E-commerce platform personalized commodity recommendation system based on user behavior sequence

Abstract

The application discloses an e-commerce platform personalized commodity recommendation system based on user behavior sequences, relates to the technical field of e-commerce platform personalized recommendation, and solves the technical problems that it is difficult to capture rollback micro-operations in user scrolling browsing and it is difficult to mine hidden interests in non-click behaviors and lack of geometric constraints of V-shaped trajectories; the system comprises the following modules: a data acquisition module synchronously acquires various data of commodity items; a trajectory mapping module maps a visible area proportion sequence into a time sequence trajectory point set; a rollback detection module detects a rollback probability by adopting a double-path time sequence alignment network, and labels a rollback inflection point; a geometric path network adopts a second-order difference driven deformable convolution to extract V-shaped trajectory features; a time sequence path network adopts a bidirectional gated recurrent unit to extract time sequence dependence; a cross alignment module fuses geometric and time sequence features by mutual attention weight; and an interest scoring module calculates a comprehensive non-click interest evaluation value according to a curvature value.

Description

Technical Field

[0001] This invention belongs to the field of personalized recommendation technology for e-commerce platforms, specifically relating to a personalized product recommendation system for e-commerce platforms based on user behavior sequences. Background Technology

[0002] With the rapid development of mobile internet and e-commerce, e-commerce platforms have accumulated massive amounts of user behavior data. How to accurately mine user interests from this data and recommend products that users may be interested in has become a core technical problem of e-commerce recommendation systems.

[0003] Existing e-commerce recommendation systems suffer from the following technical problems: Traditional recommendation systems rely solely on explicit behavioral signals such as clicks, favorites, and purchases. When users merely scroll through products without clicking, it's difficult to capture their true interests and preferences, resulting in many potentially interesting products not being recommended. Existing behavior analysis methods based on scrolling speed can only determine the speed of browsing and struggle to identify the V-shaped scrolling micro-operation of "swiping past a product and then actively scrolling back" on the product list page. When the scrolling amplitude is small, the speed change characteristics are not obvious, easily leading to missed detections. Furthermore, they do not consider the impact of time information (hours, days of the week, seasons) on user behavior patterns. Additionally, existing deep learning recommendation models only use general loss functions such as cross-entropy during training, lacking physical constraints on the geometric characteristics of the V-shaped scrolling trajectory, resulting in low accuracy in detecting scrolling behavior.

[0004] Therefore, personalized product recommendation systems for e-commerce platforms based on user behavior sequences have emerged. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a personalized product recommendation system for e-commerce platforms based on user behavior sequences. This system addresses the technical problems of difficulty in capturing users' true interests and preferences, resulting in a large number of potentially interesting products not being recommended, difficulty in recognizing the V-shaped scrolling micro-operation of users "swiping past a product and then actively swiping back" on the product list page, and lack of physical constraints on the geometric characteristics of the V-shaped scrolling trajectory.

[0006] To address the above problems, this invention provides a personalized product recommendation system for e-commerce platforms based on user behavior sequences, comprising the following modules: Data acquisition module: used to synchronously collect the visible area ratio, scrolling acceleration value and scrolling direction angle of each product item at a predetermined sampling frequency while the user scrolls the product list page; Trajectory mapping module: used to map the visible area ratio sequence to a two-dimensional plane to obtain a time-series trajectory point set with time as the horizontal axis and visible area ratio as the vertical axis; Rollback detection module: used to extract continuous time windows from the time-series trajectory point set, use a dual-path time alignment network to detect the probability that the trajectory points in the window belong to the rollback segment, mark the sampling time when the probability is greater than the probability threshold as the rollback inflection point, and backtrack to the extreme points of the first-order difference sign flipping at both ends, and mark the continuous closed interval formed as the rollback segment. Interest scoring module: used to calculate the comprehensive non-click interest evaluation value of each product item based on the number of rollback segments and the curvature value at the rollback inflection point in each rollback segment; Recommendation reordering module: used to reorder the original recommendation sequence based on the comprehensive non-click interest evaluation value.

[0007] Preferably, in the data acquisition module, the visible area ratio of each product item is collected. By calling the view visibility detection interface of the mobile terminal operating system, the rectangular boundary of the view object corresponding to the product item in the visible area of ​​the screen is obtained. The ratio of the intersection area of ​​the rectangular boundary and the screen boundary to the total area of ​​the rectangular boundary is calculated as the visible area ratio of the product item at the current sampling time.

[0008] Preferably, in the rollback detection module, before using the dual-path temporal alignment network for detection, time-aware encoding of the visible area ratio sequence is performed, including: Obtain the absolute timestamp for each sampling moment, and extract the hour value from the absolute timestamp. Weekly duty With month value ; The hour value Encoded as and The week value The encoding is and The month value Convert the vectors to seasonal vectors according to a pre-defined seasonal mapping table; By concatenating the cyclic time coordinates with the visible area ratio and the rolling acceleration value with the rolling direction angle, a time-enhanced multidimensional feature vector is formed, which serves as the input to the dual-path temporal alignment network.

