CN100412870C - Portal personalized recommendation service method and system using meta-recommendation engine - Google Patents

Abstract

The invention discloses a portal personalized recommendation service method and system using a meta-recommendation engine, and proposes the construction of a portal user's interest model, including the initial creation and subsequent update of the portal user's interest model; The personalized recommendation service architecture of the platform, the meta-recommendation engine can analyze the relationship between users and user groups and the changes in personalized interests, organize information resources and recommendation algorithms in a unified way and choose and control them reasonably, and optimize the push to generate more comprehensive and diverse personalized recommendation results ; In terms of resource display, the system for realizing portal personalized recommendation service can encapsulate various predicted and recommended Web resource content objects into portlets, which can provide vivid and intuitive personalized display to portal users and provide a higher level of personalized control . Comprehensively utilize the existing personalized resources and technical means of the portal platform to provide independent and flexible service middleware or service agents to complete personalized recommendation services.

Description

Portal personalized recommendation service method and system using meta-recommendation engine

technical field

The present invention relates to portal (Portal) technology, in particular to a portal personalized recommendation service method and system using a meta-recommendation engine.

Background technique

Personalized recommendation service can proactively provide Internet users with diversified and intelligent personalized services, display them in the fastest way, and assist users to efficiently and comprehensively obtain useful World Wide Web (Web) resource information. At present, The system that can realize personalized recommendation service has become an important part of the network information system, especially in the large-scale portal application fields such as e-commerce, digital library, and distance education. With the advent of the Web 2.0 era, Resource integration and personalized service methods based on portal platforms are becoming more and more abundant. At present, there are methods that use a central database to store all rule-based recommendation algorithms and provide dynamic and flexible choices for users to make recommendations. There is no or relatively simple analysis of , and it is mainly used in simple Web recommendation systems such as e-commerce, and cannot provide support for portal applications. Due to the different services and structures supported by the portal platform and the inflexibility of adapting to the needs, there are many problems in the field of portal personalized recommendation services.

First of all, the personalized recommendation service mechanism provided by the current portal is very dependent on the platform itself, and it needs to be loaded with customized or secondary development of the same series of service components to complete, and the coupling with the platform is very close; at the same time, using portal technology There is no comprehensive application scheme for personalized interest mining and recommendation services, and it is impossible to provide a relatively complete and flexible portal-oriented personalized recommendation service architecture. Further research is needed on system architecture integration, related algorithm strategies and interface design. Research.

Secondly, in recent years, the technical solutions of personalized recommendation services have mostly focused on the combined recommendation method that comprehensively utilizes various mature recommendation algorithms. However, how to provide users with flexible, comprehensive and high-quality recommendation results according to complex situations has always been a topic of discussion. Hot Issues. Although some combined recommender systems have been tentatively researched on the meta-recommendation model, they have not considered the role of the user interest model and its application characteristics in the realization of portal personalization. The user interest model is mainly used to describe user interests in personalized service applications, and is also used as a calculation object for recommendation services. There are many studies in the Web system in the early stage, including comprehensive Web usage mining. The current user interest model in the portal platform Create methods are less involved. Meta-recommendation refers to the synthesis of various predictive analysis algorithms to perform more comprehensive and accurate recommendation calculations. Among the existing methods that use a central database to store all recommendation algorithms and provide users with dynamic and flexible choices for recommendations, the meta-recommendation system architecture model generally emphasizes The data layer integrates the heterogeneity processing of various recommendation systems, that is, the integration of scattered and independent recommendation systems, which is relatively opaque, so it still tends to system integration, does not consider the details of the recommendation algorithm combination selection strategy, and is not suitable for easy deployment , Lightweight general requirements that are not restricted by platforms and resources.

Again, in order to meet more application scenarios, the focus should not only be on how to improve the performance or scalability of the recommendation algorithm, but should assist users in efficient and convenient portal personalized applications through innovative models and approaches. For example, the recommendation control mode should be changed from The traditional what to recommend to the user is extended to how to recommend non-repetitive and meaningful content to the user and the user's similar interest user groups, and can be displayed vividly, intuitively and quickly.

To sum up, based on the portal platform, combining personalized interest mining and recommendation services to provide users with flexible, comprehensive and high-quality recommendation results has become an urgent problem to be solved.

Contents of the invention

In view of this, an object of the present invention is to provide a portal personalized recommendation service system using a meta-recommendation engine. Another object of the present invention is to provide a portal personalized recommendation service method using a meta-recommendation engine. The mining and recommendation service are combined to provide users with flexible, comprehensive and high-quality recommendation results.

In order to achieve the above object, the portal personalized recommendation service system using a meta-recommendation engine provided by the present invention includes: a data management unit, a data storage unit, an interest mining unit, an interest model construction unit, a training classification unit, a similarity calculation unit, and a meta-recommendation engine and WWW resource index storage unit

The data management unit is used to assist in managing the data communication and calling of the training classification unit, or the similarity calculation unit and the data storage unit;

The data storage unit is used to store the interest model library of the portal user and/or the portal user group, and the interest model library includes the interest degree model library and the access transaction set of the portal user and/or the portal user group, and is further stored in the data storage unit. Recommended algorithm set;

The interest mining unit, located in the portal platform, is used to obtain personalized description files of portal users, implicitly track and capture the interest content and access behavior patterns of logged-in portal users, and provide the obtained information to the interest model construction unit;

An interest model construction unit, configured to standardize the acquired interest data, construct a portal user interest model according to the processed information, and provide the constructed portal user interest model to the training classification unit and the similarity calculation unit;

