CN111292118A - Investor portrait construction method and device based on deep learning - Google Patents
Investor portrait construction method and device based on deep learning Download PDFInfo
- Publication number
- CN111292118A CN111292118A CN202010024111.1A CN202010024111A CN111292118A CN 111292118 A CN111292118 A CN 111292118A CN 202010024111 A CN202010024111 A CN 202010024111A CN 111292118 A CN111292118 A CN 111292118A
- Authority
- CN
- China
- Prior art keywords
- investor
- investment
- behavior
- investors
- attributes
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000010276 construction Methods 0.000 title claims abstract description 29
- 238000013135 deep learning Methods 0.000 title claims abstract description 24
- 230000006399 behavior Effects 0.000 claims abstract description 143
- 238000000034 method Methods 0.000 claims abstract description 61
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 9
- 238000013528 artificial neural network Methods 0.000 claims description 17
- 230000003542 behavioural effect Effects 0.000 claims description 15
- 125000004122 cyclic group Chemical group 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 9
- 238000005295 random walk Methods 0.000 claims description 8
- 238000004590 computer program Methods 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 239000011159 matrix material Substances 0.000 claims 1
- 230000007547 defect Effects 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 17
- 230000006870 function Effects 0.000 description 11
- 238000011156 evaluation Methods 0.000 description 8
- 230000002123 temporal effect Effects 0.000 description 8
- 230000000306 recurrent effect Effects 0.000 description 7
- 230000010354 integration Effects 0.000 description 6
- 238000010801 machine learning Methods 0.000 description 5
- 238000013459 approach Methods 0.000 description 4
- 239000002131 composite material Substances 0.000 description 4
- 230000006872 improvement Effects 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 238000007637 random forest analysis Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000012417 linear regression Methods 0.000 description 3
- 230000007774 longterm Effects 0.000 description 3
- 238000007430 reference method Methods 0.000 description 3
- 238000009877 rendering Methods 0.000 description 3
- 239000000126 substance Substances 0.000 description 3
- 101100261000 Caenorhabditis elegans top-3 gene Proteins 0.000 description 2
- 230000009471 action Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000008520 organization Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 206010000117 Abnormal behaviour Diseases 0.000 description 1
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 241000283690 Bos taurus Species 0.000 description 1
- 101100510617 Caenorhabditis elegans sel-8 gene Proteins 0.000 description 1
- 101100261006 Salmonella typhi topB gene Proteins 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 238000007596 consolidation process Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000003334 potential effect Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 230000003362 replicative effect Effects 0.000 description 1
- 230000003997 social interaction Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000000638 stimulation Effects 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
- 238000012731 temporal analysis Methods 0.000 description 1
- 238000000700 time series analysis Methods 0.000 description 1
- 101150032437 top-3 gene Proteins 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/06—Asset management; Financial planning or analysis
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Accounting & Taxation (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Game Theory and Decision Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Operations Research (AREA)
- Technology Law (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
The invention provides a investor portrait construction method and device based on deep learning, wherein the method comprises the following steps: creating an investment behavior portrait by using historical investment data of an investor; constructing an investment graph by using historical investment data of investors, wherein graph nodes representing the investors are connected with corresponding investment behavior images; and simultaneously modeling the obtained investment behavior portrait and the relation influence and time dynamics in the investment map by utilizing a depth cycle map convolutional neural network, and outputting a predicted investment behavior portrait. The method of the invention gets rid of the limitation of the traditional investor portrayal method, overcomes the defects of the existing data-driven investment behavior characteristic portrayal method, obviously reduces the prediction error, and improves the portrayal prediction performance, thereby improving the effectiveness of the prediction investment behavior mode in practical application.
Description
Technical Field
The invention relates to the technical field of investor portrait construction, in particular to an investor portrait construction method and device based on deep learning.
Background
The investor portrait, namely the behaviour portrait of the investor in a future time period, is used for describing the characteristics of the investor and showing the future investment tendency of the investor, and plays a central role in the financial industry practice. Financial practitioners typically rely on investor imagery to determine appropriate investment projects and to develop effective long-term investment strategies for investors. The U.S. financial industry regulatory agency and the european union financial instrument market regulations emphasize the importance of investor figures and state that the suitability of investment recommendations depends on the particular investor figure. In addition, analyzing collective investor figures may help practitioners in market trend prediction and investment product innovation. For example, a new portfolio can be designed based on the investment profile of most investors, based on their common trends.
Traditional investor representation methods rely primarily on questionnaires to gather information such as age, gender, income, occupation, risk potential, and evasion to describe an individual's investment tendencies and behaviors. However, evidence from both academia and industry indicates that questionnaire-based investor representation methods are unreliable and have insufficient interpretability. In particular, factors derived from investor profile questionnaires, such as age, gender, and degree of risk avoidance, can account only for approximately 5% to 15% of the variation in the risk assets in an investor portfolio. Also, in practice, traditional investor portrayal is often insufficient to support appropriate recommendations and may result in significant monetary losses. It was reported that a 100 ten thousand dollar portfolio for a pair of mexico couples lost $ 519,089 due to improper recommendations of the morgans danli consultant. After the financial crisis of 2008 in 2007, there are thousands of similar complaints and directives on inappropriate investment sales in europe and the united states each year. Therefore, there is a need for new portrayal techniques that fully characterize investors and accurately predict their future investment intentions.
