Supervised Contrastive Learning for Recommending Related Content

Abstract

Aspects of the present disclosure relate to recommending related content in response to a user's search query by supervising the training of paired embeddings using contrastive learning and pairwise co-click signals. This approach combines a two-tower model architecture with a cascaded multilayer perceptron model to enable variable combinations of input features and the adoption of more representative learned paired embeddings. The learned embeddings are subjected to supervised contrastive loss training to generate a related content recommendation model, which is then evaluated using both online and offline metrics. The related content recommendation model can provide results that improve the quality of recommendations for search queries and increase user engagement, ultimately improving the long-term user experience.

Description

[Background technology]

[0001] background Relevant content recommendations have come to play an important role in modern life with the proliferation of internet search and recommendation platforms. Unfortunately, combining diverse and engaging related content recommendations for users is a challenging task. For example, recommendations based solely on semantic context may return results that lack diversity, potentially causing users to lose interest in the search and even the search platform itself.

[0002] It is with respect to these and other general considerations that the embodiments are described. Furthermore, while relatively specific problems are discussed, it should be understood that the embodiments should not be limited to solving the specific problems identified in the context. Summary of the Invention

[0003] overview Aspects of the present disclosure relate to recommending related content in response to a user's search query by supervising the training of paired embeddings using contrastive learning and pairwise co-click signals. This approach combines a two-tower model architecture with a cascaded multilayer perceptron model to enable variable combinations of input features and the adoption of more representative learned paired embeddings. The related content recommendation model is subjected to supervised contrastive loss training to generate learned embeddings, and the model is then evaluated using both online and offline metrics. The related content recommendation model can provide results that improve the quality of recommendations for search queries and increase user engagement, ultimately improving the long-term user experience.

[0004] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

[0005] BRIEF DESCRIPTION OF THE DRAWINGS Non-limiting and non-exhaustive examples are described with respect to the following figures: [Brief explanation of the drawings]

[0006] [Figure 1] 1 illustrates an overview of an example system for recommending related content in response to a user's search query. [Figure 2] 1 outlines an example method for recommending related content based on a user's search query. [Figure 3] We outline an example of a method for modeling combined feature vectors using a single cascaded multi-layer perceptron (MLP) neural network. [Figure 4] 1 is an example of a user interface of a search query result with recommended related content. [Figure 5] FIG. 1 is a block diagram illustrating an example of physical components of a computing device in which aspects of the present disclosure can be practiced. [Figure 6A] FIG. 1 is a simplified block diagram of a mobile computing device in which aspects of the present disclosure can be practiced. [Figure 6B] FIG. 1 is a simplified block diagram of a mobile computing device in which aspects of the present disclosure can be practiced. [Figure 7] FIG. 1 is a simplified block diagram of a distributed computing system in which aspects of the present disclosure may be practiced. DETAILED DESCRIPTION OF THE INVENTION

[0007] Detailed Description In the following detailed description, reference is made to the accompanying drawings which form a part hereof, and which show, by way of illustration, specific embodiments or examples. These aspects may be combined, other aspects may be utilized, and structural changes may be made without departing from the disclosure. The embodiments may be embodied as a method, system, or apparatus. Thus, the embodiments may take the form of a hardware implementation, an entirely software implementation, or an implementation combining software and hardware aspects. Therefore, the following detailed description is not to be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims and their equivalents.

[0008] In an example, individuals rely on online search platforms to generate results for search queries on various topics. In addition to initial results based on the search query, search platforms and recommender systems can generate related content recommendations in response to the user's query. Related content recommendations provide additional information on similar topics that the user may choose to explore. The task of recommending related content may require the recommender system to suggest related items for a given query or risk losing user engagement with the platform. For example, if a user is browsing images, the recommender system may generate a set of related images to provide to the user. The goal of such recommendations is to improve user engagement within the recommender system and within the broader context of the search query compared to alternative content recommendation systems. Other approaches may employ pre-trained convolutional neural network functions or pre-trained page-level language embeddings to generate visually and / or semantically similar items in response to a search query. As a result, results returned in response to a search query may tend to provide a monotonous and boring user experience that under-includes content that the user might otherwise deem relevant.

[0009] Accordingly, aspects of the present disclosure relate to recommending relevant content in response to a user's search query. This approach aims to train a recommender system by supervising image-page pair embeddings using pairwise co-click signals. As used herein, a co-click signal associates an instance of content with which a user has interacted (e.g., by clicking or tapping on the content, zooming in on the content, etc.) with content that responded to a search query, thereby indicating that the instance of content is a "positive pair" in relation to the responding content. Conversely, a co-click signal may be used to identify a "negative pair" for an instance of content with which the user does not interact. Pairwise co-click signals improve the quality of image recommendations and user engagement by generating suggested content based on quantifiable user feedback, thereby enhancing the user experience. By relying on previously selected content to generate new suggestions, the model results in a higher probability of engaging users. More specifically, this approach utilizes supervised control loss learning to guide the supervision process and improve the quality of recommendations. In aspects, the two-tower model is combined with a cascaded multi-layer perceptron (MLP) model to allow for variable combinations of input features that result in more representative learned embeddings. This approach combines the goal of supervised control learning with improved user engagement, which produces positive results in both the short and long term. Following this supervision process, users receive results of relevant content tailored to their needs. In some cases, the results can be user-specific or may be based on generalized results. The improved model then makes relevant recommendations to the user to maximize user engagement.

[0010] 1 illustrates an overview of an example system 100 for recommending related content in response to a user's search query, according to aspects described herein. As shown, the system 100 includes a user device 102, an application 104, a network 120, and a related content recommendation engine 130. In the example, the user device 102 and the related content recommendation engine 130 communicate over the network 120, which may include a local area network, a wireless network, or the Internet, or any combination thereof, among other examples.

[0011] The user device 102 can be any of a variety of devices, including, but not limited to, a desktop computer, a laptop computer, a tablet, and a wireless device. In an example, the application 104 is an application on the user device 102 that displays content for use on the user device 102 and communicates over the network 120. The application 104 can be a native application, a web-based application, or any combination thereof, among other examples. The application 104 can operate substantially local to the user device 102 or can operate according to a server / client paradigm in conjunction with one or more servers (e.g., the related content recommendation engine 130 and / or any of various other servers not shown). In certain aspects, the application 104 can be an application that includes or accesses a search function that can use a search query to obtain a list of results including content related to the search query. The search query can include and / or be directed to any of a variety of content, including, but not limited to, text, documents, image content, video content, news, websites, web addresses, and / or shopping products or information. For example, a text search query may be used to obtain image content responsive to the search query. It should be understood that the described list of potential search queries is not exhaustive and that a search query may include any type of content accessible or obtainable by the related content recommendation engine 130. In examples, a search query may be initiated on the application 104 by a user in a variety of ways, including, but not limited to, entering text and / or image content into a search field or by the user selecting an image or link displayed on the application (among other examples).

