CN112925899A

CN112925899A - Ranking model establishing method, case clue recommending device and medium

Info

Publication number: CN112925899A
Application number: CN202110181900.0A
Authority: CN
Inventors: 郑志骏; 张彦斌
Original assignee: Chongqing Zhongke Yuncong Technology Co ltd
Current assignee: Chongqing Zhongke Yuncong Technology Co ltd
Priority date: 2021-02-09
Filing date: 2021-02-09
Publication date: 2021-06-08
Anticipated expiration: 2041-02-09
Also published as: CN112925899B

Abstract

The invention relates to the technical field of clue recommendation, in particular to a ranking model establishing method, a case clue recommending device and a case clue recommending medium. The invention aims to solve the problems of single logic and poor accuracy of the conventional case clue recommendation system. To this end, the model building method of the present invention includes: obtaining multimodal data relating to a plurality of management objects; determining relationship data between the plurality of management objects based on the multimodal data; generating tags associated with the plurality of management objects based on the multimodal data and the relational data; generating a joint representation based on the relationship data and the labels; based on the joint representation, a ranking model is established for ranking the plurality of management objects. By acquiring the multi-mode data of the management object for learning and modeling, the multi-mode mass data of the business side are fully utilized, the management object is more three-dimensional in portrayal, and the recommendation accuracy and coverage rate can be greatly improved.

Description

Ranking model establishing method, case clue recommending device and medium

Technical Field

The invention relates to the technical field of clue recommendation, in particular to a ranking model establishing method, a case clue recommending device and a case clue recommending medium.

Background

A case clue recommendation system is a system for automatically recommending management objects possibly related to cases for criminal investigation personnel based on the mining of mass data when case analysis is carried out on criminal investigation scenes. The management objects include people, things, places, objects, organizations and the like.

With the development of big data technology, the current case clue recommendation system basically solves the real-time recommendation problem of mass data. However, existing case recommendation system solutions are typically based on single modality data, such as: and (3) constructing an entity relation map among people, affairs and places through case file data, and recommending an input case generation semantic tree model. The data resources available by a service party (namely a criminal investigation party) are multi-modal, and the existing case clue recommendation system does not fully utilize the abundant data of the service party, so that the accuracy is often poor after the system is on line, or the logic of a recommendation clue is too single, and the service value is not high.

Accordingly, there is a need in the art for a new solution to the above problems.

Disclosure of Invention

In order to solve at least one of the above problems in the prior art, that is, to solve the problems of single logic and poor accuracy existing in the conventional case clue recommendation system, a first aspect of the present application provides a ranking model establishing method for case clue recommendation, where the method includes:

obtaining multimodal data relating to a plurality of management objects;

determining relationship data between a plurality of the management objects based on the multimodal data;

generating tags associated with a plurality of the management objects based on the multimodal data and the relationship data;

generating a joint representation based on the relationship data and the labels;

based on the joint representation, a ranking model is established for ranking the plurality of management objects.

In a preferred embodiment of the above method for building a ranking model for case clue recommendation, after the step of "determining relationship data between a plurality of management objects", the method further includes:

and storing the relation data in a knowledge graph.

In a preferred embodiment of the above ranking model building method for case clue recommendation, the multi-modal data includes: personnel static information data, dynamic track data and case file data.

In a preferred embodiment of the above method for building a ranking model for case clue recommendation, the relationship data includes static relationship data, dynamic relationship data, and case relationship data, and the step of "determining relationship data between a plurality of management objects based on the multi-modal data" further includes:

extracting static relation data among management objects from the personnel static information data;

performing track fusion based on the dynamic track data to obtain fused dynamic track data, and extracting dynamic relation data among management objects from the fused dynamic track data;

and extracting case relation data among the management objects from the case file data.

In a preferred embodiment of the above ranking model building method for case clue recommendation, the step of "generating tags related to the plurality of management objects according to the multi-modal data and the relationship data" further includes:

and respectively generating a personal label of each management object and a relationship label between a plurality of management objects based on the personnel static information data, the fused dynamic trajectory data and the knowledge graph.

In a preferred embodiment of the above method for building a ranking model for case clue recommendation, the step of "generating a joint representation based on the relationship data and the tags" further includes:

generating a sequence structure vector with a set length in a random walk mode for each management object in the knowledge graph, and recording a label of each passing object which is randomly walked, wherein each value in the sequence structure vector is a value of a relationship label of the management object and each corresponding passing object;

determining threat degree vectors by utilizing a preset classification model based on the label of each passing object, wherein each value in the threat degree vectors is the threat degree of each passing object;

performing point multiplication on the sequence structure vector and the threat degree vector, and inputting a result after the point multiplication into a dimension reduction algorithm to obtain a knowledge graph identifier after dimension reduction, so as to generate a joint representation of the knowledge graph and the label;

wherein the classification model is used for representing the corresponding relation between the label of the passing object and the threat degree.

