A user intent prediction method fusing a time sequence model and ensemble learning

Abstract

This invention discloses a user intent prediction method integrating temporal series models and ensemble learning. The method includes the following steps: acquiring multi-source behavioral data of the target user within historical periods and the current session period. The multi-source behavioral data includes search data, browsing data, click data, dwell data, favorites data, add-to-cart data, inquiry data, return data, exit data, conversion data, page context data, entry source data, and trigger time data. This invention relates to the field of data intelligence analysis technology. This user intent prediction method integrating temporal series models and ensemble learning calculates the category deviation, intensity deviation, path deviation, context deviation, and conversion tendency deviation of the current session behavior relative to a stable intent anchor point, forming a behavioral residual signal. This enables the model to specifically model the difference between the current behavior and historical preferences, improving the ability to identify intent drift.

Description

Technical Field

[0001] This invention relates to the field of data intelligent analysis, and more specifically, to a method for predicting user intent by integrating time series models and ensemble learning. Background Technology

[0002] With the development of e-commerce, content recommendation, intelligent customer service, and precision marketing systems, user intent prediction has been widely applied in scenarios such as product recommendation, content distribution, user profile updates, and business conversion assessment. Existing technologies typically construct features based on user search, browsing, clicks, dwell time, favorites, add-to-cart, inquiries, and conversion records, and output the user's current possible intent through classification models, sequence models, or ensemble learning models. These methods are effective in scenarios where user behavior paths are stable and intent is clearly expressed. However, in real-world sessions, users often exhibit short dwell times, quick returns, repeated clicks, temporary jumps, or cross-category browsing behaviors, which do not necessarily represent true intent. Existing solutions often directly concatenate historical preferences, current session behavior, and contextual features into the model, lacking a deviation identification mechanism based on the user's normal intent, and also lacking a process for counterfactual verification and weight adjustment of suspicious behavior nodes. This makes it difficult for the model to distinguish between key intent behaviors and interfering behaviors, thus affecting the accuracy, stability, and interpretability of user intent prediction. Summary of the Invention

[0003] The purpose of this invention is to provide a user intent prediction method that integrates temporal modeling and ensemble learning. This method addresses the problem that existing solutions often directly concatenate historical preferences, current session behavior, and contextual features into the model, lacking a deviation identification mechanism based on the user's normal intent, and also lacking a process for counterfactual verification and weight correction of suspicious behavior nodes. This makes it difficult for the model to distinguish between key intent behaviors and interfering behaviors, thereby affecting the accuracy, stability, and interpretability of user intent prediction.

[0004] This invention achieves the above objective through the following technical solution: a user intent prediction method that integrates temporal models and ensemble learning, the method comprising the following steps:

[0005] Acquire multi-source behavioral data of target users within historical periods and the current session period. The multi-source behavioral data includes search data, browsing data, click data, dwell data, favorites data, add-to-cart data, inquiry data, return data, exit data, conversion data, page context data, entry source data, and trigger time data.

[0006] The multi-source behavior data is sorted according to the trigger time, and behavior type encoding, behavior object encoding, context encoding, entry source encoding, and time interval encoding are performed respectively to construct the user's historical behavior sequence and current session behavior sequence;

[0007] Based on the user's historical behavior sequence, long-term preference features that recur across historical time windows and are associated with conversion behavior are extracted to generate stable intent anchors that represent the user's normal intent.

[0008] The current session behavior sequence is matched with the stable intent anchor to obtain a behavior residual signal that includes category deviation, intensity deviation, path deviation, context deviation, and conversion tendency deviation.

[0009] The current session behavior sequence, the stable intent anchor point, and the behavior residual signal are input into the time series model, and the intent stability, intent drift, behavior conflict, and noise intensity are output.

[0010] Based on the above output, the fusion weights of multiple ensemble learning sub-models are dynamically adjusted, and counterfactual behavior perturbation verification is performed on suspicious behavior nodes in the current session behavior sequence. The intention contribution of the behavior node is determined based on the difference in intention prediction results before and after the perturbation.

[0011] Based on the intent contribution, the feature weights of interfering behavior nodes are reduced and the feature weights of key intent behavior nodes are strengthened, and the user's current intent category, intent confidence, intent migration direction, key influencing behavior nodes, and interfering behavior nodes are output.

[0012] Furthermore, constructing the user's historical behavior sequence and current session behavior sequence includes:

[0013] The historical periodic behavior data under the same user ID is divided into multiple historical sub-sequences according to a preset historical time window.

[0014] Write the trigger time difference, behavior object category, page level, and conversion action status of two adjacent behaviors within the current session cycle into the current session behavior sequence;

[0015] For behavior data with missing trigger times, fill in the missing trigger times according to the server receiving time. For behavior data that is repeatedly reported and has the same behavior object, behavior type and trigger time, retain one record.

[0016] Further, generating the stable intent anchor point includes:

[0017] Calculate the access frequency, cumulative dwell time, number of favorites, number of add-to-carts, number of inquiries, and number of conversions for each interest category within each historical subsequence;

[0018] Interest categories that appear consecutively in at least two historical subsequences and whose transformation correlation strength reaches a preset anchor threshold are identified as anchor categories.

[0019] The common entry sources, page paths, and conversion actions corresponding to the anchor point categories are combined into stable intent anchor points, wherein the preset anchor point threshold is determined by the quantile of the conversion association strength in the training samples.

[0020] Further, the behavioral residual signal is obtained, including:

[0021] The interest categories in the current session behavior sequence are matched with the anchor categories to obtain the category deviation;

[0022] The intensity deviation is obtained by matching the dwell time, number of clicks, browsing depth, favorites, add to cart and inquiry actions in the current session behavior sequence with the intensity of normal behavior in the stable intent anchor point;

[0023] The path deviation is obtained by matching the current page jump path with the normal page path in the stable intent anchor point;

[0024] By matching the page context, entry source, and conversion action separately, we can obtain context deviation and conversion tendency deviation.

[0025] Furthermore, the time-series model performs time-series state identification, including:

[0026] The current session behavior sequence, stable intent anchors, and behavior residual signals are concatenated into temporal state input features;

[0027] Using time-series models, we can identify the number of consecutive occurrences of behavioral residual signals within the current session period, the consistency of deviation direction, and the rate of change of deviation magnitude.

