Method and apparatus for predicting based on multi-source heterogeneous data

• US 9,760,832 B2
• Filed: 12/29/2014
• Issued: 09/12/2017
• Est. Priority Date: 08/27/2014
• Status: Active Grant

• Current Assignee: Baidu Online Network Technology (Beijing) Co., Ltd (Baidu Incorporated)
• Original Assignee: Baidu Online Network Technology (Beijing) Co., Ltd (Baidu Incorporated)
• Inventors: Wang, Qing, Hu, Wei, Shen, Zhiyong
• Primary Examiner(s): Holmes, Michael B
• Application Number:
  US14/585,136
• Publication Number:
  US 20160063383A1
• Time in Patent Office: 988 Days
• Field of Search: 706 52
• CPC Class Codes: G06F 17/18 for evaluating statistical data , e.g. average values, frequency distributions, probability functions, regression analysis forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04
• CPC Class Codes: G06N 5/04 Inference or reasoning models
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G06Q 50/34 Betting or bookmaking, e.g. Internet betting
• CPC Class Codes: G07F 17/3288 Betting, e.g. on live events, bookmaking

Abstract

A method and apparatus for predicting based on multi-source heterogeneous data. The method comprises: acquiring, with regard to an event of a set type, at least two types of historical data that can reflect an event result; establishing a joint likelihood model of attribute data of the event of the set type and the historical data; determining an optimal estimation of the attribute data according to a maximum posterior principle; and determining, based on a probability distribution associated with the attribute data in the joint likelihood model, a parameter in the probability distribution as a prediction result of a predicted event of the set type. Some embodiments use a hierarchical model to introduce data of different sources into different data layers, unify heterogeneous data in a joint likelihood model to perform analysis, and obtain a more accurate, instant and stable prediction result through effective fusion.

Claims

1. A computer-implemented method for event predicting via machine learning based on multi-source heterogeneous data, comprising:

acquiring at least two types of historical data associated with an event result for a first event of a predetermined type;

establishing a joint likelihood model of attribute data of the first event and the at least two types of historical data;

determining an optimal estimation of the attribute data based on the joint likelihood model according to a maximum posterior principle; and

determining a parameter in a probability distribution as a prediction result of a second event based on the probability distribution associated with the attribute data in the joint likelihood model, wherein the joint likelihood model includes one or more adjustment parameters for correcting the joint likelihood model, the adjustment parameters being determined iteratively based on an accuracy of the prediction result.

2. The method of claim 1, wherein said determining the parameter comprises determining the parameter in the probability distribution as the prediction result of an event to be predicted based on the probability distribution.

3. The method of claim 2, wherein a preselected type of the second event and the predetermined type of the first event are identical.

4. The method of claim 1, wherein said establishing comprises establishing the joint likelihood model according to a relationship between the at least two types of historical data and the attribute data and a correction function for correcting the relationship to satisfy a normal distribution.

5. The method of claim 4, wherein the relationship between the at least two types of historical data and the attribute data comprises a Poisson distribution function.

6. The method of claim 4, wherein the relationship between the at least two types of historical data and the attribute data comprises a gamma distribution function.

7. The method of claim 1, wherein the one or more adjustment parameters include one or more state adjustment parameters, a preset weight value, a first preset adjustment parameter value, a second preset adjustment parameter value, a third preset adjustment parameter value, or a combination thereof, the method further comprising:

acquiring, with regard to a match, a historical score and historical odds that can reflect a match result as the two types of historical data;

with regard to a match m, determining a home team goal parameter λ

_m,1 and a visiting team goal parameter λ

_m2 of a Poisson distribution of the match m using home odds p_m,1, average odds p_m,2 and visiting odds p_m,3 of the match m based on a relational expression as follows;

8. The method of claim 7, wherein said determining the parameter in the probability distribution comprises:

with regard to a match c to be predicted, determining a home team distribution parameter λ

_c,1 and a visiting team distribution parameter λ

_c,2 of the match c to be predicted according to offensive/defensive capability parameters θ

of a home team i and a visiting team j participating in the match based on relational expressions as follows;

λ

_c,1˜

Gamma(exp(b_kc+α

_ic−

d_jc), β

)

λ

_c,2˜

Gamma(exp(b<sub>k′

c</sub>+α

_jc−

d_ic), β

)

where b_kc is a home team correction coefficient determined according to the match c to be predicted, α

_ic is an offensive capability parameter of the home team i, d_jc is a defensive capability parameter of the visiting team j, b<sub>k′

c</sub> is a visiting team correction coefficient determined according to the match c to be predicted, α

_jc is an offensive capability parameter of the visiting team j, and d_ic is a defensive capability parameter of the home team;

with regard to the match c to be predicted, determining odds p_c,1, p_c,2 and p_c,3 according to the home team distribution parameter λ

_c,1 and visiting team distribution parameter λ

_c,2 based on a relational expression as follows;

9. The method of claim 7, further comprising performing normalization processing on the historical odds.

10. The method of claim 9, wherein said performing the normalization processing occurs after said after acquiring, with regard to the match, the historical score and historical odds.