In recent years, with the rapid development of data mining and machine learning techniques, data-driven rendering methods are becoming the first solution for many rendering tasks. Data-driven portrayal does not resort to questionnaires, but relies primarily on large numbers of digital records collected in non-reactive and non-intrusive ways, such as web browsing, advertisement clicks, and online shopping logs. Then, machine learning and data analysis techniques are used to learn features and discovery patterns to create representations for various objects. For example, in electronic commerce, customer figures are created by learning online shopping activity patterns of customers to significantly improve customer services including personalized searches, recommendations, advertisements, and the like. With the widespread use of information systems in the financial industry, people have collected a great deal of information about individual investment activities. These data provide unprecedented opportunities for efficient discovery of individual investment trends and behavioral patterns. Although data-driven portrayal approaches to various objects are encouraging, they do not capture some of the important features of an individual's investment behavior, which would degrade the predictive performance of the individual's future investment behavior and ultimately limit the practical usefulness of portrayal. In particular, for comprehensive information extraction and integration of portraits, it is necessary to determine which information should be included in one portraits to fully characterize the investor's personalized behavior. Current portrayal methods either utilize investor attributes (e.g., demographics) or aggregate statistics from investors' past activities (e.g., risk preferences, past performance) into their portrayal. However, much detailed information about investor behavior is valuable and is rarely considered to be information describing investment characteristics, such as investment item attributes (e.g., goals), investor attributes (e.g., credit rating), and external market environment (e.g., cattle or bear). Moreover, a general representation method is lacked for the valuable information, and various information in any number of investment records of individuals can be uniformly integrated into the picture so as to comprehensively describe the behavior characteristics of the individuals.
In addition, in general, an individual's investment behavior is not only dependent on individual attributes, but is also influenced by heterogeneous social relationships. In particular, individual investment decisions may be influenced by the recommendations and behavior of other investors, e.g., past collaborators, and even a large group of unfamiliar investors. Social relationships, such as friendship, between investors and investors can also affect investors' financial decisions. Therefore, when profiling the individual investment behavior characteristics and predicting their future investment trends, the impact of Heterogeneous relationships (Heterogeneous relational impacts) needs to be considered.
Also, due to a variety of factors, the investment behavior characteristics of an individual typically change over time. The risk tolerance and avoidance of investors may vary with age and macro-economic conditions. Individual investment skills may also be improved by accumulating experience and acquiring knowledge. Furthermore, the investor's social circle may evolve as the investment behavior characteristics change. For example, improving investment skills (e.g., attending investment training courses) may bring a new social circle of investment to an individual; the new social circle may influence the individual's investment decisions. With this spirit, it is inadequate and inappropriate to maintain a fixed representation for the investor without regard to the time-varying pattern. I.e. the individual's investment behaviour also has a complex time dynamics.
However, integrating all of the above features into a representation is a challenging task. First, each investor has any number of investment behavior records with rich information, and it is not easy to integrate any number of information into a unified representation. Second, while heterogeneous relationship factors affect an individual's investment behavior, relationship information (such as friendship or interaction) is often not visible in the online investment marketplace. Although various features such as centrality, similarity, social interaction strength, or structural equivalence are carefully designed to estimate social influence in a social network, it is difficult for these features to capture the impact of heterogeneous relationships when few relationships are observable. Third, the factors that drive the behavior of complex time-varying individual investments are also heterogeneous. Traditional time modeling methods, such as vector autoregression, structured time series models, etc., often assume that future data is a linear combination of past data, and these methods are not sufficient to capture complex non-linear time dynamics patterns in real investment behavior. The fourth key challenge is derived from evolutionary interweaving between relationship influence and time-varying investment behavior characteristics, most of the current deep learning methods pay attention to time modeling or relationship modeling in various applications, and the fine combined modeling design for the relationship influence and the time dynamics is a difficult problem.
Disclosure of Invention
In summary, (1) the traditional investor portrait method mostly relies on questionnaires to collect investor information, but the collected information is limited and has low quality; (2) although many factors have been individually identified by the existing literature as having a significant impact on investor behavior, few studies have been made to adequately integrate these factors to design a practical investor imaging method; (3) most of the existing investor portrait methods are weak in the aspect of simulating the time dynamics of investment behaviors, and not to mention the influence of the constantly evolving social relationship on investment decision.
The inventor of the invention aims to get rid of the limitation of the traditional investor portrait method and overcome the defects of the existing data-driven investment behavior characteristic portrait method, (1) in addition to traditional attributes (such as age), a large amount of digital footprint data left by investors on an online platform is utilized, (2) various influence factors determined in documents are considered to construct comprehensive representation of investor portrait, (3) a machine learning method is adopted, and meanwhile, time and relation modes in investor portrait are modeled, so that the investor portrait construction method and device based on deep learning are developed.
One aspect of the invention is an investor portrait construction method based on deep learning, which comprises the following steps:
creating an investment behavior portrait by using historical investment data of an investor;
constructing an investment graph by using historical investment data of investors, wherein graph nodes representing the investors are connected with corresponding investment behavior images;
and simultaneously modeling the obtained investment behavior portrait and the relation influence and time dynamics in the investment map by utilizing a deep circulation map convolution neural network (RGCN), and outputting a predicted investment behavior portrait, wherein the deep circulation map convolution neural network is a structural combination of a map convolution network (GCN) and a circulation neural network (RNN).
According to the investor portrait construction method based on deep learning, the historical investment data are the investor attributes and the behavior statistical attributes extracted from the investment behavior records of the investor in the past specific time period.
The investor portrait construction method based on deep learning of the invention is characterized in that the historical investment data further comprises complex behavior attributes of the investor extracted from the investment behavior record of the investor in a specific period of time, wherein the complex behavior attributes comprise the attributes of the project, the attributes of the investor, the relationship between the investor and external market factors.
Another aspect of the present invention is an investor portrait construction device based on deep learning, which includes the following modules:
a module for creating an investment behavior representation using historical investment data of an investor;
a module for constructing an investment map using historical investment data of investors, wherein map nodes representing investors are connected with corresponding investment behavior figures;
and a module for modeling the obtained investment behavior sketch and the relationship influence and time dynamics in the investment map by using a depth cycle map convolutional neural network, and outputting a predicted investment behavior sketch, wherein the depth cycle map convolutional neural network is a structural combination of a map convolutional network and a cyclic neural network.
The investor figure constructing device based on deep learning of the invention is characterized in that the historical investment data is the investor attribute and the behavior statistical attribute extracted from the investment behavior record of the investor in the past specific time period.
The investor figure construction device based on deep learning of the invention is characterized in that the historical investment data further comprises complex behavior attributes of the investor extracted from the investment behavior record of the investor in a specific period of time, wherein the complex behavior attributes comprise the attributes of the project, the attributes of the investor, the relationship between the investor and external market factors.
Yet another aspect of the invention is a computer comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for carrying out the steps of the method described above.
