, and latitude / longitude geographic coordinates. Create a store information table in the relational database MySQL. The table fields include: shop_id (unique store identifier, primary key), shop_name (store name), category (category, such as catering, retail, clothing, service), price_level (average order value level, such as 1-5 levels), latitude, and longitude.

[0016] Step 103: Establish a Redis cluster, create a geospatial index collection of terminal shops, import shop geocoordinates, and associate them with detailed shop information. Store shop geolocations using the GEO data type. Create a GEO collection (Key, e.g., GEO:TERMINAL:T1) for each terminal. Execute the GEOADD command to add the latitude and longitude of all shops from Step 102 to the corresponding collection, with the member being `shop_id`. Simultaneously, store the shop's name, category, and other detailed information in Redis using a hash data structure, with the key `SHOP:{shop_id}`, for fast lookup and querying. This step establishes millisecond-level response time for geolocation queries.

[0017] Step 2: Real-time passenger feature extraction and spatiotemporal modeling: Step 201: Obtain passengers' latitude and longitude location information in real time through Wi-Fi probes, Bluetooth beacons, or passenger mobile terminal applications deployed within the terminal. Using Bluetooth beacons within the terminal, the passenger's mobile app (authorization required) scans the beacon signal in the background and converts the Received Signal Strength Indication (RSSI) into latitude and longitude using triangulation or fingerprint database matching algorithms. The location information is then uploaded to the recommendation system server in real time via HTTPS protocol.

[0018] Step 202: Convert the acquired passenger's latitude and longitude coordinates into corresponding GeoHash codes to determine their gridded location unit. Use the GeoHash algorithm (6-bit precision) to encode the passenger's (longitude, latitude) coordinates into a string. This string divides the terminal floor into a grid of approximately 1.2km * 0.6km, serving as the basic unit for location calculation. This operation discretizes the continuous spatial problem, facilitating rapid matching and calculation.

[0019] Step 203: Based on the terminal building floor plan, construct an undirected graph G. Nodes are key locations such as shop entrances, path intersections, and boarding gates. Edges are walkable paths, with edge weights equal to the geometric length of the path. Pre-calculate and maintain a dynamic inter-grid walking time matrix, calculated using the following formula: (1), This calculates the shortest path distance from the center point of grid i to the center point of grid j. Walking speed v(t) is a dynamic variable, set according to real-time passenger flow density: 90 m / min for smooth traffic (density < 0.3 people / ㎡), 70 m / min for moderate traffic (0.3-0.7 people / ㎡), and 50 m / min for congested traffic (> 0.7 people / ㎡). The system updates v(t) every minute based on the latest passenger flow heatmap and recalculates or corrects T_walk(i,j) between grids, forming a dynamic time cost matrix.

[0020] Step 3: Perform collaborative filtering layer processing to generate a basic recommendation list: Step 301: Construct a passenger-store interaction matrix based on passenger historical consumption data. For frequent travelers with sufficient data, a neural collaborative filtering model is used; for new travelers with sparse data, a matrix factorization model is used. Sufficient data is defined as the number of stores interacting with the user being ≥5; otherwise, it is considered sparse. For users with sufficient data, a neural collaborative filtering model is used; for users with sparse data, a simpler matrix factorization model, which is more favorable for cold starts, is used.

[0021] Step 302: Capture the nonlinear correlation between traveler preferences and shop features using latent vector decomposition (LVM) or a multilayer neural network (MLN). Map the user ID (denoted by u) and shop ID (denoted by i) to d-dimensional dense vectors p_u and q_i respectively through an embedding layer. Concatenate p_u and q_i (p_u⊕q_i) and input the concatenation into a multilayer perceptron (MLP). The MLP can be designed with 2-3 layers, for example [64, 32], using the ReLU activation function. The output of the MLP passes through a fully connected layer and a sigmoid activation function, ultimately outputting a prediction score y_ui between 0 and 1, representing the probability of user u's preference for shop i. For sparse users, the probability can be calculated directly. , where p_u and q_i are low-dimensional latent vectors, which are learned by minimizing loss functions such as mean square error.