[0012] The related content recommendation engine 130 can receive a search query over the network 120 (e.g., from the application 104) and process it to generate related content, which can be returned over the network 120 to the user device 102 for display and / or selection within the application 104. The related content recommendation engine 130 is shown to include a search query feature generator 132, a search query cascaded MLP 134, a search query embedding engine 136, a related content feature generator 138, a related content cascaded MLP 140, a related content embedding engine 142, a dot product generator 146, a dot product training engine 148, a related content model generator 150, and a related content model evaluator 152.

[0013] In an example, the related content recommendation engine 130 can receive a user's search query from the network 120. The related content recommendation engine 130 can utilize a two-tower approach, with one machine learning tower for learning representations of the search query and another machine learning tower for related content responses to the search query (e.g., including possible related content that may be recommended by the system). Thus, in some examples, each tower can have a related content type. Each tower can be designed as a cascaded MLP model that converts input feature vectors into learned representation embeddings for recommending related content. An example of the two-tower approach is shown in FIG. 1 , where one tower for search queries is shown to include a search query feature generator 132, a search query cascaded MLP 134, and a search query embedding engine 136, and the other tower for related content is shown to include a related content feature generator 138, a related content cascaded MLP 140, and a related content embedding engine 142. The two-tower model of the related content recommendation engine 130 is beneficial because the set of input features processed by each tower may be different, and therefore there will be a different set of features for the search query tower compared to the set of features for the related content tower. It will be understood that the sets of features need not be mutually exclusive. While the following description begins with the search query tower first, it should be understood that the operating order of the system could equally well function by processing either tower first, by processing both towers simultaneously, or by some other order that maintains the linearity of each tower as a prerequisite.

[0014] Starting with the two-tower model of search query tower, the search query feature generator 132 receives or may receive raw input features related to the search query (e.g., which may be received over the network 120). The content type associated with the nature of the raw input features may vary based on the type of search query provided by the user. In aspects, the raw input features may be a high-dimensional unlabeled input dataset including multiple features. For example, if the search query is related to an image search, the raw input features may include features such as a text query (e.g., text embeddings), image features (e.g., vision embeddings, optical character recognition, attractiveness, etc.), page features (e.g., page title, surrounding text, etc.), and context features (e.g., user information, device information, timestamp, location, language, etc.).

[0015] High-dimensional raw inputs can utilize preprocessing to discover low-dimensional signature features that capture their underlying structure. In an example, the process of generating low-dimensional features may utilize feature scaling, centering, and / or dimensionality reduction. Feature scaling can be performed in several ways, including but not limited to absolute-max scaling, min-max scaling, normalization, standardization, and / or robust scaling. Dimensionality can be reduced using one or more feature selection techniques, including but not limited to missing value ratio, low-variance filter, high-correlation filter, random forest, backward feature elimination, and forward feature selection. Additionally or alternatively, dimensionality can be reduced based on factor techniques, including but not limited to factor analysis, principal component analysis, and independent component analysis, or by projection-based dimensionality reduction techniques (e.g., including t-distribution stochastic neighborhood embedding, ISOMAP, and UMAP). Once preprocessing is complete, features can be collected and / or combined into a search query combined feature vector.

[0016] The search query combined feature vector from the search query feature generator 132 is processed according to the search query cascaded MLP 134. In one example, the search query cascaded MLP 134 is a feedforward neural network that connects multiple layers in a directed graph so that signal paths through nodes pass in a single direction to facilitate training of the model's features. The search query cascaded MLP 134 starts with input data of the combined feature vector, which is provided to the first MLP layer. A cascaded MLP is one in which multiple MLP layers are stacked and processed to obtain an appropriately sized output from the input vector. At each layer of the MLP, the input can be expanded or contracted. When the input is expanded, complexity is added to the input and output to the next layer. When the input is contracted, the input is instead condensed so that the reduced final output is the vector used for search. In various embodiments, the first MLP layer can be an expansion layer, while subsequent layers can be contractions. Data scaling can occur between MLP layers. Data scaling methods, including, for example, data normalization (e.g., batch normalization, activation, dropout) or data standardization, can be applied to the output of each layer of a cascaded MLP in preparation for input to the next cascaded MLP layer. Batch normalization can include standardizing the output from various neurons in a neural network in preparation for input to another layer of the MLP. Activation adds nonlinearity to the computation, which can increase the complexity of the input data for the cascaded MLP model and subsequent MLP layers. Dropout can avoid or reduce the possibility of overfitting between layers by ignoring or dropping out the output of certain layers as subsequent inputs to the next MLP layer, thereby reducing the size of the model in the next MLP layer. In various embodiments, a cascaded MLP can also employ skip connections between the initial input and other subsequent cascaded MLP layers, allowing one or more input gradients to be passed directly to the subsequent layer without being processed by the previous MLP layer. When skip connections are utilized, the training process can be significantly accelerated in some instances. In an example, a cascaded MLP may have a structure such that an input is followed by a first MLP layer that can expand the input.Batch normalization can then be performed, which is fed into a second MLP layer to shrink the input, followed by another batch normalization, and so on through successive MLP layers. Skip connections can be made throughout the MLP (e.g., between two layers) to speed up the training process in determining a trained combined feature vector for a search query as the final output of the cascaded MLP.

[0017] The final output of the search query response (e.g., from the search query cascaded MLP 134) is sent or provided to the search query embedding engine 136 so that relationships between categories of output can be represented as embedding vectors. For example, the embedding can capture nearby similar inputs in the embedding space, which can be learned and reused subsequently or across different models. To generate related content recommendations (e.g., by the dot product generator 146), the search query embedding engine 136 can generate links in which one or more relationships between similar items as the initial search query can be passed to a related content model. Thus, the search query embedding can be one half of a pair embedding used by the dot product generator 146, while the other half of the pair embedding can be generated by the related content tower (e.g., elements 138, 140, and 142) of a two-tower model.

