Inner product retrieval methods and electronic devices
By constructing a norm-adjusted neighbor graph (NAPG), the inefficiency of inner product search methods on large-scale datasets is addressed, and the edge selection process is optimized, resulting in more efficient inner product search and improved performance of online recommendation systems.
Patent Information
- Application Number
- CN202210799585.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2022-02-18
- Filing Date
- 2022-07-06
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2042-07-06
AI Technical Summary
Existing inner product retrieval methods are inefficient on large-scale datasets, especially in online recommendation systems where they struggle to quickly and effectively return relevant data. Traditional ANN search methods and graph-based methods such as HNSW perform poorly in inner product search, and edge selection algorithms have issues that lead to inefficient index graphs.
By constructing a Norm Adjusted Proximity Graph (NAPG), adjusting the norm range of data points and estimating local and global adjustment factors, the edge selection process is optimized, resulting in a more efficient index graph suitable for inner product search.
It significantly improves the efficiency and accuracy of inner product search, especially in the retrieval of top-k results, thereby enhancing the performance of online recommendation systems.
Smart Images

Figure CN115618096B_ABST
Abstract
Description
Technical Field
[0001] This disclosure generally relates to systems and methods for computer learning, which can provide improved computer performance, features, and usability. More specifically, this disclosure relates to a computer-implemented inner product retrieval method, an electronic device implementing the method, a computer-readable medium, and a computer program product. Background Technology
[0002] In recent years, there has been a significant increase in the penetration rate of computing devices, the quantity and capacity of data storage, and data collection. As more and more computing devices are used in increasingly diverse ways and data storage costs continue to decrease, generating and / or collecting data has become easier and cheaper. This explosive growth in available data has numerous applications and is extremely valuable.
[0003] However, simply having a large amount of data is not useful; what matters is returning relevant data when users search. Here, the increase in data volume creates problems. Searching for relevant data in a very large dataset becomes more difficult and time-consuming. Furthermore, regardless of the size of the searched corpus, users expect not only high-quality results but also fast results. Therefore, researchers are constantly striving to find better systems and methods to return relevant information quickly and efficiently.
[0004] Efficient inner product search of embedded vectors is a crucial stage in online ranking services such as recommendation and information retrieval. Recommendation algorithms, such as matrix factorization, typically generate latent vectors representing users or items. Recommendation services are performed by retrieving the most relevant item vectors given a user vector, where relevance is usually defined by the inner product. Therefore, developing efficient recommendation systems often requires solving the so-called Maximum Inner Product Search (MIPS) problem, which severely impacts the performance of online recommendation systems.
[0005] The inner product does not follow the triangle inequality of metric spaces, making approximation of MIPS a unique and challenging problem. Classical approximate nearest neighbor (ANN) search methods have limited performance when extended to MIPS. The state-of-the-art graph-based ANN method, HNSW, does not perform well in MIPS because its critical part edge selection algorithm is based on the triangle inequality.
[0006] Therefore, a better method is needed to identify and return relevant data. Summary of the Invention
[0007] According to one aspect of the present disclosure, a computer-implemented inner product retrieval method is provided, comprising: given a latent vector set containing multiple latent vectors, obtaining a norm range graph, the norm range graph including multiple subranges and an estimated adjustment factor for each subrange; constructing an index graph using norm-adjusted inner products, the norm-adjusted inner products including estimated adjustment factors for the multiple subranges; and given a query vector, performing a greedy search on the constructed index graph to retrieve one or more relevant vectors.
[0008] According to another aspect of the present disclosure, an electronic device is provided, comprising: one or more processors; and a non-transitory computer-readable medium including one or more sets of instructions, which, when executed by at least one of the one or more processors, perform an inner product retrieval method according to the present disclosure.
[0009] According to another aspect of the present disclosure, a non-transitory computer-readable medium is provided, including one or more instruction sequences that, when executed by at least one processor, perform an inner product retrieval method according to the present disclosure.
[0010] According to another aspect of the present disclosure, a computer program product is provided, including a computer program stored on at least one of a readable storage medium and an electronic device, wherein when executed by a processor, the computer program performs an inner product retrieval method according to the present disclosure. Attached Figure Description
[0011] Reference will be made to embodiments of this disclosure, examples of which are shown in the accompanying drawings. These drawings are intended to be illustrative and not restrictive. Although this disclosure is generally described in the context of these embodiments, it should be understood that the scope of this disclosure is not limited to these particular embodiments. Items in the drawings may not be drawn to scale.
[0012] Figure 1 The norm distribution analysis of the two datasets is presented. In these cases, the norm of most data points is much smaller than that of the data points with the larger norm.
[0013] Figure 2A and Figure 2B A diagram illustrating edge selection is provided. Figure 2A The edge selection method for the metric space is described. Figure 2B Describing the same edge selection failed to extend to the inner product.
[0014] Figure 3 The process of estimating the global adjustment factor according to embodiments of the present disclosure is described.
[0015] Figure 4 The process of estimating local adjustment factors according to embodiments of the present disclosure is described.
[0016] Figure 5 The NAPG construction process according to embodiments of this disclosure is described.
[0017] Figure 6 The process of generating an index map according to an embodiment of the present disclosure is described.
[0018] Figures 7A to 7F The comparison of various methods based on recall and time, according to embodiments of the present disclosure, is illustrated graphically.
[0019] Figure 8 The comparison of the first 100 tags by recall and calculation results according to embodiments of the present disclosure is depicted graphically.
[0020] Figure 9 The parametric analysis (905) of the norm range number N in the NAPG embodiment according to an embodiment of the present disclosure and the study of factor α estimation (910) are described.
[0021] Figure 10 A simplified block diagram of a computing device / information processing system according to an embodiment of the present disclosure is shown. Specific Implementation
[0022] In the following description, specific details are set forth for purposes of explanation in order to provide an understanding of this disclosure. However, it will be apparent to those skilled in the art that this disclosure may be practiced without these details. Furthermore, those skilled in the art will recognize that the embodiments of this disclosure described below may be implemented in various ways, such as processes, apparatus, systems, devices, or methods, on a tangible computer-readable medium.
[0023] The components or modules illustrated in the accompanying drawings are exemplary illustrations of embodiments of this disclosure and are intended to avoid obscuring the disclosure. It should also be understood that throughout this discussion, a component may be described as a separate functional unit, which may include subunits; however, those skilled in the art will recognize that various components or portions thereof may be divided into separate components or integrated together, including, for example, within a single system or component. It should be noted that the functions or operations discussed herein may be implemented as components. Components may be implemented in software, hardware, or a combination thereof.