[0009] Preferably, in the rollback detection module, the construction of the dual-path timing alignment network includes: The dual-path temporal alignment network is composed of a geometric path network, a temporal path network, and a cross-alignment module. The input to the geometric path network is the visible area ratio at each sampling time within the window. This value is extracted from the temporally enhanced multidimensional feature vector and processed by a one-dimensional deformable convolutional layer to output the geometric feature vector at each sampling time. The shape of the convolution kernel of the one-dimensional deformable convolutional layer is dynamically adjusted by an offset field, and the input to the offset field is the second-order difference value of the visible area ratio sequence. The input to the temporal path network is the temporally enhanced multidimensional feature vector at each sampling time within the window. This vector is processed by a one-dimensional bidirectional gated recurrent unit to output the temporal feature vector at each sampling time. The cross-alignment module receives the geometric feature vector and the temporal feature vector, calculates the mutual attention weight between the geometric feature vector and the temporal feature vector at each sampling time, and uses the mutual attention weight as a fusion coefficient to perform a weighted summation of the geometric feature vector and the temporal feature vector to output the fused feature vector at each sampling time. The fused feature vector at each sampling time is input into a binary classifier, which outputs the probability that the sampling time belongs to a rollback segment.

[0010] Preferably, the offset field of the one-dimensional deformable convolutional layer is specifically: ,in, Let be the value of the second-order difference sequence representing the visible area proportion at the nth position of the convolution kernel. The preset deformation amplitude coefficient, It is the hyperbolic tangent activation function.

[0011] Preferably, during the training of the one-dimensional deformable convolutional layer, the loss function includes a cross-entropy main loss term and a physical constraint regularization term: The main cross-entropy loss term is specifically as follows: ,in, For the first The true label at each sampling time The predicted probability is the output of the binary classifier; The physical constraint regularization term includes: a constraint that the first derivative is zero at the rollback inflection point. And the sign constraint that the proportion of visible area within the rollback fragment first decreases and then increases. ,in, and These are the signs of the first-order differences before and after the inflection point, respectively. The total loss function is as follows: ,in, and These are the preset weighting coefficients.

[0012] Preferably, in the cross-alignment module, calculating the mutual attention weights between the geometric feature vector and the temporal feature vector at each sampling time includes: For the At each sampling time, its geometric feature vector With time series feature vectors Mutual attention weights between The calculation is as follows: ,in, and Let be the projection matrix. It is a sigmoid activation function. This is a transpose operation; The mutual attention weights As the fusion coefficient of geometric feature vectors, As the fusion coefficients of the temporal feature vectors, they are obtained through a linear projection matrix. Increase the dimensionality of the geometric feature vectors, according to... Calculate the first The fused feature vector at each sampling time point .

[0013] Preferably, when the binary classifier outputs the probability that the sampling time belongs to the rollback segment, it simultaneously uses the Monte Carlo dropout method to calculate the prediction confidence interval of the probability. The Monte Carlo dropout method uses a variational dropout mechanism and repeats the forward propagation T times to obtain T probability values. The standard deviation of the T probability values ​​is taken as the prediction confidence interval. For each sampling time, the prediction confidence interval is denoted as... Let the probability value output by its binary classifier be denoted as Calculate the confidence weight of the rollback segment corresponding to the sampling time, specifically as follows: ,in, This is the preset sensitivity coefficient; The confidence weight This serves as the weight multiplier for the rolled-back segment in the interest scoring module.

[0014] Preferably, in the interest scoring module, the calculation of the comprehensive non-click interest evaluation value for each product item includes: For a single item, obtain the total number of rollback segments that have been confirmed as valid in the current session. , among which, when When this happens, the overall non-click interest rating value of the product item is assigned to 0; when At that time, for the first For each rollback segment, obtain the curvature value at the rollback inflection point of the rollback segment. Obtain the confidence weight of the corresponding rollback segment. ; Calculate the first Weighted curvature value of each rollback segment Calculate the sum of the weighted curvature values ​​of all rollback segments. And the sum of the confidence weights of all rollback segments. The sum of the weighted curvature values The sum of the confidence weights The ratio of [the ratio of the ...

