Adaptive information flow recommendation method based on multi-modal behavior timing and dynamic image

Abstract

The application relates to the technical field of information flow recommendation, and discloses an adaptive information flow recommendation method based on multi-modal behavior time sequence and dynamic portraits, which comprises the following steps: performing density clustering on a user multi-modal behavior event stream, generating a time sequence context cluster label, segmenting a context exclusive behavior subsequence, and counting the event density of each mode in each time sequence context cluster; adaptively selecting a spectrum analysis method according to the mode event density in the context, generating a spectrum power distribution vector, and calculating a spectrum estimation confidence score in the context; generating an interest confidence vector through an interest phase classification network, performing cross-modal phase confidence propagation and cross-context prior shrinkage correction, and maintaining a context exclusive dynamic portrait in a partition memory unit; matching a current recommendation request to a corresponding time sequence context cluster, fusing context exclusive interest intensity and shared interest intensity, and outputting an adaptive recommendation ranking list.

Description

Technical Field

[0001] This invention relates to the field of information flow recommendation technology, specifically to an adaptive information flow recommendation method based on multimodal behavioral time series and dynamic profiling. Background Technology

[0002] Information flow recommendation systems generate dynamic user profiles by analyzing users' historical behavior sequences, and then match and rank these dynamic profiles with candidate content to push personalized content to users. Existing methods typically perform spectral analysis on all multimodal behavior sequences of users uniformly, select a spectral estimation algorithm based on the global event density of the modality, and write all behaviors into a unified memory network to maintain a single dynamic profile.

[0003] However, user behavior timelines exhibit systematic density differences across different usage scenarios. When global density is used as the basis for selecting spectral methods, it masks the true sparsity within a specific scenario, leading to some modalities employing unsuitable spectral estimation methods in particular scenarios, resulting in unreliable spectral estimation results. Simultaneously, unified dynamic profiles cannot distinguish context-specific interests; interests from different scenarios interfere with each other within the same memory network, causing context-specific interests to be erroneously attenuated. Even with partitioned memory to maintain independent profiles for each scenario, in sparse scenarios, the lack of reliable interest update signals keeps partitioned profiles outdated or contaminated by spectral noise, ultimately resulting in insufficient adaptive capability of information flow recommendations across different scenarios. Summary of the Invention

[0004] This invention provides an adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profile, which solves the technical problem in related technologies that cannot adaptively estimate the interest distribution based on the temporal context differences and modal sparsity of user behavior data, resulting in insufficient recommendation ranking accuracy.

[0005] This invention discloses an adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling, comprising: Obtain the temporal context feature vector of the user's multimodal behavior event stream, perform density clustering to generate temporal context cluster labels, segment the multimodal behavior sequence into context-specific behavior subsequences according to the temporal context cluster labels, and count the event density of each modality within each temporal context cluster; For each modal behavior subsequence within each time series context cluster, the spectral analysis method is adaptively selected based on the event density of that modality within the time series context cluster to generate a spectral power distribution vector; The spectral power distribution vector is input into the interest phase classification network, which outputs the phase classification probability distribution. The phase classification probability distributions of all modes within the same temporal context cluster are aggregated to generate the interest confidence vector of each temporal context cluster, which is then written into the corresponding partition memory unit. Extract the temporal context feature vector of the current recommendation request, match it to the nearest temporal context cluster, read the interest intensity distribution from the corresponding partition memory unit, perform weighted matching calculation with the candidate information stream content feature vector, and output the recommendation ranking list.

[0006] Furthermore, the generation process of the temporal context feature vector includes: obtaining the timestamp, device type, and interaction duration features of each event; performing encoding processing on the device type to convert it into a numerical representation; performing normalization processing on the timestamp value and the interaction duration value respectively; and concatenating the processed features into the temporal context feature vector.

[0007] Furthermore, the adaptive selection spectral analysis method includes: when the event density of a certain mode in a certain time-series context cluster is lower than the sparsity threshold within the context, the Lomb-Scargle periodogram method is used to estimate the spectral power of the irregular timestamp signal; when the event density is higher than the sparsity threshold within the context, the short-time Fourier transform is used to extract spectral features.

[0008] Furthermore, it also includes calculating the confidence score of the spectral estimation within the context. :

[0009] in, For modality In time-series context clusters The number of valid events within, For modality In time-series context clusters The dimensionless value of the normalized variance of the spectral estimation within the range. To adjust the parameters.

[0010] Furthermore, the process of aggregating and generating the interest confidence vector for each temporal context cluster introduces cross-modal phase position confidence propagation, including: multiplying the phase classification probability distribution with the context-specific spectral estimation confidence score to generate a confidence-weighted phase probability distribution; determining high-confidence and low-confidence modes based on the context-specific spectral estimation confidence score; using the confidence-weighted phase probability distribution of the high-confidence mode as a priori, performing Bayesian correction on the phase probability distribution of the low-confidence mode; and aggregating the mean of the corrected phase probability distributions for each mode to generate the propagated and corrected interest confidence vector.