[0024] Furthermore, the connections between components or systems shown in the accompanying drawings are not intended to be limited to direct connections. Instead, data between these components may be modified, reformatted, or otherwise altered by intermediate components. Additionally, additional or fewer connections may be used. It should also be noted that the terms “coupled,” “connected,” or “communicably coupled,” “connected to via an interface,” “interface,” or any derivative thereof should be understood to include direct connections, indirect connections via one or more intermediate devices, and wireless connections. It should also be noted that any communication, such as signals, responses, replies, acknowledgments, messages, queries, etc., may include one or more exchanges of information.
[0025] In this specification, references to "one or more embodiments," "preferred embodiments," "an embodiment," "an embodiment," etc., indicate that a specific feature, structure, characteristic, or function described in connection with an embodiment is included in at least one embodiment of this disclosure, and may be included in more than one embodiment. Furthermore, the appearance of the above-mentioned terms throughout this specification does not necessarily indicate the same embodiment or multiple identical embodiments.
[0026] Certain terms used throughout this specification are for illustrative purposes and should not be construed as limiting. A service, function, or resource is not limited to a single service, function, or resource; the use of these terms may indicate a distributed or aggregated grouping of related services, functions, or resources. The terms “comprising,” “including,” “containing,” and “includes” should be understood as open-ended terms, and anything listed thereafter is an instance and not intended to be limited to the listed items. A “layer” may include one or more operations. “Optimized,” “optimized,” “optimized,” etc., refer to an improvement in a result or process and do not require that the specified result or process reach an “optimal” or peak state. Memory, database, repository, data storage, table, hardware, cache, etc., may be used herein to refer to system components or components in which information can be entered or otherwise recorded. A set may contain any number of elements, including an empty set.
[0027] In one or more embodiments, stopping conditions may include: (1) a set number of iterations have been performed; (2) a processing time has been reached; (3) convergence (e.g., the difference between successive iterations is less than a first threshold); (4) deviation (e.g., performance degradation); (5) an acceptable result has been achieved; and (6) all data has been processed.
[0028] Those skilled in the art will recognize that: (1) certain steps may be performed optionally; (2) the steps may not be limited to the specific order set forth herein; (3) certain steps may be performed in a different order; and (4) certain steps may be performed simultaneously.
[0029] Any headings used herein are for organizational purposes only and should not be used to limit the scope of the specification or claims. Each reference mentioned in this disclosure is incorporated herein by reference in its entirety.
[0030] It should be noted that any experiments and results provided herein are presented in a graphical manner and performed under specific conditions using specific embodiments or multiple embodiments; therefore, neither these experiments nor their results should be used to limit the scope of disclosure of the present application.
[0031] Overview
[0032] With the widespread adoption of representation learning techniques, retrieving similar vectors in embedding spaces has become a common operation in various applications. For example, collaborative filtering (CF) projects users and items into a shared latent space, where users or items are represented as vectors of latent features. A user's previous ratings of an item are then modeled as a matching function of its latent vectors. In online recommendation services, given a user vector, the system retrieves and returns the item vector with the highest matching score. Similarly, for question-answering (QA) questions, questions and documents (containing answers) can be embedded into a common semantic space. For each query question, the system can efficiently return candidate answers based on vector comparisons in the embedding space. These candidate answers can then be further scored using a well-trained ranking model. Recently, researchers have found that, through transformer-based models and carefully designed pre-training, this vector-based retrieval method even outperforms the classic word-based BM25 (Best Match 25, an information retrieval method that ranks a set of documents based on query terms) in generating candidate answers.
[0033] The inner product is a common and efficient search / ranking measure used in recommendation models and retrieval-based question answering, for example, for retrieving similar vectors. Another common measure, cosine similarity, can be viewed as a special case of the inner product over normalized vectors. Inner product retrieval is often referred to in the literature as Maximum Inner Product Search (MIPS), and it also has wide applications in multi-label classification and natural language processing (NLP). Since the retrieval space is typically very large, such as millions or billions, linear brute-force ranking is too slow for online services. Therefore, accelerating vector retrieval will benefit the application of the model in real-world systems. In practice, approximate and sublinear complexity ranking methods can be sought, striking an appropriate trade-off between search efficiency and search effectiveness. The efficient inner product ranking problem can be viewed as an approximation of MIPS.
[0034] Approximate MIPS is considered a challenging problem because its inner product does not satisfy the triangle inequality. Traditional ANN search methods are designed for metrics such as l2 distance or cosine similarity, so directly extending them to MIPS is inefficient. Tree-based methods have been proposed to address the approximate MIPS problem. Recently, a method has been proposed to transform MIPS into a traditional ANN search by asymmetrically boosting the basic data vector and query vector to a higher dimension. After the transformation, well-developed ANN search methods can be applied to solve the MIPS problem. Other proposals for MIPS tasks include quantization methods and greedy MIPS.
[0035] Specifically, a graph-based approximation MIPS method, ip-NSW, is proposed, which solves MIPS using the currently popular ANN search algorithm, Hierarchical Navigable Small World (HNSW). Compared to HNSW, ip-NSW simply replaces the metric with an inner product. Generally, the framework of a graph-based ANN search algorithm can be summarized in two steps: (1) Constructing a neighborhood graph where each vertex represents a basic data vector. Each data vector is connected to several neighbors. (2) Given a query vector, performing a greedy search on the neighborhood graph by comparing the query vector and the basic data vector under a search metric (e.g., cosine similarity or inner product). Then, the most similar candidate is returned as the output. The key to the above two-step method is constructing a high-quality index graph, striking a proper balance between search efficiency and search effectiveness. To ensure search efficiency, the degree of each vertex is usually limited to a small number, such as 16. Simultaneously, it is desirable for its neighbors to have high diversity, thus ensuring search effectiveness.
[0036] For the graph-based MIPs method ip-NSW, although experiments show it achieves significant improvements over previous MIPS methods, problems were found with its HNSW-based algorithm. A crucial component of HNSW, the edge selection algorithm, is specifically designed for the metric space to guarantee neighborhood diversity. A coarse substitution from metric measurement to inner product can cause the edge selection method to fail, resulting in invalid proximity graphs. To address this issue, a dedicated edge selection algorithm for the inner product is proposed. However, this proposed method, namely the Inner Product Delaunay Graph (IPDG), is generally suitable for top-1 recall.
