Music search method, display method, device, storage medium, equipment and product

Abstract

This application discloses a music search method, display method, apparatus, storage medium, device, and product, including obtaining a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the user's input search terms, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. Based on an attention mechanism, the weights of the user's historical interaction feature vectors are determined using intent feature vectors, and a user's historical feature vector is generated based on the weights and historical interaction feature vectors. The intent feature vector is obtained by fusing the query feature vector and candidate song feature vectors, and the historical interaction feature vector is generated based on M historical interaction records of the user, where M is a positive integer. The query feature vector, candidate song feature vectors, and historical feature vectors are fused to generate a fused feature vector. Based on the fused feature vector, music search results are determined to improve the user experience.

Description

Technical Field

[0001] This application relates to the field of music technology, specifically to a music search method, a method for displaying music search results, a music search device, a device for displaying music search results, a storage medium, an apparatus, and a program product. Background Technology

[0002] In related technologies, with the development of internet music services, users' behavior on platforms has accumulated massive amounts of data. In music search scenarios, how to utilize users' historical behavioral data to optimize search result ranking and provide users with a personalized search experience is a problem that urgently needs to be solved in this field. Summary of the Invention

[0003] This application provides a music search method, display method, apparatus, storage medium, device, and product, which can solve at least one of the above-mentioned technical problems.

[0004] On one hand, embodiments of this application provide a music search method, the method comprising: Obtain a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the search terms input by the user, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. Based on the attention mechanism, the weights of the user's historical interaction feature vectors are determined using the intent feature vector, and the user's historical feature vector is generated based on the weights of the historical interaction feature vectors and the historical interaction feature vectors. The intent feature vector is obtained by feature vector fusion based on the query feature vector and the candidate song feature vector, and the historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer. The query feature vector, the candidate song feature vector, and the historical feature vector are fused to generate a fused feature vector; Based on the fused feature vector, music search results are determined.

[0005] On the other hand, embodiments of this application provide a method for displaying music search results, including: The music search results are obtained in response to the music search method described in any of the above embodiments; Based on the music search results, a list of music corresponding to the music search results is displayed.

[0006] On the other hand, embodiments of this application provide a music search device, the device comprising: The acquisition module is used to acquire a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the search terms input by the user, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. The generation module is used to determine the weights of the user's historical interaction feature vectors based on the attention mechanism and the intent feature vectors, and to generate the user's historical feature vectors based on the weights of the historical interaction feature vectors and the historical interaction feature vectors. The intent feature vectors are obtained by feature vector fusion based on the query feature vectors and the candidate song feature vectors, and the historical interaction feature vectors are generated based on M historical interaction records of the user, where M is a positive integer. The fusion module is used to fuse the query feature vector, the candidate song feature vector, and the historical feature vector to generate a fused feature vector; The determination module is used to determine the music search results based on the fused feature vector.

[0007] On the other hand, embodiments of this application provide a device for displaying music search results, including: A response module is configured to obtain the music search results in response to the music search method described in any of the above embodiments. The display module is used to display a list of music corresponding to the music search results.

[0008] On the other hand, embodiments of this application provide a computer-readable storage medium storing a computer program adapted for loading by a processor to execute the music search method or music search result display method as described in any of the above embodiments.

[0009] On the other hand, embodiments of this application provide a computer device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the music search method or the music search result display method as described in any of the above embodiments by calling the computer program stored in the memory.

[0010] On the other hand, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the music search method or the music search result display method as described in any of the above embodiments.

[0011] The music search method provided in this application obtains a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the user's input search terms, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. Based on an attention mechanism, the weights of the user's historical interaction feature vectors are determined using the intent feature vector. Based on the weights and historical interaction feature vectors, the user's historical feature vector is generated. The intent feature vector is obtained by fusing the query feature vector and the candidate song feature vectors, and the historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer. This avoids discarding important information in the initial screening stage and reduces the interference of irrelevant historical interaction features on subsequent calculations, thus improving the accuracy of feature representation. The query feature vector, candidate song feature vector, and historical feature vector are fused to generate a fused feature vector, achieving full information fusion of search intent, candidate content, and user's personalized historical interests, providing a complete basis for music search. Finally, based on the fused feature vector, the music search results are determined, maximizing the retention of full-dimensional information of search intent, candidate content, and historical interests, thereby improving the user's music search experience. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an example music system provided in an embodiment of this application.

[0014] Figure 2 This is a flowchart illustrating the music search method provided in an embodiment of this application.

[0015] Figure 3 This is a schematic diagram illustrating a scenario for the music search method provided in an embodiment of this application.

[0016] Figure 4 This is a schematic diagram of the structure of the music search device provided in an embodiment of this application.

[0017] Figure 5 This is a schematic diagram of the structure of the music search results display device provided in the embodiments of this application.

[0018] Figure 6 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The background technology of this application will be further explained below.

[0021] For example, in one music ranking scheme, the ranking process is divided into two steps: the first step uses simple rules or a lightweight model to filter out the Top-K behaviors most relevant to the current candidate song from the user's entire history; the second step then inputs these Top-K behaviors into a fine-tuning model. However, the hard retrieval in the first step is usually based on categories or simple vector similarity, which may incorrectly discard information valuable for the final prediction in the initial screening stage. In addition, this two-stage fragmented training method cannot directly optimize the entire process from historical sequence selection to final click-through rate prediction through gradient backpropagation.

[0022] For another example, in a solution that utilizes the STCA (Stacked Target-History Cross-Attention) structure to directly calculate the cross-attention between the "TargetItem" and the "User's Long History," and solves the computational efficiency problem through approximation algorithms and system optimization, the input in the recommendation scenario is only User and TargetItem. However, in the search scenario, the user's input contains explicit search queries. The existing STCA structure only calculates Attention(Target, History). If this structure is directly applied to search, the model will overemphasize the similarity between "historical interests" and "candidate songs," ignoring the constraints of the "search terms." For example, a user's history might include "heavy metal rock," but the current search term is "sleep-inducing light music." If existing technology is used, the model might assign a high score because "rock music" in the candidate songs highly matches the history, resulting in search results that are irrelevant to the user's search intent and severely violating the user's immediate search intent.

[0023] Furthermore, music search queries often contain complex intentions (such as searching for lyrics, genre, or emotional tone). Current technology lacks a mechanism to effectively fuse the semantic vector of the query with the content vector of the CandidateItem, and then search for corroborating evidence in long-sequence historical data. Simple concatenation, when processing 10k-level long sequences, dilutes the semantics of the query amidst massive historical noise, failing to accurately "activate" specific relevant songs in the history using the query.

