CN107180078A - A kind of method for vertical search based on user profile learning - Google Patents

Abstract

The invention discloses a vertical search method based on user interest learning, specifically: according to the content of each attribute filled in by the user for the user's personal interest needs when registering, determine the user's interest preference, and establish an initial user interest model; if the user presses Filter query by attribute category, then add the attribute category query record to the user's personal interest model to complete the model update; calculate the user's interest degree value for all entity resources, and then sort and store according to the interest degree value from high to low; personalized search Then filter the query by keyword or attribute selected by the user to generate an initial query result list; for the entity resources in the initial query result list, combined with the user's interest in these entity resources, sort the entity resources in the list by The interest degree values are sorted from high to low to generate the final list of query results. The present invention actively provides the result consistent with the user's current interest through the learning of the user's interest.

Description

A Vertical Search Method Based on User Interest Learning

technical field

The invention relates to a vertical search method based on user interest learning, in particular to a vertical search method for storing user browsing information or query behavior in a vertical search website, and belongs to the technical field of vertical search.

Background technique

Today is an era of information explosion. Information is no longer a scarce resource. With all kinds of information emerging one after another, users' attention has become scarce instead. The further development of search engine technology has given users great convenience. According to the "39th Statistical Report on Internet Development in China" released by China Internet Network Information Center CNNIC, as of December 2016, the number of search engine users in China reached 602 million, with a utilization rate of 82.4%. Smart, vertical specialization has become a development trend. On the one hand, the types of search information are more abundant; on the other hand, search engines launch more intelligent, comprehensive and professional search entities to meet users' search needs in different fields. This has triggered a new vertical and professional development trend in the search engine industry.

The user's search methods mainly include free word full-text search, keyword search, classification search and other special information search. Compared with traditional search engines, vertical search engines need to sort search results according to content rather than by analyzing the link relationship between web pages. And personalized vertical search needs to analyze user behavior, and establish user personalization model, to meet the sorting results of personalized needs. At present, most vertical search engine websites provide advanced search for entity attribute filtering, but the sorting of the filtering results is often in ascending or descending order according to certain conditions.

Among the existing personalized search methods, the mainstream search method is based on social tagging search, and the social tagging data generated through user collection, tagging, and sharing directly reflect users' quality evaluation and content understanding of web pages. Use social tagging data to rank webpages to improve webpage retrieval performance.

However, the information retrieval services provided by vertical search engines still have the following deficiencies at this stage. First, vertical search engines can only do passive query matching. Only after the user submits the query, the search engine can respond. This implies two major problems. On the one hand, it is often difficult for users to clearly summarize their needs. On the other hand, even if the query words can reflect the current needs of users, most vertical search engine websites do not take into account the interests of users. Extract and perform personalized search, or actively make information recommendation services; secondly, in vertical search engines that provide users with social annotation of resources, in general, they are easily affected by spam annotation: malicious users may use a large number of widely used The emergence of tags to improperly enhance its influence has brought great challenges to the validity of search results.

Therefore, it is necessary to pursue a personalized vertical search with higher precision in the vertical search engine.

Contents of the invention

The technical problem to be solved by the present invention is: to provide a vertical search method based on user interest learning, through the learning of user search behavior, to mine user interests, and to actively provide users with results consistent with their current interests.

The present invention adopts the following technical solutions for solving the problems of the technologies described above:

A vertical search method based on user interest learning, comprising the following steps:

Step 1, in the vertical search engine website including attribute filtering advanced search, according to the interaction with the user during initialization, determine the user's interest preference;

Step 2, according to the user's interest preference, establish an initial user interest model, and the initial user interest model is expressed as: x _i = (attribute value: the number of times the attribute value is followed by the user - the time when the attribute value was last followed by the user) , x _i represents the i-th attribute;

Step 3, if the user queries according to the attribute, the user interest model is updated in real time according to the value of the attribute selected by the user; if the user queries according to the keyword, the user interest model is not updated;

Step 4, if the user has a browsing behavior, set the browsing time as the threshold, if the browsing time of a certain entity resource exceeds the threshold, obtain the attribute value of the entity resource, and update the user interest model in real time;

Step 5, using the multi-attribute utility theory, combined with the user's preference factor and forgetting factor, calculate the interest degree value of all entity resources and sort them according to the interest degree value from high to low;

Step 6, filter and query according to keywords or attributes selected by the user, and generate an initial list of query results;

Step 7: For the entity resources in the initial query result list, combined with the user's interest in these entity resources, sort the entity resources in the initial query result list according to the interest value from high to low, and generate the final query list of results.

