A method and system for constructing a multi-channel interactive behavior sequence pattern knowledge graph

Abstract

The application discloses a kind of multi-channel interactive behavior sequence mode knowledge graph construction method and system.The method is:1) the multi-channel interactive log is preprocessed, and the generalization representation of each log record in multi-channel interactive log is obtained;2) multi-channel interactive behavior sequence mode with gap constraint is mined from multi-channel interactive log;3) for multi-channel interactive behavior sequence mode, calculate standardization support and user positive uniqueness;4) with multi-channel interactive behavior class as entity, based on multi-channel interactive behavior sequence mode, construct five-tuple set, build multi-channel interactive behavior sequence mode knowledge graph for all users, multi-channel interactive behavior sequence mode knowledge graph for target user.The application can form the structured knowledge representation covering the common law of all users and the individual preference of specific user, provide support for user experience optimization strategy in multi-channel interactive scene.

Description

Technical Field

[0001] This invention relates to the fields of data mining and human-computer interaction technology, specifically to a method for constructing a multi-channel interactive behavior sequence pattern knowledge graph. The multi-channel interactive behavior sequence pattern knowledge graph is a knowledge graph describing operational entities and their temporal relationships. It is constructed based on multi-channel operation sequence information such as touch, voice, gaze, and gesture, and can represent a user behavior knowledge network with interactive characteristics such as long duration, cross-application, multi-channel, and sequential interaction. It can be used for the analysis, generation, and prediction of user interaction behavior. Background Technology

[0002] With the development of artificial intelligence technology, intelligent interactive terminals are evolving towards multi-application parallelism and multi-channel interaction. The large-scale application of multi-channel intelligent interaction methods such as voice, gaze, and gestures enables users to flexibly switch between and more naturally operate various applications such as map navigation, audio-visual entertainment, and social networks. The massive and complex interaction log data generated during the interaction process contains user intentions and behavioral patterns. In-depth modeling and analysis of this data, describing the common patterns of user groups and the unique characteristics of individual users, is of great value for providing refined and personalized user experience optimization.

[0003] Although existing technologies can identify basic interactive behaviors from interaction logs, an increasing number of smart interactive terminals support parallel operation of multiple applications, resulting in a large number of overlapping log records of multiple application operations. At the same time, the large-scale application of multi-channel interaction methods such as voice, gaze, and gesture in smart interactive terminals has led to a lot of semantically redundant information in the log records. When combined with long periods of silence, these combined data features significantly increase the technical difficulty of analyzing the temporal relationship of interactive behaviors.

[0004] Traditional knowledge graphs construct knowledge networks using triples (entity, relation, entity), enabling the organization of general knowledge in the real world. However, they struggle to describe the temporal relationships and individual differences in user behavior. Temporal knowledge graphs can record the sequence of user actions, but they can only describe static knowledge, failing to quantify the temporal dependencies between actions and lacking descriptions of the time intervals between actions. Therefore, they are difficult to analyze user behavioral patterns and unique preferences.

[0005] The patent with publication number CN119377728A discloses an interactive behavior analysis method, system, electronic device and storage medium. It proposes a feature extraction and statistical analysis method for touch behaviors such as clicking and swiping, which is used for user experience evaluation or interface design optimization. However, it only focuses on the statistical analysis of interactive behaviors of a single application, and the interaction method is singular. It cannot analyze the temporal relationship between interactive behaviors.

[0006] Patent CN120197737A discloses a technology for application sequence prediction, including a method, device, electronic device, and storage medium. Based on application usage records, it can predict users' subsequent usage through a time-series prediction model, thereby achieving application recommendation. Its main focus is on the task of predicting the subsequent usage of applications.

[0007] Patent CN106469171A discloses a method for mining frequent sequence patterns in time series, which proposes a segment frequency index structure, mainly used to mine frequent sequence patterns from time series data. Summary of the Invention

[0008] To address the technical problems existing in the prior art, the present invention aims to provide a multi-channel interactive behavior sequence pattern knowledge graph and its construction method. Starting with a user's multi-channel interaction logs, the present invention mines multi-channel interactive behavior sequence patterns and constructs multi-channel interactive behavior sequence pattern knowledge graphs for all users and for a specific user being observed individually (i.e., the target user), providing support for user experience optimization tasks such as refined services and personalized recommendations.

[0009] The multi-channel interaction log refers to the time-series data of interactive behaviors generated when multiple users operate various applications through multiple channels such as touch, voice, gaze, and gestures. It is generally provided by event tracking programs in intelligent interactive terminals such as smartphones, tablets, and in-vehicle infotainment systems. Each data entry in the multi-channel interaction log, referred to as a log record, is a vector denoted as rec, consisting of information such as timestamp, user ID, application identifier, interaction channel, interaction action, operation object, and input content. t Where t represents the index of the log record in the multi-channel interactive log.

[0010] The multi-channel interactive behavior consists of interactive channels and interactive actions, such as touch clicks, voice commands, gaze focusing, and gesture operations. The multi-channel interactive behavior sequence is a set of multiple multi-channel interactive behaviors arranged in chronological order. The multi-channel interactive behavior sequence pattern refers to a regular arrangement of multi-channel interactive behaviors exhibiting characteristics such as repetition, stability, and recognizability in their occurrence order and time intervals. The multi-channel interactive behavior sequence pattern knowledge graph is a knowledge network composed of a quintuple (entity 1, relation, interval attribute, transition probability, entity 2). Here, an entity refers to a subcategory of multi-channel interactive behavior, a relation defines the order of entity 1 and entity 2, an interval attribute indicates the time interval between entity 1 and entity 2, and the transition probability quantifies the probability of transitioning from one type of behavior represented by entity 1 to another type of behavior represented by entity 2.

[0011] To describe the common behavioral patterns of a user group, this invention proposes to construct a knowledge graph of multi-channel interactive behavior sequence patterns for all users, aiming to represent the frequently occurring multi-channel interactive behavior sequence patterns of all users as the five-tuple; to describe the behavioral patterns of a specific user, this invention proposes to construct a knowledge graph of multi-channel interactive behavior sequence patterns for a specific user, aiming to represent the frequently occurring multi-channel interactive behavior sequence patterns of a specific user but less frequently occurring in other users as the five-tuple.

[0012] The multi-channel interactive behavior sequence pattern knowledge graph construction method proposed in this invention consists of the following five steps.

[0013] The first step involves preprocessing the multi-channel interaction logs to form a generalized representation of behavioral semantics and temporal features. Further, this step is subdivided into four sub-steps: behavior grouping, feature encoding, fine-grained clustering, and attribute generalization.

[0014] The behavior grouping refers to grouping log records in the multi-channel interaction log one by one according to the interaction channels and interaction actions in the log records. t The data is categorized into several sets of multi-channel interactive behaviors. Each set contains a group of unordered log records of the same category.