[0018] While the search query tower processes the search query embeddings, the related content tower functions in a similar manner. Initially, the related content feature generator 138 receives raw input features of the related content over the network 120. The nature of the raw input features varies based on the type of search query initiated by the user. In aspects, the raw input features may be a high-dimensional unlabeled input dataset including multiple features. For example, if the search query is an image search, the raw input features for the related content tower may include features such as a text query (e.g., text embeddings), image features (e.g., vision embeddings, optical character recognition, attractiveness, etc.), page features (e.g., page title, surrounding text, etc.), and context features (e.g., user information, device information, timestamp, location, language, etc.). The high-dimensional dataset may utilize preprocessing to generate low-dimensional features that capture the underlying structure of the raw input. In examples, the process of generating low-dimensional features may also utilize feature scaling, centering, and / or dimensionality reduction. Feature scaling can be performed in several ways, including absolute-max scaling, min-max scaling, normalization, standardization, and / or robust scaling. Dimensionality can be reduced in a variety of ways, including but not limited to feature selection techniques, including missing value ratio, low variance filters, high correlation filters, random forests, backward feature elimination, and forward feature selection. Additionally or alternatively, dimension reduction can be based on factor techniques, including factor analysis, principal component analysis, and independent component analysis, or by projection-based dimension reduction techniques, including t-distribution stochastic neighborhood embedding, ISOMAP, and UMAP. Once preprocessing is complete, features are collected and concatenated into a related content joint feature vector.

[0019] The related content combined feature vector is input to the related content cascaded MLP 140. The related content cascaded MLP 140 is a feedforward neural network that connects multiple layers in a directed graph so that signal paths through nodes only pass in a single direction to facilitate training of the model's features. The related content cascaded MLP 140 can start with input data of the combined feature vector, which is fed to the first MLP layer. A cascaded MLP is a system in which multiple MLP layers are stacked and processed to obtain an appropriately sized output from the input vector. At each layer of the MLP, the input can be expanded or contracted. When the input is expanded, it means that complexity is added to the input and output to the next layer. When the input is contracted, it is condensed so that the reduced final output is the vector used for search. In various embodiments, the first MLP layer can be an expansion layer, while subsequent layers can be contractions. Data scaling can occur between MLP layers. Data scaling methods, such as data normalization (e.g., batch normalization, activation, dropout) or data standardization, are applied to the output of each layer of a cascaded MLP in preparation for input to the next cascaded MLP layer. Batch normalization may involve standardizing the output from various neurons in a neural network in preparation for input to another layer of the MLP. Activation adds nonlinearity to the calculations to increase the complexity of the cascaded MLP model, allowing for more complex input data for subsequent MLP layers. Dropout avoids the problem of overfitting between layers by ignoring or dropping out the output of certain layers as subsequent inputs to the next MLP layer, thereby reducing the size of the model in the next MLP layer. In various embodiments, cascaded MLPs may also employ skip connections between the initial input and other subsequent cascaded MLP layers, allowing input gradients to be passed directly to subsequent layers without being processed by the previous MLP layer. When skip connections are utilized, the training process can be significantly accelerated in some instances. In an example, a cascaded MLP may have a structure such that an input is followed by a first MLP layer that can expand the input.Batch normalization can then be performed, which is fed into a second MLP layer to shrink the input, followed by another batch normalization, and so on through successive MLP layers. To speed up the training process in determining the trained combined feature vector for a search query as the final output of the cascaded MLP, skip connections can be made between each layer throughout the MLP.

[0020] The final output of related content is sent to the related content embedding engine 142 so that the relationships between categories of output can be expressed as embedding vectors. The embeddings can capture similar inputs nearby in the embedding space, which can then be learned and reused across models. To generate related content recommendations, the related content embeddings create links that can leverage relationships between items similar to the initial search query. The related content embeddings are thus half of a pair embedding that is sent to the dot product generator 146 along with the search query embedding discussed above.

[0021] In the dot product generator 146, the results of each half of the two towers (which may be generated, for example, by the search query embedding engine 136 and the related content embedding engine 142) are used to generate a dot product, which may be a single, untrained combination of the search query and related content. In an example, the dot product may have multiple instances of related content, which may be provided in association with the search query as a set of content recommendations for the initial search. For example, if the initial search query was for images of golden retriever puppies, the search query may be paired with various images corresponding to the features and embeddings derived from the related content towers. In this example, the related content portion of the dot product may include various candidate images related to golden retriever puppies (e.g., golden retriever puppies alone, golden retriever puppies running, a group of golden retriever puppies, etc.).

[0022] The dot product may be used to perform supervised training in the dot product training engine 148. As an example, the dot product training engine 148 employs self-supervised representation learning during training. In an example, a generative or discriminative method may be employed. When a discriminative method is employed for training, it may be, for example, an auxiliary task or a control loss. In an example, a supervised control loss is employed to train the model by tracking the results of pairwise co-click signals of user-selected related content derived from untrained dot products. The pairwise co-click process involves contrasting user-selected positive dot product pairs with a large dataset of negative dot product pairs based on the user's co-click signals derived from related content options presented to the user. FIG. 4 illustrates an example of a user interface that may be presented to a user. In this manner, candidate related content pairs may be provided for display to a user in response to a search query received from the user's device. For example, related or recommended content may be presented in association with content that responded to the user's search query (which may be identified, for example, by a content recommendation engine or another computing device). The user can select one of the given related content pairs, so that the user's selection can be stored for subsequent processing. Thus, a positive pair indicates related content that was selected by the user. A negative pair is an instance of related content that was not selected.

[0023] Over multiple iterations, the contrast loss supervised training techniques described herein can be used to generate a result set of high-scoring positive pairs and low-scoring negative pairs based on user selection (e.g., based on one or more of the co-click signals described above). High-scoring positive pairs are dot-product combinations that are selected many times (e.g., above a predetermined threshold) during the training process of pairwise co-click signals, and thus may strongly suggest or indicate a relationship between the two images or other types of content. Conversely, low-scoring negative pairs are dot-product combinations that are selected less frequently (e.g., below a predetermined threshold) during training, and therefore may strongly suggest or indicate a weak or no relationship between the images or other content. By comparing the high-scoring positive pairs with the low-scoring negative pairs, related content that is relatively highly correlated (e.g., may be provided as recommended content according to aspects described herein) can be determined.

[0024] The objective of contrastive loss training is to highlight the signal strength of positive pairs over multiple learning iterations compared to a large dataset of low-scoring negative pairs. In this way, after the learning process converges, contrastive loss training can generate high-scoring positive pairs that are highly relevant to the initial search query. Conversely, a large dataset of negative pairs may generate low scores indicating that the pairs are not relevant to the initial search query. By contrasting the differences between high-scoring and low-scoring related content pairs, contrastive loss with pairwise co-click signals optimizes user engagement when training a related content model for the system. To avoid noise in the co-click signal responses, one or more post-processing techniques can be employed to filter out response pairs. Post-processing techniques may include, but are not limited to, avoiding or omitting adult content, evaluating soliciting user input (from one or more human judges), and utilizing thresholds so that co-click pairs selected above or below a predetermined threshold (e.g., as positive or negative pairs) can be used for processing according to aspects described herein.