As a preferred solution of the present invention, the specific process described in step 1 is: when the user first enters the vertical search engine website including attribute screening and advanced search, interact with the user for each attribute, and obtain the user's interest preference about each attribute .

As a preferred solution of the present invention, in step 3, if the user queries according to the attribute, then according to the value of the attribute selected by the user, the specific process of updating the user interest model in real time is as follows: if the value of the attribute selected by the user has already been selected by the user In the interest model, add 1 to the number of times the value of the attribute is followed by the user, and update the time when the value of the attribute was last followed by the user, otherwise, the value of the attribute, the number of times the value of the attribute was followed by the user and The last time the value of this attribute was followed by the user is added to the user interest model.

As a preferred solution of the present invention, the specific process of the step 5 is:

Step 51, using the AF-IDF algorithm to calculate the preference factor, the calculation formula is:

k _i =AF(X/T)*log _a (N/DF(k _i ))

Among them, _ki represents the user's preference factor for the i-th attribute of the j-th entity resource, AF(X/T) is the average number of times the user pays attention to the i-th attribute of the j-th entity resource in a period T, and X is the attention The total number of times of this attribute, N is the total number of entity resources in the vertical search engine website, DF (k _i ) is the total number of entity resources identical with the i attribute value of the jth entity resource, and a is a hyperparameter;

Step 52, introduce the forgetting factor f _i into the preference factor, then the preference factor k _i ' after introducing the forgetting factor is:

k′ _i =k _i *f _i

Step 53, set the vertical search engine website to include n attributes, i.e. i=1, 2,..., n, then the user's interest degree value to the jth entity resource is represented by a multiplicative utility function:

U _j =k′ ₁ +k′ ₂ +…+k′ _n +k′ ₁ k′ ₂ +k′ ₁ k′ ₃ +…+k′ ₁ k′ _n +k′ ₂ k′ ₃ +k′ ₂ k′ ₄ +…+k′ ₂ k′ _n +k′ ₃ k′ ₄ +k′ ₃ k′ ₅ +…+k′ ₃ k′ _n +…+k′ _n-1 k′ _n

Among them, U _j represents the user's interest degree value to the jth entity resource;

Step 54, sort the interest degree values of all entity resources from high to low, and output the sorting result.

As a preferred solution of the present invention, the formula for calculating the _forgetting factor fi is:

f _i =exp(-log _b (t)/f)

Among them, t is the difference between the last time the value of the i-th attribute of the j-th entity resource was paid attention to and the current time, f is the half-life, and b is the hyperparameter.

Compared with the prior art, the present invention adopts the above technical scheme and has the following technical effects:

1. The order of magnitude of the algorithm complexity of the present invention is small. If the number of attributes is n, each entity resource can calculate the interest degree value of the entity resource only at n+ ² levels.

2. Most of the algorithm of the present invention is calculated offline, and a small part is calculated in real time. Even if the vertical search system has many physical resources and registered users, general enterprise-level servers can fully satisfy the operation and realization of the algorithm.

3. The algorithm of the present invention has strong versatility, can be implemented on various vertical search engines, and has good scalability.

Description of drawings

FIG. 1 is a schematic flowchart of a vertical search method based on user interest learning in the present invention.

detailed description

Embodiments of the invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below by referring to the figures are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

As shown in Figure 1, it is a schematic flow chart of a vertical search method based on user interest learning, and the specific steps are as follows:

Step 1: In the vertical search engine website that includes attribute filtering and advanced search, determine the user's interest preference according to the interaction with the user during initialization, so as to avoid the problem of cold start.

Step 2: Establish and initialize the user interest model, and adjust the model in real time by filtering query records and browsing records based on user attributes. Record each user's query and browsing behavior using attribute filtering, and add attribute query information and browsed entity resource attribute information to the user's personalized model.