[0024] Finally, in a 10k-length historical sequence, for a specific search term (such as "singer A"), the vast majority of historical behaviors (such as "English nursery rhymes" and "crosstalk") are pure noise. Existing technologies typically perform calculations on the entire sequence. Although approximation algorithms can accelerate this process, under the high concurrency (QPS) and low latency requirements of search scenarios, the calculation of irrelevant historical segments still results in a significant waste of computational power, and the introduced noise can interfere with the accuracy of the ranking.

[0025] To address at least one of the aforementioned technical problems, embodiments of this application provide a music search method, apparatus, storage medium, device, and program product. Specifically, the music search method of this application embodiment can be executed by a computer device, which can be a terminal or a server, etc. The terminal can be a smartphone, tablet, laptop, smart TV, wearable smart device, smart vehicle terminal, etc. The terminal can also include a client, which can be a music playback client, browser client, instant messaging client, or mini-program, etc. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0026] For example, when this music search method is used on a terminal device, the terminal device may include a display screen and a processor. The display screen is used to present the music search screen and receive instructions generated by the user interacting with the music search screen. The music screen may include a candidate list of songs, etc. The processor is used to store music applications, run the music applications, generate the music search screen, respond to instructions, and control the display of the music screen on the display screen. When the user operates the music search screen through the display screen, the music screen can control the local content of the terminal device in response to the received operation instructions. The terminal device can provide the user interface to the user in various ways, such as rendering the display on the terminal device's screen or presenting a graphical user interface through holographic projection.

[0027] For example, when this music search method runs on a server, it can be implemented and executed based on a cloud music system. A cloud music system refers to a music streaming method based on cloud computing. A cloud music system includes servers and client devices. The core of the music application and the core of the music display are separate; the storage and execution of the music search method are completed on the server. The music display is completed on the client, which is mainly used for receiving and sending user search terms and displaying the music search results. For example, the client can be a display device with data transmission capabilities located close to the user, such as a mobile terminal, television, computer, PDA, personal digital assistant, or head-mounted display device. However, the terminal device for processing music data is the server in the cloud. When playing music, the user operates the client to send instructions to the server. The server controls music playback according to the instructions, encodes and compresses the music display data, returns it to the client via the network, and finally, the client decodes and outputs the music search results.

[0028] It should be noted that, in this embodiment, the entity executing the music search method can be a terminal device or a server. The terminal device can be a local terminal device or a client device in the aforementioned cloud music platform. This embodiment does not limit the type of the entity executing the method.

[0029] For example, in conjunction with the above description, Figure 1 This application illustrates a music system 1000 for implementing a music search method, as provided in an embodiment of this application. The music system 1000 may include at least one terminal 1001, at least one server 1002, at least one database 1003, and a network. The user-held terminal 1001 can connect to different servers via the network. The terminal is any device with computing hardware capable of supporting and executing music-related software applications.

[0030] In the aforementioned music system 1000, terminal 1001 is used to install and run music applications. In some cases, terminal 1001 may not need to have the music application pre-installed; users can directly access music through a browser or other client. Users can log in to the music application using their registered music account to control music search and playback. When a user logs in, terminal 1001 sends a login request to server 1002. Server 1002 verifies the user's music account and, based on the login request, verifies the music account. If verification is successful, a login success notification is sent back to terminal 1001. During music playback through the application, terminal 1001 and server 1002 exchange data. Terminal 1001 sends the user-input search terms to server 1002. Server 1002 determines the music search data for terminal 1001 based on its stored music search mechanism and the received search term information, and sends the music search data back to terminal 1001, allowing terminal 1001 to display the music search data sent by server 1002 to the user.

[0031] In possible application scenarios, different terminals 1001 may be served by different servers 1002. Therefore, in order to distinguish the servers 1002 corresponding to different music terminals 1001, the embodiments of this application will use the terms "first" and "second" to describe them. In fact, the servers 1002 corresponding to different music terminals 1001 can be the same server 1002. Therefore, without distinguishing between "first" and "second", it can be understood that the terminals 1001 corresponding to virtual characters located in the same music scene are served by the same server 1002.

[0032] Furthermore, when the music system 1000 includes multiple terminals, multiple servers, and multiple networks, different terminals can connect to each other through different networks and servers. The network can be a wireless network or a wired network; for example, wireless networks include Wi-Fi, LAN, cellular networks, 2G, 3G, 4G, and 5G networks. Additionally, different terminals can also connect to other terminals or servers using their own Bluetooth networks or hotspot networks. Moreover, the system 100 can include multiple databases coupled to different servers, and can continuously store music-related information in the databases while different users are playing music online in a multi-user environment.

[0033] It should be noted that in this embodiment, multiple terminal devices run the same virtual music program. Therefore, data interaction between the multiple terminal devices can be achieved through the server of the virtual music program. Thus, sending data from terminal device 1 to terminal device 2 can be understood as: terminal device 1 sends data to the server of the virtual music program, and the server sends the data to terminal device 2. Receiving data from terminal device 2 can be understood as: terminal device 1 receives data sent by the server of the virtual music program, which is the data sent by terminal device 2 to the server. Alternatively, there may be no music server, and terminal device 1 directly sends music data to terminal device 2.

[0034] It should be noted that, Figure 1 The illustrated music system diagram is merely an example. The music system 1000 described in this application embodiment is intended to more clearly illustrate the technical solutions of this application embodiment and does not constitute a limitation on the technical solutions provided in this application embodiment. As those skilled in the art will know, with the evolution of music systems and the emergence of new business scenarios, the technical solutions provided in this application embodiment are also applicable to similar technical problems.

[0035] It should be noted that the triggering operations mentioned in the subsequent detailed description of the music search method provided in this application can all be regarded as triggering operations performed by the user through a finger or by controlling a medium such as a mouse, keyboard, or stylus. The specific medium used can be determined based on the type of computer device. For example, when the computer device is a touchscreen device such as a mobile phone, tablet, or music player, the user can operate the touchscreen using any suitable object or accessory such as a finger or stylus. When the terminal device is a non-touchscreen terminal device such as a desktop computer or laptop, the user can operate it using external devices such as a mouse or keyboard.