[0028] When the consistency between the current session behavior and the stable intent anchor point reaches a preset stability threshold, the corresponding intent stability is output.

[0029] When the behavior residual signal continuously increases along the same intention category direction, the corresponding intention drift degree is output.

[0030] Output the degree of behavioral conflict based on category conflict, abnormal path jump, and contradictory conversion actions;

[0031] The noise intensity is output based on the proportion of short dwell times, quick returns, repeated clicks, and behaviors with no conversion association.

[0032] Furthermore, the multiple ensemble learning sub-models include a long-term preference sub-model, a current session sub-model, a contextual scenario sub-model, a conversion behavior sub-model, and a behavior residual sub-model;

[0033] The long-term preference sub-model takes a stable intent anchor as input, the current session sub-model takes the current session behavior sequence as input, the context scenario sub-model takes the page context, entry source and trigger time as input, the conversion behavior sub-model takes the collection action, add to cart action, consultation action and conversion action as input, and the behavior residual sub-model takes the behavior residual signal as input. Each sub-model outputs the candidate user intent category and corresponding confidence level.

[0034] Furthermore, the fusion weights of each ensemble learning sub-model are dynamically adjusted, including:

[0035] When the intention stability reaches the preset stability threshold, increase the fusion weight of the long-term preference sub-model and decrease the fusion weight of the behavior residual sub-model.

[0036] When the intent drift reaches a preset drift threshold, increase the fusion weight of the current session sub-model and the behavior residual sub-model;

[0037] When the degree of behavioral conflict reaches a preset conflict threshold, the fusion weight of the sub-model that generates conflict candidate intent categories is reduced.

[0038] When the noise intensity reaches a preset noise threshold, the fusion weight of sub-models dominated by short dwell times, quick returns, repeated clicks, or no-conversion-related behaviors is reduced.

[0039] Furthermore, the counterfactual behavior perturbation verification includes:

[0040] Suspicious behavior nodes are filtered from the current session behavior sequence based on the degree of behavioral conflict and noise intensity;

[0041] For the suspicious behavior nodes, respectively perform masking processing, similar anchor point behavior replacement processing, adjacent order adjustment processing, and feature weight reduction processing to generate corresponding counterfactual behavior sequences;

[0042] The original current session behavior sequence and the counterfactual behavior sequence are respectively input into the time series model and the dynamically fused ensemble learning model. The user's current intent category, intent confidence, and intent migration direction of the two outputs are compared to obtain the intent contribution of the suspicious behavior nodes.

[0043] Furthermore, weaken interfering behavior nodes and strengthen key intent behavior nodes, including:

[0044] When the intention contribution of a behavior node is lower than the preset contribution threshold, and the intention confidence increases or the intention migration direction tends to stabilize after perturbation, the behavior node is identified as an interfering behavior node, and its feature weight is reduced according to the difference between the intention contribution and the preset contribution threshold.

[0045] When the intention contribution of a behavior node reaches the preset contribution threshold, and the intention confidence decreases after perturbation or the intention migration direction deviates from the original prediction direction, the behavior node is identified as a key intention behavior node, and its feature weight is increased according to the difference between the intention contribution and the preset contribution threshold.

[0046] Furthermore, the output of the end-user intent prediction results includes:

[0047] The current session behavior sequence, stable intent anchor point, and behavior residual signal after feature weight adjustment are input into the dynamically fused ensemble learning model, and the confidence of each candidate user intent category is weighted and summarized according to the adjusted fusion weights.

[0048] The candidate user intent category with the highest confidence after weighted aggregation is determined as the user's current intent category;

[0049] The direction of intent migration is determined based on the change in the category of the candidate user intent relative to the stable intent anchor point, and key influencing behavior nodes that contribute positively to the prediction result and interference behavior nodes whose weights are reduced are output simultaneously.

[0050] The beneficial effects of this invention are as follows:

[0051] 1. By constructing stable intent anchors, users' long-term preferences and normal intents can be extracted from historical behavior sequences, so that current conversation behavior is no longer analyzed in isolation, which helps to distinguish between changes in users' true intent and occasional behavioral fluctuations.

[0052] 2. By calculating the category deviation, intensity deviation, path deviation, context deviation, and conversion tendency deviation of the current session behavior relative to the stable intent anchor point, a behavioral residual signal is formed, which enables the model to specifically model the difference between the current behavior and historical preferences, thereby improving the ability to identify intent drift.

[0053] 3. The temporal model outputs intent stability, intent drift, behavioral conflict, and noise intensity, and dynamically adjusts the fusion weights of the long-term preference sub-model, the current session sub-model, the context scenario sub-model, the conversion behavior sub-model, and the behavioral residual sub-model accordingly, avoiding the problem that fixed-weight ensemble learning models are difficult to adapt to different session states.

[0054] 4. By performing masking, replacement of similar anchor point behaviors, adjustment of adjacent order, or reduction of feature weight on suspicious behavior nodes, counterfactual behavior sequences are generated. By comparing the differences in intent prediction results before and after the perturbation, the contribution of each behavior node to the user intent prediction results can be quantified, thereby improving the interpretability of the prediction results.

[0055] 5. Reduce the feature weights of interfering behavior nodes based on their contribution to intent, and strengthen the feature weights of key intent behavior nodes. This will enable the final prediction results to reduce the impact of noisy behaviors and highlight the impact of true intent behaviors, thereby improving the accuracy, stability and availability of user intent prediction. Attached Figure Description

[0056] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0057] Figure 1 This is a flowchart illustrating the overall method architecture of the present invention;

[0058] Figure 2 This is a flowchart illustrating the process of generating stable intent anchor points for this invention.

[0059] Figure 3 This is a flowchart of the integrated learning dynamic fusion process of the present invention. Detailed Implementation

[0060] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0061] Example 1:

[0062] Please see Figure 1-3 This invention provides a user intent prediction method that integrates temporal models and ensemble learning, applicable to e-commerce platforms, content recommendation platforms, intelligent customer service systems, advertising systems, and user profile update systems. Based on user historical behavior sequences and current session behavior sequences, this method first generates stable intent anchors to represent the user's normal intent. Then, it calculates the behavioral residual signal of the current session behavior relative to the stable intent anchors and identifies intent stability, intent drift, behavioral conflict, and noise intensity through a temporal model. On this basis, multiple ensemble learning sub-models are dynamically fused, and counterfactual behavioral perturbation verification is combined to determine the intent contribution of behavioral nodes, thereby weakening interfering behavioral nodes and strengthening key intent behavioral nodes. Finally, it outputs the user's current intent category, intent confidence, intent migration direction, key influencing behavioral nodes, and interfering behavioral nodes.

