CN116308809A

CN116308809A - Analyzer viewpoint quality assessment method and model based on event domain interaction

Info

Publication number: CN116308809A
Application number: CN202310169847.1A
Authority: CN
Inventors: 郭艳红; 蒋帅
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2023-02-27
Filing date: 2023-02-27
Publication date: 2023-06-23

Abstract

The application relates to the technical field of financial risk analysis, in particular to an analyst viewpoint quality assessment method and model based on event domain interaction. The method comprises the following steps: constructing an analyst viewpoint quality assessment model based on deep learning; classifying and determining three event domain feature representations corresponding to each view under evaluation according to the event domain of [ analyst event domain, forecast event domain, stock event domain ]; taking three event domain feature representations of the viewpoint to be evaluated as first-order event domains, performing traversal execution on the three event domain feature representations through a three-choice two-to-one complementary hierarchical interaction mode, and determining final interaction of the third-order event domains; and finally, inputting a third-order event domain, and finally, interacting with a preset viewpoint quality prediction formula to determine a quality prediction value corresponding to the viewpoint to be evaluated. According to the method, the model is built, the analysis viewpoint is disassembled and refined into the feature representation displayed in the form of [ analysis event field, prediction event field and stock event field ] for effectively screening and selecting the features.

Description

Analyzer viewpoint quality assessment method and model based on event domain interaction

Technical Field

The application relates to the technical field of financial risk analysis, in particular to an analyst viewpoint quality assessment method and model based on event domain interaction.

Background

Securities analysts are information specialists in the financial market, possessing advantageous resources and capabilities for collecting, interpreting and disseminating market information. In practice, analysts share relevant corporate and industry information to investors by publishing specialized, standardized referral reports. Researchers believe that information disclosure by analysts is more influential than profit bulletin or management layer guidance. Currently, the analyst's views contained in the referral reports provide immediate investment advice to investors, increasingly being utilized by investors to support their decisions. However, due to expertise differences, collision of interests, and dynamic changes in market environment, the quality of the analyst's view (AnalystOpinionQuality, AOQ) varies. Thus, careful consideration of whether to follow a particular view of an analyst is required to avoid potential economic loss from following an unreliable view.

In recent years, machine learning methods have been greatly developed, and remarkable achievements have been achieved in various financial technology applications. However, research in developing machine learning methods for AOQ predictions remains very limited. Existing studies have shown that analysts' professional abilities are often tied to industry knowledge and company specific referral experiences, and that differences in company qualification can also lead to differences in AOQ. The evidence above suggests that there are complex interactions between input features, and modeling such interactions can provide important implications for predicting AOQ. In this regard, one widely used automation approach is to explicitly enumerate all possible feature combinations.

However, this straightforward approach presents some problems: first, the available features may be redundant, modeling each possible interaction may result in a significant increase in model complexity, which makes it difficult for the model to determine an optimal interaction set in such a large search space. Second, modeling unnecessary interactions can adversely affect the performance of the model. In many real world applications, meaningful feature interactions are typically sparse relative to the combined space of the original features. In particular, the information expressed by some features may be homogenous, and the constructive interaction between such specific machine learning-based analyst population intelligent mining methods typically does not yield additional information value.

Furthermore, constructing a AOQ predictive model that investors can understand is an important way to make scientific decisions, requiring that the predictive model be able to provide insight into the historical perspective data of analysts and to interpret AOQ predictions that will be made. An excellent model should be a good trade-off between good predictive performance and good intelligibility.

However, the existing AOQ predictive model fails to meet the above requirements.

Disclosure of Invention

The application provides an analyst view quality assessment method and model based on event domain interaction, which can solve the problems that characteristics cannot be screened and selected, so that the characteristics have redundancy and homogeneity, and the characteristics cannot be balanced with good prediction performance and good understandability in the existing AOQ prediction model.

In a first aspect, the technical solution of the present application is an analyst perspective quality assessment method based on event domain interaction, including:

s1: constructing an analyst viewpoint quality evaluation model based on event domain interaction with the viewpoint to be evaluated as an input item and a quality prediction value corresponding to the viewpoint to be evaluated as an output item based on deep learning;

s2: acquiring a plurality of views to be evaluated of an analyst and inputting the views to the model, by which three event domain feature representations corresponding to each view to be evaluated are determined according to event domain classification of [ analyst event domain, predicted event domain, stock event domain ] and based on an attention mechanism;

s3: the three event domain feature representations of the views to be evaluated are used as first-order event domains through the model, traversing execution is carried out on the three event domain feature representations through a three-choice two-supplement one-supplement hierarchical interaction mode, and final interaction of the third-order event domains corresponding to each view to be evaluated is determined;

S4: inputting third-order event domains corresponding to each viewpoint to be evaluated through the model, and finally interacting with a preset viewpoint quality prediction formula to determine a quality prediction value corresponding to the viewpoint to be evaluated;

the preset viewpoint quality prediction formula is as follows:

in the method, in the process of the invention,

representing a quality prediction value;

β ₀ representing a bias term;

representing a third-order event domain final interaction;

f _z a fitting function representing the final interaction of the third order event domain.

Optionally, the step S2 includes:

s21: obtaining viewpoint descriptions of a plurality of viewpoints to be evaluated of an analyst and inputting the viewpoint descriptions into the model, and determining viewpoint descriptions corresponding to the event domain classifications by the model and performing the event domain classifications for the viewpoint descriptions of the viewpoints to be evaluated based on the event domain classifications of [ analyst event domains, predicted event domains, stock event domains ];

s22: for each view description corresponding to the event domain of the view under evaluation, feature mapping is performed by means of one-hot coding, binary vectors corresponding to the event domain classes are determined, feature embedding vectors corresponding to the binary vectors are determined according to a preset vector table, and three event domain feature representations corresponding to the view under evaluation are obtained from the feature embedding vectors and based on the attention mechanism.

Optionally, the step S22 includes:

s221: determining a feature classification of either category features or numerical features for each view description of the views under evaluation corresponding to the event domain classification;

S222: for the feature classification is a category feature f _c Is encoded by a one-hot encoding function

b _p =1 and b _j =0, and j+.p, determining the coding vector, and +.>

Converting the code vector to determine the class feature f corresponding to the feature classification _c The embedded vector e, e=w of view description of (a) b；

In the method, in the process of the invention,

representing class characteristics f _c Is the number of categories;

p represents the p-th feature classification for each perspective to be evaluated;

j represents the j-th feature representation of each perspective to be evaluated;

es represents the dimension of the embedded vector e;

s223: for the feature classification is a numerical feature f _n Is represented by Embedding the view description by Field Embedding, and the numerical feature f corresponding to the feature classification is determined _n The embedding vector e, e=x·ω;

in the formula, ω∈R ^es Omega represents a learnable embedded parameter vector, omega is defined by a numerical feature f _n All values in (3) are shared;

s224: determining an embedding matrix corresponding to each view to be evaluated, which comprises a plurality of characteristic embedding vectors, corresponding to the event domain classification according to the embedding vector e corresponding to each view to be evaluated, corresponding to the event domain classification;

An embedding matrix representing an event domain of an analyst, +.>

Feature embedding vectors representing the event domain of the analyst,

an embedding matrix representing a predicted event field, +.>

Feature embedding vector representing a predicted event field, +.>

An embedded matrix representing a stock event domain, +.>

Feature embedding vector representing stock event field, +.>

In the method, in the process of the invention,

an analyst event field representing an ith view under evaluation;

a predicted event field representing an i-th perspective to be evaluated;

a stock event field representing an ith perspective to be evaluated;

s225: determining event domain feature representations corresponding to event domain classifications based on an attention mechanism based on an embedding matrix corresponding to event domain classifications for each view under evaluation

Event Domain feature representation +.>

The calculation formula of (2) is as follows:

in the method, in the process of the invention,

representing attention weights corresponding to the feature embedding vectors;