[0011] Furthermore, it also includes cross-contextual prior contraction correction: calculating the weighted average of the confidence scores of the intra-contextual spectral estimates for all modalities within each time-series context cluster as an overall spectral reliability index. When the overall spectral reliability index of a certain time-series scenario cluster is lower than the partition update threshold, it is marked as a sparse scenario; the high-reliability time-series scenario cluster with the highest cosine similarity to the interest confidence vector of the sparse scenario is retrieved; using the interest confidence vector of the high-reliability time-series scenario cluster as a cross-scenario prior, the interest confidence vector of the sparse scenario is shrunk and corrected.

[0012] in, This represents the original interest confidence vector for sparse scenarios. This represents the interest confidence vector for a high-reliability time-series scenario cluster. The shrinkage coefficient, The overall spectral reliability index for high-reliability time-series scenario clusters. This is an overall spectral reliability index for sparse scenarios.

[0013] Furthermore, it also includes a context-bound labeling process: calculating the dimension-wise similarity of interest confidence vectors among all temporal context clusters; labeling interest dimensions with high cross-context similarity as context-independent interests and performing intensity synchronization updates across memory units in each partition; labeling interest dimensions with high confidence that only appear in specific temporal context clusters as context-bound interests, restricting their cross-partition propagation; and for each interest dimension... Generate context binding coefficient :

[0014] in, Theme The standard deviation of the interest confidence component across all time-series context clusters. The corresponding mean, To prevent smoothing constants with denominators of zero.

[0015] Furthermore, the process of reading the interest intensity distribution from the corresponding partition memory unit also includes: reading the interest intensity distribution of the context-specific dynamic profile and the cross-partition shared intensity distribution of context-independent interests, and utilizing the context binding coefficient. As a blending weight:

[0016] in, For the interest dimension The specific strength of the current context partition, For the interest dimension Cross-partition sharing strength, The interest weights are fused; the fused dynamic profile interest weight distribution is weighted and matched with the candidate information stream content feature vectors to calculate the recommended ranking list.

[0017] Furthermore, the theme In low confidence mode Down-corrected phase probability distribution Obtain it using the following formula:

[0018] in, Theme In modality Lower phase The original classification probability, Theme In high confidence mode Lower phase The classification probability, For modality In time-series context clusters The confidence score of the intra-context spectrum estimation is calculated, with the denominator being a normalization term. When there are multiple high-confidence modes within the time-series context cluster, the correction is performed on the low-confidence modes in descending order of their intra-context spectrum estimation confidence scores.

[0019] This invention discloses an adaptive information flow recommendation system based on multimodal behavioral time series and dynamic profiling, used to execute the above method, including: The context segmentation module is used to obtain the temporal context feature vector of the user's multimodal behavior event stream, perform density clustering to generate temporal context cluster labels, segment the multimodal behavior sequence into context-specific behavior subsequences according to the temporal context cluster labels, and count the event density of each modality within each temporal context cluster. The spectrum analysis module is used to adaptively select the spectrum analysis method for each modal behavior subsequence within each time series context cluster based on the event density, generate a spectrum power distribution vector, and calculate the confidence score of spectrum estimation within the context. The interest confidence generation and memory maintenance module is used to input the spectral power distribution vector into the interest phase classification network to generate a phase classification probability distribution, aggregate and generate interest confidence vectors for each temporal context cluster, perform cross-modal phase position confidence propagation and cross-context prior contraction correction, and maintain context-specific dynamic profiles in each partition memory unit. The recommendation ranking module is used to extract the temporal context feature vector of the current recommendation request and match it to the nearest temporal context cluster. It reads the interest intensity distribution from the corresponding partition memory unit, performs weighted matching calculation with the candidate information stream content feature vector, and outputs the recommendation ranking list.

[0020] The beneficial effects of this invention are as follows: This invention refines the granularity of spectrum analysis to the modal level within the time-series context cluster by judging the sparsity of the temporal behavior of each modality within the time-series context cluster and adaptively selecting the spectrum estimation method. This avoids the problem of global density masking the true sparsity within the time-series context cluster, which leads to the incompatibility of the spectrum method and improves the reliability of the spectrum estimation results.

[0021] Through cross-modal phase position confidence propagation within the context, the spectral results of high-confidence modes correct the phase classification bias of low-confidence modes using Bayesian priors, enabling the interest phase judgment of sparse modes to obtain reliable signal support from dense modes. Through cross-contextual prior contraction correction, the interest confidence vector of sparse scenarios obtains directional correction signals from clusters of highly reliable time-series scenarios with similar interests, and the correction magnitude is constrained by the reliability difference. This ensures that the dynamic profile of sparse scenarios no longer becomes outdated due to insufficient effective events, nor is it subject to unreliable updates due to spectral noise.