[0015] The beneficial effects of this invention are: This invention collects various physical signal data such as the visible area ratio and uses the view visibility detection interface of the mobile terminal operating system to calculate the intersection area ratio of the rectangular boundary of the product view and the screen boundary. This enables the accurate capture of the user's "swipe-look back" micro-operation behavior, and can mine the user's hidden interests from non-click browsing behavior. It solves the technical problem that existing recommendation systems have difficulty capturing non-click behavior interest signals and significantly expands the signal source dimensions of recommendation systems. This invention utilizes a deformable convolutional offset field driven by second-order difference. The second-order difference value of the visible area ratio sequence is used as the input to the offset field, causing the shape of the convolutional kernel to automatically deform according to the curvature change of the V-shaped rollback trajectory. This directly encodes the geometric prior of the V-shaped rollback trajectory into the kernel deformation, solving the technical problem that traditional fixed convolutional kernels cannot adapt to the shape of the V-shaped trajectory. This significantly improves the sensitivity of rollback detection, especially for shallow rollback scenarios with significant curvature changes. Furthermore, by introducing physical constraint regularization terms into the loss function, where the first derivative is zero, the network is forced to predict a zero derivative at the rollback inflection point, and the sign change constraint forces the network to learn the V-shaped change pattern of the visible area ratio first decreasing and then increasing, the physical geometric characteristics of the V-shaped rollback trajectory are encoded into the training process. This solves the technical problem that general loss functions cannot learn trajectory geometric characteristics, enabling the model to maintain high detection accuracy and generalization ability even with limited labeled data. This invention calculates the prediction confidence interval and generates confidence weights using the Monte Carlo dropout method. During forward propagation, some neurons are randomly dropped, and after repeated forward propagation, the standard deviation of the probability values ​​is taken as the confidence interval. The confidence weight is calculated using the confidence weight formula, providing a quantitative assessment of prediction uncertainty. This solves the technical problem of deep learning models over-relying on low-confidence predictions. When the prediction confidence interval is large, the confidence weight automatically decreases, effectively reducing the false detection rate of rollback detection. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, this invention is a personalized product recommendation system for e-commerce platforms based on user behavior sequences, comprising the following modules: Data acquisition module: used to synchronously collect the visible area ratio, scrolling acceleration value and scrolling direction angle of each product item at a predetermined sampling frequency while the user scrolls the product list page; Trajectory mapping module: used to map the visible area ratio sequence to a two-dimensional plane to obtain a time-series trajectory point set with time as the horizontal axis and visible area ratio as the vertical axis; Rollback detection module: used to extract continuous time windows from the time-series trajectory point set, use a dual-path time alignment network to detect the probability that the trajectory points in the window belong to the rollback segment, mark the sampling time when the probability is greater than the probability threshold as the rollback inflection point, and backtrack to the extreme points of the first-order difference sign flipping at both ends, and mark the continuous closed interval formed as the rollback segment. Interest scoring module: used to calculate the comprehensive non-click interest evaluation value of each product item based on the number of rollback segments and the curvature value at the rollback inflection point in each rollback segment; Recommendation reordering module: used to reorder the original recommendation sequence based on the comprehensive non-click interest evaluation value.

[0019] Specifically, during the user's scrolling of the product list page, the system synchronously collects three types of data at a predetermined sampling frequency (e.g., 30Hz): the visible area ratio of each product item, the scrolling acceleration value, and the scrolling direction angle. When collecting the scrolling acceleration value, the system listens for the trigger timestamps of touch events, calculates the time interval between two adjacent touch events based on the trigger timestamps of three consecutive touch events, and calculates the scrolling acceleration value based on the rate of change of the time interval. When collecting the scrolling direction angle, the system calculates the angle between the touch point movement vector and the vertical downward direction of the screen based on the screen coordinates of two consecutive touch events, using this angle as the scrolling direction angle. The trajectory mapping module maps the collected visible area ratio sequence to a two-dimensional plane. The specific mapping method is as follows: For each product item, a two-dimensional plane coordinate system is constructed with time as the horizontal axis and the visible area ratio as the vertical axis. Let the first... The timestamp of each sampling moment is The corresponding visible area ratio is Then the coordinates of the mapped trajectory points are Connect the trajectory points at all sampling times in chronological order to form the time-series trajectory point set for this product item. ,in, The total number of sampling points within the window represents the complete record of the continuous change in the visible area of ​​the product on the screen over time during the user's browsing process. Furthermore, since the original sampling time intervals may be uneven (due to variations in user scrolling speed), the trajectory mapping module, based on the aforementioned mapping, uses linear interpolation to resample the trajectory point set along the time axis, transforming non-uniformly sampled trajectory points into trajectory points with uniform time intervals, meaning the time intervals between adjacent trajectory points are equal, ensuring consistent numerical stability for subsequent curvature calculations. The rollback detection module extracts continuous time windows from the temporal trajectory point set and uses a dual-path temporal alignment network to detect the probability that a trajectory point within the window belongs to a rollback segment. This network consists of a geometric path network, a temporal path network, and a cross-alignment module, where rollback inflection points are identified. Using the baseline, perform time step back and forward exploration to find local maxima that satisfy the sign flip of the first-order difference: roll back the inflection point. The most recent local maximum of the visible area ratio was marked as the rollback start point. The rollback inflection point The moment when the most recent local maximum of the visible area ratio is marked as the rollback endpoint. The closed area Mark it as a complete rollback segment; The interest scoring module calculates the comprehensive non-click interest evaluation value for each product item based on the number of rollback segments and the curvature value at the rollback inflection point in each rollback segment. The recommendation reordering module reorders the original recommendation sequence based on the calculated comprehensive non-click interest evaluation value: products with higher comprehensive non-click interest evaluation values ​​are promoted to the front of the recommendation list, while products with lower evaluation values ​​are moved to the back of the list, and a personalized reordered recommendation list is output.