The investor portrait construction method and device based on deep learning 1) extracts and integrates comprehensive and meaningful information from a large amount of historical data to construct a proper investment behavior portrait, so that the accuracy of the investor portrait construction method is improved; 2) accurately predicting the future investment behavior characteristics of the individual by simultaneously capturing relationship influence and time dynamics; 3) the prediction error is obviously reduced, the image prediction performance is improved, and the effectiveness of the investment behavior prediction mode in practical application is improved.
The invention can also provide important inspiration for analyzing the investment behaviors of various stakeholders (such as investors, practitioners and supervisors) so as to support decision processes of the investors, including investment risk management, investor wealth management, abnormal behavior detection, financial product innovation and the like. It is also applicable to other fields where it is desirable for a practitioner to model their objects from temporal dynamics and relational influences to construct a profile, for example, for creating patient profiles using electronic health records for disease prediction, prevention, treatment, and the like. Therefore, the method of the present invention is expected to improve the practical effectiveness in other fields of application and to be popularized.
Drawings
Fig. 1 schematically shows the steps of the investor representation construction method of the invention and possible applications thereof.
FIG. 2 schematically illustrates the steps of the investor representation construction method of the invention.
FIG. 3(a) is a pictorial representation of a set of representations of independent investment activities conducted by an investor over a period of time in accordance with an embodiment of the present invention; FIG. 3(B) is a unified representation of a composite sketch according to an embodiment of the present invention, wherein T1 and B1 respectively calculate the investment figures for Top 10% and Bottom 10% in all listings.
Fig. 4 shows an example of constructing an investment map in the present invention.
Figure 5 schematically illustrates the architecture of a recurrent graph convolution neural network for investment behavior prediction.
FIG. 6(a) shows a basic recurrent neural network, each box representing a neuron; FIG. 6(b) is a gated-cycle cell.
Detailed Description
The following describes a specific embodiment of the present invention with reference to the drawings.
As shown in FIG. 1, the investor portrait construction method based on deep learning of the invention comprises the following steps: creating an investment behavior portrait by using historical investment data of an investor; constructing an investment map by using historical investment data of investors; and the module is used for modeling the investment behavior portrait and the relation influence and time dynamics in the investment map by utilizing the deep cycle map convolutional neural network, and outputting the predicted investment behavior portrait so as to be applied to various financial practices such as active investor prediction, investor recommendation and the like.
Investment behavior portraits
The investment behavior sketch is a comprehensive description of the attributes and the characteristics of the investment behavior of the investor. After obtaining the record of the investor's investment behavior in a specific time period, it can create the picture of their investment behavior in the corresponding time period. Portrayal creation has two main requirements, namely information richness and presentation uniformity. In particular, the created representation needs to contain as much useful information as possible to reflect the investor's investment behavior. In addition, investors can invest different projects in different time periods, have different assets, and the uniform representation is beneficial to ensuring that the portrait formats of different investors and time periods are the same.
Information of portrait
The invention determines the specific information which the investment behavior portrait should include according to the investment behavior record, and disperses the past time span into a series of time periods, and each historical investment record (S belongs to S) usually contains investors (V belongs to V) in a time period
Investment projects created internally to investor (O e.O)
Various information of investment behavior (fig. 2). The invention utilizesThe following categories of information describe arbitrary investment behavior: investor attributes and behavioral statistics attributes, project attributes, investor-investor relationships, and external market factors. Unlike traditional portrayal methods, the present invention integrates the latter four types into the portrayal information, which is referred to as complex behavior attributes (Elabonate BehaviorAttributes).
Investor attributes generally refer to investor age, education level, and the like.
The behavior statistic attributes generally refer to the number of bids, the investment amount and the like.
In addition, attributes of the investment items, such as return rate and item purpose, may reflect the investor's preferences for the items.
Similarly, the investor attributes, such as credit rating, current debt-to-income ratio, or whether it is a homeowner, etc., may also provide indicative information about the investor's propensity.
The investor's relationship with the investor strongly influences the investor's investment behavior and they may be linked to each other through various relationships, for example, the lender (investor) and the borrower (investor) may be friends in a peer-to-peer loan market, the institutional investor and the subsidized company may work together in the same area, and so on. Use of the invention avo(t) to represent the relationship attributes between investor v and investor o for time period t.
External market factors such as the quantity and quality of candidate investment items may influence investor investment decisions. For example, if the overall quality of a candidate investment project is significantly degraded (e.g., due to global economic crisis), the investor will significantly alter his investment behavior. Therefore, the present invention takes a time-varying external market factor as an attribute in the investment behavior pattern, and expresses the external market factor for a time period t as ax(t)。
Representation of an image
Investors typically conduct different amounts of investment activities over different time periods, and a problem arises because each investment activity record can be represented by a set of attributes: how to integrate different numbers of attributes into a unified representation while maintaining valuable information. There are two typical methods of synthesizing multiple attribute sets. One is to serialize the set of attributes of all investing activities of investors into a representation of the investing activities in a given time period, and although this serialized combination can retain complete information, it will result in unequal lengths of representations for different time periods, thus violating the requirement of representation uniformity. Another approach is to parallelize sets of attributes of investment behavior records, parallel combination can achieve a composite image of the same length, however, it is a challenge to maintain as much information as possible in the composite image during parallelization.
The invention adopts a parallelization method based on attribute vectorization, which uses a vector to represent each classification attribute, each element in the vector represents a unique category thereof, and the element value represents the investment quantity belonging to the category. Vectorization reserves a fixed length vector for each classification attribute, regardless of how many behavior records exist within a time period. Fig. 3 shows an example in which a credit score is a typical classification attribute, and includes 8 different categories, an investor has made 12 investments in a time period t, 8 credit scores are "a", 3 credit scores are "B", 1 credit score is "E", and thus the credit score attribute is represented as an 8-element vector, elements mapped to the categories "a", "B", and "E" are 8, 3, and 1, respectively, and the other elements are 0. Non-categorical attributes refer specifically to continuous variable attributes, such as interest rates. For non-categorical attributes, its values are first converted into a categorical format by discretization, i.e., collecting the values appearing in any investment behavior record and categorizing the values by rank, in the example of FIG. 3, the interest values are divided into four groups, Top 10% (T1), Top 10% to 50% (T5), Bottom 50% to 10% (B5), and Bottom 10% (B1), of the interest values, and then vectors are generated for the non-categorical attributes by the same vectorization method as the categorical attributes. The cutoff value (10%, 50%) may be modified to other values to achieve better performance in view of the particular application (e.g., creating a patient representation, etc.) and characteristics, while the method of the present invention is still effective.