[0037] To address this issue, this disclosure presents a novel MIPS method embodiment based on proximity graphs, which for convenience may be collectively referred to as the Norm-Adjusted Proximity Graph (NAPG) embodiment. In one or more embodiments, the l2 norm of data points is adjusted when edge selection is performed to fit the inner product. One of the motivations for this method is based on the long-tailed l2 norm distribution in real datasets. The inventors of this disclosure observed that the l2 norm of most basis vectors is much smaller than that of data points with large norms, such as... Figure 1 The example shown. Figure 1 Norm distribution analysis is depicted for two datasets (dataset 1 and dataset 2). In the datasets shown, the norm of most data points is much smaller than that of data points with larger norms. For example, when performing edge selection via ip-NSW, a data vector x with a smaller l2 norm has neighbor candidate points {p i}, but only connect the vertex p with which x has the largest inner product, and not connect other vertices with a large inner product, because x and {p} i The inner product between} and {p i The inner product values between} are completely incomparable. This leads to inefficient graphics structures and severely impacts MIPS performance. In one or more embodiments, the data vector can be rescaled by adjusting the l2 norm of data points in the smaller l2 norm using a carefully chosen factor α, so that... and They have the same expected value. Thus, the edge selection method is reapplied to the inner product, allowing for the construction of more meaningful (and diverse) edges. The final indexed graph improves connectivity and delivers performance gains. The key adjustment factor α can be estimated based on small sample data. Furthermore, the factor can be refined to a finer granularity for each norm subrange.
[0038] In experiments, embodiments of the introduced method NAPG were evaluated on six datasets. These datasets had different sizes and were extracted using different embedding models. Compared to competing baselines, the NAPG embodiments demonstrated significant advantages in different test cases. Some key contributions include, but are not limited to, the following:
[0039] A norm adjustment factor is introduced for MIPS. In one or more embodiments, this factor bridges the gap between the metric and the inner product, thereby effectively adapting to graph-based MIPS methods.
[0040] An example of an estimation method for the global adjustment factor is introduced, which can be refined to a finer granularity for each norm subrange.
[0041] Extensive experiments were conducted to evaluate embodiments of this estimation method. The results demonstrate the effectiveness and efficiency of this embodiment for approximate MIPS problems.
[0042] MIPS issues:
[0043] In terms of form, in space In the context of query / user latent vector set S, u ={q1, ..., q n} and the basic / project latent vector set S i ={x1, ..., x m The ranking model defines a continuous symmetric similarity function. And the goal of optimal similarity vector retrieval is to find:
[0044]
[0045] Here, we discuss common cases of vector retrieval using the inner product, namely... The problem concerning the inner product in Equation (1) is often referred to as Maximum Inner Product Search (MIPS), which is closely related to but significantly different from the traditional nearest neighbor search task. This is because the inner product has different similarity to l2 distance or cosine similarity and does not satisfy the triangle inequality.
[0046] Formula (1) can be reformulated as a negative l2 distance problem or a cosine similarity problem:
[0047]
[0048]
[0049] It is important to note that querying the L2 norm of q does not affect the ranking.
[0050] When x∈S i When all l2 norms are constant, the search problems defined in equations (2) and (3) are equivalent to MIPs. However, the l2 norms of the data points may differ significantly from each other in practice, such as Figure 1 The examples in the text make MIPs a task that distinguishes them from traditional ANN search.
[0051] Brute-force methods for solving MIPS are computationally expensive. Some researchers have attempted to accelerate this task through novel indexing and pruning techniques. These techniques fall into two categories: (1) fast and accurate MIPS; and (2) fast approximate MIPS, which will be described in detail in the following subsections.
[0052] In practice, the goal is not only to retrieve the most relevant vectors, but also to retrieve the top k vectors. Simultaneously, due to latency constraints, another objective is to retrieve these vectors as quickly as possible. Therefore, it becomes:
[0053] make For the following set:
[0054]
[0055] f(x, q) ≥ f(x′, q)
[0056] Within time t, the goal is to find k candidate vectors C that satisfy the following conditions. k :
[0057] In satisfying In the case of maximize|C k ∩T k |
[0058] Where T k C represents the top k best-fit vectors. k It is the candidate vector returned by the method within a limited time t.
[0059] Existing solutions for accurate MIPs
[0060] Efficient and accurate MIPS primarily utilizes matrix computation and pruning techniques. LEMP (derived from finding large entries in matrix multiplication) transforms the MIPS problem into numerous smaller cosine similarity search problems. This transformation allows LEMP to quickly prune most of the search space and select an appropriate search algorithm for each remaining problem. Fast and accurate inner product retrieval based on sequential scanning (FEXIPRO) has been proposed. FEXIPRO applies a singular value decomposition (SVD) transformation to the basis vector matrix, then captures most of the inner product in the first few dimensions. This prunes the basis vectors by computing partial inner products between the basis vectors and the query vector. Furthermore, it has been pointed out that the fastest indexes today are not always superior to block matrix multiplication, leading to the proposal of MAXIMUS, a simple hardware-optimized indexing scheme that utilizes a linear algebra kernel to improve hardware efficiency while pruning computation. OPTIMUS has also been designed, a system that claims to efficiently choose between using indexes and block matrix multiplication. While these methods appear to be much more efficient than brute-force scanning, they are only suitable for medium-sized data and data dimensionality.
[0061] Existing solutions similar to MIPS
[0062] Because the inner product is a non-metric similarity measure that does not conform to the triangle inequality, it renders many well-studied ANN search techniques ineffective or inefficient. To overcome the variation of the l2 norm, previous work simplified MIPS to an ANN search problem in a metric space by performing asymmetric preprocessing on the basic data and queries. Recently, researchers have found that these methods have performance limitations due to the long-tailed distribution of data norms. Based on these observations, a method called Norm Range Local Sensitive Hashing (LHS) has been devised. In this method, the dataset is first divided into small subsets according to the l2 norm level, and then the data is normalized using a local maximum l2 norm, with a transformation applied to each subset. Similarly, alternative query transformations have been proposed to avoid distortion errors by partitioning the data based on the norm distribution.
[0063] The motivation behind the embodiments in this paper is also to observe the long-tail distribution in the data norm, but the embodiments in this paper address this issue through different graph-based indexing methods.
