WO2022126873A1 - Intelligent financial information recommendation system - Google Patents

Abstract

The present invention relates to an intelligent financial information recommendation system, comprising a news feature vector calculation module configured to calculate a feature vector of each piece of news; a user feature vector calculation module configured to calculate a feature vector of each user; a multi-dimensional news pool creation module configured to create news pools in a plurality of dimensions, and to sort news in the news pools; and a news recommendation module configured to calculate a sampling probability of each news pool, to sample each news pool according to the sampling probability, and to recommend, to the users, the top news in the news pool obtained by means of sampling. According to the present invention, the problems in the prior art that the information recommendation function is single and that information that has diversity, novelty, a high accuracy rate, personalization, and high time efficiency at the same time cannot be recommended are solved. According to the present invention, news information that conforms to the investment preference of a user can be actively pushed to the user, all kinds of content-rich timely information can also be provided to the user and rapidly changing opportunities of investment can also be captured in a timely manner.

Description

An intelligent recommendation system for financial information technical field

The invention relates to the technical field of information recommendation, in particular to an intelligent recommendation system for financial information.

Background technique

With the development of economic society and related technologies such as the Internet, various news and information in the financial field have exploded. Faced with the massive amount of information generated every day, it is difficult for investors to find the useful information they need and make investment decisions quickly based on this information. In order to allow investors to quickly obtain financial information that is useful to them, it is a good way to actively recommend news information that meets investors' preferences through the news recommendation system. However, due to the strong timeliness of financial information (such as emergencies and sudden investment hotspots) and multi-dimensionality (such as macro and micro), it is difficult for traditional recommendations based on relevant content to provide timely and useful information to investors. Investment information and investment opportunities.

The existing recommendation algorithms mainly include: recommendation algorithm based on content relevance, recommendation algorithm based on collaborative filtering, recommendation algorithm based on popularity, and recommendation algorithm based on model.

Recommendation algorithm based on content relevance: By analyzing user behavior and the text content of news, users and news are marked with keywords (subject headings) that can represent the characteristics of users and news, and then these keywords are calculated by tf-idf and other words. The weighting algorithm constitutes the feature vectors of users and news respectively, and these feature vectors represent the characteristics of each user and each news. Then, the similarity between the user vector and the news vector is calculated by methods such as cosine similarity, and then the news that is more similar to the user is recommended to the user according to these similarities. One of the biggest drawbacks of the recommendation algorithm based on content relevancy is that it is easy to generate homogeneous recommendation, that is, it always recommends news of the same type of content to users and enters an infinite loop, losing the diversity and novelty of the recommended content.

Recommendation algorithm based on collaborative filtering: analyze each user's evaluation of the item (through browsing records and purchase records, etc.); calculate the similarity between all users based on the user's evaluation of the item; select the N most similar to the current user. users; recommend the items that the N users have the highest evaluation and the current user has not browsed to the current user. The recommendation algorithm based on collaborative filtering also has many shortcomings: 1. The cold start problem, that is, for new users and new items, the recommendation cannot be carried out; 2. The accuracy of the algorithm depends on a large amount of accurate user data, if the user data is small. 3. In some systems with short item life cycles (such as news and advertisements), due to the fast update speed, a large number of items will not have user ratings, resulting in a sparse rating matrix, which is not conducive to Recommendations for these content.

Popularity-based recommendation algorithm: Sort by popularity according to data such as clicks, page views, unique visitors, and sharing rates, and recommend them to users. The advantage of this algorithm is that it is simple and suitable for new users who have just registered. The disadvantage is that it cannot provide personalized recommendations for users.

Model-based recommendation algorithm: build a model through methods such as machine learning, and then use a large amount of existing user behavior data, purchase records and various user characteristics to train and fit the built model, and then apply the training A good model inputs the feature attributes of the user to be recommended, and the model outputs the final recommendation result. The disadvantage of this algorithm is that it requires a large amount of user historical behavior data, and requires repeated manual intervention for attribute combination and screening (ie, feature engineering). At the same time, due to the timeliness of news, the model also needs to be repeatedly trained and updated to adapt to changes.