[0022] Context-binding coefficients are used to distinguish between context-independent and context-bound interests, and cross-regional propagation boundaries are controlled separately to prevent correction signals from high-reliability temporal context clusters from improperly covering the unique interest features of sparse contexts. During the recommendation phase, the temporal context feature vector of the current recommendation request is automatically matched to the corresponding temporal context cluster, and the cross-regional shared intensity distribution of context-specific interests and context-independent interests is fused. This allows the information flow recommendation to automatically switch to the corresponding dynamic profile partition under different usage scenarios, achieving context-adaptive information flow recommendation ranking. Attached Figure Description

[0023] Figure 1 This is a flowchart of an adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profile provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the density statistics of each modal event within each temporal context cluster provided in the embodiments of the present invention; Figure 3 This is a schematic diagram comparing the time span and click density of each temporal context cluster provided in the embodiments of the present invention; Figure 4 This is a schematic diagram illustrating the confidence scores of the spectrum estimation for each mode in scenarios A and C provided in the embodiments of the present invention; Figure 5 This is a schematic diagram comparing the phase probability of the technology-themed commentary modality before and after correction in Scenario A provided by an embodiment of the present invention; Figure 6 This is a schematic diagram of the overall spectral reliability index of each time-series scenario cluster provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the context binding coefficients for each topic dimension provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the fusion result of the interest intensity of each topic in scenario C provided by the embodiment of the present invention. Detailed Implementation

[0024] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0025] At least one embodiment of the present invention discloses an adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling, see [link to relevant documentation]. Figure 1 This includes the following steps: Step 1, Context Segmentation and Event Density Statistics: The system acquires the timestamp, device type, and interaction duration features of each event in the user's multimodal behavior event stream, and concatenates these features into a temporal context feature vector. Device type is non-numerical categorical data, and must be encoded and converted to a numerical representation before concatenation. Timestamp values ​​and interaction duration values ​​have different dimensions, and must be normalized separately before concatenation to eliminate the impact of these dimension differences on subsequent clustering operations. Density clustering is performed on the temporal context feature vector to generate several temporal context cluster labels. The multimodal behavior sequence is segmented into context-specific behavior subsequences according to the temporal context cluster labels, and the event density of each behavior modality within each temporal context cluster is calculated.

[0026] It should be noted that density-based clustering employs a density-based spatial clustering method. Its input is the temporal context feature vector of each event, and its output is the temporal context cluster label to which each event belongs. The density-based spatial clustering method automatically discovers sets of events with similar temporal context features and labels them as the same temporal context cluster. Different temporal context clusters correspond to user behavior patterns in different usage scenarios; for example, morning commutes, lunch breaks, evening deep reading, and weekend browsing are clustered into different temporal context clusters.

[0027] It should be noted that the event density of each behavioral modality is the ratio of the number of events of a certain modality within that temporal context cluster to the time span of that temporal context cluster. Behavioral modalities here include, but are not limited to, interaction types such as clicking, sharing, commenting, saving, and browsing duration records.

[0028] Step 2, Context-dependent adaptive spectrum analysis: For each modal behavior subsequence within each time-series context cluster, the spectral analysis method is adaptively selected based on the event density of that modality within the time-series context cluster to generate the spectral power distribution vector of each modality and topic within each time-series context cluster.

[0029] It should be noted that before performing spectral analysis, the behavioral subsequences of each modality are clustered and projected along the topic dimension, mapping the behavioral events to the corresponding topic space, and obtaining the intensity signal of each topic with irregular timestamps in each modality. Spectral analysis is applied to the intensity signal of each topic dimension, and the generated spectral power distribution vector represents the periodic power characteristics of each topic in each modality.

[0030] It should be noted that the rule for adaptively selecting the spectral analysis method is as follows: when the event density of a certain mode within a certain time series context is lower than the sparsity threshold within the context, the Lomb-Scargle periodogram method is used to directly process the irregular timestamp signal for spectral power estimation; when the event density is higher than the sparsity threshold within the context, the short-time Fourier transform is used to extract spectral features. The sparsity threshold within the context is a preset density parameter used to distinguish between sparse and dense modes within that time series context. The Lomb-Scargle periodogram method does not require interpolation and resampling of irregularly sampled signals and is suitable for sparse event sequences. The short-time Fourier transform is suitable for dense modes where the event density is sufficient and the signal can be approximated as uniformly sampled.

[0031] In this embodiment, to quantify the reliability of the spectrum estimation results to support subsequent correction processes, an in-context spectrum estimation confidence score is calculated based on the inverse relationship between the number of valid events in each modality within each time-series scenario cluster and the variance of the spectrum estimation. In-context spectrum estimation confidence score The calculation formula is as follows:

[0032] in, For modality In time-series context clusters The number of valid events within, For modality In time-series context clusters The dimensionless value of the normalized variance of the spectral estimation within the range. To adjust the parameters, For modal indexing, This is a time-series context cluster index. The larger the number of valid events and the smaller the variance of the spectral estimation, the closer the confidence score of the spectral estimation within the context is to 1; conversely, it is closer to 0.

[0033] Step 3, Interest Confidence Vector Generation and Partition Memory Maintenance: The spectral power distribution vectors of each modality within each temporal context cluster are input into the interest phase classification network, which outputs the phase classification probability distribution of each topic under each modality. The phase classification probability distributions of all modalities within the same temporal context cluster are aggregated by mean summation to generate the interest confidence vector for that temporal context cluster. The interest confidence vectors of each temporal context cluster are written into the corresponding partitioned memory units to maintain a context-specific dynamic profile.