[0064] Non-reducible MIPS methods have also been proposed and evaluated in the MIPS benchmark. A tree-based approach for MIPS has been proposed. This method does not construct a dedicated tree for MIPS. Instead, a new bound is introduced for the inner product during the search phase. This bound is likely key to the tree-based approach, determining the conditions under which subtrees can be ignored to speed up the search process. Recently, randomized partitioning trees have also been used to solve MIPS problems. A MIPS method based on product quantization (PQ) has been proposed, extending the application of PQ techniques. An upper bound of the inner product has been used as an approximation of MIPS, and a greedy search algorithm has been designed to find this approximation, called Greedy MIPS. Similarly, an improved inner product bound under random projections has been provided, which matches the optimal bound of the Euclidean distance.
[0065] Graph-based indexing methods have demonstrated superior efficiency and effectiveness in ANN search using cosine similarity or L2 distance. Therefore, the idea of searching via neighbor graphs has been extended to other similarity metrics, such as inner products. A theoretical basis has been provided for using neighbor graphs in MIPS: if the constructed graph contains a Delaunay graph with respect to inner products, then a greedy search on the graph yields the exact true answer to MIPS. In its implementation, ip-NSW uses an algorithm to construct a hierarchical navigable small-world (HNSW) graph for indexing. To adapt to the MIPS problem, ip-NSW replaces the metric in HNSW with the inner product. To construct a better Delaunay graph with respect to inner products, a novel edge selection algorithm, the Inner Product Delaunay Graph (IPDG), has been proposed. IPDG performs surprisingly well in top-1 recall, but it is not suitable for top-k scenarios (e.g., k=10 or k=100).
[0066] Nearest neighbor graph-based search
[0067] This section explains why HNSW and ip-NSW have problems solving MIPS.
[0068] 1. Search for approximate Delaunay graphs
[0069] HNSW has two main components: (1) proximity graph construction (method 1) and (2) greedy search on the graph (method 2).
[0070]
[0071] Method 2: Greedy search on the graph SoG(q, G, k, f)
[0072] 1. Input: query element q, similarity graph G = (V, E), priority queue size k, and similarity metric f.
[0073] 2. Randomly select a vertex p∈V as the starting point and initialize the set A←{p}.
[0074] 3. Set p to selected and the other vertices to unselected.
[0075] 4. In the case that A does not converge
[0076] 5. Add the unselected neighboring vertices in A to A.
[0077] 6. Select the vertices in A.
[0078] 7. Following the order f(v, q), select the first k elements of A ← v∈A.
[0079] 8. Output: A.
[0080] The graph of HNSW is constructed in a greedy manner. As shown in Method 1, for each newly added data point x, it first searches for the neighbor candidate set {p} on the current graph using a greedy search algorithm. i Then, x is inserted into graph G by connecting to a carefully selected set of M neighbors. It should be noted that in one or more embodiments, graph G may be initialized as an empty graph. The first data point is then inserted as an isolation node. When inserting the second node, the first edge is constructed, and so on. After the graph is constructed, given a query point, method 2 performs a greedy search on the graph and returns the item that maximizes the similarity metric f.
[0081] In one or more embodiments, to find the exact nearest neighbor via greedy search, the graph contains a Delaunay graph as a subgraph. This implementation can be extended to the MIPS problem. Although Delaunay graphs have shown potential in similarity search, directly constructing Delaunay graphs in large-scale high-dimensional datasets is often infeasible due to the exponential growth in the number of edges in high-dimensional data. To address this issue, practical methods can be used to approximate Delaunay graphs, as shown in Method 1 (HNSW / ip-NSW).
[0082] 2. Select while
[0083] To achieve an efficient and effective approximate Delaunay graph, the key to graph construction is neighbor / edge selection (as shown in lines 5-16 of Method 1), ensuring that the finite selection of neighbors is representative (e.g., the directions of the metric cases are different). This largely facilitates the search trade-off between efficiency and effectiveness. For example, Figure 2A The diagram illustrates the similarity metric. Here, the similarity metric is negative Euclidean distance. Assume the current insertion point is x, and the returned candidate neighbors are {a, b, c}. HNSW first adds x's nearest neighbor a to c and connects x to a. Then, for the other vertices in the neighbor set, when f(x, c) > f(a, c) and f(x, b) < f(a, b), HNSW chooses c instead of b for connection, where b is more similar to a, and c is more different from a in direction. This greedy search navigation, especially in the initial steps, can be more efficient.
[0084] As the successor to HNSW, ip-NSW shares Method 1 and Method 2. The difference lies in the similarity metric. HNSW is designed for metrics such as cosine similarity or negative Euclidean distance, while ip-NSW sets... As mentioned above, the inner product poses a challenge to the edge selection method of Method 1 because it does not follow the triangle inequality of the metric space. Figure 2B As shown in the example, a small-norm data point x retrieves a set of nodes with a larger inner product value, such as {a, b, c}. As mentioned earlier, {a, b, c} typically consists of nodes with a larger norm that are in the same Voronoi unit as x. Figure 2B In the diagram, the gray area represents the corresponding Voronoi unit. The l2 norm of {a, b, c} is much larger than the l2 norm of x. For this case, it may be necessary to re-examine the edge selection method in Method 1. Since the inner product between x and {a, b, c} is much smaller than the inner product between {a, b, c}, when as well as x is not connected to b or c. Therefore, x is only connected to the pair with the largest inner product x. TNode a of node a. Ultimately, most data points have no (incoming) edges and do not function in query searches. Therefore, the edge selection method in Method 1 leads to inefficiency of the index graph and significantly limits retrieval performance, especially for top-k results where k >> 1, e.g., 100.
[0085] The above analysis of the norm distribution of the experimental dataset reveals that... Figure 2B The situation shown is very common. The analysis results of the two datasets are as follows: Figure 1 As shown. Due to the distribution of the long-tailed l2 norm, only a few data points have a large l2 norm, while most other data points have a relatively small norm. In the next section, an example of a norm-adjusted edge selection method more suitable for constructing graph indexes based on inner products is given.
[0086] Norm Adjusted Proximity Graph Example
[0087] This section describes an embodiment of the graph-based MIPS method, which, for convenience, can be referred to as the Norm Adjusted Proximity Graph (NAPG) embodiment. Return to References Figure 2B In the example shown, the norm of x is much smaller than that of its neighboring candidates a, b, and c, which makes Always less than the value in line 9 of method 1 This leads to inefficient index graphs and impacts search performance. In practical applications that consider the top 10 or top 100 MIPS results, this impact can become even more pronounced.