W _d,1 and W is _d,2 All represent transformation matrix parameters, W _d,1 ∈R ^2es×es ，W _d,2 ∈R ^es ；

b _d,1 And b _d,2 All represent bias terms, b _d,1 ∈R ^es ，b _d,2 ∈R；

ReLU represents an activation function;

q _d representing a query vector in an attention mechanism, q _d ∈R ^es 。

Optionally, the step S3 includes:

s31: three event domains of the viewpoint to be evaluated are used as first-order event domains, any two event domains are selected in the three event domains to interact, and three second-order event domains are determined Event domain interactions

And->

The interaction procedure is as follows:

wherein d ₁ 、d ₂ Representing any two of an analyst event field, a forecast event field, and a stock event field;

representation d ₁ 、d ₂ Is>

Representation d ₁ 、d ₂ Bias item of->

tanh (x) represents the activation function,

representing a ha Ma Deji operator;

s32: selecting any one second-order event domain interaction to interact with a first-order event domain which does not participate in the corresponding second-order event domain interaction, and determining three third-order event domain initial interactions

And->

The interaction procedure is as follows:

in which W is _a∩r←s 、

And->

All represent model parameters that can be learned;

W _a∩r←s 、

b _a∩r←]s 、b _a∩s←r and b _r∩s←a All represent model parameters that can be learned;

b _a∩r←s 、b _a∩s←r 、b _r∩s←a ∈R ^es ；

s33: initial interaction according to three third-order event domains

And->

And determining a third-order event domain final interaction corresponding to each view to be evaluated based on the attentiveness mechanism +.>

The determination process of (2) is as follows:

in which W is _Υ,1 And W is _Υ,2 All represent transformation matrix, W _Υ,1 ∈R ^2es×es ，W _γ,2 ∈R ^es×es ；

b _r,1 And b _γ,2 All represent bias terms, b _γ,1 ∈R ^es ，b _γ,2 ∈R；

q _γ Representing a query vector in an attention mechanism, q _γ ∈R ^es ；

T represents the set of indices of the interaction, T= { a n r≡s, a n s≡r, r n s≡a }.

Optionally, a fitting function f in the model _z Comprising the following steps: seven residual multi-layer perceptrons;

the forward propagation process of the residual multi-layer perceptron is as follows:

In the method, in the process of the invention,

input representing the first hidden layer in the residual multi-layer perceptron,/>

Representing a third-order event domain final interaction;

l _z representing depth representing residual multi-layer perceptron;

representing the first concealment in a residual multi-layer perceptronModel parameters of the layer->

Bias term representing the first hidden layer in residual multi-layer perceptron, < >>

And, a fitting function f in the model _z Output scalar of (2)

The calculation formula of (2) is as follows:

optionally, the learning objective function of the event domain interaction-based analyst perspective quality assessment model is a binary cross entropy loss ζ;

wherein δ represents a Sigmoid function;

m represents the number of training samples;

y _i representing analyst's perspective x _i Is the true quality of (3);

lambda is a superparameter representing L ₂ The weight coefficient of the regular term;

representing all of the trainable parameters in the model.

In a second aspect, the present application provides an analyst perspective quality assessment model based on event domain interactions, comprising: the system comprises an event domain characterization module, a cross-domain interactive learning module and an attribution quality prediction module which are connected in sequence;

the event domain characterization module is used for acquiring a plurality of views to be evaluated of an analyst, classifying event domains according to [ analyst event domains, predicted event domains, stock event domains ] and determining three event domain feature representations corresponding to each view to be evaluated based on an attention mechanism;

The cross-domain interaction learning module is used for taking three event domain feature representations of views to be evaluated as first-order event domains, performing traversal execution on the three event domain feature representations through a three-choice two-supplement hierarchical interaction mode, and determining third-order event domain final interaction corresponding to each view to be evaluated;

the attribution quality prediction module is used for inputting a third-order event domain corresponding to each viewpoint to be evaluated and finally interacting with a preset viewpoint quality prediction formula to determine a quality prediction value corresponding to the viewpoint to be evaluated;

the preset viewpoint quality prediction formula is as follows:

in the method, in the process of the invention,

representing a quality prediction value;

β ₀ representing a bias term;

representing a third-order event domain final interaction;

f _z fitting function representing third-order event domain final interactions

The beneficial effects are that:

according to the method, the model is built, automatic analysis of the views of the analysts can be achieved, in the implementation process of the model, the views of the analysts are disassembled and refined into feature representations displayed in the form of [ analyst event fields, prediction event fields and stock event fields ] for effectively screening and selecting the features, the three event field feature representations are traversed and executed through a three-choice two-one-supplement hierarchical interaction mode, the three-order event field final interaction corresponding to each view to be evaluated is determined, and finally evaluation is conducted through a view quality prediction formula, so that efficient utilization and balanced utilization of the features can be guaranteed, and good prediction performance and comprehensibility are guaranteed;

In summary, the method and the device can solve the problems that the existing AOQ prediction model cannot select the features, so that the features have redundancy and homogeneity, and the good prediction performance and the good understandability cannot be balanced.

Drawings

In order to more clearly illustrate the technical solutions of the present application, the drawings that are needed in the embodiments will be briefly described below, and it will be obvious to those skilled in the art that other drawings can be obtained from these drawings without inventive effort.

FIG. 1 is a flow chart of an analyst perspective quality assessment method based on event domain interactions in an embodiment of the present application;

fig. 2 is a schematic structural diagram of an EDIAN model in an embodiment of the application;

FIG. 3 is a schematic diagram illustrating an implementation of the EDIAN model according to an embodiment of the present application;

FIG. 4 is a schematic diagram illustrating the attention mechanism in an embodiment of the present application;

fig. 5 is a schematic diagram of an interaction process of the EDIAN model in the embodiment of the application;

FIG. 6 is a schematic diagram showing the profitability of different prediction models according to an embodiment of the present application;

FIG. 7 is a schematic diagram of an analysis of attribution effects of event domain interactions in a AOQ prediction in an embodiment of the present application;

FIG. 8 is a comparison of future point of view quality for selected experts in the embodiments of the present application;

FIG. 9 is a schematic diagram of a local attention weight based importance visualization in an embodiment of the present application;

FIG. 10 is a schematic diagram of the impact of MLP connection on model performance in an embodiment of the application;

FIG. 11 is a schematic diagram showing the comparison of the effects of different model settings in an embodiment of the present application;

in the figure, a 1-event domain characterization module; 2-cross-domain interactive learning module; 3-due to quality prediction module.

Detailed Description

Reference will now be made in detail to the embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The embodiments described in the examples below do not represent all embodiments consistent with the present application. Merely as examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

In a first aspect, the present application provides a method for evaluating quality of an analyst's perspective based on event domain interaction, as shown in fig. 1, fig. 1 is a flow chart of a method for evaluating quality of an analyst's perspective based on event domain interaction in an embodiment of the present application, including:

S1: an analyst perspective quality assessment model based on event domain interactions having a perspective to be assessed as an input item and a quality prediction value corresponding to the perspective to be assessed as an output item is constructed based on deep learning.

Specifically, an analyst perspective quality assessment model based on event domain interactions is first outlined.

In the model, x= { X is used ₁ ,x ₂ ,…,x _m -representing a view data set of an analyst, m representing the number of view samples; x is x _i A feature vector representing the ith view;

using y _i E {0,1} represents the viewpoint x _i When the viewpoint is reliable, y _i =1, otherwise 0.

According to the event embedding framework, each analyst perspective may be represented as a triplet of event domains, embodiments of the present application use d= { a, r, s } to represent a set of event domains, a to represent an analyst event domain, r to represent a predicted event domain, s to represent a stock event domain.

Further, the analyst's perspective is expressed as

n _a 、n _r And n _s And each represents the number of features in each event domain.