[0025] As an example, let (i,j) denote the positive pair collected from the co-click signal,

number

number

[0026] The related content recommendation model is generated by the related content model generator 150. The trained model parameters generated by the dot product training engine 148 can be utilized by the related content model generator 150 to generate candidate models and calculate feature vectors. The feature vectors can be calculated using the trained model parameters in combination with a neural network architecture. The related content recommendation engine 130 can apply the calculated feature vectors to a ranking algorithm to recommend related content to a user in response to a search. The resulting related content recommendation model can be usable by the related content recommendation engine 130 in accordance with aspects described herein. High-scoring positive pairs from the training can serve as templates for feature vectors for the model to apply to a user's search query when generating related content recommendations. Similarly, low-scoring negative pairs from the training can serve as templates for feature vectors for the model to ignore or remove in response to a user's search query when generating related content recommendations. In other words, after the learning process converges, the related content model generated by the related content model generator 150 can generate a high score if the piece of content is highly relevant to a given search query, while generating a low score if the piece of content is not highly relevant to the search query.

[0027] The related content model is evaluated by the related content model evaluator 152. Generally, the related content model is evaluated by applying the model to a user's search query (which may be received, for example, from the user device 102), generating a ground truth prediction of the user's next click for the search query, recording the results of the user's next click, and comparing the results to the prediction to determine if the prediction was correct. Both offline and / or online metrics may be used to evaluate the effectiveness of the trained related content model.

[0028] Examples of offline metrics that can be employed alone or in combination with other offline or online metrics include, but are not limited to, precision, top-K search recall, fallout, F-score, average precision, defect rate, precision at k, R-precision, average precision, and / or depreciated cumulative gain. For example, top-K search recall can be employed to reflect how many ground truth targets the relevance model correctly predicted among the top-K ranked candidates when calculated for a moderate-sized dataset. In this example, a typical selection for the dataset size is 1 million relevant content pairs, with K ranging from 5 to 200. Furthermore, after completing the evaluation using top-K search recall and obtaining a reasonable related content model, a human-based evaluation metric known as defect rate can be applied. The defect rate tracks the number of defective items in the top recommendation results from the related content model based on the opinions of trained human raters. The size of the top results considered can vary based on the amount of data evaluated. The defect rate attempts to provide a very close approximation to the true performance of the trained model, unbiased by noise in the co-click signal.

[0029] Examples of online metrics that can be employed alone or in combination with other offline or online metrics include, but are not limited to, click-through rate, online A / B testing, session success rate, and / or zero-result rate. For example, after obtaining two offline metrics for the model described above, an online A / B test can be conducted on the related content model. During this test, two flights are constructed. The first, flight A, is a control flight that generates results for search queries using the basic methodology of the current production model. The second, flight B, is a treatment flight that generates results for search queries using the best related content model selected from the offline metric benchmark. Approximately equal traffic is directed between the two flights, and the A / B test run is conducted for a certain period of time. Upon completion of the test, a scorecard is generated showing the performance difference between control flight A and treatment flight B. The scorecard may include various performance metrics comparing the two flights, including, but not limited to, click-through rate, the percentage of unique users who clicked on related content, and the number of clicks per unique user. The scorecard results can highlight model inefficiencies that may need to be addressed by validating the related content model or through subsequent training. The evaluated related content models can then be utilized to make related content recommendations in response to user search queries over a network 120 that communicates the recommended related content to the user device 102 .

[0030] It should be understood that the various methods, devices, applications, nodes, functions, etc. described with respect to FIG. 1 or any of the figures described herein are not intended to limit the system to being performed by the particular applications and functions described. For example, multiple user devices and / or related content recommendation engines may be used. As another example, the related content recommendation engine 130 may generate content recommendations for another computing device (e.g., as may be requested and / or provided by an application programming interface (API), among other examples). Accordingly, additional components may be used to practice the methods and systems herein, and / or described functions and applications may be excluded without departing from the methods and systems disclosed herein.

[0031] FIG. 2 illustrates an example method 200 for recommending related content based on a user's search query. The general sequence of operations of method 200 is shown in FIG. 2. Generally, method 200 begins at start operation 202 and ends at end operation 218. Method 200 may include more or fewer steps, or the order of steps may be arranged differently than shown in FIG. 2. Method 200 may be implemented as computer-executable instructions executed by a computer system, or may be encoded or stored on a computer-readable medium. Furthermore, method 200 may be performed by gates or circuits associated with a processor, ASIC, FPGA, SOC, or other hardware device. Method 200 will be described below with respect to the systems, components, devices, modules, software, data structures, data characteristic representations, signal diagrams, methods, etc., described with respect to FIGS. 1, 3, 4, 5, 6A, 6B, and 7.

[0032] Following start operation 202, method 200 continues with generate operation 204, in which search query features are generated from input data (which may be based on a search query, for example, received from a user device, such as user device 102). This generate operation is the first step in the search query tower of the two-tower cascaded MLP model (e.g., operations 204, 206, and 208). The input data may be a high-dimensional dataset that may utilize preprocessing to discover low-dimensional features that capture the underlying structure of the search query. In examples, preprocessing may also utilize feature scaling, centering, and / or dimensionality reduction. Once preprocessing is complete, the features are collected and concatenated into a search query combined feature vector.

[0033] In operation 206, the search query is modeled as a cascaded MLP (e.g., as may be generated by search query cascaded MLP 134). At each layer of the MLP, the input may be expanded or contracted. Data scaling may occur between MLP layers, including, but not limited to, data normalization (e.g., batch normalization, activation, dropout) and / or data standardization. Cascaded MLPs may also employ skip connections between the initial input and other subsequent cascaded MLP layers, allowing input gradients to be passed directly to subsequent layers without being processed by the previous MLP layer.

[0034] Flow proceeds to operation 208 where the search query vector is embedded. For example, using the output of a cascaded MLP model (e.g., generated in operation 206), the embedding operation can convert the input feature vector into a learned representation embedding specialized for subsequent related content recommendations. Flow then proceeds to operation 216, discussed below.

[0035] A generate operation 210 generates features of related content from the input data. As shown, operation 210 is the first operation in the related content tower (e.g., comprised of operations 210, 212, and 214) of the two-tower cascaded MLP model. The input data may be a high-dimensional dataset (e.g., based on a search query that may be received from a user device, such as user device 102) that may utilize preprocessing to discover low-dimensional features that capture the underlying structure of the input data. In examples, preprocessing may also utilize scaling, centering, and / or dimensionality reduction of the features. Once preprocessing is complete, the features are collected and may be concatenated into a related content joint feature vector.

[0036] In operation 212, the related content (e.g., as may be generated by related content cascaded MLP 140) is modeled as a cascaded MLP. At each layer of the MLP, the input may be expanded or contracted. Data scaling may occur between MLP layers, including, but not limited to, data normalization (e.g., batch normalization, activation, dropout) and / or data standardization. Cascaded MLPs may also employ skip connections between the initial input and other subsequent cascaded MLP layers, allowing input gradients to be passed directly to subsequent layers without being processed by the previous MLP layer.