[0036] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0037] Please see Figures 2 to 5 , Figure 2 This is a flowchart illustrating the music search method provided in an embodiment of this application. Figures 3 to 5 This is a schematic diagram illustrating an application scenario of the music search method provided in this application embodiment. It should be noted that the steps shown may be executed in a logical order different from that shown in the flowchart. The method provides a display interface through a terminal device, the interface containing recommended music, and the method may include the following steps: Step 011: Obtain the query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the user's input search terms, and the candidate song feature vectors are obtained by filtering the preset music library based on the search terms.

[0038] The query feature vector can be the music search term (Query) input by the user, which is then transformed into a high-dimensional dense vector, Eq, by a pre-trained encoder. The query feature vector can express the semantic information of the search term. For example, taking the music search term "singer A" as an example, the query feature vector can be used to represent the semantic information of "singer A".

[0039] The preset music library can be a music resource library that includes a variety of music.

[0040] Among them, the candidate song feature vector can be the feature vector transformed from the candidate music resources (Candidate Item) obtained after filtering from the preset music library based on music search terms. It can be used to represent the candidate song's identifier, audio features, audio type, etc.

[0041] Optionally, the method further includes: Step 014: Based on the search terms and historical interaction feature vectors, filter the preset music library to obtain multiple candidate songs; Step 015: Extract the audio fingerprint features and song attribute features of the candidate songs, and concatenate the audio fingerprint features and song attribute features to obtain the song concatenation vector; Step 016: Map the concatenated song vectors to candidate song feature vectors using a multilayer perceptron.

[0042] The historical interaction feature vector can be a feature vector (or vector matrix, etc.) obtained by encoding the user behavior sequence generated by user interactions on music platforms, music clients, etc. Interaction behaviors can include listening to music, adding music to favorites (indicating that the user likes the music), sharing music, and ignoring music (indicating that the user dislikes the music). For example, users can interact by clicking the favorites control on the music display interface that indicates "likes the music".

[0043] Among them, audio fingerprint features can be low-level acoustic features extracted from music audio, such as tempo, mood, melody, timbre, beat and other dimensions. In the music field, audio fingerprint features can be regarded as an objective quantitative representation of the audio attributes of a song.

[0044] Optionally, song attribute features include song category features and song metadata features. Song category features can be feature vectors obtained by quantifying the musical style or genre of candidate songs (e.g., rock music, pop music, sleep aid music, etc.) and the subject matter of candidate songs (e.g., children's songs, etc.).

[0045] Among them, the metadata features of a song can include information such as the artist and album to which the song belongs, as well as identification features such as the song ID embedding.

[0046] Among them, the song fusion vector can be a combined feature vector obtained by fusing or splicing the song category features, audio fingerprint features, and song attribute features of the candidate songs according to the feature dimensions. It can be used to represent the fusion of multi-dimensional features of a song.

[0047] Among them, the Multilayer Perceptron (MLP) can be a deep learning model composed of multiple fully connected layers. It has non-linear mapping capabilities and can map high-dimensional, multi-dimensional spliced ​​features into dense vectors with uniform dimensions and concise semantics.

[0048] This can be achieved by fusing the user's explicit search intent (the user's search terms) and implicit historical interests (the user's historical interaction feature vectors), and then filtering from a pre-set music library based on the fusion results. This ensures that the candidate songs obtained after filtering not only match the user's current real-time search needs but also fit the user's long-term music preferences.

[0049] Because music is multi-dimensional and multimodal, a single feature is insufficient to fully represent the information of a song. However, song category features can reflect the style of the music, audio fingerprint features can reflect the acoustic nature of the music, and song metadata features can reflect the singer and related information of the music. Therefore, by fusing and combining the song category features, audio fingerprint features, and song metadata features of candidate songs, a comprehensive and complete characterization of the candidate music can be achieved, avoiding the one-sided problem caused by single feature representation.

[0050] Furthermore, different types of songs have different feature dimensions and representational meanings. We can first obtain a song splicing vector by concatenation to achieve the fusion of multimodal features of the song. Then, we can use a multilayer perceptron (MLP) to map and condense the spliced ​​high-dimensional features to generate candidate song feature vectors with unified dimensions and dense semantics. This makes the candidate song feature vectors more suitable for the computational requirements of subsequent music search models.

[0051] Specifically, the system first receives the user's search terms. Since these natural language search terms possess abstract semantic information, a pre-trained language model (such as a Natural Language Processing model (Bidirectional Encoder Representations from Transformers, BERT) or a Transformer encoder) can be used to transform the abstract textual semantics into a high-dimensional, dense query feature vector. This yields candidate song feature vectors, making the vectors of semantically similar search terms closer together, thus quantifying the user's explicit search intent. Then, using a rule-based model similar to those used in search engine retrieval, a preliminary screening of a pre-defined music library is performed based on the search terms. Simple matching rules quickly filter candidate songs (Candidate Items) related to the search terms from the massive music library, avoiding excessive computation caused by involving the entire music library. Finally, a multimodal feature extraction process, such as using a multilayer perceptron, is employed to obtain the candidate song feature vectors for each song.

[0052] Step 012: Based on the attention mechanism, the weights of the user's historical interaction feature vectors are determined using the intent feature vector. Based on the weights and historical interaction feature vectors, the user's historical feature vector is generated. The intent feature vector is obtained by fusing the query feature vector and the candidate song feature vector. The historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer.

[0053] The composite intent vector can be obtained by fusing the query feature vector and candidate song feature vectors. The intent vector contains constraints on the user's immediate search intent and content information of the candidate songs. For example, the query feature vector and candidate song feature vectors can be input into a dual-tower fusion layer, using a gating mechanism or nonlinear transformation.

[0054] in, This represents vector concatenation. This is a learnable weight matrix.

[0055] obtained in this way This is the intent feature vector, which contains information about what the candidate song is (Candidate Info) and constraints on the user's search intent (Query Info).

[0056] Among them, the historical interaction feature vector can be a feature vector or feature vector matrix obtained by encoding M historical interaction records of the user, where M is a positive integer greater than 0. For example, M is an ultra-long historical interaction feature vector sequence on the order of 10k, which can include user interaction behaviors such as listening to music, collecting music, and sharing music, and can reflect the user's historical music interests.