Further, embodiments of the present application use

Representing the j-th feature within the analyst event field a e D, accordingly, use +.>

Representing the jth feature within the predicted event domain r.epsilon.D, using

Representing the jth feature within the stock event domain s e D;

Further, the feature set included in the perspective data set X is expressed as

Based on the above settings, the AOQ prediction problem to be solved by the embodiments of the present application is specifically "determine mapping function f: x → {0,1},

”。

fig. 2 is a schematic structural diagram of an EDIAN model in an embodiment of the application, and fig. 3 is an implementation schematic diagram of the EDIAN model in an embodiment of the application, where the model includes an event domain characterization module 1, a cross-domain interaction learning module 2, and an attribution quality prediction module 3 connected in sequence.

The analyst viewpoint quality assessment model (Event Domain Interaction Attribution Network, EDIAN) is simply referred to as EDIAN model, and is used for accurately predicting the quality of the analyst viewpoint.

The event domain characterization module 1 converts each input feature of the analyst's perspective into a dense real-valued vector through the embedding layer, and then integrates feature embedding using an attention mechanism (Attention Mechanism) to learn the overall representation of each event domain. In particular, the event domain representation may also be considered a first order interaction (First OrderInteraction). The attention mechanism may automatically capture more important features in the aggregate and reduce feature redundancy within each event domain.

The cross-domain interaction learning module 2 adopts advanced deep interaction learning technology to generate meaningful event domain interactions, and specifically comprises three second-order event domain interactions and one third-order event domain interaction.

The attribution quality prediction module 3 predicts AOQ using the constructed event domain interactions as high order features. The module follows the interpretation framework of the generalized additive model, and helps to improve the understandability of the model at the event domain level.

S2: several views to be evaluated of an analyst are acquired and input to a model by which three event domain feature representations corresponding to each view to be evaluated are determined from event domain classifications of [ analyst event domain, predicted event domain, stock event domain ] and based on an attention mechanism.

Wherein, step S2 includes:

s21: the method comprises the steps of obtaining viewpoint descriptions of a plurality of viewpoints to be evaluated of an analyst, inputting the viewpoint descriptions to a model, classifying the viewpoint descriptions corresponding to the event domain classification by the model and classifying the event domain based on the event domain classification of [ analyst event domain, prediction event domain, stock event domain ].

Specifically, feature embedding (Feature Embedding) techniques map feature values to a low-dimensional continuous space by capturing semantic feature information or similarity.

Wherein, step S22 includes:

s221: for each view description of the views to be evaluated, which corresponds to the event domain classification, a feature classification, which is either a category feature or a numerical feature, is determined.

b _p =1 and b _j =0, and j+.p, determining the coding vector, and +.>

Converting the code vector to determine the class feature f corresponding to the feature classification _c The embedding vector e, < >>

In the method, in the process of the invention,

representing class characteristics f _c Is a number of categories of (a).

p represents the p-th feature class for each perspective to be evaluated.

j represents the j-th feature representation of each perspective to be evaluated.

es denotes the dimension of the embedded vector e.

Specifically, for a category type feature, the standard practice is One-Hot Encoding, where the feature is first mapped to a sparse binary vector using One-Hot Encoding (One-Hot Encoding), and then an embedded representation is looked up in a vector table. The EDIAN model employs this approach to embed class-type features in the analyst's view.

EDIAN uses Vfc to represent the number of categories covered in a category characteristic fc e F.

For the p-th category in category feature fc, the present embodiment uses a single pieceThe thermal encoding function encodes it as

bp=1 and bj=0, and j+.p.

EDIAN reuses a leachable transformation matrix

To generate an embedded vector e, e=w corresponding to the category b。

In the method, in the process of the invention,

representing class characteristics f _c Is a number of categories of (a).

p represents the p-th feature class for each perspective to be evaluated.

es denotes the dimension of the embedded vector e.

S223: for the feature classification is a numerical feature f _n Is represented by Embedding the view description by Field Embedding, and the numerical feature f corresponding to the feature classification is determined _n The embedding vector e, e=x·ω from the viewpoint of (a).

In the formula, ω∈R ^es Omega represents a learnable embedded parameter vector, omega is defined by a numerical feature f _n Is shared by all the values in the table.

Specifically, for numerical features, the EDIAN model employs Field Embedding techniques to embed the learning into the representation. With Field Embedding, all values in the same feature share a unified feature Embedding, which is then multiplied by their feature values.

For numerical feature f _n One eigenvalue x of e F, embedded is denoted e=x·ω.

S224: from the embedding vectors e corresponding to the event domain classifications for each of the views under evaluation, an embedding matrix corresponding to the event domain classifications for each of the views under evaluation is determined that includes a number of feature embedding vectors.

An embedding matrix representing an event domain of an analyst, +.>

Feature embedding vector representing an analyst event field, < >>

An embedding matrix representing a predicted event field, +.>

Feature embedding vector representing a predicted event field, +.>

An embedded matrix representing a stock event domain, +.>

Feature embedding vector representing stock event field, +.>

In the method, in the process of the invention,

an analyst event field representing the ith view under evaluation.

A predicted event field representing an i-th perspective to be evaluated.

A stock event field representing the ith perspective to be evaluated.

Event Domain feature representation +.>

The calculation formula of (2) is as follows:

In the method, in the process of the invention,

representing the attention weights corresponding to the feature embedding vectors.

W _d,1 And W is _d,2 All represent transformation matrix parameters, W _d,1 ∈R ^2es×es ，W _d,2 ∈R ^es 。

b _d,1 And b _d,2 All represent bias terms, b _d,1 ∈R ^es ，b _d,2 ∈R。

ReLU represents an activation function.

q _d Representing a query vector in an attention mechanism, q _d ∈R ^es 。

Specifically, each vector in the feature embedding matrix describes an important feature of the event domain. To fully represent an event domain, EDIAN employs a unique attention mechanism to aggregate all feature embeddings within the domain. The basic idea of the attention mechanism is to allow different parts to have different contributions when combined into one single characterization, so that the contribution of the valuable part is more.

View x for analyst _i Is a specific event domain D epsilon D, and the feature embedding matrix is as follows

Accordingly, the attention representation (attention domain representation) of the event field can be determined by:

in the method, in the process of the invention,

Attention weighting

The value of (2) depends on the attention value +.>

In which W is _d,1 And W is _d,2 All represent transformation matrix parameters, W _d,1 ∈R ^2es×es ，W _d,2 ∈R ^es 。

b _d,1 And b _d,2 All represent Bias Term (Bias Term), b _d,1 ∈R ^es ，b _d,2 ∈R。

q _d Representing a query vector in an attention mechanism, q _d ∈R ^es 。

The parameters may all be determined by deep learning.

[ ·|· ] represents a join operator (Concatenation Operator);

ReLU (x) =max (0, x) represents the activation function.

By adding attention value->

Normalization was performed using a softmax function, the calculation formula was as follows:

FIG. 4 is a schematic diagram illustrating the attention mechanism according to the embodiment of the present application, and the EDIAN model learns the viewpoint x by applying the attention mechanism to the feature embedding matrix of the event domain _i Representation of each event field in a system, in particular comprising an analyst event field

Predictive event Domain +.>

And +.>

S3: and traversing and executing the three event domain feature representations through a model by taking the three event domain feature representations of the views to be evaluated as first-order event domains and using a three-choice two-to-one-supplement hierarchical interaction mode to determine the final interaction of the third-order event domains corresponding to each view to be evaluated.

Specifically, the EDIAN model then builds cross-domain interactions to effectively capture AOQ predicted implicit knowledge and potential patterns based on the view-based event domain representation (which may be considered first-order event domain interactions). Since there are three event fields in each analyst's view event, the EDIAN model can build three

Second order event domain interaction and one +. >

Third order event domain interactions.

Technically, the EDIAN model builds event domain interaction for AOQ prediction by taking advantage of Bilinear-FFM 119 modeling interaction.

Fig. 5 is a schematic diagram of an interaction process of the EDIAN model in the embodiment of the application, and as shown in fig. 5 (a), the EDIAN model uses a first-order and first-order interaction module (First order to First order Interaction, FFI) to construct a second-order event domain interaction. The FFI is first represented by integrating two different event domains with one linear layer and one nonlinear activation function. In this way, each original event domain is mapped to a new interaction space by a transformation matrix with shared parameters, which transformation allows each event domain to behave differently when interacting with different domains.