[0088] To address this limitation, an embodiment of a novel method for constructing appropriate graphs for MIPS is developed and described herein. The basic idea is to re-select edges, choosing representative neighbors for nodes with smaller norms (the vast majority). Specifically, in one or more embodiments, the norm of the inserted node x can be adjusted by an adjustment factor α, and the inequality in line 9 of Method 1 can be replaced with:
[0089]
[0090] In one or more embodiments, the adjustment factor can be set to α > 1, and it can be made and The scales are comparable. Thus, not only do the nodes have the maximum inner product with x (i.e., ...), Figure 2B In addition to the above, we can also select other candidate nodes as output neighbors of x. The remaining key issue is how to choose an appropriate value for the adjustment factor α. In the next section, an example of a flexible method for estimating α based on sampled data is given.
[0091] 1. Example of Adjustment Factor Estimation
[0092] The factor α serves to adjust both sides of inequality (4) to a comparable scale. In one or more embodiments, for each indexed data point x and its first n MIPS results {p i} (corresponding to adjacent candidate sets), we hope that the adjusted x and {p} i The expected value of the inner product of} and the data point {p i The expected values of the inner products between} are equal:
[0093]
[0094] Since α is a constant, formula (5) can be rewritten as:
[0095]
[0096] The expected value can be obtained by taking the average of the samples. Know An estimation is performed. Specifically, in one or more embodiments, methods such as... Figure 3 The steps shown are used to estimate the global adjustment factor α:
[0097] (1) Sample a (305) small subset {x} from the entire indexed dataset;
[0098] (2) For each sampled data point x, retrieve (310) its first n (e.g., n = 100) MIPS results {p} from the entire dataset. i};
[0099] (3) Calculate all x and their first n MIPS results {p} of (315) i The average inner product between}
[0100] (4) Calculate {p} for all x in (320). i The average inner product of} as well as
[0101] (5) Through Divide by Estimate the global adjustment factor α (325), for example
[0102] After estimating the global adjustment factor α, the inequality in line 9 of Method 1 can be replaced with inequality (4) to construct a better approximation of the Delaunay plot.
[0103] 2. Example of adjustment factor based on norm range
[0104] It should be noted that the global adjustment factor α may not be optimal because the inserted data points have different norms. A finer-grained approach is to estimate the local adjustment factor for each norm range of data points. This norm range concept can be considered similar to the idea of simplifying the MIPS problem to a traditional ANN search problem by preprocessing the data within each norm range. However, this paper explores the norm range in a different way for local adjustment factors based on norm ranges.
[0105] Figure 4 The process of estimating a local adjustment factor according to an embodiment of the present disclosure is described. To obtain the local adjustment factor, in one or more embodiments, the data points are first partitioned according to the l2 norm of the data points. Given a vector set, a global norm range of the vector set based on the l2 norm values is obtained (405). The global norm range (from minimum to maximum) is then divided (410) into multiple (N) subranges based on the l2 norm values. Each subrange can be considered a stratum. Stratified sampling can be used to extract (415) samples from the dataset such that each subrange receives the same number of samples. For each subrange, using the samples in each subrange, a local adjustment factor α for the subrange is obtained (420) according to formula (6). r (As shown in lines 7-9 of Method 3). In graph construction, in one or more embodiments, the corresponding α of the current insertion point can be located based on the norm value of the current insertion point. r Then, α r It can be used as a factor in edge selection.
[0106] 3. Implementation Method Examples
[0107] At the algorithmic level, in one or more embodiments, NAPG can share the greedy search method shown in Method 2, but utilize a unique graph construction method (i.e., Method 4). Before constructing NAPG, adjustment factors for each norm range can be estimated. An embodiment of the norm range-based adjustment factor estimation method is shown in Method 3.
[0108] Based on the estimated factors, NAPG construction can be implemented in one or more embodiments as shown in Method 4. It should be noted that one of the main differences between HNSW / ip-NSW and NAPG lies in the edge selection step. Using a norm-range-based adjustment factor, NAPG can construct a more efficient and effective index graph for MIPS compared to ip-NSW. As the experimental results in Section E show, the tested NAPG embodiments perform better than ip-NSW thanks to an appropriate index graph.
[0109] Method 3: Example of Adjustment Factor Estimation Based on Norm Range
[0110]
[0111] Method 4: NAPG Construction Example
[0112]
[0113]
[0114] Figure 5 A NAPG construction process according to an embodiment of this disclosure is described. Given a set of potential vectors, a norm range graph R is obtained (505) using the process shown in Method 3, which includes multiple subranges and an estimated adjustment factor for each subrange. An index graph G is constructed (510) using norm-adjusted inner products containing the estimated adjustment factors for the multiple subranges. In one or more embodiments, the index graph G may be generated involving at least one process shown in Method 4. Once the index graph G is generated, a greedy search (515) of a given query vector is performed on the constructed index graph G to retrieve one or more relevant vectors. In one or more embodiments, the retrieved one or more relevant vectors are sorted according to the inner product between each relevant vector and the query vector.
[0115] In one or more embodiments, the following can be used: Figure 6 The steps shown generate the index graph G. As shown in line 2 of Method 4, the index graph (605) is initialized with an empty graph. For each latent vector x in the latent vector set S, a greedy search (610) is performed to obtain the k vertices {p} on the index graph G. i}, where k is an integer greater than 1, and these vertices have the first k largest inner product values with x. For example, the selected vertex set C of the latent vector x is initialized using an empty set (615). As shown in line 5 of Method 4, the adjustment factor α of the estimated (620) is obtained based on the subrange to which the latent vector x belongs. r The selected vertex set C is updated (625) by adding one or more selected vertices to the k vertices, wherein the norm-adjusted inner product between each added vertex and the latent vector x is not less than the inner product between vertices in the selected vertex set C. The index graph G is updated (630) by adding one or more edges connecting the latent vector x and one or more selected vertices. Once the number of vertices in the selected vertex set C reaches a predetermined threshold (M), updating the index graph G at the latent vector x is stopped, and the NAPG process iterates to another latent vector in the latent vector set S and continues.
[0116] experiment
[0117] It should be noted that these experiments and results are provided in an illustrative manner and are performed under specific conditions using one or more specific embodiments; therefore, these experiments and their results should not be used to limit the scope of disclosure of the present publication.
[0118] 1. Dataset
[0119] Implementations of the NAPG method are evaluated on MIPS vector datasets from various sources. First, different methods for fast item retrieval in online recommender systems are evaluated. Two recommender datasets, Dataset 1 and Dataset 2, are used. Dataset 1 is unfiltered, while Dataset 2 is filtered to ensure users maintain at least 30 interactions. These two datasets contain significantly more items than other recommender datasets, making them more suitable for studying item retrieval efficiency.