[0057] The attention mechanism can be a cross-attention mechanism, used to calculate the correlation between the intent feature vector and the historical interaction feature vector, assigning differentiated weights to different historical interaction feature vectors to achieve the effect of focusing on historical behaviors related to the current search intent. For example, the above... As the Query in the attention mechanism, the historical feature vector is used as the Key and Value:

[0058] Based on a multi-head cross-attention mechanism, this process automatically identifies segments in the historical sequence that both match the current search intent and are similar in style to candidate songs, assigning them extremely high weights. The calculation can be performed using stacked L layers of Transformer Blocks, incorporating residual connections and layer normalization to support gradient propagation in deep networks. The historical feature vector can be obtained by aggregating the weighted historical interaction feature vectors. It can be viewed as a condensed representation of the user's historical music interests, guided by their current search intent and after removing historical noise irrelevant to that intent.

[0059] It should be noted that long sequences corresponding to historical interaction feature vectors can also be processed using RNNs (such as variants of GRU4Rec) to generate historical feature vectors. Furthermore, before generating the intent feature vector, a small cross-attention interaction can be performed between the Query and the Candidate, allowing the Query to focus on specific attributes of the Candidate, which can also yield the intent feature vector.

[0060] Specifically, for example, in a scheme that truncates a user's listening history to the most recent 50 to 100 records and uses a short-sequence model to calculate recommended music based on the embedding vectors of these records and the current candidate songs, the truncation operation leads to the loss of this important information. This means that when a user searches for related content again, the short-sequence model, having "forgotten" long-term historical interests, cannot make accurate personalized recommendations, resulting in a degraded user search experience. This application, however, addresses this by using intent feature vectors as the query and historical interaction feature vectors as the key and value. It then calculates the similarity between vectors using a dot product, assigning a corresponding weight value to each historical interaction feature vector. Historical behaviors with higher relevance to the intent feature vector receive higher weights, thus achieving the filtering of historical interaction feature vectors. Then, the weights of the historical interaction feature vectors are summed and pooled to generate historical feature vectors. This avoids the problem of discarding important information in the initial screening stage, reduces the interference of irrelevant historical interaction features on subsequent calculations, and improves the accuracy of feature representation.

[0061] Step 013: Fuse the query feature vector, candidate song feature vector, and historical feature vector to generate a fused feature vector.

[0062] The fused feature vector can be data that incorporates at least the user's current search intent, candidate song content, and the user's intent-driven historical interests.

[0063] Specifically, by fusing query feature vectors (representing the user's current search intent), candidate song feature vectors (representing the specific content of the candidate songs), and historical feature vectors (representing the user's historical preferences under the current intent), a comprehensive information fusion of search intent, candidate content, and the user's personalized historical interests can be achieved, providing a basis for music search. This can be achieved by fusing query feature vectors (representing the user's current search intent), candidate song feature vectors (representing the specific content of the candidate songs), and historical feature vectors (e.g., concatenating the query feature vector, candidate song feature vector, and historical feature vector to obtain a fused feature vector; or by weighting and summing the query feature vector, candidate song feature vector, and historical feature vector, where the weights can be learned by the model or set based on experience; or by using a lightweight attention network, taking one vector as the query, calculating the attention weights with other vectors, and then performing a weighted fusion).

[0064] Step 014: Determine the music search results based on the fused feature vector.

[0065] Optionally, the method further includes: Step 017: Display the music list generated from the music search results.

[0066] Music search can involve sorting multiple candidate songs and displaying the sorted results to the user. For example, multiple candidate songs could be sorted to generate a music list, which is then displayed to the user.

[0067] Thus, by acquiring query feature vectors and multiple candidate song feature vectors, the query feature vector is generated in response to the user's input search terms, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. Based on an attention mechanism, the weights of the user's historical interaction feature vectors are determined using the intent feature vector, and the user's historical feature vector is generated based on the weights and historical interaction feature vectors. The intent feature vector is obtained by fusing the query feature vector and candidate song feature vectors, and the historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer. This avoids discarding important information in the initial screening stage while reducing the interference of irrelevant historical interaction features on subsequent calculations, thus improving the accuracy of feature representation. The query feature vector, candidate song feature vector, and historical feature vector are fused to generate a fused feature vector, achieving full information fusion of search intent, candidate content, and the user's personalized historical interests, providing a complete basis for music search. Finally, based on the fused feature vector, the music search results are determined, maximizing the retention of full-dimensional information of search intent, candidate content, and historical interests, thereby improving the user's music search experience.

[0068] In some implementations, the method further includes: Step 018: Obtain the user's historical interaction sequence, which includes the song encoding feature vector, timestamp feature vector, and interaction type feature vector of the historical songs; Step 019: Generate position vectors for historical songs based on their order in the historical interaction sequence; Step 020: Perform feature fusion on the location vector, song encoding feature vector, timestamp feature vector, and interaction type feature vector to generate historical interaction feature vector.

[0069] Among them, the historical interaction sequence can be the user's behavior sequence (Ultra-Long User BehaviorSequence).

[0070] Among them, the song encoding feature vector can be obtained by encoding and mapping the song's ItemID Embedding, audio fingerprint features, metadata features, etc., and serves as a quantitative expression of the content attributes of historical candidate songs.

[0071] Among them, the timestamp feature vector is a feature vector obtained by quantifying the timestamp (the specific time information of the interaction) of historical song interaction behavior.

[0072] Among them, the interaction type feature vector can be a feature vector after vectorizing the user's interaction behavior type with historical songs (such as play, favorite, share, skip, comment, etc.), which can represent the user's preference for historical songs (such as favorite represents high preference, skip represents low preference).

[0073] The position vector can be a feature vector generated based on the order of historical songs in the historical interaction sequence, i.e., positional encoding, which can be used to represent the chronological order of historical behaviors in the sequence.

[0074] Among them, the historical interaction feature vector can be a feature vector generated by fusing the location vector, song encoding feature vector, timestamp feature vector, and interaction type feature vector. It serves as a multi-dimensional information integration expression for a single historical interaction behavior, including full quantification information of historical song content, interaction time, behavior type, and sequence position.