The training classification unit is used to call the interest model stored in the data storage unit through the data management unit, and perform the feedback learning of the nearest neighbor clustering on the interest model from the interest model construction unit and the interest model from the data storage unit, and then feed back the learning result Update the portal user interest model stored in the data storage unit through the data management unit and provide it to the similarity calculation unit;

The similarity calculation unit is used to call the interest model stored in the data storage unit through the data management unit, and perform the calculation according to the interest model from the interest model construction unit, the feedback learning update result from the training classification unit, and other interest models from the data storage unit. Similarity calculation, and then provide the similarity calculation result to the meta recommendation engine;

The meta-recommendation engine is used to call the interest model stored in the data storage unit through the data management unit, and determine the selection and selection of the recommendation control strategy and the recommendation algorithm according to the interest model from the data storage unit and the similarity calculation result from the similarity calculation unit. Combining, and then performing predictive filtering analysis according to the similarity calculation results from the similarity calculation unit, and performing calculations to determine the recommended results according to the predicted analysis results, recommended control strategies and recommendation algorithms, and calling the World Wide Web WWW resource index storage unit according to the determined recommended results The WWW resource index stored in the WWW resource is packaged as a portlet containing Web page content, and pushed to the portal user;

The WWW resource index storage unit is used for storing the WWW resource index.

The meta-recommendation engine includes: a recommendation selector, a predictive analysis unit and a recommended resource presentation unit,

The recommendation selector is used to call the interest model stored in the data storage unit through the data management unit, and determine the selection and selection of the recommended control strategy and the recommendation algorithm according to the interest model from the data storage unit and the similarity calculation result from the similarity calculation unit. Combining, and then providing to the predictive analysis unit, and providing the similarity calculation result from the similarity calculation unit to the predictive analysis unit;

The predictive analysis unit is used to perform predictive filtering analysis according to the similarity calculation results from the similarity calculation unit, and perform calculations to determine the recommended results according to the predictive analysis results and the recommended control strategy and recommendation algorithm from the recommended selector, and the determined recommended results Provide the recommended resource presentation unit by calling the WWW resource index stored in the WWW resource index storage unit;

The recommended resource presentation unit is used to package the WWW resources from the predictive analysis unit into portlets containing web page content and push them to portal users.

The recommended resource presentation unit includes: a Portlet configuration management unit, a Portlet session management unit, a request command analysis unit, a Web page acquisition unit, a response tag processing unit and a WSRP interface encapsulation unit,

The portlet configuration management unit is used to maintain the metadata of all portlets provided by the encapsulation mechanism WA2WP of the current World Wide Web web application to conform to the encapsulation mechanism of the remote portlet component Web service Portlet;

Portlet session management unit, used to manage the entire life cycle of session objects;

The request command analysis unit is used to receive the encapsulation display request of the resource link contained in the recommendation result and the user request for accessing the resource, analyze the request parameters and session data to determine the target resource to be accessed, locate the target URL, obtain and prepare for access Request parameters and session data required by the target resource;

The web page acquisition unit is used to access the web application according to the target URL, request parameters and session data from the request command analysis unit, obtain the returned page mark content and cookie data, and provide it to the response mark processing unit;

The response tag processing unit is used to preprocess the hypertext tag information returned by the Web page acquisition unit before encapsulation, obtain the Web resource page fragment, and then provide it to the WSRP interface encapsulation unit;

The WSRP interface encapsulation unit is used for encapsulating web resource page fragments as portlets and displaying them on the portal personalized desktop.

The training and classification unit is further used to: store the user or user group ID for which the interest model has been established, and if no user or user group ID is stored, provide the trained and classified interest model to the data storage unit for storage through the data management unit.

The system for realizing portal personalized recommendation service further includes a privacy protection unit; the interest mining unit is used to provide the obtained information to the privacy protection unit; the privacy protection unit is used to embed a security mark in the information from the interest mining unit, For privatization filtering protection, and then provided to the interest model building unit.

The portal personalized recommendation service method using the meta-recommendation engine provided by the present invention includes:

A. Mining the interests of portal users, obtaining personalized description files of portal users, implicitly tracking and capturing the interests and access behavior patterns of logged-in portal users;

B. Perform standardized processing, extract information related to portal user interests, and determine whether to create a new portal user interest model, if so, create a new portal user interest model, otherwise, update the existing portal user interest model ;

C. Training and classifying the constructed portal user interest model and the stored portal user interest model;

D. Carry out similarity calculation according to the constructed portal user interest model, the stored portal user interest model and the feedback learning results;

E. According to the stored interest model and similarity calculation results, determine the selection and combination of recommended control strategies and recommendation algorithms, perform predictive filtering analysis according to the similarity calculation results, and then according to the predicted analysis results and the determined recommended control strategies and recommendation algorithms, Executing calculations to determine the recommended results, and calling the stored WWW resource index according to the determined recommended results;

F. Encapsulate the invoked WWW resource index into a portlet containing web page content, and push it to the portal user.

Between step A and step B, it further includes: embedding a security mark into the acquired information.

The step C is: perform feature training according to the constructed portal user interest model, extract interest content and behavior characteristics to preliminarily divide the category of the interest model and the category of interest resource content, and continuously update the portal user interest model.

The step D is: on the basis of the existing classification, similar matching and comparison between user interest models are performed to generate a neighbor set of target portal users.

The predictive filtering analysis described in step E is: based on the selected target portal user's neighbor set, predict the resources that the target portal user has not browsed or has unknown interests.

