CN113011942A - Customized product demand collaborative filtering recommendation method based on three-layer neighbor selection framework - Google Patents

Abstract

The invention discloses a customized product demand collaborative filtering recommendation method based on a three-layer neighbor selection framework, which divides a time domain scored by a user into a plurality of time windows, and each window is endowed with a dynamic forgetting factor conforming to a customized product demand preference information attenuation rule and a suggested failure factor conforming to a suggested failure rule; and then combining the two factors to respectively construct a capability evaluation module, a sensitive trust module and a suggestion evaluation module. The three modules are integrated by a three-layer neighbor selection framework, and neighbors with strong prediction capability, trusty capability and suggestion capability are screened out by combining user scoring information; and finally, predicting the item score of the target user through the neighbor score information, and recommending items which may be interested by the user. The method and the device can capture the preference change of the user, select the item score of the proper neighbor prediction target user and recommend the item, relieve the problem of reduced item recommendation precision caused by sparse data, and improve the satisfaction degree of the user on the recommended customized product or scheme.

Description

A collaborative filtering recommendation method for customized product requirements based on a three-layer neighbor selection framework

technical field

The invention relates to a personalized customized product recommendation technology, in particular to a customized product demand collaborative filtering recommendation method based on a three-layer neighbor selection framework.

Background technique

In the "Internet +" environment, users need to determine the needs of the products they need in the huge and confusing product information. By analyzing explicit (such as ratings) or implicit (such as web browsing history) user behavior data, recommender systems can not only shorten the time for users to find the desired products, but also help users explicitly customize products by recommending products that users may be interested in need.

Collaborative filtering (CF) algorithm is widely used in recommender systems because it does not require domain-related feature extraction, and it is very convenient to process unstructured data. CF algorithms can be divided into two categories: (1) Item similarity-based methods (IBCF), which recommend items that are similar to items previously liked by the target user. (2) User similarity-based method (UBCF), which recommends items of neighbor users with similar preferences to the target user. However, the existing methods are difficult to deal with the sparsity of the rating matrix caused by the insufficient number of user rating items and the dynamic change of user preference and recommendation ability affected by time factors. Aiming at this problem, this method constructs three neighbor selection modules on the basis of considering the user's prediction ability, recommendation ability, and credible ability, and adopts the three-layer neighbor selection strategy to dynamically evaluate the user's ability, so as to realize that the user can still be selected in a sparse data environment. Reliable neighbors.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a customized product demand collaborative filtering recommendation method based on a three-layer neighbor selection framework, which can accurately analyze the user's preference through the three-layer neighbor selection framework, select the most suitable neighbors, and alleviate the problems caused by data sparseness. The problem of the decline of recommendation accuracy can better realize the recommendation and improve the user's satisfaction with the recommended items.

In order to achieve the above object, the technical scheme adopted in the present invention is as follows:

A collaborative filtering recommendation method for customized product requirements based on a three-layer neighbor selection framework, the method includes the following steps:

(1) Obtain user rating information, construct a user rating matrix, and users include neighbor users and target users.

(2) Divide the time domain of neighbor user ratings into multiple time windows, and each time window is assigned a dynamic forgetting factor that conforms to the decay law of customized product demand preference information and a suggested failure factor that conforms to the law of suggested failure; wherein, the preference information The decay law is that preference information does not decay in the instantaneous memory stage, rapidly decays in the short-term memory stage, and decays slowly or even no longer in the long-term memory stage. The proposed failure rule is that it is suggested that failure does not occur in the instantaneous memory stage, and it continuously fails in the short-term memory stage and the long-term memory stage.

(3) According to the number of common evaluation items and the asymmetric suggestion ability of neighbor users and target users, a capability evaluation module and evaluation formula are constructed to calculate the prediction ability of each neighbor user.

(4) Combined with the scoring information under each time window and the assigned dynamic forgetting factor, a sensitive trust module is constructed to capture the latest preferences of each neighbor user and the target user, judge their preference consistency, and use the degree of preference consistency as each Criteria for judging the trustworthiness of neighbor users.

(5) According to the number of scoring items used for suggestion and the given suggestion failure factor under each time window, construct a suggestion evaluation module and evaluation formula, calculate the effective suggestion information amount of each neighbor user, and evaluate each neighbor user's Ability to advise.