The investment behavior image is composed of a series of attributes and behaviors corresponding to investorsVector components of fixed length related to attributes of statistical attributes, project attributes, investor and investor relationships, external market factors, and the like. Here, x is usedv(t) to represent a representation of investors v's investment behaviour over a period of time t, each investor v holding a series of representations xv＝{xv(t0),xv(t1),…,xv(ti-1) Is used for expressing the set of all investor investment behavior characteristics in a time period t, namely
X represents a representation of the investment behavior of all investors during all time periods.
Construction of time-varying investment maps
In the present invention, relationship factors such as investment impact and similarity between investors are captured by constructing an investment map. Investors are the focus and are considered nodes in the constructed investment map. The links may then indicate investment impact or similarity between investors. Relationships such as friends, past partners, and culturally similar investors may affect the investment behavior of the associated investors, however, associating an investor with any single type of relationship (e.g., friend) may lose indicative information from other factors (e.g., partner). Furthermore, the impact between investors is difficult to quantify directly, as relationships such as friendship are typically rarely recorded in investment behavior data. Given that all types of effective relationship effects may ultimately be turned to investment behavior, the present invention relates investors through their "follow-on investment" behavior to fully capture the various potential effects among investors.
To construct an investment map, investors can be connected sequentially with directed links relating to investment time, given a project. In particular, if a series of investors
viE.g. V invests a project j successively, then
Follow up
And through a directed edge will
And
are connected together. All investors for a project then form a directed path in the investment map, with early investors at the front and late investors at the rear. Investors with co-investments at similar times will be closer to each other in the figure and investors with more co-investments will share more paths.
The invention builds an investment map for each time period t based on the investment projects occurring at and before time period t. Figure 4 shows an example of two investment maps for two consecutive time periods. In the first time period, investors 2, 1, 3 invest in project A in turn, investor 3 invests in project B after investor 4, and then uses directed edge e<1,2>、e<3,1>、e<3,4>These investors are connected to construct an investment map. In the second time period, based on the graph of the first time period, according to the new investment behavior of the project C, investors 5 are added in the investment graph as new nodes, and some edges e are inserted<5,4＞、e<1,5>. That is, by adding nodes and edges corresponding to new investment behavior, the investment map for a period of time may grow based on the investment map for a previous period of time. It is to be understood that the series investment maps
Not only is the complete historical behavioral relationship stored, but also the change in relationship between investors over time.
In addition, the method can be used for producing a composite materialThe present invention distinguishes the weighting of edges based on different impact/similarity between investors. First, if investor viFollowing investors v on many projectsjThen investor vjMay be to investor viA number of strong effects occur. Then, from viTo vjAre essentially weighted by their subsequent investment frequency. Thus, in FIG. 4, during the second time period, e<3,1>Is greater than e<1,5>This is because the investor 3 follows the investor 1 in both time periods, while the investor 5 follows the investor 1 only in the second time period. Second, the time of occurrence of the trailing investment is also taken into account, with closer trailing investment relationships being more important and therefore being given higher weight. In fig. 4, the follow-up investment between investors 1 and 2 occurs in the first time period, so e<1,2>Decays in the second time period (fig. 4 exemplarily sets the decay parameter 0.8). Edge e is normalized here using sigmoid function<vi,vj>Weight at time period t:
wherein if v isiFollowing v at a time period tjThen I (v)i,vjT') is 1, i.e. e<vi,vj>∈Gt′And is and
otherwise I (v)i,vjAnd t') is 0.δ is the time decay parameter of the edge weight.
As can be seen from the above, according to the historical investment behavior record, the invention constructs a directed weighted investment graph G for each time period tt. That is, each node in the graph represents an investor, and the linked investors successively invest in at least one of the same items. It will be appreciated that since each investor has a representation of investment behavior during each time period, the nodes in the investment map are associated with the representation of investor behavior during the corresponding time period.
Cyclic graph convolution networkTime and relationship modeling of
After a series of time-varying investment behavior portraits and investment maps are obtained by the method, the time and relationship characteristics of the investment behaviors of investors are captured, all information is finally integrated, and the future investment behavior portraits of the investors can be accurately predicted. The invention provides a cyclic graph and volume network (RGCN) which is used for simultaneously capturing the time and relationship characteristics of investment behaviors of investors by utilizing the advantages of the graph and volume network (GCN) and the Recurrent Neural Network (RNN).
Figure 5 illustrates the architecture of RGCN, which implements both time and relationship modeling using both the important functions of random walk graph convolution and convolutional represented cyclic neural networks. Specifically, the input is a series of investment graphs in which all graph nodes (i.e., investors) are associated with their investment behavior profiles over respective time periods. And carrying out convolution operation on each node based on the investment graph of each time period, and encoding the investment behavior sketch of the node into convolution representation through a graph convolution function which is randomly walked. Briefly, the convolved representation of a node is an integrated representation that contains the investment behavior characteristics of the node and its neighbors. With a series of node convolution representations as inputs, the recurrent neural network captures the relationship factors and time-dynamic patterns by alternately stacking Gated Recurrent Units (GRUs) and dense layers. These two functions are described below.