[0020] In one embodiment of the present invention, the data acquisition module collects the visible area ratio of each product item, obtains the rectangular boundary of the view object corresponding to the product item in the visible area of ​​the screen by calling the view visibility detection interface of the mobile terminal operating system, and calculates the ratio of the intersection area of ​​the rectangular boundary and the screen boundary to the total area of ​​the rectangular boundary as the visible area ratio of the product item at the current sampling time.

[0021] Specifically, when a user scrolls through the product list page, the mobile terminal operating system maintains a corresponding view object for each product item in the list. The system obtains the rectangular boundary of the view object corresponding to the current product item within the visible area of ​​the screen by calling the operating system's view visibility detection interface, such as the getLocalVisibleRect() method in Android or the CGRectIntersection function in iOS. Specifically, let the complete rectangular boundary of the product item view object be Rect. total Its parameter is the left boundary. total Right boundary total , top boundary total and the bottom boundary total The rectangular boundary of the visible area of ​​the screen is called Rect. screen Its parameter is the left boundary. screen (Usually 0), right boundary screen (Screen width), top boundary screen (The area below the status bar) and the bottom boundary screen (Bottom of screen); The system calculates the intersection rectangle (Rect) of the product item view rectangle and the visible screen area rectangle. intersection Left boundary of intersection: left inter =max(left total left screen ); Right boundary of intersection: right inter =min(right total , right screen ); Upper boundary of intersection: top inter =max(top total top screen ); Lower boundary of intersection: bottom inter =min(bottom total bottom screen If left inter <right> inter And top inter <bottom inter Then there exists a non-empty intersection, and the area of ​​the intersection is: If the above conditions are not met (i.e., the product view is not in the screen at all), then ; Get the full area of ​​the item view object ; Calculate the visible area ratio ,in, The value ranges from 0 to 1, where 0 indicates that the product item is completely invisible and 1 indicates that the product item is completely visible. This ratio is the visible area ratio of the product item at the current sampling time. As the product view gradually enters or leaves the screen during the user's scrolling, this ratio value changes continuously over time, forming a visible area ratio sequence. The data acquisition module uses a predetermined sampling frequency. Synchronous data acquisition is performed. In one embodiment, the sampling frequency is preset to 30Hz, which means 30 samples are collected per second. The sampling frequency is determined based on the fact that the sampling rate of mobile terminal touch screens is usually 60Hz-120Hz. 30Hz can ensure that the details of trajectory changes during the user's rapid scrolling process are captured without generating excessive computing overhead and data storage pressure. The sampling frequency can be parameterized through the system configuration file to adapt to mobile terminal devices with different performance levels. The acquisition of scroll acceleration and scroll direction angle is achieved by touch event listening. The system listens to the onTouchEvent callback function to obtain the timestamp and screen coordinates of each touch event. The time interval between two consecutive touch events is calculated based on the timestamps of three consecutive touch events, and the scroll acceleration value is calculated based on the rate of change of the time interval. The angle between the touch point movement vector and the vertical downward direction of the screen is calculated based on the screen coordinates of two consecutive touch events and is used as the scroll direction angle.

[0022] In one embodiment of the present invention, the rollback detection module performs time-aware encoding on the visible area ratio sequence before using a dual-path temporal alignment network for detection, including: Obtain the absolute timestamp for each sampling moment, and extract the hour value from the absolute timestamp. Weekly duty With month value ; The hour value Encoded as and The week value The encoding is and The month value Convert the vectors to seasonal vectors according to a pre-defined seasonal mapping table; By concatenating the cyclic time coordinates with the visible area ratio and the rolling acceleration value with the rolling direction angle, a time-enhanced multidimensional feature vector is formed, which serves as the input to the dual-path temporal alignment network.