[0015] The feature encoding refers to establishing a feature extractor and a feature encoder to extract semantic and temporal features from each log record, forming a feature code. The feature extractor consists of a bidirectional encoder representation model (BERT) and a feature fusion layer. It converts information such as interaction channels, interaction actions, and operation objects in the log record into word embedding vectors, which are then combined with the record's index in the sequence and the time difference between the record and the previous record to form a fused feature vector. The feature encoder is composed of four stacked Transformer layers. Using the fused feature vector as input data, it performs a self-supervised learning task of mask prediction, enabling the feature encoder to learn the temporal and semantic features, thereby obtaining the rec value of any log record. t The feature encoding, denoted as feat t The self-supervised learning task for mask prediction involves randomly occluding the interaction channels, actions, and objects of a subset of log records in the interaction log, and training an encoder to predict the occluded content. Specifically, 15% of the log records are randomly selected, and 40% of the interaction channel representation vectors, 40% of the interaction action representation vectors, and 20% of the object representation vectors are replaced with the [MASK] label. The loss function for the self-supervised learning task of mask prediction is...

[0016]

[0017] Where M represents the set of interactive log records for which a mask is applied, and mask(rec m ) represents the log record after the mask, p(rec m |mask(rec m )) represents the probability predicted by the feature encoder.

[0018] The refined clustering refers to performing density clustering on the feature codes of log records in a multi-channel interaction behavior set as a unit, obtaining several sub-category clusters, and calculating the probability of each log record belonging to each sub-category cluster. Taking the "touch click" behavior set as an example, density clustering of all log records in this set can obtain N sub-category clusters, denoted as C1, C2, ..., C6. N A binary tuple (behavior, cluster) consisting of multi-channel interaction behavior and any cluster represents the sub-category to which a log record belongs after fine clustering, denoted as an item, such as (touch click, C1), (touch click, C2), (touch swipe, C1), (voice command, C1), (voice command, C2), (eye gaze, C1), (eye gaze, C2), (eye gaze, C3), etc.

[0019] Furthermore, the calculation process for the attribution probability includes cluster center calculation and similarity calculation. The cluster center calculation refers to calculating the cluster center for any cluster C obtained after density clustering. n The cluster center is obtained by averaging the feature codes of all the log records it contains. n ,Right now

[0020]

[0021] Among them, C n Indicates cluster C k The number of log records included. The similarity calculation refers to calculating the feature code (feat) of each log record in each multi-channel interaction behavior set. t to each cluster center n cosine similarity n (t), and convert it into a similarity score. n (t), that is

[0022]

[0023] Wherein, β represents the influence coefficient of the cosine similarity on the similarity score, denoted as cluster C. n All log records and cluster center n The average cosine similarity, i.e. The attribution probability p n(t) calculation refers to the calculation of log record rec t Clusters with similarity scores higher than parameter ρ are selected to form a candidate cluster set. Then, the similarity scores in the set are normalized to obtain rec. t Belonging to each candidate cluster C n The probability p n (t), that is

[0024]

[0025] Wherein, the probability of belonging p n The value range of (t) is [0,1]; the log record rec t The sum of the probability of belonging to all subcategories is 1; if and only if rec t When it does not belong to a certain set of behaviors, rec t The probability of belonging to any of the subcategories of this behavior set is 0. For example, rec t If it belongs to the "touch click" behavior set, then rec t The probability of belonging to the subcategories such as (touch swipe, C_1), (voice command, C_1), (voice command, C_2), (eye gaze, C_1), (eye gaze, C_2), and (eye gaze, C_3) is 0.

[0026] The attribute generalization refers to applying the generalization of any log record rec. t The set of two tuples {(item, p)} consisting of the item's subcategories and their probabilities of belonging are appended to the original log record. The log record after attribute generalization can be represented as:

[0027]

[0028] The attribute generalization enables the original log records to not only contain basic information, but also implicitly contain their probability distribution characteristics in the behavioral pattern space.

[0029] The second step is to mine multi-channel interaction behavior sequence patterns with gap constraints, that is, to discover sequence structures composed of multiple different multi-channel interaction behavior categories from the interaction logs, for example... <item i ,item j ,item k In the sequence structure, the interval between log records contained in any two adjacent multi-channel interaction behavior categories does not exceed the preset parameter τ. For example, if there are subsequences in the interaction log. <rec a ,rec b ,rec c > and meets the following conditions:

[0030] a)rec aFor item i probability of belonging

[0031] b)rec b For item j probability of belonging

[0032] c)rec c For item k probability of belonging

[0033] d)rec a ,rec b ,rec c They have a temporal relationship but are not necessarily adjacent in the interaction logs, and rec a With rec b rec b With rec c The time intervals between them do not exceed the preset parameter τ.

[0034] but <item i ,item j ,item k >It is a multi-channel interaction behavior sequence pattern with gap constraints mined from interaction logs.

[0035] Furthermore, the mining of multi-channel interactive behavior sequence patterns with gap constraints includes three sub-steps: fragment segmentation, pattern retrieval, and pattern probability calculation.

[0036] The segmentation refers to first grouping the interaction logs according to the user ID, and then dividing each user's interaction logs into several non-overlapping segments according to the segmentation principle that "if the interval between two adjacent log records exceeds the parameter ω, they should be divided into two segments".

[0037] The pattern retrieval process works on a segment-by-segment basis, traversing the multi-channel interactive behavior category to which each log record in the segment belongs, forming a category candidate set, and generating a multi-channel interactive behavior sequence pattern of length 1, i.e., containing only one multi-channel interactive behavior category. Based on this, for each generated multi-channel interactive behavior sequence pattern, any multi-channel interactive behavior category not yet appearing in the current multi-channel interactive behavior sequence pattern is selected from the category candidate set and linked to the end of that multi-channel interactive behavior sequence pattern, thus forming a longer multi-channel interactive behavior sequence pattern. This process is repeated until no new multi-channel interactive behavior sequence patterns can be generated. To avoid generating excessively long multi-channel interactive behavior sequence patterns, a pre-set parameter μ is used to constrain the maximum length of the multi-channel interactive behavior sequence pattern and limit the number of multi-channel interactive behavior categories contained in the multi-channel interactive behavior sequence pattern. The pattern retrieval stage will retain multi-channel interactive behavior sequence patterns with a length between 2 and μ.

