Transformer-based hybrid recommendation model with contextual feature support
In a computer-implemented embodiment, an interaction machine learning model is trained based on many interactions on many resources. A context lexical token is inferred that represents a current operational context of a user. The context lexical token is inserted into a sequence of other inferred lexical tokens. From the context lexical token within the sequence of tokens, the interaction machine learning model infers a predicted resource that will be accessed next. In an embodiment, accelerated matchmaking entails suitability measurement by a dot product of a) a dynamically inferred user embedding that is based on the context lexical token and b) a statically inferred item embedding.
1 . A method comprising:
training, based on a plurality of interactions on a plurality of resources, an interaction large language model (LLM);
generating a context lexical token that represents an operational context, wherein the operational context occurs after said training the interaction LLM; and
inferring by the interaction LLM, from an input, a predicted resource of said plurality of resources that will be accessed next, wherein the input comprises the context lexical token, a contiguous plurality of padding tokens, a temporal sequence of interaction lexical tokens that each represents a respective distinct interaction that was already caused by a particular user, and a lexical token that represents the particular user in the operational context.
2 . The method of claim 1 wherein said inferring the predicted resource that will be accessed is based on a sequence of lexical tokens that comprises:
the context lexical token,
a temporal sequence of interaction lexical tokens that each represents a respective distinct interaction that was already caused by a particular user, and
between the context lexical token and said temporal sequence of interaction lexical tokens, a predefined lexical token that does not represent data.
3 . The method of claim 1 further comprising inferring a particular lexical token that is a fixed-size embedding of an object, wherein the particular lexical token is selected from a group consisting of the context lexical token and an interaction lexical token that represents a respective distinct access of a respective resource by a particular user.
4 . The method of claim 3 further comprising dropout training an artificial neural network to perform said inferring the particular lexical token.
5 . The method of claim 1 further comprising inferring a respective fixed-size embedding of each resource in the plurality of resources.
6 . The method of claim 5 further comprising generating a multidimensional index of said fixed-size embeddings of the plurality of resources.
7 . The method of claim 6 further comprising using the predicted resource that will be accessed as a lookup key to retrieve, from the multidimensional index, a nearest neighbors subset of the fixed-size embeddings of the plurality of resources.
8 . The method of claim 7 wherein:
said inferring the predicted resource that will be accessed comprises generating a fixed-size embedding of the predicted resource;
said using the predicted resource that will be accessed as the lookup key comprises using the fixed-size embedding of the predicted resource as the lookup key.
9 . The method of claim 8 further comprising selecting, filtering, or ranking the nearest neighbors subset of the plurality of resources based on a respective dot product of the fixed-size embedding of the predicted resource and each fixed-size embeddings of the nearest neighbors subset of the fixed-size embeddings of the plurality of resources.
10 . The method of claim 1 wherein said inferring the predicted resource that will be accessed is based on a temporal sequence of interactions that occurred after said plurality of interactions.
11 . One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause:
training, based on a plurality of interactions on a plurality of resources, an interaction large language model (LLM);
generating a context lexical token that represents an operational context, wherein the operational context occurs after said training the interaction LLM; and
inferring by the interaction LLM, from an input, a predicted resource of said plurality of resources that will be accessed next, wherein the input comprises the context lexical token, a contiguous plurality of padding tokens, a temporal sequence of interaction lexical tokens that each represents a respective distinct interaction that was already caused by a particular user, and a lexical token that represents the particular user in the operational context.
12 . The one or more non-transitory computer-readable media of claim 11 wherein said inferring the predicted resource that will be accessed is based on a sequence of lexical tokens that comprises:
the context lexical token,
a temporal sequence of interaction lexical tokens that each represents a respective distinct interaction that was already caused by a particular user, and
between the context lexical token and said temporal sequence of interaction lexical tokens, a predefined lexical token that does not represent data.
13 . The one or more non-transitory computer-readable media of claim 11 wherein:
the instructions further cause inferring a particular lexical token that is a fixed-size embedding of an object;
the particular lexical token is selected from a group consisting of the context lexical token and an interaction lexical token that represents a respective distinct access of a respective resource by a particular user.
14 . The one or more non-transitory computer-readable media of claim 13 wherein the instructions further cause dropout training an artificial neural network to perform said inferring the particular lexical token.
15 . The one or more non-transitory computer-readable media of claim 11 wherein the instructions further cause inferring a respective fixed-size embedding of each resource in the plurality of resources.
16 . The one or more non-transitory computer-readable media of claim 15 wherein the instructions further cause generating a multidimensional index of said fixed-size embeddings of the plurality of resources.
17 . The one or more non-transitory computer-readable media of claim 16 wherein the instructions further cause using the predicted resource that will be accessed as a lookup key to retrieve, from the multidimensional index, a nearest neighbors subset of the fixed-size embeddings of the plurality of resources.
18 . The one or more non-transitory computer-readable media of claim 11 wherein said inferring the predicted resource that will be accessed is based on a temporal sequence of interactions that occurred after said plurality of interactions.