(6) Based on three modules, a three-layer neighbor selection strategy is introduced, and a three-layer neighbor selection framework is constructed to obtain a neighbor set with strong prediction ability, trustworthiness, and strong recommendation ability;

(7) According to the item scores of each neighbor user in the neighbor set and the ability index value in steps 3-5, predict the target user's score for each item, and generate a recommendation list for the items with the highest predicted score for the user to choose.

Further, in step (2), the time domain of dividing neighbor user scores is a plurality of time windows, wherein, the larger the value of the index mark θ of the time window, the farther the time period corresponding to the window is from the current time, and falls within the time window. Window's ratings are evaluated earlier. The dynamic forgetting factor is divided into the consistent preference information forgetting factor f _con and the divergent preference information forgetting factor f _dif , which are specifically expressed as:

In the formula, γ _i =T _i , γ _s =T _i +T _s , Ti is the instantaneous memory stage, T _s is the short-term memory stage, and T _l is the long-term memory stage _. t _w is the length of the time window, λ ₁ is the unanimous preference information forgetting rate factor, λ ₂ is the divergent preference information forgetting rate factor, the larger their values, the faster the information forgetting rate.

Further, in step (2), when the index of the time window w _θ is marked as θ, it is suggested that the failure factor inv satisfies the function:

where ε is the failure rate factor, which determines the recommended failure rate.

Further, in step (3), the evaluation formula is:

In the formula, |I _u | represents the total number of items rated by target user u; |I _ue | represents the number of items jointly evaluated by target user u and neighbor user e, pcc(u,e) represents the Pearson of target user u and neighbor user e correlation coefficient.

Further, step (4) specifically includes the following substeps:

计算公式为：(4.1) The preference information is divided into consistent preference information and divergent preference information, among which the consistent/divided preference information scored by a single user falling within the time window w _θ

The calculation formula is:

表示邻居用户e在落入时间窗口w_θ的项目v上的评分，R_uv表示目标用户u对项目v的评分。邻居用户e可以在窗口w_θ中与目标用户u有多个共同项目的评分，在每个窗口w_θ下的邻居用户e的所有一致/分歧偏好部分评定分数为：where _Rmax and _Rmin represent the maximum and minimum item ratings, respectively.

represents the rating of the neighbor user e on the item v falling within the time window w _θ , and R _uv represents the rating of the target user u on the item v. Neighbor user e can have scores of multiple items in common with target user u in the window w _θ , and all the consensus/divergence preference partial evaluation scores of neighbor user e under each window w _θ are:

(4.2) Whenever a new score is generated, the previous score preference information will be attenuated. For a time window w _θ , the total amount of consensus/divergence preference information C ^θ (u,e)/D ^θ (u,e) is calculated by:

C ^θ (u,e)=C ^θ+1 (u,e)×(1-f _con (θ))+c ^θ (u,e)D ^θ (u,e)=D ^θ+1 (u, e)×(1-f _dif (θ))+d ^θ (u,e)

(4.3) The sensitive trust module judges the trustworthiness of neighbor users according to the consistency of preference information among users. Under the time window θ=1, evaluate the trustworthiness of the target user u and the neighbor user e. The evaluation formula is as follows:

Further, step (5) is specifically:

In the neighbor user set, the maximum number of items I _rmax for recommendation can be expressed as

I _rmax =max|I _er |

In the formula, |I _er | represents the number of ratings that neighbor user e uses for recommendations, e ∈ N'(u), and N'(u) represents the set of neighbor users.

The suggested information amount m ^θ of neighbor user e under each time window w _θ can be expressed as:

为邻居用户e的评分中落入时间窗口w_θ的建议评分数量。评分携带的建议信息量会随时间减少。对于时间窗口w_θ，；邻居用户的建议信息总量通过下式计算：In the formula,

is the number of suggested ratings that fall within the time window w _θ among the ratings of neighbor user e. The amount of suggested information carried by the score decreases over time. For the time window w _θ , the total amount of suggested information of neighbor users is calculated by the following formula:

M ^θ =M ^θ+1 ×(1-inv(θ))+m ^θ

The proposed capability evaluation formula of neighbor user e is:

rec(u,e)=M ¹ /I _rmax

In the formula, M ¹ is the total amount of suggestion information of the neighbor user e under the time window w ₁ .