After obtaining the converted event field representation, selecting an appropriate interaction operator is another problem. The two classical interaction operators are the Inner Product (Inner Product) and the Hadamard Product (Hadamard Product). In general, inner products are widely used for shallow models, such as FM and FFM, while hadamard products are commonly used for deep models, such as AFM and NFM. Therefore, the FFI module provided by the embodiment of the application adopts Hadamard product to learn corresponding second-order event domain interaction.

Wherein, step S3 includes:

s31: from the point of view to be evaluatedThree event domains are used as first-order event domains, any two event domains are selected in the three event domains to interact, and three second-order event domain interactions are determined

And->

The interaction procedure is as follows:

wherein d ₁ 、d ₂ Representing any two of an analyst event field, a predicted event field, and a stock event field.

Representation d ₁ 、d ₂ Is used for the conversion matrix of (a). />

Representation d ₁ 、d ₂ Is included. />

tanh (x) represents the activation function,

representing the ha Ma Deji operator.

In particular, for analyst perspective x _i Is a two event domain of (2)

And->

d ₁ 、d ₂ E D, FFI generates a corresponding second order event domain interaction by:

in the method, in the process of the invention,

representation d ₁ 、d ₂ Is a learnable conversion matrix of->

/>

Representation d ₁ 、d ₂ Is +.>

Since each analyst view contains three event fields d= { a, r, s }, a combination of three different event fields can be used to construct a second order interaction.

According to the formula

View x for analyst _i The three 2-order event domain interactions generated by the EDIAN model are +.>

And->

And->

The interaction procedure is as follows:

in which W is _a∩r←s 、

And->

All represent model parameters that can be learned.

W _a∩r←s 、

b _a∩r←s 、b _a∩s←r And b _r∩s←a All represent model parameters that can be learned.

b _a∩r←s 、b _a∩s←r 、b _r∩s←a ∈R ^es 。

Specifically, EDIAN proposes a First-order and Second-order interaction module (FSI) to generate a third-order interaction by interacting a First-order interaction (i.e., an original event domain representation) with a Second-order interaction, the specific mechanism of which is shown in fig. 5 (b).

The first order interaction and the second order interaction of the inputs in the FSI must satisfy information complementary constraints, i.e. the first order input must be different from the two event domains in the second order interaction, which can prevent information interactions within the same event domain. The FSI then maps the first order interactions to interaction space with linear layers and nonlinear activation functions, and then the converted results will interact with the input second order interactions by hadamard products.

Will second order interaction

Interaction with first order->

Input FSI module, output third-order interaction +.>

Can be determined by the following formula:

correspondingly, the dia can obtain other two third-order event domain interactions through the FSI module, and the corresponding formulas are as follows:

in which W is _a∩r←s 、

And->

All represent model parameters that can be learned.

W _a∩r←s 、

b _a∩r←s 、b _a∩s←r 、b _r∩s←a ∈R ^es 。

S33: initial interaction according to three third-order event domains

And->

The determination process of (2) is as follows:

in which W is _Υ,1 And W is _Υ,2 All represent transformation matrix, W _Υ,1 ∈R ^2es×es ，W _γ,2 ∈R ^es×es 。

b _γ,1 And b _γ,2 All represent bias terms, b _γ,1 ∈R ^es ，b _γ,2 ∈R。

q _γ Representing a query vector in an attention mechanism, q _γ ∈R ^es 。

Specifically, in order to obtain the third-order event domain final interaction, the EDIAN model aggregates the three generated event domain initial interactions by using the attention mechanism illustrated in fig. 3, and the specific calculation method is as follows:

s4: and inputting a third-order event domain corresponding to each viewpoint to be evaluated through the model, finally interacting with a preset viewpoint quality prediction formula, and determining a quality prediction value corresponding to the viewpoint to be evaluated.

The preset viewpoint quality prediction formula is as follows:

in the method, in the process of the invention,

representing the quality prediction value.

β ₀ Representing the bias term.

Representing the third order event domain final interaction.

Wherein the fitting function f in the model _z Comprising the following steps: seven residual multi-layer perceptrons.

in the method, in the process of the invention,

Representing the third order event domain final interaction.

l _z Representing the depth of the residual multi-layer perceptron.

And the model parameters of the first hidden layer in the residual multi-layer perceptron are represented.

And the bias term of the first hidden layer in the residual multi-layer perceptron is represented.

And, fitting function f in the model _z Output scalar of (2)

The calculation formula of (2) is as follows:

in particular, EDIAN does not merely predict AOQ, but rather aims to give the model understandability in order to extract implicit knowledge from the analyst's point of view data. To this end, EDIAN builds AOQ a predictive module in accordance with the Generalized Additive Model (GAM) framework.

The embodiment of the application expresses the interaction index set as Z = { a, r, s, a ∈r, a ∈s, r ∈s, a ∈r ∈s }, EDIAN predicts the viewpoint quality y through the following formula _i ：

In the method, in the process of the invention,

representing the quality prediction value.

β ₀ Representing the bias term.

Representing the third order event domain final interaction.

The frame has two unique characteristics: (1) GAM mathematically relates the additivity of features to the mean of the hypothetical distribution using a linking function g, which can be flexibly selected according to the type of result; (2) Flexible fitting function f _z (.) relaxes the linear constraint in the simple regression model, but instead assumes that the result can be modeled by the sum of arbitrary functions for each feature. The core of GAM is still the summation of characteristic effects, but it allows some nonlinear relationship between characteristics and output. EDIAN helps to quantitatively account for which event domain interaction AOQ is caused by designing the prediction module in a GAM manner.

As shown in fig. 3, the EDIAN model uses 7 Residual Multi-Layer perceptions (Residual Multi-Layer perceptions) as a fitting function, which uses Skip Connection (Skip Connection) to alleviate the gradient vanishing problem and send past information through deeper layers.

Interaction for a particular event domain

EDIAN represents the corresponding residual multi-layer perceptron function as f _z And use->

Represents f _z Input of the first hidden layer of (1) is performed by using l _z Representing depth representing residual multi-layer perceptron;

in the method, in the process of the invention,

model parameters representing the first hidden layer in a residual multi-layer perceptron, < >>

ReLU is an activation function and

after this, the output of the residual multi-layer perceptron corresponding to the time domain interaction Z e Z is a scalar +. >

Wherein->

And->

Are all model parameters that can be learned.

(II) in a second aspect, embodiments of the present application provide an analyst perspective quality assessment model based on event domain interactions, comprising: an event domain characterization module 1, a cross-domain interactive learning module 2 and an attribution quality prediction module 3 which are connected in sequence.

An event domain characterization module 1 for acquiring several views to be evaluated of an analyst and classifying according to [ analyst event domain, predicted event domain, stock event domain ] event domain and determining three event domain feature representations corresponding to each view to be evaluated based on an attention mechanism.

The cross-domain interaction learning module 2 is configured to take three event domain feature representations of views to be evaluated as first-order event domains, and perform traversal execution on the three event domain feature representations through a three-choice two-supplement hierarchical interaction mode, so as to determine final interactions of the third-order event domains corresponding to each view to be evaluated.

The attribution quality prediction module 3 is configured to input a third-order event domain corresponding to each viewpoint to be evaluated and finally interact with a preset viewpoint quality prediction formula to determine a quality prediction value corresponding to the viewpoint to be evaluated.

The preset viewpoint quality prediction formula is as follows:

In the method, in the process of the invention,

representing the quality prediction value. />

β ₀ Representing the bias term.

Representing the third order event domain final interaction.

Specifically, the learning objective function of the EDIAN model during training is a binary cross entropy loss (Binary Cross Entropy Loss), which is defined as follows:

in the formula, δ represents a Sigmoid function.

m represents the number of training samples.

y _i Representing analyst's perspective x _i Is a true quality of (c).

Lambda is a superparameter representing L ₂ The weighting coefficients of the regularization term.

Representing all of the trainable parameters in the model.