[0063] In this implementation, to avoid dimensional differences in data such as dwell time, click count, browsing depth, number of favorites, number of add-to-carts, number of inquiries, number of returns, number of exits, conversion count, page context data, and entry source data, all features involved in stable intent anchor point calculation, behavioral residual signal calculation, temporal state recognition, dynamic fusion, and counterfactual perturbation verification are first converted into intervals. Dimensionless eigenvalues. For the target user At any moment The first primitive behavioral characteristics The normalization process is as follows:

[0064] (1)

[0065] in, Represents the normalized i-th Class behavioral feature values; and These respectively represent the training samples or rolling statistical samples in the same business scenario. Minimum and maximum values ​​of class features; To prevent the correction amount from being zero in the denominator, its value is [value missing]. One part per million. Since the numerator and denominator have the same dimensions, therefore The values ​​are dimensionless. For page context, entry source, behavior object category and interest category, one-hot encoding or category embedding encoding is first used, and then converted into dimensionless vectors according to equation (1).

[0066] Acquire multi-source behavioral data generated by target users within historical periods and the current session period. Multi-source behavioral data includes search data, browsing data, click data, dwell time data, favorites data, add-to-cart data, inquiry data, return data, exit data, conversion data, page context data, entry source data, and trigger time data.

[0067] The historical period is used to characterize the user's long-term behavioral state prior to the current session. In this implementation, the historical period is set to thirty days prior to the start of the current session; when the business system exhibits significant seasonality or activity cycles, the historical period is set to ninety days prior to the start of the current session. The current session period is used to characterize the continuous interaction process formed from the user's entry into the business system to their exit from the business system; when the user does not actively exit, if the time interval between two adjacent actions exceeds thirty minutes, the continuous interaction process prior to the previous action is defined as a current session period.

[0068] After acquiring multi-source behavioral data, the data is sorted according to trigger time and then encoded in four ways: behavior type, behavior object, context, entry point source, and time interval. Behavior type encoding indicates whether the user performed an action such as searching, browsing, clicking, staying, adding to favorites, adding to cart, inquiring, returning, exiting, or converting. Behavior object encoding indicates the product, content, service, question, or page object that the user's behavior refers to. Context encoding indicates the page, channel, section, recommendation slot, or interaction entry point the user is on. Entry point source encoding indicates whether the user originated from organic search, in-site recommendations, external redirects, activity pages, or customer service entry points. Time interval encoding indicates the time difference between adjacent behaviors and the position of the behavior within the session cycle.

[0069] For historical periodic behavior data under the same user identifier, it is segmented according to a preset historical time window to generate multiple historical sub-sequences. In this embodiment, the preset historical time window is seven days; when the historical period is thirty days, four complete historical sub-sequences and one remaining historical sub-sequence are generated. If the data volume of the remaining historical sub-sequence is less than 30% of the data volume of the complete historical sub-sequence, it will not participate in the calculation of stable intent anchors. For behavior data within the current session period, the trigger time difference, behavior object category, page level, and conversion action status of two adjacent behaviors are written into the current session behavior sequence. For behavior data with missing trigger times, it is supplemented according to the server receiving time; for behavior data that is repeatedly reported and has the same behavior object, behavior type, and trigger time, only one record is retained. Thus, the user's historical behavior sequence and the current session behavior sequence are obtained.

[0070] Based on the user's historical behavior sequence, long-term preference features that recur across historical time windows and are associated with conversion behavior are extracted, and stable intent anchors that represent the user's normal intent are generated.

[0071] Specifically, the user's historical behavior sequence is divided into multiple historical sub-sequences according to a preset historical time window. The frequency of visits, cumulative dwell time, number of favorites, number of add-to-carts, number of inquiries, and number of conversions are calculated for different interest categories within each historical sub-sequence. Interest Categories Stable anchor point score Determine as follows:

[0072] (2)

[0073] in, Indicates interest categories The stable anchor point score is a dimensionless value; Indicates interest categories Normalized access frequency across multiple historical time windows; Indicates interest categories Normalized cumulative dwell time; Indicates interest categories The corresponding normalized behavioral intensity is determined by the number of times the user favorites, adds to cart, and inquires. Indicates interest categories The corresponding normalized transformation correlation strength; , , and These represent the weights for visit frequency, cumulative dwell time, behavioral intensity, and conversion correlation strength, respectively, satisfying the following conditions: .

[0074] In this embodiment, , , and The initial values ​​are set as follows: , , and This ensures that the conversion correlation strength directly related to real business conversions has a high influence. The rules for determining the above weights are as follows: when there are insufficient labeled samples in the early stages of the business system's launch, the above initial values ​​are used; when the number of training samples with real conversion labels reaches 10,000 user sessions, the weighted result of user intent category prediction accuracy and conversion prediction accuracy is used as the optimization objective, and the above weights are fine-tuned on a validation set. During tuning, the search range for each weight is... to Step size is And keep the sum of the four weights at 1 .

[0075] The threshold for determining the anchor point category is determined based on the sample distribution of stable anchor point scores, as follows:

[0076] (3)

[0077] in, This represents the threshold for determining the anchor point category; it is a dimensionless value. The score of the stable anchor point in the training samples or rolling statistics samples represents the score of the first stable anchor point. percentile; This is the minimum stability limit value. If a certain interest category... Stable anchor point score Not less than If an interest category appears consecutively in at least two historical time windows, then that interest category is identified as an anchor category.

[0078] After determining the anchor point category, the commonly used entry points, page paths, and conversion actions corresponding to that anchor point category are combined to form stable intent anchors. Stable intent anchors are used to characterize a user's normal interest categories, normal entry points, normal page paths, and normal conversion tendencies under non-temporary disturbances. Through stable intent anchors, a reference benchmark can be provided for the current session behavior, enabling the system to determine whether the current session behavior continues the user's long-term preferences or generates a new intent shift relative to long-term preferences.