[0038] The pattern probability calculation refers to a given fragment s = <rec 1 ,rec 2 ,…,rec N >and multi-channel interactive behavior sequence pattern α= <item i ,item j ,item k Define a derivation function F(α,x,y) to compute the subsequence from segment s. <rec 1 ,rec 2 ,…,rec y Pattern α is found in (where (1≤y≤N)) and in the subsequence of fragment s. <rec x ,rec x+1 ,…,rec y >The last multi-channel interactive behavior category item was found to be in pattern α. k The probability of finding pattern α in fragment s. F(α,1,N) represents the probability of finding pattern α in fragment s if and only if x=1 and y=N. The actual calculation method of the derived function F(α,x,y) is as follows:

[0039]

[0040] in, This indicates that the log record is stored in the rec. y Item belonging to the multi-channel interaction behavior category k The probability; α prefix The prefix pattern representing pattern α, i.e., α prefix = <item i ,item j >; SP(y) represents the segment s that satisfies "with log record recy The smallest index for log records with an interval ≤ τ, i.e.

[0041] SP(y) = min{t|t <y∧timestamp(rec y )-timestamp(rec t )≤τ}(7)

[0042] function F(α) prefix SP(y), y-1) represents the subsequence <rec 1 ,rec 2 ,…,rec y-1 Prefix pattern α was found in > prefix Meanwhile, in the subsequence <rec SP(y) ,rec SP(y)+1 ,…,rec y-1 Prefix pattern α was found in > prefix The last multi-channel interactive behavior category item j The probability of . The boundary conditions of the derived function F(α,x,y) include:

[0043] a) When y≤0 or x>y, the value of the function is 0, that is, F(α,x,y)=0;

[0044] b) When the multi-channel interactive behavior sequence pattern α is empty, i.e., it does not contain a multi-channel interactive behavior category, the value of this function is 1.

[0045] That is, F(α,x,y)=1.

[0046] The third step is to calculate the standardization support and positive user uniqueness for multi-channel interaction behavior sequence patterns.

[0047] Traditional support refers to the average probability of a multi-channel interactive behavior sequence pattern occurring in all segments. However, the probability of a multi-channel interactive behavior sequence pattern decreases as the pattern length increases, making it impossible to directly compare the support of patterns of different lengths. This invention uses an nth-power root transform to standardize the occurrence probability, thereby reducing the impact of pattern length on support.

[0048] The method for calculating the standardized support is as follows:

[0049]

[0050] Where scsupp(α) represents the normalized support of pattern α across all fragments of all users; scsupp(α,u) represents the normalized support of pattern α across all fragments of a specific user u. The pattern α is represented in fragment s rThe probability of occurrence in; |S| represents the total number of segments for all users; |S u | represents the number of fragments for user u, and |α| represents the number of multi-channel interactive behavior categories in the multi-channel interactive behavior sequence pattern α.

[0051] The user positive uniqueness is used to quantify the ability of multi-channel interaction behavior sequence patterns to represent specific user behavior characteristics, thereby measuring whether the pattern can serve as a significant behavioral feature of a specific user. Its calculation method is as follows:

[0052] ps(α,u)=scsupp(α,u)-scsupp(α)(10)

[0053]

[0054] Wherein, Φ represents the set of multi-channel interactive behavior sequence patterns mined from all users; |Φ| represents the number of multi-channel interactive behavior sequence patterns contained in the set Φ; ps(α,u) represents the degree of personalization deviation of pattern α from user u, quantifying the difference in the standardization support of pattern α in user u relative to all users. The larger the value, the more concentrated the pattern α appears in user u's log records; pc(α,u) represents the positive personalization contribution of pattern α to user u, quantifying the increment of the degree of personalization deviation of pattern α in user u relative to the average deviation level of other patterns of user u. The larger the value, the more the degree of personalization deviation of pattern α from user u exceeds the average level, that is, the more effective pattern α is in characterizing the individual behavioral characteristics of user u; pr(α,u) represents the positive user uniqueness of user u with respect to pattern α, quantifying the relative proportion of pattern α in the set of all patterns that can bring positive personalization contribution to user u. The larger the value, the higher the contribution proportion of pattern α in the set of patterns that contribute positive personalization, that is, the more critical pattern α is in characterizing the individual behavioral characteristics of user u.

[0055] The fourth step is to construct a knowledge graph of multi-channel interaction behavior sequence patterns for all users. This knowledge graph uses multi-channel interaction behavior categories as entities and, based on the multi-channel interaction behavior sequence patterns with gap constraints mined in the previous steps, constructs a set of five-tuples (entity 1, relation, interval attribute, transition probability, entity 2) that can represent the knowledge network structure. Specifically, given a multi-channel interaction behavior sequence pattern of length 2, the two multi-channel interaction behavior categories in the pattern are respectively designated as entity 1 and entity 2; the relation between entities is explicitly a temporal relation, i.e., entity 1 occurs before entity 2; the gap constraint in the interval attribute is ≤τ; and the transition probability from entity 1 to entity 2... The calculation is based on the standardization support of the pattern. The sequence pattern knowledge graph constructed using the above rules can reflect the temporal dependencies and statistical regularities of user interaction behaviors at the group level.

[0056] The fifth step is to construct a knowledge graph of multi-channel interaction behavior sequence patterns for a specific user. This knowledge graph of multi-channel interaction behavior sequence patterns for a specific user can be used for tasks such as predicting the user's next interaction behavior and recommending applications that the user is interested in. Structurally, this graph is still a set of five tuples, where the transition probability from entity 1 to entity 2 is... The calculation is based on user positive uniqueness. Specifically, based on a set of multi-channel interaction behavior sequence patterns with a length of 2 or greater, all multi-channel interaction behavior sequence patterns that can form the target quintuple are integrated. The mean user positive uniqueness of these patterns is calculated and weighted with the standardized support of the corresponding patterns to obtain the transition probability in the target quintuple. i , first occurs at, ≤τ, item j For example, the transition probability. The calculation method is as follows:

[0057]

[0058] Where u represents the ID of the specific user being targeted; Ψ (i,j) Indicates the presence of subsequences <item i ,item j A collection of multi-channel interactive behavior sequence patterns. Standardized support is introduced to avoid the special case where a user's positive uniqueness is 0, resulting in an edge weight of 0. Calculating the transition probability between entities based on the user's positive uniqueness of the pattern can significantly enhance the uniqueness of a specific user's behavioral patterns in cross-application interaction scenarios.

[0059] The advantages of this invention are as follows:

[0060] This invention provides a novel knowledge graph and its construction method. By modeling the temporal dependencies of multi-channel interaction behaviors based on complex information from interaction logs, a structured knowledge representation covering common patterns of all users and individual preferences of specific users is formed, providing support for user experience optimization strategies in multi-channel interaction scenarios.

[0061] 1. This invention constructs a knowledge graph by mining multi-channel interactive behavior sequence patterns, focusing on analyzing the temporal relationship and interval time between behaviors, which can support more refined interaction design optimization.

[0062] 2. The interaction log data used in this invention not only includes application usage records, but also user operation behaviors within the application. The constructed multi-channel interaction behavior sequence pattern knowledge graph can describe the user's behavior patterns when using a specific application, as well as the user's behavior patterns when switching between multiple applications, and is not limited to the order of application usage.