Step 3: Using the multi-attribute utility theory, combined with the user's preference factor, calculate the personalized interest degree value of all entity resources in the background, and generate a sorted list of entity resources of user's personalized interest. Specifically:

Step 301: Given n attributes in the vertical search engine, the user's personalized preference utility equation can be described by a multiplicative utility function: U _j =k ₁ +k ₂ +...+k _n +k ₁ k ₂ +k ₂ k ₃ +k ₁ k ₃ +...+k _n-1 k _n . Where U _j represents the interest degree value of the j-th entity in the current user, and _ki represents the user's preference factor for the i-th attribute value of the entity. The formula is expressed as the addition of the independent calculation results of each entity attribute value preference factor, and each entity attribute value preference factor must be multiplied by each other entity attribute value preference factor and then added.

Step 302: Calculating the preference factor _ki Based on the user's different interest in the entity, the present invention uses the average Average Frequency/Inverse Document Frequency (AF-IDF) algorithm to characterize the user's preference for the i-th attribute value of the entity, _ki = AF(X/T)*log _a (N/DF(k _i )), where AF(X/T) is the average number of times the user pays attention to the i-th attribute value of the entity in the T period (query or browsing times ), X is the total number of attribute values concerned, n is the total number of entities in the vertical search system, DF(k _i ) is the number of entities whose attribute value is k _i , after calculating the weight of each attribute in turn, the user based on An interest preference vector for attribute classification.

Step 303: The user's interest often changes, and the interest forgetting factor f is introduced into the preference factor, namely: k′ _i =k _i *f _i , f _i =exp(-log _b (t)/f), where t is the time difference from the last time the attribute value was paid attention to, and f is the half-life.

Step 304: Calculate the personalized interest degree value for each entity in the vertical search system in turn until the calculation of the last entity is completed.

Step 305: Sorting from high to low according to the personalized interest value, outputting the sorting result, and storing it in the user's personal interest model.

Step 4: According to the user's search conditions, filter the query according to keywords or attributes, and generate an initial list of query results.

Step 5: For the entity resources in the generated initial query result list, combined with the entity resource sorting list of the user's personalized interest, reorder the entity resources in the initial query result list according to the interest degree value from high to low.

Take the search of a large domestic automobile vertical website as an example (but not limited to this example). In addition to a search box for users to search by keyword, the website also provides a precise car selection module by category, including by price and by price. Model, by brand, by displacement 4 attributes.

S1. To realize personalized search, it is necessary to learn user interests. First, user registration information is required to establish a personalized model. When the user registers, fill in the content of the 4 basic attributes for the user's personal interest needs. According to the interaction with the user during initialization, the user's interest preference is determined to avoid the problem of cold start. The user's initial personalized interest model consists of vectors of 4 basic attributes.

S2. Assume that the initial personalized interest model of user 1 is:

Price = (80,000-100,000: 1-t), model = (SUV: 1-t), brand = (Great Wall: 1-t) displacement = (1.1-1.6L: 1-t).

Note: The value of the price attribute vector is "80,000-100,000", "1" is its value, which means it has been followed by a user once, "-" is a connector, and "t" is the time when it was last followed by a user.

S3. There are two query methods for the user. If the query is based on the attribute category, it indicates that the user pays attention to the attribute category, reflects the user's preference, and needs to update the user's personal interest model. If it is not filtered by attribute category, that is, keyword query, the user's preference will not be recorded, and the user's personal interest model will not be updated.

The update process is to record each user's search or browsing behavior and store it in the personalized information record to complete the update of personalized information. Individual interests are not fixed but dynamic. User preferences shift with the user's own personal and external influences. Taking the scenario as an example, when a user is initialized, it may be that the user has just joined the work, and the requirement for the car is only for transportation. The price requirement may not be high, but with the improvement of personal growth and quality of life, it is necessary to improve the requirement for the car. At this time, the user's personalized interests will change, and the user's search and browsing content will also change.

S4. Add the selected attribute category to the user personalization model. In the case of an existing category, its value is increased by 1, and the last attention time is updated.

Assume that user 1 initiates a filter query for attribute categories as follows:

Price: RMB 80,000-100,000; Model: Hatchback; Brand: Ford; Displacement: 1.1-1.6L.

Then the corresponding personalized model is updated as: price=(8-100,000: 2-t), model=(SUV: 1-t, hatchback: 1-t), brand=(Great Wall: 1-t, Ford : 1-t), displacement = (1.1-1.6L: 2-t).