Further, in step (6), the three-layer neighbor selection strategy is to set the number of neighbor selection layers to three, and each layer selects a certain number of neighbors through a key module. Assuming that the number of final neighbor selections is K, and the neighbor scale coefficient is ζ, the first layer selects the first ζ ² K neighbors with strong predictive ability through the capability evaluation module, and the second layer selects the first ζK through the sensitive trust module. Neighbors, the third layer selects the top K neighbors with strong recommendation ability through the recommendation evaluation module.

Further, in step (7), after the neighbor selection framework selects neighbors, the following formula is used to predict the score of the target user u on the item v:

为邻居用户e所做所有项目评分的平均值，R_ev为邻居用户e在项目v上的评分，

为目标用户u所做所有项目评分的平均值，N(u)为步骤6筛选获得的邻居集合，per(u,e)为预测能力指标，tru(u,e)为可信能力指标，rec(u,e)为建议能力指标。in,

is the average of all item ratings made by neighbor user e, R _ev is the score of neighbor user e on item v,

is the average of all items scored by the target user u, N(u) is the neighbor set obtained by screening in step 6, per(u,e) is the prediction ability index, tru(u,e) is the credible ability index, rec (u, e) are the suggested capability indicators.

Compared with the current technology, the present invention dynamically analyzes the user's preference through a three-layer neighbor selection framework, selects the most suitable neighbor set, and uses the neighbor data to predict the item score of the target user, so as to alleviate the problem of decreasing recommendation accuracy caused by data sparseness, and improve the performance of the system. Implement recommendations well and improve user satisfaction with recommended products or solutions.

Description of drawings

Figure 1 is a flowchart of a collaborative filtering recommendation method for customized product requirements based on a three-layer neighbor selection framework.

Figure 2 is an architecture diagram of a collaborative filtering recommendation system for customized product requirements based on a three-layer neighbor selection framework.

Figure 3 is a process diagram of the implementation process of collaborative filtering recommendation for elevator customized product decoration requirements.

Detailed ways

The specific embodiments of the present invention will be described more completely and clearly below in conjunction with the accompanying drawings. Figure 1 is a flowchart of a method for collaborative filtering and recommending customized product requirements based on a three-layer neighbor selection framework proposed by the present invention, including the following steps:

(1) Obtain user rating information and construct a user rating matrix;

The user's rating information can be collected from the user's browsing history on the website, historical evaluation data, etc., and converted into a user rating matrix containing multiple users and corresponding multiple item rating information, as the data input of this method.

(2) Divide the time domain into multiple time windows, and each window is assigned a dynamic forgetting factor and a suggested failure factor that conform to the decay law of the preference information and the recommended failure law; the decay period of the rating information refers to the concept of human brain memory and is divided into instantaneous In the memory stage T _i , the short-term memory stage T _s and the long-term memory stage T _l , the calculation formulas of the time intersection nodes are respectively γ _i =T _i and γ _s =T _i +T _s . The decay law of preference information is that preference information does not decay in the instantaneous memory stage, decays rapidly in the short-term memory stage, and decays slowly or even no longer in the long-term memory stage. The suggested failure rule is that the suggestion does not fail in the instantaneous memory stage, and fails continuously in the short-term memory stage and the long-term memory stage. In this embodiment, the recommended value range of T _i is 24h-72h, and the recommended value range of T _s is 72h-144h. The length of the time window is selected according to the actual situation, and the recommended value range is 12h-72h.

The dynamic forgetting factor is divided into a consistent preference information forgetting factor f _con and a divergent preference information forgetting factor f _dif . The index of the time window w _θ is marked as θ, and the index of the current time window is marked as 1, which increases sequentially, that is, the larger the value of the index mark of the time window, the farther the time period corresponding to the window is from the current time, and falls within the Window's ratings are evaluated earlier. The dynamic forgetting factor of each window satisfies the function:

In the formula, t _w is the length of the time window, λ ₁ is the unanimous preference information forgetting rate factor, and λ ₂ is the divergent preference information forgetting rate factor. The larger their values, the faster the information forgetting rate. In this embodiment, the recommended value range of λ ₁ is 0.07-0.072, and the recommended value range of λ ₂ is 0.068-0.07.

The corresponding proposed failure factor inv satisfies the function:

where ε is the failure rate factor, which determines the recommended failure rate. In this embodiment, the recommended value range of ε is 0.004-0.006.