[0023] Specifically, the absolute timestamp (Unix timestamp) of each sampling moment is obtained, and three time-dimensional features are parsed from this timestamp: hour value. Values ​​range from 0 to 23, representing the hour of the day corresponding to the sampling time; weekday value. Values ​​range from 0 to 6, where 0 represents Monday and 6 represents Sunday; month value. Values ​​range from 1 to 12; due to the cyclical nature of hourly values ​​(23:00 and 0:00 are adjacent), sine-cosine encoding is used to map hourly values ​​to a two-dimensional continuous space. and ;in, The hour values ​​are mapped to the angular interval [0, 2π). For example, when hour=0, the output is (0, 1); when hour=6, the output is (1, 0); when hour=12, the output is (0, -1); and when hour=18, the output is (-1, 0). This encoding method ensures the continuity of the distance between adjacent hour values ​​in the encoding space, avoiding the unreasonable problem of "23:00 being far from 0:00" caused by One-Hot encoding. The same sine-cosine encoding method as the hour values ​​is used to map the weekday values ​​to a two-dimensional continuous space. and This encoding makes Monday (0) and Sunday (6) adjacent in the encoding space, which conforms to the cyclical nature of the week. The encoding of the seasonal dimension adopts a preset seasonal mapping table. The seasonal division rules of China are as follows: spring is March, April, and May; summer is June, July, and August; autumn is September, October, and November; and winter is December, January, and February. The system converts the month values ​​into a four-dimensional One-Hot seasonal vector. , among which, if ,but ;like ,but And so on; combine the above encoding results into a loop time coordinate vector. The dimension is 2+2+4=8; the cyclic time coordinate vector and the physical signal vector are combined. The parts are then assembled. The proportion of visible area. This is the rolling acceleration value. The rolling direction angle is used to form a time-enhanced multidimensional feature vector. The total number of dimensions is 11. Arrange the time-enhanced multidimensional feature vectors of each sampling time point in chronological order to form a window-level input feature sequence. ,in, This represents the total number of sampling points within the time window; this sequence serves as the input to the dual-path temporal alignment network for subsequent rollback probability detection and classification.

[0024] In one embodiment of the present invention, the construction of the dual-path timing alignment network in the rollback detection module includes: The dual-path temporal alignment network is composed of a geometric path network, a temporal path network, and a cross-alignment module. The input to the geometric path network is the visible area ratio at each sampling time within the window. This value is extracted from the temporally enhanced multidimensional feature vector and processed by a one-dimensional deformable convolutional layer to output the geometric feature vector at each sampling time. The shape of the convolution kernel of the one-dimensional deformable convolutional layer is dynamically adjusted by an offset field, and the input to the offset field is the second-order difference value of the visible area ratio sequence. The input to the temporal path network is the temporally enhanced multidimensional feature vector at each sampling time within the window. This vector is processed by a one-dimensional bidirectional gated recurrent unit to output the temporal feature vector at each sampling time. The cross-alignment module receives the geometric feature vector and the temporal feature vector, calculates the mutual attention weight between the geometric feature vector and the temporal feature vector at each sampling time, and uses the mutual attention weight as a fusion coefficient to perform a weighted summation of the geometric feature vector and the temporal feature vector to output the fused feature vector at each sampling time. The fused feature vector at each sampling time is input into a binary classifier, which outputs the probability that the sampling time belongs to a rollback segment.

[0025] Specifically, the dual-path temporal alignment network consists of three branches: a geometric path network, a temporal path network, and a cross-alignment module. The input is the time window... Time-enhanced multidimensional feature vector sequence at each sampling time point The input to the geometric path network is... Visible area proportion component sequence The network employs a single one-dimensional deformable convolutional layer with a kernel size of K=3 and an output channel count of D=64. Unlike conventional convolution, the sampling position of the deformable convolution kernel is not a fixed grid point, but rather determined by an offset field. Make dynamic adjustments: ,in, This is the index position at the current sampling time. For the input sequence at position The value, For the first The weight parameters at each convolutional kernel position are obtained through network training. For the first The offset of each sampling point; the geometric path network outputs the geometric feature vector at each sampling time. ; The input to the temporal path network is a complete... It includes the visible area ratio, scrolling acceleration value, scrolling direction angle, and time-aware encoding vector. The network uses a single-layer bidirectional gated recurrent unit (BiGRU) with a hidden layer dimension of 64. The BiGRU consists of a forward GRU and a backward GRU: Forward GRU: Backward GRU: ;in, and The first and The forward hidden state at each sampling time point has a dimension of 64. This is a forward-gated loop unit function. For the first The 11-dimensional input feature vector at each sampling time point, and The first and The backward hidden state at each sampling time; The temporal feature vector at each sampling time point is the concatenation of the two: ,in, For the concatenation operation; the temporal path network outputs the temporal feature vector at each sampling time. ; The cross-alignment module receives a sequence of geometric feature vectors. and time series feature vector sequence Calculate the mutual attention weights at each sampling time and perform weighted fusion; for the ... At each sampling time point, geometric features and temporal features are first mapped to the same space using a projection matrix: ; ;in, It is a 64×64-dimensional projection matrix. The projection matrix is ​​64×128 dimensional; calculate the mutual attention weights between geometric features and temporal features: ,in, It is a sigmoid activation function, and the output value is in the interval (0, 1). For the transpose operation, the physical meaning of this formula is: when the geometric eigenvector and the temporal eigenvector have the same direction in the projection space, the dot product value is larger. Approaching 1; when the two directions are opposite or orthogonal, Approaching 0; weighting mutual attention As the fusion coefficient of geometric feature vectors, As fusion coefficients for temporal feature vectors, a weighted fusion feature vector is calculated. Since the dimension of the geometric feature vector (e.g., 64-dimensional) differs from the dimension of the temporal feature vector (e.g., 128-dimensional), a learnable linear projection matrix is ​​used. Upscale the low-dimensional geometric feature vector to a higher dimension to align it with the dimension of the temporal feature vector: ,in, For the first The original geometric feature vector (64-dimensional real column vector) at each sampling time. The projective matrix is ​​a learnable linear projection matrix with a size of 128*64, which maps a 64-dimensional vector to a 128-dimensional vector. This projection matrix is ​​trained together with other network parameters. For the first The projected geometric feature vectors (128-dimensional real column vectors) at each sampling time point are obtained; the projected geometric feature vectors are then weighted and summed with the temporal feature vectors. .