[0079] The current session behavior sequence is matched with stable intent anchors to obtain behavioral residual signals that include category deviation, intensity deviation, path deviation, context deviation, and conversion tendency deviation.

[0080] Specifically, the interest categories in the current session behavior sequence are matched with the anchor categories in the stable intent anchors to obtain category deviation; the dwell time, number of clicks, browsing depth, favorite actions, add-to-cart actions, and inquiry actions in the current session behavior sequence are matched with the intensity of normal behaviors in the stable intent anchors to obtain intensity deviation; the current page jump path is matched with the normal page path in the stable intent anchors to obtain path deviation; and the page context, entry source, and conversion action are matched with the normal context, normal entry source, and normal conversion tendency in the stable intent anchors to obtain context deviation and conversion tendency deviation.

[0081] Current session time behavioral residuals Calculate as follows:

[0082] (4)

[0083] in, Indicates time The behavioral residual value is a dimensionless value. Indicates category deviation; Indicates a deviation in intensity; Indicates path deviation; Indicates a deviation from the context; This indicates a deviation in the tendency to transform; , , , and The residual weights for the above five types of deviation characteristics are respectively, satisfying... .

[0084] In this embodiment, the residual weights are initially set as follows: , , , , Among them, category deviation has a strong indicative effect on changes in user intent, and therefore is given a higher weight; context deviation is greatly affected by the entry source and page activity, and therefore is given a relatively lower weight. The above weights are determined as follows: when training samples with true intent labels exist, they are determined through a validation set grid search, with the search range being [missing information]. to Step size is The optimization goal is to use the accuracy of the prediction of the user's current intent category. When there is a lack of real intent tags, the above initial values ​​are used and periodically updated after subsequent business conversion feedback is generated.

[0085] The behavioral residual signal obtained by equation (4) is used to characterize the direction, magnitude and duration of the current session behavior relative to the user's normal intention.

[0086] The current session behavior sequence, stable intent anchor point, and behavior residual signal are concatenated into temporal state input features and input into the temporal model. The temporal model outputs intent stability, intent drift, behavior conflict, and noise intensity.

[0087] In this embodiment, the temporal model employs a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), a temporal convolutional network, or a Transformer temporal coding model. The temporal model is used to identify the number of consecutive occurrences of the behavioral residual signal within the current session period, the consistency of the deviation direction, and the rate of change of the deviation magnitude.

[0088] Let the length of the current session's sliding observation window be... The length of the sliding observation window represents the number of the most recent behavior nodes involved in continuous state recognition. Determined by the median length of the effective behavioral path before conversion in the training samples; when the median is less than Time to take When the median is greater than Time to take Based on recent Individual behavioral residuals, intentional stability and intentional drift Determine as follows:

[0089] (5)

[0090] in, This represents the intended stability and is a dimensionless numerical value. This represents the intentional drift degree, and is a dimensionless numerical value. Indicates the first The behavioral residual value of each behavioral node; Indicates the first The deviation direction corresponding to each behavior node; This indicates the direction of deviation that appears most frequently within the current sliding observation window; For indicator functions, when and Take when consistent Otherwise take .

[0091] According to Equation (5), if the residual of the most recent behavior node relative to the stable intention anchor point is generally low, the intention stability is high; if the most recent behavior node continues to change along the same deviation direction, the intention drift is high.

[0092] Behavioral conflict and noise intensity Calculate as follows:

[0093] (6)

[0094] in, This represents the degree of behavioral conflict and is a dimensionless numerical value. This represents noise intensity and is a dimensionless numerical value. This indicates the number of nodes with conflicting categories within the sliding observation window; Indicates the number of nodes exhibiting abnormal path redirection behavior; Indicates the number of contradictory behavior nodes in the transformation action; This indicates the number of nodes exhibiting short-stay behavior. This indicates the number of nodes that quickly return the behavior. Indicates the number of nodes that are repeated clicks; This indicates the number of nodes with no conversion-related behavior. The counting unit above is "number", and the denominator is... Similarly, the number of behavior nodes, therefore and All values ​​are dimensionless.

[0095] Among them, the threshold for determining short-stay behavior nodes is determined by the dwell time of the node under the same page type. Percentiles are determined; the threshold for quickly returning behavior nodes is determined by the first percentile of the time interval between returned behaviors in the same business scenario. Percentiles are determined; repeated click behavior nodes are determined based on the number of times the same behavior object is clicked repeatedly within a preset short time window, and the preset short time window is [value missing]. Instant The time interval is determined by the concentrated distribution range of repeated click behavior in the platform logs; non-conversion related behavior nodes are determined based on the path association between the behavior node and subsequent collection, add-to-cart, inquiry or conversion actions.

[0096] By recognizing the temporal state as described above, the subsequent ensemble learning model can no longer use fixed fusion weights, but instead adaptively determine the degree of influence of each sub-model based on the user's current session state.

[0097] Construct multiple ensemble learning sub-models. These sub-models include a long-term preference sub-model, a current session sub-model, a contextual scenario sub-model, a conversion behavior sub-model, and a behavior residual sub-model.

[0098] The long-term preference sub-model takes stable intent anchors as input and outputs candidate user intent categories and corresponding confidence levels based on long-term preference judgments. The current session sub-model takes the current session behavior sequence as input and outputs candidate user intent categories and corresponding confidence levels based on continuous behavior judgments within the current session. The context scenario sub-model takes page context, entry source, and trigger time as input and outputs candidate user intent categories and corresponding confidence levels based on the access scenario judgment. The conversion behavior sub-model takes favorites, add-to-cart, inquiry, and conversion actions as input and outputs candidate user intent categories and corresponding confidence levels based on conversion tendency judgments. The behavior residual sub-model takes behavior residual signals as input and outputs candidate user intent categories and corresponding confidence levels based on the deviation of the current session's intent from the normal state.

[0099] In this implementation, each ensemble learning sub-model outputs candidate user intent categories and corresponding confidence scores within the same intent category set. This same intent category set is pre-defined by the business system and remains consistent throughout the model training phase. By constructing these multiple ensemble learning sub-models, user intent can be predicted from five perspectives: long-term preferences, current session, contextual scenario, conversion behavior, and behavioral deviation. This avoids prediction bias caused by a single model over-reliance on a particular feature.