(3) Considering the number of common evaluations and asymmetric suggestion capabilities of neighbors and target users, a capability evaluation module and evaluation formula are constructed. The formula is as follows:

In the formula, |I _u | represents the total number of items rated by target user u; |I _ue | represents the number of items jointly evaluated by target user u and neighbor user e, pcc(u,e) represents the Pearson of target user u and neighbor user e correlation coefficient.

(4) Combine the dynamic forgetting factor, build a sensitive trust module and a trustworthy capability evaluation formula, capture the latest preferences of neighbor users and target users, and judge the preference consistency between users;

和

)，计算公式为：The sensitive trust module evaluates the trustworthiness of neighbor users according to the consistency of preference information among users. Preference information can be divided into consistent preference information and divergent preference information. consensus/divergence preference information for individual user ratings falling within the time window w _θ (marked as

and

), the calculation formula is:

表示邻居用户e在落入时间窗口w_θ的项目v上的评分，R_uv表示目标用户u对项目v的评分。用户e可以在窗口w_θ中与目标用户u有多个共同项目的评分，在窗口w_θ下的用户e的所有一致/分歧偏好部分评定分数为：where _Rmax and _Rmin represent the maximum and minimum item ratings, respectively.

represents the rating of the neighbor user e on the item v falling within the time window w _θ , and R _uv represents the rating of the target user u on the item v. User e can have scores of multiple items in common with target user u in the window w _θ , and the evaluation scores of all the consensus/divergence preference parts of user e under the window w _θ are:

Each time a new rating is generated, the prior rating preference information is attenuated. For a time window w _θ , the total amount of consensus/divergence preference information (ie, C ^θ (u,e) and D ^θ (u,e)) can be calculated by:

C ^θ (u,e)=C ^θ+1 (u,e)×(1-f _con (θ))+c ^θ (u,e)D ^θ (u,e)=D ^θ+1 (u, e)×(1-f _dif (θ))+d ^θ (u,e)

Under the time window θ=1, evaluate the trustworthiness of the target user u and the neighbor user u. The evaluation formula is as follows:

(5) Combined with the recommendation failure factor, construct a recommendation evaluation module and evaluation formula, and calculate the user's effective recommendation information.

Further, in step (5), in the neighbor user set, the maximum number of recommended items I _rmax can be expressed as

I _rmax =max|I _er |

In the formula, |I _er | is the number of ratings that neighbor user e uses for recommendations, e∈N'(u), N'(u) represents the set of neighbor users.

The suggested information amount m ^θ of neighbor user e under each time window w _θ can be expressed as:

为邻居用户e的评分中落入时间窗口w_θ的建议评分数量。评分携带的建议信息量会随时间减少。对于时间窗口w_θ，；邻居用户的建议信息总量通过下式计算：In the formula,

is the number of suggested ratings that fall within the time window w _θ among the ratings of neighbor user e. The amount of suggested information carried by the score decreases over time. For the time window w _θ , the total amount of suggested information of neighbor users is calculated by the following formula:

M ^θ =M ^θ+1 ×(1-inv(θ))+m ^θ

The proposed capability evaluation formula of neighbor user e is:

rec(u,e)=M ¹ /I _rmax

In the formula, M ¹ is the total amount of suggestion information of the neighbor user e under the time window w ₁ .

(6) Introduce a three-layer neighbor selection strategy, build a neighbor selection framework based on three key modules, and select neighbors with strong prediction ability, trustworthiness, and strong recommendation ability.

The three-layer neighbor selection strategy is to set the number of neighbor selection layers to three, and each layer selects a certain number of neighbors through a key module. Assuming that the number of final neighbor selections is K, and the neighbor scale coefficient is ζ, the first layer selects the first ζ ² K neighbors with strong predictive ability through the capability evaluation module, and the second layer selects the first ζK through the sensitive trust module. Neighbors, the third layer selects the top K neighbors with strong recommendation ability through the recommendation evaluation module, and the final set is represented as N(u).

(7) Predict the item rating of the target user through the neighbor rating data, and generate a recommendation list for the items with the highest predicted rating for the user to choose.

After the neighbor selection framework selects neighbors, the target user u's rating on item v is predicted by the following formula:

为邻居的项目评分平均值，R_ev为邻居用户e在项目v上的评分，

为目标用户u的项目评分平均值。最终，将预测评分最高的多个项目生成推荐列表给目标用户。in,

is the average item rating of neighbors, R _ev is the rating of neighbor user e on item v,

Average item ratings for target user u. Finally, multiple items with the highest predicted scores are generated to recommend lists to target users.