[0034] It should be noted that the interest phase classification network is a multi-layer fully connected network. The input to the network is the spectral power distribution vector of a single modality within a single temporal context cluster, and the output is the phase classification probability distribution of each topic within that modality. The output layer uses the Softmax activation function. The phase classification probability distribution represents the periodic stage of a user's interest in each topic. During training, the real interest phase labels of each topic within each modality are used as supervision signals, and the cross-entropy loss function is employed. The Adam optimization algorithm is used to update the parameters of the interest phase classification network.

[0035] It should be noted that each partitioned memory unit is an independent storage area divided according to the time-series context cluster labels. Each partitioned memory unit stores the interest confidence vector corresponding to a time-series context cluster and its historical update records. By default, the partitioned memory units are isolated from each other, and the dynamic profile of each time-series context cluster is maintained independently.

[0036] In this embodiment, to correct the classification bias of the low-confidence mode using the spectral results of the high-confidence mode, a cross-modal phase position confidence propagation process is introduced when generating the interest confidence vector. Step 3 may specifically include: Phase classification probability distribution and in-context spectral estimation confidence score Multiply to generate a confidence-weighted phase probability distribution. Within each time-series context cluster, perform cross-modal phase position confidence propagation for different modalities of the same topic. The cross-modal phase position confidence propagation process uses the confidence-weighted phase probability distribution of the high-confidence modality as a prior and performs Bayesian correction on the low-confidence modality.

[0037] Let the mode For high-confidence modes, the mode For low-confidence modes, the time-series context clusters are labeled as follows: Both the high-confidence and low-confidence modes are estimated based on the in-context spectrum obtained in step 2 to determine their confidence scores. The system determines whether a mode has a confidence score for its spectral estimation above a preset threshold as a high-confidence mode, or a mode has a confidence score below the preset threshold as a low-confidence mode. (Topic) In modality Down-corrected phase probability distribution Obtain it using the following formula:

[0038] in, Theme In modality Lower phase The original classification probability, Theme In modality Lower phase The classification probability, For modality In time-series context clusters Confidence score for in-context spectral estimation. For phase category index, To sum the phase category indices during the traversal, a normalization term is used in the denominator to ensure that the sum of the corrected phase probability distributions is 1. When multiple high-confidence modes exist within the time-series context cluster, the above correction is applied to low-confidence modes sequentially based on their in-context spectral estimation confidence scores, from highest to lowest. The corrected phase probability distributions for each mode are then aggregated by mean to generate a propagation-corrected in-context interest confidence vector.

[0039] It should be noted that each component of the interest confidence vector corresponds to a topic, and its value is a scalar obtained by aggregating the mean of the phase probability distributions of that topic after correction across all modalities. This reflects the user's overall confidence level in the topic. The dimension of the interest confidence vector is equal to the total number of topics. That is, vector ,in The total number of themes in a pre-determined theme space. For temporal context clusters Interest confidence vector.

[0040] In this embodiment of the application, to address the problem of unreliable interest confidence vectors due to insufficient valid events in sparse scenarios, a cross-scenario prior contraction correction process is further included in step 3. The cross-scenario prior contraction correction process specifically includes: Calculate the overall spectral reliability index for each time series scenario cluster. Overall spectrum reliability indicators Defined as the weighted average of the confidence scores of the spectral estimates within all modalities in this time series context cluster:

[0041] in, For modality The weights satisfy , Based on the presupposition of the importance of modality in interest representation For modal indices, sum and iterate through all modalities. For modality In time-series context clusters Confidence score for in-context spectral estimation.

[0042] When a certain time-series context cluster Overall spectrum reliability index When the data falls below the partition update threshold, the time-series scenario cluster is marked as a sparse scenario. Calculate the sparse scenario. The cosine similarity between the interest confidence vector and the interest confidence vectors of all other high-reliability time-series context clusters is used to retrieve the high-reliability time-series context cluster with the highest cosine similarity. Clusters of high-reliability time-series scenarios. Interest confidence vectors serve as cross-contextual priors for sparse contexts. The interest confidence vector is shrunk and corrected.

[0043] in, For sparse scenarios The original interest confidence vector, For high-reliability timing scenario clusters Interest confidence vectors, both of which are based on the total number of topics. A vector of dimension 1. The shrinkage coefficient, For high-reliability timing scenario clusters The overall spectrum reliability index, For sparse scenarios The overall spectrum reliability index, The magnitude of the shrinkage correction is proportional to the reliability difference, representing the reliability difference. For the sparse situation after shrinkage correction Interest confidence vectors are generated. Reliability-enhanced interest confidence vectors for each temporal context cluster are generated, and these enhanced vectors are then updated to the corresponding partitioned memory units.