In a preferred embodiment of the above method for building a ranking model for case lead recommendation, the step of "building a ranking model based on the joint representation" further includes:

and inputting the joint representation into a deep learning classification algorithm for training to obtain the sequencing model.

In a second aspect of the present application, there is also provided an ordering model building apparatus for case thread recommendation, the apparatus including:

a data acquisition module for acquiring multimodal data relating to a plurality of management objects;

a relationship determination module for determining relationship data between a plurality of the management objects based on the multimodal data;

a tag generation module for generating tags associated with the plurality of management objects based on the multimodal data and the relationship data;

a joint representation module for generating a joint representation based on the relationship data and the labels;

a model building module for building a ranking model for ranking the plurality of management objects based on the joint representation.

In a preferred embodiment of the above ranking model building apparatus for case clue recommendation, the relationship determining module is further configured to store the relationship data in a knowledge graph after determining the relationship data among the plurality of management objects.

In a preferred embodiment of the above ranking model building apparatus for case clue recommendation, the multi-modal data includes: personnel static information data, dynamic track data and case file data.

In a preferred embodiment of the above ranking model building apparatus for case clue recommendation, the relationship data includes static relationship data, dynamic relationship data, and case relationship data, and the relationship determining module determines the relationship data between the plurality of management objects based on the multi-modal data by:

In a preferred embodiment of the above ranking model building apparatus for case clue recommendation, the tag generating module generates tags related to the plurality of management objects according to the multi-modal data and the relationship data as follows:

In a preferred embodiment of the above ranking model building apparatus for case clue recommendation, the joint representation module generates a joint representation based on the relationship data and the tags as follows:

generating a sequence structure vector with a set length in a random walk mode for each management object in the knowledge graph, and recording a label of each passing object which passes by random walk, wherein each value in the sequence structure vector is a value of a relationship label of the management object and each corresponding passing object;

In a preferred embodiment of the above ranking model establishing apparatus for case clue recommendation, the model establishing module establishes a ranking model based on the joint representation in the following manner:

In a third aspect of the present application, there is further provided a case clue recommendation method, including:

extracting key factors aiming at the input case files and determining a plurality of management objects meeting the conditions based on the key factors;

ranking the plurality of management objects based on a ranking model;

screening the plurality of sorted management objects based on a business strategy;

and outputting the screened management object.

In a fourth aspect of the present application, there is also provided a case thread recommending apparatus, including:

the recalling module is used for extracting key factors of the input case files and determining a plurality of management objects meeting conditions based on the key factors;

a ranking module to rank the plurality of management objects based on a ranking model;

the screening module is used for screening the sorted management objects based on the business strategy;

and the output module is used for outputting the screened management objects.

In a fifth aspect of the present application, there is further provided a processing apparatus, comprising a processor and a memory, wherein the memory is adapted to store a plurality of program codes, and the program codes are adapted to be loaded and executed by the processor to perform the arrangement model establishing method for leads recommendation of cases of any of the above preferred technical solutions or the lead recommendation method of the above preferred technical solutions.

In a sixth aspect of the present application, there is further provided a computer-readable storage medium, which stores a plurality of program codes, wherein the program codes are adapted to be loaded and executed by a processor to execute the arrangement model establishing method for leads recommendation of cases according to any of the above-mentioned preferred embodiments or the lead recommendation method in the above-mentioned preferred embodiments.

One or more technical schemes of the invention at least have one or more of the following beneficial effects:

by acquiring multi-mode data of the management object for learning and modeling, mass data of the business side in the multi-mode are fully utilized, and the management object is depicted more three-dimensionally. When the established model is applied to the case clue recommendation system, compared with a recommendation system established by adopting a single mode, the recommendation system can give a more three-dimensional recommendation result by using more dimensional data, so that the accuracy and the coverage rate of recommendation can be greatly improved.

Furthermore, by adopting static personnel information data, dynamic track data and case file data as multi-mode data, the management object can be depicted from multiple dimensions, and the management object can be depicted more three-dimensionally. For example, basic attributes of the management object can be described through the personnel static information data, and the real behavior pattern of the management object can be described through the dynamic track data.

Drawings

The ranking model establishing method, case clue recommending method, device and medium of the present invention will be described with reference to the accompanying drawings. In the drawings:

FIG. 1 is a flowchart of a ranking model building method for case thread referral according to the present invention;

FIG. 2 is a data flow diagram of the ranking model building process for case clue recommendation of the present invention;

FIG. 3 is a block diagram illustrating an embodiment of an apparatus for building a ranking model for case lead referral according to the present invention;

FIG. 4 is a flowchart illustrating a case thread referral method according to the present invention;

FIG. 5 is a block diagram of an embodiment of a case thread recommender according to the present invention.