[0063] 3. This invention proposes a novel frequent user interaction sequence pattern and its mining method, as well as two new indicators: standardized support and positive user uniqueness. It can not only mine the frequent behavior patterns of user groups, but also the unique behavior patterns of specific users. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the basic process of a multi-channel interactive behavior graph construction method based on pattern mining in one embodiment.

[0065] Figure 2 This is a schematic diagram of the basic process of interactive log preprocessing in the embodiment.

[0066] Figure 3 This is a schematic diagram of a sequence pattern knowledge graph for all users in the embodiment.

[0067] Figure 4 This is a schematic diagram of the sequence pattern knowledge graph for user1 in the embodiment.

[0068] Figure 5 This is a schematic diagram of the sequence pattern knowledge graph for user2 in the embodiment. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0070] Taking mobile phone users as an example, the original interaction logs of users1 and user2 are exported from a certain brand of Android phone. The log records are represented as a seven-tuple (timestamp, user ID, application identifier, interaction channel, interaction action, operation object, input content). The interaction logs collected in this embodiment are as follows:

[0071] rec 1 =(9:00:03, user1, Gaode Map, touch, click, search box, West Lake)

[0072] rec 2 =(9:00:05, user1, Gaode Map, voice, speak, / , navigate to West Lake)

[0073] rec3 = (9:00:22, user1, Gaode Map, touch, click, route list, shortest route)

[0074] rec 4 =(9:01:28, user1, system, gaze, attention, notification bar, / )

[0075] rec 5 = (9:01:45, user1, NetEase Cloud Music, voice, speak, / , play popular songs)

[0076] rec 6 =(9:02:01, user1, NetEase Cloud Music, gesture, swipe left, / , / )

[0077] rec 7 =(9:02:25, user1, system, touch, swipe down, taskbar, / )

[0078] rec 8 =(9:02:28, user2, system, gaze, attention, notification bar, / )

[0079] rec 9 =(9:02:31, user2, system, voice, speak, / , music)

[0080] rec 10 =(9:02:53, user2, NetEase Cloud Music, touch, click, pause button, / )

[0081] rec 11 =(9:03:15, user2, Gaode Map, touch, click, search box, week)

[0082] rec 12 = (9:03:17, user2, NetEase Cloud Music, voice, speak, / , play Jay Chou's song)

[0083] rec 13 =(9:06:52, user2, system, touch, swipe down, taskbar, / )

[0084] rec 14 =(9:07:05, user2, WeChat, touch, click, chat list, / )

[0085] rec 15 =(9:07:28, user2, WeChat, voice, speak, / , send me location sharing)

[0086] rec 16=(9:07:50, user2, system, touch, swipe down, taskbar, / )

[0087] rec 17 =(9:07:53, user2, system, touch, click, player, next track)

[0088] rec 18 =(10:01:01, user1, WeChat, touch, click, chat list, / )

[0089] rec 19 =(10:01:23, user1, WeChat, voice message, speaking, / , arrived at the agreed location)

[0090] rec 20 =(10:01:46, user1, WeChat, touch, click, map card, / )

[0091] rec 21 =(10:01:52, user1, WeChat, touch, click, location sharing, / )

[0092] rec 22 =(10:02:01, user1, WeChat, voice message, speak, / , estimated arrival time)

[0093] rec 23 =(10:02:08, user1, WeChat, voice, speak, / , I have arrived).

[0094] The first step is to preprocess the interaction logs. First, based on the interaction channel and action, the interaction logs are divided into five categories of multi-channel interaction behaviors: touch click, touch swipe down, voice speech, eye gaze, and left swipe gesture. The log records for each category are as follows:

[0095] Touch click = {rec 1 ,rec 3 ,rec 10 ,rec 11 ,rec 14 ,rec 17 ,rec 18 ,rec 20 ,rec 21}

[0096] Touch swipe down = {rec 7 ,rec 13 ,rec 16}

[0097] Voice speech = {rec 2 ,rec5 ,rec 9 ,rec 12 ,rec 15 ,rec 19 ,rec 22 ,rec 23}

[0098] Eye contact = {rec 4 ,rec 8}

[0099] Swipe left = {rec 6}

[0100] Next, the interaction logs are converted into feature codes. For each log record, the application identifier, interaction channel, interaction action, operation object, input content, etc., are extracted to form a word list. For example, from the first log record "rec",... 1 The extracted word list is [Gaode Map, touch, click, search box, West Lake]; a pre-trained BERT-Base-Chinese model is used to transform the word list into a 768-dimensional semantic vector, and the vector is linked to the log record rec. 1 The position index (32-dimensional, with the position index of the first log record being 0) and the time difference with the previous log record (32-dimensional, with the time difference of the first log record being 0) constitute a fused feature vector. The fused feature vectors of all interaction logs are fed into a feature encoder, which is trained by performing a self-supervised learning task of mask prediction to obtain the feature encoding of the interaction logs. In this embodiment, the feature encoder operates for approximately 30 epochs with a learning rate of 1×10⁻⁶. -4 It converges under the condition of batch size 512.

[0101] Then, using a set of multi-channel interaction behaviors as units, the similarity between log records in the set is calculated using cosine distance, and the DBSCAN algorithm (with a neighborhood radius of 0.6 and a minimum sample size of 10) is used to perform density clustering on the log records in the set. For example, the log records in the set "touch click" are clustered into 3 clusters C1, C2, and C3. According to the above steps, the interaction logs in this embodiment can be divided into 8 subcategories: (touch click, C1), (touch click, C2), (touch click, C3), (touch swipe down, C1), (voice talk, C1), (voice talk, C2), (eye gaze, C1), and (gesture left swipe, C1). For simplicity, the symbol "item" is used. iThis refers to multi-channel interactive behavior categories, such as: item1←(touch click, C1), item2←(touch click, C2), item3←(touch click, C3), item4←(touch swipe down, C1), item5←(voice talk, C1), item6←(voice talk, C2), item7←(eye gaze, C1), item8←(gesture swipe left, C1).

[0102] Next, calculate the probability of association between log records and their respective subcategories. (Using log record rec...) 3 For example, the possible subcategories include (touch click, C1), (touch click, C2), and (touch click, C3). Assuming that the cosine similarity between its feature encoding and the center point of cluster C1 is 0.08, and the average distance from all log records in cluster C1 to the center point of cluster C1 is 0.1, rec is calculated according to formula (4). 3 The similarity score s1(1) with cluster C1 is exp(-1×0.08÷0.1)=0.449. In this embodiment, the similarity score parameter ρ=0.2 is set, and clusters with similarity scores higher than 0.2 are retained. Then, the record rec is calculated according to formula (5). 3 The probability of belonging to clusters C1, C2, C3, for example: p1(3) = 0.78, p2(3) = 0.22, p3(3) = 0.