[0100] The fusion weights of each ensemble learning sub-model are dynamically adjusted based on intent stability, intent drift, behavioral conflict, and noise intensity.

[0101] Let the multiple ensemble learning sub-models be the long-term preference sub-model, the current session sub-model, the context sub-model, the conversion behavior sub-model, and the behavior residual sub-model, and their numbers be ______. Dynamic fusion weights of each sub-model The final intent category confidence score is determined as follows, and based on this weight, it is output:

[0102]

[0103] (7)

[0104] in, Indicates the first The dynamic fusion weights of the five ensemble learning sub-models are dimensionless values, and the sum of the dynamic fusion weights of the five sub-models is... ; Indicates the first The basic weight bias of each sub-model; , , and These represent the effects of intention stability, intention drift, behavioral conflict, and noise intensity on the first... The adjustment coefficient of each sub-model; Indicates the first Sub-models for user intent categories The confidence level of the output; Indicates the user intent category after dynamic fusion The final confidence level.

[0105] In this embodiment, The basic weight bias is determined based on the single-model accuracy of each sub-model on the validation set. The higher the accuracy, the higher the basic weight bias. , , and Determined through training on a validation set. Specifically, when the long-term preference sub-model corresponds to... When the value is positive, the higher the intention stability, the higher the weight of the sub-model; when the value is positive, the weight of the sub-model corresponding to the behavioral residual is higher. When the value is positive, the higher the intention drift, the higher the weight of the sub-model; when a sub-model is susceptible to conflict behavior and noise behavior, its corresponding weight... and Setting it to a positive value will decrease the weight of the sub-model when the degree of behavioral conflict and the noise intensity increase.

[0106] When intent stability reaches a preset stability threshold, it indicates that the current session behavior is largely consistent with the user's long-term preferences. In this case, the fusion weight of the long-term preference sub-model is increased, while the fusion weight of the behavior residual sub-model is decreased. When intent drift reaches a preset drift threshold, it indicates that the current session behavior has a continuous deviation from the stable intent anchor point. In this case, the fusion weight of the current session sub-model and the behavior residual sub-model is increased. When behavior conflict reaches a preset conflict threshold, it indicates that there are category conflicts, abnormal path jumps, or contradictory conversion actions between different session behaviors. In this case, the fusion weight of the sub-model that generates conflicting candidate intent categories is decreased. When noise intensity reaches a preset noise threshold, it indicates that there are many short pauses, quick returns, repeated clicks, or behaviors without conversion relevance in the current session. In this case, the fusion weight of the sub-model dominated by the above behaviors is decreased.

[0107] Among them, the stability threshold, drift threshold, conflict threshold, and noise threshold are all determined based on the distribution of training samples or rolling statistical samples. The stability threshold is taken as the th value in the distribution of the intended stability samples. Percentile; the drift threshold is taken as the th percentile of the intended drift sample distribution. Percentile; the conflict threshold is taken as the th percentile of the conflict degree sample distribution. Percentile; the noise threshold is taken as the th percentile of the noise intensity sample distribution. Percentiles. All the above thresholds are dimensionless values, and are recalculated based on the most recent period's sample after each statistical period to adapt to changes in user behavior patterns under different business scenarios.

[0108] After adjusting the fusion weights, the candidate user intent categories and corresponding confidence scores output by each ensemble learning sub-model are weighted and summarized according to the adjusted fusion weights to form the dynamically fused ensemble learning model output. This processing method can automatically change the model's focus according to different session states, improving the adaptability and stability of intent prediction results.

[0109] Suspicious behavioral nodes are filtered from the current session behavior sequence based on behavioral conflict level and noise intensity. Suspicious behavioral nodes include those that increase behavioral conflict level, those that increase noise intensity, and those that deviate significantly from the stable intent anchor point.

[0110] Suspicious behavior nodes are processed by masking, replacing similar anchor points, adjusting adjacent order, and reducing feature weights to generate corresponding counterfactual behavior sequences. Masking removes the feature corresponding to the behavior node from the model input; replacing similar anchor points with a reference behavior node consistent with the stable intent anchor point; adjusting adjacent order adjusts the order between the behavior node and its neighbors; and reducing feature weights retains the behavior node but reduces its influence on the input features.

[0111] The original current conversation behavior sequence is input into the temporal model and the dynamically fused ensemble learning model to obtain the first intent prediction result; the counterfactual behavior sequence is input into the temporal model and the dynamically fused ensemble learning model to obtain the second intent prediction result. For the... The intention contribution of each suspicious behavior node. Calculate as follows:

[0112] (8)

[0113] in, Indicates the first The intention contribution of each suspicious behavior node is a dimensionless value. This represents the original sequence of current session actions; Indicates the first The behavior sequence obtained after counterfactually perturbing a suspicious behavior node; This indicates the user's current intent category, derived from the original sequence of current session behaviors. This indicates the user's current intent category obtained based on the counterfactual behavior sequence; Indicates the user's current intent category within the original current session behavior sequence. Confidence level; Indicates the user's current intent category under the counterfactual behavior sequence. Confidence level; This represents the category change penalty coefficient, which is a dimensionless value. As an indicator function, it is taken when the intention category changes before and after the counterfactual perturbation. Otherwise take .

[0114] In this embodiment, Pick to And determined through a validation set; when the business system places greater emphasis on sensitivity to changes in intent categories, Take close The numerical value; when the business system places more emphasis on the magnitude of confidence level changes. Take close The value.

[0115] By verifying the counterfactual behavior disturbances described above, we can determine the actual impact of a certain behavior node on the final intention prediction result, thereby providing a basis for weakening subsequent interference behaviors and strengthening key intention behaviors.

[0116] Based on the intent contribution of suspicious behavior nodes, feature weights are adjusted for behavior nodes in the current session behavior sequence.

[0117] For contribution threshold The first step is to use the intention contribution of behavioral nodes in the training samples. Percentiles are used as initial values; if the system places greater emphasis on stability, then the 1st percentile is used. Percentiles; if the system places greater emphasis on key recall actions, then the percentiles will be used. Percentile. This contribution threshold is a dimensionless value.