In the training process, the embodiment of the application adopts a feature normalization technology to perform standardized processing on the digital feature values in the data preprocessing stage. Examples: a digital eigenvalue x is normalized to x++x-mean (x)/std (x).

The embodiment of the application takes second-order event domain interaction as an example to analyze the difference between event domain interaction and feature interaction.

Let the analyst point of view x _i Is characterized by the feature embedding matrix of

n represents the number of features, n=n _a +n _r +n _s 。

The feature interactions explicitly built by enumerating each feature combination can be represented by the following formula:

e _i,j e _i,k ,j∈{1,2,…,n-1},k∈{j+1,…,n}；

therefore, construction is required

The interaction of the features, the surface time complexity is O (n ² )＝O[(n _a +n _r +n _s ) ² ]。

In addition, the weights of each interaction constructed are the same, and this approach is ineffective in distinguishing meaningful interactions from noise, as opposed to event domain interactions proposed by embodiments of the present application can be represented as follows:

Wherein d ₁ 、d ₂ E D represents the event domain of two interactions,

representing the total two feature embedding vectors of the event field +.>

Is determined by end-to-end learning.

From the above, it can be seen that the time complexity of constructing event domain interactions is O (n _a n _r +n _a n _s +n _r n _s ) O (n) of specific enumeration feature interaction method ² )＝O[(n _a +n _r +n _s ) ² ]Much smaller (especially in high dimensional situations).

The reduction in complexity comes mainly from the EDIAN model ignoring interactions between features within the same event domain, as their homogeneity is high, and often does not provide additional information value. The importance of the interaction may be represented by the attention weight of the interaction features. Therefore, event domain interaction constructed by the EDIAN model provided by the embodiment of the application is beneficial to identifying meaningful interaction.

(III) the construction process and related analysis of an analyst perspective quality assessment model based on event domain interactions are now shown with specific data examples.

1. Experimental data set

The present embodiments collect a perspective dataset of security analysts from the national security financial database (China Stock Market & Accounting Research, CSMAR), which is one of the most standard, authoritative financial databases in the chinese securities market. The dataset contained a total of 308167 instances of analyst view over a 10 year time span (from 2010 to 2019). According to the distribution time sequence of each view event, views distributed in the first eight years are set as training sets, and views distributed in the remaining one and a half years are used for testing. Specific data partitions are summarized in table 1. In addition, for calculation AOQ, embodiments of the present application collect daily price data for related stocks and market indexes on the chinese securities market from the Wind financial database. To handle the predictive tasks of AOQ, embodiments of the present application collect features of perspective based on the event embedding framework.

Table 1 dataset description and partitioning thereof

Data set	Time period	Number of views	Ratio from a reliability standpoint
				Training set	2010/01/01-2018/06/30	251569	35.23％
Test set	2018/07/01-2019/12/31	56598	39.49％
				All of	2010/01/01-2019/12/31	308167	36.02％

The data set included 25 features for AOQ prediction, as shown in table 2. The features are divided according to the event domains to which they belong, wherein the analyst event domain contains 11 features, the predicted event domain has 5 features, and the remaining 9 features belong to the stock event domain. All features are of the continuous numerical type or of the discrete class type, marked with N and C respectively. For numeric features, the fourth column of table 6.2 shows their average value, while for category features, the fourth column is the number of categories contained in the feature. The numbers in table 6.2 show: the data set covers 3961 analysts from 113 securities companies and 2667 stocks from 20 industries.

The predictive goal of embodiments of the present application is analyst perspective quality AOQ, which is a measure of whether the analyst's rating of the referring stock can correctly predict future revenue trends for the target stock. This view is considered reliable if the actual yield of the stock meets the expectations of the analyst, otherwise unreliable.

TABLE 2 characterization of AOQ prediction

AOQ indicates whether the analyst's rating of the referring stock correctly predicts the return trend of the target stock during the pre-specified future period, and thus the calculated idea and implementation steps of AOQ are as follows:

(1) Representing an expected value of the analyst's referral share rating by an expected value ARE;

ARE＝{S,T,B,R _min ,R _max }；

wherein S represents target stocks, T represents an observation window, and B represents a profit standard;

R _min and R is _max Representing the maximum and minimum of expected revenue, respectively.

(R _min ,R _max ) Is an interval indicating that the expected yield of analysts is from R _min To R _max 。

In some embodiments, an example of criteria for an analyst to rate the strands is shown in the following table, where the rating criteria is the relative market performance of the company's share price (or industry index) versus the contemporaneous benchmark index within 6 months of the report release date. Wherein the A market is based on the Shanghai Deck 300 index; the new three-board market is based on three-board index (for protocol transfer targets) or three-board market index (for market transfer targets); the Chinese hong Kong market is based on the Morganisdan China index, and the Meihan market is based on the standard 500 index or the Nastrake comprehensive index.

Table 3 analyst referral rating criteria example

Rating of the referral strand	Rating criteria
		Buying and putting in	The rise amplitude of the reference index is more than 15 percent relative to the synchronous
Increase the holding	The amplitude of the reference index is 5% to 15% relative to the synchronous
		Hold and hold	The amplitude of the reference index is between-5% +5% relative to the contemporaneous
Reduction of hold	The falling amplitude of the reference index is more than 5 percent relative to the synchronous reference index

The particular referral rating reported by the analyst corresponding to Table 3 is a buy and the analyst's rating's expectation may be formally determined as shown at 3 according to the above equation.

Table 4 expected benefits of the rating of the referral strands in Table 3

Parameters (parameters)	Value taking
		Target stock (S)	688383.SH
Viewing window (T)	6months
		Revenue benchmark (B)	CSI300Index
Minimum benefit (Rmin)	5％
		Maximum benefit (Rmax)	15％

(2) After deducing the implicit analyst's expectations from the rating criteria, calculating the actual benefits of the target stock is the core task of determining AOQ, i.e., the actual benefits that need to be calculated can be divided into two types: absolute benefit and relative benefit.

Let are= { S, T, B, R _min ,R _max The } represents an implied expected value in the target referral rating, and if the rating predicts only the return trend of the stock itself within the observation window, meaning b=null, the return that needs to be calculated is an absolute return. The absolute profit only needs to calculate the accumulated profit rate of the target stock S in the observation window T, and the specific calculation formula is as follows:

in the method, in the process of the invention,

indicating the price of the target stock on the date of the rated release, < ->

Representing the stock price of the target stock at the t-th trading day after the release of the rating;

if the analyst predicts the rate of return of the target stock relative to the market index within the observation window, then the return to be calculated is the relative return. In this case, in addition to the cumulative rate of return of the target stock, the cumulative rate of return of the market index B should be calculated by the following formula:

In the method, in the process of the invention,

price representing market index on rated release day, +.>

Representing the price of the market index at the t-th trade day after the release of the rating;

in summary, the calculation formula of the return of the target stock relative to the market index is as follows:

the calculation formula of the actual benefits AR of the target stock corresponding to the rating of the analyst in the observation window is as follows:

(3) Because of the significant differences in the rating criteria and levels of the referring shares of each stock company, the referring share ratings of analysts are generally divided into three categories: positive, negative and neutral. The evaluation method has the advantages that the evaluation stock is in a positive view, the evaluation stock is in a negative view, and the neutral view is between the two conditions. For example, for the ratings appearing in Table 4, "buy" and "hold" are positive, "hold" is negative, and "hold" is neutral.

For a particular rating of the referral strand, the analyst expects the benefits ep= { S, T, B, R _min ,R _max The } can be obtained by ARE= { S, T, B, R _min ,R _max Determination of actual benefit ar= { AR _t |t∈[0,T]The } can be determined by:

thus, for positive rating, if R _min ≤max(AR)≤R _max It is satisfied that it is to be considered reliable; otherwise, it is unreliable;

For negative ratings, if R _min ≤min(AR)≤R _max It is satisfied that it will be considered reliable; otherwise, it is unreliable;

for neutral rating, if R _min ≤min(AR)≤max(AR)≤R _max It is satisfied that it will be considered reliable; otherwise, it is unreliable.