In the present invention, the construction of the portal user interest model is proposed, including the initial creation of the portal user interest model and the subsequent update of the portal user interest model; a personalized recommendation service architecture independent of the portal platform using a meta-recommendation engine, meta-recommendation The engine can analyze the relationship between users and user groups and changes in personalized interests, organize information resources and recommendation algorithms in a unified manner and rationally select and control them, and optimize pushes to generate more comprehensive and diverse personalized recommendation results; in terms of resource display, realize portal personalized recommendations The service system can encapsulate various predicted and recommended Web resource content objects into portlets, which can be vividly and intuitively displayed to portal users and provide a higher level of personalized control. Comprehensively utilize the existing personalized resources and technical means of the portal platform to provide independent and flexible service middleware or service agents to complete personalized recommendation services.

Description of drawings

Fig. 1 shows a schematic structural diagram of a system realizing portal personalized recommendation service in the present invention;

Fig. 2 shows the flow chart of realizing portal personalized recommendation service among the present invention;

Fig. 3 shows a schematic diagram of the construction process of the portal user interest model in the present invention;

Fig. 4 shows the data structure diagram of data set among the present invention;

Fig. 5 shows a schematic diagram of the meta-recommendation control strategy in the present invention;

Fig. 6 shows a schematic diagram of the presentation mechanism of recommended resources in the present invention;

Fig. 7 shows a flow chart of implementing recommended resource presentation in the present invention.

Detailed ways

In the present invention, the construction of the portal user interest model is proposed, including the initial creation of the portal user interest model and the subsequent update of the portal user interest model; a personalized recommendation service architecture independent of the portal platform using a meta-recommendation engine, meta-recommendation The engine can analyze the relationship between users and user groups and changes in personalized interests, organize information resources and recommendation algorithms in a unified manner and rationally select and control them, and optimize pushes to generate more comprehensive and diverse personalized recommendation results; in terms of resource display, realize portal personalized recommendations The service system can encapsulate various predicted and recommended Web resource content objects into portlets, which can be vividly and intuitively displayed to portal users and provide a higher level of personalized control.

Through the offline processing process, it provides early data maintenance guarantee for the online processing process and reduces the complexity of online calculation. It can be composed of three parts: training classification unit, data management unit and data storage unit. Based on the portal user interest content model and historical access transaction information, perform neighbor clustering and training learning, classify the data according to information related to various interests, store them in the interest degree model library of the data set and access transaction collection, and perform training These data are used for classification and similarity calculation. The data set adopts a lightweight data organization method, and complex unstructured data can be used for data communication through configuration connections, which is convenient for the flexible deployment and application of services. In addition, the recommendation algorithm set required for the portal personalized recommendation service is also centrally stored in Data centralization. Lightweight data organization refers to small databases that only retain storage and reading functions, and try not to use specialized large databases with large resource usage.

The online processing process includes three steps: mining interest of portal users, creating and updating interest configuration, and pushing recommended content by meta-recommendation engine.

First, mine the interests of portal users, obtain personalized description files of portal users, implicitly track and capture the content of interest and access behavior patterns of portal users. Since the interest information of portal users is acquired implicitly, the security of user privacy should be guaranteed during the process from acquisition to normalization, and privatization and filtering protection can be carried out by embedding security tags into the acquired information.

Secondly, standardize the personalized description files and access transaction sets of portal users, construct the interest model of portal users and their user groups, dynamically adjust and update each portal user's interest decay, and continuously use it for training Classified feedback learning, and based on the interest model library in the data set, perform more accurate clustering of users or user groups and calculation of interest similarity.

Then, after obtaining the interest model and similarity classification of portal users, the selection and combination of recommendation algorithms are dynamically performed through the recommendation control strategy of portal users and portal user groups, and then corresponding prediction and filtering calculations are performed. The specific content of the recommendation results comes from the The classified index library retrieved from World Wide Web (WWW) resources is finally transformed and packaged into a portlet containing Web page content and pushed to portal users.

Fig. 1 has shown the system structural diagram that realizes portal personalized recommendation service in the present invention, as shown in the figure: as shown, the system that realizes portal personalized recommendation service includes interest mining unit 101, interest model construction unit 103, training classification unit 104, Data management unit 105, similarity calculation unit 106, recommendation selector 107, data storage unit 108, predictive analysis unit 109, WWW resource index storage unit 110 and recommended resource presentation unit 111.

The data management unit 105 is used to assist in the management of data communication and calls between the training classification unit 104 , or the similarity calculation unit 106 and the data storage unit 108 .

Data storage unit 108 is used for storing the interest model storehouse of portal user and/or portal user group, and this interest model storehouse comprises portal user and/or portal user group's degree of interest model storehouse and access transaction set, further stores in data storage unit 108 There are sets of recommended algorithms.

The interest mining unit 101 is located in the portal platform, and is used to obtain the personalized description files of portal users, implicitly track and capture the interest content and access behavior patterns of logged-in portal users, and provide the obtained information to the interest model construction unit 103.

The interest model construction unit 103 is used to normalize the acquired interest data, construct the portal user interest model according to the processed information, and provide the constructed portal user interest model to the training classification unit 104 and the similarity calculation unit 106.

If the interest model of the portal user does not exist yet, then the training classification unit 104 is first used to provide the interest model after the training classification to the data storage unit 108 for storage by the data management unit 105. No matter whether the interest model of the portal user already exists, the training The classification unit 104 is used to call the interest model stored in the data storage unit 108 through the data management unit 105, and the interest model from the interest model construction unit 103 and the interest model from the data storage unit 108 are used for feedback learning of neighbor clustering, and then The feedback learning result is updated by the data management unit 105 through the portal user interest model stored in the data storage unit 108 and provided to the similarity calculation unit 106. The training classification unit 104 can store the user or user group identification of the established interest model, like this , the training classification unit 104 can determine whether the interest model from the interest model construction unit 103 already exists through the stored identifier.