Since financial information is highly timely and multi-dimensional, and different investors have different investment preferences and investment logics, it is difficult for any of the above recommendation algorithms to provide investors with timely and useful investment information and investment opportunities. .

Therefore, it is necessary to provide an intelligent financial information recommendation system to solve the problem that the information recommendation function in the prior art is single, and it is impossible to recommend information with diversity, novelty, high accuracy, personalization and strong timeliness.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an intelligent financial information recommendation system to solve the problem of single information recommendation function in the prior art, unable to recommend information with diversity, novelty, high accuracy, personalization and strong timeliness.

In order to solve the problems existing in the prior art, the present invention provides a financial information intelligent recommendation system, including:

A news feature vector calculation module, configured to calculate the feature vector of each news;

a user feature vector calculation module, configured to calculate the feature vector of each user;

The multi-dimensional news pool creation module is configured to create news pools of multiple dimensions and sort the news in each news pool;

The news recommendation module is configured to calculate the sampling probability of each news pool, sample each news pool according to the sampling probability, and recommend the news ranked first in the sampled news pool to the user.

Optionally, in the financial information intelligent recommendation system, the method of calculating the feature vector of each news is:

Extract all the subject words in each news in the database;

Calculate the weight of each subject term and the dynamic inverse document frequency within a period;

Calculate the feature vector V of each news, the feature vector V is an N-dimensional vector, N is the total number of all the topic words in the database, and one bit of the feature vector V corresponds to a topic word, wherein, the bit of the vector corresponding to any topic word in each news The value is equal to the product of the weight of the keyword and the dynamic inverse document frequency within a period of the keyword.

Optionally, in the financial information intelligent recommendation system,

A cycle consists of 20, 30 or 40 days;

The inverse document frequency is Inverse Document Frequency, which is idf.

Optionally, in the financial information intelligent recommendation system, in the user feature vector calculation module, the calculation method of each user's feature vector is:

V _s =Normalize(V _p )-B·Normalize(V _d )·η(||V _d || ₂ )-E·V _t ;

Among them, V _s is the feature vector of each user, V _p is the feature vector of the user's historical reading news, V _d is the feature vector of the news that the user has clicked and disliked, and V _t is the feature vector of the keyword that the user has clicked and disliked, ||V _d || ₂ is the 2-norm of the eigenvector V _d , Normalize(V _p ) and Normalize(V _d ) are the normalized vectors of the eigenvectors V _p and V _d respectively, B and E are the calculation parameter, η is the penalty function of the number of news.

Optionally, in the financial information intelligent recommendation system,

V _p , V _d and V _t are all N-dimensional feature vectors, N is the total number of all subject words in the database, and one bit of the feature vector corresponds to a subject word;

Any one of the _Vp feature vectors is equal to the weight of the corresponding topic word in the user's historical reading news multiplied by the dynamic inverse document frequency of the topic word within a period;

Any one of the V _d feature vectors is equal to the weight of the corresponding topic word in the news that the user has clicked and disliked multiplied by the dynamic inverse document frequency of the topic word within a period;

Any one of the V _t feature vectors is equal to the weight of the keyword that the user has clicked and disliked multiplied by the dynamic inverse document frequency within a period of the keyword.

Optionally, in the financial information intelligent recommendation system, the dimension types in the multi-dimensional news pool creation module include but are not limited to: macro, pre-market, afternoon comment, capital inflow and outflow, investment hotspot, hot topic news, click list, Self-selected stocks, international current affairs and finance, fund channels, Hong Kong stocks and external markets.

Optionally, in the financial information intelligent recommendation system,

Hot topic news pool, first calculate the hot probability of each hot topic in the hot topic news pool, sample each hot topic in the hot topic news pool according to the hot probability, and sort the latest news corresponding to the sampled hot topics in order;

The news in the click list news pool is sorted according to the user's click volume;

The news in the news pool with strong timeliness is sorted in reverse order according to the release time of the news, that is, the news with the newer release time is ranked in the front, and the news pool with strong timeliness includes pre-market, afternoon comments, and capital inflows and outflows;

The news in the rest of the news pools are sorted according to the similarity of the feature vector of the news and the feature vector of the user, and sorted according to the degree of similarity.