S5. Using the multi-attribute utility theory and combining the user's preference factors, calculate the personalized interest degree value of all physical resources in the background, and generate a sorted list of physical resources of the user's personalized interest. Specifically:

Given n attributes in the vertical search engine, the user’s personalized preference utility equation can be described by multiplicative utility function: U _j =k ₁ +k ₂ +...+k _n +k ₁ k ₂ +k ₂ k ₃ +k ₁ k ₃ +...+k _n-1 k _n . Where U _j represents the interest degree value of the j-th entity in the current user, and _ki represents the user's preference factor for the i-th attribute value of the entity. The formula is expressed as the addition of the independent calculation results of each entity attribute value preference factor, and each entity attribute value preference factor must be multiplied by each other entity attribute value preference factor and then added.

Calculating the preference factor _ki is based on the user's interest in the entity. The present invention uses the Average Frequency/Inverse Document Frequency (AF-IDF) algorithm to characterize the user's preference for the i-th attribute value of the entity, _ki =AF(X/ T)*log _a (N/DF(k _i )), where AF(X/T) is the average number of times the user pays attention to the i-th attribute value of the entity within the T period (queries or browsing times), and X is The total number of attribute values concerned, n is the total number of entities in the vertical search system, DF(k _i ) is the number of entities whose attribute value is k _i , after calculating the weight of each attribute in turn, the user's interest based on attribute classification is obtained preference vector.

The user's interest often changes, and the interest forgetting factor f _i is introduced into the preference factor, namely: k′ _i =k _i *f _i , f _i =exp(-log _b (t)/f), where t is an attribute The value is the time difference from the last time the value was concerned to the current time, and f is the half-life.

Calculate the personalized interest degree value for each entity in the vertical search system in turn until the calculation of the last entity is completed.

Sorting from high to low is performed according to the personalized interest value, and the sorting results are output and stored in the user's personal interest model. The result sorting is to calculate the score of each entity resource according to the algorithm itself, and sort them. Taking the scenario as an example, the scores of all entity resources are calculated for each user in the background. When the user opens the homepage of the website and does not initiate a query, the entities can be displayed in the recommendation column in descending order of scores. If the user initiates a filter query by keyword or attribute, the candidate entity resources are obtained first, and then reordered according to the scores of the entity resources and displayed in the search result column.

In order to better describe the algorithm, the following simplified assumptions are made:

The interest model of user 1 happens to be updated in step S4, and the vectors of the 4 attributes are:

Price = (80,000-100,000: 2-t),

Model = (SUV: 1-t, Hatchback: 1-t),

Brand = (Great Wall: 1-t, Ford: 1-t),

Displacement = (1.1-1.6L:2-t).

A certain entity U ₁ happens to be the same type of car concerned by step S4, and its attributes are: price: 80,000-100,000; model: hatchback; brand: Ford; displacement: 1.1-1.6L.

The total number of car entities in the vertical website is n=1000.

The price is 80,000-100,000 and the quantity is 500,

The number of models is hatchback sedan is 250,

The brand is Ford and the quantity is 25,

The quantity with a displacement of 1.1-1.6L is 250,

The period T is temporarily set to 1.

Since the attention time t is the same day as the query and there is no time difference, the forgetting factor is 1.

Calculate the value of U ₁ =k ₁ +k ₂ +k ₃ +k ₄ +k ₁ k ₂ +k ₁ k ₃ +k ₁ k ₄ +k ₂ k ₃ +k ₂ k ₄ +k ₃ k ₄ , indicating that the user Degree of interest in entity U ₁ . In this example, the value of a is 2.

k ₁ =2*log ₂ (1000/500)=2

k ₂ =2*log ₂ (1000/250)=4

k ₃ =2*log ₂ (1000/25)=10.64

k ₄ =2*log ₂ (1000/250)=4

U ₁ ＝2+4+10.64+4+2*4+2*10.64+2*4+4*10.64+4*4+10.64*4＝159.04

The weight _of entity U1 in user1's personalized model is 159.04.

The weight of an entity in the user personalization model is described above through a simple example. All data settings are artificial assumptions, and the implementation needs to be based on the data in the specific vertical entity website. Parameters such as period T and half-life f need to be set more effectively through experiments.