2. Baseline model and metric

To verify the effectiveness of EDIAN in predicting AOQ, the present embodiments compare it to a series of competitive machine learning baseline models, which are often used to handle classification tasks in various research areas and exhibit excellent performance. The baseline model adopted by the embodiment of the application comprises the following steps:

(1)LR(Logistic Regres-sion)；

(2)FM(Factorization Machine)；

(3)RFs(Random Forests)；

(4)LGBM(Light Gra-dient Boosting Machine)；

(5)PlainMLP(Plain Multi-layer Perceptron)；

(6)DenseMLP(Dense Multi-layer Perceptron)；

(7)ResMLP(Residual Multi-layer Perceptron)；

(8)WAD(Wide an Deep)；

(9)DeepFM。

to obtain a comprehensive and reliable comparison between the EDIAN model and the baseline model presented in the examples of the present application, six well-established performance assessment metrics were employed in this study, including Accuracy, precision, recall, F1-score, AUC and Log-loss. The definition of the first five indices is as follows:

the AUC curve is calculated and determined by ROC curve, balances Precision and Recall, and measures the comprehensive performance of a model. ROC curves with false positive rate (FPR, equivalent to 1-Precision) and true positive rate (TPR, i.e., recall) as x-axis and y-axis show the relative trade-offs of FPR and TPR, respectively.

Log-loss, also known as logic loss or cross entropy loss, is defined as the negative logarithm of the probability that the model returns the property for its training data y.

For an analyst perspective sample with a true label y/in 0,1 and probability estimate p=pr (y=1), the log loss is shown as follows:

l(y,p)＝-[ylog(p)+(1-y)log(1-p)]；

furthermore, to estimate the relative improvement achieved by the EDIAN model relative to the compared baseline model, embodiments of the present application herein define the calculation formula for the relative improvement RI for a particular index M as follows:

3. results of excess syndrome

3.1 predictive Performance comparison

The performance of the AOQ predictive model is reported in table 5 to explore the predictive capabilities of the EDIAN model presented in the examples of the present application. From the experimental results in this table the following conclusions can be drawn:

1) In conventional non-deep learning methods, integrated models (RFs, LGBM) are superior to individual models (LR, FM) in most metrics, which illustrates the superiority of using population wisdom to build machine learning models.

2) Deep FM performs better overall than WAD in the Deep learning approach, with the main difference that Deep FM explicitly builds 2-order feature interactions, the above results indicating the importance of building low-order interactions.

3) The MLPs model does not focus on low-order feature interactions, but rather implies that high-order feature interactions are built. It performs better than Deep FM and WAD, indicating the importance of higher order feature interactions. In addition, res MLP and Dense MLP show better performance than Plain MLP, which shows that proper connection mode is designed in MLPs to solve the problems of gradient disappearance, information loss and the like, and the prediction capability is improved.

(4) Compared with the adopted baseline model, the EDIAN model provided by the embodiment of the application can obtain the best performance on all evaluation indexes. In particular, EDIAN was improved by 2.32% Accicy, 1.72% Precision,3.21% Recall,4.10% F1-Score,1.56% AUC and 1.30% Log-loss over the strongest baseline. This result verifies that modeling interactions at the event domain level is a more advantageous way to model interactions at the feature level than for the AOQ predictive task. The interaction of the event fields effectively captures the valuable knowledge implicit in the analyst's perspective data as it reduces interactions between redundant and homogenous information.

Table 5 comparison of AOQ predictive performance for different machine learning methods

/>

3.2 investment returns comparison

The ability to obtain revenue from stocks recommended by investment analysts is a major concern for investors when making decisions using the analyst's perspective. Unlike AOQ, which relies on a complete sample of analyst views, the return on investment is more dependent on the best quality analyst view. To explore whether the improved AOQ predictive performance of EDIAN can help investors achieve higher yields in decisions using analyst views, embodiments of the present application construct a practical investment strategy to select the stocks recommended by the analysts for simulated investment.

Investment strategy: for an analyst's perspective event x, the machine learning model may output its predicted quality

I.e. the reliability probability of the viewpoint.

The return of the target stock in the analyst' S view is expected to be are= { S, T, B, R _min ,R _max }；

Wherein ARE represents an expected value of an analyst' S rating of a referring stock, S represents a target stock, T represents an observation window, B represents a revenue benchmark, R _min And R is _max Representing the maximum and minimum of expected revenue, respectively. (R) _min ,R _max ) Is an interval representing that the expected yield of analysts is from R _min To R _max 。

The present embodiments assume that investors are Risk aversions (Risk averses) and tend to take careful investment strategies. By using

As an index for measuring the investment value of a stock recommendation, the minimum expected return per unit time is expressed.

When the recommended stock has a high Score, it is worth the investment. Thus, embodiments of the present application rank the stocks recommended by the analysts according to Score, and the top-ranked stocks on this list may be selected as investment targets.

Based on the investment policies described above, different machine learning AOQ predictive models can generate a ranked list of recommended stocks by different analysts. In application practice, the revenue performance of EDIAN is always continuously superior to other comparative investment strategies.

Furthermore, the above schemes have described how to calculate the actual benefits of a selected analyst's referral share ratings, which embodiments of the present application use as another measure of the profitability of an investment strategy. As shown in fig. 6, fig. 6 is a schematic diagram comparing the profitability of different prediction models in the embodiment of the present application, and DeepFM, resMLP, denseMLP, LGBM and DEIAN are sequentially shown from left to right in each group of graphs in fig. 6, that is, fig. 6 summarizes the comparative profitability results of different investment strategies. EDIAN always achieves the best profitability compared to the other baselines, regardless of the highest percentage set. Overall, the results reported in this section are fully indicative that the improvement in predictive power achieved by EDIAN can translate into an improvement in investment benefits.

3.3 expert identification study

3.3.1 analysis of perspective attribution

The EDIAN model follows the framework of the generalized additive model to learn how to utilize event domain interaction predictions AOQ. Since AOQ is modeled as a summation of the effect values of the different event domain interactions, EDIAN can attribute AOQ at the level of event domain interactions, thereby quantitatively analyzing which event domain interactions actually dominate the predicted AOQ.

The embodiments of the present application randomly select an example of an analyst's perspective to illustrate the attribution process of EDIAN model to AOQ. The referral rating and rating criteria specified in the analyst's view indicate that the analyst expects that the price of the target stock will rise 5% to 15% within the next 12 months. In reality, the price of the stock rises 9.80% within 12 months after being evaluated, and this analyst's view is reliable according to the definition of AOQ. Accordingly, the EDIAN model predicts that this view is f (x) =sigmoid (0.3662) =0.5905 >0.5, correctly predicting that it is reliable.

As shown in fig. 7, fig. 7 is a schematic diagram of analysis of attribution effects of event domain interactions in AOQ prediction in the embodiment of the present application, which can prove that EDIAN predicts that the analyst's view is reliable, and the waterfall diagram in fig. 7 visualizes attribution effect values of constructed event domain interactions. The results in fig. 7 show that: first-order domain interaction e ^r First order domain interaction e ^s Interaction with second order domain e ^r∩s Negative attribution effects are caused, resulting in the view that model prediction is unreliable. However, first-order domain interaction e ^a In particular e ^a 3 rd order domain interaction e ^a∩r∩s Generates very highPositive attribution effect, which completely twists the predicted result.

Thus, these two event domain interactions are the primary contributors to the model's predicted outcome. Based on detailed investigation, the analyst has sent 450 reports of referral strands in the past, 70% of which are reliable. In addition, the analyst published 22 reports of referral to the industry of target stocks, with 15 (nearly 70%) reports of referral ratings being reliable. The above facts indicate that the analyst is an expert in analyzing target stocks.