[0103] Finally, based on the subcategories and their belonging probabilities, each interaction log can be expanded to include a set of binary tuples (category code, probability) as a new attribute. In this embodiment, the interaction log can be expanded to the following form:

[0104] rec 1 =rec 1 ∪{(item2,1.0)};

[0105] rec 2 =rec 2 ∪{(item5:1.0)};

[0106] rec 3 =rec 3 ∪{(item1:0.78),(item2,0.22)};

[0107] rec 4 =rec 4 ∪{(item7:1.0)};

[0108] rec 5 =rec 5 ∪{(item5:0.75),(item6:0.25)};

[0109] rec 6 =rec 6 ∪{(item8:1.0)};

[0110] rec 7 =rec 7 ∪{(item4:1.0)};

[0111] rec 8 =rec 8 ∪{(item7:1.0)};

[0112] rec 9 =rec 9 ∪{(item5:1.0)};

[0113] rec 10 =rec 10 ∪{(item4:1.0)};

[0114] rec 11 =rec 11 ∪{(item2:1.0)};

[0115] rec 12 =rec 12 ∪{(item5:0.83),(item6:0.17)};

[0116] rec 13 =rec 13 ∪{(item4:1.0)};

[0117] rec 14 =rec 14 ∪{(item1:1.0)};

[0118] rec 15 =rec 15 ∪{(item5:0.2),(item6:0.8)};

[0119] rec 16 =rec 16 ∪{(item4:1.0)};

[0120] rec 17 =rec 17 ∪{(item3:1.0)};

[0121] rec 18 =rec 18 ∪{(item2:1.0)};

[0122] rec 19 =rec 19 ∪{(item5:0.17),(item6:0.83)};

[0123] rec 20 =rec 20 ∪{(item1:1.0)};

[0124] rec 21 =rec 21 ∪{(item1:1.0)};

[0125] rec 22 =rec 22 ∪{(item5:0.25),(item6:0.75)};

[0126] rec 23 =rec 23 ∪{(item5:0.17),(item6:0.83)}.

[0127] The second step is to mine multi-channel interactive behavior sequence patterns with gap constraints. In this embodiment, the interval parameter ω = 60 seconds, the maximum length parameter μ of the multi-channel interactive behavior sequence pattern = 3, and the gap constraint parameter τ = 20 seconds are set.

[0128] First, the interaction log is divided into segments. The interaction log is divided into multiple ordered subsequences according to user ID. Then, according to the segmentation principle defined in this invention: if the time difference between two adjacent log entries exceeds 60 seconds (parameter ω), they should be divided into two segments. User 1's interaction log is divided into two interaction segments: User2's interaction log was split into two interaction segments:

[0129] Then, for each interactive segment as well as Perform pattern retrieval. Using interactive fragments. For example, iterate through the set of tuples (category code, probability) for each log record to generate a multi-channel interaction behavior sequence pattern of length 1, including: <item1> 、 <item2> 、 <item4> 、 <item5> 、 <item6> 、 <item7> 、 <item8>For existing patterns, use patterns <item2>For example, from the interactive segment The first log record in the middle containing the multi-channel interactive behavior category item2 1 Start searching backwards, only retrieving subsequent log records and rec. 1 If the interval between events does not exceed 20 seconds (parameter τ), and the log records contain multi-channel interactive behavior categories different from item2, then the new multi-channel interactive behavior category can be appended to the end of the existing pattern to generate a longer multi-channel interactive behavior sequence pattern, for example:<item2,item5> ,<item2,item1> Repeat the above process until no new patterns satisfying the gap constraint and with a length not exceeding 3 can be generated. All multi-channel interactive behavior sequence patterns with a length not less than 2 and not greater than 3 constitute the pattern set Φ.

[0130] Next, for the retrieved multi-channel interactive behavior sequence patterns, the probability of their occurrence in each interactive segment is calculated. (This is based on the multi-channel interactive behavior sequence patterns.)<item1,item6> For example, interactive segments There are three log record sequences related to this pattern: <20,22>, <21,22>, and <21,23>. According to formula (6), their occurrence probability is 0.83×1.0+(1-0.83)×(0.75×1.0)≈0.958; Interactive fragment This pattern does not exist in the game; its probability of occurrence is 0. (Interactive segment) This pattern does not exist in the game; its probability of occurrence is 0. (Interactive segment) This pattern does not exist in the game, and its probability of occurrence is 0.

[0131] Based on the above calculation method, the multi-channel interactive behavior sequence patterns that can be mined in this embodiment and their probability of occurrence in interactive segments are as follows:

[0132] Multi-channel interactive behavior sequence pattern<item1,item5> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.378; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0133] Multi-channel interactive behavior sequence pattern<item1,item6> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.958; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0134] Multi-channel interactive behavior sequence pattern<item1,item5,item6> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.208; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0135] Multi-channel interactive behavior sequence pattern<item1,item6,item5> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.128; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0136] Multi-channel interactive behavior sequence pattern<item2,item1> exist The probability of it appearing in the middle is 0.78; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0137] Multi-channel interactive behavior sequence pattern<item2,item5> exist The probability of it appearing in the middle is 1.0; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.83; in The probability of it appearing in the middle is 0.0.

[0138] Multi-channel interactive behavior sequence pattern<item2,item5,item1> exist The probability of it appearing in the middle is 0.78; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0139] Multi-channel interactive behavior sequence pattern<item2,item6> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.17; in The probability of it appearing in the middle is 0.0.

[0140] Multi-channel interactive behavior sequence pattern<item4,item1> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing is 1.0.

[0141] Multi-channel interactive behavior sequence pattern<item4,item3> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing is 1.0.

[0142] Multi-channel interactive behavior sequence pattern<item5,item1> exist The probability of it appearing in the middle is 0.78; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0143] Multi-channel interactive behavior sequence pattern<item5,item2> exist The probability of it appearing in the middle is 0.22; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0144] Multi-channel interactive behavior sequence pattern<item5,item6> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.208; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0145] Multi-channel interactive behavior sequence pattern<item5,item8> exist The probability of it appearing in the middle is 0.75; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0146] Multi-channel interactive behavior sequence pattern<item6,item5> exist The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.128; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0147] Multi-channel interactive behavior sequence pattern<item6,item8> exist The probability of it appearing in the middle is 0.25; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0148] Multi-channel interactive behavior sequence pattern<item7,item5> exist The probability of it appearing in the middle is 0.75; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 1.0; in The probability of it appearing in the middle is 0.0.

[0149] Multi-channel interactive behavior sequence pattern<item7,item5,item8> exist The probability of it appearing in the middle is 0.75; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0150] Multi-channel interactive behavior sequence pattern<item7,item6> exist The probability of it appearing in the middle is 0.25; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0151] Multi-channel interactive behavior sequence pattern<item7,item6,item8> exist The probability of it appearing in the middle is 0.25; in The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0; The probability of it appearing in the middle is 0.0.