[0022] Step 303: Calculate the preference prediction score for each shop and generate a Top-N basic recommendation list containing shop ID, preference score, shop category, and average order value. For target user u, use the trained model (NCF or MF) to calculate their prediction score S_cf(i) = y_ui for all shops (or the initially screened candidate shops). Sort by S_cf(i) in descending order and select the Top-N (e.g., N=50) shops to form the basic recommendation list L_cf. Each record in the list contains {shop_id, S_cf, category, price_level}.

[0023] Step 4: Perform geofencing filtering to select nearby shops: Step 401: Invoke the range query command of the geospatial index set to query all nearby shops within a radius of R meters, centered on the traveler's real-time location. Send the command to the Redis cluster: `GEORADIUS GEO:TERMINAL:T1 {longitude} {latitude} 100 m WITHDIST WITHCOORD`. The radius R is set to 100 meters. The command will return all `shop_id`s within the 100-meter range, along with their straight-line distance and coordinates. This operation typically completes within 10 milliseconds.

[0024] Step 402: Calculate the straight-line distance between each candidate shop and the traveler's location, and generate a spatial fit score. The distance dist returned by the GEORADIUS command is the straight-line distance D(i, user). Calculate the spatial fit score: (2) When D<= 100 meters, the score is between 0 and 1, and the closer the distance, the higher the score; when D>100 meters, the score is 0.

[0025] Step 403: Based on preset distance and walking time thresholds, remove shops with a spatial fit score of zero or excessively long walking times. First, remove shops where S_geo(i) = 0 (i.e., distance > 100 meters). Then, for each remaining shop i, query the walking time T_walk(i) from the traveler's current location grid to the grid where shop i is located from the walking time matrix in Step 203. If T_walk(i) > 15 minutes, the walking time is considered too long and the shop is removed. The shops obtained through this step form a geographic filtering list L_geo.

[0026] Step 5: Perform dynamic time warping layer processing and evaluate time adaptability: Step 501: Obtain the remaining waiting time T_remain based on the passenger's flight information, and obtain the estimated walking time T_walk to each candidate shop based on the walking time matrix in Step 2.

[0027] (3), For a candidate shop i from L_cf or L_geo, query T_walk(i) from the walking time matrix.

[0028] Step 502: Calculate the time adaptation score S_time for each candidate shop. Define a time warping function f. This function takes into account T_remain (remaining time), T_walk(i) (one-way time), and Category(i) (shop type).

[0029] (1) Calculate the basic time pressure coefficient: (4) If time_pressure>=1, then S_time(i) = 0 (completely infeasible).

[0030] (2) Time estimation based on store type: Define a typical service time T_serve(category) for each type of store, such as 5 minutes for fast food, 25 minutes for full-service meals, and 10 minutes for retail.

[0031] Calculate the time adaptation score: (5), Here, w(category) is the category weight, which can be dynamically adjusted based on T_remain. For example, when T_remain < 30 minutes, the w value for fast food is increased, and the w value for full-service meals is decreased. This function ensures that when time is tight, the system will prioritize recommending shops with shorter processing times and matching categories.

[0032] Step 503: Dynamically adjust the recommendation priority using a time warping function to filter shops that do not meet the time feasibility condition. Calculate S_time(i) for all candidate shops. Set a minimum threshold (S_time(i) > 0.1); shops below this threshold are considered time-infeasible and removed from the candidate list. The remaining shops form the time-feasible list L_time.

[0033] Step 6: Perform hierarchical fusion based on adaptive weights to generate the final recommendation list, fusing the scores from the three different dimensions to obtain the comprehensive optimal solution: Step 601: Configure dynamic weights W_cf, W_geo, and W_time for the collaborative filtering layer, geofencing layer, and dynamic time warping layer, respectively, with the sum of the weights being 1. The initial weight values ​​can be set based on business experience, for example, W_cf=0.4, W_geo=0.3, W_time=0.3. These weights are user-level or scenario-level parameters and will be dynamically adjusted in step 7.