According to classical attribution theory (Attribution Theory), key factors that explain success or failure behavior (including ability, effort, task difficulty, and fortune) can be classified into two categories by "control point" (Locus of control): intrinsic reasons describing personal characteristics (including competence and effort), and extrinsic reasons controlled by environmental factors (including task difficulty and fortune). Attributing current success to intrinsic causes enhances the expectation of future behavior success. A reliable analyst view may be considered a successful behavior through which the question of "why it was successful" may be answered by causal attribution studies. The attribution theory mentioned above provides a very promising perspective for identifying real experts from a population of analysts using the attribution capabilities of EDIAN.

Specifically, AOQ can be attributed by the seven event domain interactions and the effect of one bias term built by the EDIAN model. According to control point attribution theory, event domain interactions associated with analysts may be considered intrinsic causes, while the remaining event domain interactions and bias terms are extrinsic causes. In EDIAN, AOQ is a combined effect of both intrinsic event domain interactions and extrinsic event domain interactions. For each reliable analyst perspective, EDIAN may perform attribution analysis (according to the example in fig. 7) to determine whether AOQ is dominated by intrinsic event domain interactions. Intuitively, if a reliable analyst's view is dominated by intrinsic event domain interactions, it is more likely to demonstrate that the success of the view is due to the analyst's capabilities, expertise, and other personal features. Thus, embodiments of the present application should have a higher expectation for analysts who have put forward more reliable views due to inherent interactions. The present embodiments measure the expertise of an analyst in the EDIAN model as a percentage of the reliable views that can be attributed to intrinsic interactions (denoted AttrScore).

3.3.2 future Performance verification by expert

To verify whether AttrScore constructed in the embodiments of the present application can help investors identify real experts from a population of analysts, all analysts who published more than 100 ratings of referral were selected as candidates (co 397). These analysts are then ranked according to their AttrScore and the top ranked analysts are selected as the actual expert. Subsequently, the embodiment of the present application verifies the reliability of the above identified expert by checking their future point of view quality, and the specific results are shown in fig. 8, where fig. 8 is a comparative schematic diagram of the future point of view quality of the selected expert in the embodiment of the present application, and Experience, performance and attributon are represented in order from left to right in each group of diagrams in fig. 8. In fig. 8, the benchmark labeled "experiential" refers to ranking analysts by the number of referral reports they issue, and the benchmark labeled "Performance" is ranking them according to the percentage of reliable views they issue, both benchmarks being commonly used in practice by investors. Expert recognition results obtained due to the score AttrScore according to the embodiments of the present application are shown as "attributo" in fig. 8, and future performances of analysts selected based on AttrScore are always superior to those of analyses selected by other criteria. The above results demonstrate that EDIAN-based attribution analysis can help investors identify real experts from a population of analysts.

3.4 feature importance analysis

3.4.1 localized attention visualization

The present embodiment first performs a case analysis using the same example of analyst's perspective as analyzed in fig. 7. EDAIN aggregates feature embedding within each event domain to obtain an overall representation of the corresponding domain, where learned attention weights reflect the importance of features within the event domain. For the above example, as shown in fig. 9, fig. 9 is a schematic diagram of an importance visualization based on local attention weights in the embodiment of the present application, fig. 9 (a) shows attention weights of event domains of analysts, fig. 9 (b) shows attention weights of predicted event domains, fig. 9 (c) shows attention weights of stock event domains, and fig. 9 (d) is an attention weight graph of third-order event domain interactions; fig. 9 (a) visualizes the attention weight of each feature in the analysis-event domain. From this it can be observed that stock_report is given the most attention weight by the EDIAN model, which indicates that the analyst contributes most to the overall characterization of the analyst event domain for the target stock's referral experience. In contrast, gener and team contribute relatively little to obtaining an overall characterization of the analyst event domain.

Similarly, FIG. 9 (b) shows the attention weight distribution of relevant features of this perspective example within the predicted event domain. It can be seen that std_rating and period obtain the greatest weight, while the contribution of the remaining features is smaller. Fig. 9 (c) shows the attention weight of the relevant features of this example in the stock event domain, and it can be seen that stock_id gets the highest weight, and features (including eps, pe, companysize) representing the status of the business of the marketing company get higher weight. FIG. 9 (d) then visually shows the attention weights of three provisional interactions for constructing a third order event domain interaction, it being seen that for this analyst perspective, second order interactions e are constructed first ^a∩s Then interact with the first order e ^r It is important to perform high order interactions.

3.4.2 Global attention visualization

In practice, the features analysis_id, stock_ report, broker _id and mean/per get the highest attention weight when learning the characterization of the analyst event domain, indicating that the analyst itself, the analyst's historical referral records and the brokerage company to which they belong are the most important information characterizing the analyst event domain. The possible reasons behind this are that the dealer provides the analyst with a basic working environment, while the historical referral records reflect the analyst's capabilities. In contrast, the attention weights of the features gener and team are relatively low.

Furthermore, in practice, the features std_rating and period are the primary contributors to the representation of the predicted event domain, perhaps because they are critical to inferring the expectations of the analyst. The attention weight obtained by the feature stock_id is highest in the stock event domain. Three financial indicators (including ps, eps, pe) are also important in obtaining a representation of the stock event domain. .

And in practice, when constructing a third-order event domain interaction, e will be ^a∩r And e ^s Interaction is more efficient. The reason behind this may be that the ratings of most analysts are positive, whereas the ratings criteria used by the dealer companies to which the analysts belong are generally fixed, which results in less easy changes in the analyst's referral ratings behavior.

3.5 ablation experiments

The part uses the ablation experimental result of the EDIAN model to better understand the contribution of each main component of the model.

3.5.1 event Domain interactions

Because the core of the EDIAN model is to construct event domain interactions of each order, in order to determine whether modeling event domain interactions of each order is necessary to improve performance of EDIAN, the embodiment of the application constructs variants of the following two models:

EDIAN-S/T: there is no EDIAN model of second order event domain interactions and no EDIAN model of third order event domain interactions.

EDIAN-T: there is no EDIAN model for third order event domain interactions.

Table 6 ablation experiments for event domain interactions

Table 6 reports the results of the AOQ predicted performance comparisons of the above models, from which it can be observed that EDIAN-S/T performed the worst, EDIAN performed the best, and EDIAN-T performed between the two (except for being better in accuracy than EDIAN). The result shows that constructing the second-order event domain interaction and the third-order event domain interaction jointly improves the prediction capability of the EDIAN model. Furthermore, the improvement of EDIAN-T over EDIAN-S/T is significantly higher than EDIAN-T, which suggests that second-order event domain interactions are critical to the improvement of model performance, probably because it is a "relay" hub, i.e., receiving information from a first-order interaction and passing it to a subsequent third-order interaction.

3.5.2 attention mechanism

The EDIAN model adopts an attention mechanism to realize information aggregation when constructing first-order event domain interaction and third-order event domain interaction. To verify the necessity of the attention mechanism, the present embodiments replace it with an average aggregation strategy, respectively, and thus construct variants of the following three models:

EDIAN-FA/TA: the attention mechanism is replaced by an average aggregation strategy in the first-order event domain interaction and the third-order event domain interaction respectively.

EDIAN-FA: the attention mechanism is replaced with the average aggregation policy only in the first order event domain interactions.

EDIAN-TA: the attention mechanism is replaced with the average aggregation policy only in third order event domain interactions.

TABLE 7 ablation experiments of attention mechanisms

It can be observed from table 7 that information aggregation using the attention mechanism can achieve the best performance on most metrics when modeling first and third order event domain interactions, which justifies using the attention mechanism in both modules. Furthermore, EDIAN-TA achieved optimal performance on both Recall and F1Score, and outperforms the other two variants on other metrics, indicating that the attention mechanism in first order event domain interactions is relatively more necessary.

3.6 robustness analysis

This section uses additional analysis to further study the modeling performance of the model setup on EDIAN, including in particular the structure of the MLP and the dimensions of the embedded vector.

3.6.1 hidden layer connection mode

In the EDIAN model, the default MLP architecture is an equal-width MLP with Residual Connection. Here, the embodiments of the present application also explore two other ways of connection, including Plain MLP (direct connection of adjacent hidden layers) and DenseMLP (connecting the outputs of all hidden layers before each layer as input to each layer).