The similarity calculation unit 106 is used to call the interest model stored in the data storage unit 108 through the data management unit 105, according to the interest model from the interest model construction unit 103, the feedback learning update result from the training classification unit 104 and other information from the data storage unit The interest model of 108 performs more accurate similarity calculation, and then provides the similarity calculation result to the recommendation selector 107.

The recommendation selector 107 is used to call the interest model stored in the data storage unit 108 through the data management unit 105, and determine the recommended control strategy and recommendation according to the interest model from the data storage unit 108 and the similarity calculation result from the similarity calculation unit 106. The selection and combination of algorithms are then provided to the predictive analysis unit 109 , and the similarity calculation results from the similarity calculation unit 106 are provided to the predictive analysis unit 109 .

The predictive analysis unit 109 is used to perform predictive filtering analysis according to the similarity calculation results from the similarity calculation unit 106, and perform calculations to determine the recommended results according to the predictive analysis results and the recommended control strategy and recommendation algorithm from the recommended selector 107, and the determined The recommendation result is provided to the recommended resource presentation unit 111 by calling the WWW resource index stored in the WWW resource index storage unit 110 .

The WWW resource index storage unit 110 is used for storing the WWW resource index.

The recommended resource presentation unit 111 is used to package the WWW resources from the predictive analysis unit 109 into portlets containing Web page content, and push them to portal users.

The above-mentioned recommendation selector 107, predictive analysis unit 109 and recommended resource presentation unit 111 constitute a meta-recommendation engine.

Between the interest mining unit 101 and the interest model construction unit 103, a privacy protection unit 102 can be further included, and the interest mining unit 101 is used to provide the obtained information to the privacy protection unit 102; The information of is embedded in the security mark for privatization filtering protection, and then provided to the interest model construction unit 103.

Fig. 2 shows the flow chart of realizing portal personalized recommendation service among the present invention, as shown in Fig. 2, the concrete process of realizing portal personalized recommendation service comprises the following steps:

Step 201: Mining the interests of portal users, obtaining personalized description files of portal users, implicitly tracking and capturing interest content and access behavior patterns of portal users.

Step 202: Since the portal user's interest information is acquired implicitly, the security of user privacy should be ensured during the process from acquisition to normalization processing, and privatization filtering protection can be performed by embedding security tags into the acquired information.

Step 203: Standardize the information protected by privatization filtering, and extract information related to portal user interests.

Step 204: Determine whether to create a new portal user's interest model, if yes, then execute step 205; otherwise, execute step 206. The identifier of the portal user who has created an interest model can be stored. In this way, if the current portal user has been stored Portal user ID, it means that the interest model has been created for the corresponding portal user, and there is no need to create a new portal user’s interest model; if the current portal user’s ID is not stored, it means that the interest model has not been created for the corresponding portal user. Create new portal user interest models.

Step 205: Create a new portal user interest model, and then proceed to step 207.

Step 206: Update the existing portal user interest model, and then proceed to step 207.

The portal user interest model is a computable description of the portal user's interest preferences and usage behavior patterns. The description objects refer to various users who log in to the portal with personalized service permissions and registered users who log in. In terms of structure, individual portal users and There are two types of portal user groups. The portal user group described in the present invention is a more flexible and dynamic virtual concept that is different from the organizational structure of portal users, and is clustered according to the actual interest similarity of portal users. As the interest decay of portal users changes, its The portal user group to which it belongs will also change accordingly. Relatively, the interest maintained by the portal user group is more stable and durable than that of a single portal user, so it can also be used as a reference for the meta-recommendation engine in predicting calculations.

For steps 201 to 206, the process of creating and updating the portal user interest model is to implicitly realize the dynamic interest mining process that combines the portal user's interested content and access behavior, including the following steps as shown in Figure 3 , firstly, obtain the portal platform interest description file (UserProfile, UP) of the portal user, then perform privacy protection on UP, perform privatization filtering protection, and embed security tags into UP; secondly, perform data preprocessing on UP, perform feature expansion, Mining interest categories and standardizing access transaction sets; thirdly, building a portal user interest model, extending UP to UP′, and establishing a multigroup <U, I(A+C), G>; finally, performing normalization processing for dimensionality reduction, reducing Computational complexity to generate a portal user interest model.

The specific operation shown in FIG. 3 will be described in more detail below.

If the portal user u sets up and accesses his personalized desktop sequentially within T time period, and browses the Tab collection {t ₁ , t ₂ ..., t _M } of M different pages and N portlet components Collection of Portlets {p ₁ , p ₂ ..., p _N }.

On the one hand, breadth first extracts the corresponding interest content topics for feature description and expansion, and sets InterestContent(p, t) to describe the interest degree function of the portal user's interest content, then InterestContent(p, t) can be expressed as

InterestContent(p, t) = F((Feature(p, t), Weight(p, t)), FeatureExpand(p, t)) (1)

Among them, Feature() and Weight() are the extraction feature function and weight function respectively, and the extraction feature refers to the subject, keywords, etc. of the extracted content; FeatureExpand() is used to expand the description of related subject features. The weighting process is to extract The features are weighted according to the importance of interest and the degree of relevance, and can usually be expressed in grades.