[0026] The fused feature vector at each sampling time point is input into a binary classifier, which outputs the probability that the sampling time point belongs to a rollback segment. This classifier contains a fully connected layer (input dimension 128 → output dimension 1) and a Sigmoid activation function. ,in, For belonging to the first The predicted probability of belonging to the rollback segment at each sampling time. This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, when > The probability threshold (this threshold is determined by the ROC curve on the validation set to maximize the F1 score. In this embodiment, the value is 0.7. The probability threshold is determined by the following method: on the validation set, traverse the threshold range from 0.1 to 0.9 with a step size of 0.01, calculate the F1 score (harmonic average of precision and recall) under each threshold, and select the highest F1 score as the fixed threshold), and mark the sampling time as the rollback inflection point.

[0027] In one embodiment of the present invention, the offset field of the one-dimensional deformable convolutional layer is specifically: ,in, Let be the value of the second-order difference sequence representing the visible area proportion at the nth position of the convolution kernel. The preset deformation amplitude coefficient, It is the hyperbolic tangent activation function.

[0028] Specifically, non-uniformly sampled trajectory points are converted into points with a constant time step through linear interpolation. Based on the uniform sampling sequence, the second-order difference approximation of the visible area ratio is calculated as follows: ,in, , and The first , and The proportion of visible area at each sampling time. The second-order difference represents the time interval between adjacent trajectory points after resampling. Positive values ​​of this second-order difference indicate a convexity in the sequence (V-shaped bottom), while negative values ​​indicate a concaveness (Λ-shaped top). The offset is then calculated using a deformation function. ,in, The preset deformation amplitude coefficient is set to 2 in one embodiment. The core of the formula is that the offset is directly driven by the curvature of the input signal itself, rather than being learned by the network. When the visible area ratio sequence presents a V-shaped trajectory, the second-order difference is a large positive value, the tanh output is close to 1, and the convolution kernel produces the maximum deformation to fit the bottom of the V-shape. When the sequence is flat, the second-order difference is close to 0, and the convolution kernel maintains the standard shape.

[0029] In one embodiment of the present invention, during the training process of the one-dimensional deformable convolutional layer, the loss function includes a cross-entropy main loss term and a physical constraint regularization term: The main cross-entropy loss term is specifically as follows: ,in, For the first The true label at each sampling time The predicted probability is the output of the binary classifier; The physical constraint regularization term includes: a constraint that the first derivative is zero at the rollback inflection point. And the sign constraint that the proportion of visible area within the rollback fragment first decreases and then increases. ,in, and These are the signs of the first-order differences before and after the inflection point, respectively. The total loss function is as follows: ,in, and These are the preset weighting coefficients.

[0030] Specifically, during training, a loss function with physical constraints is used, assuming that the training samples contain the labeled true labels. The predicted probabilities of and . The physical constraint regularization term consists of two parts: First, the constraint that the first derivative is zero at the rollback inflection point: Let the fitted function be... (Obtained through cubic spline interpolation), the rollback inflection point is... This yields a constraint that the first derivative is zero. ,in, For the fitted function in The first derivative value at the point forces the network's prediction derivative to approach 0 at the rollback inflection point; second, the sign constraint of the visible area ratio within the rollback segment first decreasing and then increasing: let the first difference before the inflection point be... The first difference after the inflection point is The sign constraint was obtained where the visible area ratio within the rollback fragment first decreased and then increased. ,when and hour, , This results in a penalty item; At that time, the penalty item is 0; Among them, in the total loss function and In this embodiment, the preset weighting coefficients are used. and The values ​​are 0.1 and 0.05, respectively, which are determined on the validation set through grid search. The network is trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 iterations. During training, the F1 score of the rollback detection is calculated on the validation set after each round, and the model with the highest F1 score is selected as the final model.

[0031] In one embodiment of the present invention, the cross-alignment module calculates the mutual attention weight between the geometric feature vector and the temporal feature vector at each sampling time, including: For the At each sampling time, its geometric feature vector With time series feature vectors Mutual attention weights between The calculation is as follows: ,in, and Let be the projection matrix. It is a sigmoid activation function. This is a transpose operation; The mutual attention weights As the fusion coefficient of geometric feature vectors, As the fusion coefficients of the temporal feature vectors, they are obtained through a linear projection matrix. Increase the dimensionality of the geometric feature vectors, according to... Calculate the first The fused feature vector at each sampling time point .