[0044] In this embodiment, to distinguish the differentiated behavior of different interest dimensions during cross-context propagation and to prevent the correction signal of a high-reliability temporal context cluster from improperly covering the unique interest features of a sparse context, context binding tags are introduced during the partitioned memory maintenance process. The context binding tagging process may specifically include: Calculate the dimension-wise similarity of interest confidence vectors among all temporal context clusters. Interest dimensions with high cross-context similarity are labeled as context-independent interests. For context-independent interest dimensions, intensity synchronization updates are performed across partition memory units, ensuring that context-independent interests share the latest intensity values. Interest dimensions with high confidence only appearing in specific high-reliability temporal context clusters are labeled as context-bound interests, restricting the cross-partition propagation of context-bound interests. For each interest dimension... Generate context binding coefficient , The value ranges from 0 to 1. Indexed by topic.

[0045] It should be noted that the context binding coefficient By calculating the topic The ratio of the cross-contextual standard deviation to the mean of the corresponding interest dimension in the interest confidence vector of each temporal context cluster is obtained. Let the topic be... In time-series context clusters Interest confidence component Then calculate the standard deviation of this dimension across all time series context clusters. with the mean The context binding coefficient is defined as:

[0046] in, Theme The standard deviation of the interest confidence component across all time-series context clusters. The corresponding mean, To prevent smoothing constants with denominators of zero, Theme In time-series context clusters Interest confidence component. A value close to 0 indicates that the interest dimension has little difference across different temporal context clusters, and is a context-independent interest. A value close to 1 indicates that the interest dimension varies significantly across different temporal context clusters, thus binding the interest to the context. Output the updated dynamic profile memory state for each partition.

[0047] Step 4, Context-Adaptive Recommendation Ranking: Extract the temporal context feature vector of the current recommendation request and match it to the nearest temporal context cluster. Read the interest intensity distribution of the context-specific dynamic profile from the corresponding partition memory unit, perform weighted matching calculation with the candidate information stream content feature vector, and output an adaptive recommendation ranking list.

[0048] It should be noted that the method for matching to the nearest temporal context cluster is to calculate the distance between the temporal context feature vector of the current recommendation request and the center of each temporal context cluster, and select the temporal context cluster with the smallest distance as the current context. The temporal context feature vector of the current recommendation request and the temporal context feature vectors of each event in step 1 use the same feature construction method, that is, perform the same encoding processing on the device type, and perform the same normalization processing on the timestamp value and the interaction duration value, so as to ensure the consistency of the feature space and make the distance calculation effective.

[0049] In this embodiment of the application, given that context binding coefficients have already been generated, the interest intensity fusion process in step 4 further includes: reading the interest intensity distribution of the context-specific dynamic profile from the corresponding partition memory unit, and simultaneously reading the cross-partition shared intensity distribution of context-independent interests, and utilizing the context binding coefficients. The interest intensity distribution of context-specific dynamic profiles is fused with the cross-regional shared intensity distribution of context-independent interests as a hybrid weight to generate the dynamic profile interest weight distribution in the current context. The fusion formula is as follows:

[0050] in, For the interest dimension The specific strength of the current context partition, For the interest dimension Cross-partition sharing strength, For the merged interest weights, For the interest dimension Context binding coefficient, Indexed by topic. When When the value approaches 1, the fusion result is mainly based on the interest intensity distribution of context-specific dynamic profiles; when... When the value approaches 0, the fusion result is dominated by the cross-regional shared intensity distribution of context-independent interests. The fused dynamic profile interest weight distribution is then weighted and matched with the candidate information stream content feature vectors to calculate an adaptive recommendation ranking list.

[0051] On the other hand, this invention also proposes an adaptive information flow recommendation system based on multimodal behavioral time series and dynamic profiles, including: The context segmentation module is used to obtain the temporal context features of the user's multimodal behavior event stream, perform density clustering to generate temporal context cluster labels, segment the multimodal behavior sequence into context-specific behavior subsequences according to the temporal context cluster labels, and count the event density of each modality within each temporal context cluster.

[0052] The spectrum analysis module is used to adaptively select the spectrum analysis method for each modal behavior subsequence within each time series context cluster based on the modal event density within the time series context cluster, generate the spectrum power distribution vector for each modality and topic within each time series context cluster, and calculate the confidence score of the spectrum estimation within the context.

[0053] The interest confidence generation and memory maintenance module is used to input the spectral power distribution vector into the interest phase classification network to generate the phase classification probability distribution, aggregate and generate the interest confidence vector of each temporal context cluster, perform cross-modal phase position confidence propagation and cross-context prior contraction correction, maintain the context-specific dynamic profile in each partition memory unit, and generate context binding coefficients.

[0054] The recommendation ranking module is used to extract the temporal context feature vector of the current recommendation request and match it to the nearest temporal context cluster. It reads the context-specific dynamic portrait interest intensity distribution of the corresponding partition memory unit, uses the context binding coefficient to fuse the interest intensity distribution of the context-specific dynamic portrait with the cross-partition shared intensity distribution of context-independent interests, performs weighted matching calculation with the candidate information stream content feature vector, and outputs an adaptive recommendation ranking list.