Optionally, in the financial information intelligent recommendation system, the calculation method of the hot probability of each hot topic in the hot topic news pool is:

Calculate the hot probability of each hot topic;

Normalize the calculated hotspot probability of hot topics;

The calculation method of the hotspot probability of each hot topic is:

The calculation method of normalizing the hotspot probability of hot topics is as follows:

Among them, K is the number of hot topics to be sampled, j is the hot topic ranked jth in the hot value,

is the unnormalized sampling probability of hot topic j, power is the calculation parameter, q(j) is the normalized sampling probability of hot topic j, h _j is the hotness value of hot topic j, and h _K is the K-th The hotness value of the hot topic, that is, h _K is the hotness value of the hot topic corresponding to the lowest hot topic in the hot topic.

Optionally, in the financial information intelligent recommendation system, the method of calculating the sampling probability of each news pool is:

Calculate the similarity probability between the financial information user to be recommended and the news ranked first in each news pool;

define additional probabilities;

Calculate the sampling probability of each news pool according to the similarity probability and the additional probability;

sampling each news pool according to the sampling probability;

The news that ranks first in the sampled news pool is recommended to the user.

Optionally, in the financial information intelligent recommendation system,

The similarity probability is calculated as:

Among them, i represents any news pool, Padjust _i is the similarity probability between the financial information user to be recommended and the news ranked first in the news pool, sim _i is the financial information user to be recommended and the news ranked first in the news pool. The similarity of the news in bits, Primary _i is the preset initial probability of the news pool, C and ε are the calculation parameters;

The additional probability is calculated as:

Among them, i represents any news pool, Padditional _i is the normalized additional probability of the news pool, and m is the total number of current news pools that can be sampled;

The sampling probability is calculated as:

Among them, i represents any news pool, P _i is the sampling probability of the news pool, Padjust _i is the similarity probability between the financial information user to be recommended and the news ranked first in the news pool, Padditional _i is the news pool return The normalized additional probability.

In the financial information intelligent recommendation system provided by the present invention, by establishing news pools of various dimensions that meet the characteristics of the financial field and investment needs, it is possible to push multi-dimensional, multi-level and diverse news to users; The statistical method of probability can dynamically sample news from multiple dimensions (multiple news pools). The sampling probability can be dynamically determined by factors such as user similarity (ie user preference), news popularity and business logic. While pushing news that conforms to user preferences and interests to users, it can also push accurate, timely and useful investment information and investment opportunities to users.

Description of drawings

1 is a block diagram of a financial information intelligent recommendation system provided by an embodiment of the present invention;

FIG. 2 is a recommendation flowchart in the financial information intelligent recommendation system provided by an embodiment of the present invention;

Detailed ways

The specific embodiments of the present invention will be described in more detail below with reference to the schematic diagrams. The advantages and features of the present invention will become more apparent from the following description. It should be noted that, the accompanying drawings are all in a very simplified form and in inaccurate scales, and are only used to facilitate and clearly assist the purpose of explaining the embodiments of the present invention.

In the following, if a method described herein includes a series of steps, the order of the steps presented herein is not necessarily the only order in which the steps may be performed, and some of the steps described may be omitted and/or some not described herein Additional steps can be added to the method.

Because financial information is highly timely and multi-dimensional, and different investors have different investment preferences and investment logics, it is difficult for existing recommendation algorithms to provide investors with timely and useful investment information and investment opportunities.

Therefore, it is necessary to provide a financial information intelligent recommendation system, as shown in Figures 1 and 2. Figure 1 is a block diagram of a financial information intelligent recommendation system provided by an embodiment of the present invention; Figure 2 is a financial information intelligent recommendation system provided by an embodiment of the present invention. The recommendation flow chart in the recommendation system, the financial information intelligent recommendation system includes:

A news feature vector calculation module, configured to calculate the feature vector of each news;

a user feature vector calculation module, configured to calculate the feature vector of each user;

The multi-dimensional news pool creation module is configured to create news pools of multiple dimensions and sort the news in each news pool;

The news recommendation module is configured to calculate the sampling probability of each news pool, sample each news pool according to the sampling probability, and recommend the news ranked first in the sampled news pool to the user.