[0032] In one embodiment of the present invention, when the binary classifier outputs the probability that the sampling time belongs to the rollback segment, it simultaneously uses the Monte Carlo dropout method to calculate the prediction confidence interval of the probability. The Monte Carlo dropout method uses a variational dropout mechanism and repeats the forward propagation T times to obtain T probability values. The standard deviation of the T probability values ​​is taken as the prediction confidence interval. For each sampling time, the prediction confidence interval is denoted as... Let the probability value output by its binary classifier be denoted as Calculate the confidence weight of the rollback segment corresponding to the sampling time, specifically as follows: ,in, This is the preset sensitivity coefficient; The confidence weight This serves as the weight multiplier for the rolled-back segment in the interest scoring module.

[0033] Specifically, when the binary classifier outputs the probability that the sampling time belongs to a rollback segment, it simultaneously employs a Monte Carlo dropout method with a variational dropout mechanism to generate random masks for the linear layers and BiGRU recurrent layers in the network during forward propagation, according to a fixed dropout rate. Specifically, for the BiGRU module in the temporal path network, the same random dropout mask is applied to all sampling times within the same sequence window to maintain the temporal continuity of the recurrent neural network's hidden states. Between each complete forward propagation (a total of T times), sampling is performed using independent and different dropout masks. Since the location of the dropped neuron is different each time, T forward propagations yield T slightly different output probability values; let the th... The probability value obtained from the second forward propagation is Then, after T propagations, we obtain the set of probability values. ; Calculate the mean and standard deviation of T probability values. In this embodiment, T is 20 times, that is, the same input sample is subjected to 20 forward propagations with random dropouts during inference, and the standard deviation of the 20 probability values ​​is calculated as the prediction confidence interval. For each sampling time, its prediction confidence interval is denoted as... Let the probability value output by its binary classifier be denoted as (The average of T probability values ​​can be taken as the final output probability), and the confidence weight of the rollback segment corresponding to this sampling time is calculated according to the following formula: , The confidence weights range from 0 to 1. To predict the confidence interval, where, The preset sensitivity coefficient is used to control the confidence weights. The degree of sensitivity, The larger the value, the higher the confidence weight. The faster the decay; The smaller the value, the slower the decay; in this embodiment, The value is 5.0, and this value is determined by testing on the validation set. Within the range of 1.0 to 10.0, the option that maximizes the overall recommendation accuracy is selected. Value; the physical meaning of this formula is: when When the prediction is completely certain, The confidence weight is at its maximum value; when When it increases, Monotonically decreasing; when When it approaches infinity, The formula tends to 0, and by predicting rollback segments with higher uncertainty, the confidence weight is lower.

[0034] In one embodiment of the present invention, the interest scoring module calculates the comprehensive non-click interest evaluation value for each product item, including: For a single item, obtain the total number of rollback segments that have been confirmed as valid in the current session. , among which, when When this happens, the overall non-click interest rating value of the product item is assigned to 0; when At that time, for the first For each rollback segment, obtain the curvature value at the rollback inflection point of the rollback segment. Obtain the confidence weight of the corresponding rollback segment. ; Calculate the first Weighted curvature value of each rollback segment Calculate the sum of the weighted curvature values ​​of all rollback segments. And the sum of the confidence weights of all rollback segments. The sum of the weighted curvature values The sum of the confidence weights The ratio of [the ratio of the ...

[0035] Specifically, the comprehensive non-click interest rating of the product item. This formula ensures that rollback segments with high confidence weights dominate the final score, while rollback segments with low confidence weights have a weakened impact on the technical effect.

[0036] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A personalized product recommendation system for e-commerce platforms based on user behavior sequences, characterized in that, Includes the following modules: Data acquisition module: used to synchronously collect the visible area ratio, scrolling acceleration value and scrolling direction angle of each product item at a predetermined sampling frequency while the user scrolls the product list page; Trajectory mapping module: used to map the visible area ratio sequence to a two-dimensional plane to obtain a time-series trajectory point set with time as the horizontal axis and visible area ratio as the vertical axis; Rollback detection module: used to extract continuous time windows from the time-series trajectory point set, use a dual-path time alignment network to detect the probability that the trajectory points in the window belong to the rollback segment, mark the sampling time when the probability is greater than the probability threshold as the rollback inflection point, and backtrack to the extreme points of the first-order difference sign flipping at both ends, and mark the continuous closed interval formed as the rollback segment. Interest scoring module: used to calculate the comprehensive non-click interest evaluation value of each product item based on the number of rollback segments and the curvature value at the rollback inflection point in each rollback segment; Recommendation reordering module: used to reorder the original recommendation sequence based on the comprehensive non-click interest evaluation value; The construction of the dual-path timing alignment network in the rollback detection module includes: The dual-path temporal alignment network is composed of a geometric path network, a temporal path network, and a cross-alignment module. The input to the geometric path network is the visible area ratio at each sampling time within the window. This value is extracted from the temporally enhanced multidimensional feature vector and processed by a one-dimensional deformable convolutional layer to output the geometric feature vector at each sampling time. The shape of the convolution kernel of the one-dimensional deformable convolutional layer is dynamically adjusted by an offset field, and the input to the offset field is the second-order difference value of the visible area ratio sequence. The input to the temporal path network is the temporally enhanced multidimensional feature vector at each sampling time within the window. This vector is processed by a one-dimensional bidirectional gated recurrent unit to output the temporal feature vector at each sampling time. The cross-alignment module receives the geometric feature vector and the temporal feature vector, calculates the mutual attention weight between the geometric feature vector and the temporal feature vector at each sampling time, and uses the mutual attention weight as a fusion coefficient to perform a weighted summation of the geometric feature vector and the temporal feature vector to output the fused feature vector at each sampling time. The fused feature vector at each sampling time is input into a binary classifier, which outputs the probability that the sampling time belongs to a rollback segment.

2. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, In the data acquisition module, the visible area ratio of each product item is collected. By calling the view visibility detection interface of the mobile terminal operating system, the rectangular boundary of the view object corresponding to the product item in the visible area of ​​the screen is obtained. The ratio of the intersection area of ​​the rectangular boundary and the screen boundary to the total area of ​​the rectangular boundary is calculated as the visible area ratio of the product item at the current sampling time.

3. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, In the rollback detection module, before employing a dual-path temporal alignment network for detection, time-aware encoding is performed on the visible area ratio sequence, including: Obtain the absolute timestamp for each sampling moment, and extract the hour value from the absolute timestamp. Weekly duty With month value ; The hour value Encoded as and The week value The encoding is and The month value Convert the vectors to seasonal vectors according to a pre-defined seasonal mapping table; By concatenating the cyclic time coordinates with the visible area ratio and the rolling acceleration value with the rolling direction angle, a time-enhanced multidimensional feature vector is formed, which serves as the input to the dual-path temporal alignment network.

4. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, The offset field of the one-dimensional deformable convolutional layer is specifically as follows: ,in, Let be the value of the second-order difference sequence representing the visible area proportion at the nth position of the convolution kernel. The preset deformation amplitude coefficient, It is the hyperbolic tangent activation function.

5. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, During the training of the one-dimensional deformable convolutional layer, the loss function includes a cross-entropy main loss term and a physical constraint regularization term: The main cross-entropy loss term is specifically as follows: ,in, For the first The true label at each sampling time The predicted probability is the output of the binary classifier; The physical constraint regularization term includes: a constraint that the first derivative is zero at the rollback inflection point. And the sign constraint that the proportion of visible area within the rollback fragment first decreases and then increases. ,in, and , respectively, are the signs of the first-order differences before and after the inflection point; where, For the fitted function in The first derivative value at; The total loss function is as follows: ,in, and These are the preset weighting coefficients.

6. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, In the cross-alignment module, the mutual attention weights between the geometric feature vector and the temporal feature vector at each sampling time are calculated, including: For the At each sampling time, its geometric feature vector With time series feature vectors Mutual attention weights between The calculation is as follows: ,in, and For the projection matrix, It is a sigmoid activation function. This is a transpose operation; The mutual attention weights As the fusion coefficient of geometric feature vectors, As the fusion coefficients of the temporal feature vectors, they are obtained through a linear projection matrix. Increase the dimensionality of the geometric feature vectors, according to... Calculate the first The fused feature vector at each sampling time point .

7. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, When the binary classifier outputs the probability that the sampling time belongs to the rollback segment, it simultaneously uses the Monte Carlo dropout method to calculate the prediction confidence interval of the probability. The Monte Carlo dropout method uses a variational dropout mechanism and repeats the forward propagation T times to obtain T probability values. The standard deviation of the T probability values ​​is taken as the prediction confidence interval. For each sampling time, the prediction confidence interval is denoted as... Let the probability value output by its binary classifier be denoted as Calculate the confidence weight of the rollback segment corresponding to the sampling time, specifically as follows: ,in, This is the preset sensitivity coefficient; The confidence weight This serves as the weight multiplier for the rolled-back segment in the interest scoring module.

8. The personalized product recommendation system for e-commerce platforms based on user behavior sequences according to claim 1, characterized in that, The interest scoring module calculates the comprehensive non-click interest rating for each product item, including: For a single item, obtain the total number of rollback segments that have been confirmed as valid in the current session. , among which, when When this happens, the overall non-click interest rating value of the product item is assigned to 0; when At that time, for the first For each rollback segment, obtain the curvature value at the rollback inflection point of the rollback segment. Obtain the confidence weight of the corresponding rollback segment. ; Calculate the first Weighted curvature value of each rollback segment Calculate the sum of the weighted curvature values ​​of all rollback segments. And the sum of the confidence weights of all rollback segments. The sum of the weighted curvature values The sum of the confidence weights The ratio of [the ratio of the ...

Images