On the other hand, by standardizing the behavior patterns and access process of portal users, we can focus on dynamic tracking and capture of several behavior operations such as click, layout, editing, and reference, which approximately reflect typical portal user interest behaviors. Setting InterestAction( u, p, t) is the interest degree function describing portal user behavior, then InterestAction(u, p, t) can be expressed as

InterestAction(u, p, t) = G(u, Click(p), Arrange(p), Edit(p), Quate(p), Freq(t), Duration(t)) (2)

Among them, Click(p), Arrange(p), Edit(p) and Quate(p) are respectively used to describe the behaviors of portal users clicking, layouting, editing and quoting portlets, Freq() is the number of return visits, and Duration( ) is the dwell time of the return visit.

Considering the interactive adaptability between portal user behavior and content interest changes, the graph theory definition can be used to generate access transaction sequences, and the access transaction of each portal user is defined as an access path of portal users to the portal as={p,t,Feature (p, t), InterestAction(u, p, t)}, the portal user access transaction set is the access path set AS={u, {as}, T} of each portal user to the portal in different time periods, and then Comprehensively compare the similarity of the portal users' interest content, interest behavior and access transactions, and set the portal user group UserGroup category to which the portal users belong.

After the obtained UP is privatized, filtered, protected, and preprocessed by data cleaning, an extended description of interests that combines interest content and behavior, and stability and outstanding interests is carried out. Based on the semantic structure, a relatively complete interest description suitable for portal users is established. File UP', UP or UP' are mostly resource definition framework (Resource Definition Framework, RDF) files based on Extensible Markup Language (XML), and are constructed by extracting feature tuples <User, <InterestContent, InterestAction>, UserGroup> Vector mockup of portal user interests.

In addition, the Fibonacci Numbers description function Fibo() is introduced, and the method of combining progressive forgetting and sliding windows is used to solve the problem of model update due to the drift of portal user interests. Limit the number of windows L of user interest categories , and select the time interval for portal users to visit the same related content, such as the number of days, and dynamically move the interest with the least attention of portal users out of the window to ensure timely and effective update of the portal user interest model. Define q=Interval for a certain path (as, as′), and obtain the access time interval of portal users, the weight update relationship can be expressed as

Weight'(p,t)＝Weight(p,t)+Feedback(q)/Fibo(L) (3)

Among them, Feedback() is the feedback function describing the interest drift of portal users, expressed as

$\mathrm{<span\; class="notranslate">\; Feedback</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; ifq</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Fibo</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; if</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}\end{array}\right.$

All interest degree model libraries and access transaction sets are loaded into the data set through the training classification module for centralized maintenance, among them, the interest degree function InterestContent(p, t) describing the portal user's interest content and the interest degree function InterestAction(p, t) describing the portal user behavior u, p, t) can be stored in the interest degree model library, the function as describing the access path of the portal user and the function AS describing the access path set of the portal user can be stored in the access transaction set. The granularity and method of this data processing are sufficient Considering the completeness of the user interest model and the characteristics of the portal, it is easy to expand, it is convenient for the comparison and calculation of portal user similarity, and it is also conducive to the compatibility and expansion of portal integration.

Step 207: Train and classify the constructed portal user interest model and the stored portal user interest model. The constructed portal user interest model includes the initially created portal user interest model and the updated portal user interest model. The training classification is based on The constructed portal user interest model conducts feature training, extracts the content of interest, behavioral characteristics, etc. to initially classify the categories of interest models and the categories of interest resource content, and continuously updates the portal user interest model. The division methods include portal user interest models, Similarity comparison between resources. It is necessary to comprehensively consider the description of the portal user interest model in terms of interest content, behavior, and preliminary user groups.

Step 208: According to the constructed portal user interest model, the stored portal user interest model and the feedback learning results, perform more accurate similarity calculation. The basis for the neighbor calculation in step 210 is the similarity calculation algorithm. Similar matching and comparison between user interest models are carried out on the basis of classification. The higher the similarity, the greater the probability of generating close neighbors, so it is a clustering process. At the same time, since the drift update result of the portal user interest model returned by the front end is considered, the similarity calculation process in this step is more accurate and sufficient. Finally, a neighbor set of the target portal user is generated.

Step 209: According to the stored interest model and similarity calculation results, determine the selection and combination of recommended control strategies and recommended algorithms.

Step 210: Carry out predictive filtering analysis according to the similarity calculation results, and then perform calculations to determine the recommendation results according to the predictive analysis results and the determined recommendation control strategy and recommendation algorithm, and call the stored WWW resource index according to the determined recommendation results, specifically referring to The prediction process is to predict the resources that the target portal user has not browsed or unknown interest on the basis of the selected target portal user u’s neighbor set, usually based on the neighbor’s related interest history or similar interest content rules, and then from the predicted From the results, the resources that the system thinks the target portal user will be interested in are selected and recommended to the target portal user.

Step 211: Encapsulate the invoked WWW resource index into a portlet containing web page content, and push it to the portal user.

In the present invention, meta-recommendation in the personalized recommendation service refers to the process of organizing and controlling information resources and recommendation algorithms in a unified manner and selecting and pushing information resources and recommendation algorithms by comprehensively considering various needs of portal users’ personalized interests, so as to realize high-level management and control of data and computing .Different recommendation algorithm models can be the input of other recommendation models, which is different from the concept of feature mutual input in combined recommendation, that is to say, instead of using each calculation result as the next input, the algorithm model is directly The whole is used as input, and finally the calculation results are considered comprehensively.