The present invention solves the problem of single information recommendation function in the prior art, unable to recommend information with diversity, novelty, high accuracy, personalization and strong timeliness. It can also provide users with a variety of content-rich and timely information to capture the ever-changing investment opportunities in a timely manner.

Specifically, the way to calculate the feature vector of each news is:

Extract all the subject words in each news in the database through algorithms such as TextRank;

Calculate the weight of each subject word and the dynamic inverse document frequency in a period, wherein the weight of each subject word can be obtained according to calculation or experience, a period includes 20 days, 30 days or 40 days, preferably a period is 30 days, The inverse document frequency is Inverse Document Frequency, ie idf;

Finally, the feature vector V of each news is calculated, the feature vector V is an N-dimensional vector, N is the total number of all subject words in the database, and one bit of the feature vector V corresponds to a subject word, wherein, the vector corresponding to any subject word in each news The bit value is equal to the product of the weight of the topic word and the dynamic inverse document frequency within a period of the topic word, which is expressed by the formula: V[w]=theme_weight(w)×idf(w), where V[w] is The bit value of the vector corresponding to any topic word in each news, theme_weight(w) is the weight of the topic word, and idf(w) is the dynamic inverse document frequency of the topic word within a period.

Further, in the user feature vector calculation module, the calculation method of the feature vector of each user is:

V _s= Normalize(V _p )-B·Normalize(V _d )·η(||V _d || ₂ )-E·V _t ;

Among them, V _s is the feature vector of each user, V _p is the feature vector of the user's historical reading news, V _d is the feature vector of the news that the user has clicked and disliked, and V _t is the feature vector of the keyword that the user has clicked and disliked, ||V _d || ₂ is the 2-norm of the feature vector V _d , Normalize(V _p ) and Normalize(V _d ) are the normalized vectors of the feature vectors V _p and V _d respectively, B and E are the calculation parameter, η is the penalty function of the number of news.

Further, the calculation formula of ||V _d || ₂ is:

Among them, v _d1 , v _d2 ,..., v _dr are r components of V _d ;

The calculation formulas of Normalize(V _p ) and Normalize(V _d ) are:

Normalize(V _d )=[v _d1 /||V _d || ₂ , v _d2 /||V _d || ₂ , …, v _dr /||V _d || ₂ ];

Among them, v _p1 , v _p2 ,..., v _pr are the r components of V _p , v _d1 , v _d2 ,..., v _dr are the r components of V _d , ||V _p || ₂ is the feature vector V _p The 2-norm of , ||V _d || ₂ is the 2-norm of the feature vector V _d .

The calculation formula of η(||V _d || ₂ ) is:

Among them, θ and α are the calculation parameters. The reason for designing the penalty function is: when the user first uses the function of disliking news, he only clicks a few news. At this time, due to the small vector norm, the normalization After transformation, each component is too large, resulting in too strong shielding at the beginning, so it is necessary to multiply the vector by a small penalty function to make each component smaller. With the increase in the use of the dislike news function, the norm of the dislike news vector increases, so that η(||V _d || ₂ ) approaches 1, and the punishment will be reduced quickly.

Preferably, among the users, the user's historical reading news, the news that the user has clicked and disliked, and the subject words that the user has clicked and disliked will decay before updating the user's feature vector, or decay over time, so that new users can The behavior is given a greater weight in order to adapt to the changes in the user's reading interest, and the news corresponding to the news that has been pushed, the news that the user has clicked and disliked, and the subject words that the user has clicked dislike can also be removed from the recommendation process. .

Still further, in the financial information intelligent recommendation system,

The following V _p , V _d and V _t are all N-dimensional feature vectors, N is the total number of all subject words in the database, and one bit of the feature vector corresponds to a subject word;

Any one of the _Vp feature vectors is equal to the weight of the corresponding topic word in the user's historical reading news multiplied by the dynamic inverse document frequency of the topic word within a period;

Any one of the V _d feature vectors is equal to the weight of the corresponding topic word in the news that the user has clicked and disliked multiplied by the dynamic inverse document frequency of the topic word within a period;

Any one of the V _t feature vectors is equal to the weight of the keyword that the user has clicked and disliked multiplied by the dynamic inverse document frequency within a period of the keyword.