Convolution of a pattern with random walk
The invention applies convolution operation to the investment graph, and gathers adjacent information to an investor node for embedding the relationship factors such as investment influence, similarity and the like. In integrating neighbor information, it is important to distinguish neighbors of a node, since an investor may be affected differently by his/her different neighbors. Thus, the present invention utilizes a restart Random Walk (RWR) method to measure the importance of neighbors. An investment map G is giventAnd an investor v from which the random walk will start and iteratively pass through G with a probability according to the edge weighttThe random walk may also return v and restart with a probability αvThe importance of other investors to v is demonstrated by solving the following equation, where u is the other node in the graph (other investors) that is not equal to v:
wherein A istShows an investment diagram GtQ denotes a query vector corresponding to all nodes in the graph, where q isv1 and qu＝0(u∈V,u≠v)。
The neighbor importance information is then convolved by RWR scores and an integrated character representation is generated for each investor
(i.e., convolution representation):
wherein the content of the first and second substances,
viis e.g. V, and
represents viAn investment behavior profile over a time period t. It is to be understood that the present invention will generate a series of convolution representations for each investor. In order to capture the time characteristics of the relationship factors, the invention does not directly apply a dense neural network to the convolution representation of the investors like the common GCN method, but takes the convolution representation sequence of the investors as input, and adopts a recurrent neural network to simultaneously capture the relationship and time characteristics of investment behavior modeling.
Recurrent neural network for convolutional representation of investors
RNN is a powerful tool that allows a priori knowledge to persist in predictions by using directed cyclic joining that feeds outputs of past time periods to neurons asHidden input for future time period prediction (fig. 6 a). The cyclic concatenation of RNNs can be spread out in time series into a string of replicating neurons. Given a variable length sequence x (t)0),x(t1),x(t2) …, RNN updates its hidden state by the following equation
Wherein f is a nonlinear function, such as a logic sigmoid function, a hyperbolic tangent function, a gated cyclic unit, and the like. Based on state
The RNN can predict the probability distribution of the next element in the sequence by:
wherein g may be a common activation function, such as a sigmoid function, a tangent function, or a softmax function, etc.
Considering that investment behavior is influenced by long-term and short-term factors, in RNN units, the present invention may preferably employ Gated Round Units (GRUs) that contain more complex hidden states h that may better carry long-term information (fig. 6 b). In order to obtain a time period tiOf the kth hidden GRU
The GRU outputs the former GRU
And newly calculated GRU output
Linear summation:
wherein the content of the first and second substances,
called update gate, which determines the amount of past information to be passed to the current hidden state. The update gate is calculated as follows:
obtaining newly calculated GRU output using tangent function
As a general purpose unit:
wherein the content of the first and second substances,
i.e. a reset gate, for forgetting/discarding irrelevant information when calculating the current GRU hidden state, the calculation formula is:
the RNN structure of the present invention (lower part of fig. 5) alternately stacks dense and GRU layers with GRU as a main unit for extracting potential patterns from the investor's time-series convolution representation. Specifically, to understand the relationship factors between investors per time period, the present invention applies a dense layer on the convolutional representation of each investor that encodes investor and its neighbors characteristics, and then the GRU layer is used as a core neuron to capture temporal dynamics over time. It is to be appreciated that because the GRU acts on the investor's embedded convolution representation, the neural network architecture of the present invention is able to model not only the investor's investment behavior patterns over time, but also the temporal dynamics of the relationship factors. Algorithm 1 below shows the pseudo code of the time and relationship model used by the RGCN of the present invention for investment behavior prediction.
Algorithm 1: temporal and relational investment behavior profiling using RGCN
Inputting: investor's investment over past time period
The historical data S of (1), the depth parameter K;
and (3) outputting: for all investors
At a subsequent time period
Predicted investment behavior portrayal
The invention designs a structural combination RGCN of GCN and RNN, which is a new deep learning structure of time and relationship, can comprehensively capture the relationship influence and time attribute of investors-investors, particularly comprises the relationship dynamic with time, and predicts the investment behavior portrayal of investors according to the past behavior record of the investors. However, in the conventional technique of representing investors, only limited information is used, and the influence of investors on relationships and time attributes are not considered at the same time, so that an inadequate representation is generated, and finally the investors complain about services such as inappropriate investment recommendation. Table 1 below lists the present invention in comparison to existing investor representation methods.
TABLE 1
Examples
The following further illustrates embodiments of the invention by way of specific examples.
Data set
The investor can be an individual or an organization, so the invention utilizes two representative real investment market data sets of Prosper and Crunchbase respectively to more fully evaluate the effect of the invention on predicting the investor image.
Prosper is one of the most well-known online point-to-point loan markets in the united states, and investors can bid and finance on individual loan lists requested by borrowers (investors) through the internet. Embodiments of the present invention employ Prosper data over a period of two years, from 2009-9 to 2011-8, with an average credit amount of $ 7,225 per listing and an average bid frequency of 17.2 per month per investor. Here, the two-year Prosper data is divided into 24 time periods, one time period being one month, and a representation is generated monthly for the Prosper investor. The Prosper data described above consists of individual investors taking bidding action for at least 6 months, with a total of 7,455 investors, 39,362 lists and 528,485 bids in the Prosper data set.
Crunchbase is a platform that collects various business information from global companies, tracking global investment and financing records for the company (i.e., the institutional investor). Embodiments of the present invention sample the institutional investor investment records from 2003 to 2012, the 10 years. Companies in Crunchbase typically need to raise more funds, from 10 thousand dollars to 10 billion dollars, than Prosper borrowers typically require thousands of dollars. Here, the time period is one year granular, as institutional investors typically make only a small investment per year. The data set retains institutional investors who engaged in investment activities for at least three years. In the Crunchbase dataset, 2,458 institutional investors, 15,254 subsidized companies and 51,898 successful investment actions were included.
According to the investor representation method of the invention, the investment representation information in the two data sets is classified, and comprises investor attributes (such as age), behavior statistic attributes (such as average investment amount), project attributes, investor-investor relationship and external market factors. Then, the information of the investment behavior of the public investors in a specific time period is stacked into the portrait by the portrait information representation method of the present invention. It is to be understood that the specific characteristics of creating a representation using these two data sets may be different, for example, in Prosper, social factors determine the relationship between investors and borrowers (e.g., friendship), while in Crunchbase, social factors between institutional investors and subsidized companies may be measured by their commonality (e.g., location commonality). The present embodiment employs the following specific features.
Prosper
The investor attributes are: age of investment; investor roles such as borrower, lender, and team leader.
Investor behavior statistics attributes: the number of bids; risk performance; past performance.
Item attributes: interest rate; project purposes, such as business purposes, home stimulation purposes, etc.; a duration; an amount; length of investment list.