[0152] The third step is to calculate the normalized support and positive uniqueness of the interaction sequence pattern. First, according to formulas (8) and (9), the normalized support for all users and for specific users is calculated respectively. The calculation results are as follows:

[0153] Multi-channel interactive behavior sequence pattern<item1,item5> The standardization support for all users is 0.154;

[0154] The standardization support for user1 is 0.307; the standardization support for user2 is 0.0.

[0155] Multi-channel interactive behavior sequence pattern<item1,item6> The standardization support for all users is 0.245;

[0156] The standardization support for user1 is 0.489; the standardization support for user2 is 0.0.

[0157] Multi-channel interactive behavior sequence pattern<item1,item5,item6> The standardization support for all users is 0.148; the standardization support for user1 is 0.296; and the standardization support for user2 is 0.0.

[0158] Multi-channel interactive behavior sequence pattern<item1,item6,item5> The standardization support for all users is 0.126; the standardization support for user1 is 0.252; and the standardization support for user2 is 0.0.

[0159] Multi-channel interactive behavior sequence pattern<item2,item1> The standardization support for all users is 0.221;

[0160] The standardization support for user1 is 0.442; the standardization support for user2 is 0.0.

[0161] Multi-channel interactive behavior sequence pattern<item2,item5> The standardization support for all users is 0.478;

[0162] The standardization support for user1 is 0.5; the standardization support for user2 is 0.456.

[0163] Multi-channel interactive behavior sequence pattern<item2,item5,item1> The standardization support for all users is 0.23; the standardization support for user1 is 0.46; and the standardization support for user2 is 0.0.

[0164] Multi-channel interactive behavior sequence pattern<item2,item6> The standardization support for all users is 0.103;

[0165] The standardization support for user1 is 0.0; the standardization support for user2 is 0.206.

[0166] Multi-channel interactive behavior sequence pattern<item4,item1> The standardization support for all users is 0.25; the standardization support for user1 is 0.0; and the standardization support for user2 is 0.5.

[0167] Multi-channel interactive behavior sequence pattern<item4,item3> The standardization support for all users is 0.25;

[0168] The standardization support for user1 is 0.0; the standardization support for user2 is 0.5.

[0169] Multi-channel interactive behavior sequence pattern<item5,item1> The standardization support for all users is 0.221;

[0170] The standardization support for user1 is 0.442; the standardization support for user2 is 0.0.

[0171] Multi-channel interactive behavior sequence pattern<item5,item2> The standardization support for all users is 0.117;

[0172] The standardization support for user1 is 0.235; the standardization support for user2 is 0.0.

[0173] Multi-channel interactive behavior sequence pattern<item5,item6> The standardization support for all users is 0.114;

[0174] The standardization support for user1 is 0.228; the standardization support for user2 is 0.0.

[0175] Multi-channel interactive behavior sequence pattern<item5,item8> The standardization support for all users is 0.217;

[0176] The standardization support for user1 is 0.433; the standardization support for user2 is 0.0.

[0177] Multi-channel interactive behavior sequence pattern<item6,item5> The standardization support for all users is 0.089;

[0178] The standardization support for user1 is 0.179; the standardization support for user2 is 0.0.

[0179] Multi-channel interactive behavior sequence pattern<item6,item8> The standardization support for all users is 0.125;

[0180] The standardization support for user1 is 0.25; the standardization support for user2 is 0.0.

[0181] Multi-channel interactive behavior sequence pattern<item7,item5> The standardization support for all users is 0.467;

[0182] The standardization support for user1 is 0.433; the standardization support for user2 is 0.5.

[0183] Multi-channel interactive behavior sequence pattern<item7,item5,item8> The standardization support for all users is 0.227; the standardization support for user1 is 0.454; and the standardization support for user2 is 0.0.

[0184] Multi-channel interactive behavior sequence pattern<item7,item5,item4> The standardization support for all users is 0.25; the standardization support for user1 is 0.0; and the standardization support for user2 is 0.5.

[0185] Multi-channel interactive behavior sequence pattern<item7,item6> The standardization support for all users is 0.125;

[0186] The standardization support for user1 is 0.25; the standardization support for user2 is 0.0.

[0187] Multi-channel interactive behavior sequence pattern<item7,item6,item8> The standardization support for all users is 0.158; the standardization support for user1 is 0.315; and the standardization support for user2 is 0.0.

[0188] Then, according to formulas (10), (11), and (12), the positive uniqueness of each interaction sequence pattern for users user1 and user2 is calculated, and the results are as follows:

[0189] Multi-channel interactive behavior sequence pattern<item1,item5> The positive uniqueness of user1 is 0.158;

[0190] The positive uniqueness of user2 is 0.0.

[0191] Multi-channel interactive behavior sequence pattern<item1,item6> The positive uniqueness of user1 is 0.268;

[0192] The positive uniqueness of user2 is 0.0.

[0193] Multi-channel interactive behavior sequence pattern<item1,item5,item6> The positive uniqueness of user1 is 0.149; the positive uniqueness of user2 is 0.0.

[0194] Multi-channel interactive behavior sequence pattern<item1,item6,item5> The positive uniqueness of user1 is 0.286; the positive uniqueness of user2 is 0.0.

[0195] Multi-channel interactive behavior sequence pattern<item2,item1> The positive uniqueness of user1 is 0.246;

[0196] The positive uniqueness of user2 is 0.0.

[0197] Multi-channel interactive behavior sequence pattern<item2,item5> The positive uniqueness of user1 is 0.0;

[0198] The positive uniqueness of user2 is 0.176.

[0199] Multi-channel interactive behavior sequence pattern<item2,item5,item1> The positive uniqueness of user1 is 0.255; the positive uniqueness of user2 is 0.0.

[0200] Multi-channel interactive behavior sequence pattern<item2,item6> The positive uniqueness of user1 is 0.0; the positive uniqueness of user2 is 0.31.

[0201] Multi-channel interactive behavior sequence pattern<item4,item1> The positive uniqueness of user1 is 0.0; the positive uniqueness of user2 is 0.364.

[0202] Multi-channel interactive behavior sequence pattern<item4,item3> The positive uniqueness of user1 is 0.0; the positive uniqueness of user2 is 0.364.

[0203] Multi-channel interactive behavior sequence pattern<item5,item1> The positive uniqueness of user1 is 0.246;

[0204] The positive uniqueness of user2 is 0.0.

[0205] Multi-channel interactive behavior sequence pattern<item5,item2> The positive uniqueness of user1 is 0.087;

[0206] The positive uniqueness of user2 is 0.0.

[0207] Multi-channel interactive behavior sequence pattern<item5,item6> The positive uniqueness of user1 is 0.079;

[0208] The positive uniqueness of user2 is 0.0.

[0209] Multi-channel interactive behavior sequence pattern<item5,item8> The positive uniqueness of user1 is 0.241;

[0210] The positive uniqueness of user2 is 0.0.