[0120] As shown in Table 1, four different vector datasets were generated from Dataset 1 and Dataset 2, named Dataset 1 Matrix Factorization (MF), Dataset 1 Deep Matrix Factorization (DMF), Dataset 2 MF, and Dataset 2 DMF, respectively. The following will explain how these four datasets in Table 1 were generated based on MF and DMF.
[0121] Collaborative filtering (DMF) is the most representative collaborative filtering framework, designed to generate vectors for users and items. Furthermore, DMF is used to generate alternative vector datasets. It's important to note that the matching function here is set to inner product instead of cosine similarity, and the loss function is set to squared loss. in Y is the predicted rating score, while Y is the actual score. + It is positive training data, Y - This is sampled negative data. The network has 2 layers.
[0122] For these four datasets, the length of all vectors was set to 64. It should be noted that the vectors generated by MF and DMF are quite different. MF-generated vectors are dense, containing both positive and negative real numbers. Vectors from DMF, on the other hand, are sparse and, due to the ReLU activity function, contain only non-negative numbers.
[0123] In addition to these four datasets, Table 1 includes two larger vector datasets from other domains: Dataset 3 and Dataset 4. Unlike the recommendation datasets, which require a method to generate latent vectors for users and items, the raw vectors from these two datasets are used. 1000 additional vectors are randomly selected as queries (corresponding to these user vectors). The statistics for all datasets are listed below in Table 1. It can be seen that the experimental datasets differ in dimensionality, source, and extraction methods, which is sufficient for a fair comparison.
[0124] Table 1: Dataset Statistics
[0125]
[0126] To generate evaluation labels, the items / base vectors of each user / query vector are sorted by a brute-force scan (in the inner product). The top k items are considered as labels (k = 10 and 100).
[0127] 2. Baseline
[0128] As a baseline, some representative, state-of-the-art MIP methods were selected for comparative evaluation with the embodiments of this disclosure.
[0129] Greedy MIPs are the most advanced MIP algorithms.
[0130] ip-NSW is also a MIPS method. Similar to the tested embodiment, ip-NSW is also a graph-based method.
[0131] Range LSH is a hash-based method. For MIPs, Range LSH outperforms previous hash-based methods. Therefore, it was compared as a representative of this approach. It's important to note that H2-ALSH (an asymmetric locality-sensitive hashing scheme based on concentric hypersphere partitioning) also uses a data partitioning method (and query transformation). Initial H2-ALSH experimental results showed that H2-ALSH performs well primarily because it uses a partitioning method rather than a query transformation. Therefore, Range LSH and H2-ALSH are similar methods. Range LSH was chosen as the baseline because its implementation is publicly accessible.
[0132] 3. Experimental setup
[0133] All comparison methods have adjustable parameters. The tested NAPG and ip-NSW implementations share three common parameters: M, k 构建 and k 搜索 These parameters control the degree of each node and the number of search attempts. The tested NAPG embodiment also has a unique parameter for the norm range number N. Greedy MIPS has a key parameter budget, and the range LSH has two parameters: code length and the number of subsets. To obtain a fair comparison, all parameters of all methods were changed on a fine grid.
[0134] There are two popular approaches to evaluating ANN / MIPS methods: (a) recall and time; (b) recall and computation. Both assess the trade-off between search efficiency and search effectiveness. Recall and time reports the number of queries per second that the method can process at each recall level. Recall and computation reports the number or percentage of pairwise distance / similarity computations consumed by the search method at each recall level. For a comprehensive evaluation, both perspectives are illustrated in the following experiments.
[0135] 4. Results of the recall and its timing
[0136] The comparison results obtained through recall and time are first in Figures 7A to 7F The table shows the results. All methods can be evaluated from this perspective. Each row represents each of the datasets 1MF, 1DMF, 2MF, 2DMF, 3, and 4. The two columns show the results for the top-10 and top-100 labels. Figures 7A to 7F As shown, the tested NAPG implementation performs best in most cases, especially at higher recall levels, which is more meaningful in practical applications. Furthermore, the NAPG implementation consistently performs well on labels of varying sizes. Different datasets have extremely different vectors, and the NAPG implementation excels on all these types of vector data. These comparisons demonstrate the efficiency and robustness of the NAPG implementation.
[0137] Conversely, the baseline does not consistently perform well on labels of varying sizes. For example, ip-NSW and greedy MIPS deteriorate from top-10 to top-100. Compared to ip-NSW, the NAPG implementation improves the edge selection method by flexibly adjusting the norm of data points. Clearly, on these datasets, the tested NAPG implementation delivers significantly better performance than ip-NSW, especially for top-100 labels. ip-NSW performs poorly on top-100 labels because it filters out too many meaningful neighbors in edge selection, severely impacting performance. The hash-based method, LSH, performs poorly in this regard compared to other methods. It can be seen that its performance cannot compare with other products. Aside from the recommended dataset, it should be noted that the NAPG implementation consistently performs well on two larger datasets (dataset 3 and dataset 4). ip-NSW performs poorly close to the baseline. Particularly for the top-100 dataset in dataset 3, ip-NSW only achieves around 40% recall.
[0138] 5. Results of recall and calculation
[0139] Figure 8 The corresponding comparison results obtained through recall and computation are shown. From this perspective, greedy MIPS cannot be evaluated. Therefore, the results of three other methods are presented. The results for the top-100 labels are only available in... Figure 8 The report also noted that the NAPG implementation (in most cases) delivered better results compared to the baseline. Quantitatively, to achieve high recall, the tested NAPG implementation required small partitions of pairwise computations. For example, for the top-100 labels of the 2MF dataset, the NAPG implementation requested only 1% of the computations to achieve 95% recall. The baseline, on the other hand, struggled to even achieve 95% recall.
[0140] Surprisingly, Range LSH performs comparably to other methods in terms of recall and computation. As described in Section E.3, recall and computation do not take into account the cost of different index structures. Although Range LSH performs well in this respect, its overall time cost is significantly higher than the other methods shown in Section E.4. One possible reason is that the table-based indexes used in Range LSH are inefficient in search. Clearly, graph-based index structures have a significant advantage in search efficiency.