[0037] Flow continues to operation 214 where relevant content vectors are embedded. Using the output of the cascaded MLP model (e.g., as generated in operation 212), the embedding operation converts the input feature vectors into learned representation embeddings specialized for subsequent relevant content recommendations.

[0038] At operation 216, a dot product is generated from the search query vector (e.g., generated at operation 208) and the related content vector (e.g., generated at operation 214). Although method 200 illustrates operations 204, 206, and 208 as occurring simultaneously with operations 210, 212, and 214, it will be understood that in other examples, at least some of the operations may occur sequentially or according to any of a variety of other orders. The dot product may be a single, untrained combination of the search query and the related content. In an example, the dot product may have multiple related content pieces that are provided in combination with the search query as related content recommendations for the initial search.

[0039] At operation 218, the dot product is trained (e.g., as may be performed by the dot product training engine 148 described above with respect to FIG. 1). For example, a supervised control loss may be employed to train the model by tracking the results of pairwise co-click signals of user-selected related content derived from untrained dot products. The pairwise co-click process may include contrasting user-selected high-scoring positive dot product pairs with a large dataset of low-scoring negative dot product pairs. The result is recommendations of highly correlated related content in response to a search query.

[0040] Flow continues to operation 220, where a related content model is generated based on the results of the supervised training of operation 218. For example, the set of high-scoring positive pairs can become templates for feature vectors of the model to apply to a user's search query when generating related content recommendations. Similarly, the set of low-scoring negative pairs can become templates for feature vectors of the model to ignore or remove in response to a user's search query when generating related content recommendations. The resulting related content recommendation model can be usable in accordance with aspects described herein (e.g., by a related content recommendation engine, such as related content recommendation engine 130 described above with respect to FIG. 1).

[0041] At operation 222, the related content model is evaluated. For example, the related content model is evaluated by applying the model to the user's search query, generating a ground truth next click prediction for the search query, recording the results of the user's next click (e.g., positive and / or negative interactions with the recommended content), and / or comparing the results to the prediction to determine if the prediction was correct. Both offline and / or online metrics may be used to evaluate the effectiveness of the trained related content model. Method 200 ends at operation 224.

[0042] Method 200, as described above, begins with operation 202 and continues with operation 204, although operation 202 may be followed by either generate operation 204 or generate operation 210. Method 200 includes a two-tower cascaded MLP, represented by steps 204 through 214. Steps 204 through 208 represent a search query tower, while steps 210 through 214 represent a related content tower. The steps of either tower may be performed sequentially, as described above, and / or in any other order that maintains the functionality of each tower and a linear input-to-output flow. In this regard, it is contemplated that the steps of each tower may be performed simultaneously, such that steps 204 and 210 may be performed simultaneously, as well as steps 208 and 214. Similarly, these steps may be performed with step 210 performed first, followed sequentially through step 214, then steps 204 through 208, or in some other order based on user preference.

[0043] FIG. 3 illustrates an example of a method 300 for modeling a combined feature vector using a single cascaded MLP neural network. The general sequence of operations for method 300 is shown in FIG. 3. Generally, method 300 begins with a start operation 302 and ends with an end operation 318. Method 300 may include more or fewer steps, or the order of steps may be arranged differently than shown in FIG. 3. Method 300 may be implemented as a single-tower cascaded MLP or may be utilized in combination with other cascaded MLPs in a multi-tower model as described above. In examples, aspects of method 300 may be performed as part of operation 206 and / or operation 212 described above with respect to method 200 of FIG. 2. Method 300 may be implemented as computer-executable instructions executed by a computer system and may be encoded or stored on a computer-readable medium. Additionally, method 300 may be performed by gates or circuits associated with a processor, ASIC, FPGA, SOC, or other hardware device. The method 300 will be described below with respect to the systems, components, devices, modules, software, data structures, data characteristic representations, signal diagrams, methods, etc. described with respect to Figures 1, 2, 4, 5, 6A, 6B, and 7.

[0044] Following start operation 302, method 300 continues with operation 304, which involves obtaining a combined feature vector for either the search query or the related content tower. The combined feature vector may be the result of preprocessing the raw input dataset to discover low-dimensional features that can be applied to a feed-forward MLP neural network to discover feature embeddings.

[0045] In operation 306, MLP layers in a cascaded MLP neural network are generated. For example, at each layer of a cascaded MLP, the input can be expanded or contracted. When an input is expanded, complexity is added to the input and output to the next layer. When an input is contracted, the input is condensed so that the final output is a vector of a reasonable size for use in search.

[0046] Flow continues to operation 308, where one or more data scaling operations are performed on the output of the previous MLP layer. A data scaling method, such as data normalization (e.g., batch normalization, activation, dropout) or data standardization, is applied to the output of each layer of the cascaded MLP in preparation for input to the next cascaded MLP layer.

[0047] At decision 310, the method determines whether to insert skip connections. Skip connections may be utilized to facilitate direct gradient backpropagation within a neural network. The benefit of introducing skip connections is that early MLP layers of a deep neural network can be trained more directly using the loss gradients of later MLP layers, thereby reducing the time to parameter convergence. One or more skip connections may be introduced based on design characteristics and / or scenario complexity. For example, decision 310 may include deciding to insert skip connections when the number of cascaded MLP layers increases beyond a predetermined threshold (e.g., 20 layers or more), thereby reducing the time to parameter convergence. If it is determined not to insert skip connections, the method branches "no" and proceeds to decision 314, discussed below.

[0048] Alternatively, if it is determined that a skip connection should be inserted, the method branches "yes" and proceeds to operation 312, where a skip connection operation is performed. At operation 312, the output from operation 308 and the output from operation 306 (e.g., as indicated by dashed arrow 320) may be combined (e.g., using element-wise addition) to create an output vector. The output vector may be used as an input to a subsequent MLP layer in operation 306. Flow then proceeds to decision 314.

[0049] At decision 314, it is determined whether additional MLP layers should be generated. If a more complex model is desired, additional MLP layers can be generated. In instances where there is relatively high variability in user behavior and training data, additional MLP layers can be generated to address such complexity. As an example, the more complex a scenario is, the more likely it is that a deeper neural network (e.g., with additional cascaded MLP layers) will be required to provide accurate recommendations. Conversely, in situations where the complexity of the scenario is relatively low, a shallower neural network with fewer cascaded MLP layers may be utilized and can still provide accurate recommendations in such scenarios.

[0050] If it is determined that additional MLP layers should be generated, the method branches "yes" and returns to operation 306, which generates additional MLP layers as described above. If no skip connections are inserted, the input to operation 306 can be the output from operation 308. Alternatively, if skip connections are inserted, the input to operation 306 can be the output of the element-wise sum from operation 312. If it is decided not to generate additional MLP layers, flow instead branches "no," the training of the cascaded MLP is complete, and the method proceeds to output operation 316, which outputs the trained combined feature vector for the search query or related content resulting from the cascaded MLP. Method 300 ends at end operation 318.