The data set uniformly stores and maintains the attribute set variables related to the meta-recommendation service, and uses the data management module to uniformly operate and call. The interface of the basic data structure is shown in Figure 4. It includes the interest model library, access transaction set, recommendation algorithm set, and recommendation records , recommended content index, user index, and resource display records, etc., and introduce the context triple <Content, User, TimeStamp> to ensure the flexible selection of the meta-recommendation engine.

Table InterestModel and AccessSequence correspond to Interest Model Library and Access Transaction Set respectively. Table User maintains the basic information of portal users as a reference for update and similar calculation. Table Recommendation Record (RecomRecord) It is used to record the algorithm selection and prediction push results of each recommendation process, among which, the attributes user name (User), recommendation algorithm (RecomAlgorithem), user recommendation content (UserContent) and user group recommendation content (UserGroupContent) are all external to the auxiliary context The key identifier is used as the foreign key of the database, the time stamp (TimeStamp) records the recommended time stamp, and whether it is recommended (IfPresented) indicates whether the resource is displayed on the portal. The table recommended content (Content) is a synchronous mapping of the WWW resource index library After predictive analysis, resource links and other information are extracted as configuration parameters of the recommended resource display module, and recorded in the table (recommended display) Presentation.

The recommendation algorithm is the logical structure of a specific calculation method that realizes the recommendation service function, and is the core of the recommendation task. According to the input of interest mining, the recommendation results are calculated through corresponding predictive analysis. The meta-recommendation service architecture in this article does not limit the category and number of recommendation algorithms. The initial key value of each algorithm is used to start the relevant recommendation algorithm. The maximum key value (MaxKey) in the table RecomAlgorithem sets the maximum threshold of the initial key value. The maximum threshold is used to distinguish the level of each algorithm. In the prototype, by comprehensively considering the analysis of the characteristics of the portal user/portal user group’s interest content and behavior, the recommendation algorithm is defined as follows.

Content-based filtering: Do not directly cluster pages, but extract portlet content features for clustering. The calculation method of content feature weight uniform processing is as follows:

$\mathrm{<span\; class="notranslate">\; Weight</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Weight</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Weight</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

The weight setting method, that is, the initial key value is calculated and obtained according to the similarity between indexes of WWW resource content, and then the level is set to facilitate the selection of the value.

Pattern matching based on access transactions: The process of searching and matching paths through the sequence characteristics of access transactions is similar to rule-based predictive calculations. The access patterns among users in the same transaction cluster are similar, and the access patterns among users in different transaction clusters are different. Among them, the access transaction sequence feature weight represents the access strength, which is related to the relevant sequence feature, and the calculation method is as follows:

$\mathrm{<span\; class="notranslate">\; Weight</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; as</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; as</span>}<span\; class="notranslate">\; |</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; as</span>}<span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; Weight</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; Interest\; Action</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Features</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Item-based collaborative filtering: Based on similar item interests, construct a set of k-nearest neighbors {UserGroup} _k , and discover natural clusters of interest based on the k-neighborhood relationship, and generate reference recommendations to target portal users based on the nearest neighbor score. Define Rate(u , p, t) = R(InterestAction(u, p, t)), and obtain portal user behavior feedback mapping to express score implicitly, then the collaborative prediction algorithm for content obtained by portal user u through the nearest neighbor set is as follows:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; Rate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; UserGroup</span>}{<span\; class="notranslate">\; \}</span>}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; sin</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; Rate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; Rate</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span>{<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; UserGroup</span>}<span\; class="notranslate">\; \}</span>}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; Sim</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Among them, v is the neighbor set belonging to portal user u, that is, users similar to portal user u; Sim() indicates the similarity between portal user u and v, and Rate() indicates the average score of portal users. The initial key value can be defined in combination with the joint frequency of occurrence of the content of interest.

Comprehensive filtering and sorting (Top-N): comprehensively consider the interest interaction between portal users and their portal user groups, filter and screen according to the principle of comprehensive prioritization, and sort and sort according to the Top-N method.

The meta-recommendation control strategy is the core of the recommendation selector 107. The connection combination of portal user interest model and recommendation algorithm is established through policy configuration, including portal user/portal user group recommendation control and recommendation algorithm combination control. As shown in Figure 5 The parallel combined scheduling approach shown provides flexible control and comprehensive novel resource prediction, where ① denotes content-based filtering and/or matching based on access transaction patterns, ② denotes community-based partial matching, and ③ denotes item-based collaborative filtering , ④ means comprehensive filtering and sorting. The combination ideas adopted by the prototype in this paper include mixed, cascade and feature augmentation. Among them, mixing refers to using multiple technologies to give multiple recommendation results at the same time; cascading It means that a rough algorithm is first generated by one recommendation technology, and another recommendation technology performs further precise calculation on this basis; feature expansion refers to the addition of feature embedding to the results obtained by one recommendation technology and another recommendation technology as input.

The portal user/portal user group recommendation control strategy decomposes the portal user interest model into two sub-parts that act on the portal user's private and portal user groups, and then extracts the prominent and changing personalized interests of portal users and the stability of representative portal user groups. Finally, it is combined with the reference for predictive analysis. For the optimization of the portal user interest model, feature expansion and cascading methods can be adopted step by step, and Singular Value Decomposition (SVD) can be simplified. Get the neighbor set from rough to precise, reduce the computational complexity, and solve the problems of sparsity and scalability.