[0141] 6. Parameter Analysis
[0142] This section analyzes the key parameter N, the number of norm ranges, in the NAPG implementation. In one or more implementations, N is a data-dependent parameter that depends on the size of the dataset and the norm distribution. In the experiments described above, N was treated as an adjustable parameter along with other parameters. Here, we will analyze N, M, and k... 构建 The values are fixed at 16 and 100 respectively. The results for dataset 1MF(top-10) are as follows: Figure 9 As shown in Figure 905, the optimal N value for this dataset is approximately 3 or 5, which is far superior to a single norm range. Splitting the dataset into more ranges than this optimal approach, such as 10, can also impact performance. This is because more ranges may result in some ranges having very small norm data, potentially leading to very large adjustment factors. Excessively large ranges can over-adjust the data and build too many unnecessary edges.
[0143] 7. Estimation Factor Research
[0144] Adjusting the factor α is important for the tested NAPG implementation. A flexible factor estimation method is designed based on the sample data (Section D). Alternatively, α can be set as an adjustable parameter (either for all indexed data or for each norm range). In this section, the factor estimation method is empirically evaluated. The results are as follows... Figure 9 As shown in Figure 910, only the top-10 results for the MF dataset are shown. In the example shown, the cases of NAPG with the norm range number N=1 and α adjusted to 0.95, 1, 10, and 100 were also investigated. The estimated α value for NAPG is 4.33. It should be noted that α=1 is the scheme for ip-NSW. If α is very large, such as 100, the inequality in line 9 of Method 1 may always be false, and edge selection may become invalid. It can be seen that the NAPG implementation learns a good estimate of α and achieves the best performance among almost all selections. In one or more embodiments, a smaller adjustable parameter will simplify the application without sacrificing performance. The advantage may be even more pronounced for multiple norm ranges N>1.
[0145] Some conclusions
[0146] Fast inner product search is an important task in online recommender systems and retrieval-based question answering services. In the literature, this task is often referred to as the Approximate Maximum Inner Product Search (MIPs) problem. While solving MIPS accurately is expensive and even infeasible, developing approximate MIPS algorithms is also challenging in industrial practice because the inner product is a non-metric measure, which distinguishes the MIPS problem from traditional ANN search problems (which typically deal with metrics such as cosine similarity). This disclosure describes an embodiment of a novel graph-based search method, termed Norm Adjusted Neighborhood Graph (NAPG), for approximating MIPS. The NAPG embodiment adjusts the norm of data points when inserting them into the graph while constructing an appropriate index graph for the inner product. Flexible adjustment factor estimation methods and improved embodiments based on norm ranges are also introduced, making the embodiment more flexible and robust. Empirical evaluations on a range of datasets demonstrate that the tested NAPG embodiment is indeed effective and efficient compared to common baselines.
[0147] Computing System Implementation Examples
[0148] In one or more embodiments, aspects of this disclosure may relate to, include, or be implemented on one or more information processing systems (or computing systems). An information processing system / computing system may include any means or combination of means operable to calculate, compute, determine, classify, process, transmit, receive, retrieve, initiate, route, exchange, store, display, communicate, exhibit, detect, record, reproduce, dispose of, or utilize any form of information, intelligence, or data. For example, a computing system may be or may include a personal computer (e.g., a laptop computer), a tablet computer, a mobile device (e.g., a personal digital assistant (PDA), a smartphone, a phablet, a tablet, etc.), a smartwatch, a server (e.g., a blade server or a rack server), a network storage device, a camera, or any other suitable device, and may vary in size, shape, performance, functionality, and price. A computing system may include random access memory (RAM), one or more processing resources (such as a central processing unit (CPU) or hardware or software control logic), read-only memory (ROM), and / or other types of memory. Other components of the computing system may include one or more drives (e.g., hard disk drives, solid-state drives, or both), one or more network ports for communicating with external devices, and various input and output (I / O) devices (e.g., keyboard, mouse, touchscreen, stylus, microphone, camera, touchpad, display, etc.). The computing system may also include one or more buses operable for transmitting communication between various hardware components.
[0149] Figure 10A simplified block diagram of an information processing system (or computing system) according to embodiments of the present disclosure is depicted. Although it should be understood that the computing system may be configured differently and include different components, including, for example... Figure 10 The system shown may contain fewer or more components, but it should be understood that the functionality shown for system 1000 is operable to support various implementations of the computing system.
[0150] like Figure 10 As shown, the computing system 1000 includes one or more central processing units (CPUs) 1001, which provide computing resources and control the computer. The CPU 1001 may be implemented as a microprocessor, etc., and may also include one or more graphics processing units (GPUs) 1002 and / or floating-point coprocessors for mathematical calculations. In one or more embodiments, the one or more GPUs 1002 may be integrated into a display controller 1009, for example, as part of one or more graphics cards. The system 1000 may also include system memory 1019, which may include RAM, ROM, or both.
[0151] like Figure 10As shown, multiple controllers and peripheral devices may also be provided. Input controller 1003 represents an interface to various input devices 1004. The computing system 1000 may also include a storage controller 1007 for interfacing with one or more storage devices 1008, each of which includes a storage medium (such as magnetic tape or disk) or optical medium (which may be used to record programs for instructions for operating systems, utilities, and applications, and may include embodiments of programs implementing aspects of this disclosure). Storage devices 1008 may also be used to store processed data or data to be processed according to this disclosure. The system 1000 may also include a display controller 1009 for providing an interface to a display device 1011, which may be a cathode ray tube (CRT) display, a thin-film transistor (TFT) display, an organic light-emitting diode (OLED) display, an electroluminescent panel, a plasma panel, or any other type of display. The computing system 1000 may also include one or more peripheral device controllers or interfaces 1005 for one or more peripheral devices 1006. Examples of peripheral devices may include one or more printers, scanners, input devices, output devices, sensors, etc. The communication controller 1014 may interface with one or more communication devices 1015, enabling the system 1000 to connect to remote devices via any of various networks, including the Internet, cloud resources (such as Ethernet cloud, Fibre Channel over Ethernet (FCoE) / Data Center Bridge (DCB) cloud, etc.), local area networks (LAN), wide area networks (WAN), storage area networks (SAN)), or via any suitable electromagnetic carrier signal (including infrared signals). As depicted in the embodiments, the computing system 1000 includes one or more fans or fan trays 1018 and one or more cooling subsystem controllers 1017 that monitor the thermal temperature of the system 1000 (or its components) and operate the fans / fan trays 1018 to help regulate the temperature.