List of reference numerals

11. A data acquisition module; 12. a relationship determination module; 13. a tag generation module; 14. a joint representation module; 15. a model building module;

21. a recall module; 22. a sorting module; 23. a screening module; 24. and an output module.

Detailed Description

Preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only for explaining the technical principle of the present invention, and are not intended to limit the scope of the present invention.

In the description of the present invention, a "module" or "processor" may include hardware, software, or a combination of both. A module may comprise hardware circuitry, various suitable sensors, communication ports, memory, may comprise software components such as program code, or may be a combination of software and hardware. The processor may be a central processing unit, microprocessor, image processor, digital signal processor, or any other suitable processor. The processor has data and/or signal processing functionality. The processor may be implemented in software, hardware, or a combination thereof. Non-transitory computer readable storage media include any suitable medium that can store program code, such as magnetic disks, hard disks, optical disks, flash memory, read-only memory, random-access memory, and the like. The term "a and/or B" denotes all possible combinations of a and B, such as a alone, B alone or a and B. The term "at least one A or B" or "at least one of A and B" means similar to "A and/or B" and may include only A, only B, or both A and B. The singular forms "a", "an" and "the" may include the plural forms as well.

First, referring to fig. 1 and fig. 2, a method for building an arrangement model for case clue recommendation according to the present invention will be described. FIG. 1 is a flowchart illustrating a method for building a ranking model for case clue recommendation according to the present invention; FIG. 2 is a data flow diagram of the ranking model building process for case clue recommendation of the present invention.

As shown in fig. 1 and fig. 2, in order to solve the problems of single logic and poor accuracy of the conventional case clue recommendation system, the method for establishing a ranking model for case clue recommendation mainly includes:

s101, multi-modal data related to a plurality of management objects are acquired.

In one embodiment, the multimodal data includes: personnel static information data, dynamic track data and case file data. The personal static information data refers to basic information of a person, such as name, age, sex, place of residence, native place, and the like of a management object, and can be obtained from a system of a business party, such as a household system of a public security department. The dynamic trajectory data refers to data capable of reflecting the spatiotemporal trajectory of the management object, such as a snapshot record of the management object (e.g., a face snapshot record, vehicle snapshot data, etc.), a trajectory of the communication terminal, a transaction flow, a call record, a store record, a train/airplane trip record, and the like.

The snapshot record of the management object can be determined by data snapshot of cameras and video cameras installed at intersections, doorways, elevators and other positions. The communication terminal may be a mobile phone or other device with communication function, and has a unique identifier, such as an IMSI (international mobile subscriber identity), a MAC address, and the communication terminal track may be determined by the sensed MAC address and IMSI. The position of the mobile phone can be determined through the MAC address, and therefore the track of the mobile phone is obtained. The MAC address may be obtained by a WIFI probe. The stream of transactions may be called by bank, by record may be called by operator, by hotel management system, by train/plane travel record may be called by transportation department, etc.

Case file data refers to data that records the incident process, which may also be obtained from the business party's system, such as from file archives in the police department.

Of course, the present application is not limited to the specific composition of the three types of data in the multimodal data, and those skilled in the art can adjust the specific range of the three types of data based on the specific application scenario, such as deleting the data from the current data or replacing the data with other relevant data.

S103, determining relation data among the management objects based on the multi-modal data, and storing the relation data in a knowledge graph.

In one possible implementation, still taking as an example that the multi-modal data includes personnel static information data, dynamic track technical bureau, and case file data, the relationship data includes static relationship data, dynamic relationship data, and case relationship data, the step S103 may further include:

s1031, aiming at the static information data of the personnel, extracting static relation data established by the management object from the static information data; for example, from the names of persons and the information of residence places, "zhang san → xiang → li si" relationship data is extracted. The extraction method is simple and will not be described herein.

S1032, extracting case relation data among the management objects from the case file data. For example, by existing natural language processing algorithms, including but not limited to: the language structure of the case description in the case file is analyzed by an LTAG algorithm, a CCG algorithm, a RST tree structure analysis algorithm, a reference resolution algorithm and the like, and entities and relations are extracted from the language structure, for example, in the case description that '10/month, 1/day and 15/year 2020, Zhang III hits Li IV in Zhang river artificial intelligence island', the relation of 'Zhang III → traffic onset → Li IV' is extracted, and the occurrence time of the case (10/month, 1/day and 15/year 2020) and the place (Zhang river artificial intelligence island) are taken as the attributes of the relations.

