A fast cloud service recall recommendation method based on multi-model fusion

Abstract

The present invention provides a fast cloud service recall recommendation method based on multi-model fusion. First, the recommended items are cleaned and preprocessed, including but not limited to de-redundancy, denoising, and zero-filling. The numerical, text, and image features of the items are extracted. A local sensitive hash multi-bucket strategy cloud service recall method based on normal distribution is designed to solve the problems of high latency and easy loss of neighboring points in large-scale cloud service recommendations, thereby improving the real-time performance of recommendations. At the same time, an integrated learning method is used to fuse the strengths of multiple models to solve the problems of lack of diversity in cloud service recommendation results and limitations of the models themselves, thereby improving the model recommendation effect.

Description

Technical Field

[0001] The present invention relates to the field of mobile Internet technology, and in particular to a fast cloud service recall recommendation method based on multi-model fusion Background Art

[0002] Recommendation methods of existing technologies: In existing cloud service recall methods, there are problems with bucket boundary solidification or unreasonable generalization parameters, which makes it easy to lose neighboring candidate points (cloud services); on the other hand, single recommendation models have their own limitations, and there is still room for improvement in the model recommendation effect.

[0003] By using the improved cloud service fast recall method and multi-model fusion method, the real-time recommendation and recommendation effect of the model are improved. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a fast cloud service recall recommendation method based on multi-model fusion to solve the above-mentioned technical problems.

[0005] In order to achieve the above objectives, the present invention is implemented through the following technical solutions: First, the recommended items are cleaned and preprocessed, including but not limited to redundancy removal, denoising, zero-filling, and extracting the numerical, text, and image features of the items.

[0006] For further optimization, an improved coarse sorting algorithm based on the normally distributed local sensitive hashing strategy is used to roughly sort and recall the recommended candidate set, and candidate items at the million level or even higher are initially screened and reduced to the ten thousand level or below, and then input into the fine ranking layer for fine sorting and recommendation. This section mainly improves the locally sensitive hashing recall method. Since large-scale cloud service recommendations often have problems of high latency and slow response, and the locally sensitive hashing multi-bucket recall strategy has the problem of easy loss of neighboring points due to the solidification of bucket boundaries, the present invention adds a normally distributed random bias term on the basis of locally sensitive hashing, generalizes the bucket boundaries, and improves the recall rate of cloud service neighboring candidate points. The expression of the improved locally sensitive hashing multi-bucket recall strategy is shown in the formula:

[0007]

[0008] in, is a floor operation, ^ is an AND operation, x is a k-dimensional embedding vector in a high-dimensional space, v is a randomly generated k-dimensional mapping vector, w is the bucket width, and bn is arrive A normally distributed random variable between , to avoid the solidification of bucket boundaries.

[0009] Further optimization is performed by inputting into the fine ranking layer (multi-model fusion recommendation model), scoring and fine ranking the items, and finally returning the topN recommendation result list. The multi-model fusion recommendation model has its own strengths and limitations for a single model. At the same time, in order to improve the diversity of cloud service recommendation results, through the ensemble learning method, multiple weak click-through rate prediction models are integrated to form a strong click-through rate prediction model to improve the model recommendation effect.

[0010] In summary, the present invention provides a fast cloud service recall recommendation method based on multi-model fusion. The present invention solves the problems of high latency and easy loss of neighboring points in large-scale cloud service recommendations by designing a local sensitive hash multi-bucket strategy cloud service recall method based on normal distribution, thereby improving the real-time performance of recommendations. At the same time, by using an integrated learning method to fuse the strengths of multiple models, the present invention solves the problems of lack of diversity in cloud service recommendation results and the limitations of the models themselves, thereby improving the model recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the fast cloud service recall layer processing flow of the present invention;

[0012] Figure 2 It is a schematic diagram of the structure of the multi-model fusion method of the present invention; DETAILED DESCRIPTION

[0013] The present invention will be further described below in conjunction with the accompanying drawings.

[0014] like Figure 1-2 As shown, the present invention is a fast cloud service recall recommendation method based on multi-model fusion. First, the recommended items are cleaned and preprocessed, including but not limited to de-redundancy, denoising, and zero-value filling, and the numerical, text, and image features of the items are extracted.

[0015] For further optimization, an improved coarse sorting algorithm based on the normally distributed local sensitive hashing strategy is used to roughly sort and recall the recommended candidate set, and candidate items at the million level or even higher are initially screened and reduced to the ten thousand level or below, and then input into the fine ranking layer for fine sorting and recommendation. This section mainly improves the locally sensitive hashing recall method. Since large-scale cloud service recommendations often have problems of high latency and slow response, and the locally sensitive hashing multi-bucket recall strategy has the problem of easy loss of neighboring points due to the solidification of bucket boundaries, the present invention adds a normally distributed random bias term on the basis of locally sensitive hashing, generalizes the bucket boundaries, and improves the recall rate of cloud service neighboring candidate points. The expression of the improved locally sensitive hashing multi-bucket recall strategy is shown in the formula:

[0016]

[0017] in, is a floor operation, ^ is an AND operation, x is a k-dimensional embedding vector in a high-dimensional space, v is a randomly generated k-dimensional mapping vector, w is the bucket width, and bn is arrive A normally distributed random variable between , to avoid the solidification of bucket boundaries.

[0018] Further optimization is performed by inputting into the fine ranking layer (multi-model fusion recommendation model), scoring and fine ranking the items, and finally returning the topN recommendation result list. The multi-model fusion recommendation model has its own strengths and limitations for a single model. At the same time, in order to improve the diversity of cloud service recommendation results, through the ensemble learning method, multiple weak click-through rate prediction models are integrated to form a strong click-through rate prediction model to improve the model recommendation effect.

[0019] In summary, the present invention provides a fast cloud service recall recommendation method based on multi-model fusion. The present invention solves the problems of high latency and easy loss of neighboring points in large-scale cloud service recommendations by designing a local sensitive hash multi-bucket strategy cloud service recall method based on normal distribution, thereby improving the real-time performance of recommendations. At the same time, by using an integrated learning method to fuse the strengths of multiple models, the present invention solves the problems of lack of diversity in cloud service recommendation results and the limitations of the models themselves, thereby improving the model recommendation effect.

Claims

1. A fast cloud service recall recommendation method based on multi-model fusion, characterized in that: This recall recommendation method first cleans and preprocesses the recommended items, including redundancy removal, denoising, and zero-value filling, and extracts the numerical, text, and image features of the items; The improved local sensitive hashing strategy rough sorting algorithm based on normal distribution is used to roughly sort and recall the recommended candidate set, and the candidate items at the million level and above are initially screened and reduced to the ten thousand level or below. On the basis of local sensitive hashing, a random bias term of normal distribution is added to generalize the bucket boundary and improve the recall rate of the cloud service's neighboring candidate points. The improved local sensitive hashing multi-bucket recall strategy expression is shown in the formula: in, is a floor operation, ^ is an AND operation, x is a k-dimensional embedding vector in a high-dimensional space, v is a randomly generated k-dimensional mapping vector, w is the bucket width, and bn is arrive A normally distributed random variable to avoid the solidification of bucket boundaries; The items are input into the fine ranking layer, scored and sorted, and finally a list of topN recommended results is returned. The fine ranking layer model adopts a multi-model fusion recommendation model through an integrated learning method to integrate multiple weak click-through rate prediction models to form a strong click-through rate prediction model, thereby improving the model recommendation effect.

Images