[0051] FIG. 4 illustrates an example of a user interface for a search query result with recommended related content. User interface 400 is an example of content that can be presented to a user as a web interface for related content (e.g., during a pairwise co-click training session or when a user is searching for content). In this example, the initial search query was an image of a dog. Center image 402 may be content responsive to the user's initial search query. Side images 404, 406, and 408 are candidate related content recommendations (e.g., potentially generated by a model as the other half of a dot product, e.g., according to aspects described herein). Each of images 402, 404, 406, and / or 408 may be selected by the user. If the user selects a side image of related content, e.g., image 406, the selected image can be used as a positive pair, indicating a potential positive relationship between the center image and the selected related content. Conversely, the remaining non-selected images (e.g., images 404 and 408) can be used as negative pairs, indicating that a positive relationship is unlikely to exist between the center image and the non-selected content. Over multiple iterations (e.g., by the same user or across multiple users), control loss supervised training can generate a result set of high-scoring positive pairs and low-scoring negative pairs based on user selections. By comparing such high-scoring positive pairs and low-scoring negative pairs, highly correlated related content can be determined in accordance with aspects described herein.

[0052] 5 is a block diagram illustrating an example of the physical components of a computing device in which aspects of the present disclosure may be practiced. FIG. 5 is a block diagram illustrating the physical components (e.g., hardware) of a computing device 500 in which aspects of the present disclosure may be practiced. The computing device components described below may be suitable for the computing devices described above, including user device 102, as well as one or more of the devices discussed above with respect to FIG. 1. In a basic configuration, computing device 500 may include at least one processing unit 502 and system memory 504. Depending on the configuration and type of computing device, system memory 504 may include, but is not limited to, volatile storage (e.g., random access memory), non-volatile storage (e.g., read-only memory), flash memory, or any combination of such memory.

[0053] The system memory 504 may include an operating system 505 and one or more program modules 506 suitable for executing software applications 520, such as one or more components supported by the system described herein. By way of example, the system memory 504 may store a dot product generator 524 and an associated content model generator 526. The operating system 505 may be suitable for controlling the operation of the computing device 500, for example.

[0054] Furthermore, embodiments of the present disclosure may be practiced with graphics libraries, other operating systems, or any other application programs and are not limited to any particular application or system. This basic configuration is illustrated in FIG. 5 by the components within dashed line 508. Computing device 500 may have additional features or functionality. For example, computing device 500 may also include additional data storage devices (removable and / or non-removable), such as, for example, magnetic disks, optical disks, or tape. Such additional storage is illustrated in FIG. 5 by removable storage 509 and non-removable storage 510.

[0055] As mentioned above, several program modules and data files may be stored within the system memory 504. While executing on the processing unit 502, the program modules 506 (e.g., applications 520) may perform processes including, but not limited to, the aspects described herein. Other program modules that may be used in accordance with aspects of the present disclosure may include email and contact applications, word processing applications, spreadsheet applications, database applications, slide presentation applications, drawing or computer-aided application programs, etc.

[0056] Furthermore, embodiments of the present disclosure may be practiced in electrical circuits including discrete electronic elements, packaged or integrated electronic chips including logic gates, circuits utilizing a microprocessor, or on a single chip including electronic elements or a microprocessor. For example, embodiments of the present disclosure may be practiced by a system-on-chip (SOC), in which each or many of the components shown in FIG. 5 may be integrated onto a single integrated circuit. Such an SOC device may include one or more processing units, graphics units, communications units, system virtualization units, and various application functions, all of which are integrated (or "burned") onto the chip substrate as a single integrated circuit. When operated by a SOC, the functionality described herein with respect to the client's ability to switch protocols may be operated by application-specific logic integrated with other components of the computing device 500 on a single integrated circuit (chip). Embodiments of the present disclosure may also be practiced using other technologies capable of performing logical operations such as AND, OR, and NOT, including, but not limited to, mechanical, optical, fluidic, and quantum technologies. Additionally, embodiments of the present disclosure may be practiced within a general-purpose computer or any other circuit or system.

[0057] The computing device 500 may also have one or more input devices 512, such as a keyboard, mouse, pen, voice or audio input device, touch or swipe input device, etc. Output devices 514, such as a display, speakers, printer, etc., may also be included. The aforementioned devices are examples, and other devices may be used. The computing device 500 may include one or more communication connections 516 that enable communication with other computing devices 550. Examples of suitable communication connections 516 include, but are not limited to, radio frequency (RF) transmitter, receiver, and / or transceiver circuitry, universal serial bus (USB), parallel and / or serial ports.

[0058] As used herein, the term computer-readable medium may include computer storage media. Computer storage media may include volatile and nonvolatile, removable and non-removable media implemented by any method or technology for storage of information, such as computer-readable instructions, data structures, or program modules. System memory 504, removable storage 509, and non-removable storage 510 are all examples of computer storage media (e.g., memory storage). Computer storage media may include RAM, ROM, electrically erasable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other article of manufacture that can be used to store information and that can be accessed by computing device 500. Any such computer storage media may be part of computing device 500. Computer storage media do not include carrier waves or other propagated or modulated data signals.

[0059] Communication media may be embodied by computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term "modulated data signal" may refer to a signal that has one or more characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media may include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, radio frequency (RF), infrared and other wireless media.

[0060] 6A and 6B illustrate a mobile computing device 600, such as a mobile phone, a smartphone, a wearable computer (such as a smart watch), a tablet computer, a laptop computer, etc., on which embodiments of the present disclosure may be practiced. In some aspects, a client may be a mobile computing device. Referring to FIG. 6A, one aspect of a mobile computing device 600 for implementing aspects is shown. In a basic configuration, the mobile computing device 600 is a handheld device having both input and output elements. The mobile computing device 600 typically includes a display 605 and input buttons 610 that allow a user to input information into the mobile computing device 600. The display 605 of the mobile computing device 600 may also function as an input device (e.g., a touchscreen display).

[0061] If included, an optional side input element 615 allows for further user input. The side input element 615 may be a rotary switch, a button, or any other type of manual input element. In alternative aspects, the mobile computing device 600 may incorporate more or fewer input elements. For example, the display 605 may not be a touchscreen in some embodiments. In another example, the mobile computing device 600 may also include an optional keypad (not shown), which may be a physical keypad or a "soft" keypad generated on a touchscreen display.