[0055] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0056] This invention determines the sparsity of modal behavior at the level of the temporal context cluster, rather than the global modality level, and adaptively selects the spectral estimation method, thus refining the granularity of spectral analysis from the global modality level to the modality level within the temporal context cluster. Therefore, the same modality can be analyzed using a spectral method that matches its actual event density in different temporal context clusters, avoiding the problem of global density masking the true sparsity within the temporal context cluster, which leads to an unsuitable spectral method.

[0057] Through cross-modal phase position confidence propagation within the context, the spectral results of high-confidence modes correct the phase classification bias of low-confidence modes using Bayesian priors, ensuring that the interest phase judgment of sparse modes receives reliable signal support from dense modes. Through cross-contextual prior contraction correction, the interest confidence vector of sparse scenarios obtains directional correction signals from clusters of high-reliability time-series scenarios with similar interests, and the correction magnitude is constrained by the reliability difference. Therefore, the regional dynamic profile of sparse scenarios is no longer outdated due to insufficient effective events, nor is it subject to unreliable updates due to spectral noise.

[0058] Context-bound interests are distinguished from context-independent interests by a context-binding coefficient, and cross-partition propagation boundaries are controlled separately. Context-independent interests are shared and updated across all partitions, while context-bound interests are confined to their corresponding partitions. Therefore, the correction signal of the high-reliability temporal context cluster will not unduly cover the unique interest features of sparse contexts. During the recommendation phase, the temporal context feature vector of the current recommendation request is automatically matched to the corresponding temporal context cluster. The context-specific dynamic profile of that partition's memory unit is read, and the interest intensity distribution of the context-specific dynamic profile is fused with the cross-partition shared intensity distribution of context-independent interests. This allows the information flow recommendation to automatically switch to the corresponding dynamic profile partition under different usage scenarios, realizing context-adaptive information flow recommendation driven by multimodal behavior time series.

[0059] See Figures 2-8 The following application examples are proposed: A short video feed platform, targeting a large number of daily active users, needs to retrieve dynamic profiles of users in different usage scenarios in real time and generate adaptive recommendation lists. Taking user A (ID: user4821)'s behavior record on a certain workday as an example, this user generated multimodal interaction events in four typical usage scenarios that day, covering four behavioral modalities: clicking, sharing, commenting, and saving. The theme space is preset to five thematic dimensions: technology, food, sports, finance, and entertainment.

[0060] Step 1: Context segmentation and event density statistics.

[0061] Extract the timestamp, device type, and interaction duration for each event from User A's event stream. Perform one-hot encoding on the device type (phone, tablet), and normalize the timestamp and interaction duration separately, concatenating them into a temporal context feature vector. Perform density-based spatial clustering on all feature vectors to obtain four temporal context clusters: morning commute (Context A), lunch break (Context B), evening in-depth reading (Context C), and weekend fragmented browsing (Context D). After segmenting the behavior sequences according to context cluster labels, calculate the event density of each modality within each context.

[0062] Table 1. Event density statistics for each modality within each temporal context cluster.

[0063] As shown in Table 1, the click density in scenarios A and D is much higher than that of the other modalities, while the share, comment, and favorite modalities all exhibit sparse characteristics in these two scenarios. Scenario C has a relatively low overall event density, but the distribution of each modality is relatively balanced. This difference provides an input basis for the selection of the adaptive spectrum method in step 2.

[0064] Step 2: Context-adaptive spectrum analysis.

[0065] The behavioral subsequences of each modality within each scenario are clustered and projected along five thematic dimensions to obtain the intensity signal of each theme in each modality. Then, the event density obtained in Table 1 is compared with a preset sparsity threshold (0.20 times / minute) within the scenario. For modalities with densities below the threshold, the Lomb-Scargle periodogram method is used; for modalities with densities above the threshold, short-time Fourier transform is used. After completing the spectral estimation, the formula is used... Calculate the in-context spectral estimation confidence score for each modality within each context, where... Use the default value.

[0066] Table 2. Method selection and confidence scores for each modality spectrum in scenarios A and C.

[0067] In scenario A, the click modality has a confidence score of 0.91, while the confidence scores for the share, comment, and favorite modalities are all below the preset high confidence threshold (0.50). Therefore, these modalities will be marked as low-confidence modalities in step 3 and will undergo cross-modal correction. In scenario C, the modal densities are low, but the confidence scores for each modality are relatively close, resulting in low overall spectral reliability. This will trigger cross-scenario prior contraction correction.

[0068] Step 3: Generation of interest confidence vector and maintenance of partition memory.

[0069] The spectral power distribution vectors of each modality within each scenario are input into the interest phase classification network to obtain the phase classification probability distribution of each topic. Taking scenario A as an example, the click modality has a confidence score of 0.91 and is classified as a high-confidence modality. The share, comment, and favorite modalities are then subjected to Bayesian correction in sequence. The correction formula is as follows: The confidence-weighted phase probability distribution of the high-confidence mode is used as a priori to correct the classification bias of the low-confidence mode. After correction, the phase probability distributions of each mode are aggregated by mean to generate the interest confidence vector for scenario A.