[0118] When the intention contribution of the behavior node Below the contribution threshold Furthermore, when the confidence of intent increases or the direction of intent migration stabilizes after perturbation, the behavior node is identified as a perturbation behavior node, and its feature weight is reduced according to the difference between the intent contribution and the contribution threshold. This processing can weaken the impact of short dwell times, quick returns, accidental clicks, temporary jumps, or behaviors without conversion association on the prediction results.

[0119] When the intention contribution of the behavior node Not lower than the contribution threshold Furthermore, when the confidence level of intent decreases after perturbation or the direction of intent migration deviates from the original prediction direction, the behavior node is identified as a key intent behavior node, and its feature weight is increased according to the difference between the intent contribution and the contribution threshold. This processing can strengthen search, deep browsing, favorites, add-to-cart, inquiry, or conversion behaviors that have a decisive impact on the user's true intent.

[0120] After the above processing, the influence of interfering behavior nodes in the current session behavior sequence is weakened, while the influence of key intent behavior nodes is strengthened, making the behavioral features of the ensemble learning model after dynamic fusion of inputs closer to the user's true intent expression.

[0121] The current session behavior sequence, stable intent anchors, and behavior residual signals, adjusted by feature weights, are input into the dynamically fused ensemble learning model, and the confidence scores of each candidate user intent category are weighted and summarized according to the adjusted fusion weights.

[0122] The candidate user intent category with the highest confidence after weighted aggregation is determined as the user's current intent category; the intent migration direction is determined based on the category change of this candidate user intent category relative to the stable intent anchor point; at the same time, the key influencing behavior nodes that have a positive contribution to the prediction result, as well as the interfering behavior nodes whose weights have been reduced, are output.

[0123] The final output includes the user's current intent category, intent confidence, intent migration direction, key influencing behavior nodes, and interfering behavior nodes. By simultaneously outputting key influencing behavior nodes and interfering behavior nodes, this invention not only obtains user intent prediction results but also explains the basis for these predictions, improving the interpretability and business usability of the system in recommendation, customer service, marketing, and user profile update scenarios.

[0124] Example 2:

[0125] This embodiment uses an office equipment purchase intent prediction scenario on an e-commerce platform as an example to further illustrate the method of the present invention. In this embodiment, the target user browses printer, scanner, office supplies, and conference equipment products multiple times within a historical period, and engages in favorites, add-to-cart, and inquiry behaviors in some historical sessions. Within the current session, the target user first enters "wireless printer" through the site search entry, then browses multiple printer product pages, briefly clicks on the "gaming headset" product page and quickly returns, then continues browsing printer supplies pages and initiates a customer service inquiry. The system needs to determine whether the user's current true intent is still to purchase office equipment, or whether a new interest migration has occurred.

[0126] The system obtains the target user's historical behavior data for the 30 days prior to the start of the current session and divides it into seven-day historical time windows. Within each historical time window, the system calculates the user's access frequency, cumulative dwell time, number of favorites, number of add-to-carts, number of inquiries, and number of conversions for each interest category. In this embodiment, the normalized access frequency of the user under the "Office Equipment" category is: Normalized cumulative stay duration is Normalized behavioral intensity is The normalized transformation correlation strength is The normalized access frequency under the "Digital Accessories" category is: Normalized cumulative stay duration is Normalized behavioral intensity is The normalized transformation correlation strength is .

[0127] The stable anchor score is calculated according to equation (2), where the access frequency weight is... Pick Cumulative stay time weight Pick Action intensity weight Pick Transformation of correlation strength weights Pick After calculation, the stable anchor score for the "Office Equipment" category is: The stable anchor score for the "Digital Accessories" category is If the score of the stable anchor point in the current rolling statistical sample is the first... percentiles Then, the anchor point category determination threshold is determined according to equation (3). for Because the stable anchor score for the "Office Equipment" category is no less than [a certain value], Since it appears consecutively in at least two historical time windows, "office equipment" is identified as the anchor category for target users, and its commonly used entry sources, page paths, and conversion actions are combined into stable intent anchors.

[0128] The system acquires behavioral data of the target user within the current session period. The current session behavior sequence includes: searching for "wireless printer", clicking on the printer list page, browsing the first printer product page, adding the first printer product to favorites, briefly clicking on the gaming headset product page, quickly returning to the printer list page, browsing the printer consumables page, and initiating a customer service inquiry. The system encodes the above behaviors by behavior type, behavior object, page context, entry source, and time interval, and converts the dwell time, number of clicks, browsing depth, and inquiry action into dimensionless feature values.

[0129] The system matches the current session behavior sequence with stable intent anchors. For behaviors such as searching for "wireless printer," browsing printer product pages, adding printer products to favorites, browsing printer consumables pages, and initiating customer service inquiries, the category deviation is low because the behavior category matches the "office equipment" anchor category. However, for the brief click on the gaming headset product page, the behavior category deviates from the "office equipment" anchor category, and the page dwell time is only the second longest among similar pages. The node is below the percentile and then experiences a rapid return, resulting in high class deviation, path deviation, and noise characteristics.

[0130] During the behavior residual signal calculation phase, the system calculates the behavior residual value of each behavior node in the current session according to equation (4). In this embodiment, the behavior residual value for searching the "wireless printer" behavior node is... The behavioral residual value of the behavior node browsing the first printer product page is The residual value of the behavior node for collecting the first printer product is The behavioral residual value of the brief click on the gaming headset product page action node is The behavior residual value of the behavior node for quickly returning to the printer list page is The behavioral residual value of the behavior node browsing the printer consumables page is The behavioral residual value of the node that initiates customer service consultation is .

[0131] The system inputs the current session behavior sequence, stable intent anchors, and behavior residual signals into the time series model. In this embodiment, the sliding observation window length... Pick The overall intent stability of the current session is calculated according to equation (5). Intended drift degree is The degree of behavioral conflict is calculated according to equation (6). The noise intensity is If the stability threshold is The drift threshold is The conflict threshold is The noise threshold is If the current session is identified as having "high intent stability, no continuous intent drift, and local suspicious behavior but overall noise not exceeding the threshold", then the current session is identified as having "high intent stability, no continuous intent drift, and local suspicious behavior but overall noise not exceeding the threshold".