S1033, aiming at the dynamic track data, the tracks are fused to obtain fused dynamic track data, and then dynamic relation data among the management objects are extracted from the fused dynamic track data. Specifically, firstly, through existing algorithms such as a face detection algorithm, a face feature extraction algorithm, a feature clustering algorithm and the like, snapshot-based associated trajectory data of the portrait and the vehicle is obtained from the portrait and the vehicle snapshot data. Taking a portrait as an example, the primary data structure may include: face id, snapshot time, snapshot longitude, snapshot latitude, snapshot height, snapshot device address, and the like. And secondly, fusing the associated track data based on the snapshot and other track data (such as transaction flow, call records, store records and train and plane trip records) in the dynamic track data through an image code joint detection algorithm to obtain a fused dynamic track data track. The steps of the graph code joint detection algorithm are briefly introduced as follows:

and S1033-1, supplementing historical space-time tracks of the management objects by using a sequence supplementing method to form the supplemented historical space-time tracks, wherein the sequence supplementing method comprises but is not limited to generation of a countermeasure network, tensor decomposition, linear interpolation, LSTM and the like.

S1033-2, converting the supplemented historical space-time trajectory into word vectors by using a word vector model. The word vector model includes, but is not limited to, word2vec, VAE, WAE, Auto-Encoders, SEQ-GAN, etc.

S1033-3, clustering the word vectors by using the topic clustering model. The topic clustering model includes, but is not limited to LDA, DBSCAN, etc.

S1033-4, constructing a subgraph for the management objects corresponding to the word vectors under each cluster and historical space-time trajectories of the management objects. The way of constructing the subgraph is as follows: and acquiring pairwise combinations of the management objects, and calculating the space-time similarity of historical space-time trajectories between the pairwise combinations as the weight of the edges. The calculation method of the space-time similarity includes, but is not limited to, euclidean distance, LCS distance, and the like.

S1033-5, rejecting the sub-graph from the isolated point, wherein the isolated point is defined in such a way that the weight of all edges led out from the point is less than a set threshold value.

And S1033-6, for each subgraph, selecting a representative node according to the centrality of the node.

S1033-7, calculating space-time similarity of pairwise combination of the representative nodes of each subgraph, and fusing the subgraphs of which the space-time similarity exceeds a set threshold.

The process of fusing the two spatiotemporal trajectories is to use the same mark to identify the two spatiotemporal trajectories, for example, the mark 1 is used to identify a plurality of spatiotemporal trajectories, for example, the mark 1 is used to identify the spatiotemporal trajectory a, the spatiotemporal trajectory B, and the spatiotemporal trajectory C, so that the spatiotemporal trajectory a, the spatiotemporal trajectory B, and the spatiotemporal trajectory C can be considered to be the same management object trajectory.

While the above-mentioned graph code joint detection algorithm is briefly introduced, it is understood by those skilled in the art that in some other embodiments, some of the above-mentioned steps may be replaced or omitted, or those skilled in the art may also perform fusion by using other graph code joint detection algorithms, for example, trajectory fusion may be performed by using a fusion method shown in publication numbers CN111797295A or CN 111695511A.

And extracting the dynamic relation relationship between the management objects in the fused dynamic track data. The extraction method of the dynamic relationship data can be divided into the following three types: a. the relation of Zhangsan → possession → mobile phone is directly extracted by the image code joint detection algorithm; b. the relationship discovery system automatically excavates the relationship based on the equivalence relationship, such as: extracting the relation of 'Zhangsan → Tong Liquan'; c. based on the relationship found by the business policy and statistics, for example, the business policy is: the call records of a certain managed object in the last week are counted, wherein the call records have the call times more than 20, and therefore the relation of Zhang three → frequent call → Li four is found in the call records.

S1034, keeping the relationship data between the management objects generated in the steps S1031-S1033 in a knowledge graph.

Compared with the conventional image code joint detection algorithm in the prior art, the image code joint detection algorithm can realize the association of various management objects by using the knowledge graph and the semantic topic clustering technology, for example, the association of various objects of human face-human body-vehicle-mobile phone-financial account numbers is realized, so that the aims of describing the management objects more three-dimensionally and recommending more accurately are realized.

And S105, generating labels related to a plurality of management objects according to the multi-modal data and the knowledge graph.