The recommendation algorithm combination control strategy is used to automatically select the appropriate recommendation algorithm for prediction and analysis in each link, and each generates a recommendation result as the input of the next step, and finally obtains the prediction results of outstanding interest of portal users and the prediction results of interest of portal user groups, and then mixes them To filter irrelevant and meaningless recommendations, the selection key value with limited priority can be introduced to control the priority, and the personalized interest prediction results of portal users can be obtained. Among them, in order to expand the interest content of portal user groups, the classification method can be improved and optimized, using The group-based partial similarity matching method increases the breadth of item selection and the novelty of unknown content, solves the problem of singularity discovery, and recommends a more accurate and comprehensive neighbor prediction recommendation set. The recommendation results can be used for other similar user groups.

The basic idea of comprehensive filtering and sorting selection of prediction results is: introduce threshold Threshold as an auxiliary threshold to ensure recommendation efficiency, use user category, time condition, and whether to display, etc. And perform Top-N sorting and selection according to the key value KeyValue. After successful push, the IfPresented flag should be set to TRUE.

Different meta-recommendation control strategies use different models and recommendation algorithms to provide different recommendation services, and are promoted by the meta-recommendation engine. In order to meet different recommendation requirements, the meta-recommendation engine can start multiple meta-recommendation control strategies at the same time, by loading the policy configuration , to start different recommendation processes. The control process of the meta-recommendation engine includes the start or stop of the engine and the start or stop of the recommendation algorithm.

As the executor of the recommendation algorithm, the predictive analysis unit 109 runs the relevant recommendation algorithm according to the calling strategy of the recommendation controller 107. The optimization and improvement strategy mentioned above is used to solve hot issues such as sparsity, scalability, cold start and singularity discovery.

Since the prediction needs to consider the novelty of the results and the timing of the recommendation, it cannot be repeated and affect the presentation of other recommendations. Therefore, a learning feedback mechanism based on the updated portal user interest model can be introduced to make appropriate dynamic adjustments. The principle is that content and weight are better than time factor.

Since most of the recommended resources are common Web applications, the transformation and expression of portal-oriented recommended resources is a key issue. The recommended resource presentation unit 111 of the present invention can adopt a method of encapsulating a Web application into a web service ( The encapsulation mechanism of Web Services for Remote Portlets.WSRP) Portlet (Web Application to WSRP Portlet, WA2WP), as shown in Figure 6. By implementing a WSRP producer agent service independent of the portal platform, the recommended target resources are mapped and packaged into corresponding portlets, and released in a manner conforming to the WSRP interface specification, so as to achieve seamless integration and intuitive display with the Portal.

WA2WP consists of Portlet configuration management unit, Portlet session management unit, WSRP interface encapsulation unit, request command analysis unit, Web page acquisition unit and response mark processing unit. Among them, the portlet configuration management unit is used to maintain the metadata of all portlets provided by the current WA2WP. It can extract the corresponding resource parameters from the data set table Presentation, and use XML format files for dynamic configuration, as shown in the figure; the portlet session management unit uses To realize the management of the entire life cycle of the session object; the request command analysis unit is used to analyze the encapsulation display request of the resource link contained in the recommendation result and the resource access user request received, locate the target Portlet, and then locate the target unified resource location character (Uniform Resource Locator, URL), to obtain and prepare the request parameters and session data required to access the target resource; the Web page acquisition unit is used to access the Web application according to the target URL request parameters and session data from the request command analysis unit, and obtain The returned page tag content and cookie data are provided to the response tag processing unit; the response tag processing unit is used to process the obtained page tag content to make it a legal and valid Portlet tag fragment that complies with the WSRP specification; the WSRP interface encapsulation unit It is used to implement a service interface that conforms to the WSRP specification and provides Portal or other aggregator access.

The basic workflow and data interaction process are shown in Figure 7: the request command analysis unit receives the encapsulation display request of the resource link contained in the recommendation result and the user request for accessing the resource, analyzes the request parameters and session data to determine the target resource to be accessed, and sends the request through the Web The page acquisition unit accesses and obtains the web resource page, which may include page tag content and cookie data, and the response tag processing unit preprocesses the hypertext tag information returned by the web page acquisition unit before encapsulation, obtains the web resource page fragment, and then provides it to The WSRP interface encapsulation unit, the WSRP interface encapsulation unit finally encapsulates the processing result, that is, the web resource page fragment into a portlet and displays it on the portal personalized desktop, and the web resource page fragment is a legal and effective Portlet markup fragment conforming to the WSRP specification.

In terms of resource display updates, considering the user's individual interest differences and usage habits, the organization form is basically based on the push weight distribution. The push weight can be weighted based on time importance, novelty, etc., by recommending column channels and identifying the update time Information is pushed step by step. If the portal user modifies the layout or deletes an item, it can be dynamically adjusted and adapted according to the updated feedback of the user interest model.

Obviously, those skilled in the art can carry out various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Like this, if these modifications and variations of the present invention belong to the claims of the present invention and within the scope of equivalent technologies thereof, It is intended that the present invention also encompasses such changes and modifications.