[0132] The system constructs and invokes the long-term preference sub-model, the current session sub-model, the contextual scenario sub-model, the conversion behavior sub-model, and the behavior residual sub-model. The long-term preference sub-model outputs the confidence level of "office equipment purchase intention" as follows: The confidence level of the current conversation sub-model output "office equipment purchase intention" is . The confidence level of the context-based sub-model outputting "office equipment purchase intention" is: The conversion behavior sub-model outputs the "office equipment purchase intention" with a confidence level based on collection and consultation behaviors. The behavioral residual sub-model is influenced by the brief click behavior of "gaming headset," and the confidence level of the output "digital accessories browsing intent" is [value missing]. And output the confidence level of "office equipment purchase intention" as . .

[0133] During the dynamic fusion phase, the system adjusts the fusion weights of each integrated learning sub-model according to equation (7). Since the current session intent stability reaches the stability threshold, the system increases the fusion weights of the long-term preference sub-model and the conversion behavior sub-model; since the intent drift does not reach the drift threshold, the system does not increase the dominant role of the behavior residual sub-model; since neither the behavior conflict nor the noise intensity reaches the corresponding threshold, the system only performs subsequent counterfactual verification on local behaviors affected by short stays and rapid returns, without reducing the overall weight of the current session sub-model. After dynamic fusion, the overall confidence level of "office equipment purchase intent" is... The overall confidence level of "digital accessories browsing intent" is 1. .

[0134] The system performs counterfactual behavior perturbation verification on suspicious behavior nodes. In this embodiment, "briefly clicking on the gaming headset product page" and "quickly returning to the printer list page" are identified as suspicious behavior nodes. For the "briefly clicking on the gaming headset product page" behavior node, the system performs masking processing, similar anchor point behavior replacement processing, adjacent order adjustment processing, and feature weight reduction processing to generate a counterfactual behavior sequence. After inputting the original current session behavior sequence and the counterfactual behavior sequence into the time series model and the dynamically fused ensemble learning model, respectively, the system finds that after masking this behavior node, the confidence level of "office equipment purchase intention" is reduced from... Rise to The user's current intent category has not changed, but the intent migration direction has been adjusted from "slight deviation followed by return to office equipment" to "continuous office equipment intent".

[0135] The intent contribution of the suspicious behavior node is calculated according to equation (8). Let the category change penalty coefficient be... Pick Since the user's current intent category remains unchanged before and after the disturbance, the indicator function takes... Then the intention contribution of this behavior node is If the current contribution threshold for If the intention contribution of the behavior node is lower than the contribution threshold, and the intention confidence increases and the intention migration direction tends to stabilize after the disturbance, the system will determine "briefly clicking on the gaming headset product page" as an interfering behavior node and reduce its feature weight.

[0136] For the "Initiate customer service inquiry" action node, after the system performs masking, the confidence level of "office equipment purchase intention" is reduced from... Descending to Furthermore, the user's current intent has weakened from "intent to purchase office equipment" to "intent to browse office equipment". The intent contribution of this behavior node is calculated according to equation (8). Above the contribution threshold Furthermore, the confidence of the intent decreases after the perturbation and the direction of intent migration deviates from the original prediction direction. Therefore, the system identifies "initiating customer service consultation" as a key intent behavior node and increases its feature weight.

[0137] The system inputs the current session behavior sequence (adjusted with feature weights), stable intent anchors, and behavior residual signals into the dynamically fused ensemble learning model. It then weights and summarizes the confidence scores of each candidate user intent category according to the adjusted fusion weights. The final output is that the user's current intent category is "office equipment purchase intent," with an intent confidence score of [insert confidence score here]. The intent migration direction is "continuing the stable intent anchor direction". Key influencing behavior nodes include "searching for wireless printers", "favoriting the first printer product", "browsing the printer consumables page" and "initiating customer service inquiries". Interfering behavior nodes include "briefly clicking on the gaming headset product page" and "quickly returning to the printer list page".

[0138] As can be seen from this embodiment, when both genuine intent behaviors and occasional interfering behaviors exist in the user's current session, the present invention can identify the user's normal intent using stable intent anchor points, discover local deviation behaviors using behavioral residual signals, dynamically fuse weights using time-series states, and determine the actual impact of each behavioral node on the prediction result through counterfactual behavior perturbation verification. This avoids misjudging brief clicks, quick returns, or temporary jumps as new user intents, thereby improving the accuracy, stability, and interpretability of user intent prediction.

[0139] In summary, all variables involved in weighted summation, threshold comparison, and model fusion in this embodiment are dimensionless numerical values. Original dwell time, number of behaviors, number of path nodes, and category features are all converted to a uniform scale before being included in the calculation, thus avoiding model bias caused by directly adding data of different dimensions. Furthermore, the stable intent anchor point threshold, stability threshold, drift threshold, conflict threshold, noise threshold, and contribution threshold are all determined through the distribution of training samples or rolling statistical samples, and the weights are jointly determined through initial empirical values ​​and validation set optimization. This ensures the reproducibility and adjustability of this invention across different platforms, user groups, and business cycles.

[0140] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A user intent prediction method fusing a time series model and ensemble learning, characterized in that, The method includes the following steps: Acquire multi-source behavioral data of target users within historical periods and the current session period. The multi-source behavioral data includes search data, browsing data, click data, dwell data, favorites data, add-to-cart data, inquiry data, return data, exit data, conversion data, page context data, entry source data, and trigger time data. The multi-source behavior data is sorted according to the trigger time, and behavior type encoding, behavior object encoding, context encoding, entry source encoding, and time interval encoding are performed respectively to construct the user's historical behavior sequence and current session behavior sequence; Based on the user's historical behavior sequence, long-term preference features that recur across historical time windows and are associated with conversion behavior are extracted to generate stable intent anchors that represent the user's normal intent. The current session behavior sequence is matched with the stable intent anchor to obtain a behavior residual signal that includes category deviation, intensity deviation, path deviation, context deviation, and conversion tendency deviation. The current session behavior sequence, the stable intent anchor point, and the behavior residual signal are input into the time series model, and the intent stability, intent drift, behavior conflict, and noise intensity are output. Based on the above output, the fusion weights of multiple ensemble learning sub-models are dynamically adjusted, and counterfactual behavior perturbation verification is performed on suspicious behavior nodes in the current session behavior sequence. The intention contribution of the behavior node is determined based on the difference in intention prediction results before and after the perturbation. Based on the intent contribution, the feature weights of interfering behavior nodes are reduced and the feature weights of key intent behavior nodes are strengthened, and the user's current intent category, intent confidence, intent migration direction, key influencing behavior nodes, and interfering behavior nodes are output. 2.The user intent prediction method of fusing a time sequence model and ensemble learning according to claim 1, wherein, Constructing the user's historical behavior sequence and current session behavior sequence includes: The historical periodic behavior data under the same user ID is divided into multiple historical sub-sequences according to a preset historical time window. Write the trigger time difference, behavior object category, page level, and conversion action status of two adjacent behaviors within the current session cycle into the current session behavior sequence; For behavior data with missing trigger times, fill in the missing trigger times according to the server receiving time. For behavior data that is repeatedly reported and has the same behavior object, behavior type and trigger time, retain one record. 3.The user intent prediction method of fusing a time sequence model and ensemble learning according to claim 2, characterized in that, Generating the stable intent anchor point includes: Calculate the access frequency, cumulative dwell time, number of favorites, number of add-to-carts, number of inquiries, and number of conversions for each interest category within each historical subsequence; Interest categories that appear consecutively in at least two historical subsequences and whose transformation correlation strength reaches a preset anchor threshold are identified as anchor categories. The common entry sources, page paths, and conversion actions corresponding to the anchor point categories are combined into stable intent anchor points, wherein the preset anchor point threshold is determined by the quantile of the conversion association strength in the training samples.