In one possible implementation, the label factory system may be utilized to calculate the personal label of the management object and the relationship label between the management object based on the business strategy from the personnel static information data, the fused dynamic trajectory data and the knowledge graph. For example, for a single management object, tags such as "elderly person", "high school calendar" and the like are generated from age and school calendar data in the personal static information data; for another example, for a single management object, based on a business strategy that "the number of occurrences at night is higher than 70% in the recent occurrence number of a certain object", a label such as "daytime and nighttime" is formed from the fused dynamic trajectory data; for another example, labels of "frequently contacting with a person involved in a virus", "first degree relationship of a key person (i.e., direct contact with the key person)", "second degree relationship of the key person (i.e., indirect contact with the key person)" are formed from the knowledge graph for the relationship between a certain management object and other management objects; for another example, relationship labels such as "affinity", "light relationship", and "attention-needed relationship" are generated for relationships between different management objects, and different weights are given to the relationships between the management objects according to the relationship labels, where the weight of "attention-needed relationship" is 1, the weight of "affinity" is 0.8, and the weight of "light relationship" is 0.3. Tag factory systems are common in the art, and the principle and process of generating tags are not described herein.

And S107, generating a joint representation based on the knowledge graph and the label.

Because the final sequencing model needs to input structured data, the labels in the obtained data belong to the structured data and can be directly input into the sequencing model, and the knowledge graph belongs to unstructured data and needs to be converted into the structured data. Thus, in a possible implementation, step S107 may further include: and acquiring the joint representation of the knowledge graph between the management objects and the relationship labels between the management objects through a graph embedding algorithm based on the relationship labels between the knowledge graph and the management objects. The graph embedding algorithm specifically comprises the following steps:

s1071, a sequence structure vector of a specified length is generated by random walk for a certain node in the knowledge graph, that is, a certain management object, and the label (hereinafter, referred to as "id") of each passing management object (that is, passing object) is sequentially recorded. Each value in the sequence structure vector is a label value of each edge that the node randomly walks (i.e., a relationship label value of the management object and each corresponding passing object). The value of the relationship label is the weight, and if the "relationship needing attention" is marked as 1, the "close relationship" is marked as 0.8, and the "light relationship" is marked as 0.3 according to the weight. If the random walk is 5 times, the generated sequence structure vector may be, for example: (1,0.3,0.8,0.3,0.3), and the id of each passing object is recorded as: (id1, id2, id3, id2, id 4).

S1072, based on the label carried by each passing object, determining the threat degree of each passing object by using a preset classification model to obtain a threat degree vector. In a preferred embodiment, the classification model may be an intelligent score card model, and the specific establishment method of the intelligent score card model may refer to a patent application with a publication number of CN111563810A or CN111898675A, where the present application is different from the above two applications in that the historical volume data of the criminal investigator is used to perform model establishment and training, and the trained classification model is used to determine the threat degree of each passing object. The threat level is a quantitative index, such as a value between 0 and 1, which can be used to refer to the suspicious level of passing through the object. For example, the threat degree vector composed of the threat degrees of the randomly wandering passing objects is: (0,0.3,0.7,0.3,1)

S1073, carrying out point multiplication on the sequence structure vector and the threat degree vector, and inputting the result after the point multiplication into a dimension reduction algorithm to obtain the knowledge graph identification after dimension reduction. For example, the two vectors are point-multiplied by (1,0.3,0.8,0.3,0.3) · (0,0.3,0.7,0.3,1) · (0,0.09,0.56,0.09,0.3), and the result of the point multiplication is input into the PCA automatic dimension reduction algorithm to obtain the reduced-dimension knowledge graph identifier.

Of course, in addition to the graph embedding algorithm described above, one skilled in the art may substitute the above algorithm with other graph embedding algorithms, or substitute certain steps of the above graph embedding algorithm, to adjust the interpretability of the ranking model. For example, the Node2Vec algorithm replaces the graph embedding algorithm of the present application as a whole, or the classification model is replaced by using a logistic regression model and a DNN model, or the PCA dimension reduction algorithm is replaced by using an algorithm such as an automatic coding machine and VAE. However, compared with the algorithm provided by the present application, the use of the above alternative algorithm may bring about the defects of poor interpretability, incapability of combining with an actual security scene, and the like.

Criminal investigation often requires that the results of a recommendation for a lead be interpretable and traceable. In the implementation process of the graph embedding algorithm, the algorithm is configurable and the characteristics can be processed automatically in the model building process by combining the random walk, the classification model and the dimension reduction algorithm based on the data characteristics and the service characteristics of the criminal investigation industry. By further adopting the technical scheme of combining the random walk model, the intelligent scoring card model and the PCA dimension reduction algorithm, the random walk model, the intelligent scoring card model and the PCA dimension reduction algorithm have interpretability in the application process, so that the whole graph embedding algorithm has interpretability, and the development and training processes of the algorithm are highly combined with criminal investigation data, so that the algorithm model is closer to the actual use scene of criminal investigation while high interpretability is kept, and the accuracy equivalent to or higher than that of the existing model is realized in the criminal investigation field.

And S109, establishing a sequencing model based on the joint representation.