11. An apparatus for event predicting via machine learning based on multi-source heterogeneous data, comprising:

a processor; and

a memory having one or more programs stored thereon for instructing said processor, the one or more programs including;

instruction for acquiring, with regard to an event of a set type, at least two types of historical data that can reflect an event result;

instruction for establishing a joint likelihood model of attribute data of the event of the set type and the at least two types of historical data and determining an optimal estimation of the attribute data based on the joint likelihood model according to a maximum posterior principle; and

instruction for determining, with regard to an event to be predicted which is of the set type, based on a probability distribution associated with the attribute data in the joint likelihood model, a parameter in the probability distribution as a prediction result of the event to be predicted, wherein the joint likelihood model includes one or more adjustment parameters for correcting the joint likelihood model, the adjustment parameters being determined iteratively based on an accuracy of the prediction result.

12. The apparatus of claim 11, wherein the one or more programs include instruction for establishing the joint likelihood model according to a relationship between the at least two types of historical data and the attribute data and a correction function for correcting the relationship to satisfy a normal distribution.

13. The apparatus of claim 12, wherein the relationship between the historical data and the attribute data comprises a Poisson distribution function.

14. The apparatus of claim 12, wherein the relationship between the historical data and the attribute data comprises a gamma distribution function.

15. The apparatus of claim 11, wherein the one or more adjustment parameters include one or more state adjustment parameters, a preset weight value, a first preset adjustment parameter value, a second preset adjustment parameter value, a third preset adjustment parameter value, or a combination thereof, wherein the one or more programs include:

instruction for acquiring, with regard to a match, a historical score and historical odds which can reflect a match result as two types of historical data;

instruction for;

with regard to a match m, determining a home team goal parameter λ

_m,1 and a visiting team goal parameter λ

_m,2 of the Poisson distribution of said match m using home odds p_m,1, average odds p_m,2 and visiting odds p_m,3 of said match m based on a relational expression as follows;

16. The apparatus of claim 14, wherein the one or more programs include:

instruction for, with regard to a match c to be predicted, determining a home team distribution parameter λ

_c,1 and a visiting team distribution parameter λ

_c,2 of said match c to be predicted according to offensive/defensive capability parameters θ

of a home team i and a visiting team j participating in the match based on relational expressions as follows;

θ

_c,1˜

Gamma(exp(b_kc+α

_ic−

d_jc),β

)

θ

_c,2˜

Gamma(exp(b<sub>k′

c</sub>+α

_jc−

d_ic),β

)

where b_kc is a home team correction coefficient determined according to said match c to be predicted, α

_ic is an offensive capability of the home team i, d_jc is a defensive capability of the visiting team j, b<sub>k′

c</sub> is a visiting team correction coefficient determined according to said match c to be predicted, α

_jc is an offensive capability of the visiting team j, and d_ic is a defensive capability of the home team;

instruction for, with regard to the match c to be predicted, determining odds p_c,1, p_c,2 and p_c,3 according to said home team distribution parameter λ

_c,1 and visiting team distribution parameter λ

_c,2 based on a relational expression as follows;

17. The apparatus of claim 14, wherein the one or more programs include instruction for performing normalization processing on the historical odds after acquiring, with regard to a match, a historical score and historical odds which can reflect a match result as two types of historical data.

18. A non-transitory computer storage medium including at least one program for event predicting via machine learning based on multi-source heterogeneous data when implemented by a processor, comprising:

instruction for acquiring at least two types of historical data associated with an event result for a first event of a predetermined type;

instruction for establishing a joint likelihood model of attribute data of the first event and the at least two types of historical data;

instruction for determining an optimal estimation of the attribute data based on the joint likelihood model according to a maximum posterior principle; and

instruction for determining a parameter in a probability distribution as a prediction result of a second event based on the probability distribution associated with the attribute data in the joint likelihood model, the second event to be predicted based on the probability distribution and having a preselected type of the second event that is identical to the predetermined type of the first event, wherein the joint likelihood model includes one or more adjustment parameters for correcting the joint likelihood model, the adjustment parameters being determined iteratively based on an accuracy of the prediction result.

19. The computer storage medium of claim 18, wherein said instruction for establishing comprises instruction for establishing the joint likelihood model according to a relationship between the at least two types of historical data and the attribute data and a correction function for correcting the relationship to satisfy a normal distribution.

20. The computer storage medium of claim 19, wherein the relationship between the at least two types of historical data and the attribute data comprises at least one of a Poisson distribution function and a gamma distribution function.