4. The user intent prediction method that integrates temporal models and ensemble learning according to claim 3, characterized in that, Obtaining the behavioral residual signal includes: The interest categories in the current session behavior sequence are matched with the anchor categories to obtain the category deviation; The intensity deviation is obtained by matching the dwell time, number of clicks, browsing depth, favorites, add to cart and inquiry actions in the current session behavior sequence with the intensity of normal behavior in the stable intent anchor point; The path deviation is obtained by matching the current page jump path with the normal page path in the stable intent anchor point; By matching the page context, entry source, and conversion action separately, we can obtain context deviation and conversion tendency deviation.

5. The user intent prediction method that integrates temporal models and ensemble learning according to claim 1, characterized in that, The time series model performs time series state identification, including: The current session behavior sequence, stable intent anchors, and behavior residual signals are concatenated into temporal state input features; Using time-series models, we can identify the number of consecutive occurrences of behavioral residual signals within the current session period, the consistency of deviation direction, and the rate of change of deviation magnitude. When the consistency between the current session behavior and the stable intent anchor point reaches a preset stability threshold, the corresponding intent stability is output. When the behavior residual signal continuously increases along the same intention category direction, the corresponding intention drift degree is output. Output the degree of behavioral conflict based on category conflict, abnormal path jump, and contradictory conversion actions; The noise intensity is output based on the proportion of short dwell times, quick returns, repeated clicks, and behaviors with no conversion association.

6. The user intent prediction method that integrates temporal models and ensemble learning according to claim 5, characterized in that: The multiple ensemble learning sub-models include a long-term preference sub-model, a current session sub-model, a contextual scenario sub-model, a conversion behavior sub-model, and a behavior residual sub-model. The long-term preference sub-model takes a stable intent anchor as input, the current session sub-model takes the current session behavior sequence as input, the context scenario sub-model takes the page context, entry source and trigger time as input, the conversion behavior sub-model takes the collection action, add to cart action, consultation action and conversion action as input, and the behavior residual sub-model takes the behavior residual signal as input. Each sub-model outputs the candidate user intent category and corresponding confidence level.

7. The user intent prediction method that integrates temporal models and ensemble learning according to claim 6, characterized in that, Dynamically adjust the fusion weights of each ensemble learning sub-model, including: When the intention stability reaches the preset stability threshold, increase the fusion weight of the long-term preference sub-model and decrease the fusion weight of the behavior residual sub-model. When the intent drift reaches a preset drift threshold, increase the fusion weight of the current session sub-model and the behavior residual sub-model; When the degree of behavioral conflict reaches a preset conflict threshold, the fusion weight of the sub-model that generates conflict candidate intent categories is reduced. When the noise intensity reaches a preset noise threshold, the fusion weight of sub-models dominated by short dwell times, quick returns, repeated clicks, or no-conversion-related behaviors is reduced.

8. The user intent prediction method that integrates temporal models and ensemble learning according to claim 7, characterized in that, The counterfactual behavior perturbation verification includes: Suspicious behavior nodes are filtered from the current session behavior sequence based on the degree of behavioral conflict and noise intensity; For the suspicious behavior nodes, respectively perform masking processing, similar anchor point behavior replacement processing, adjacent order adjustment processing, and feature weight reduction processing to generate corresponding counterfactual behavior sequences; The original current session behavior sequence and the counterfactual behavior sequence are respectively input into the time series model and the dynamically fused ensemble learning model. The user's current intent category, intent confidence, and intent migration direction of the two outputs are compared to obtain the intent contribution of the suspicious behavior nodes.

9. The user intent prediction method that integrates temporal models and ensemble learning according to claim 8, characterized in that, Weaken disruptive behavior nodes and strengthen key intent behavior nodes, including: When the intention contribution of a behavior node is lower than the preset contribution threshold, and the intention confidence increases or the intention migration direction tends to stabilize after perturbation, the behavior node is identified as an interfering behavior node, and its feature weight is reduced according to the difference between the intention contribution and the preset contribution threshold. When the intention contribution of a behavior node reaches the preset contribution threshold, and the intention confidence decreases after perturbation or the intention migration direction deviates from the original prediction direction, the behavior node is identified as a key intention behavior node, and its feature weight is increased according to the difference between the intention contribution and the preset contribution threshold.

10. The user intent prediction method that integrates temporal models and ensemble learning according to claim 9, characterized in that, Output the end-user intent prediction results, including: The current session behavior sequence, stable intent anchor point, and behavior residual signal after feature weight adjustment are input into the dynamically fused ensemble learning model, and the confidence of each candidate user intent category is weighted and summarized according to the adjusted fusion weights. The candidate user intent category with the highest confidence after weighted aggregation is determined as the user's current intent category; The direction of intent migration is determined based on the change in the category of the candidate user intent relative to the stable intent anchor point, and key influencing behavior nodes that contribute positively to the prediction result and interference behavior nodes whose weights are reduced are output simultaneously.

Images