[0152] In the illustrated system, all major system components can be connected to bus 1016, which may represent more than one physical bus. However, the various system components may or may not be physically close to each other. For example, input data and / or output data may be remotely transferred from one physical location to another. Additionally, programs implementing aspects of this disclosure can be accessed from a remote location (e.g., a server) via a network. Such data and / or programs can be transmitted via any of a variety of machine-readable media, including, for example: magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as optical discs (CDs) and holographic devices; magneto-optical media; and hardware devices specifically configured to store or store and execute program code, such as application-specific integrated circuits (ASICs), programmable logic devices (PLDs), flash memory devices, other non-volatile memory (NVM) devices (e.g., 3D XPoint-based devices), and ROM and RAM devices.
[0153] Various aspects of this disclosure can be encoded on one or more non-transitory computer-readable media having instructions for one or more processors or processing units to perform steps. It should be noted that the one or more non-transitory computer-readable media should include both volatile and non-volatile memory. It should be noted that alternative implementations are possible, including hardware implementations or software / hardware implementations. Hardware implementations can be implemented using ASICs, programmable arrays, digital signal processing circuits, etc. Therefore, the term "means" in any claim is intended to cover both software and hardware implementations. Similarly, the term "computer-readable medium or media" as used herein includes software and / or hardware having instructions implemented thereon, or a combination thereof. Utilizing these contemplated alternative implementations, it should be understood that the accompanying drawings and description provide information required by those skilled in the art to write program code (i.e., software) and / or manufacture circuitry (i.e., hardware) to perform the desired processing.
[0154] It should be noted that embodiments of this disclosure may also relate to computer products having a non-transitory tangible computer-readable medium storing computer code for performing operations of various computer implementations. The medium and computer code may be specifically designed and constructed for the purposes of this disclosure, or may be known or available to those skilled in the art. Examples of tangible computer-readable media include, for example: magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CDs and holographic devices; magneto-optical media; and hardware devices specifically configured to store or store and execute program code, such as ASICs, PLDs, flash memory devices, other non-volatile memory devices (e.g., 3DXPoint-based devices), and ROM and RAM devices. Examples of computer code include machine code (e.g., compiler-generated code) and files containing higher-level code that can be executed by a computer using an interpreter. Embodiments of this disclosure may be implemented, in whole or in part, as machine-executable instructions in a program module that can be executed by a processing device. Examples of program modules include libraries, programs, routines, objects, components, and data structures. In a distributed computing environment, program modules may be physically located in a local, remote, or both setting.
[0155] Those skilled in the art will recognize that the computing system or programming language is not important to the practice of this disclosure. They will also recognize that the various elements described above may be physically and / or functionally divided into modules and / or sub-modules or combined together.
[0156] Those skilled in the art will understand that the examples and embodiments described above are exemplary and do not limit the scope of this disclosure. It is intended to illustrate that all substitutions, enhancements, equivalents, combinations, or modifications of this disclosure that will be apparent to those skilled in the art upon reading this specification and studying the accompanying drawings are included within the true spirit and scope of this disclosure. It should also be noted that the elements of any claim may be arranged differently, including having multiple dependencies, configurations, and combinations.
Claims
1. A computer-implemented question-answering method, comprising: Given a set of latent vectors containing multiple latent vectors, a norm range graph is obtained, which includes multiple subranges and an estimated adjustment factor for each subrange, wherein the latent vectors are used to represent documents; An index graph is constructed using norm-adjusted inner products, which include the estimated adjustment factors for the plurality of subranges; Given a query vector, a greedy search is performed on the constructed index graph to retrieve one or more relevant vectors, where the query vector represents the query question; as well as Based on one or more documents represented by the one or more relevance vectors, candidate answers are returned, wherein the one or more documents include the candidate answers. The estimated adjustment factor for each subrange is obtained through the following steps: For each sample in each subrange, retrieve the top n vectors from the latent vector set that have the largest n inner products with each sample, where n is an integer greater than 1; Calculate the average of the inner product values between all samples in each subrange and the top n vectors retrieved for that subrange; Calculate the average of the inner product values among the first n vectors retrieved for all samples in each subrange; as well as The estimated adjustment factor for each subrange is obtained by dividing the average of the inner product values of the first n retrieved vectors by the average of the inner product values of all samples in each subrange with the first n retrieved vectors for that subrange.
2. The computer-implemented method according to claim 1, wherein, The multiple sub-ranges are obtained through the following steps: The global norm range is obtained based on the l2 norm value of the potential vector set; The global norm range is divided into the multiple sub-ranges on an average basis; as well as Stratified sampling is used to extract samples from the latent vector set, such that each subrange receives the same number of samples.
3. The computer-implemented method according to claim 1, wherein, Building an index graph using norm-adjusted inner products involves the following steps: Initialize the index graph; and For each latent vector in the set of latent vectors: A greedy search is performed to obtain the k vertices on the index graph that have the top k largest inner product values with each potential vector, where k is an integer greater than 1; Initialize the selected vertex set for each potential vector; The estimated adjustment factor is obtained based on the subrange to which each potential vector belongs; The selected vertex set is updated by adding one or more selected vertices to the k vertices, wherein the norm-adjusted inner product between each added vertex and each latent vector is not less than the inner product between vertices in the selected vertex set; and The index graph is updated by adding one or more edges until the number of vertices in the selected vertex set reaches a predetermined threshold, wherein the one or more edges connect each latent vector to the one or more selected vertices.
4. The computer-implemented method according to claim 3, wherein, The index graph is initialized using an empty graph.
5. The computer-implemented method according to claim 3, wherein, The norm-adjusted inner product between each added vertex and each latent vector is the product of the inner product between each added vertex and each latent vector and the estimated adjustment factor.
6. The computer-implemented method according to claim 1, wherein, The retrieved one or more relevant vectors are sorted according to the inner product between each relevant vector and the query vector.
7. An electronic device, comprising: One or more processors; as well as A non-transitory computer-readable medium comprising one or more sets of instructions that, when executed by at least one of the one or more processors, perform the method according to any one of claims 1 to 6.
8. A non-transitory computer-readable medium comprising one or more instruction sequences that, when executed by at least one processor, perform the method of any one of claims 1 to 6.
9. A computer program product comprising a computer program stored on at least one of a readable storage medium and an electronic device, the computer program implementing the method according to any one of claims 1 to 6 when executed by a processor.
Citation Information
Patent Citations
Efficient retrieval of top similarity representations
CN112685603A
Transformation for fast inner product search on graph
CN112784115A