Sponsor attributes: real estate information when creating a manifest; debt and income proportion; the credit rating.
Investor-investor relationship: a friendship; organizing the membership; past investment relations.
External market factors: the number of competitors; the number of items.
Crunchbase
The investor attributes are: the age of the organization; a milestone event; important member learning calendar; major member specialization; top college graduate proportion.
Investor behavior statistics attributes: an amount funded each year; annual supply times; total number of supplies.
Item attributes: required financing; required financing.
Sponsor attributes: a company type; a milestone event; age of the company; a company country; home Human Development Index (HDI); important member learning calendar; major member specialization; top college graduate proportion.
Investor-investor relationship: a regional relationship; membership of investors and investors; past investment relations.
External market factors: financing company total; total number of market investors.
1. The invention adopts RGCN to construct the performance evaluation of investor portrait
Deriving investors' past time periods based on the data set
Using RGCN to model the influence of the relationship and the temporal dynamics at the same time to predict the investor in the ith time period tiThe investment behavior of (1).
The predictive method will get a better evaluation if the predicted investment behaviour pattern is more similar to what is actually the case. Here, the investment behavior image is predicted by quantization using Root Mean Square Error (RMSE)
And actual xv(ti) The difference between the two to evaluate the predicted performance. The smaller the RMSE, the better the model.
To train the RGCN investment behavior profile model, we here follow the general parameter set of the deep learning method, combining three GRU layers and two dense layers, where each layer includes 256 hidden channels. And is preferably optimized using the Adam algorithm because it reduces errors faster and produces better results in the model used in the present invention than the gradient descent algorithm.
In addition, for comparison with RGCN, five types of reference methods are used here, including Shallow Static Modeling (SSM), Shallow Relational Modeling (SRM), Shallow Time Modeling (STM), Deep Relational Modeling (DRM), and Deep Time Modeling (DTM), as follows:
[ SSM ] Average (AVG): the AVG uses the average of the investor figures over the historical period to predict their future figures, regardless of temporal dynamics;
[ SSM ] K-nerest periods (KNP): KNP assumes that the investor's future behavior is primarily related to their recent behavior, and then predicts their future portrayal using the average of the investor's portrayal over the recent K past time periods, where K is set to 3;
[ SRM ] Random Walk with Restart (RWR): the RWR distinguishes the importance of the investor's neighbors from the investment map and then predicts the investor's future representation through a weighted combination of historical representations of the investor's own and his/her neighbors;
[ STM ] Vector Autoregensive (VAR): VAR is a linear regression model commonly used in time series analysis that predicts the future from the past with weighting;
[ STM ] Bayesian Structured Time Series (BSTS): BSTS is a set prediction method, averaging different combinations of predictors, where local trend and autoregressive (lag 3) components are added to the BSTS model;
[ STM ] Hidden Markov Model (HMM): the HMM predicts the probability of a future state of the sequence based on the current state of the sequence by following the Markov property;
[ DRM ] Graph conditional Neural Networks (GCN): the GCN averages the investor's historical portraits as their features and uses the GCN to generate embeddings for the investor to predict their portraits over the next time period.
[ DTM ] Current Neural Networks (RNN): the RNN models the temporal dynamics of investor behavior, but ignores the time-varying relationship factors between investors.
The results of the comparison are shown in table 2 below, where the RMSE based on the various methods described above is shown:
from the above table, it can be seen that the relational modeling of RWR can significantly reduce RSME on both data sets compared to AVG and KNP, and thus it can be seen that the present invention introduces the relational impact between investors, and can significantly improve the effectiveness of investment behavior profiling.
Also, the deep learning method consistently produces much lower RMSE values than the shallow reference method over all time periods of both data sets. For example, the RNN reduces the RMSE by an average of 34.24% and 35.12% compared to the best shallow benchmark methods, namely Prosper's VAR and Crunchbase's BSTS. In addition, GCN can also lift 7.72% and 18.71% of RWR, respectively.
In addition, comparing RGCN with the depth-time modeling method RNN and the depth-relationship modeling method GCN, it can be seen that, although RNN and GCN significantly reduce the RMSE of the shallow modeling method, RGCN can further significantly reduce the RMSE of RNN by 11.55% and 11.89% and reduce the RMSE of GCN by 38.30% and 17.26% on Prosper and Crunchbase. This result demonstrates the significant effect of the time and relationship model of the present invention in improving performance in investment behavior portrayal.
2. 3-aspect evaluation of the invention in two applications
Application 1: active investor prediction
This application aims to identify active investors who may invest in the next time period, and by distinguishing active investors from inactive investors, financial institutions and online peer-to-peer loan platforms can focus on providing better personalized services for highly active investors, such as product recommendation or portfolio customization, and reduce wasteful investment for inactive investors, and furthermore, the original enterprise can start financing from active institutional investors, and the like.
Specifically, the investor figures are used as independent variables to predict the activity of investors in future time periods. By comparing the predictive performance of active investors, the effectiveness of various portrait methods in creating portraits was examined. Investors who make at least one investment activity over a period of time are marked as "active", otherwise marked as "inactive". The model was trained using classical machine learning methods including linear regression, support vector machine and random forest, 5-fold cross validation was performed, and the prediction performance was compared in terms of Accuracy (Accuracy) and F1 scores. In fact, all machine learning methods (i.e., linear regression, support vector machine, and random forest) achieved similar results, while random forest results were better, so the following evaluations used investors of random forest training models to predict performance aggressively and compare average results over different time periods.
Application 2: investor recommendations
The purpose of this application is to properly relate investors and investees based on their information (e.g., characteristics), where one of the most popular recommendation methods, i.e., based on investor similarity, is used to compare the actual effects of various portrayal methods.
In particular, given a investor and a time period t, assuming that some investors have already invested in the investor, but the investors are not sufficiently funded to meet the investor's needs, the recommendation task is to select a further number of investors who may invest in the investor during the time period t. To accomplish this task, we compute the image similarity between any active candidate investor and the funded investor by cosine similarity, and choose the most similar N (topN) investors to recommend to the investor. The time segment lengths of the historical data were varied in predicting the picture and the performance of the top3, top5 and top7 recommendations, respectively, was assessed by the average F1 scores for the different time segment lengths. The portrayal approach is considered more effective if the predicted portrayal of investors can achieve better recommendation performance.