In a possible implementation, the step S109 further includes: and inputting the joint representation into a deep learning classification algorithm for training to obtain a sequencing model for sequencing a plurality of management objects. The process of inputting the joint representation into the deep learning classification algorithm for training to obtain the ranking model is conventional, and is not described herein again.

Of course, those skilled in the art can substitute the above modeling algorithm to obtain a ranking model more suitable for a specific application scenario.

In summary, the method for establishing the ranking model for case clue recommendation changes data of different modes into relational data and labels through operations such as track fusion, relational calculation, label calculation and the like, respectively stores the relational data and the labels in the knowledge graph and the structured database, obtains joint representation of the knowledge graph and the labels through graph embedding algorithm learning, and establishes the ranking model based on the joint representation. The interpretability and accuracy of the ranking model can be adjusted by changing the selection of the graph embedding algorithm and the modeling algorithm in the process.

When the ranking model is applied to clue recommendation, compared with a single-modal recommendation system, data with more dimensions can be utilized, and a recommendation result is more three-dimensional. Such as: the personnel static information data can depict the basic attribute of the management object, and the dynamic track data can depict the real behavior mode of the management object, so that the recommendation accuracy and the recommendation coverage rate can be improved. Through a mode of combining multi-mode learning and a service strategy of a criminal investigation party, a perception cognitive decision is communicated, the recommendation is more diversified, and the recommendation effect is better.

It should be noted that although the foregoing embodiments describe each step as being in the foregoing sequence, those skilled in the art will understand that, in order to achieve the effect of the present embodiments, different steps need not be executed in such sequence, and they may be executed simultaneously (in parallel) or in reverse sequence, and these simple changes are all within the scope of the present invention. For example, steps S1031 to S1033 may be performed simultaneously, or may be performed in any order.

In an alternative embodiment, although the above embodiment is described with a plurality of relationship data determined being stored in the knowledge graph, this embodiment is not invariable, and the person skilled in the art can select whether to store the relationship data in the knowledge graph based on a specific application scenario. For example, when the amount of data is not large, the identified relationship data may not be stored in the knowledge graph, but may directly participate in the calculation of the tag and the generation of the joint representation in the form of the relationship table. Of course, the abundance of relational data is the characteristic that criminal investigation industry data is different from other industry data, and if a knowledge map is not used, the calculation speed of the model can be influenced.

Referring to fig. 3, a ranking model establishing apparatus for case clue recommendation according to the present application will be described. Fig. 3 is a block diagram illustrating an embodiment of an apparatus for building a ranking model for case clue recommendation according to the present invention.

As shown in fig. 3, the ranking model establishing apparatus for case clue recommendation of the present application mainly includes: a data acquisition module 11 for acquiring multimodal data relating to a plurality of management objects; a relationship determination module 12 for determining relationship data between the plurality of management objects based on the multimodal data and saving the relationship data in a knowledge graph; a tag generation module 13 for generating tags related to the plurality of management objects based on the multimodal data and the knowledge graph; a joint representation module 14 for generating a joint representation based on the knowledge-graph and the labels; a model building module 15 for building a ranking model based on the joint representation. In one embodiment, the detailed implementation function can be described in reference to steps S101 to S109.

In one embodiment, the multimodal data includes: personnel static information data, dynamic track data and case file data.

In one embodiment, the relationship data includes static relationship data, dynamic relationship data, and case relationship data, and the relationship determination module 12 determines the relationship data between the plurality of management objects based on the multimodal data by: extracting static relation data among management objects from the personnel static information data; based on the dynamic track data, carrying out track fusion to obtain fused dynamic track data, and extracting dynamic relation data among the management objects from the fused dynamic track data; case relation data among the management objects is extracted from the case volume data. The description of the specific implementation function may refer to the description of step S103.

In one embodiment, tag generation module 13 generates tags associated with a plurality of management objects based on the multimodal data and the knowledge graph by: and respectively generating a personal label of each management object and a relationship label among a plurality of management objects based on the personnel static information data, the fused dynamic track data and the knowledge graph. The description of the specific implementation function can be found in step S105.

In one embodiment, the joint representation module 14 generates the joint representation based on the knowledge-graph and the tags by: generating a sequence structure vector with a set length in a random walk mode aiming at each management object in the knowledge graph, and recording the label of each passing object passing through the random walk mode, wherein each value in the sequence structure vector is the value of a relationship label of the management object and each corresponding passing object; determining a threat degree vector by utilizing a preset classification model based on the label of each passing object, wherein each value in the threat degree vector is the threat degree of each passing object; performing point multiplication on the sequence structure vector and the threat degree vector, and inputting a result after the point multiplication into a dimension reduction algorithm to obtain a knowledge graph identifier after dimension reduction, so as to generate a joint representation of the knowledge graph and the label; the classification model is used for representing the corresponding relation between the label of the passing object and the threat degree. The description of the specific implementation function may refer to the description of step S107 above.