[0075] Specifically, taking the feature vector obtained through deep learning model processing as an example, deep learning models (such as attention and Transformer) do not have the ability to identify sequence positions. The processing results of unordered feature vectors are consistent. By generating a unique position vector for each candidate song in the historical interaction sequence, the model learns the temporal order of the user's historical behavior and preserves the temporal characteristics of the sequence. By retrieving M historical music interaction sequences from the user, including the song ID, timestamp of the action, and interaction type, and then vectorizing each dimension of information in each interaction record, song encoding feature vector, timestamp feature vector, and interaction type feature vector are generated. For example, the song encoding feature vector can be generated based on the song ID, audio fingerprint, metadata, etc. of historical songs, through encoding or MLP mapping, representing the attributes of the song itself; as another example, the timestamp feature vector can be a dense vector with a time decay coefficient converted from the timestamp of the user's interaction, representing the time freshness of the interaction behavior (e.g., the shorter the time difference from the current moment, the higher the weight of the behavior); as yet another example, the interaction type feature vector can be a numerical vector generated by embedding and encoding interaction types such as play, favorite, skip, and share, representing the user's preference for the song. By associating the song encoding feature vector, timestamp feature vector, and interaction type feature vector of each historical song, and arranging them chronologically according to the occurrence of the interaction, a user's historical interaction sequence with time decay information can be obtained (e.g., a historical sequence vector matrix H = [h1, h2, ..., hn], where n can reach 10000). Each element in the sequence is a combination of the three basic feature vectors of a single historical song. The historical interaction sequence does not require manual truncation, fully preserving the user's long-term, multi-type music interaction behavior information, reflecting the periodicity, diversity, and variability of the user's music tastes, and providing complete data support for extracting the user's personalized historical interests.

[0076] Since the characteristics of a user's historical music interaction behavior are multi-dimensional and complementary (for example, song encoding features reflect the attributes of the interactive object, timestamp features reflect the time characteristics of the interaction, interaction type features reflect the user's subjective preferences, and position vectors reflect the sequence characteristics of the interaction), the position vector, song encoding feature vector, timestamp vector, and interaction type feature vector are fused to obtain the historical interaction feature vector, so that the model can perform unified calculations on multi-dimensional historical information.

[0077] In some implementations, the method further includes: Step 021: Sort the historical feature vectors according to their interaction time to generate a historical feature vector matrix; Step 022: Divide the historical feature vector matrix into blocks to obtain multiple historical feature vector sub-blocks; Step 023: Calculate the similarity parameter between each historical feature vector sub-block and the query feature vector, and update the historical interaction feature vector based on the historical feature vector sub-blocks whose similarity parameter is greater than the preset parameter threshold.

[0078] The interaction time can be the exact moment a user's historical interaction occurred (the timestamp information of the specific interaction). For example, suppose a user added song M to their favorites list at 10:10:10 on a certain day, and skipped song N at 12:10:10 on the same day. Then, the interaction time for song M is 10:10:10, and the interaction time for song N is 12:10:10. Based on the interaction time, the historical interaction feature vector corresponding to song M is sorted before the historical interaction feature vector of song N.

[0079] The historical feature vector matrix can be a two-dimensional feature matrix formed by sorting all the user's historical interaction feature vectors in the order of interaction time. The row dimension can be the number of historical interaction behaviors (e.g., on the order of 10k), and the column dimension can be the feature vector dimension.

[0080] The block processing can be an operation that divides a high-dimensional, long sequence of historical feature vector matrix into several sub-matrices of equal length (or approximately equal length) according to a fixed number of lines. Each sub-matrix is ​​a historical feature vector sub-block (Chunk).

[0081] The historical feature vector sub-block can be a sub-matrix obtained by dividing the historical feature vector matrix into blocks, and each sub-block contains a fixed number of historical interaction feature vectors (e.g., 64).

[0082] The similarity parameter can be calculated by an algorithm and represents the degree of semantic or feature matching between the historical feature vector sub-block and the query feature vector. For example, it can be the cosine similarity with a value range of [-1, 1]. The closer the cosine similarity value is to 1, the higher the matching degree.

[0083] The preset parameter threshold (α) can be a similarity threshold set for filtering relevant historical feature vector sub-blocks, and a quantitative standard for determining whether historical sub-blocks are related to the current search intent.

[0084] Specifically, to address the enormous computational overhead of 10k sequences and remove noise irrelevant to the current search intent (e.g., when searching for "children's songs," historical "heavy metal" records would be considered noise), a pre-filtering mechanism can be introduced. Users' historical music interaction behavior is typically temporally correlated. Sorting historical interaction feature vectors by interaction time preserves the temporal regularity of user behavior. The resulting sub-blocks after block processing also possess temporal continuity, aligning with the user's actual interest changes. Directly calculating the similarity of each historical interaction feature vector (e.g., on the order of 10k) would consume enormous computational resources. However, by dividing the long sequence into several sub-blocks and performing coarse-grained similarity calculations on a sub-block basis, the number of calculations can be significantly reduced, achieving low-cost preliminary filtering. For example, using cosine similarity as the similarity parameter, the degree of matching can be represented by calculating the cosine of the angle between two vectors, measuring the semantic or feature correlation between high-dimensional feature vectors to determine the style and content matching degree between historical feature sub-blocks and the query feature vector (the user's explicit search intent). Then, a binary classification filter is applied to the similarity results using preset parameter thresholds. Historical sub-blocks with similarity below the threshold (which can be considered as severely conflicting with the current search intent) are removed, while only historical sub-blocks with similarity above the threshold (which can be considered as relevant to the current search intent) are retained and integrated into a new historical interaction feature vector. This achieves noise removal and sequence compression, overcoming the computational redundancy and noise interference problems caused by excessively long sequences. In this way, sequence pruning guided by the query feature vector (the user's explicit search intent) ensures that all subsequent calculations are only performed on historical features relevant to the current search intent. This can compress the effective sequence length to 20% to 30% of the original, significantly improving computational speed while ensuring that the extraction of historical information is strictly constrained by the search terms, avoiding irrelevant historical noise interfering with model judgment.

[0085] In some implementations, step 012: Based on an attention mechanism, determining the weights of the user's historical interaction feature vectors using the intent feature vector includes: Step 0121: Based on the attention mechanism, the intent feature vector is used as the query end of the attention mechanism, and the historical interaction feature vector is used as the key and value end of the attention mechanism. The similarity between the intent feature vector and the user's historical interaction feature vector is calculated. Step 0122: Determine the weights of historical interaction feature vectors based on similarity, where similarity and weight are positively correlated.

[0086] In this attention mechanism, the query component can be the feature vector used to initiate a similarity query. The key component can be the feature vector used to calculate similarity with the query component, assigning weights to its corresponding features based on the matching degree with the query component. The value component can be the final weighted feature vector in the attention mechanism, representing the object of weight allocation.