[0211] Multi-channel interactive behavior sequence pattern<item6,item5> The positive uniqueness of user1 is 0.0; the positive uniqueness of user2 is 0.0.

[0212] Multi-channel interactive behavior sequence pattern<item6,item8> The positive uniqueness of user1 is 0.105;

[0213] The positive uniqueness of user2 is 0.0.

[0214] Multi-channel interactive behavior sequence pattern<item7,item5> The positive uniqueness of user1 is 0.0; the positive uniqueness of user2 is 0.25.

[0215] Multi-channel interactive behavior sequence pattern<item7,item5,item8> The positive uniqueness of user1 is 0.252; the positive uniqueness of user2 is 0.0.

[0216] Multi-channel interactive behavior sequence pattern<item7,item6> The positive uniqueness of user1 is 0.105;

[0217] The positive uniqueness of user2 is 0.0.

[0218] Multi-channel interactive behavior sequence pattern<item7,item6,item8> The positive uniqueness of user1 is 0.165; the positive uniqueness of user2 is 0.0.

[0219] The fourth step is to construct a sequence pattern knowledge graph for all users. Based on the multi-channel interaction behavior sequence patterns of length 2 in set Φ, a set of quintuples (entity 1, relation, interval attribute, transition probability, entity 2) is constructed to represent the sequence pattern knowledge graph for all users. This is based on the multi-channel interaction behavior sequence patterns.<item1,item5> For example, two multi-channel interaction behavior categories, item1 and item5, are represented as entity 1 and entity 2 of the quintuple, respectively; the relationship between the entities is a temporal relationship, represented as "preceded by"; the gap constraint in the relationship attribute is ≤20 seconds; the transition probability between entities is equal to the normalized support of this pattern for all users, 0.154. The quintuple constituted in this embodiment is as follows:

[0220] (item1, occurred first, ≤20 seconds, 0.154, item5)

[0221] (item1, occurred first, ≤20 seconds, 0.245, item6)

[0222] (item2, occurred first, ≤20 seconds, 0.221, item1)

[0223] (item2, occurred first, ≤20 seconds, 0.478, item5)

[0224] (item2, occurred first, ≤20 seconds, 0.103, item6)

[0225] (item4, occurs first, ≤20 seconds, 0.25, item1)

[0226] (item4, occurs first, ≤20 seconds, 0.25, item3)

[0227] (item5, occurred first, ≤20 seconds, 0.221, item1)

[0228] (item5, occurred first, ≤20 seconds, 0.117, item2)

[0229] (item5, occurred first, ≤20 seconds, 0.114, item6)

[0230] (item5, occurred first, ≤20 seconds, 0.217, item8)

[0231] (item6, occurred first, ≤20 seconds, 0.089, item5)

[0232] (item6, occurred first, ≤20 seconds, 0.125, item8)

[0233] (item7, occurred first, ≤20 seconds, 0.467, item5)

[0234] (item7, occurred first, ≤20 seconds, 0.125, item6)

[0235] Based on step three, the sequence pattern knowledge graph constructed in this embodiment for all users is as follows: Figure 3 As shown.

[0236] Step 5: Construct a sequence pattern knowledge graph for specific users. Based on the multi-channel interaction behavior sequence patterns with a length greater than 2 in set Φ, construct sequence pattern knowledge graphs for users user1 and user2 respectively. For user1, the constructed quintuple is as follows:

[0237] (item1, occurred first, ≤20 seconds, 0.414, item5)

[0238] (item1, occurred first, ≤20 seconds, 0.631, item6)

[0239] (item2, occurred first, ≤20 seconds, 0.579, item1)

[0240] (item2, occurred first, ≤20 seconds, 0.627, item5)

[0241] (item5, occurred first, ≤20 seconds, 0.581, item1)

[0242] (item5, occurred first, ≤20 seconds, 0.301, item2)

[0243] (item5, occurred first, ≤20 seconds, 0.316, item6)

[0244] (item5, occurred first, ≤20 seconds, 0.573, item8)

[0245] (item6, occurred first, ≤20 seconds, 0.179, item5)

[0246] (item6, occurred first, ≤20 seconds, 0.351, item8)

[0247] (item7, occurred first, ≤20 seconds, 0.433, item5)

[0248] (item7, occurred first, ≤20 seconds, 0.351, item6)

[0249] For user2, the quintuple is constructed as follows:

[0250] (item2, occurred first, ≤20 seconds, 0.551, item5)

[0251] (item2, occurred first, ≤20 seconds, 0.452, item6)

[0252] (item4, occurred first, ≤20 seconds, 0.682item1)

[0253] (item4, occurred first, ≤20 seconds, 0.682, item3)

[0254] (item7, occurred first, ≤20 seconds, 0.625, item5)

[0255] Correspondingly, the sequence pattern knowledge graph for user1 is as follows: Figure 4 As shown; the sequence pattern knowledge graph for user2 is as follows: Figure 5 As shown.

[0256] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims. < / item7> < / item6> < / item5> < / item4> < / item2> < / item1>

Claims

1. A method for constructing a knowledge graph of multi-channel interactive behavior sequence patterns, comprising the following steps: 1) Preprocess the multi-channel interaction logs to obtain a generalized representation of each log record in the multi-channel interaction logs; 2) Mining multi-channel interaction behavior sequence patterns with gap constraints from the multi-channel interaction logs; wherein, the method for obtaining the multi-channel interaction behavior sequence patterns is as follows: 21) Grouping the log records in the multi-channel interaction logs according to the user ID to obtain the interaction log set corresponding to each user ID; if the interval between two adjacent log records in the interaction log set exceeds a set threshold... Then, using these two log records as the dividing point, the log records in the interaction log set are divided into several non-overlapping segments; 22) Taking each segment as a unit, the multi-channel interaction behavior category to which each log record in the segment belongs is traversed to form a category candidate set, and a multi-channel interaction behavior sequence pattern of length 1 is generated, that is, it contains only 1 multi-channel interaction behavior category; then, for each generated multi-channel interaction behavior sequence pattern, any multi-channel interaction behavior category that has not yet appeared in the current multi-channel interaction behavior sequence pattern is selected from the category candidate set and linked to the end of the current multi-channel interaction behavior sequence pattern, until no new multi-channel interaction behavior sequence pattern can be generated or the maximum length of the current multi-channel interaction behavior sequence pattern reaches the preset parameter. ;23) For a given segment and multi-channel interactive behavior sequence patterns Using the derived function Calculate the subsequence from fragment s Discovering patterns And in the subsequence Multi-channel interaction behavior sequence patterns were discovered. The last multi-channel interactive behavior category The probability when; and hour, Indicates in fragment Discovery of multi-channel interaction behavior sequence patterns The probability of; where, Representing fragments The number of log records; For fragments The xth log record in; This represents three categories of multichannel interactive behaviors with a sequential order. 3) For multi-channel interactive behavior sequence patterns, calculate the standardized support and positive user uniqueness; 4) Using multi-channel interactive behavior categories as entities, construct quintuples based on the multi-channel interactive behavior sequence patterns with gap constraints. A knowledge graph of multi-channel interactive behavior sequence patterns for all users is constructed from the set of data. Entity 1 and Entity 2 are two entities in a multi-channel interactive behavior sequence pattern with gap constraints. The relationship is a temporal relationship, meaning entity 1 occurs before entity 2. The gap constraint of the interval attribute is... ; Calculate the first transition probability from entity 1 to entity 2 based on the standardized support. 5) Using multi-channel interactive behavior categories as entities, construct quintuples based on the multi-channel interactive behavior sequence patterns with gap constraints. A knowledge graph of multi-channel interactive behavior sequence patterns targeting a user is constructed from the set of entities. Entities 1 and 2 are two entities in a multi-channel interactive behavior sequence pattern with gap constraints. The relationship is a temporal relationship, meaning entity 1 occurs before entity 2. The gap constraint of the interval attribute is... The second transition probability from entity 1 to entity 2 is calculated based on the user's positive uniqueness.