S6. If the user has a browsing behavior, the browsing time can be set as the threshold here. When browsing a certain resource exceeds the threshold, it indicates that the user pays attention to the attribute category of this entity resource, reflecting the interest preference, and updates the user's personal interest according to the step S4 Model.

The above embodiments are only to illustrate the technical ideas of the present invention, and can not limit the protection scope of the present invention with this. All technical ideas proposed in accordance with the present invention, any changes made on the basis of technical solutions, all fall within the protection scope of the present invention. Inside.

Claims (5)

1. a vertical search method based on user interest learning, is characterized in that, comprises the steps: Step 1, in the vertical search engine website including attribute filtering advanced search, according to the interaction with the user during initialization, determine the user's interest preference; Step 2, according to the user's interest preference, establish an initial user interest model, and the initial user interest model is expressed as: x _i = (attribute value: the number of times the attribute value is followed by the user - the time when the attribute value was last followed by the user) , x _i represents the i-th attribute; Step 3, if the user queries according to the attribute, the user interest model is updated in real time according to the value of the attribute selected by the user; if the user queries according to the keyword, the user interest model is not updated; Step 4, if the user has a browsing behavior, set the browsing time as the threshold, if the browsing time of a certain entity resource exceeds the threshold, obtain the attribute value of the entity resource, and update the user interest model in real time; Step 5, using the multi-attribute utility theory, combined with the user's preference factor and forgetting factor, calculate the interest degree value of all entity resources and sort them according to the interest degree value from high to low; Step 6, filter and query according to keywords or attributes selected by the user, and generate an initial list of query results; Step 7: For the entity resources in the initial query result list, combined with the user's interest in these entity resources, sort the entity resources in the initial query result list according to the interest value from high to low, and generate the final query list of results. 2. according to the described vertical search method based on user interest learning of claim 1, it is characterized in that, the specific process described in step 1 is: when the user enters the vertical search engine website that comprises attribute screening advanced search for the first time, for each attribute and The user interacts to obtain the user's interest preferences for each attribute. 3. The vertical search method based on user interest learning according to claim 1, characterized in that, in step 3, if the user queries according to the attribute, the user interest model is updated in real time according to the value of the attribute selected by the user. It is: if the value of the attribute selected by the user is already in the user interest model, add 1 to the number of times the value of the attribute is followed by the user, and update the time when the value of the attribute was last followed by the user; otherwise, take the value of the attribute value, the number of times the value of the attribute is followed by the user, and the time when the value of the attribute was last followed by the user are added to the user interest model. 4. according to the described vertical search method based on user interest learning of claim 1, it is characterized in that, the specific process of described step 5 is: Step 51, using the AF-IDF algorithm to calculate the preference factor, the calculation formula is: k _i =AF(X/T)*log _a (N/DF(k _i )) Among them, _ki represents the user's preference factor for the i-th attribute of the j-th entity resource, AF(X/T) is the average number of times the user pays attention to the i-th attribute of the j-th entity resource in a period T, and X is the attention The total number of times of this attribute, N is the total number of entity resources in the vertical search engine website, DF (k _i ) is the total number of entity resources identical with the i attribute value of the jth entity resource, and a is a hyperparameter; Step 52, introduce the forgetting factor f _i into the preference factor, then the preference factor k _i ' after introducing the forgetting factor is: k′ _i =k _i *f _i Step 53, set the vertical search engine website to include n attributes, i.e. i=1, 2,..., n, then the user's interest degree value to the jth entity resource is represented by a multiplicative utility function: U _j =k′ ₁ +k′ ₂ +…+k′ _n +k′ ₁ k′ ₂ +k′ ₁ k′ ₃ +…+k′ ₁ k′ _n +k′ ₂ k′ ₃ +k′ ₂ k′ ₄ +…+k′ ₂ k′ _n + k′ ₃ k′ ₄ +k′ ₃ k′ ₅ +…+k′ ₃ k′ _n +…+k′ _n-1 k′ _n Among them, U _j represents the user's interest degree value to the jth entity resource; Step 54, sort the interest degree values of all entity resources from high to low, and output the sorting result. 5. according to the described vertical search method based on user interest learning of claim 4, it is characterized in that, described forgetting factor _fi calculation formula is: f _i =exp(-log _b (t)/f) Among them, t is the difference between the last time the value of the i-th attribute of the j-th entity resource was paid attention to and the current time, f is the half-life, and b is the hyperparameter.