Evaluation 1: validity of information integration
In the present invention, in addition to investor attributes and behavioral statistics attributes, the inventors of the present invention newly propose a series of complex behavioral attributes to integrate into investor pictorial representations. To evaluate the effectiveness of information consolidation for complex behavioral attributes, comparison is made here using RGCN but using only investor attributes and behavioral statistics attributes to construct investor figures. Tables 3 and 4 show the results of the validity evaluation of the information integration in the two applications.
Table 3: effectiveness of information integration in active investor prediction
Table 4: effectiveness of information integration in investor recommendations
From the above table data, it can be seen that in both applications, performance is continually improved by adding complex behavior attributes to the investment representation. More specifically, in the active investor prediction (table 5), complex behavioral attributes may help Prosper and Crunchbase to improve F1 scores by 2.79% and 6.14%, respectively. In addition, the RGCN already has much information in the generated sketch for active investor prediction, including investor attribute and behavior statistics attribute information, so the prediction F1 score on both datasets approaches 0.9, and the performance on both datasets can still be significantly improved by integrating complex behavior attributes into the sketch representation. The improvement of integrating complex behavioral attributes is more pronounced in the investor recommendations (table 6). For example, in the top5 recommendations of Prosper and Crunchbase, the F1 score incorporating complex behavior attributes increased by 41.29% and 120.14%, respectively. The above results verify that integration into complex behavioral attributes can effectively enrich the representation of the portrait, thus benefiting practical applications.
Evaluation 2: effectiveness of RGCN
In addition to the above validation that RGCN is more effective than other modeling methods, there is a continuing check here of the actual effectiveness of RGCN in predicting images in both of the above applications. Specifically, investor figures were predicted in both applications using RGCN and all of the benchmark methods listed above, and then their performance was compared, as shown in tables 5 and 6.
Table 5: effectiveness of RGCN in active investor prediction
Table 5 shows that the RGCN generated representations on both data sets perform optimally in terms of accuracy and F1 score to predict the active investors in the future. In Prosper, the RGCN image can predict the active investors with an accuracy of 0.95, while the best shallow reference imaging method (i.e., VAR) can only achieve an accuracy of 0.821, i.e., the RGCN of the present invention is 12.64% higher in accuracy than the shallow reference method, and has a significant improvement of 28.25% in F1 score compared to VAR. In Crunchbase, the accuracy and F1 score of the best shallow benchmark method (i.e., RWR) were 0.805 and 0.838, respectively, while the accuracy and F1 score of RGCN increased by 16.82% and 13.49%, respectively. Moreover, RGCN improves active investor prediction performance compared to GCN and RNN, which verifies the practical effectiveness of simultaneous time and relationship modeling in investment behavior profiling.
Table 6: effectiveness of RGCN in investor recommendations
Table 6 compares the recommended performance with an F1 score, indicating that RGCN can significantly outperform all baseline methods. The average F1 score of the RGCN is higher than the best shallow base method (BSTS) of Prosper and Crunchbase by more than 30 percent, and compared with a depth method only considering time dynamics (RNN) or relationship influence (GCN), the portrait predicted by the RGCN is continuously and remarkably improved, namely the portrait of the RGCN can greatly improve the recommendation performance of investors.
Evaluation 3: the superiority of the investor portrait construction method of the invention
In both applications, the present invention is compared to existing investor representation methods. The existing investor portrayal uses investor attributes (Attribute, Attr) and Behavior Statistics (BS) attributes such as investor past performance, risk preference and past experience, and a past Behavior List (Behavior List, BL) of people. The results are shown in tables 7 and 8 below.
Table 7: the superiority of the investor sketch construction method in active investor prediction
Table 7 shows the superiority of the investor representation construction method of the invention over two data sets. The BL exhibits the lowest prediction accuracy, which means that simply listing all of the investor's past behavior does not accurately reflect her/his future investment positivity. The combination of Attr + BS + BL can only provide very limited performance improvement compared to the best single information reference, e.g. in Prosper, Attr + BS + BL only achieves 0.52% higher accuracy than Attr. The slight increase means that there is information redundancy between investor attributes, behavioral statistical attributes and behavioral lists. By containing more behavior attributes and modeling time and relationship factors, the investor portrait construction method can better understand behavior patterns of investors and enables active investors to predict and obtain remarkable performance improvement. In particular, the method of the invention improves the accuracy of active investor prediction and F1 score by 18.07% and 38.04%, respectively, on Prosper and 40.51% and 30.02% on Crunchbase, respectively, compared to the best benchmark method (Attr + BS + BL).
Table 8: the superiority of the investor portrait construction method in investor recommendation
Table 8 lists the performance of the present invention and the existing portrait method from the perspective of investor recommendations. The results show that the present invention consistently exhibits the best performance recommended by Top3, Top5 and Top 7. On Prosper, the average F1 score was increased by 14.24% over BS alone and by more than 20% over other baseline methods. At Crunchbase, Attr + BS + BL performed the highest in baseline testing, whereas in Top3, Top5 and Top7 recommendations, the present invention was reliably 21.97%, 28.52% and 32.54% higher than Attr + BS + BL, respectively. In addition, as is clear from the above-described benchmark test results, simply stacking various types of information does not always improve the performance in applications. For example, at Propper, Attr + BS + BL is inferior to any single information rendering method (Attr, BS, or BL). This highlights that the invention is significantly superior to all other benchmark methods by combining complex behavioral attributes while learning temporal dynamics and relational effects.
Claims (9)
1. An investor portrait construction method based on deep learning comprises the following steps:
creating an investment behavior portrait by using historical investment data of an investor;
constructing an investment graph by using historical investment data of investors, wherein graph nodes representing the investors are connected with corresponding investment behavior images;
and simultaneously modeling the obtained investment behavior sketch and the relation influence and time dynamics in the investment map by utilizing a depth cycle map convolutional neural network, and outputting a predicted investment behavior sketch, wherein the depth cycle map convolutional neural network is a structural combination of a map convolutional network and a cyclic neural network.