In the financial information intelligent recommendation system, in order to push news with rich content and different dimensions to users, the present invention automatically divides news into multiple categories through a classification algorithm or classifier. Therefore, the dimension types in the dimension news pool creation module include: But not limited to: macro, pre-market, afternoon review, capital inflows and outflows, investment hotspots, hot topic news, click list, self-selected stocks, international current affairs and finance, fund channels, Hong Kong stocks and external markets.

Preferably, in the hot topic news pool, the hot probability of each hot topic in the hot topic news pool is first calculated, and each hot topic in the hot topic news pool is sampled according to the hot probability, and the latest hot topic corresponding to the sampled hot topic is sampled. The news is sorted in order. The higher the popularity value of the hot topic, the greater the sampling probability, and the greater the possibility that the news of the hot topic will be recommended first; the news in the click list news pool is sorted according to the user's click volume, and the higher the click volume The news indicates that the news is more popular or important, and the ranking is higher; the news in the news pool with strong timeliness is sorted in reverse order according to the release time of the news, that is, the news with the newer release time is ranked first, so as to ensure that Major financial news can be pushed to users as soon as possible. News pools with strong timeliness include pre-market, afternoon reviews, and capital inflows and outflows; the news in other news pools are sorted according to the similarity between the feature vector of the news and the feature vector of the user, and the similarity is based on the similarity. The higher the similarity is, the better the news is in line with the user's reading or investment preferences, and the higher the ranking is, so that the news recommendation module can give priority to pushing these news to the user as much as possible. For each news of , the cosine similarity between each news and each user is calculated through the news feature vector and the user feature vector.

Further, the calculation method of the hot probability of each hot topic in the hot topic news pool is:

Calculate the hot probability of each hot topic;

Normalize the calculated hotspot probability of hot topics;

Among them, the calculation method of the hotspot probability of each hot topic is:

The calculation method of normalizing the hotspot probability of hot topics is as follows:

Among them, K is the number of hot topics to be sampled, j is the hot topic ranked jth in the hot value,

is the unnormalized sampling probability of hot topic j, power is the calculation parameter, q(j) is the normalized sampling probability of hot topic j, h _j is the hotness value of hot topic j, and h _K is the K-th The hotness value of the hot topic, that is, h _K is the hotness value of the hot topic corresponding to the lowest hot topic in the hot topic.

Optionally, in the financial information intelligent recommendation system, the news recommendation module is equivalent to the brain center of the entire financial information intelligent recommendation system, and the news recommendation module adopts probability sampling technology combined with time to determine when and to which users what news is finally recommended. . Further, please continue to refer to Figure 2. The way to calculate the sampling probability of each news pool is:

Calculate the similarity probability between the financial information user to be recommended and the news ranked first in each news pool;

define additional probabilities;

Calculate the sampling probability of each news pool according to the similarity probability and the additional probability;

sampling each news pool according to the sampling probability;

The news that ranks first in the sampled news pool is recommended to the user.

Optionally, in the financial information intelligent recommendation system,

The similarity probability is calculated as:

Among them, i represents any news pool, Padjust _i is the similarity probability between the financial information user to be recommended and the news ranked first in the news pool, sim _i is the financial information user to be recommended and the news ranked first in the news pool. The similarity of bits of news, Pinitial _i is the preset initial probability of the news pool, the specific value of Pintial _i is determined according to the business scenario and the importance of each news pool, C and ε are calculation parameters; the sampling probability of each news pool will be It increases rapidly with the increase of the similarity probability of the first-ranked news, that is, if the similarity probability of the first-ranked news of a certain news pool is significantly larger than that of the first-ranked news of other news pools. If the probability is similar, the news in the news pool will be preferentially sampled, because a large similarity indicates that the news is more in line with the user's reading and investment preferences.