In one embodiment, the model building module 15 builds the ranking model based on the joint representation by: and inputting the joint representation into a deep learning classification algorithm for training to obtain a sequencing model. The description of the specific implementation function may refer to the description of step S109.

The above described ranking model building apparatus for case clue recommendation is used to execute the above described embodiment of the ranking model building method for case clue recommendation, and the technical principles, solved technical problems and generated technical effects of the two are similar, and it can be clearly understood by those skilled in the art that for convenience and brevity of description, the specific working process and related description of the ranking model building apparatus for case clue recommendation may refer to the content described in the embodiment of the ranking model building method for case clue recommendation, and will not be described herein again.

Referring to fig. 4, a case clue recommendation method according to the present application will be described. FIG. 4 is a flowchart illustrating a case clue recommending method according to the present invention.

As shown in fig. 4, the case clue recommendation method of the present application mainly includes the following steps:

s201, aiming at the input case files, extracting key factors and determining a plurality of management objects meeting conditions based on the key factors.

In one possible implementation mode, for an input case file, key factors are extracted through a syntax tree model, and management objects which possibly meet conditions are screened for the key factors. Among these, key factors include, but are not limited to: case information such as identity certification, alarm receiving record, time and place in the record and the like of the suspect.

S203, sorting the plurality of management objects based on the sorting model.

In a possible implementation manner, the management objects which are preliminarily screened are input into the ranking model obtained by the ranking model establishing method for case clue recommendation, and the conformity degree of each management object is output through the ranking model and ranked according to the conformity degree. Of course, before using the ranking model, the ranking model needs to be trained. The training mode may be to train the model by using the case data already filed as the input and output of the model.

S205, screening the sorted management objects based on the business strategy.

In a possible implementation manner, after the possible management objects are sorted, not all the management objects are needed by the service side, and at this time, the service policy provided by the service side needs to be used to further filter the preliminarily screened management objects so as to narrow the scope of the management objects. The service policy is the screening condition provided by the service party, such as screening the management objects that are in daytime and nighttime, or screening the management objects that frequently communicate with each other.

And S207, outputting the screened management objects.

In a possible implementation manner, after the management objects are further screened, the screened management object list may be output in a form of conformity sorting, so as to implement the recommendation of case clues.

According to the case clue recommendation method, by adopting the sorting model, compared with a single-mode recommendation system, data with more dimensions can be utilized, and the recommendation result is more three-dimensional, so that the recommendation accuracy and coverage rate can be improved. By screening the sorted management objects based on the service strategy, multi-mode data and the service strategy of a criminal investigation party can be combined, perception and cognition decisions are communicated, the provided recommendation is more diversified, and the recommendation effect is better.

Referring to fig. 5, a case clue recommending apparatus according to the present application will be described. Fig. 5 is a block diagram illustrating an embodiment of a case thread recommending apparatus according to the present invention.

As shown in fig. 5, the case clue recommendation device of the present application mainly includes: a recall module 21 for extracting key factors from the inputted case files and determining a plurality of management objects meeting the conditions based on the key factors; a ranking module 22 for ranking the plurality of management objects based on a ranking model; a screening module 23, configured to screen the sorted management objects based on the service policy; and an output module 24, configured to output the screened management object. In one embodiment, the detailed description of the implementation function may refer to steps S201 to S205.

The technical principles, the technical problems solved and the technical effects produced by the above case clue recommending device for implementing the above case clue recommending method embodiments are similar, and it is clear to those skilled in the art that for convenience and simplicity of description, the specific working process and related description of the case clue recommending device may refer to the contents described in the case clue recommending method embodiments, and the details are not repeated here.

It will be understood by those skilled in the art that all or part of the flow of the method according to the above-described embodiment may be implemented by a computer program, which is stored in a computer-readable storage medium and used for instructing related hardware, and when the computer program is executed by a processor, the steps of the method embodiments may be implemented. Wherein the computer program comprises computer program code, which may be in the form of source code, object code, an executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying computer program code, media, U-disk, removable hard disk, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier wave signals, telecommunications signals, software distribution media, and the like. It should be noted that the computer readable medium may contain other components which may be suitably increased or decreased as required by legislation and patent practice in jurisdictions, for example, in some jurisdictions, in accordance with legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunications signals.

The various component embodiments of the invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that a microprocessor or Digital Signal Processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in a server, client, or the like, according to embodiments of the present invention. The present invention may also be embodied as an apparatus or device program (e.g., PC program and PC program product) for carrying out a portion or all of the methods described herein. Such a program implementing the invention may be stored on a PC readable medium or may be in the form of one or more signals. Such a signal may be downloaded from an internet website or provided on a carrier signal or in any other form.