[0062] In various embodiments, the output elements include a display 605 for displaying a graphical user interface (GUI), a visual indicator 620 (e.g., a light emitting diode), and / or an audio transducer 625 (e.g., a speaker). In some aspects, the mobile computing device 600 incorporates a vibration transducer for providing tactile feedback to the user. In yet other aspects, the mobile computing device 600 incorporates input and / or output ports, such as an audio input (e.g., a microphone jack), an audio output (e.g., a headphone jack), and a video output (e.g., an HDMI port) for transmitting signals to an external source or receiving signals from an external device.

[0063] 6B is a block diagram illustrating the architecture of one aspect of a mobile computing device. That is, the mobile computing device 600 can incorporate a system (e.g., architecture) 602 for implementing some aspects. In one embodiment, the system 602 is implemented as a "smartphone" capable of running one or more applications (e.g., a browser, email, calendar, contact manager, messaging client, games, and media client / player). In some aspects, the system 602 is integrated into computing devices such as an integrated personal digital assistant (PDA) and wireless telephone.

[0064] One or more application programs 666 may be loaded into memory 662 and execute on or in association with operating system 664. Examples of application programs include a phone dialer program, an email program, a personal information manager (PIM) program, a word processing program, a spreadsheet program, an internet browser program, a messaging program, etc. System 602 includes a non-volatile storage area 668 within memory 662. Non-volatile storage area 668 may be used to store persistent information that should not be lost even if system 602 is powered down. Application programs 666 may use and store information in non-volatile storage area 668, such as emails or other messages used by an email application. A synchronization application (not shown) is also present on system 602 and is programmed to interact with a corresponding synchronization application on a host computer to keep information stored in non-volatile storage area 668 synchronized with corresponding information stored at the host computer. As should be understood, other applications may be loaded into memory 662 and executed on the mobile computing device 600 described herein.

[0065] The system 602 includes a power supply 670, which may be implemented as one or more batteries. The power supply 670 may also include an external power source, such as an AC adapter or a powered base, that supplements or charges the batteries.

[0066] The system 602 may include a radio interface layer 672 that performs the function of transmitting and receiving radio frequency communications. The radio interface layer 672 facilitates wireless connectivity between the system 602 and the "outside world" via a communications carrier or service provider. Transmissions to and from the radio interface layer 672 occur under the control of the operating system 664. In other words, communications received by the radio interface layer 672 may be disseminated by the operating system 664 to the application programs 666, and vice versa.

[0067] The visual indicator 620 can be used to provide a visual notification, and / or the audio interface 674 can be used to produce an audible notification via the audio transducer 625. In the illustrated embodiment, the visual indicator 620 is a light-emitting diode (LED) and the audio transducer 625 is a speaker. These devices can be directly coupled to the power source 670 to remain on at startup for a period of time dictated by the notification mechanism, even though the processor 660 and other components may shut down to conserve battery power. The LED can be programmed to remain on indefinitely to indicate a powered-on state of the device until the user takes action. The audio interface 674 is used to provide audible signals to and receive audible signals from the user. For example, in addition to being coupled to the audio transducer 625, the audio interface 674 can also be coupled to a microphone to receive audible input to facilitate telephone conversations. According to embodiments of the present disclosure, the microphone can also function as an audio sensor to facilitate control of notifications, as described below. The system 602 may further include a video interface 676 that enables operation of the on-board camera 630 for recording still images, video streams, and the like.

[0068] Mobile computing device 600 implementing system 602 may have additional features or functionality. For example, mobile computing device 600 may also include additional data storage devices (removable and / or non-removable) such as magnetic disks, optical disks, or tape. Such additional storage is illustrated in FIG. 6B by non-volatile storage area 668.

[0069] The data / information generated or captured by the mobile computing device 600 and stored by the system 602 may be stored locally on the mobile computing device 600, as described above, or the data may be stored on any number of storage media that may be accessed by the device via the wireless interface layer 672 or via a wired connection between the mobile computing device 600 and another computing device associated with the mobile computing device 600, such as a server computer in a distributed computing network such as the Internet. As should be understood, such data / information may be accessed by the mobile computing device 600 via the wireless interface layer 672 or via a distributed computing network. Similarly, such data / information may be readily transferred between computing devices for storage and use in accordance with known data / information transfer and storage means, including email and collaborative data / information sharing systems.

[0070] 7 illustrates one aspect of a system architecture for processing data received at a computing system from a remote source, such as a personal computer 704, a tablet computing device 706, or a mobile computing device 708, as described above. Content displayed at the server device 702 may be stored via different communication channels or other storage types. For example, various documents may be stored using a directory service 722, a web portal 724, a mailbox service 726, an instant messaging store 728, or a social networking site 730.

[0071] The application 720 may be employed by a client in communication with the server device 702, and / or the related content recommendation engine 721 may be employed by the server device 702. The server device 702 may provide data to and from client computing devices, such as a personal computer 704, a tablet computing device 706, and / or a mobile computing device 708 (e.g., a smartphone), over a network 715. By way of example, the computing system described above may be embodied by a personal computer 704, a tablet computing device 706, and / or a mobile computing device 708 (e.g., a smartphone). Any of these implementations of a computing device may obtain content from the store 716 in addition to receiving graphical data that can be used for pre-processing in a graphics generating system or post-processing in a receiving computing system.

[0072] In examples, aspects and functionality described herein may operate on a distributed system (e.g., a cloud-based computing system) in which application functionality, memory, data storage, and search and various processing functions may operate remotely from one another over a distributed computing network such as the Internet or an intranet. User interfaces and various types of information may be displayed by on-board computing device displays or by remote display units associated with one or more computing devices. For example, user interfaces and various types of information may be displayed and interacted with on a wall surface onto which the user interfaces and various types of information are projected. Interactions with multiple computing systems in which implementations of the present invention may be practiced include keystroke input, touchscreen input, voice or other audio input, gesture input, etc., if the associated computing device has detection (e.g., camera) capabilities for capturing and interpreting user gestures to control computing device functions.

[0073] As can be understood from the above disclosure, one aspect of the present technology relates to a system including at least one processor and a memory storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations. The set of operations includes obtaining a search query; generating a set of features for a search query tower and a related content tower, where the search query tower and the related content tower are based on the obtained search query, respectively; training the search query tower and the related content tower; generating trained feature vectors for each of the search query tower and the related content tower; training a dot product based on the search query trained feature vector and the related content trained feature vector; and generating a related content model based on the trained dot product for generating a set of recommended content based on the user's search query. In one example, training the search query tower and the related content tower further includes utilizing a cascaded multilayer perceptron model composed of multiple layers. In another example, the layers of the cascaded multilayer perceptron model include one or more expansion layers or contraction layers, and data scaling is performed between the multilayer perceptron layers. In a further example, the data scaling includes one or more of data standardization or data normalization including batch normalization, activation, and dropout. In yet another example, a skip connection is utilized between the initial input layer and another multilayer perceptron layer of the cascaded multilayer perceptron model. In yet a further example, generating the set of features further includes preprocessing the search query to generate a set of low-dimensional features using at least one of feature scaling, centering, or dimensionality reduction. In another example, the dot product is trained using self-supervised representation learning employing a contrastive loss by tracking pairwise co-click signals associated with multiple search queries.