[0087] Specifically, the cross-attention mechanism completes weight calculation through a ternary input of query (Q), key (K), and value (V). Guided by Q, it calculates the similarity between Q and each K, then converts the similarity into weights and applies them to the corresponding V, achieving differentiated weighting of V. In this application, K and V reuse historical interaction feature vectors, simplifying calculation while ensuring the accuracy of weight allocation. Using an intent feature vector that integrates the explicit intent of the search term and the content information of candidate songs as Q allows similarity calculation to revolve around the current search intent and candidate content, ensuring that only historical interactions highly related to both are calculated with high similarity. This addresses the shortcomings of traditional models that only focus on candidate content and ignore search intent. Finally, by normalizing the similarity parameter into weight values ​​and setting a positive correlation between similarity and weight, high-similarity historical interaction feature vectors receive a higher weight proportion, while low-similarity features are weakened, achieving soft filtering of historical features.

[0088] In some implementations, step 013: fusing the query feature vector, candidate song feature vector, and historical feature vector to generate a fused feature vector, including: Step 0131: Obtain the user's profile feature vector. The profile feature vector is used to represent the user's listening preferences when listening to music, including at least one of the following: song category features, song melody features, and singer preference features. Step 0132: Fuse the portrait feature vector, query feature vector, candidate song feature vector, and historical feature vector to generate a fused feature vector.

[0089] Among them, the user's profile feature vector can be a high-dimensional dense numerical vector used to represent the user's inherent and long-term listening preferences in music listening. The feature dimension can include at least one of the following: song category, melody, singer preference, etc., which can be regarded as the user's music preference features that have not changed within a preset time period.

[0090] Among them, the song category feature can be a quantitative vector of the user's long-term preferred music style and category features (such as rock, pop, traditional Chinese style, sleep aid, etc.), reflecting the user's preference for music categories.

[0091] Among them, the melody features of a song can be a quantitative vector of the musical melody attributes that users have long preferred (such as soothing, exciting, light, and deep), reflecting users' preferences for the acoustic melody of music.

[0092] Among them, the singer preference feature can be a quantitative vector of the relevant features of singers or bands that users have been paying attention to and listening to frequently, reflecting the user's preference for specific music creators.

[0093] Listening preference can be a relatively stable music selection tendency formed by users during long-term music interaction.

[0094] Specifically, a fused feature vector is obtained by directly concatenating user profile feature vectors, query feature vectors, candidate song feature vectors, and historical feature vectors according to feature dimensions. User inherent preferences, immediate search intent, candidate content attributes, and intent-guided historical interests are also concatenated to obtain a fused feature vector. Without changing the feature values ​​and semantic information of the original vectors, the original information of each dimension of features is preserved to the greatest extent, while a standardized single feature vector is generated to adapt to subsequent calculation requirements.

[0095] It is understandable that a user's music interaction behavior is determined by both stable preferences and temporary dynamic interests. The historical feature vector is the dynamic historical interest guided by the current search intent, while the profile feature vector is the user's long-term stable listening preferences. The combination of the two can more accurately depict the full picture of a user's music interests, solving the problem that dynamic features alone cannot reflect the user's inherent preferences.

[0096] In some implementations, the music search results include a music list. Step 014: Determine the music search results based on the fused feature vector, including: Step 0141: Perform probability calculations on the fused feature vectors to determine the probability values ​​of the user selecting each candidate song feature vector; Step 0142: Sort the candidate songs corresponding to each candidate song feature vector according to the probability value to generate a music list. In the music list, the sorting number and the probability value are negatively correlated.

[0097] The probability value can be used to quantify the likelihood of a user selecting a candidate song. The higher the probability value of a candidate song, the greater the likelihood that the user will select that candidate song.

[0098] The music list can be a recommendation list generated by sorting all candidate songs according to the user's selection probability value (from highest to lowest). The higher the probability value of a user selecting a candidate song, the higher the ranking of the candidate song in the music list; the lower the probability value, the lower the ranking of the candidate song.

[0099] Specifically, the concatenated feature vector integrates multi-dimensional information such as user's inherent preferences, immediate search intent, candidate song content, and intent-driven historical interests. Based on the concatenated feature vector, the probability value of the user selecting candidate songs is calculated. According to the negative correlation rule that the higher the probability value, the higher the ranking, the more prominent the candidate songs that best match the user's needs appear in the recommendation list, maximizing the matching of the user's search and listening needs and improving the user's click and listening experience.

[0100] In some implementations, the method further includes: Step 024: Construct a music search model. The music search model includes an attention layer, a feature aggregation layer, and a probability output layer. The attention layer is used to determine the weights of the user's historical interaction feature vectors. The feature aggregation layer is used to generate fused feature vectors. The probability output layer is used to calculate the probability of the fused feature vectors to obtain the probability values ​​of the user selecting each candidate song feature vector. The music search model is trained on the attention layer, feature aggregation layer, and probability output layer through gradient backpropagation.

[0101] The music search model can be a deep learning model that integrates an attention layer, a feature aggregation layer, and a probability output layer.

[0102] The attention layer can be a feature weighting layer of the model, based on a two-stage stacked cross attention (STCA) mechanism, which uses the intent feature vector to determine the attention weights of the user's historical interaction feature vector.

[0103] Among them, the feature aggregation layer can be the feature fusion layer of the music search model, which can concatenate user profile feature vectors, query feature vectors, candidate song feature vectors, and historical feature vectors to generate a fused feature vector containing full-dimensional information.

[0104] The probability output layer can be the prediction output layer of the music search model. It can perform nonlinear transformation and probability mapping on the fused feature vector, calculate and output the probability value (CTR Score) of the user selecting each candidate song feature vector.

[0105] Backpropagation is a training optimization method for deep learning models. It can propagate the error (loss value) between the model's predicted value and the true label from the output layer to the input layer. By calculating the gradient of the parameters of each layer and updating the parameters, it can achieve collaborative optimization of each layer of the model, allowing the model's prediction results to continuously approach the true value.