Claims (10)

1. A portal personalized recommendation service system that adopts a meta-recommendation engine, is characterized in that the system includes at least: The data management unit is used to assist in managing the data communication and calling of the training classification unit, or the similarity calculation unit and the data storage unit; The data storage unit is used to store the interest model library of the portal user and/or the portal user group, and the interest model library includes the interest degree model library and the access transaction set of the portal user and/or the portal user group, and is further stored in the data storage unit. Recommended algorithm set; The interest mining unit, located in the portal platform, is used to obtain personalized description files of portal users, implicitly track and capture the interest content and access behavior patterns of logged-in portal users, and provide the obtained information to the interest model construction unit; An interest model construction unit, configured to standardize the acquired interest data, construct a portal user interest model according to the processed information, and provide the constructed portal user interest model to the training classification unit and the similarity calculation unit; The training classification unit is used to call the interest model stored in the data storage unit through the data management unit, and perform the feedback learning of the nearest neighbor clustering on the interest model from the interest model construction unit and the interest model from the data storage unit, and then feed back the learning result Update the portal user interest model stored in the data storage unit through the data management unit and provide it to the similarity calculation unit; The similarity calculation unit is used to call the interest model stored in the data storage unit through the data management unit, and perform the calculation according to the interest model from the interest model construction unit, the feedback learning update result from the training classification unit, and other interest models from the data storage unit. Similarity calculation, and then provide the similarity calculation result to the meta recommendation engine; The meta-recommendation engine is used to call the interest model stored in the data storage unit through the data management unit, and determine the selection and selection of the recommendation control strategy and the recommendation algorithm according to the interest model from the data storage unit and the similarity calculation result from the similarity calculation unit. Combining, and then performing predictive filtering analysis according to the similarity calculation results from the similarity calculation unit, and performing calculations to determine the recommended results according to the predicted analysis results, recommended control strategies and recommendation algorithms, and calling the World Wide Web WWW resource index storage unit according to the determined recommended results The WWW resource index stored in the WWW resource is packaged as a portlet containing Web page content, and pushed to the portal user; The WWW resource index storage unit is used for storing the WWW resource index. 2. system according to claim 1, is characterized in that, described meta recommendation engine comprises: The recommendation selector is used to call the interest model stored in the data storage unit through the data management unit, and determine the selection and selection of the recommended control strategy and the recommendation algorithm according to the interest model from the data storage unit and the similarity calculation result from the similarity calculation unit. combined, and then provided to the predictive analysis unit and provided to the predictive analysis unit with the similarity calculation result from the similarity calculation unit; The predictive analysis unit is used to perform predictive filtering analysis according to the similarity calculation results from the similarity calculation unit, and perform calculations to determine the recommended results according to the predictive analysis results and the recommended control strategy and recommendation algorithm from the recommended selector, and the determined recommended results Provide the recommended resource presentation unit by calling the WWW resource index stored in the WWW resource index storage unit; The recommended resource presentation unit is used to package the WWW resources from the predictive analysis unit into a portlet containing web page content, and push it to the portal user. 3. The system according to claim 2, wherein the recommended resource presentation unit includes: The portlet configuration management unit is used to maintain the metadata of all portlets provided by the encapsulation mechanism WA2WP that conforms to the encapsulation mechanism WA2WP of the current World Wide Web web application; Portlet session management unit, used to manage the entire life cycle of session objects; The request command analysis unit is used to receive the encapsulation display request of the resource link contained in the recommendation result and the user request for accessing the resource, analyze the request parameters and session data to determine the target resource to be accessed, locate the target URL, obtain and prepare for access Request parameters and session data required by the target resource; Web page obtaining unit is used to access the Wed application according to the target URL, request parameters and session data from the request command analysis unit, obtain the returned page mark content and Cookie data, and provide it to the response mark processing unit; The response tag processing unit is used to preprocess the hypertext tag information returned by the Web page acquisition unit before encapsulation, obtain the Wed resource page fragment, and then provide it to the WSRP interface encapsulation unit; The WSRP interface encapsulation unit is used to encapsulate Wed resource page fragments as portlets and display them on the portal personalized desktop. 4. The system according to claim 1, wherein the training and classification unit is further used for: storing the user or user group identification of the established interest model, if the user or user group identification is not stored, the data management unit The interest model after training and classification is provided to the data storage unit for storage. 5. The system according to claim 1, 2 or 3, further comprising a privacy protection unit; The interest mining unit is used to provide the obtained information to the privacy protection unit; The privacy protection unit is used to embed security marks in the information from the interest mining unit for privatization filtering protection, and then provide it to the interest model construction unit. 6. A portal personalized recommendation service method adopting meta-recommendation engine, is characterized in that, the method comprises the following steps: A. Mining the interests of portal users, obtaining personalized description files of portal users, implicitly tracking and capturing the interests and access behavior patterns of logged-in portal users; B. Perform standardized processing, extract information related to portal user interests, and determine whether to create a new portal user interest model, if so, create a new portal user interest model, otherwise, update the existing portal user interest model ; C. Training and classifying the constructed portal user interest model and the stored portal user interest model; D. Carry out similarity calculation according to the constructed portal user interest model, the stored portal user interest model and the feedback learning results; E. According to the stored interest model and similarity calculation results, determine the selection and combination of recommended control strategies and recommendation algorithms, perform predictive filtering analysis according to the similarity calculation results, and then according to the predicted analysis results and the determined recommended control strategies and recommendation algorithms, Executing calculations to determine the recommended results, and calling the stored WWW resource index according to the determined recommended results; F. Encapsulate the invoked WWW resource index into a portlet containing web page content, and push it to the portal user. 7. The method according to claim 6, wherein, between the steps A and B, further comprising: embedding a security mark into the obtained information. 8. The method according to claim 6, wherein said step C is: perform feature training according to the constructed portal user interest model, extract interest content, behavior characteristics and preliminarily divide the category of interest model and the category of interest resource content , and continuously update the portal user interest model. 9. The method according to claim 6, wherein said step D is: on the basis of existing classifications, similar matching and comparison between user interest models are performed to generate a neighbor set of target portal users. 10. The method according to claim 9, wherein the predictive filtering analysis described in step E is: on the basis of the selected target portal user neighbor set, the target portal user has no browsing or unknown interests resources are forecasted.

Images