2. The method according to claim 1, characterized in that, The method for obtaining the generalized representation of the log records is as follows: 11) Based on the interaction channels and actions in the log records, classify the log records in the multi-channel interaction log one by one to form several sets of multi-channel interaction behaviors; 12) Extract the semantic and temporal features of each log record to form the feature encoding of the log record; 13) Perform density clustering on the feature encoding of each log record in each set of multi-channel interaction behaviors to obtain several sub-category clusters, and calculate the probability of belonging between each log record and the sub-category of its sub-category cluster; 14) For each log record, append its subcategories and the probability of belonging to the log record to obtain a generalized representation of the log record.

3. The method according to claim 2, characterized in that, The method for forming the feature encoding of the log record is as follows: the interaction channel, interaction action, and operation object in the log record are extracted by the feature extractor and converted into word embedding vectors. Then, the log record is combined with the index of the log record in the multi-channel interaction log and the time difference between the log record and the previous log record to generate the fused feature vector of the log record, which is used as the feature encoding of the log record.

4. The method according to claim 1, characterized in that, The derivation function The calculation method is as follows: ;in, Represents log recording It belongs to the category of multi-channel interaction behavior. The probability of; Represents a multi-channel interactive behavior sequence pattern The prefix pattern, i.e. ; Representing fragments Satisfaction with log recording Interval time Minimum index of conditional log records, function Indicates in subsequence Prefix patterns were found in And in the subsequence of fragment s Prefix patterns were found in The last multi-channel interactive behavior category The probability; the derived function Boundary conditions include: a) When or hour, b) When multi-channel interactive behavior sequence pattern When empty, .

5. The method according to claim 1, characterized in that, The method for calculating the standardized support is as follows: , ;in, Represents a multi-channel interactive behavior sequence pattern Standardized support across all segments for all users; Represents a multi-channel interactive behavior sequence pattern In users Standardized support across all segments; Represents a multi-channel interactive behavior sequence pattern In the fragment The probability of it appearing in; This represents the total number of segments for all users. Indicates user The number of fragments, Multi-channel interactive behavior sequence mode The number of multi-channel interactive behavior categories.

6. The method according to claim 5, characterized in that, The method for calculating the positive uniqueness of a user is as follows: , , ;in, This represents a set of multi-channel interactive behavior sequence patterns mined from all users; Represents a set The number of multi-channel interactive behavior sequence patterns included; Represents a multi-channel interactive behavior sequence pattern For users The degree of personalization deviation; Representation pattern For users The personalized positive contribution; Indicates user For the pattern Positive user uniqueness.

7. The method according to claim 6, characterized in that, Five-tuple ( , first occurred in, , , The transition probability of ) ;in, Indicates the user's ID; Indicates the presence of subsequences A collection of multi-channel interactive behavior sequence patterns.

8. A system for constructing a knowledge graph of multi-channel interactive behavior sequence patterns, characterized in that, It includes a preprocessing module, a pattern mining module, a feature calculation module, and a knowledge graph generation module; The preprocessing module is used to preprocess the multi-channel interactive logs to obtain a generalized representation of each log record in the multi-channel interactive logs. The pattern mining module is used to mine multi-channel interaction behavior sequence patterns with gap constraints from the multi-channel interaction logs; wherein, the method for obtaining the multi-channel interaction behavior sequence patterns is as follows: 21) Grouping the log records in the multi-channel interaction logs according to the user ID to obtain an interaction log set corresponding to each user ID; if the interval between two adjacent log records in the interaction log set exceeds a set threshold... Then, using these two log records as the dividing point, the log records in the interaction log set are divided into several non-overlapping segments; 22) Taking each segment as a unit, the multi-channel interaction behavior category to which each log record in the segment belongs is traversed to form a category candidate set, and a multi-channel interaction behavior sequence pattern of length 1 is generated, that is, it contains only 1 multi-channel interaction behavior category; then, for each generated multi-channel interaction behavior sequence pattern, any multi-channel interaction behavior category that has not yet appeared in the current multi-channel interaction behavior sequence pattern is selected from the category candidate set and linked to the end of the current multi-channel interaction behavior sequence pattern, until no new multi-channel interaction behavior sequence pattern can be generated or the maximum length of the current multi-channel interaction behavior sequence pattern reaches the preset parameter. ;23) For a given segment and multi-channel interactive behavior sequence patterns Using the derived function Calculate the subsequence from fragment s Discovering patterns And in the subsequence Multi-channel interaction behavior sequence patterns were discovered. The last multi-channel interactive behavior category The probability when; and hour, Indicates in fragment Discovery of multi-channel interaction behavior sequence patterns The probability of; where, Representing fragments The number of records in the log; For fragments The xth log record in; This represents three categories of multichannel interactive behaviors with a sequential order. The feature calculation module is used to calculate the standardization support and user positive uniqueness for multi-channel interaction behavior sequence patterns. The knowledge graph generation module is used to construct quintuples based on the multi-channel interaction behavior sequence patterns with gap constraints, using multi-channel interaction behavior categories as entities. The dataset is used to construct a multi-channel interaction behavior sequence pattern knowledge graph for all users and a multi-channel interaction behavior sequence pattern knowledge graph for target users. Entity 1 and Entity 2 are two entities in a multi-channel interaction behavior sequence pattern with gap constraints. The relationship is a temporal relationship, meaning entity 1 occurs before entity 2. The gap constraint of the interval attribute is... ; Calculate the first transition probability from entity 1 to entity 2 based on the standardized support; Calculate the second transition probability from entity 1 to entity 2 based on the user's positive uniqueness.

Images