As shown in fig. 10, fig. 10 is a schematic diagram illustrating the effect of the MLP Connection mode on the performance of the model in the embodiment of the present application, in which the effect of different MLP Connection modes on the EDIAN prediction performance is explored by using Accuracy, AUC and Log-loss as performance evaluation indexes, the correlation results are shown in fig. 10, fig. 10 (a) shows Accuracy, fig. 10 (b) shows AUC, and fig. 10 (c) shows Log-loss, and it can be seen from fig. 10 that the performance of the EDIAN model using Residual Connection or Dense Connection is significantly higher than Plain terms of average and variance of performance. In addition, although the variance of the performance of the Dense Connection was slightly less than Residual Connection, the average performance of Residual Connection was better than that of the Dense Connection. The above results demonstrate that EDIAN selection Residual Connection is correct.

3.6.2 depth of hidden layer

The embodiment of the application sets the hidden layer to {2,3, & gtand, 11 to explore how the number of hidden layers of the MLP affects the AOQ predictive performance of the EDIAN model. In fact, increasing the number of hidden layers increases the complexity of the model. As shown in fig. 11, fig. 11 is a schematic diagram showing the comparison of the effects of different model settings in the embodiment of the present application, where fig. 11 (a) shows the MLP depth and fig. 11 (b) shows the feature embedding dimension.

From the experimental results in fig. 11 (a), it is shown that increasing the number of layers can improve the model performance at the beginning, but the performance decreases with increasing number of layers. This is because an overly complex model is easily overfitted, setting the number of hidden layers to 9 is a good choice for the analyst's view dataset of the present embodiment.

3.6.3 feature embedding dimension

The present embodiment changes the dimension of the embedded feature from 50 to 200 in steps of 25 and reports the comparison of the predicted performance in fig. 11 (b). It can be seen that the performance of the proposed model generally increases with increasing embedding dimensions. The reason behind this is that increasing the capacity of the model enhances the representation learning ability of the model. However, using an excessively large embedding dimension may not only increase the computational burden, but may also frustrate performance. In order to simultaneously consider both predictive performance and learning efficiency, the embodiment of the present application sets the embedding dimension to 175.

Effective assessment AOQ is critical to investors making scientific decisions using analyst's views. The embodiment of the application provides a new research view angle: a machine learning model was developed to predict AOQ. To this end, embodiments of the present application propose a deep learning model EDIAN to capture expert knowledge and potential patterns implicit in analyst perspective data using information interactions. Unlike traditional methods of constructing feature level interactions, the EDIAN model constructs event domain interactions, which greatly reduces redundant and homogenous information interactions. The result of the related demonstration shows that the prediction performance of AOQ is remarkably improved by the EDIAN model. To further illustrate the superiority of the EDIAN model, the present examples demonstrate that improved predictive performance achieved by EDIAN can be translated into higher investment benefits. Furthermore, the EDIAN model also has advantages in producing meaningful AOQ attribution results, which may allow investors to better identify real experts and their views. Finally, the EDIAN model employs an attention mechanism to assess the importance of the information, which allows the model to exhibit good performance in terms of intelligibility.

Theoretically, the embodiments of the present application verify that developing a machine learning model to predict AOQ is a direction of further investigation in the field of quantitative evaluation AOQ. Furthermore, embodiments of the present application utilize an event embedding framework to represent analyst views as event domain triples (analysts, forecasts, stocks), which lay a solid foundation for the attribution analysis of AOQ. Finally, the information interaction may capture the value knowledge implicit in the analyst's perspective data, which in turn may facilitate the predictive performance of AOQ. Up to the present, the embodiment of the application is the first study to introduce deep interaction learning into AOQ evaluation task, and the proposed event domain interaction is an innovative extension of the conventional method to learn interaction on the feature level.

The foregoing detailed description of the embodiments of the present application has been provided for the purpose of illustrating the preferred embodiments of the present application and is not to be construed as limiting the scope of the present application. All equivalent changes and modifications within the scope of the present application should be made within the scope of the present application.

Claims

1. An analyst perspective quality assessment method based on event domain interactions, comprising:

the preset viewpoint quality prediction formula is as follows:

in the method, in the process of the invention,

representing a quality prediction value;

β ₀ representing a bias term;

representing a third-order event domain final interaction;

2. The event domain interaction based analyst perspective quality assessment model of claim 1, wherein step S2 comprises:

3. The event domain interaction based analyst perspective quality assessment model of claim 2, wherein step S22 comprises:

b _p =1 and b _j =0, and j+.p, determining the coding vector, and passing through a learnable transformation matrix

In the method, in the process of the invention,

representing class characteristics f _c Is the number of categories;

es represents the dimension of the embedded vector e;

an embedding matrix representing an event domain of an analyst, +.>

Feature embedding vector representing an analyst event field, < >>

An embedding matrix representing a predicted event field, +.>

Feature embedding vector representing a predicted event field, +.>

An embedded matrix representing a stock event domain, +.>

Feature embedding vector representing stock event field, +.>

In the method, in the process of the invention,

an analyst event field representing an ith view under evaluation;

a predicted event field representing an i-th perspective to be evaluated;

a stock event field representing an ith perspective to be evaluated;

Event Domain feature representation +.>

The calculation formula of (2) is as follows:

in the method, in the process of the invention,

representing attention weights corresponding to the feature embedding vectors;

b _d,1 And b _d,2 All represent bias terms, b _d,1 ∈R ^es ，b _d,2 ∈R；

ReLU represents an activation function;

q _d Representing a query vector in an attention mechanism, q _d ∈R ^es 。

4. The event domain interaction based analyst perspective quality assessment model of claim 3, wherein step S3 comprises:

s31: three event domains of the viewpoint to be evaluated are used as first-order event domains, any two event domains are selected in the three event domains to interact, and three second-order event domain interactions are determined

And->

The interaction procedure is as follows:

representation d ₁ 、d ₂ Is>

Representation d ₁ 、d ₂ Bias item of->

tanh (x) represents the activation function,

representing a ha Ma Deji operator;

And->

The interaction procedure is as follows:

in which W is _a∩r←s 、

And->

All represent model parameters that can be learned;

W _a∩r←s 、

b _a∩r←s 、b _a∩s←r and b _r∩s←a All represent model parameters that can be learned;

b _a∩r←s 、b _a∩s←r 、b _r∩s←a ∈R ^es ；

s33: initial interaction according to three third-order event domains

And->

The determination process of (2) is as follows:

In which W is _γ,1 And W is _γ,2 All represent transformation matrix, W _γ,1 ∈R ^2es×es ，W _γ,2 ∈R ^es×es ；

b _γ,1 And b _γ,2 All represent bias terms, b _γ,1 ∈R ^es ，b _γ,2 ∈R；

q _γ Representing a query vector in an attention mechanism, q _γ ∈R ^es ；

5. According to the weightsAn event domain interaction based analyst perspective quality assessment model according to claim 3, wherein the fitting function f in the model _z Comprising the following steps: seven residual multi-layer perceptrons;

in the method, in the process of the invention,

Representing a third-order event domain final interaction;

l _z representing depth representing residual multi-layer perceptron;

And, a fitting function f in the model _z Output scalar of (2)

The calculation formula of (2) is as follows:

6. the event domain interaction based analyst perspective quality assessment model of claim 5, wherein the learning objective function of the event domain interaction based analyst perspective quality assessment model is a binary cross entropy loss ζ;

wherein δ represents a Sigmoid function;

m represents the number of training samples;

y _i representing analyst's perspective x _i Is the true quality of (3);

representing all of the trainable parameters in the model.

7. An analyst perspective quality assessment model based on event domain interactions, comprising: the system comprises an event domain characterization module, a cross-domain interactive learning module and an attribution quality prediction module which are connected in sequence;

The preset viewpoint quality prediction formula is as follows:

in the method, in the process of the invention,

representing a quality prediction value;

β ₀ representing a bias term;

representing a third-order event domain final interaction;