2. The deep learning-based investor representation construction method according to claim 1, wherein the historical investment data is investor attributes and behavioral statistical attributes extracted from investor's investment behavior records in past specific time periods.
3. The deep learning based investor representation construction method according to claim 2, wherein the historical investment data further comprises complex behavior attributes of the investor extracted from the investor's investment behavior record in a specific period of time in the past, wherein the complex behavior attributes comprise the attributes of the project, the investor's attributes, the investor's relationship with the investor, and external market factors.
4. The investor portrait construction method based on deep learning of any one of claims 1-3, wherein the algorithm for modeling the relationship influence and time dynamics by using the deep cycle graph convolutional neural network is as follows:
where S is investor v' S investment over all time periods in the past
X represents an investment behavior profile of all investors in all past time periods, G represents a series of investment profiles of all past time periods, i is a time period number, tiIndicating that the ith time period has elapsed,
is at tiThe importance of the other investors to this investor v, α is the probability of restarting the random walk,
is shown at tiA adjacency matrix for the time slot investment graph G, q represents the query vector corresponding to all nodes in the investment graph,
is at tiA convolution representation generated for the investor by convolving the importance information of the neighbors with the restart random walk score, K being a depth parameter, hkRefers to the hidden state of the kth e K gated loop elements,
for the subsequent time period for the totality of investors v ∈ v
To predict an investment behavior profile.
5. An investor portrait construction device based on deep learning comprises the following modules:
a module for creating an investment behavior representation using historical investment data of an investor;
a module for constructing an investment map using historical investment data of investors, wherein map nodes representing investors are connected with corresponding investment behavior figures;
and a module for modeling the obtained investment behavior sketch and the relationship influence and time dynamics in the investment map by using a depth cycle map convolutional neural network, and outputting a predicted investment behavior sketch, wherein the depth cycle map convolutional neural network is a structural combination of a map convolutional network and a cyclic neural network.
6. The deep learning-based investor representation construction device according to claim 4, wherein the historical investment data is the investor attributes and behavioral statistical attributes extracted from investor's investment behavior records in past specific time periods.
7. The deep learning based investor representation construction device according to claim 5, wherein the historical investment data further comprises complex behavior attributes of the investor extracted from investor's investment behavior record in past specific time period, wherein the complex behavior attributes comprise item attributes, investor's relationship with investor and external market factors.
8. A computer comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for carrying out the steps of the method according to any one of claims 1 to 3.
9. A computer comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for carrying out the steps of the method of claim 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010024111.1A CN111292118A (en) | 2020-01-10 | 2020-01-10 | Investor portrait construction method and device based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010024111.1A CN111292118A (en) | 2020-01-10 | 2020-01-10 | Investor portrait construction method and device based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111292118A true CN111292118A (en) | 2020-06-16 |
Family
ID=71023401
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010024111.1A Pending CN111292118A (en) | 2020-01-10 | 2020-01-10 | Investor portrait construction method and device based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111292118A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112862182A (en) * | 2021-02-04 | 2021-05-28 | 北京百度网讯科技有限公司 | Investment prediction method and device, electronic equipment and storage medium |
| CN113656702B (en) * | 2021-08-27 | 2023-07-14 | 建信基金管理有限责任公司 | User behavior prediction method and device |
-
2020
- 2020-01-10 CN CN202010024111.1A patent/CN111292118A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112862182A (en) * | 2021-02-04 | 2021-05-28 | 北京百度网讯科技有限公司 | Investment prediction method and device, electronic equipment and storage medium |
| CN113656702B (en) * | 2021-08-27 | 2023-07-14 | 建信基金管理有限责任公司 | User behavior prediction method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109977151B (en) | Data analysis method and system | |
| Baesens et al. | Credit risk analytics: Measurement techniques, applications, and examples in SAS | |
| Fader et al. | RFM and CLV: Using iso-value curves for customer base analysis | |
| Perez et al. | Digital social capital and performance of initial coin offerings | |
| Herrero-Lopez | Social interactions in P2P lending | |
| Huang | Mark Twain’s Cat: Investment experience, categorical thinking, and stock selection | |
| CN106875206A (en) | Acquisition of information, assessment, questionnaire method, device and server | |
| Magni et al. | AN ALTERNATIVE APPROACH TO FIRMS'EVALUATION: EXPERT SYSTEMS AND FUZZY LOGIC | |
| Fitzpatrick et al. | How can lenders prosper? Comparing machine learning approaches to identify profitable peer-to-peer loan investments | |
| Dumitrescu et al. | Machine Learning or Econometrics for Credit Scoring: Let's Get the Best of Both Worlds | |
| Koumba et al. | Does uncertainty predict cryptocurrency returns? A copula-based approach | |
| Cosma et al. | Assessing corporate governance quality: Substance over form | |
| Florez-Lopez | Modelling of insurers’ rating determinants. An application of machine learning techniques and statistical models | |
| CN111292118A (en) | Investor portrait construction method and device based on deep learning | |
| Sarfaraz et al. | Analyzing the Investment Behavior in the Iranian Stock Exchange during the COVID‐19 Pandemic Using Hybrid DEA and Data Mining Techniques | |
| Jain | Macro forecasting using alternative data | |
| Wang et al. | Economic development program spending in the US: is there club convergence? | |
| Stone | Computational analytics for venture finance | |
| Parr-Rud | Business Analytics Using SAS Enterprise Guide and SAS Enterprise Miner: A Beginner's Guide | |
| Panneerselvam et al. | Bias‐driven marketing that instigates pledging to a crowdfunding campaign: An experimental consideration of behavioral anomalies | |
| Liu | Predicting Labor Market Competition and Employee Mobility--a Machine Learning Approach | |
| Bothner et al. | A model of robust positions in social structure | |
| Saha | Exploring the Impact of Behavioral Anomalies in Investment Decision of Investors of Capital Market in Bangladesh: A Behavioral Finance Approach | |
| Loterman | Predicting loss given default | |
| Adekunle et al. | Models towards Risk Behavior Prediction and Analysis: A Netherlands Case study |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200616 |