Further, the invention also provides a computer readable storage medium. In one embodiment of the computer-readable storage medium according to the present invention, the computer-readable storage medium may be configured to store a program for implementing the arrangement model building method for case lead recommendation and/or the case lead recommendation method of the above-described method embodiment, and the program may be loaded and executed by a processor to implement the arrangement model building method for case lead recommendation and/or the case lead recommendation method. For convenience of explanation, only the parts related to the embodiments of the present invention are shown, and details of the specific techniques are not disclosed. The computer readable storage medium may be a storage device formed by including various electronic devices, and optionally, the computer readable storage medium is a non-transitory computer readable storage medium in the embodiment of the present invention.

Furthermore, the invention also provides a processing device. In an embodiment of the processing apparatus according to the present invention, the processing apparatus includes a processor and a memory, the memory may be configured to store a program for performing the arrangement model establishment method for case cue recommendation and/or the case cue recommendation method of the above-mentioned method embodiment, and the processor may be configured to execute a program in the memory, the program including but not limited to a program for performing the arrangement model establishment method for case cue recommendation and/or the case cue recommendation method of the above-mentioned method embodiment. For convenience of explanation, only the parts related to the embodiments of the present invention are shown, and details of the specific techniques are not disclosed. The processing device may be a device apparatus formed including various electronic apparatuses.

It should be understood that, since the configuration of each module is only for explaining the functional units of the apparatus of the present invention, the corresponding physical devices of the modules may be the processor itself, or a part of software, a part of hardware, or a part of a combination of software and hardware in the processor. Thus, the number of individual modules in the figures is merely illustrative.

Those skilled in the art will appreciate that the various modules in the apparatus may be adaptively split or combined. Such splitting or combining of specific modules does not cause the technical solutions to deviate from the principle of the present invention, and therefore, the technical solutions after splitting or combining will fall within the protection scope of the present invention.

So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.

Claims

1. A ranking model building method for case clue recommendation is characterized by comprising the following steps:

obtaining multimodal data relating to a plurality of management objects;

2. The method of claim 1, wherein after the step of determining relationship data between a plurality of management objects, the method further comprises:

and storing the relation data in a knowledge graph.

3. The method of claim 2, wherein the multi-modal data comprises: personnel static information data, dynamic track data and case file data.

4. The method according to claim 3, wherein the relationship data comprises static relationship data, dynamic relationship data and case relationship data, and the step of determining relationship data between the plurality of management objects based on the multi-modal data further comprises:

5. The method of claim 4, wherein the step of generating labels associated with the plurality of management objects according to the multi-modal data and the relationship data further comprises:

6. The method of claim 2, wherein the step of generating a joint representation based on the relationship data and the tags further comprises:

generating a sequence structure vector with a set length by aiming at each management object in the knowledge graph in a random walk mode, and recording a label of each passing object which is randomly walked, wherein each value in the sequence structure vector is a value of a relationship label of the management object and each corresponding passing object;

7. The method of claim 1, wherein the step of building a ranking model based on the joint representation further comprises:

8. An ordering model building apparatus for case clue recommendation, the apparatus comprising:

9. The apparatus of claim 8, wherein the relationship determining module is further configured to store the relationship data in a knowledge graph after determining the relationship data between a plurality of the management objects.

10. The apparatus of claim 9, wherein the multi-modal data comprises: personnel static information data, dynamic track data and case file data.

11. The apparatus of claim 10, wherein the relationship data comprises static relationship data, dynamic relationship data, and case relationship data, and the relationship determination module determines the relationship data between the plurality of management objects based on the multimodal data by:

12. The apparatus of claim 11, wherein the tag generation module generates tags related to the plurality of management objects according to the multi-modal data and the relationship data by:

13. The apparatus of claim 9, wherein the joint representation module generates a joint representation based on the relationship data and the tags by:

14. The apparatus of claim 8, wherein the model building module builds the ranking model based on the joint representation by:

15. A case clue recommendation method is characterized by comprising the following steps:

ranking the plurality of management objects based on a ranking model;

and outputting the screened management object.

16. A case thread recommender, the recommender comprising:

and the output module is used for outputting the screened management objects.

17. A processing apparatus comprising a processor and a memory, the memory being adapted to store a plurality of program codes, wherein the program codes are adapted to be loaded and run by the processor to perform the arrangement model establishing method for a clue recommendation of any one of claims 1 to 7 or the clue recommendation method of claim 15.

18. A computer-readable storage medium storing a plurality of program codes, wherein the program codes are adapted to be loaded and executed by a processor to perform the arrangement model establishing method for leads recommendation according to any one of claims 1 to 7 or the leads recommendation method according to claim 15.