[0034] Step 602: Calculate the final recommendation score for candidate shops that have passed each layer of filtering. Take the intersection of the three lists (or process the union; for shops not appearing in a list, set the score for that layer to 0 or a very small baseline value). For shop i, the final score is calculated using the following formula: (6), where S_cf(i), S_geo(i), and S_time(i) have all been normalized to the interval [0, 1].

[0035] Step 603: Sort all candidate shops in descending order of their final recommendation score S_final, and generate and output the Top-N final recommendation list. Sort the shops in descending order of S_final(i), and take the top M (e.g., M=10) as the final recommendation list, returning it to the user terminal. The list information includes shop ID, name, category, estimated walking time, and overall rating.

[0036] Step 7: Real-time monitoring of passenger behavior feedback, and execution of weight adaptive optimization, achieving adaptive recommendation strategy: Step 701: Set a monitoring period and collect real-time data on travelers' clicks, browsing, and consumption behavior related to the recommended results. Set a sliding time window. Within this window, the system records the user's feedback on each recommendation: whether they clicked (click=0 / 1), whether they visited the store (determined by geofencing, visit=0 / 1), and whether they made a purchase (purchase=0 / 1 and amount). Each feedback sample is associated with the weights (W_cf, W_geo, W_time) used to generate that recommendation and the scores (S_cf, S_geo, S_time) for each layer on the recommended store.

[0037] Step 702: Calculate the weight adjustment amount for each recommendation layer based on the behavioral feedback data. Define a comprehensive revenue function R: (7), Where α, β, and γ are adjustment coefficients, avg_amount is the average consumption amount, and baseline is the average return over a recent period.

[0038] The I(k) indicator function can be defined as: (8), This refers to the proportion of the score of layer k in the total score. The weight adjustment ΔW_k can then be calculated according to the following rule: When a recommendation generates positive feedback (R>baseline), the layer with the largest contribution (larger I(k)) receives positive incentive: (9), When feedback is unsatisfactory, a negative penalty is applied. η is the learning rate, which controls the adjustment range.

[0039] Step 703: Dynamically update the weights of each layer based on the weight adjustment amount to ensure that the total weights are normalized. The weight update formula is: (10) After the update, the weight vector [W_cf_new, W_geo_new, W_time_new] is normalized so that its sum is 1 again.

[0040] This update can be performed on a user, user group, or globally, enabling the system to automatically learn the importance of each dimension in different scenarios such as weekdays / weekends and morning / evening rush hours, achieving an optimal balance between personalized and contextualized recommendations. It should be noted that the above content merely illustrates the technical concept of the invention and should not be construed as limiting the scope of protection of the invention. For those skilled in the art, various improvements and modifications can be made without departing from the principles of the invention, and all such improvements and modifications fall within the scope of protection of the claims of this invention.

Claims

1. A personalized shop recommendation method based on multi-dimensional spatiotemporal perception and hierarchical fusion, characterized in that, Includes the following steps: Step 1: System initialization and multi-source data integration to build a data infrastructure platform that supports real-time recommendation calculation; Step 2: Real-time passenger feature extraction and spatiotemporal modeling to obtain the dynamic location of passengers and construct a walking time cost model that integrates passenger flow density; Step 3: Perform collaborative filtering layer processing to mine deep preferences based on users' historical behavior data and generate a preliminary list of recommended shops; Step 4: Perform geofencing filtering, and filter shops that do not meet the geographical constraints based on the real-time location of the traveler. Step 5: Perform dynamic time warping layer processing, combining the passenger's remaining waiting time with the characteristics of shop services, to evaluate and filter out recommended options with feasible times; Step 6: Execute a hierarchical fusion mechanism based on adaptive weights to dynamically integrate scores from different processing layers and generate the final personalized recommendation list; Step 7: Monitor passenger behavior feedback in real time and dynamically optimize the weight configuration of each processing layer through online learning strategies.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 101: Build a unified data warehouse and integrate heterogeneous data sources from the airport passenger relationship management system, flight information system and passenger flow monitoring system; Step 102: Construct a structured database of shop information, and record in detail the unique identifier, category, average order value range, and precise geographic coordinates of each shop. Step 103: Establish a high-performance geospatial database cluster and create a geospatial index for all shops within the terminal.

3. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 201: Capture the passenger's precise latitude and longitude location information in real time and proactively through Wi-Fi probes, Bluetooth beacon networks, or passenger-authorized mobile applications deployed in the airport environment. Step 202: Convert the obtained raw latitude and longitude coordinates into standardized GeoHash grid codes, thereby discretizing the continuous location information and mapping it to predefined grid cells; Step 203: Pre-calculate and dynamically maintain a walking time matrix between grids, which combines the actual path structure of the terminal with real-time passenger flow density information.

4. The method according to claim 1, characterized in that, Step 3 includes the following steps: Step 301: Construct a passenger-shop interaction matrix based on passenger historical consumption and browsing records, and adaptively select a neural collaborative filtering or matrix factorization model for training and prediction for users with different data richness. Step 302: Using the selected recommendation model, learn the implicit feature vectors of users and shops, and calculate the user's preference prediction score for shops not visited in order to capture their potential interests. Step 303: Sort the calculated preference scores in descending order to generate a basic recommendation list containing the Top-N candidate shops for each traveler, which will serve as the input for subsequent processing layers.

5. The method according to claim 1, characterized in that, Step 4 includes the following steps: Step 401: Using the established geospatial index, and centering on the real-time grid code of the traveler, perform a high-efficiency geofence range query to quickly obtain all potential accessible shops within a radius of R meters. Step 402: For each candidate shop within the fence, calculate its precise spherical distance from the traveler's current location, and generate a quantified spatial fit score based on this distance; the closer the distance, the higher the score. Step 403: Based on the preset hard distance threshold and the estimated time obtained from the walking time matrix, a second screening is performed on the shops within the fence, eliminating options that are too far away or are not expected to be reached within a reasonable time.

6. The method according to claim 1, characterized in that, Step 5 includes the following steps: Step 501: Obtain the passenger's remaining waiting time from the flight information and query the specific one-way travel time required to reach each candidate shop from the walking time matrix; Step 502: Design a time warping function that comprehensively considers the remaining time, walking time, and the service type of the shop, and calculates a time-dimensional suitability score for each shop to quantify its recommendation priority under time constraints. Step 503: Based on the time suitability score, re-sort the shops that have passed the previous steps and directly filter out those shops that are completely unfeasible in terms of time.

7. The method according to claim 1, characterized in that, Step 6 includes the following steps: Step 601: Initialize a set of dynamically adjustable weight parameters for the three processing layers representing user preference, spatial proximity, and temporal feasibility, respectively, whose sum is always 1. Step 602: For the final candidate store set that has passed both geographical and time filtering, the scores calculated at each layer are linearly weighted according to the current weight to calculate the final comprehensive recommendation score for each store. Step 603: Sort all candidate shops globally based on the final comprehensive score, select the Top-N shops with the highest scores, and generate and output the final personalized recommendation list for the traveler.

8. The method according to claim 1, characterized in that, Step 7 includes the following steps: Step 701: Set a fixed monitoring time window and continuously collect actual interaction data of travelers with the recommendation results, including clicks, in-store navigation, and key feedback signals of consumption conversion; Step 702: Based on the aggregated feedback data within a time window, analyze the contribution of each recommendation layer to generating positive feedback, and calculate the adjustment amount and direction of the weight of each layer accordingly. Step 703: Update the weights of each layer in a smooth manner based on the calculated adjustment amount.

Images