[0106] Specifically, please refer to Figure 3 , Figure 3An example using 10,000 user historical interaction sequences illustrates the music search model. An attention layer can be built based on stacked target-history cross-attention (STCA) (or directly using standard multi-head attention), including a multi-head cross-attention calculation module and an L-layer TransformerBlock stacking module, while introducing residual connections and layer normalization. The model takes the intent feature vector (Query) and pruned historical interaction feature vectors (Key / Value) as input, and the attention-weighted historical feature vector as output, achieving intent-guided historical feature weighting. The user profile feature vector, query feature vector, candidate song feature vector, and the historical feature vector output from the attention layer are used as input to the feature aggregation layer, and the high-dimensional fusion feature vector after four-dimensional feature concatenation is used as output, achieving full-dimensional feature aggregation. Finally, multiple fully connected layers (FCLayers) and Sigma... The probability output layer is composed of an osi activation function, a fully connected layer is used for feature condensation and dimension mapping, and a sigmoid function is used for probability mapping. The input of the probability output layer is the fused feature vector output by the feature aggregation layer. After nonlinear transformation by the fully connected layer, the output value is mapped to the 0~1 range by the sigmoid activation function. The final output is the probability value (CTRScore) of the feature vector of each candidate song selected by the user. The model can be trained using the cross-entropy loss function (Log-Loss): the predicted probability value of the probability output layer and the user's actual selected label are used as inputs, and the error (loss value) between the model's predicted value and the actual value is calculated until the loss value is less than a preset threshold to determine the music search model training fit. By propagating the loss error through gradient backpropagation to all layers of the model, end-to-end parameter optimization is achieved from historical feature weighting to feature aggregation and probability prediction. At the same time, the gradient can be further propagated to the preceding feature encoding and sequence pruning stages, solving the problem that the two-stage fragmented training in existing technologies cannot optimize the entire process, resulting in better overall model performance. The feature aggregation layer integrates multi-dimensional features through direct dimensional concatenation without changing the original feature values ​​and semantic information, maximizing the preservation of full-dimensional information such as user profile, search intent, candidate content, and historical interests. This provides complete and accurate feature input for the probability output layer. Moreover, the model is trained directly using raw, untrunculated ultra-long sequence data (i.e., end-to-end training), which significantly improves the prediction accuracy of user selection probability values.

[0107] It is understood that in this embodiment, the coarse screening of historical blocks based on the query vector can eliminate 70%-80% of historical noise, significantly reducing inference time while ensuring accuracy. The improved two-stage STCA architecture, adapted to search scenarios, solves the feature dilution problem caused by simply concatenating Query, Candidate, and History. Through layered processing, it ensures the dominance of explicit search intent in long sequence modeling, improving the relevance of search results and personalized click-through rate (CTR).

[0108] Please see Figure 4 The present application also proposes a method for displaying music search results, the method including: Step 031: In response to the music search method described in any of the above embodiments, obtain music search results; Step 032: Based on the music search results, display the music list corresponding to the music search results.

[0109] Specifically, after receiving a user's search command, the music server or terminal can generate music search results by calling a music search model and display the music list corresponding to the search results on a screen such as the terminal's display screen. The displayed music list matches the user's immediate search intent and long-term listening preferences, improving the relevance of the results, simplifying the search operation, and enhancing the user's experience and satisfaction.

[0110] All of the above technical solutions can be combined in any way to form optional embodiments of this application, and will not be described in detail here.

[0111] To facilitate better implementation of the music search method of this application embodiment, this application embodiment also provides a music search device. Please refer to... Figure 4 , Figure 4 A schematic diagram of the structure of a music search device provided in an embodiment of this application. The music search device 200 may include: The acquisition module 201 is used to acquire query feature vectors and multiple candidate song feature vectors. The query feature vectors are generated in response to the search terms input by the user, and the candidate song feature vectors are obtained by filtering the preset music library based on the search terms. The generation module 202 is used to determine the weight of the user's historical interaction feature vector based on the attention mechanism and the intention feature vector, and generate the user's historical feature vector based on the weight of the historical interaction feature vector and the historical interaction feature vector. The intention feature vector is obtained by feature vector fusion based on the query feature vector and the candidate song feature vector, and the historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer. The fusion module 203 is used to fuse the query feature vector, the candidate song feature vector, and the historical feature vector to generate a fused feature vector; Module 204 is used to determine music search results based on the fused feature vector.

[0112] Each unit in the aforementioned music search device can be implemented entirely or partially through software, hardware, or a combination thereof. These units can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can invoke and execute the corresponding operations of each unit.

[0113] The music search device 200 can be integrated into a terminal or server that has storage and a processor and thus computing power, or the music search device 200 can be the terminal or server.

[0114] To facilitate better implementation of the music search method of this application embodiment, this application embodiment also provides a music search device. Please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of a music search result display device provided in an embodiment of this application. The music search result display device 300 may include: Response module 301 is configured to obtain the music search results in response to the music search method described in any of the above embodiments; The display module 302 is used to display a music list corresponding to the music search results based on the music search results.

[0115] Each unit in the aforementioned music search device can be implemented entirely or partially through software, hardware, or a combination thereof. These units can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can invoke and execute the corresponding operations of each unit.

[0116] The music search results display device 300 can be integrated into a terminal or server that has storage and a processor and thus computing power, or the music search results display device 300 can be the terminal or server.

[0117] Optionally, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0118] Figure 6This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device may be a terminal or a server. As shown in the figure, the computer device 400 includes a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, and a computer program stored on the memory 402 and executable on the processor. The processor 401 and the memory 402 are electrically connected. Those skilled in the art will understand that the computer device structure shown in the figure does not constitute a limitation on the computer device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0119] The processor 401 is the control center of the computer device 400. It connects various parts of the computer device 400 through various interfaces and lines. By running or loading software programs and / or modules stored in the memory 402, and calling data stored in the memory 402, it performs various functions of the computer device 400 and processes data, thereby performing overall processing of the computer device 400.

[0120] In this embodiment, the processor 401 in the computer device 400 loads the instructions corresponding to the processes of one or more computer programs into the memory 402 according to the following steps, and the processor 401 runs the computer programs stored in the memory 402 to realize various functions: Obtain a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the search terms input by the user, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. Based on the attention mechanism, the weights of the user's historical interaction feature vectors are determined using the intent feature vector, and the user's historical feature vector is generated based on the weights of the historical interaction feature vectors and the historical interaction feature vectors. The intent feature vector is obtained by feature vector fusion based on the query feature vector and the candidate song feature vector, and the historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer. The query feature vector, the candidate song feature vector, and the historical feature vector are fused to generate a fused feature vector; Based on the fused feature vector, music search results are determined.

[0121] or: The music search results are obtained in response to the music search method described in any of the above embodiments; Based on the music search results, a list of music corresponding to the music search results is displayed.