Further, because the news in the hot topic news pool and the click list news pool may not be very similar to the user's feature vector, in order to let the user know the current hot spots in time or let the user discover new points of interest, it is also necessary to send Users actively push this type of news, that is, the present invention defines an additional probability to increase the recommendation rate of this type of news.

The additional probability is calculated as:

Among them, i represents any news pool, Padditional _i is the normalized additional probability of the news pool, m is the total number of current news pools that can be sampled; the additional probability is the normalized probability, so this additional probability will not be affected by The influence of the sampling probability of other news pools ensures that such news can be actively pushed to users under any circumstances, and its push strength mainly depends on this additional probability.

The sampling probability is calculated as:

Among them, i represents any news pool, P _i is the sampling probability of the news pool, Padjust _i is the similarity probability between the financial information user to be recommended and the news ranked first in the news pool, Padditional _i is the news pool return After normalizing the additional probability, after the sampling probability of each news pool is obtained, each news pool is sampled according to the sampling probability, and the news ranked first in the sampled news pool is recommended to the user.

Finally, the news recommendation module recommends news to users according to the following logic and steps: For news pools with strong timeliness (such as pre-market, afternoon comment, and next week's outlook, etc.), set a specific push time, such as the pre-market push time is From 0:00 on the stock market trading day to before the opening, the afternoon is rated as 11:30-13:00 on the stock market trading day. During these time periods, the news will be pushed to users first, and these news will not be pushed to users in other time periods; then, according to the above calculation According to the sampling probability of each news pool, the news pool is sampled. If a news pool is sampled, the first news in the news pool will be pushed to the user first.

Please refer to FIG. 3 . FIG. 3 is a display diagram of financial information recommendation provided by an embodiment of the present invention. As can be seen from the figure, the recommendation system provided by the present invention pushes news information that meets the user's interest preference (house price and real estate), It also pushes various content-rich and timely news information (such as current hot events, the latest important macroeconomic data and unexpected financial events, etc.) to users.

To sum up, in the financial information intelligent recommendation system provided by the present invention, by establishing news pools of various dimensions that meet the characteristics of the financial field and investment needs, it is possible to push multi-dimensional, multi-level and diverse news to users; The invention dynamically samples news from multiple dimensions (multiple news pools) through the statistical method of sampling probability, and the sampling probability can be dynamically determined by factors such as user similarity (ie user preference), news popularity and business logic. , so that it can push accurate, timely and useful investment information and investment opportunities to users while pushing news that conforms to users' preferences and interests.

The above are only preferred embodiments of the present invention, and do not have any limiting effect on the present invention. Any person skilled in the art, within the scope of not departing from the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, all belong to the technical solution of the present invention. content still falls within the protection scope of the present invention.

Claims (8)

1. An intelligent recommendation system for financial information, comprising:

   A news feature vector calculation module, configured to calculate the feature vector of each news;

   a user feature vector calculation module, configured to calculate the feature vector of each user;

   The multi-dimensional news pool creation module is configured to create news pools of multiple dimensions and sort the news in each news pool;

   The news recommendation module is configured to calculate the sampling probability of each news pool. The method of calculating the sampling probability of each news pool is: calculate the similarity probability between the financial information user to be recommended and the news ranked first in each news pool, and define the additional probability , calculate the sampling probability of each news pool according to the similarity probability and the additional probability, sample each news pool according to the sampling probability, and recommend the news ranked first in the sampled news pool to the user;

   Among them, the calculation method of similarity probability is:

   

   Among them, i represents any news pool, Padjust _i is the similarity probability between the financial information user to be recommended and the news ranked first in the news pool, sim _i is the financial information user to be recommended and the news ranked first in the news pool. The similarity of the news in bits, Primary _i is the preset initial probability of the news pool, C and ε are the calculation parameters;

   The additional probability is calculated as:

   

   Among them, i represents any news pool, Padditional _i is the normalized additional probability of the news pool, and m is the total number of current news pools that can be sampled;

   The sampling probability is calculated as:

   

   Among them, i represents any news pool, P _i is the sampling probability of the news pool, Padjust _i is the similarity probability between the financial information user to be recommended and the news ranked first in the news pool, Padditional _i is the news pool return The normalized additional probability.
2. The financial information intelligent recommendation system as claimed in claim 1, is characterized in that, the mode of calculating the feature vector of each news is:

   Extract all the subject words in each news in the database;

   Calculate the weight of each subject term and the dynamic inverse document frequency within a period;

   Calculate the feature vector V of each news, the feature vector V is an N-dimensional vector, N is the total number of all the topic words in the database, and one bit of the feature vector V corresponds to a topic word, wherein, the bit of the vector corresponding to any topic word in each news The value is equal to the product of the weight of the keyword and the dynamic inverse document frequency within a period of the keyword.
3. The financial information intelligent recommendation system according to claim 2, wherein,

   A cycle consists of 20, 30 or 40 days;

   The inverse document frequency is Inverse Document Frequency, which is idf.
4. The financial information intelligent recommendation system according to claim 3, wherein, in the user feature vector calculation module, the calculation method of the feature vector of each user is:

   V _s =Normalize(V _p )-B·Normalize(V _d )·η(||V _d || ₂ )-E·V _t ;

   Among them, V _s is the feature vector of each user, V _p is the feature vector of the user's historical reading news, V _d is the feature vector of the news that the user has clicked and disliked, and V _t is the feature vector of the keyword that the user has clicked and disliked, ||V _d || ₂ is the 2-norm of the feature vector V _d , Normalize(V _p ) and Normalize(V _d ) are the normalized vectors of the feature vectors V _p and V _d respectively, B and E are the calculation parameter, η is the penalty function of the number of news.
5. The financial information intelligent recommendation system according to claim 4, wherein,

   V _p , V _d and V _t are all N-dimensional feature vectors, N is the total number of all subject words in the database, and one bit of the feature vector corresponds to a subject word;

   Any one of the _Vp feature vectors is equal to the weight of the corresponding topic word in the user's historical reading news multiplied by the dynamic inverse document frequency of the topic word within a period;

   Any one of the V _d feature vectors is equal to the weight of the corresponding topic word in the news that the user has clicked and disliked multiplied by the dynamic inverse document frequency of the topic word within a period;

   Any one of the V _t feature vectors is equal to the weight of the keyword that the user has clicked and disliked multiplied by the dynamic inverse document frequency within a period of the keyword.
6. The financial information intelligent recommendation system according to claim 1, wherein the dimension types in the multi-dimensional news pool creation module include but are not limited to: macro, pre-market, afternoon comment, capital inflow and outflow, investment hotspot, hot topic news, Click list, self-selected stocks, international current affairs and finance, fund channel, Hong Kong stocks and external markets.
7. The financial information intelligent recommendation system according to claim 6, wherein,

   Hot topic news pool, first calculate the hot probability of each hot topic in the hot topic news pool, sample each hot topic in the hot topic news pool according to the hot probability, and sort the latest news corresponding to the sampled hot topics in order;

   The news in the click list news pool is sorted according to the user's click volume;

   The news in the news pool with strong timeliness is sorted in reverse order according to the release time of the news, that is, the news with the newer release time is ranked in the front, and the news pool with strong timeliness includes pre-market, afternoon comments, and capital inflows and outflows;

   The news in the rest of the news pools are sorted according to the similarity of the feature vector of the news and the feature vector of the user, and sorted according to the degree of similarity.
8. The financial information intelligent recommendation system according to claim 7, wherein the calculation method of the hot probability of each hot topic in the hot topic news pool is:

   Calculate the hot probability of each hot topic;

   Normalize the calculated hotspot probability of hot topics;

   The calculation method of the hotspot probability of each hot topic is:

   

   The calculation method of normalizing the hotspot probability of hot topics is as follows:

   

   Among them, K is the number of hot topics to be sampled, j is the hot topic ranked jth in the hot value,

   

   is the unnormalized sampling probability of hot topic j, power is the calculation parameter, q(j) is the normalized sampling probability of hot topic j, h _j is the hotness value of hot topic j, and h _K is the K-th The hotness value of the hot topic, that is, h _K is the hotness value of the hot topic corresponding to the lowest hot topic in the hot topic.

Images