[0070] The table below shows the changes in the phase probability of the low-confidence modality (comment) before and after the correction of the technology topic in Scenario A, to illustrate the effect of cross-modal propagation.

[0071] Table 3 Comparison of Modal Phase Probabilities Before and After Correction for Scenario A Science and Technology Theme Commentary

[0072] After the correction, the probability of the comment modality rising in the technology topic increased from 0.41 to 0.68, which is consistent with the high confidence judgment of the click modality, and the correction effect is as expected.

[0073] After completing cross-modal correction and mean aggregation, the overall spectral reliability index of each time series scenario cluster is calculated. ,in The preset values ​​are retrieved sequentially based on clicks, shares, comments, and favorites. Context C's overall spectral reliability index is below the partition update threshold, so it is marked as a sparse context. The cosine similarity between C and the interest confidence vectors of contexts A, B, and D is calculated. Context B is found to be the high-reliability context with the highest cosine similarity. Using the interest confidence vector of context B as a priori, a shrinkage correction is performed. .

[0074] Table 4 Overall Spectral Reliability Indicators and Sparse Context Identification for Each Temporal Context Cluster

[0075] After the contraction correction in scenario C, its interest confidence vector receives a directional correction in the financial and technological dimensions, similar to that in scenario B. The magnitude of the correction is influenced by the reliability difference. The constraints will not excessively pull the distribution towards scenario B.

[0076] Then execute the context binding tag, according to the formula. Calculate the context binding coefficients for the five thematic dimensions.

[0077] Table 5. Contextual Binding Coefficients for Each Theme Dimension

[0078] The sports and entertainment themes have high context binding coefficients, indicating that user interest in these two themes varies significantly across different contexts, and their cross-regional propagation should be restricted. The technology and finance themes have low context binding coefficients, and their intensity should be updated synchronously across all regions. After each region's memory unit is updated, the latest memory state of the context-specific dynamic profile is output.

[0079] Step 4: Context-adaptive recommendation ranking.

[0080] User A initiates a recommendation request in the evening of a weekday. The temporal context feature vector of this request (including mobile device, evening time period, and normalized interaction duration) is extracted and its distance is calculated with the centers of four temporal context clusters, matching it to context C (evening deep reading). The context-specific interest intensity distribution is read from the corresponding partition memory unit of context C, and the cross-partition shared intensity of context-independent interests (technology, finance) is also read, and then calculated according to the formula... The dynamic profile interest weight distribution in the current context is generated by fusion.

[0081] Table 6 Interest Weights After Integration of Each Topic in Scenario C

[0082] After fusion, the sports dimension has a weight of 0.55, primarily driven by context-specific intensity, preserving users' unique preferences for sports content in the context of in-depth evening reading. Technology and finance are mainly contributed by cross-section sharing intensity, with fusion weights of 0.62 and 0.59 respectively, reflecting the user's consistently stable interests across multiple contexts. The fused interest weight distribution is then weighted and matched with the feature vectors of candidate information stream content to generate a recommendation ranking list, prioritizing in-depth technology analysis, evening financial news, and sports event recaps.

[0083] Throughout the implementation process, data originates from user A's original multimodal event stream. Step 1 involves density clustering to obtain four temporal context clusters and the event density for each modality. Step 2 adaptively selects a spectral method based on the intra-context density and outputs a confidence score. Step 3 uses the confidence score as a link to drive cross-modal Bayesian correction, cross-context shrinkage correction, and context binding labeling. Finally, in step 4, context-specific interests and context-independent interests are fused using the context binding coefficient as a hybrid weight to output an adaptive recommendation list. The core intermediate quantities (event density, confidence score, reliability index, and context binding coefficient) are progressively passed through the data stream, with each subsequent step using the calculation results of the preceding step as input, ensuring the logical coherence of the entire data stream.

[0084] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling, characterized in that, include: Obtain the temporal context feature vector of the user's multimodal behavior event stream, perform density clustering to generate temporal context cluster labels, segment the multimodal behavior sequence into context-specific behavior subsequences according to the temporal context cluster labels, and count the event density of each modality within each temporal context cluster; For each modal behavior subsequence within each time series context cluster, the spectral analysis method is adaptively selected based on the event density of that modality within the time series context cluster to generate a spectral power distribution vector; The spectral power distribution vector is input into the interest phase classification network, which outputs the phase classification probability distribution. The phase classification probability distributions of all modes within the same temporal context cluster are aggregated to generate the interest confidence vector of each temporal context cluster, which is then written into the corresponding partition memory unit. Extract the temporal context feature vector of the current recommendation request, match it to the nearest temporal context cluster, read the interest intensity distribution from the corresponding partition memory unit, perform weighted matching calculation with the candidate information stream content feature vector, and output the recommendation ranking list.

2. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 1, characterized in that, The process of generating the temporal context feature vector includes: The timestamp, device type, and interaction duration features of each event are obtained; the device type is encoded and converted into a numerical representation; the timestamp value and the interaction duration value are normalized respectively; and the processed features are concatenated into the temporal context feature vector.

3. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 1, characterized in that, The adaptive selection spectrum analysis method includes: When the event density of a certain mode within a certain time-series context cluster is lower than the sparsity threshold within the context, the Lomb-Scargle periodogram method is used to estimate the spectral power of irregular timestamp signals; when the event density is higher than the sparsity threshold within the context, short-time Fourier transform is used to extract spectral features.

4. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 1, characterized in that, It also includes calculating the confidence score of the spectral estimation within the context. : ； in, For modality In time-series context clusters The number of valid events within, For modality In time-series context clusters The dimensionless value of the normalized variance of the spectral estimation within the range. To adjust the parameters.

5. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 4, characterized in that, The process of aggregating and generating interest confidence vectors for each temporal context cluster introduces cross-modal phase position confidence propagation, including: The phase classification probability distribution is multiplied by the confidence score of the in-context spectrum estimation to generate a confidence-weighted phase probability distribution; high-confidence mode and low-confidence mode are determined based on the confidence score of the in-context spectrum estimation; the confidence-weighted phase probability distribution of the high-confidence mode is used as a priori to perform Bayesian correction on the phase probability distribution of the low-confidence mode; the mean of the phase probability distributions of each mode is aggregated to generate the propagation-corrected interest confidence vector.

6. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 4, characterized in that, It also includes cross-contextual prior contraction correction: The weighted average of the confidence scores of the intra-context spectral estimation for all modalities within each time-series context cluster is calculated as the overall spectral reliability index. When the overall spectral reliability index of a certain time-series scenario cluster is lower than the partition update threshold, it is marked as a sparse scenario; the high-reliability time-series scenario cluster with the highest cosine similarity to the interest confidence vector of the sparse scenario is retrieved; using the interest confidence vector of the high-reliability time-series scenario cluster as a cross-scenario prior, the interest confidence vector of the sparse scenario is shrunk and corrected. ； in, This represents the original interest confidence vector for sparse scenarios. This is the interest confidence vector for a high-reliability time-series scenario cluster. The shrinkage coefficient, The overall spectral reliability index for high-reliability time-series scenario clusters. This is an overall spectral reliability index for sparse scenarios.

7. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 6, characterized in that, It also includes the context binding tagging process: Calculate the dimension-wise similarity of interest confidence vectors among all temporal context clusters; mark interest dimensions with high cross-context similarity as context-independent interests and perform intensity synchronization updates across memory units in each partition; mark interest dimensions with high confidence only appearing in specific temporal context clusters as context-bound interests and restrict their cross-partition propagation; for each interest dimension... Generate context binding coefficient : ； in, Theme The standard deviation of the interest confidence component across all time-series context clusters. The corresponding mean, To prevent smoothing constants with a denominator of zero.

8. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 7, characterized in that, The process of reading the interest intensity distribution from the corresponding partition memory unit also includes: The interest intensity distribution of context-specific dynamic profiles and the cross-regional shared intensity distribution of context-independent interests are read, and the context binding coefficient is used. As a blending weight: ； in, For the interest dimension The specific strength of the current context partition, For the interest dimension Cross-partition sharing strength, The interest weights are fused; the fused dynamic profile interest weight distribution is weighted and matched with the candidate information stream content feature vectors to calculate the recommended ranking list.

9. The adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profiling according to claim 5, characterized in that, theme In low confidence mode Down-corrected phase probability distribution Obtain it using the following formula: ； in, Theme In modality Lower phase The original classification probability, Theme In high confidence mode Lower phase The classification probability, For modality In time-series context clusters The confidence score of the intra-context spectrum estimation is calculated, with the denominator being a normalization term. When there are multiple high-confidence modes within the time-series context cluster, the correction is performed on the low-confidence modes in descending order of their intra-context spectrum estimation confidence scores.

10. An adaptive information flow recommendation system based on multimodal behavioral temporal sequence and dynamic profile, used to execute the adaptive information flow recommendation method based on multimodal behavioral temporal sequence and dynamic profile as described in any one of claims 1 to 9, characterized in that, include: The context segmentation module is used to obtain the temporal context feature vector of the user's multimodal behavior event stream, perform density clustering to generate temporal context cluster labels, segment the multimodal behavior sequence into context-specific behavior subsequences according to the temporal context cluster labels, and count the event density of each modality within each temporal context cluster. The spectrum analysis module is used to adaptively select the spectrum analysis method for each modal behavior subsequence within each time series context cluster based on the event density, generate a spectrum power distribution vector, and calculate the confidence score of spectrum estimation within the context. The interest confidence generation and memory maintenance module is used to input the spectral power distribution vector into the interest phase classification network to generate a phase classification probability distribution, aggregate and generate interest confidence vectors for each temporal context cluster, perform cross-modal phase position confidence propagation and cross-context prior contraction correction, and maintain context-specific dynamic profiles in each partition memory unit. The recommendation ranking module is used to extract the temporal context feature vector of the current recommendation request and match it to the nearest temporal context cluster. It reads the interest intensity distribution from the corresponding partition memory unit, performs weighted matching calculation with the candidate information stream content feature vector, and outputs the recommendation ranking list.

Images