[0122] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0123] Optional, such as Figure 6 As shown, the computer device 400 also includes a display screen 403, a radio frequency circuit 404, an audio circuit 405, an input unit 406, and a power supply 407. The processor 401 is electrically connected to the display screen 403, the radio frequency circuit 404, the audio circuit 405, the input unit 406, and the power supply 407. Those skilled in the art will understand that the computer device structure shown in the figures does not constitute a limitation on the computer device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0124] The display screen 403 can be used to display a graphical user interface (GUI) and receive operation commands generated by the user interacting with the GUI. The display screen 403 may include a display panel and a touch panel. The display panel can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the computer device. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. The touch panel can be used to collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near the touch panel), generate corresponding operation commands, and execute the corresponding program. Optionally, the touch panel may include a touch detection device and a touch controller. The touch detection device detects the user's touch location and the signal generated by the touch operation, and transmits the signal to the touch controller. The touch controller receives touch information from the touch detection device, converts it into touch point coordinates, sends it to the processor 401, and can receive and execute commands from the processor 401. The touch panel can cover the display panel. When the touch panel detects a touch operation on or near it, it transmits the information to the processor 401 to determine the type of touch event. Subsequently, the processor 401 provides corresponding visual output on the display panel according to the type of touch event. In this embodiment, the touch panel and the display panel can be integrated into the display screen 403 to achieve input and output functions. However, in some embodiments, the touch panel and the display screen 403 can be implemented as two independent components to achieve input and output functions. That is, the display screen 403 can also be used as part of the input unit 406 to achieve input functions.

[0125] The radio frequency circuit 404 can be used to transmit and receive radio frequency signals to establish wireless communication with network devices or other computer devices, and to transmit and receive signals with network devices or other computer devices.

[0126] Audio circuitry 405 can be used to provide an audio interface between a user and a computer device via a speaker and a microphone. Audio circuitry 405 can convert received audio data into electrical signals and transmit them to the speaker, where the speaker converts them into sound signals for output. Conversely, the microphone converts collected sound signals into electrical signals, which are then received by audio circuitry 405, converted back into audio data, and then processed by processor 401 before being transmitted via radio frequency circuitry 404 to, for example, another computer device, or output to memory 402 for further processing. Audio circuitry 405 may also include an earphone jack to facilitate communication between peripheral headphones and the computer device.

[0127] The input unit 406 can be used to receive input numbers, characters, or object feature information (such as fingerprints, iris, facial information, etc.), and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control.

[0128] Power supply 407 is used to supply power to various components of computer device 400. Optionally, power supply 407 can be logically connected to processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. Power supply 407 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0129] although Figure 6 As not shown in the diagram, computer equipment 400 may also include a camera, sensor, wireless fidelity module, Bluetooth module, etc., which will not be described in detail here.

[0130] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding process in the music search method of the embodiments of this application; for brevity, further details are omitted here.

[0131] This application also provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding process in the music search method of this application embodiment. For simplicity, further details are omitted here.

[0132] This application also provides a computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding process in the music search method of this application; for brevity, this will not be elaborated further.

[0133] It should be understood that the processor in this application may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0134] It is understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0135] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0136] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0137] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0138] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0139] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0140] In addition, the functional units in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0141] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer or a server) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0142] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A music search method, characterized in that, include: Obtain a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the search terms input by the user, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. Based on the attention mechanism, the weights of the user's historical interaction feature vectors are determined using the intent feature vector, and the user's historical feature vector is generated based on the weights of the historical interaction feature vectors and the historical interaction feature vectors. The intent feature vector is obtained by feature vector fusion based on the query feature vector and the candidate song feature vector, and the historical interaction feature vector is generated based on the user's M historical interaction records, where M is a positive integer. The query feature vector, the candidate song feature vector, and the historical feature vector are fused to generate a fused feature vector; Based on the fused feature vector, music search results are determined.

2. The music search method according to claim 1, characterized in that, The method of determining the weights of a user's historical interaction feature vectors based on an attention mechanism, utilizing intent feature vectors, includes: Based on the attention mechanism, the intent feature vector is used as the query end of the attention mechanism, and the historical interaction feature vector is used as the key and value end of the attention mechanism to calculate the similarity between the intent feature vector and the user's historical interaction feature vector. The weights of the historical interaction feature vectors are determined based on the similarity, wherein the similarity and the weights are positively correlated.

3. The music search method according to claim 1, characterized in that, The step of fusing the query feature vector, the candidate song feature vector, and the historical feature vector to generate a fused feature vector includes: Obtain the user's profile feature vector, which is used to characterize the user's listening preferences when listening to music, including at least one of song category features, song melody features, and singer preference features; The portrait feature vector, the query feature vector, the candidate song feature vector, and the historical feature vector are fused to generate a fused feature vector.

4. The music search method according to claim 1 or 3, characterized in that, The music search results include a music list, and determining the music search results based on the fused feature vector includes: Probability calculations are performed on the fused feature vectors to determine the probability values ​​for each of the candidate song feature vectors selected by the user. Based on the probability values, the candidate songs corresponding to each candidate song feature vector are sorted to generate the music list. In the music list, the sorting sequence number and the probability value are negatively correlated.

5. A method for displaying music search results, characterized in that, include: In response to the music search method according to any one of claims 1-4, the music search results are obtained; Based on the music search results, a list of music corresponding to the music search results is displayed.

6. A music search device, characterized in that, The device includes: The acquisition module is used to acquire a query feature vector and multiple candidate song feature vectors. The query feature vector is generated in response to the search terms input by the user, and the candidate song feature vectors are obtained by filtering a preset music library based on the search terms. The generation module is used to determine the weights of the user's historical interaction feature vectors based on the attention mechanism and the intent feature vectors, and to generate the user's historical feature vectors based on the weights of the historical interaction feature vectors and the historical interaction feature vectors. The intent feature vectors are obtained by feature vector fusion based on the query feature vectors and the candidate song feature vectors, and the historical interaction feature vectors are generated based on M historical interaction records of the user, where M is a positive integer. The fusion module is used to fuse the query feature vector, the candidate song feature vector, and the historical feature vector to generate a fused feature vector; The determination module is used to perform music search based on the fused feature vector.

7. A device for displaying music search results, characterized in that, include: A response module is configured to obtain the music search results in response to the music search method according to any one of claims 1-4; The display module is used to display a list of music corresponding to the music search results.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted for loading by a processor to perform the music search method as described in any one of claims 1-4 or the music search results display method as described in claim 5.

9. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, and the processor executing the music search method according to any one of claims 1-4 or the music search result display method according to claim 5 by calling the computer program stored in the memory.

10. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the music search method according to any one of claims 1-4 or the music search result display method according to claim 5.

Images