[0074] In another aspect, the technology relates to a method including receiving a search request including a search query for content; using a related content model including a two-tower cascaded multilayer perceptron model to generate a set of recommended content related to both the search query and an instance of content responsive to the search query; and providing the generated set of recommended content related to the instance of content responsive to the search query in response to the search request. In one example, the instance of content is a first instance of content, and the related content model is trained using a set of co-click signals including a positive pair between a second instance of content and the first instance of recommended content and a negative pair between the second instance of content and the second instance of recommended content. In another example, a first tower of the two-tower cascaded multilayer perceptron model is related to a first content type, and a second tower of the two-tower cascaded multilayer perceptron model is related to a second content type. In a further example, the first content type is a text content type related to the search query, and the second content type is an image content type. In yet another example, the search request further includes an indication of the instances of content responsive to the search query. In yet a further example, the method further includes identifying the instances of content responsive to the search query.

[0075] In a further aspect, the technology relates to another method including obtaining a search query; generating a set of features for a search query tower and a related content tower, where the search query tower and the related content tower are based on the obtained search query, respectively; training the search query tower and the related content tower; generating a trained feature vector for each of the search query tower and the related content tower; training a dot product based on the search query trained feature vector and the related content trained feature vector; and generating a related content model based on the trained dot product for generating a set of recommended content based on the user's search query. In one example, training the search query tower and the related content tower further includes utilizing a cascaded multilayer perceptron model composed of multiple layers. In another example, the layers of the cascaded multilayer perceptron model include one or more expansion layers or contraction layers, and data scaling is performed between the multilayer perceptron layers. In a further example, the data scaling includes one or more of data standardization or data normalization, including batch normalization, activation, and dropout. In yet another example, a skip connection is utilized between the initial input layer and another multilayer perceptron layer of the cascaded multilayer perceptron model. In yet a further example, generating the set of features further includes preprocessing the search query to generate a set of low-dimensional features using at least one of feature scaling, centering, or dimensionality reduction. In another example, the dot product is trained using self-supervised representation learning employing a contrastive loss by tracking pairwise co-click signals associated with multiple search queries.

[0076] Aspects of the present disclosure are described above with reference to block diagrams and / or operational illustrations of, for example, methods, systems, and computer program products according to aspects of the present disclosure. The functions / acts noted in the blocks may occur out of the order noted in any flowchart. For example, two blocks shown in succession may in fact be executed substantially concurrently, or the blocks may be executed in the reverse order, depending on the functions / acts involved.

[0077] The description and illustrations of one or more aspects set forth herein are not intended to in any way limit or restrict the scope of the present disclosure as set forth in the claims. The aspects, examples, and details provided herein are believed to be sufficient to transfer ownership and enable others to make and use aspects of the present disclosure as set forth in the claims. The present disclosure as set forth in the claims should not be construed as limited to any aspect, example, or detail provided herein. It is intended that various features (both structural and methodological) be selectively included or omitted to create embodiments having a particular set of characteristics, whether shown and described in combination or separately. Having provided the description and illustrations herein, those skilled in the art will be able to devise variations, modifications, and alternative embodiments within the spirit of the broader aspects of the general inventive concepts embodied herein without departing from the broader scope of the present disclosure as set forth in the claims.

Claims

1. at least one processor; a memory storing instructions that, when executed by said at least one processor, cause the system to perform a set of operations; wherein the set of operations comprises: Obtaining a search query; generating a set of features for a search query tower and a related content tower, the search query tower and the related content tower being based on the obtained search query, respectively; training the search query tower and the related content tower; generating a trained feature vector for each of the search query tower and the related content tower; training a dot product based on the search query trained feature vector and the related content trained feature vector; generating a related content model based on the trained dot product to generate a set of recommended content based on a user's search query; Including, the system.

2. 2. The system of claim 1, wherein generating the set of features further comprises preprocessing the search query to generate a set of reduced-dimensional features using at least one of feature scaling, centering, or dimensionality reduction.

3. The system of claim 1 , wherein the dot product is trained using self-supervised representation learning employing contrastive loss by tracking pairwise co-click signals associated with multiple search queries.

4. receiving a search request including a search query for content; generating a set of recommended content that is related to both the search query and the instances of content responsive to the search query using a relevant content model that includes a two-tower cascaded multi-layer perceptron model; responding to the search request, providing the generated set of recommended content related to the instances of content responsive to the search query; and A method comprising:

5. the instance of content is a first instance of content; The related content model is a positive pair between the second instance of the content and the first instance of the recommended content; and a negative pair between the second instance of the content and a second instance of the recommended content; is trained using a set of co-click signals including The method of claim 4.

6. 5. The method of claim 4, wherein a first tower of the two-tower cascaded multi-layer perceptron model is associated with a first content type and a second tower of the two-tower cascaded multi-layer perceptron model is associated with a second content type.

7. Obtaining a search query; generating a set of features for a search query tower and a related content tower, the search query tower and the related content tower being based on the obtained search query, respectively; training the search query tower and the related content tower; generating a trained feature vector for each of the search query tower and the related content tower; training a dot product based on the search query trained feature vector and the related content trained feature vector; generating a related content model based on the trained dot product to generate a set of recommended content based on a user's search query; A method comprising:

8. The method of claim 7 , wherein training the search query tower and related content tower further comprises utilizing a cascaded multi-layer perceptron model configured in multiple layers.

9. The method of claim 8 , wherein the layers of the cascaded multi-layer perceptron model include one or more of an expansion layer or a contraction layer, and data scaling is performed between the multi-layer perceptron layers.

10. The method of claim 9 , wherein the data scaling comprises one or more of data standardization or data normalization including batch normalization, activation, and dropout.

11. The system of claim 1 , wherein training the search query tower and related content tower further comprises utilizing a cascaded multi-layer perceptron model configured in multiple layers.

12. The system of claim 11 , wherein the layers of the cascaded multi-layer perceptron model include one or more of an expansion layer or a contraction layer, and data scaling is performed between the multi-layer perceptron layers.

13. The method of claim 6 , wherein the first content type is a text content type related to the search query and the second content type is an image content type.

14. The method of claim 8 , wherein a skip connection is utilized between an initial input layer and another multi-layer perceptron layer of the cascaded multi-layer perceptron model.

15. The method of claim 7 , wherein the dot product is trained using self-supervised representation learning employing contrastive loss by tracking pairwise co-click signals associated with multiple search queries.

Images