As shown in Figure 2, the present invention also designs a collaborative filtering recommendation system based on a three-layer neighbor selection framework, including:

(1) Interactive interface: Users can select products and express their needs through the interactive interface.

(2) Data extraction module: extract the user's personal information and demand information in the interactive page, retrieve the information of similar users and similar items in the user database, and output the user rating matrix.

(3) Ability neighbor selection module: Calculate the ability evaluation scores of similar users according to the prediction ability evaluation formula, and filter out the neighbor sets with strong prediction ability in the first layer.

(4) Trusted neighbor selection module: Calculate the trustworthiness of the first-layer neighbors according to the trustworthy capability evaluation formula, and screen out the neighbors with strong second-tier trustworthy capabilities.

(5) Suggestion neighbor selection module: According to the suggestion ability evaluation formula, the effective suggestion information amount of the second-layer neighbors is calculated, and the set of neighbors with strong third-layer suggestion ability is screened out.

(6) Prediction score calculation module: predicts the target user's score for each item through the item score and ability index value of each neighbor user in the neighbor set.

(7) Item recommendation module: generate a recommendation list for multiple items with the highest predicted scores to the user.

FIG. 3 is a process diagram for realizing the corresponding elevator product decoration requirement recommendation according to the present invention.

(1) The user selects the favorite item on the elevator product decoration selection page, and obtains the user's demand information and user information after processing by the data extraction module. Combined with the rating information of related users and related items, a user rating matrix is jointly constructed.

(2) According to the user score matrix, the neighbors with strong prediction ability, trustworthiness and strong recommendation ability are screened layer by layer, and the neighbor item score data of each layer is retained through the data storage module to realize asynchronous screening of neighbors and improve recommendation efficiency.

(3) Calculate the predicted scores of the unrated items of the target user through the predicted score calculation module and the neighbor item score matrix, and generate a recommendation list for the items with the highest predicted scores, transmit them to the interactive interface, and recommend them to the target user.

The above are preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Contents that are conceived and equivalently replaced or changed should all be included within the protection scope of the present invention.

Claims (8)

1. a customized product requirement collaborative filtering recommendation method based on three-layer neighbor selection framework, is characterized in that, this method comprises the steps: (1) Obtain user rating information, construct a user rating matrix, and users include neighbor users and target users. (2) Divide the time domain of neighbor user ratings into multiple time windows, and each time window is given a dynamic forgetting factor that conforms to the decay law of customized product demand preference information and a suggested failure factor that conforms to the law of suggested failure; wherein, the preference information The decay law is that preference information does not decay in the instantaneous memory stage, rapidly decays in the short-term memory stage, and decays slowly or even no longer in the long-term memory stage. The proposed failure rule is that it is suggested that failure does not occur in the instantaneous memory stage, and it continuously fails in the short-term memory stage and the long-term memory stage. (3) According to the number of common evaluation items and the asymmetric suggestion ability of neighbor users and target users, a capability evaluation module and evaluation formula are constructed to calculate the prediction ability of each neighbor user. (4) Combined with the scoring information under each time window and the assigned dynamic forgetting factor, a sensitive trust module is constructed to capture the latest preferences of each neighbor user and the target user, judge their preference consistency, and use the degree of preference consistency as each Criteria for judging the trustworthiness of neighbor users. (5) According to the number of scoring items used for suggestion and the given suggestion failure factor under each time window, build a suggestion evaluation module and an evaluation formula, calculate the effective suggestion information amount of each neighbor user, and evaluate each neighbor user's Ability to advise. (6) Based on three modules, a three-layer neighbor selection strategy is introduced, and a three-layer neighbor selection framework is constructed to obtain a neighbor set with strong prediction ability, trustworthiness, and strong recommendation ability; (7) According to the item scores of each neighbor user in the neighbor set and the ability index value in steps 3-5, predict the target user's score for each item, and generate a recommendation list for the items with the highest predicted score for the user to choose. 2. a kind of customized product demand collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: in step (2), the time domain of dividing neighbor user score is a plurality of time windows, wherein, The larger the value of the index mark θ of the time window, the farther the time period corresponding to the window is from the current time, and the earlier the evaluation time of the scores falling into the window. The dynamic forgetting factor is divided into the consistent preference information forgetting factor f _con and the divergent preference information forgetting factor f _dif , which are specifically expressed as:

In the formula, γ _i =T _i , γ _s =T _i +T _s , Ti is the instantaneous memory stage, T _s is the short-term memory stage, and T _l is the long-term memory stage _. t _w is the length of the time window, λ ₁ is the unanimous preference information forgetting rate factor, λ ₂ is the divergent preference information forgetting rate factor, the larger their values, the faster the information forgetting rate. 3. a kind of customized product requirement collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: in step (2), when the index of time window w _θ is marked as θ, it recommends failure factor inv satisfies the function:

where ε is the failure rate factor, which determines the recommended failure rate. 4. a kind of customized product demand collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: in step (3), evaluation formula is:

In the formula, |I _u | represents the total number of items rated by target user u; |I _ue | represents the number of items jointly evaluated by target user u and neighbor user e, pcc(u,e) represents the Pearson of target user u and neighbor user e correlation coefficient. 5. a kind of customized product requirement collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: step (4) specifically comprises the following substeps: (4.1) The preference information is divided into consistent preference information and divergent preference information, among which the consistent/divided preference information scored by a single user falling within the time window w _θ

The calculation formula is:

where _Rmax and _Rmin represent the maximum and minimum item ratings, respectively.

represents the rating of the neighbor user e on the item v falling within the time window w _θ , and R _uv represents the rating of the target user u on the item v. Neighbor user e can have scores of multiple items in common with target user u in the window w _θ , and all the consensus/divergence preference partial evaluation scores of neighbor user e under each window w _θ are:

(4.2) Whenever a new score is generated, the previous score preference information will be attenuated. For a time window w _θ , the total amount of consensus/divergence preference information C ^θ (u,e)/D ^θ (u,e) is calculated by: C ^θ (u,e)=C ^θ+1 (u,e)×(1-f _con (θ))+c ^θ (u,e)D ^θ (u,e)=D ^θ+1 (u, e)×(1-f _dif (θ))+d ^θ (u,e) (4.3) The sensitive trust module judges the trustworthiness of neighbor users according to the consistency of preference information among users. Under the time window θ=1, evaluate the trustworthiness of the target user u and the neighbor user e. The evaluation formula is as follows:

6. a kind of customized product demand collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: step (5) is specifically: In the neighbor user set, the maximum number of items I _rmax for recommendation can be expressed as I _rmax =max|I _er | In the formula, |I _er | represents the number of ratings that neighbor user e uses for recommendations, e ∈ N'(u), and N'(u) represents the set of neighbor users. The suggested information amount m ^θ of neighbor user e under each time window w _θ can be expressed as:

In the formula,

is the number of suggested ratings that fall within the time window w _θ among the ratings of neighbor user e. The amount of suggested information carried by the score decreases over time. For the time window w _θ , the total amount of suggested information of neighbor users is calculated by the following formula: M ^θ =M ^θ+1 ×(1-inv(θ))+m ^θ The proposed capability evaluation formula of neighbor user e is: rec(u,e)=M ¹ /I _rmax In the formula, M ¹ is the total amount of suggestion information of the neighbor user e under the time window w ₁ . 7. a kind of customized product requirement collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: in step (6), three-layer neighbor selection strategy is to set neighbor selection layer number to be three layers, Each layer filters a certain number of neighbors through a key module. Assuming that the number of final neighbor selections is K, and the neighbor scale coefficient is ζ, the first layer selects the first ζ ² K neighbors with strong predictive ability through the capability evaluation module, and the second layer selects the first ζK through the sensitive trust module. Neighbors, the third layer selects the top K neighbors with strong recommendation ability through the recommendation evaluation module. 8. a kind of customized product demand collaborative filtering recommendation method based on three-layer neighbor selection framework according to claim 1, is characterized in that: in step (7), after neighbor selection framework selects neighbor, predicts target user u by following formula Rating on item v:

in,

is the average of all item ratings made by neighbor user e, R _ev is the score of neighbor user e on item v,

is the average of all items scored by the target user u, N(u) is the neighbor set obtained by screening in step 6, per(u,e) is the prediction ability index, tru(u,e) is the credible ability index, rec (u, e) are the suggested capability indicators.

Images